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) 2011, 2012 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 <lustre_mdt.h>
  53 #include "mdt_internal.h"
  54 #ifdef HAVE_QUOTA_SUPPORT
  55 # include <lustre_quota.h>
  56 #endif
  57 #include <lustre_acl.h>
  58 #include <lustre_param.h>
  59 #include <lustre_fsfilt.h>
  60
  61 struct mds_device {
  62         /* super-class */
  63         struct md_device           mds_md_dev;
  64         struct ptlrpc_service     *mds_regular_service;
  65         struct ptlrpc_service     *mds_readpage_service;
  66         struct ptlrpc_service     *mds_setattr_service;
  67         struct ptlrpc_service     *mds_mdsc_service;
  68         struct ptlrpc_service     *mds_mdss_service;
  69         struct ptlrpc_service     *mds_fld_service;
  70 };
  71
  72 /*
  73  *  * Initialized in mdt_mod_init().
  74  *   */
  75 static unsigned long mdt_num_threads;
  76 CFS_MODULE_PARM(mdt_num_threads, "ul", ulong, 0444,
  77                 "number of MDS service threads to start "
  78                 "(deprecated in favor of mds_num_threads)");
  79
  80 static unsigned long mds_num_threads;
  81 CFS_MODULE_PARM(mds_num_threads, "ul", ulong, 0444,
  82                 "number of MDS service threads to start");
  83
  84 static char *mds_num_cpts;
  85 CFS_MODULE_PARM(mds_num_cpts, "c", charp, 0444,
  86                 "CPU partitions MDS threads should run on");
  87
  88 static unsigned long mds_rdpg_num_threads;
  89 CFS_MODULE_PARM(mds_rdpg_num_threads, "ul", ulong, 0444,
  90                 "number of MDS readpage service threads to start");
  91
  92 static char *mds_rdpg_num_cpts;
  93 CFS_MODULE_PARM(mds_rdpg_num_cpts, "c", charp, 0444,
  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 CFS_MODULE_PARM(mds_attr_num_threads, "ul", ulong, 0444,
  99                 "number of MDS setattr service threads to start");
 100
 101 static char *mds_attr_num_cpts;
 102 CFS_MODULE_PARM(mds_attr_num_cpts, "c", charp, 0444,
 103                 "CPU partitions MDS setattr threads should run on");
 104
 105 #define DEFINE_RPC_HANDLER(base, flags, opc, fn, fmt)                   \
 106 [opc - base] = {                                                        \
 107         .mh_name        = #opc,                                         \
 108         .mh_fail_id     = OBD_FAIL_ ## opc ## _NET,                     \
 109         .mh_opc         = opc,                                          \
 110         .mh_flags       = flags,                                        \
 111         .mh_act         = fn,                                           \
 112         .mh_fmt         = fmt                                           \
 113 }
 114
 115 /* Request with a format known in advance */
 116 #define DEF_MDT_HDL(flags, name, fn)                                    \
 117         DEFINE_RPC_HANDLER(MDS_GETATTR, flags, name, fn, &RQF_ ## name)
 118
 119 /* Request with a format we do not yet know */
 120 #define DEF_MDT_HDL_VAR(flags, name, fn)                                \
 121         DEFINE_RPC_HANDLER(MDS_GETATTR, flags, name, fn, NULL)
 122
 123 /* Map one non-standard request format handler.  This should probably get
 124  * a common OBD_SET_INFO RPC opcode instead of this mismatch. */
 125 #define RQF_MDS_SET_INFO RQF_OBD_SET_INFO
 126
 127 static struct mdt_handler mdt_mds_ops[] = {
 128 DEF_MDT_HDL(0,                          MDS_CONNECT,      mdt_connect),
 129 DEF_MDT_HDL(0,                          MDS_DISCONNECT,   mdt_disconnect),
 130 DEF_MDT_HDL(0,                          MDS_SET_INFO,     mdt_set_info),
 131 DEF_MDT_HDL(0,                          MDS_GET_INFO,     mdt_get_info),
 132 DEF_MDT_HDL(0           | HABEO_REFERO, MDS_GETSTATUS,    mdt_getstatus),
 133 DEF_MDT_HDL(HABEO_CORPUS,               MDS_GETATTR,      mdt_getattr),
 134 DEF_MDT_HDL(HABEO_CORPUS| HABEO_REFERO, MDS_GETATTR_NAME, mdt_getattr_name),
 135 DEF_MDT_HDL(HABEO_CORPUS,               MDS_GETXATTR,     mdt_getxattr),
 136 DEF_MDT_HDL(0           | HABEO_REFERO, MDS_STATFS,       mdt_statfs),
 137 DEF_MDT_HDL(0           | MUTABOR,      MDS_REINT,        mdt_reint),
 138 DEF_MDT_HDL(HABEO_CORPUS,               MDS_CLOSE,        mdt_close),
 139 DEF_MDT_HDL(HABEO_CORPUS,               MDS_DONE_WRITING, mdt_done_writing),
 140 DEF_MDT_HDL(0           | HABEO_REFERO, MDS_PIN,          mdt_pin),
 141 DEF_MDT_HDL_VAR(0,                      MDS_SYNC,         mdt_sync),
 142 DEF_MDT_HDL(HABEO_CORPUS| HABEO_REFERO, MDS_IS_SUBDIR,    mdt_is_subdir),
 143 DEF_MDT_HDL(0,                          MDS_QUOTACHECK,   mdt_quotacheck),
 144 DEF_MDT_HDL(0,                          MDS_QUOTACTL,     mdt_quotactl),
 145 DEF_MDT_HDL(0           | HABEO_REFERO, MDS_HSM_PROGRESS, mdt_hsm_progress),
 146 DEF_MDT_HDL(0           | HABEO_REFERO, MDS_HSM_CT_REGISTER,
 147                                                 mdt_hsm_ct_register),
 148 DEF_MDT_HDL(0           | HABEO_REFERO, MDS_HSM_CT_UNREGISTER,
 149                                                 mdt_hsm_ct_unregister),
 150 DEF_MDT_HDL(HABEO_CORPUS| HABEO_REFERO, MDS_HSM_STATE_GET,
 151                                                 mdt_hsm_state_get),
 152 DEF_MDT_HDL(HABEO_CORPUS| HABEO_REFERO, MDS_HSM_STATE_SET,
 153                                                 mdt_hsm_state_set),
 154 DEF_MDT_HDL(HABEO_CORPUS| HABEO_REFERO, MDS_HSM_ACTION,
 155                                                 mdt_hsm_action),
 156 };
 157
 158 #define DEF_OBD_HDL(flags, name, fn)                                    \
 159         DEFINE_RPC_HANDLER(OBD_PING, flags, name, fn, NULL)
 160
 161 static struct mdt_handler mdt_obd_ops[] = {
 162 DEF_OBD_HDL(0,                          OBD_PING,         mdt_obd_ping),
 163 DEF_OBD_HDL(0,                          OBD_LOG_CANCEL,   mdt_obd_log_cancel),
 164 DEF_OBD_HDL(0,                          OBD_QC_CALLBACK,  mdt_obd_qc_callback),
 165 DEF_OBD_HDL(0,                          OBD_IDX_READ,     mdt_obd_idx_read)
 166 };
 167
 168 #define DEF_DLM_HDL_VAR(flags, name, fn)                                \
 169         DEFINE_RPC_HANDLER(LDLM_ENQUEUE, flags, name, fn, NULL)
 170 #define DEF_DLM_HDL(flags, name, fn)                                    \
 171         DEFINE_RPC_HANDLER(LDLM_ENQUEUE, flags, name, fn, &RQF_ ## name)
 172
 173 static struct mdt_handler mdt_dlm_ops[] = {
 174 DEF_DLM_HDL    (HABEO_CLAVIS,           LDLM_ENQUEUE,     mdt_enqueue),
 175 DEF_DLM_HDL_VAR(HABEO_CLAVIS,           LDLM_CONVERT,     mdt_convert),
 176 DEF_DLM_HDL_VAR(0,                      LDLM_BL_CALLBACK, mdt_bl_callback),
 177 DEF_DLM_HDL_VAR(0,                      LDLM_CP_CALLBACK, mdt_cp_callback)
 178 };
 179
 180 #define DEF_LLOG_HDL(flags, name, fn)                                   \
 181         DEFINE_RPC_HANDLER(LLOG_ORIGIN_HANDLE_CREATE, flags, name, fn, NULL)
 182
 183 static struct mdt_handler mdt_llog_ops[] = {
 184 DEF_LLOG_HDL(0,         LLOG_ORIGIN_HANDLE_CREATE,        mdt_llog_create),
 185 DEF_LLOG_HDL(0,         LLOG_ORIGIN_HANDLE_NEXT_BLOCK,    mdt_llog_next_block),
 186 DEF_LLOG_HDL(0,         LLOG_ORIGIN_HANDLE_READ_HEADER,   mdt_llog_read_header),
 187 DEF_LLOG_HDL(0,         LLOG_ORIGIN_HANDLE_WRITE_REC,     NULL),
 188 DEF_LLOG_HDL(0,         LLOG_ORIGIN_HANDLE_CLOSE,         NULL),
 189 DEF_LLOG_HDL(0,         LLOG_ORIGIN_CONNECT,              NULL),
 190 DEF_LLOG_HDL(0,         LLOG_CATINFO,                     NULL),
 191 DEF_LLOG_HDL(0,         LLOG_ORIGIN_HANDLE_PREV_BLOCK,    mdt_llog_prev_block),
 192 DEF_LLOG_HDL(0,         LLOG_ORIGIN_HANDLE_DESTROY,       mdt_llog_destroy),
 193 };
 194
 195 #define DEF_SEC_HDL(flags, name, fn)                                    \
 196         DEFINE_RPC_HANDLER(SEC_CTX_INIT, flags, name, fn, NULL)
 197
 198 static struct mdt_handler mdt_sec_ctx_ops[] = {
 199 DEF_SEC_HDL(0,                          SEC_CTX_INIT,     mdt_sec_ctx_handle),
 200 DEF_SEC_HDL(0,                          SEC_CTX_INIT_CONT,mdt_sec_ctx_handle),
 201 DEF_SEC_HDL(0,                          SEC_CTX_FINI,     mdt_sec_ctx_handle)
 202 };
 203
 204 #define DEF_QUOTA_HDL(flags, name, fn)                          \
 205         DEFINE_RPC_HANDLER(QUOTA_DQACQ, flags, name, fn, &RQF_ ## name)
 206
 207 static struct mdt_handler mdt_quota_ops[] = {
 208 DEF_QUOTA_HDL(HABEO_REFERO,             QUOTA_DQACQ,      mdt_quota_dqacq),
 209 };
 210
 211 struct mdt_opc_slice mdt_regular_handlers[] = {
 212         {
 213                 .mos_opc_start  = MDS_GETATTR,
 214                 .mos_opc_end    = MDS_LAST_OPC,
 215                 .mos_hs         = mdt_mds_ops
 216         },
 217         {
 218                 .mos_opc_start  = OBD_PING,
 219                 .mos_opc_end    = OBD_LAST_OPC,
 220                 .mos_hs         = mdt_obd_ops
 221         },
 222         {
 223                 .mos_opc_start  = LDLM_ENQUEUE,
 224                 .mos_opc_end    = LDLM_LAST_OPC,
 225                 .mos_hs         = mdt_dlm_ops
 226         },
 227         {
 228                 .mos_opc_start  = LLOG_ORIGIN_HANDLE_CREATE,
 229                 .mos_opc_end    = LLOG_LAST_OPC,
 230                 .mos_hs         = mdt_llog_ops
 231         },
 232         {
 233                 .mos_opc_start  = SEC_CTX_INIT,
 234                 .mos_opc_end    = SEC_LAST_OPC,
 235                 .mos_hs         = mdt_sec_ctx_ops
 236         },
 237         {
 238                 .mos_opc_start  = QUOTA_DQACQ,
 239                 .mos_opc_end    = QUOTA_LAST_OPC,
 240                 .mos_hs         = mdt_quota_ops
 241         },
 242         {
 243                 .mos_hs         = NULL
 244         }
 245 };
 246
 247 /* Readpage/readdir handlers */
 248 static struct mdt_handler mdt_readpage_ops[] = {
 249 DEF_MDT_HDL(0,                  MDS_CONNECT,  mdt_connect),
 250 DEF_MDT_HDL(HABEO_CORPUS | HABEO_REFERO, MDS_READPAGE, mdt_readpage),
 251 /* XXX: this is ugly and should be fixed one day, see mdc_close() for
 252  * detailed comments. --umka */
 253 DEF_MDT_HDL(HABEO_CORPUS,               MDS_CLOSE,        mdt_close),
 254 DEF_MDT_HDL(HABEO_CORPUS,               MDS_DONE_WRITING, mdt_done_writing),
 255 };
 256
 257 static struct mdt_opc_slice mdt_readpage_handlers[] = {
 258         {
 259                 .mos_opc_start = MDS_GETATTR,
 260                 .mos_opc_end   = MDS_LAST_OPC,
 261                 .mos_hs = mdt_readpage_ops
 262         },
 263         {
 264                 .mos_opc_start = OBD_FIRST_OPC,
 265                 .mos_opc_end   = OBD_LAST_OPC,
 266                 .mos_hs = mdt_obd_ops
 267         },
 268         {
 269                 .mos_hs = NULL
 270         }
 271 };
 272
 273 /* Sequence service handlers */
 274 #define DEF_SEQ_HDL(flags, name, fn)                                    \
 275         DEFINE_RPC_HANDLER(SEQ_QUERY, flags, name, fn, &RQF_ ## name)
 276
 277 static struct mdt_handler mdt_seq_ops[] = {
 278 DEF_SEQ_HDL(0,                          SEQ_QUERY,        (void *)seq_query),
 279 };
 280
 281 struct mdt_opc_slice mdt_seq_handlers[] = {
 282         {
 283                 .mos_opc_start = SEQ_QUERY,
 284                 .mos_opc_end   = SEQ_LAST_OPC,
 285                 .mos_hs = mdt_seq_ops
 286         },
 287         {
 288                 .mos_hs = NULL
 289         }
 290 };
 291
 292 /* FID Location Database handlers */
 293 #define DEF_FLD_HDL(flags, name, fn)                                    \
 294         DEFINE_RPC_HANDLER(FLD_QUERY, flags, name, fn, &RQF_ ## name)
 295
 296 static struct mdt_handler mdt_fld_ops[] = {
 297 DEF_FLD_HDL(0,                          FLD_QUERY,        (void *)fld_query),
 298 };
 299
 300 struct mdt_opc_slice mdt_fld_handlers[] = {
 301         {
 302                 .mos_opc_start = FLD_QUERY,
 303                 .mos_opc_end   = FLD_LAST_OPC,
 304                 .mos_hs = mdt_fld_ops
 305         },
 306         {
 307                 .mos_hs = NULL
 308         }
 309 };
 310
 311 static int mds_regular_handle(struct ptlrpc_request *req)
 312 {
 313         return mdt_handle_common(req, mdt_regular_handlers);
 314 }
 315
 316 static int mds_readpage_handle(struct ptlrpc_request *req)
 317 {
 318         return mdt_handle_common(req, mdt_readpage_handlers);
 319 }
 320
 321 static int mds_mdsc_handle(struct ptlrpc_request *req)
 322 {
 323         return mdt_handle_common(req, mdt_seq_handlers);
 324 }
 325
 326 static int mds_mdss_handle(struct ptlrpc_request *req)
 327 {
 328         return mdt_handle_common(req, mdt_seq_handlers);
 329 }
 330
 331 static int mds_fld_handle(struct ptlrpc_request *req)
 332 {
 333         return mdt_handle_common(req, mdt_fld_handlers);
 334 }
 335
 336 /* device init/fini methods */
 337 static void mds_stop_ptlrpc_service(struct mds_device *m)
 338 {
 339         ENTRY;
 340         if (m->mds_regular_service != NULL) {
 341                 ptlrpc_unregister_service(m->mds_regular_service);
 342                 m->mds_regular_service = NULL;
 343         }
 344         if (m->mds_readpage_service != NULL) {
 345                 ptlrpc_unregister_service(m->mds_readpage_service);
 346                 m->mds_readpage_service = NULL;
 347         }
 348         if (m->mds_setattr_service != NULL) {
 349                 ptlrpc_unregister_service(m->mds_setattr_service);
 350                 m->mds_setattr_service = NULL;
 351         }
 352         if (m->mds_mdsc_service != NULL) {
 353                 ptlrpc_unregister_service(m->mds_mdsc_service);
 354                 m->mds_mdsc_service = NULL;
 355         }
 356         if (m->mds_mdss_service != NULL) {
 357                 ptlrpc_unregister_service(m->mds_mdss_service);
 358                 m->mds_mdss_service = NULL;
 359         }
 360         if (m->mds_fld_service != NULL) {
 361                 ptlrpc_unregister_service(m->mds_fld_service);
 362                 m->mds_fld_service = NULL;
 363         }
 364         EXIT;
 365 }
 366
 367 static int mds_start_ptlrpc_service(struct mds_device *m)
 368 {
 369         static struct ptlrpc_service_conf conf;
 370         struct obd_device *obd = m->mds_md_dev.md_lu_dev.ld_obd;
 371         cfs_proc_dir_entry_t *procfs_entry;
 372         int rc = 0;
 373         ENTRY;
 374
 375         procfs_entry = obd->obd_proc_entry;
 376         LASSERT(procfs_entry != NULL);
 377
 378         conf = (typeof(conf)) {
 379                 .psc_name               = LUSTRE_MDT_NAME,
 380                 .psc_watchdog_factor    = MDT_SERVICE_WATCHDOG_FACTOR,
 381                 .psc_buf                = {
 382                         .bc_nbufs               = MDS_NBUFS,
 383                         .bc_buf_size            = MDS_BUFSIZE,
 384                         .bc_req_max_size        = MDS_MAXREQSIZE,
 385                         .bc_rep_max_size        = MDS_MAXREPSIZE,
 386                         .bc_req_portal          = MDS_REQUEST_PORTAL,
 387                         .bc_rep_portal          = MDC_REPLY_PORTAL,
 388                 },
 389                 /*
 390                  * We'd like to have a mechanism to set this on a per-device
 391                  * basis, but alas...
 392                  */
 393                 .psc_thr                = {
 394                         .tc_thr_name            = LUSTRE_MDT_NAME,
 395                         .tc_thr_factor          = MDS_THR_FACTOR,
 396                         .tc_nthrs_init          = MDS_NTHRS_INIT,
 397                         .tc_nthrs_base          = MDS_NTHRS_BASE,
 398                         .tc_nthrs_max           = MDS_NTHRS_MAX,
 399                         .tc_nthrs_user          = mds_num_threads,
 400                         .tc_cpu_affinity        = 1,
 401                         .tc_ctx_tags            = LCT_MD_THREAD,
 402                 },
 403                 .psc_cpt                = {
 404                         .cc_pattern             = mds_num_cpts,
 405                 },
 406                 .psc_ops                = {
 407                         .so_req_handler         = mds_regular_handle,
 408                         .so_req_printer         = target_print_req,
 409                         .so_hpreq_handler       = ptlrpc_hpreq_handler,
 410                 },
 411         };
 412         m->mds_regular_service = ptlrpc_register_service(&conf, procfs_entry);
 413         if (IS_ERR(m->mds_regular_service)) {
 414                 rc = PTR_ERR(m->mds_regular_service);
 415                 CERROR("failed to start regular mdt service: %d\n", rc);
 416                 m->mds_regular_service = NULL;
 417
 418                 RETURN(rc);
 419         }
 420
 421         /*
 422          * readpage service configuration. Parameters have to be adjusted,
 423          * ideally.
 424          */
 425         memset(&conf, 0, sizeof(conf));
 426         conf = (typeof(conf)) {
 427                 .psc_name               = LUSTRE_MDT_NAME "_readpage",
 428                 .psc_watchdog_factor    = MDT_SERVICE_WATCHDOG_FACTOR,
 429                 .psc_buf                = {
 430                         .bc_nbufs               = MDS_NBUFS,
 431                         .bc_buf_size            = MDS_BUFSIZE,
 432                         .bc_req_max_size        = MDS_MAXREQSIZE,
 433                         .bc_rep_max_size        = MDS_MAXREPSIZE,
 434                         .bc_req_portal          = MDS_READPAGE_PORTAL,
 435                         .bc_rep_portal          = MDC_REPLY_PORTAL,
 436                 },
 437                 .psc_thr                = {
 438                         .tc_thr_name            = LUSTRE_MDT_NAME "_rdpg",
 439                         .tc_thr_factor          = MDS_RDPG_THR_FACTOR,
 440                         .tc_nthrs_init          = MDS_RDPG_NTHRS_INIT,
 441                         .tc_nthrs_base          = MDS_RDPG_NTHRS_BASE,
 442                         .tc_nthrs_max           = MDS_RDPG_NTHRS_MAX,
 443                         .tc_nthrs_user          = mds_rdpg_num_threads,
 444                         .tc_cpu_affinity        = 1,
 445                         .tc_ctx_tags            = LCT_MD_THREAD,
 446                 },
 447                 .psc_cpt                = {
 448                         .cc_pattern             = mds_rdpg_num_cpts,
 449                 },
 450                 .psc_ops                = {
 451                         .so_req_handler         = mds_readpage_handle,
 452                         .so_req_printer         = target_print_req,
 453                 },
 454         };
 455         m->mds_readpage_service = ptlrpc_register_service(&conf, procfs_entry);
 456         if (IS_ERR(m->mds_readpage_service)) {
 457                 rc = PTR_ERR(m->mds_readpage_service);
 458                 CERROR("failed to start readpage service: %d\n", rc);
 459                 m->mds_readpage_service = NULL;
 460
 461                 GOTO(err_mds_svc, rc);
 462         }
 463
 464         /*
 465          * setattr service configuration.
 466          *
 467          * XXX To keep the compatibility with old client(< 2.2), we need to
 468          * preserve this portal for a certain time, it should be removed
 469          * eventually. LU-617.
 470          */
 471         memset(&conf, 0, sizeof(conf));
 472         conf = (typeof(conf)) {
 473                 .psc_name               = LUSTRE_MDT_NAME "_setattr",
 474                 .psc_watchdog_factor    = MDT_SERVICE_WATCHDOG_FACTOR,
 475                 .psc_buf                = {
 476                         .bc_nbufs               = MDS_NBUFS,
 477                         .bc_buf_size            = MDS_BUFSIZE,
 478                         .bc_req_max_size        = MDS_MAXREQSIZE,
 479                         .bc_rep_max_size        = MDS_MAXREPSIZE,
 480                         .bc_req_portal          = MDS_SETATTR_PORTAL,
 481                         .bc_rep_portal          = MDC_REPLY_PORTAL,
 482                 },
 483                 .psc_thr                = {
 484                         .tc_thr_name            = LUSTRE_MDT_NAME "_attr",
 485                         .tc_thr_factor          = MDS_SETA_THR_FACTOR,
 486                         .tc_nthrs_init          = MDS_SETA_NTHRS_INIT,
 487                         .tc_nthrs_base          = MDS_SETA_NTHRS_BASE,
 488                         .tc_nthrs_max           = MDS_SETA_NTHRS_MAX,
 489                         .tc_nthrs_user          = mds_attr_num_threads,
 490                         .tc_cpu_affinity        = 1,
 491                         .tc_ctx_tags            = LCT_MD_THREAD,
 492                 },
 493                 .psc_cpt                = {
 494                         .cc_pattern             = mds_attr_num_cpts,
 495                 },
 496                 .psc_ops                = {
 497                         .so_req_handler         = mds_regular_handle,
 498                         .so_req_printer         = target_print_req,
 499                         .so_hpreq_handler       = NULL,
 500                 },
 501         };
 502         m->mds_setattr_service = ptlrpc_register_service(&conf, procfs_entry);
 503         if (IS_ERR(m->mds_setattr_service)) {
 504                 rc = PTR_ERR(m->mds_setattr_service);
 505                 CERROR("failed to start setattr service: %d\n", rc);
 506                 m->mds_setattr_service = NULL;
 507
 508                 GOTO(err_mds_svc, rc);
 509         }
 510
 511         /*
 512          * sequence controller service configuration
 513          */
 514         memset(&conf, 0, sizeof(conf));
 515         conf = (typeof(conf)) {
 516                 .psc_name               = LUSTRE_MDT_NAME "_seqs",
 517                 .psc_watchdog_factor    = MDT_SERVICE_WATCHDOG_FACTOR,
 518                 .psc_buf                = {
 519                         .bc_nbufs               = MDS_NBUFS,
 520                         .bc_buf_size            = MDS_BUFSIZE,
 521                         .bc_req_max_size        = SEQ_MAXREQSIZE,
 522                         .bc_rep_max_size        = SEQ_MAXREPSIZE,
 523                         .bc_req_portal          = SEQ_CONTROLLER_PORTAL,
 524                         .bc_rep_portal          = MDC_REPLY_PORTAL,
 525                 },
 526                 .psc_thr                = {
 527                         .tc_thr_name            = LUSTRE_MDT_NAME "_seqs",
 528                         .tc_nthrs_init          = MDS_OTHR_NTHRS_INIT,
 529                         .tc_nthrs_max           = MDS_OTHR_NTHRS_MAX,
 530                         .tc_ctx_tags            = LCT_MD_THREAD,
 531                 },
 532                 .psc_ops                = {
 533                         .so_req_handler         = mds_mdsc_handle,
 534                         .so_req_printer         = target_print_req,
 535                         .so_hpreq_handler       = NULL,
 536                 },
 537         };
 538         m->mds_mdsc_service = ptlrpc_register_service(&conf, procfs_entry);
 539         if (IS_ERR(m->mds_mdsc_service)) {
 540                 rc = PTR_ERR(m->mds_mdsc_service);
 541                 CERROR("failed to start seq controller service: %d\n", rc);
 542                 m->mds_mdsc_service = NULL;
 543
 544                 GOTO(err_mds_svc, rc);
 545         }
 546
 547         /*
 548          * metadata sequence server service configuration
 549          */
 550         memset(&conf, 0, sizeof(conf));
 551         conf = (typeof(conf)) {
 552                 .psc_name               = LUSTRE_MDT_NAME "_seqm",
 553                 .psc_watchdog_factor    = MDT_SERVICE_WATCHDOG_FACTOR,
 554                 .psc_buf                = {
 555                         .bc_nbufs               = MDS_NBUFS,
 556                         .bc_buf_size            = MDS_BUFSIZE,
 557                         .bc_req_max_size        = SEQ_MAXREQSIZE,
 558                         .bc_rep_max_size        = SEQ_MAXREPSIZE,
 559                         .bc_req_portal          = SEQ_METADATA_PORTAL,
 560                         .bc_rep_portal          = MDC_REPLY_PORTAL,
 561                 },
 562                 .psc_thr                = {
 563                         .tc_thr_name            = LUSTRE_MDT_NAME "_seqm",
 564                         .tc_nthrs_init          = MDS_OTHR_NTHRS_INIT,
 565                         .tc_nthrs_max           = MDS_OTHR_NTHRS_MAX,
 566                         .tc_ctx_tags            = LCT_MD_THREAD | LCT_DT_THREAD
 567                 },
 568                 .psc_ops                = {
 569                         .so_req_handler         = mds_mdss_handle,
 570                         .so_req_printer         = target_print_req,
 571                         .so_hpreq_handler       = NULL,
 572                 },
 573         };
 574         m->mds_mdss_service = ptlrpc_register_service(&conf, procfs_entry);
 575         if (IS_ERR(m->mds_mdss_service)) {
 576                 rc = PTR_ERR(m->mds_mdss_service);
 577                 CERROR("failed to start metadata seq server service: %d\n", rc);
 578                 m->mds_mdss_service = NULL;
 579
 580                 GOTO(err_mds_svc, rc);
 581         }
 582
 583         /* FLD service start */
 584         memset(&conf, 0, sizeof(conf));
 585         conf = (typeof(conf)) {
 586                 .psc_name            = LUSTRE_MDT_NAME "_fld",
 587                 .psc_watchdog_factor = MDT_SERVICE_WATCHDOG_FACTOR,
 588                 .psc_buf                = {
 589                         .bc_nbufs               = MDS_NBUFS,
 590                         .bc_buf_size            = MDS_BUFSIZE,
 591                         .bc_req_max_size        = FLD_MAXREQSIZE,
 592                         .bc_rep_max_size        = FLD_MAXREPSIZE,
 593                         .bc_req_portal          = FLD_REQUEST_PORTAL,
 594                         .bc_rep_portal          = MDC_REPLY_PORTAL,
 595                 },
 596                 .psc_thr                = {
 597                         .tc_thr_name            = LUSTRE_MDT_NAME "_fld",
 598                         .tc_nthrs_init          = MDS_OTHR_NTHRS_INIT,
 599                         .tc_nthrs_max           = MDS_OTHR_NTHRS_MAX,
 600                         .tc_ctx_tags            = LCT_DT_THREAD | LCT_MD_THREAD
 601                 },
 602                 .psc_ops                = {
 603                         .so_req_handler         = mds_fld_handle,
 604                         .so_req_printer         = target_print_req,
 605                         .so_hpreq_handler       = NULL,
 606                 },
 607         };
 608         m->mds_fld_service = ptlrpc_register_service(&conf, procfs_entry);
 609         if (IS_ERR(m->mds_fld_service)) {
 610                 rc = PTR_ERR(m->mds_fld_service);
 611                 CERROR("failed to start fld service: %d\n", rc);
 612                 m->mds_fld_service = NULL;
 613
 614                 GOTO(err_mds_svc, rc);
 615         }
 616
 617         EXIT;
 618 err_mds_svc:
 619         if (rc)
 620                 mds_stop_ptlrpc_service(m);
 621
 622         return rc;
 623 }
 624
 625 static inline struct mds_device *mds_dev(struct lu_device *d)
 626 {
 627         return container_of0(d, struct mds_device, mds_md_dev.md_lu_dev);
 628 }
 629
 630 static struct lu_device *mds_device_fini(const struct lu_env *env,
 631                                          struct lu_device *d)
 632 {
 633         struct mds_device *m = mds_dev(d);
 634         struct obd_device *obd = d->ld_obd;
 635         ENTRY;
 636
 637         mds_stop_ptlrpc_service(m);
 638         lprocfs_obd_cleanup(obd);
 639         RETURN(NULL);
 640 }
 641
 642 static struct lu_device *mds_device_free(const struct lu_env *env,
 643                                          struct lu_device *d)
 644 {
 645         struct mds_device *m = mds_dev(d);
 646         ENTRY;
 647
 648         md_device_fini(&m->mds_md_dev);
 649         OBD_FREE_PTR(m);
 650         RETURN(NULL);
 651 }
 652
 653 static struct lu_device *mds_device_alloc(const struct lu_env *env,
 654                                           struct lu_device_type *t,
 655                                           struct lustre_cfg *cfg)
 656 {
 657         struct mds_device        *m;
 658         struct obd_device        *obd;
 659         struct lu_device          *l;
 660         int rc;
 661
 662         OBD_ALLOC_PTR(m);
 663         if (m == NULL)
 664                 return ERR_PTR(-ENOMEM);
 665
 666         md_device_init(&m->mds_md_dev, t);
 667         l = &m->mds_md_dev.md_lu_dev;
 668
 669         obd = class_name2obd(lustre_cfg_string(cfg, 0));
 670         LASSERT(obd != NULL);
 671
 672         l->ld_obd = obd;
 673         /* set this lu_device to obd, because error handling need it */
 674         obd->obd_lu_dev = l;
 675
 676         rc = lprocfs_obd_setup(obd, lprocfs_mds_obd_vars);
 677         if (rc != 0) {
 678                 mds_device_free(env, l);
 679                 l = ERR_PTR(rc);
 680                 return l;
 681         }
 682
 683         rc = mds_start_ptlrpc_service(m);
 684
 685         if (rc != 0) {
 686                 mds_device_free(env, l);
 687                 l = ERR_PTR(rc);
 688                 return l;
 689         }
 690
 691         return l;
 692 }
 693
 694 /* type constructor/destructor: mdt_type_init, mdt_type_fini */
 695 LU_TYPE_INIT_FINI(mds, &mdt_thread_key);
 696
 697 static struct lu_device_type_operations mds_device_type_ops = {
 698         .ldto_init = mds_type_init,
 699         .ldto_fini = mds_type_fini,
 700
 701         .ldto_start = mds_type_start,
 702         .ldto_stop  = mds_type_stop,
 703
 704         .ldto_device_alloc = mds_device_alloc,
 705         .ldto_device_free  = mds_device_free,
 706         .ldto_device_fini  = mds_device_fini
 707 };
 708
 709 static struct lu_device_type mds_device_type = {
 710         .ldt_tags     = LU_DEVICE_MD,
 711         .ldt_name     = LUSTRE_MDS_NAME,
 712         .ldt_ops      = &mds_device_type_ops,
 713         .ldt_ctx_tags = LCT_MD_THREAD
 714 };
 715
 716 static struct obd_ops mds_obd_device_ops = {
 717         .o_owner           = THIS_MODULE,
 718 };
 719
 720 int mds_mod_init(void)
 721 {
 722         int rc;
 723
 724         if (mdt_num_threads != 0 && mds_num_threads == 0) {
 725                 LCONSOLE_INFO("mdt_num_threads module parameter is deprecated, "
 726                               "use mds_num_threads instead or unset both for "
 727                               "dynamic thread startup\n");
 728                 mds_num_threads = mdt_num_threads;
 729         }
 730
 731         rc = class_register_type(&mds_obd_device_ops, NULL,
 732                                  lprocfs_mds_module_vars, LUSTRE_MDS_NAME,
 733                                  &mds_device_type);
 734         return rc;
 735 }
 736
 737 void mds_mod_exit(void)
 738 {
 739         class_unregister_type(LUSTRE_MDS_NAME);
 740 }