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