lustre/ptlrpc/niobuf.c

   1 /* -*- mode: c; c-basic-offset: 8; indent-tabs-mode: nil; -*-
   2  * vim:expandtab:shiftwidth=8:tabstop=8:
   3  *
   4  * GPL HEADER START
   5  *
   6  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
   7  *
   8  * This program is free software; you can redistribute it and/or modify
   9  * it under the terms of the GNU General Public License version 2 only,
  10  * as published by the Free Software Foundation.
  11  *
  12  * This program is distributed in the hope that it will be useful, but
  13  * WITHOUT ANY WARRANTY; without even the implied warranty of
  14  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  15  * General Public License version 2 for more details (a copy is included
  16  * in the LICENSE file that accompanied this code).
  17  *
  18  * You should have received a copy of the GNU General Public License
  19  * version 2 along with this program; If not, see
  20  * http://www.sun.com/software/products/lustre/docs/GPLv2.pdf
  21  *
  22  * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
  23  * CA 95054 USA or visit www.sun.com if you need additional information or
  24  * have any questions.
  25  *
  26  * GPL HEADER END
  27  */
  28 /*
  29  * Copyright  2008 Sun Microsystems, Inc. All rights reserved
  30  * Use is subject to license terms.
  31  */
  32 /*
  33  * This file is part of Lustre, http://www.lustre.org/
  34  * Lustre is a trademark of Sun Microsystems, Inc.
  35  */
  36
  37 #define DEBUG_SUBSYSTEM S_RPC
  38 #ifndef __KERNEL__
  39 #include <liblustre.h>
  40 #endif
  41 #include <obd_support.h>
  42 #include <lustre_net.h>
  43 #include <lustre_lib.h>
  44 #include <obd.h>
  45 #include "ptlrpc_internal.h"
  46
  47 static int ptl_send_buf (lnet_handle_md_t *mdh, void *base, int len,
  48                          lnet_ack_req_t ack, struct ptlrpc_cb_id *cbid,
  49                          struct ptlrpc_connection *conn, int portal, __u64 xid,
  50                          unsigned int offset)
  51 {
  52         int              rc;
  53         lnet_md_t         md;
  54         ENTRY;
  55
  56         LASSERT (portal != 0);
  57         LASSERT (conn != NULL);
  58         CDEBUG (D_INFO, "conn=%p id %s\n", conn, libcfs_id2str(conn->c_peer));
  59         md.start     = base;
  60         md.length    = len;
  61         md.threshold = (ack == LNET_ACK_REQ) ? 2 : 1;
  62         md.options   = PTLRPC_MD_OPTIONS;
  63         md.user_ptr  = cbid;
  64         md.eq_handle = ptlrpc_eq_h;
  65
  66         if (ack == LNET_ACK_REQ &&
  67             OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_ACK | OBD_FAIL_ONCE)) {
  68                 /* don't ask for the ack to simulate failing client */
  69                 ack = LNET_NOACK_REQ;
  70                 obd_fail_loc |= OBD_FAIL_ONCE | OBD_FAILED;
  71         }
  72
  73         rc = LNetMDBind (md, LNET_UNLINK, mdh);
  74         if (rc != 0) {
  75                 CERROR ("LNetMDBind failed: %d\n", rc);
  76                 LASSERT (rc == -ENOMEM);
  77                 RETURN (-ENOMEM);
  78         }
  79
  80         CDEBUG(D_NET, "Sending %d bytes to portal %d, xid "LPD64", offset %u\n",
  81                len, portal, xid, offset);
  82
  83         rc = LNetPut (conn->c_self, *mdh, ack,
  84                       conn->c_peer, portal, xid, offset, 0);
  85         if (rc != 0) {
  86                 int rc2;
  87                 /* We're going to get an UNLINK event when I unlink below,
  88                  * which will complete just like any other failed send, so
  89                  * I fall through and return success here! */
  90                 CERROR("LNetPut(%s, %d, "LPD64") failed: %d\n",
  91                        libcfs_id2str(conn->c_peer), portal, xid, rc);
  92                 rc2 = LNetMDUnlink(*mdh);
  93                 LASSERTF(rc2 == 0, "rc2 = %d\n", rc2);
  94         }
  95
  96         RETURN (0);
  97 }
  98
  99 int ptlrpc_start_bulk_transfer(struct ptlrpc_bulk_desc *desc)
 100 {
 101         struct ptlrpc_connection *conn = desc->bd_export->exp_connection;
 102         int                       rc;
 103         int                       rc2;
 104         lnet_md_t                 md;
 105         __u64                     xid;
 106         ENTRY;
 107
 108         if (OBD_FAIL_CHECK_ONCE(OBD_FAIL_PTLRPC_BULK_PUT_NET))
 109                 RETURN(0);
 110
 111         /* NB no locking required until desc is on the network */
 112         LASSERT (!desc->bd_network_rw);
 113         LASSERT (desc->bd_type == BULK_PUT_SOURCE ||
 114                  desc->bd_type == BULK_GET_SINK);
 115         desc->bd_success = 0;
 116         desc->bd_sender = LNET_NID_ANY;
 117
 118         md.user_ptr = &desc->bd_cbid;
 119         md.eq_handle = ptlrpc_eq_h;
 120         md.threshold = 2; /* SENT and ACK/REPLY */
 121         md.options = PTLRPC_MD_OPTIONS;
 122         ptlrpc_fill_bulk_md(&md, desc);
 123
 124         LASSERT (desc->bd_cbid.cbid_fn == server_bulk_callback);
 125         LASSERT (desc->bd_cbid.cbid_arg == desc);
 126
 127         /* NB total length may be 0 for a read past EOF, so we send a 0
 128          * length bulk, since the client expects a bulk event. */
 129
 130         rc = LNetMDBind(md, LNET_UNLINK, &desc->bd_md_h);
 131         if (rc != 0) {
 132                 CERROR("LNetMDBind failed: %d\n", rc);
 133                 LASSERT (rc == -ENOMEM);
 134                 RETURN(-ENOMEM);
 135         }
 136
 137         /* Client's bulk and reply matchbits are the same */
 138         xid = desc->bd_req->rq_xid;
 139         CDEBUG(D_NET, "Transferring %u pages %u bytes via portal %d "
 140                "id %s xid "LPX64"\n", desc->bd_iov_count,
 141                desc->bd_nob, desc->bd_portal,
 142                libcfs_id2str(conn->c_peer), xid);
 143
 144         /* Network is about to get at the memory */
 145         desc->bd_network_rw = 1;
 146
 147         if (desc->bd_type == BULK_PUT_SOURCE)
 148                 rc = LNetPut (conn->c_self, desc->bd_md_h, LNET_ACK_REQ,
 149                               conn->c_peer, desc->bd_portal, xid, 0, 0);
 150         else
 151                 rc = LNetGet (conn->c_self, desc->bd_md_h,
 152                               conn->c_peer, desc->bd_portal, xid, 0);
 153
 154         if (rc != 0) {
 155                 /* Can't send, so we unlink the MD bound above.  The UNLINK
 156                  * event this creates will signal completion with failure,
 157                  * so we return SUCCESS here! */
 158                 CERROR("Transfer(%s, %d, "LPX64") failed: %d\n",
 159                        libcfs_id2str(conn->c_peer), desc->bd_portal, xid, rc);
 160                 rc2 = LNetMDUnlink(desc->bd_md_h);
 161                 LASSERT (rc2 == 0);
 162         }
 163
 164         RETURN(0);
 165 }
 166
 167 /* Server side bulk abort. Idempotent. Not thread-safe (i.e. only
 168  * serialises with completion callback) */
 169 void ptlrpc_abort_bulk(struct ptlrpc_bulk_desc *desc)
 170 {
 171         struct l_wait_info       lwi;
 172         int                      rc;
 173
 174         LASSERT(!in_interrupt());               /* might sleep */
 175
 176         if (!ptlrpc_server_bulk_active(desc))   /* completed or */
 177                 return;                         /* never started */
 178
 179         /* Do not send any meaningful data over the wire for evicted clients */
 180         if (desc->bd_export && desc->bd_export->exp_failed)
 181                 ptl_rpc_wipe_bulk_pages(desc);
 182
 183         /* The unlink ensures the callback happens ASAP and is the last
 184          * one.  If it fails, it must be because completion just happened,
 185          * but we must still l_wait_event() in this case, to give liblustre
 186          * a chance to run server_bulk_callback()*/
 187
 188         LNetMDUnlink(desc->bd_md_h);
 189
 190         for (;;) {
 191                 /* Network access will complete in finite time but the HUGE
 192                  * timeout lets us CWARN for visibility of sluggish NALs */
 193                 lwi = LWI_TIMEOUT_INTERVAL(cfs_time_seconds(LONG_UNLINK),
 194                                            cfs_time_seconds(1), NULL, NULL);
 195                 rc = l_wait_event(desc->bd_waitq,
 196                                   !ptlrpc_server_bulk_active(desc), &lwi);
 197                 if (rc == 0)
 198                         return;
 199
 200                 LASSERT(rc == -ETIMEDOUT);
 201                 CWARN("Unexpectedly long timeout: desc %p\n", desc);
 202         }
 203 }
 204
 205 int ptlrpc_register_bulk(struct ptlrpc_request *req)
 206 {
 207         struct ptlrpc_bulk_desc *desc = req->rq_bulk;
 208         lnet_process_id_t peer;
 209         int rc;
 210         int rc2;
 211         lnet_handle_me_t  me_h;
 212         lnet_md_t         md;
 213         ENTRY;
 214
 215         if (OBD_FAIL_CHECK_ONCE(OBD_FAIL_PTLRPC_BULK_GET_NET))
 216                 RETURN(0);
 217
 218         /* NB no locking required until desc is on the network */
 219         LASSERT (desc->bd_nob > 0);
 220         LASSERT (!desc->bd_network_rw);
 221         LASSERT (desc->bd_iov_count <= PTLRPC_MAX_BRW_PAGES);
 222         LASSERT (desc->bd_req != NULL);
 223         LASSERT (desc->bd_type == BULK_PUT_SINK ||
 224                  desc->bd_type == BULK_GET_SOURCE);
 225
 226         desc->bd_success = 0;
 227         desc->bd_sender = LNET_NID_ANY;
 228
 229         peer = desc->bd_import->imp_connection->c_peer;
 230
 231         md.user_ptr = &desc->bd_cbid;
 232         md.eq_handle = ptlrpc_eq_h;
 233         md.threshold = 1;                       /* PUT or GET */
 234         md.options = PTLRPC_MD_OPTIONS |
 235                      ((desc->bd_type == BULK_GET_SOURCE) ?
 236                       LNET_MD_OP_GET : LNET_MD_OP_PUT);
 237         ptlrpc_fill_bulk_md(&md, desc);
 238
 239         LASSERT (desc->bd_cbid.cbid_fn == client_bulk_callback);
 240         LASSERT (desc->bd_cbid.cbid_arg == desc);
 241
 242         /* XXX Registering the same xid on retried bulk makes my head
 243          * explode trying to understand how the original request's bulk
 244          * might interfere with the retried request -eeb */
 245         LASSERTF (!desc->bd_registered || req->rq_xid != desc->bd_last_xid,
 246                   "registered: %d  rq_xid: "LPU64" bd_last_xid: "LPU64"\n",
 247                   desc->bd_registered, req->rq_xid, desc->bd_last_xid);
 248         desc->bd_registered = 1;
 249         desc->bd_last_xid = req->rq_xid;
 250
 251         rc = LNetMEAttach(desc->bd_portal, peer,
 252                          req->rq_xid, 0, LNET_UNLINK, LNET_INS_AFTER, &me_h);
 253         if (rc != 0) {
 254                 CERROR("LNetMEAttach failed: %d\n", rc);
 255                 LASSERT (rc == -ENOMEM);
 256                 RETURN (-ENOMEM);
 257         }
 258
 259         /* About to let the network at it... */
 260         desc->bd_network_rw = 1;
 261         rc = LNetMDAttach(me_h, md, LNET_UNLINK, &desc->bd_md_h);
 262         if (rc != 0) {
 263                 CERROR("LNetMDAttach failed: %d\n", rc);
 264                 LASSERT (rc == -ENOMEM);
 265                 desc->bd_network_rw = 0;
 266                 rc2 = LNetMEUnlink (me_h);
 267                 LASSERT (rc2 == 0);
 268                 RETURN (-ENOMEM);
 269         }
 270
 271         CDEBUG(D_NET, "Setup bulk %s buffers: %u pages %u bytes, xid "LPX64", "
 272                "portal %u\n",
 273                desc->bd_type == BULK_GET_SOURCE ? "get-source" : "put-sink",
 274                desc->bd_iov_count, desc->bd_nob,
 275                req->rq_xid, desc->bd_portal);
 276         RETURN(0);
 277 }
 278
 279 /* Disconnect a bulk desc from the network. Idempotent. Not
 280  * thread-safe (i.e. only interlocks with completion callback). */
 281 int ptlrpc_unregister_bulk(struct ptlrpc_request *req, int async)
 282 {
 283         struct ptlrpc_bulk_desc *desc = req->rq_bulk;
 284         cfs_waitq_t             *wq;
 285         struct l_wait_info       lwi;
 286         int                      rc;
 287         ENTRY;
 288
 289         LASSERT(!in_interrupt());     /* might sleep */
 290
 291         /* Let's setup deadline for reply unlink. */
 292         if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_BULK_UNLINK) &&
 293             async && req->rq_bulk_deadline == 0)
 294                 req->rq_bulk_deadline = cfs_time_current_sec() + LONG_UNLINK;
 295
 296         if (!ptlrpc_client_bulk_active(req))  /* completed or */
 297                 RETURN(1);                    /* never registered */
 298
 299         LASSERT(desc->bd_req == req);  /* bd_req NULL until registered */
 300
 301         /* the unlink ensures the callback happens ASAP and is the last
 302          * one.  If it fails, it must be because completion just happened,
 303          * but we must still l_wait_event() in this case to give liblustre
 304          * a chance to run client_bulk_callback() */
 305
 306         LNetMDUnlink(desc->bd_md_h);
 307
 308         if (!ptlrpc_client_bulk_active(req))  /* completed or */
 309                 RETURN(1);                    /* never registered */
 310
 311         /* Move to "Unregistering" phase as bulk was not unlinked yet. */
 312         ptlrpc_rqphase_move(req, RQ_PHASE_UNREGISTERING);
 313
 314         /* Do not wait for unlink to finish. */
 315         if (async)
 316                 RETURN(0);
 317
 318         if (req->rq_set != NULL)
 319                 wq = &req->rq_set->set_waitq;
 320         else
 321                 wq = &req->rq_reply_waitq;
 322
 323         for (;;) {
 324                 /* Network access will complete in finite time but the HUGE
 325                  * timeout lets us CWARN for visibility of sluggish NALs */
 326                 lwi = LWI_TIMEOUT_INTERVAL(cfs_time_seconds(LONG_UNLINK),
 327                                            cfs_time_seconds(1), NULL, NULL);
 328                 rc = l_wait_event(*wq, !ptlrpc_client_bulk_active(req), &lwi);
 329                 if (rc == 0) {
 330                         ptlrpc_rqphase_move(req, req->rq_next_phase);
 331                         RETURN(1);
 332                 }
 333
 334                 LASSERT(rc == -ETIMEDOUT);
 335                 DEBUG_REQ(D_WARNING, req, "Unexpectedly long timeout: desc %p",
 336                           desc);
 337         }
 338         RETURN(0);
 339 }
 340
 341 int ptlrpc_send_reply(struct ptlrpc_request *req, int flags)
 342 {
 343         struct ptlrpc_service     *svc = req->rq_rqbd->rqbd_service;
 344         struct ptlrpc_reply_state *rs = req->rq_reply_state;
 345         struct ptlrpc_connection  *conn;
 346         int                        service_time;
 347         unsigned int               offset = 0;
 348         int                        rc;
 349
 350         /* We must already have a reply buffer (only ptlrpc_error() may be
 351          * called without one).  We must also have a request buffer which
 352          * is either the actual (swabbed) incoming request, or a saved copy
 353          * if this is a req saved in target_queue_final_reply(). */
 354         LASSERT (req->rq_reqmsg != NULL);
 355         LASSERT (req->rq_repmsg != NULL);
 356         LASSERT (rs != NULL);
 357         LASSERT (req->rq_repmsg == rs->rs_msg);
 358         LASSERT ((flags & PTLRPC_REPLY_MAYBE_DIFFICULT) || !rs->rs_difficult);
 359         LASSERT (rs->rs_cb_id.cbid_fn == reply_out_callback);
 360         LASSERT (rs->rs_cb_id.cbid_arg == rs);
 361
 362         /* There may be no rq_export during failover */
 363
 364         if (req->rq_export && req->rq_export->exp_obd &&
 365             req->rq_export->exp_obd->obd_fail) {
 366                 /* Failed obd's only send ENODEV */
 367                 req->rq_type = PTL_RPC_MSG_ERR;
 368                 req->rq_status = -ENODEV;
 369                 CDEBUG(D_HA, "sending ENODEV from failed obd %d\n",
 370                        req->rq_export->exp_obd->obd_minor);
 371         }
 372
 373         if (req->rq_type != PTL_RPC_MSG_ERR)
 374                 req->rq_type = PTL_RPC_MSG_REPLY;
 375
 376         lustre_msg_set_type(req->rq_repmsg, req->rq_type);
 377         lustre_msg_set_status(req->rq_repmsg, req->rq_status);
 378         lustre_msg_set_opc(req->rq_repmsg, lustre_msg_get_opc(req->rq_reqmsg));
 379
 380         service_time = max_t(int, cfs_time_current_sec() -
 381                              req->rq_arrival_time.tv_sec, 1);
 382         if (!(flags & PTLRPC_REPLY_EARLY) &&
 383             (req->rq_type != PTL_RPC_MSG_ERR) &&
 384             !(lustre_msg_get_flags(req->rq_reqmsg) &
 385               (MSG_RESENT | MSG_REPLAY | MSG_LAST_REPLAY))) {
 386                 /* early replies, errors and recovery requests don't count
 387                  * toward our service time estimate */
 388                 int oldse = at_add(&svc->srv_at_estimate, service_time);
 389                 if (oldse != 0)
 390                         DEBUG_REQ(D_ADAPTTO, req,
 391                                   "svc %s changed estimate from %d to %d",
 392                                   svc->srv_name, oldse,
 393                                   at_get(&svc->srv_at_estimate));
 394         }
 395         /* Report actual service time for client latency calc */
 396         lustre_msg_set_service_time(req->rq_repmsg, service_time);
 397         /* Report service time estimate for future client reqs, but report 0
 398          * (to be ignored by client) if it's a error reply during recovery.
 399          * (bz15815) */
 400         if (req->rq_type == PTL_RPC_MSG_ERR &&
 401             (req->rq_export == NULL || req->rq_export->exp_obd->obd_recovering))
 402                 lustre_msg_set_timeout(req->rq_repmsg, 0);
 403         else
 404                 lustre_msg_set_timeout(req->rq_repmsg,
 405                                        at_get(&svc->srv_at_estimate));
 406
 407         target_pack_pool_reply(req);
 408
 409         if (lustre_msghdr_get_flags(req->rq_reqmsg) & MSGHDR_AT_SUPPORT) {
 410                 /* early replies go to offset 0, regular replies go after that*/
 411                 if (flags & PTLRPC_REPLY_EARLY) {
 412                         offset = 0;
 413                         /* Include a checksum on early replies - must be done
 414                            after all other lustre_msg_set_* */
 415                         lustre_msg_set_cksum(req->rq_repmsg,
 416                                          lustre_msg_calc_cksum(req->rq_repmsg));
 417                 } else {
 418                         offset = lustre_msg_early_size(req);
 419                 }
 420         } else {
 421                 CDEBUG(D_ADAPTTO, "No early reply support: flags=%#x "
 422                        "req_flags=%#x magic=%d:%x/%x len=%d\n",
 423                       flags, lustre_msg_get_flags(req->rq_reqmsg),
 424                       lustre_msg_is_v1(req->rq_reqmsg),
 425                       lustre_msg_get_magic(req->rq_reqmsg),
 426                       lustre_msg_get_magic(req->rq_repmsg), req->rq_replen);
 427         }
 428
 429         if (req->rq_export == NULL || req->rq_export->exp_connection == NULL)
 430                 conn = ptlrpc_connection_get(req->rq_peer, req->rq_self, NULL);
 431         else
 432                 conn = ptlrpc_connection_addref(req->rq_export->exp_connection);
 433
 434         if (conn == NULL) {
 435                 CERROR("not replying on NULL connection\n"); /* bug 9635 */
 436                 return -ENOTCONN;
 437         }
 438
 439         atomic_inc (&svc->srv_outstanding_replies);
 440         ptlrpc_rs_addref(rs);                   /* +1 ref for the network */
 441         req->rq_sent = cfs_time_current_sec();
 442
 443         rc = ptl_send_buf (&rs->rs_md_h, req->rq_repmsg, req->rq_replen,
 444                            rs->rs_difficult ? LNET_ACK_REQ : LNET_NOACK_REQ,
 445                            &rs->rs_cb_id, conn, svc->srv_rep_portal,
 446                            req->rq_xid, offset);
 447         if (rc != 0) {
 448                 atomic_dec (&svc->srv_outstanding_replies);
 449                 ptlrpc_req_drop_rs(req);
 450         }
 451         ptlrpc_connection_put(conn);
 452         return rc;
 453 }
 454
 455 int ptlrpc_reply (struct ptlrpc_request *req)
 456 {
 457         return (ptlrpc_send_reply (req, 0));
 458 }
 459
 460 int ptlrpc_send_error(struct ptlrpc_request *req, int may_be_difficult)
 461 {
 462         int rc;
 463         ENTRY;
 464
 465         if (!req->rq_repmsg) {
 466                 rc = lustre_pack_reply(req, 1, NULL, NULL);
 467                 if (rc)
 468                         RETURN(rc);
 469         }
 470
 471         req->rq_type = PTL_RPC_MSG_ERR;
 472
 473         rc = ptlrpc_send_reply(req, may_be_difficult);
 474         RETURN(rc);
 475 }
 476
 477 int ptlrpc_error(struct ptlrpc_request *req)
 478 {
 479         return ptlrpc_send_error(req, 0);
 480 }
 481
 482 int ptl_send_rpc(struct ptlrpc_request *request, int noreply)
 483 {
 484         int rc;
 485         int rc2;
 486         struct ptlrpc_connection *connection;
 487         lnet_handle_me_t  reply_me_h;
 488         lnet_md_t         reply_md;
 489         struct obd_device *obd = request->rq_import->imp_obd;
 490         ENTRY;
 491
 492         OBD_FAIL_RETURN(OBD_FAIL_PTLRPC_DROP_RPC, 0);
 493
 494         LASSERT (request->rq_type == PTL_RPC_MSG_REQUEST);
 495
 496         /* If this is a re-transmit, we're required to have disengaged
 497          * cleanly from the previous attempt */
 498         LASSERT(!request->rq_receiving_reply);
 499
 500         if (request->rq_import->imp_obd &&
 501             request->rq_import->imp_obd->obd_fail) {
 502                 CDEBUG(D_HA, "muting rpc for failed imp obd %s\n",
 503                        request->rq_import->imp_obd->obd_name);
 504                 /* this prevents us from waiting in ptlrpc_queue_wait */
 505                 request->rq_err = 1;
 506                 request->rq_status = -ENODEV;
 507                 RETURN(-ENODEV);
 508         }
 509
 510         connection = request->rq_import->imp_connection;
 511
 512         if (request->rq_bulk != NULL) {
 513                 rc = ptlrpc_register_bulk (request);
 514                 if (rc != 0)
 515                         RETURN(rc);
 516         }
 517
 518         lustre_msg_set_handle(request->rq_reqmsg,
 519                               &request->rq_import->imp_remote_handle);
 520         lustre_msg_set_type(request->rq_reqmsg, PTL_RPC_MSG_REQUEST);
 521         lustre_msg_set_conn_cnt(request->rq_reqmsg,
 522                                 request->rq_import->imp_conn_cnt);
 523         lustre_msghdr_set_flags(request->rq_reqmsg,
 524                                 request->rq_import->imp_msghdr_flags);
 525
 526         if (request->rq_resend)
 527                 lustre_msg_add_flags(request->rq_reqmsg, MSG_RESENT);
 528
 529         if (!noreply) {
 530                 LASSERT (request->rq_replen != 0);
 531                 if (request->rq_repbuf == NULL)
 532                         OBD_ALLOC(request->rq_repbuf, request->rq_replen);
 533                 if (request->rq_repbuf == NULL) {
 534                         /* this prevents us from looping in ptlrpc_queue_wait */
 535                         request->rq_err = 1;
 536                         request->rq_status = -ENOMEM;
 537                         GOTO(cleanup_bulk, rc = -ENOMEM);
 538                 }
 539                 request->rq_repmsg = NULL;
 540
 541                 rc = LNetMEAttach(request->rq_reply_portal,/*XXX FIXME bug 249*/
 542                                   connection->c_peer, request->rq_xid, 0,
 543                                   LNET_UNLINK, LNET_INS_AFTER, &reply_me_h);
 544                 if (rc != 0) {
 545                         CERROR("LNetMEAttach failed: %d\n", rc);
 546                         LASSERT (rc == -ENOMEM);
 547                         GOTO(cleanup_repmsg, rc = -ENOMEM);
 548                 }
 549         }
 550
 551         spin_lock(&request->rq_lock);
 552         /* If the MD attach succeeds, there _will_ be a reply_in callback */
 553         request->rq_receiving_reply = !noreply;
 554         /* We are responsible for unlinking the reply buffer */
 555         request->rq_must_unlink = !noreply;
 556         /* Clear any flags that may be present from previous sends. */
 557         request->rq_replied = 0;
 558         request->rq_err = 0;
 559         request->rq_timedout = 0;
 560         request->rq_net_err = 0;
 561         request->rq_resend = 0;
 562         request->rq_restart = 0;
 563         request->rq_rep_swab_mask = 0;
 564         spin_unlock(&request->rq_lock);
 565
 566         if (!noreply) {
 567                 reply_md.start     = request->rq_repbuf;
 568                 reply_md.length    = request->rq_replen;
 569                 /* Allow multiple early replies */
 570                 reply_md.threshold = LNET_MD_THRESH_INF;
 571                 /* Manage remote for early replies */
 572                 reply_md.options   = PTLRPC_MD_OPTIONS | LNET_MD_OP_PUT |
 573                         LNET_MD_MANAGE_REMOTE;
 574                 reply_md.user_ptr  = &request->rq_reply_cbid;
 575                 reply_md.eq_handle = ptlrpc_eq_h;
 576
 577                 /* We must see the unlink callback to unset rq_must_unlink,
 578                    so we can't auto-unlink */
 579                 rc = LNetMDAttach(reply_me_h, reply_md, LNET_RETAIN,
 580                                  &request->rq_reply_md_h);
 581                 if (rc != 0) {
 582                         CERROR("LNetMDAttach failed: %d\n", rc);
 583                         LASSERT (rc == -ENOMEM);
 584                         spin_lock(&request->rq_lock);
 585                         /* ...but the MD attach didn't succeed... */
 586                         request->rq_receiving_reply = 0;
 587                         spin_unlock(&request->rq_lock);
 588                         GOTO(cleanup_me, rc = -ENOMEM);
 589                 }
 590
 591                 CDEBUG(D_NET, "Setup reply buffer: %u bytes, xid "LPU64
 592                        ", portal %u\n",
 593                        request->rq_replen, request->rq_xid,
 594                        request->rq_reply_portal);
 595         }
 596
 597         /* add references on request for request_out_callback */
 598         ptlrpc_request_addref(request);
 599         if (obd->obd_svc_stats != NULL)
 600                 lprocfs_counter_add(obd->obd_svc_stats, PTLRPC_REQACTIVE_CNTR,
 601                                     request->rq_import->imp_inflight.counter);
 602
 603         OBD_FAIL_TIMEOUT(OBD_FAIL_PTLRPC_DELAY_SEND, request->rq_timeout + 5);
 604
 605         request->rq_sent = cfs_time_current_sec();
 606         do_gettimeofday(&request->rq_arrival_time);
 607         /* We give the server rq_timeout secs to process the req, and
 608            add the network latency for our local timeout. */
 609         request->rq_deadline = request->rq_sent + request->rq_timeout +
 610                     ptlrpc_at_get_net_latency(request);
 611
 612         ptlrpc_pinger_sending_on_import(request->rq_import);
 613
 614         DEBUG_REQ(D_INFO, request, "send flg=%x",
 615                   lustre_msg_get_flags(request->rq_reqmsg));
 616         rc = ptl_send_buf(&request->rq_req_md_h,
 617                           request->rq_reqmsg, request->rq_reqlen,
 618                           LNET_NOACK_REQ, &request->rq_req_cbid,
 619                           connection,
 620                           request->rq_request_portal,
 621                           request->rq_xid, 0);
 622         if (rc == 0) {
 623                 ptlrpc_lprocfs_rpc_sent(request);
 624                 RETURN(rc);
 625         }
 626
 627         ptlrpc_req_finished(request);
 628         if (noreply)
 629                 RETURN(rc);
 630
 631  cleanup_me:
 632         /* MEUnlink is safe; the PUT didn't even get off the ground, and
 633          * nobody apart from the PUT's target has the right nid+XID to
 634          * access the reply buffer. */
 635         rc2 = LNetMEUnlink(reply_me_h);
 636         LASSERT (rc2 == 0);
 637         /* UNLINKED callback called synchronously */
 638         LASSERT(!request->rq_receiving_reply);
 639
 640  cleanup_repmsg:
 641         OBD_FREE(request->rq_repbuf, request->rq_replen);
 642         request->rq_repbuf = NULL;
 643         request->rq_repmsg = NULL; //remove
 644
 645  cleanup_bulk:
 646         /* We do sync unlink here as there was no real transfer here so
 647          * the chance to have long unlink to sluggish net is smaller here. */
 648         ptlrpc_unregister_bulk(request, 0);
 649         return rc;
 650 }
 651
 652 int ptlrpc_register_rqbd(struct ptlrpc_request_buffer_desc *rqbd)
 653 {
 654         struct ptlrpc_service   *service = rqbd->rqbd_service;
 655         static lnet_process_id_t  match_id = {LNET_NID_ANY, LNET_PID_ANY};
 656         int                      rc;
 657         lnet_md_t                 md;
 658         lnet_handle_me_t          me_h;
 659
 660         CDEBUG(D_NET, "LNetMEAttach: portal %d\n",
 661                service->srv_req_portal);
 662
 663         if (OBD_FAIL_CHECK_ONCE(OBD_FAIL_PTLRPC_RQBD))
 664                 return (-ENOMEM);
 665
 666         rc = LNetMEAttach(service->srv_req_portal,
 667                           match_id, 0, ~0, LNET_UNLINK, LNET_INS_AFTER, &me_h);
 668         if (rc != 0) {
 669                 CERROR("LNetMEAttach failed: %d\n", rc);
 670                 return (-ENOMEM);
 671         }
 672
 673         LASSERT(rqbd->rqbd_refcount == 0);
 674         rqbd->rqbd_refcount = 1;
 675
 676         md.start     = rqbd->rqbd_buffer;
 677         md.length    = service->srv_buf_size;
 678         md.max_size  = service->srv_max_req_size;
 679         md.threshold = LNET_MD_THRESH_INF;
 680         md.options   = PTLRPC_MD_OPTIONS | LNET_MD_OP_PUT | LNET_MD_MAX_SIZE;
 681         md.user_ptr  = &rqbd->rqbd_cbid;
 682         md.eq_handle = ptlrpc_eq_h;
 683
 684         rc = LNetMDAttach(me_h, md, LNET_UNLINK, &rqbd->rqbd_md_h);
 685         if (rc == 0)
 686                 return (0);
 687
 688         CERROR("LNetMDAttach failed: %d; \n", rc);
 689         LASSERT (rc == -ENOMEM);
 690         rc = LNetMEUnlink (me_h);
 691         LASSERT (rc == 0);
 692         rqbd->rqbd_refcount = 0;
 693
 694         return (-ENOMEM);
 695 }