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