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 /* Server side bulk abort. Idempotent. Not thread-safe (i.e. only
 166  * serialises with completion callback) */
 167 void ptlrpc_abort_bulk(struct ptlrpc_bulk_desc *desc)
 168 {
 169         struct l_wait_info       lwi;
 170         int                      rc;
 171
 172         LASSERT(!in_interrupt());               /* might sleep */
 173
 174         if (!ptlrpc_server_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_INTERVAL(cfs_time_seconds(LONG_UNLINK),
 192                                            cfs_time_seconds(1), NULL, NULL);
 193                 rc = l_wait_event(desc->bd_waitq,
 194                                   !ptlrpc_server_bulk_active(desc), &lwi);
 195                 if (rc == 0)
 196                         return;
 197
 198                 LASSERT(rc == -ETIMEDOUT);
 199                 CWARN("Unexpectedly long timeout: desc %p\n", desc);
 200         }
 201 }
 202
 203 int ptlrpc_register_bulk(struct ptlrpc_request *req)
 204 {
 205         struct ptlrpc_bulk_desc *desc = req->rq_bulk;
 206         lnet_process_id_t peer;
 207         int rc;
 208         int rc2;
 209         lnet_handle_me_t  me_h;
 210         lnet_md_t         md;
 211         ENTRY;
 212
 213         if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_BULK_GET_NET))
 214                 RETURN(0);
 215
 216         /* NB no locking required until desc is on the network */
 217         LASSERT (desc->bd_nob > 0);
 218         LASSERT (!desc->bd_network_rw);
 219         LASSERT (desc->bd_iov_count <= PTLRPC_MAX_BRW_PAGES);
 220         LASSERT (desc->bd_req != NULL);
 221         LASSERT (desc->bd_type == BULK_PUT_SINK ||
 222                  desc->bd_type == BULK_GET_SOURCE);
 223
 224         desc->bd_success = 0;
 225
 226         peer = desc->bd_import->imp_connection->c_peer;
 227
 228         md.user_ptr = &desc->bd_cbid;
 229         md.eq_handle = ptlrpc_eq_h;
 230         md.threshold = 1;                       /* PUT or GET */
 231         md.options = PTLRPC_MD_OPTIONS |
 232                      ((desc->bd_type == BULK_GET_SOURCE) ?
 233                       LNET_MD_OP_GET : LNET_MD_OP_PUT);
 234         ptlrpc_fill_bulk_md(&md, desc);
 235
 236         LASSERT (desc->bd_cbid.cbid_fn == client_bulk_callback);
 237         LASSERT (desc->bd_cbid.cbid_arg == desc);
 238
 239         /* XXX Registering the same xid on retried bulk makes my head
 240          * explode trying to understand how the original request's bulk
 241          * might interfere with the retried request -eeb */
 242         LASSERTF (!desc->bd_registered || req->rq_xid != desc->bd_last_xid,
 243                   "registered: %d  rq_xid: "LPU64" bd_last_xid: "LPU64"\n",
 244                   desc->bd_registered, req->rq_xid, desc->bd_last_xid);
 245         desc->bd_registered = 1;
 246         desc->bd_last_xid = req->rq_xid;
 247
 248         rc = LNetMEAttach(desc->bd_portal, peer,
 249                          req->rq_xid, 0, LNET_UNLINK, LNET_INS_AFTER, &me_h);
 250         if (rc != 0) {
 251                 CERROR("LNetMEAttach failed: %d\n", rc);
 252                 LASSERT (rc == -ENOMEM);
 253                 RETURN (-ENOMEM);
 254         }
 255
 256         /* About to let the network at it... */
 257         desc->bd_network_rw = 1;
 258         rc = LNetMDAttach(me_h, md, LNET_UNLINK, &desc->bd_md_h);
 259         if (rc != 0) {
 260                 CERROR("LNetMDAttach failed: %d\n", rc);
 261                 LASSERT (rc == -ENOMEM);
 262                 desc->bd_network_rw = 0;
 263                 rc2 = LNetMEUnlink (me_h);
 264                 LASSERT (rc2 == 0);
 265                 RETURN (-ENOMEM);
 266         }
 267
 268         CDEBUG(D_NET, "Setup bulk %s buffers: %u pages %u bytes, xid "LPX64", "
 269                "portal %u\n",
 270                desc->bd_type == BULK_GET_SOURCE ? "get-source" : "put-sink",
 271                desc->bd_iov_count, desc->bd_nob,
 272                req->rq_xid, desc->bd_portal);
 273         RETURN(0);
 274 }
 275
 276 /* Disconnect a bulk desc from the network. Idempotent. Not
 277  * thread-safe (i.e. only interlocks with completion callback). */
 278 int ptlrpc_unregister_bulk(struct ptlrpc_request *req, int async)
 279 {
 280         struct ptlrpc_bulk_desc *desc = req->rq_bulk;
 281         cfs_waitq_t             *wq;
 282         struct l_wait_info       lwi;
 283         int                      rc;
 284         ENTRY;
 285
 286         LASSERT(!in_interrupt());     /* might sleep */
 287
 288         /* Let's setup deadline for reply unlink. */
 289         if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_BULK_UNLINK) &&
 290             async && req->rq_bulk_deadline == 0)
 291                 req->rq_bulk_deadline = cfs_time_current_sec() + LONG_UNLINK;
 292
 293         if (!ptlrpc_client_bulk_active(req))  /* completed or */
 294                 RETURN(1);                    /* never registered */
 295
 296         LASSERT(desc->bd_req == req);  /* bd_req NULL until registered */
 297
 298         /* the unlink ensures the callback happens ASAP and is the last
 299          * one.  If it fails, it must be because completion just happened,
 300          * but we must still l_wait_event() in this case to give liblustre
 301          * a chance to run client_bulk_callback() */
 302
 303         LNetMDUnlink(desc->bd_md_h);
 304
 305         if (!ptlrpc_client_bulk_active(req))  /* completed or */
 306                 RETURN(1);                    /* never registered */
 307
 308         /* Move to "Unregistering" phase as bulk was not unlinked yet. */
 309         ptlrpc_rqphase_move(req, RQ_PHASE_UNREGISTERING);
 310
 311         /* Do not wait for unlink to finish. */
 312         if (async)
 313                 RETURN(0);
 314
 315         if (req->rq_set != NULL)
 316                 wq = &req->rq_set->set_waitq;
 317         else
 318                 wq = &req->rq_reply_waitq;
 319
 320         for (;;) {
 321                 /* Network access will complete in finite time but the HUGE
 322                  * timeout lets us CWARN for visibility of sluggish NALs */
 323                 lwi = LWI_TIMEOUT_INTERVAL(cfs_time_seconds(LONG_UNLINK),
 324                                            cfs_time_seconds(1), NULL, NULL);
 325                 rc = l_wait_event(*wq, !ptlrpc_client_bulk_active(req), &lwi);
 326                 if (rc == 0) {
 327                         ptlrpc_rqphase_move(req, req->rq_next_phase);
 328                         RETURN(1);
 329                 }
 330
 331                 LASSERT(rc == -ETIMEDOUT);
 332                 DEBUG_REQ(D_WARNING, req, "Unexpectedly long timeout: desc %p",
 333                           desc);
 334         }
 335         RETURN(0);
 336 }
 337
 338 static void ptlrpc_at_set_reply(struct ptlrpc_request *req, int flags)
 339 {
 340         struct ptlrpc_service *svc = req->rq_rqbd->rqbd_service;
 341         int service_time = max_t(int, cfs_time_current_sec() -
 342                                  req->rq_arrival_time.tv_sec, 1);
 343
 344         if (!(flags & PTLRPC_REPLY_EARLY) &&
 345             (req->rq_type != PTL_RPC_MSG_ERR) &&
 346             !(lustre_msg_get_flags(req->rq_reqmsg) &
 347               (MSG_RESENT | MSG_REPLAY | MSG_LAST_REPLAY))) {
 348                 /* early replies, errors and recovery requests don't count
 349                  * toward our service time estimate */
 350                 int oldse = at_add(&svc->srv_at_estimate, service_time);
 351                 if (oldse != 0)
 352                         DEBUG_REQ(D_ADAPTTO, req,
 353                                   "svc %s changed estimate from %d to %d",
 354                                   svc->srv_name, oldse,
 355                                   at_get(&svc->srv_at_estimate));
 356         }
 357         /* Report actual service time for client latency calc */
 358         lustre_msg_set_service_time(req->rq_repmsg, service_time);
 359         /* Report service time estimate for future client reqs, but report 0
 360          * (to be ignored by client) if it's a error reply during recovery.
 361          * (bz15815) */
 362         if (req->rq_type == PTL_RPC_MSG_ERR &&
 363             (req->rq_export == NULL || req->rq_export->exp_obd->obd_recovering))
 364                 lustre_msg_set_timeout(req->rq_repmsg, 0);
 365         else
 366                 lustre_msg_set_timeout(req->rq_repmsg,
 367                                        at_get(&svc->srv_at_estimate));
 368
 369         if (req->rq_reqmsg &&
 370             !(lustre_msghdr_get_flags(req->rq_reqmsg) & MSGHDR_AT_SUPPORT)) {
 371                 CDEBUG(D_ADAPTTO, "No early reply support: flags=%#x "
 372                        "req_flags=%#x magic=%d:%x/%x len=%d\n",
 373                        flags, lustre_msg_get_flags(req->rq_reqmsg),
 374                        lustre_msg_is_v1(req->rq_reqmsg),
 375                        lustre_msg_get_magic(req->rq_reqmsg),
 376                        lustre_msg_get_magic(req->rq_repmsg), req->rq_replen);
 377         }
 378 }
 379
 380 int ptlrpc_send_reply(struct ptlrpc_request *req, int flags)
 381 {
 382         struct ptlrpc_service     *svc = req->rq_rqbd->rqbd_service;
 383         struct ptlrpc_reply_state *rs = req->rq_reply_state;
 384         struct ptlrpc_connection  *conn;
 385         int                        rc;
 386
 387         /* We must already have a reply buffer (only ptlrpc_error() may be
 388          * called without one). The reply generated by sptlrpc layer (e.g.
 389          * error notify, etc.) might have NULL rq->reqmsg; Otherwise we must
 390          * have a request buffer which is either the actual (swabbed) incoming
 391          * request, or a saved copy if this is a req saved in
 392          * target_queue_final_reply().
 393          */
 394         LASSERT (req->rq_no_reply == 0);
 395         LASSERT (req->rq_reqbuf != NULL);
 396         LASSERT (rs != NULL);
 397         LASSERT ((flags & PTLRPC_REPLY_MAYBE_DIFFICULT) || !rs->rs_difficult);
 398         LASSERT (req->rq_repmsg != NULL);
 399         LASSERT (req->rq_repmsg == rs->rs_msg);
 400         LASSERT (rs->rs_cb_id.cbid_fn == reply_out_callback);
 401         LASSERT (rs->rs_cb_id.cbid_arg == rs);
 402
 403         /* There may be no rq_export during failover */
 404
 405         if (unlikely(req->rq_export && req->rq_export->exp_obd &&
 406                      req->rq_export->exp_obd->obd_fail)) {
 407                 /* Failed obd's only send ENODEV */
 408                 req->rq_type = PTL_RPC_MSG_ERR;
 409                 req->rq_status = -ENODEV;
 410                 CDEBUG(D_HA, "sending ENODEV from failed obd %d\n",
 411                        req->rq_export->exp_obd->obd_minor);
 412         }
 413
 414         if (req->rq_type != PTL_RPC_MSG_ERR)
 415                 req->rq_type = PTL_RPC_MSG_REPLY;
 416
 417         lustre_msg_set_type(req->rq_repmsg, req->rq_type);
 418         lustre_msg_set_status(req->rq_repmsg, req->rq_status);
 419         lustre_msg_set_opc(req->rq_repmsg,
 420                 req->rq_reqmsg ? lustre_msg_get_opc(req->rq_reqmsg) : 0);
 421
 422         target_pack_pool_reply(req);
 423
 424         ptlrpc_at_set_reply(req, flags);
 425
 426         if (req->rq_export == NULL || req->rq_export->exp_connection == NULL)
 427                 conn = ptlrpc_connection_get(req->rq_peer, req->rq_self, NULL);
 428         else
 429                 conn = ptlrpc_connection_addref(req->rq_export->exp_connection);
 430
 431         if (unlikely(conn == NULL)) {
 432                 CERROR("not replying on NULL connection\n"); /* bug 9635 */
 433                 return -ENOTCONN;
 434         }
 435         atomic_inc (&svc->srv_outstanding_replies);
 436         ptlrpc_rs_addref(rs);                   /* +1 ref for the network */
 437
 438         rc = sptlrpc_svc_wrap_reply(req);
 439         if (unlikely(rc))
 440                 goto out;
 441
 442         req->rq_sent = cfs_time_current_sec();
 443
 444         rc = ptl_send_buf (&rs->rs_md_h, rs->rs_repbuf, rs->rs_repdata_len,
 445                            (rs->rs_difficult && !rs->rs_no_ack) ?
 446                            LNET_ACK_REQ : LNET_NOACK_REQ,
 447                            &rs->rs_cb_id, conn, svc->srv_rep_portal,
 448                            req->rq_xid, req->rq_reply_off);
 449 out:
 450         if (unlikely(rc != 0)) {
 451                 atomic_dec (&svc->srv_outstanding_replies);
 452                 ptlrpc_req_drop_rs(req);
 453         }
 454         ptlrpc_connection_put(conn);
 455         return rc;
 456 }
 457
 458 int ptlrpc_reply (struct ptlrpc_request *req)
 459 {
 460         if (req->rq_no_reply)
 461                 return 0;
 462         else
 463                 return (ptlrpc_send_reply(req, 0));
 464 }
 465
 466 int ptlrpc_send_error(struct ptlrpc_request *req, int may_be_difficult)
 467 {
 468         int rc;
 469         ENTRY;
 470
 471         if (req->rq_no_reply)
 472                 RETURN(0);
 473
 474         if (!req->rq_repmsg) {
 475                 rc = lustre_pack_reply(req, 1, NULL, NULL);
 476                 if (rc)
 477                         RETURN(rc);
 478         }
 479
 480         req->rq_type = PTL_RPC_MSG_ERR;
 481
 482         rc = ptlrpc_send_reply(req, may_be_difficult);
 483         RETURN(rc);
 484 }
 485
 486 int ptlrpc_error(struct ptlrpc_request *req)
 487 {
 488         return ptlrpc_send_error(req, 0);
 489 }
 490
 491 int ptl_send_rpc(struct ptlrpc_request *request, int noreply)
 492 {
 493         int rc;
 494         int rc2;
 495         struct ptlrpc_connection *connection;
 496         lnet_handle_me_t  reply_me_h;
 497         lnet_md_t         reply_md;
 498         struct obd_device *obd = request->rq_import->imp_obd;
 499         ENTRY;
 500
 501         if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_DROP_RPC))
 502                 RETURN(0);
 503
 504         LASSERT(request->rq_type == PTL_RPC_MSG_REQUEST);
 505         LASSERT(request->rq_wait_ctx == 0);
 506
 507         /* If this is a re-transmit, we're required to have disengaged
 508          * cleanly from the previous attempt */
 509         LASSERT(!request->rq_receiving_reply);
 510
 511         if (request->rq_import->imp_obd &&
 512             request->rq_import->imp_obd->obd_fail) {
 513                 CDEBUG(D_HA, "muting rpc for failed imp obd %s\n",
 514                        request->rq_import->imp_obd->obd_name);
 515                 /* this prevents us from waiting in ptlrpc_queue_wait */
 516                 request->rq_err = 1;
 517                 request->rq_status = -ENODEV;
 518                 RETURN(-ENODEV);
 519         }
 520
 521         connection = request->rq_import->imp_connection;
 522
 523         lustre_msg_set_handle(request->rq_reqmsg,
 524                               &request->rq_import->imp_remote_handle);
 525         lustre_msg_set_type(request->rq_reqmsg, PTL_RPC_MSG_REQUEST);
 526         lustre_msg_set_conn_cnt(request->rq_reqmsg,
 527                                 request->rq_import->imp_conn_cnt);
 528         lustre_msghdr_set_flags(request->rq_reqmsg,
 529                                 request->rq_import->imp_msghdr_flags);
 530
 531         rc = sptlrpc_cli_wrap_request(request);
 532         if (rc)
 533                 RETURN(rc);
 534
 535         /* bulk register should be done after wrap_request() */
 536         if (request->rq_bulk != NULL) {
 537                 rc = ptlrpc_register_bulk (request);
 538                 if (rc != 0)
 539                         RETURN(rc);
 540         }
 541
 542         if (!noreply) {
 543                 LASSERT (request->rq_replen != 0);
 544                 if (request->rq_repbuf == NULL) {
 545                         LASSERT(request->rq_repdata == NULL);
 546                         LASSERT(request->rq_repmsg == NULL);
 547                         rc = sptlrpc_cli_alloc_repbuf(request,
 548                                                       request->rq_replen);
 549                         if (rc) {
 550                                 /* this prevents us from looping in
 551                                  * ptlrpc_queue_wait */
 552                                 request->rq_err = 1;
 553                                 request->rq_status = rc;
 554                                 GOTO(cleanup_bulk, rc);
 555                         }
 556                 } else {
 557                         request->rq_repdata = NULL;
 558                         request->rq_repmsg = NULL;
 559                 }
 560
 561                 rc = LNetMEAttach(request->rq_reply_portal,/*XXX FIXME bug 249*/
 562                                   connection->c_peer, request->rq_xid, 0,
 563                                   LNET_UNLINK, LNET_INS_AFTER, &reply_me_h);
 564                 if (rc != 0) {
 565                         CERROR("LNetMEAttach failed: %d\n", rc);
 566                         LASSERT (rc == -ENOMEM);
 567                         GOTO(cleanup_bulk, rc = -ENOMEM);
 568                 }
 569         }
 570
 571         spin_lock(&request->rq_lock);
 572         /* If the MD attach succeeds, there _will_ be a reply_in callback */
 573         request->rq_receiving_reply = !noreply;
 574         /* We are responsible for unlinking the reply buffer */
 575         request->rq_must_unlink = !noreply;
 576         /* Clear any flags that may be present from previous sends. */
 577         request->rq_replied = 0;
 578         request->rq_err = 0;
 579         request->rq_timedout = 0;
 580         request->rq_net_err = 0;
 581         request->rq_resend = 0;
 582         request->rq_restart = 0;
 583         spin_unlock(&request->rq_lock);
 584
 585         if (!noreply) {
 586                 reply_md.start     = request->rq_repbuf;
 587                 reply_md.length    = request->rq_repbuf_len;
 588                 /* Allow multiple early replies */
 589                 reply_md.threshold = LNET_MD_THRESH_INF;
 590                 /* Manage remote for early replies */
 591                 reply_md.options   = PTLRPC_MD_OPTIONS | LNET_MD_OP_PUT |
 592                         LNET_MD_MANAGE_REMOTE;
 593                 reply_md.user_ptr  = &request->rq_reply_cbid;
 594                 reply_md.eq_handle = ptlrpc_eq_h;
 595
 596                 /* We must see the unlink callback to unset rq_must_unlink,
 597                    so we can't auto-unlink */
 598                 rc = LNetMDAttach(reply_me_h, reply_md, LNET_RETAIN,
 599                                   &request->rq_reply_md_h);
 600                 if (rc != 0) {
 601                         CERROR("LNetMDAttach failed: %d\n", rc);
 602                         LASSERT (rc == -ENOMEM);
 603                         spin_lock(&request->rq_lock);
 604                         /* ...but the MD attach didn't succeed... */
 605                         request->rq_receiving_reply = 0;
 606                         spin_unlock(&request->rq_lock);
 607                         GOTO(cleanup_me, rc = -ENOMEM);
 608                 }
 609
 610                 CDEBUG(D_NET, "Setup reply buffer: %u bytes, xid "LPU64
 611                        ", portal %u\n",
 612                        request->rq_repbuf_len, request->rq_xid,
 613                        request->rq_reply_portal);
 614         }
 615
 616         /* add references on request for request_out_callback */
 617         ptlrpc_request_addref(request);
 618         if (obd->obd_svc_stats != NULL)
 619                 lprocfs_counter_add(obd->obd_svc_stats, PTLRPC_REQACTIVE_CNTR,
 620                                     request->rq_import->imp_inflight.counter);
 621
 622         OBD_FAIL_TIMEOUT(OBD_FAIL_PTLRPC_DELAY_SEND, request->rq_timeout + 5);
 623
 624         do_gettimeofday(&request->rq_arrival_time);
 625         request->rq_sent = cfs_time_current_sec();
 626         /* We give the server rq_timeout secs to process the req, and
 627            add the network latency for our local timeout. */
 628         request->rq_deadline = request->rq_sent + request->rq_timeout +
 629                 ptlrpc_at_get_net_latency(request);
 630
 631         ptlrpc_pinger_sending_on_import(request->rq_import);
 632
 633         DEBUG_REQ(D_INFO, request, "send flg=%x",
 634                   lustre_msg_get_flags(request->rq_reqmsg));
 635         rc = ptl_send_buf(&request->rq_req_md_h,
 636                           request->rq_reqbuf, request->rq_reqdata_len,
 637                           LNET_NOACK_REQ, &request->rq_req_cbid,
 638                           connection,
 639                           request->rq_request_portal,
 640                           request->rq_xid, 0);
 641         if (rc == 0) {
 642                 ptlrpc_lprocfs_rpc_sent(request);
 643                 RETURN(rc);
 644         }
 645
 646         ptlrpc_req_finished(request);
 647         if (noreply)
 648                 RETURN(rc);
 649
 650  cleanup_me:
 651         /* MEUnlink is safe; the PUT didn't even get off the ground, and
 652          * nobody apart from the PUT's target has the right nid+XID to
 653          * access the reply buffer. */
 654         rc2 = LNetMEUnlink(reply_me_h);
 655         LASSERT (rc2 == 0);
 656         /* UNLINKED callback called synchronously */
 657         LASSERT(!request->rq_receiving_reply);
 658
 659  cleanup_bulk:
 660         /* We do sync unlink here as there was no real transfer here so
 661          * the chance to have long unlink to sluggish net is smaller here. */
 662         ptlrpc_unregister_bulk(request, 0);
 663         return rc;
 664 }
 665
 666 int ptlrpc_register_rqbd(struct ptlrpc_request_buffer_desc *rqbd)
 667 {
 668         struct ptlrpc_service   *service = rqbd->rqbd_service;
 669         static lnet_process_id_t  match_id = {LNET_NID_ANY, LNET_PID_ANY};
 670         int                      rc;
 671         lnet_md_t                 md;
 672         lnet_handle_me_t          me_h;
 673
 674         CDEBUG(D_NET, "LNetMEAttach: portal %d\n",
 675                service->srv_req_portal);
 676
 677         if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_RQBD))
 678                 return (-ENOMEM);
 679
 680         rc = LNetMEAttach(service->srv_req_portal,
 681                           match_id, 0, ~0, LNET_UNLINK, LNET_INS_AFTER, &me_h);
 682         if (rc != 0) {
 683                 CERROR("LNetMEAttach failed: %d\n", rc);
 684                 return (-ENOMEM);
 685         }
 686
 687         LASSERT(rqbd->rqbd_refcount == 0);
 688         rqbd->rqbd_refcount = 1;
 689
 690         md.start     = rqbd->rqbd_buffer;
 691         md.length    = service->srv_buf_size;
 692         md.max_size  = service->srv_max_req_size;
 693         md.threshold = LNET_MD_THRESH_INF;
 694         md.options   = PTLRPC_MD_OPTIONS | LNET_MD_OP_PUT | LNET_MD_MAX_SIZE;
 695         md.user_ptr  = &rqbd->rqbd_cbid;
 696         md.eq_handle = ptlrpc_eq_h;
 697
 698         rc = LNetMDAttach(me_h, md, LNET_UNLINK, &rqbd->rqbd_md_h);
 699         if (rc == 0)
 700                 return (0);
 701
 702         CERROR("LNetMDAttach failed: %d; \n", rc);
 703         LASSERT (rc == -ENOMEM);
 704         rc = LNetMEUnlink (me_h);
 705         LASSERT (rc == 0);
 706         rqbd->rqbd_refcount = 0;
 707
 708         return (-ENOMEM);
 709 }