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