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(!cfs_in_interrupt());           /* might sleep */
 173
 174         if (!ptlrpc_server_bulk_active(desc))   /* completed or */
 175                 return;                         /* never started */
 176
 177         /* We used to poison the pages with 0xab here because we did not want to
 178          * send any meaningful data over the wire for evicted clients (bug 9297)
 179          * However, this is no longer safe now that we use the page cache on the
 180          * OSS (bug 20560) */
 181
 182         /* The unlink ensures the callback happens ASAP and is the last
 183          * one.  If it fails, it must be because completion just happened,
 184          * but we must still l_wait_event() in this case, to give liblustre
 185          * a chance to run server_bulk_callback()*/
 186
 187         LNetMDUnlink(desc->bd_md_h);
 188
 189         for (;;) {
 190                 /* Network access will complete in finite time but the HUGE
 191                  * timeout lets us CWARN for visibility of sluggish NALs */
 192                 lwi = LWI_TIMEOUT_INTERVAL(cfs_time_seconds(LONG_UNLINK),
 193                                            cfs_time_seconds(1), NULL, NULL);
 194                 rc = l_wait_event(desc->bd_waitq,
 195                                   !ptlrpc_server_bulk_active(desc), &lwi);
 196                 if (rc == 0)
 197                         return;
 198
 199                 LASSERT(rc == -ETIMEDOUT);
 200                 CWARN("Unexpectedly long timeout: desc %p\n", desc);
 201         }
 202 }
 203
 204 int ptlrpc_register_bulk(struct ptlrpc_request *req)
 205 {
 206         struct ptlrpc_bulk_desc *desc = req->rq_bulk;
 207         lnet_process_id_t peer;
 208         int rc;
 209         int rc2;
 210         lnet_handle_me_t  me_h;
 211         lnet_md_t         md;
 212         ENTRY;
 213
 214         if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_BULK_GET_NET))
 215                 RETURN(0);
 216
 217         /* NB no locking required until desc is on the network */
 218         LASSERT (desc->bd_nob > 0);
 219         LASSERT (!desc->bd_network_rw);
 220         LASSERT (desc->bd_iov_count <= PTLRPC_MAX_BRW_PAGES);
 221         LASSERT (desc->bd_req != NULL);
 222         LASSERT (desc->bd_type == BULK_PUT_SINK ||
 223                  desc->bd_type == BULK_GET_SOURCE);
 224
 225         desc->bd_success = 0;
 226
 227         peer = desc->bd_import->imp_connection->c_peer;
 228
 229         md.user_ptr = &desc->bd_cbid;
 230         md.eq_handle = ptlrpc_eq_h;
 231         md.threshold = 1;                       /* PUT or GET */
 232         md.options = PTLRPC_MD_OPTIONS |
 233                      ((desc->bd_type == BULK_GET_SOURCE) ?
 234                       LNET_MD_OP_GET : LNET_MD_OP_PUT);
 235         ptlrpc_fill_bulk_md(&md, desc);
 236
 237         LASSERT (desc->bd_cbid.cbid_fn == client_bulk_callback);
 238         LASSERT (desc->bd_cbid.cbid_arg == desc);
 239
 240         /* XXX Registering the same xid on retried bulk makes my head
 241          * explode trying to understand how the original request's bulk
 242          * might interfere with the retried request -eeb */
 243         LASSERTF (!desc->bd_registered || req->rq_xid != desc->bd_last_xid,
 244                   "registered: %d  rq_xid: "LPU64" bd_last_xid: "LPU64"\n",
 245                   desc->bd_registered, req->rq_xid, desc->bd_last_xid);
 246         desc->bd_registered = 1;
 247         desc->bd_last_xid = req->rq_xid;
 248
 249         rc = LNetMEAttach(desc->bd_portal, peer,
 250                          req->rq_xid, 0, LNET_UNLINK, LNET_INS_AFTER, &me_h);
 251         if (rc != 0) {
 252                 CERROR("LNetMEAttach failed: %d\n", rc);
 253                 LASSERT (rc == -ENOMEM);
 254                 RETURN (-ENOMEM);
 255         }
 256
 257         /* About to let the network at it... */
 258         desc->bd_network_rw = 1;
 259         rc = LNetMDAttach(me_h, md, LNET_UNLINK, &desc->bd_md_h);
 260         if (rc != 0) {
 261                 CERROR("LNetMDAttach failed: %d\n", rc);
 262                 LASSERT (rc == -ENOMEM);
 263                 desc->bd_network_rw = 0;
 264                 rc2 = LNetMEUnlink (me_h);
 265                 LASSERT (rc2 == 0);
 266                 RETURN (-ENOMEM);
 267         }
 268
 269         CDEBUG(D_NET, "Setup bulk %s buffers: %u pages %u bytes, xid "LPU64", "
 270                "portal %u\n",
 271                desc->bd_type == BULK_GET_SOURCE ? "get-source" : "put-sink",
 272                desc->bd_iov_count, desc->bd_nob,
 273                req->rq_xid, desc->bd_portal);
 274         RETURN(0);
 275 }
 276
 277 /* Disconnect a bulk desc from the network. Idempotent. Not
 278  * thread-safe (i.e. only interlocks with completion callback). */
 279 int ptlrpc_unregister_bulk(struct ptlrpc_request *req, int async)
 280 {
 281         struct ptlrpc_bulk_desc *desc = req->rq_bulk;
 282         cfs_waitq_t             *wq;
 283         struct l_wait_info       lwi;
 284         int                      rc;
 285         ENTRY;
 286
 287         LASSERT(!cfs_in_interrupt());     /* might sleep */
 288
 289         /* Let's setup deadline for reply unlink. */
 290         if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_BULK_UNLINK) &&
 291             async && req->rq_bulk_deadline == 0)
 292                 req->rq_bulk_deadline = cfs_time_current_sec() + LONG_UNLINK;
 293
 294         if (!ptlrpc_client_bulk_active(req))  /* completed or */
 295                 RETURN(1);                    /* never registered */
 296
 297         LASSERT(desc->bd_req == req);  /* bd_req NULL until registered */
 298
 299         /* the unlink ensures the callback happens ASAP and is the last
 300          * one.  If it fails, it must be because completion just happened,
 301          * but we must still l_wait_event() in this case to give liblustre
 302          * a chance to run client_bulk_callback() */
 303
 304         LNetMDUnlink(desc->bd_md_h);
 305
 306         if (!ptlrpc_client_bulk_active(req))  /* completed or */
 307                 RETURN(1);                    /* never registered */
 308
 309         /* Move to "Unregistering" phase as bulk was not unlinked yet. */
 310         ptlrpc_rqphase_move(req, RQ_PHASE_UNREGISTERING);
 311
 312         /* Do not wait for unlink to finish. */
 313         if (async)
 314                 RETURN(0);
 315
 316         if (req->rq_set != NULL)
 317                 wq = &req->rq_set->set_waitq;
 318         else
 319                 wq = &req->rq_reply_waitq;
 320
 321         for (;;) {
 322                 /* Network access will complete in finite time but the HUGE
 323                  * timeout lets us CWARN for visibility of sluggish NALs */
 324                 lwi = LWI_TIMEOUT_INTERVAL(cfs_time_seconds(LONG_UNLINK),
 325                                            cfs_time_seconds(1), NULL, NULL);
 326                 rc = l_wait_event(*wq, !ptlrpc_client_bulk_active(req), &lwi);
 327                 if (rc == 0) {
 328                         ptlrpc_rqphase_move(req, req->rq_next_phase);
 329                         RETURN(1);
 330                 }
 331
 332                 LASSERT(rc == -ETIMEDOUT);
 333                 DEBUG_REQ(D_WARNING, req, "Unexpectedly long timeout: desc %p",
 334                           desc);
 335         }
 336         RETURN(0);
 337 }
 338
 339 static void ptlrpc_at_set_reply(struct ptlrpc_request *req, int flags)
 340 {
 341         struct ptlrpc_service *svc = req->rq_rqbd->rqbd_service;
 342         int service_time = max_t(int, cfs_time_current_sec() -
 343                                  req->rq_arrival_time.tv_sec, 1);
 344
 345         if (!(flags & PTLRPC_REPLY_EARLY) &&
 346             (req->rq_type != PTL_RPC_MSG_ERR) &&
 347             (req->rq_reqmsg != NULL) &&
 348             !(lustre_msg_get_flags(req->rq_reqmsg) &
 349               (MSG_RESENT | MSG_REPLAY |
 350                MSG_REQ_REPLAY_DONE | MSG_LOCK_REPLAY_DONE))) {
 351                 /* early replies, errors and recovery requests don't count
 352                  * toward our service time estimate */
 353                 int oldse = at_add(&svc->srv_at_estimate, service_time);
 354                 if (oldse != 0)
 355                         DEBUG_REQ(D_ADAPTTO, req,
 356                                   "svc %s changed estimate from %d to %d",
 357                                   svc->srv_name, oldse,
 358                                   at_get(&svc->srv_at_estimate));
 359         }
 360         /* Report actual service time for client latency calc */
 361         lustre_msg_set_service_time(req->rq_repmsg, service_time);
 362         /* Report service time estimate for future client reqs, but report 0
 363          * (to be ignored by client) if it's a error reply during recovery.
 364          * (bz15815) */
 365         if (req->rq_type == PTL_RPC_MSG_ERR &&
 366             (req->rq_export == NULL || req->rq_export->exp_obd->obd_recovering))
 367                 lustre_msg_set_timeout(req->rq_repmsg, 0);
 368         else
 369                 lustre_msg_set_timeout(req->rq_repmsg,
 370                                        at_get(&svc->srv_at_estimate));
 371
 372         if (req->rq_reqmsg &&
 373             !(lustre_msghdr_get_flags(req->rq_reqmsg) & MSGHDR_AT_SUPPORT)) {
 374                 CDEBUG(D_ADAPTTO, "No early reply support: flags=%#x "
 375                        "req_flags=%#x magic=%d:%x/%x len=%d\n",
 376                        flags, lustre_msg_get_flags(req->rq_reqmsg),
 377                        lustre_msg_is_v1(req->rq_reqmsg),
 378                        lustre_msg_get_magic(req->rq_reqmsg),
 379                        lustre_msg_get_magic(req->rq_repmsg), req->rq_replen);
 380         }
 381 }
 382
 383 int ptlrpc_send_reply(struct ptlrpc_request *req, int flags)
 384 {
 385         struct ptlrpc_service     *svc = req->rq_rqbd->rqbd_service;
 386         struct ptlrpc_reply_state *rs = req->rq_reply_state;
 387         struct ptlrpc_connection  *conn;
 388         int                        rc;
 389
 390         /* We must already have a reply buffer (only ptlrpc_error() may be
 391          * called without one). The reply generated by sptlrpc layer (e.g.
 392          * error notify, etc.) might have NULL rq->reqmsg; Otherwise we must
 393          * have a request buffer which is either the actual (swabbed) incoming
 394          * request, or a saved copy if this is a req saved in
 395          * target_queue_final_reply().
 396          */
 397         LASSERT (req->rq_no_reply == 0);
 398         LASSERT (req->rq_reqbuf != NULL);
 399         LASSERT (rs != NULL);
 400         LASSERT ((flags & PTLRPC_REPLY_MAYBE_DIFFICULT) || !rs->rs_difficult);
 401         LASSERT (req->rq_repmsg != NULL);
 402         LASSERT (req->rq_repmsg == rs->rs_msg);
 403         LASSERT (rs->rs_cb_id.cbid_fn == reply_out_callback);
 404         LASSERT (rs->rs_cb_id.cbid_arg == rs);
 405
 406         /* There may be no rq_export during failover */
 407
 408         if (unlikely(req->rq_export && req->rq_export->exp_obd &&
 409                      req->rq_export->exp_obd->obd_fail)) {
 410                 /* Failed obd's only send ENODEV */
 411                 req->rq_type = PTL_RPC_MSG_ERR;
 412                 req->rq_status = -ENODEV;
 413                 CDEBUG(D_HA, "sending ENODEV from failed obd %d\n",
 414                        req->rq_export->exp_obd->obd_minor);
 415         }
 416
 417         if (req->rq_type != PTL_RPC_MSG_ERR)
 418                 req->rq_type = PTL_RPC_MSG_REPLY;
 419
 420         lustre_msg_set_type(req->rq_repmsg, req->rq_type);
 421         lustre_msg_set_status(req->rq_repmsg, req->rq_status);
 422         lustre_msg_set_opc(req->rq_repmsg,
 423                 req->rq_reqmsg ? lustre_msg_get_opc(req->rq_reqmsg) : 0);
 424
 425         target_pack_pool_reply(req);
 426
 427         ptlrpc_at_set_reply(req, flags);
 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 (unlikely(conn == NULL)) {
 435                 CERROR("not replying on NULL connection\n"); /* bug 9635 */
 436                 return -ENOTCONN;
 437         }
 438         cfs_atomic_inc (&svc->srv_outstanding_replies);
 439         ptlrpc_rs_addref(rs);                   /* +1 ref for the network */
 440
 441         rc = sptlrpc_svc_wrap_reply(req);
 442         if (unlikely(rc))
 443                 goto out;
 444
 445         req->rq_sent = cfs_time_current_sec();
 446
 447         rc = ptl_send_buf (&rs->rs_md_h, rs->rs_repbuf, rs->rs_repdata_len,
 448                            (rs->rs_difficult && !rs->rs_no_ack) ?
 449                            LNET_ACK_REQ : LNET_NOACK_REQ,
 450                            &rs->rs_cb_id, conn, svc->srv_rep_portal,
 451                            req->rq_xid, req->rq_reply_off);
 452 out:
 453         if (unlikely(rc != 0)) {
 454                 cfs_atomic_dec (&svc->srv_outstanding_replies);
 455                 ptlrpc_req_drop_rs(req);
 456         }
 457         ptlrpc_connection_put(conn);
 458         return rc;
 459 }
 460
 461 int ptlrpc_reply (struct ptlrpc_request *req)
 462 {
 463         if (req->rq_no_reply)
 464                 return 0;
 465         else
 466                 return (ptlrpc_send_reply(req, 0));
 467 }
 468
 469 int ptlrpc_send_error(struct ptlrpc_request *req, int may_be_difficult)
 470 {
 471         int rc;
 472         ENTRY;
 473
 474         if (req->rq_no_reply)
 475                 RETURN(0);
 476
 477         if (!req->rq_repmsg) {
 478                 rc = lustre_pack_reply(req, 1, NULL, NULL);
 479                 if (rc)
 480                         RETURN(rc);
 481         }
 482
 483         req->rq_type = PTL_RPC_MSG_ERR;
 484
 485         rc = ptlrpc_send_reply(req, may_be_difficult);
 486         RETURN(rc);
 487 }
 488
 489 int ptlrpc_error(struct ptlrpc_request *req)
 490 {
 491         return ptlrpc_send_error(req, 0);
 492 }
 493
 494 int ptl_send_rpc(struct ptlrpc_request *request, int noreply)
 495 {
 496         int rc;
 497         int rc2;
 498         struct ptlrpc_connection *connection;
 499         lnet_handle_me_t  reply_me_h;
 500         lnet_md_t         reply_md;
 501         struct obd_device *obd = request->rq_import->imp_obd;
 502         ENTRY;
 503
 504         if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_DROP_RPC))
 505                 RETURN(0);
 506
 507         LASSERT(request->rq_type == PTL_RPC_MSG_REQUEST);
 508         LASSERT(request->rq_wait_ctx == 0);
 509
 510         /* If this is a re-transmit, we're required to have disengaged
 511          * cleanly from the previous attempt */
 512         LASSERT(!request->rq_receiving_reply);
 513
 514         if (request->rq_import->imp_obd &&
 515             request->rq_import->imp_obd->obd_fail) {
 516                 CDEBUG(D_HA, "muting rpc for failed imp obd %s\n",
 517                        request->rq_import->imp_obd->obd_name);
 518                 /* this prevents us from waiting in ptlrpc_queue_wait */
 519                 request->rq_err = 1;
 520                 request->rq_status = -ENODEV;
 521                 RETURN(-ENODEV);
 522         }
 523
 524         connection = request->rq_import->imp_connection;
 525
 526         lustre_msg_set_handle(request->rq_reqmsg,
 527                               &request->rq_import->imp_remote_handle);
 528         lustre_msg_set_type(request->rq_reqmsg, PTL_RPC_MSG_REQUEST);
 529         lustre_msg_set_conn_cnt(request->rq_reqmsg,
 530                                 request->rq_import->imp_conn_cnt);
 531         lustre_msghdr_set_flags(request->rq_reqmsg,
 532                                 request->rq_import->imp_msghdr_flags);
 533
 534         if (request->rq_resend)
 535                 lustre_msg_add_flags(request->rq_reqmsg, MSG_RESENT);
 536
 537         rc = sptlrpc_cli_wrap_request(request);
 538         if (rc)
 539                 RETURN(rc);
 540
 541         /* bulk register should be done after wrap_request() */
 542         if (request->rq_bulk != NULL) {
 543                 rc = ptlrpc_register_bulk (request);
 544                 if (rc != 0)
 545                         RETURN(rc);
 546         }
 547
 548         if (!noreply) {
 549                 LASSERT (request->rq_replen != 0);
 550                 if (request->rq_repbuf == NULL) {
 551                         LASSERT(request->rq_repdata == NULL);
 552                         LASSERT(request->rq_repmsg == NULL);
 553                         rc = sptlrpc_cli_alloc_repbuf(request,
 554                                                       request->rq_replen);
 555                         if (rc) {
 556                                 /* this prevents us from looping in
 557                                  * ptlrpc_queue_wait */
 558                                 request->rq_err = 1;
 559                                 request->rq_status = rc;
 560                                 GOTO(cleanup_bulk, rc);
 561                         }
 562                 } else {
 563                         request->rq_repdata = NULL;
 564                         request->rq_repmsg = NULL;
 565                 }
 566
 567                 rc = LNetMEAttach(request->rq_reply_portal,/*XXX FIXME bug 249*/
 568                                   connection->c_peer, request->rq_xid, 0,
 569                                   LNET_UNLINK, LNET_INS_AFTER, &reply_me_h);
 570                 if (rc != 0) {
 571                         CERROR("LNetMEAttach failed: %d\n", rc);
 572                         LASSERT (rc == -ENOMEM);
 573                         GOTO(cleanup_bulk, rc = -ENOMEM);
 574                 }
 575         }
 576
 577         cfs_spin_lock(&request->rq_lock);
 578         /* If the MD attach succeeds, there _will_ be a reply_in callback */
 579         request->rq_receiving_reply = !noreply;
 580         /* We are responsible for unlinking the reply buffer */
 581         request->rq_must_unlink = !noreply;
 582         /* Clear any flags that may be present from previous sends. */
 583         request->rq_replied = 0;
 584         request->rq_err = 0;
 585         request->rq_timedout = 0;
 586         request->rq_net_err = 0;
 587         request->rq_resend = 0;
 588         request->rq_restart = 0;
 589         request->rq_reply_truncate = 0;
 590         cfs_spin_unlock(&request->rq_lock);
 591
 592         if (!noreply) {
 593                 reply_md.start     = request->rq_repbuf;
 594                 reply_md.length    = request->rq_repbuf_len;
 595                 /* Allow multiple early replies */
 596                 reply_md.threshold = LNET_MD_THRESH_INF;
 597                 /* Manage remote for early replies */
 598                 reply_md.options   = PTLRPC_MD_OPTIONS | LNET_MD_OP_PUT |
 599                         LNET_MD_MANAGE_REMOTE |
 600                         LNET_MD_TRUNCATE; /* allow to make EOVERFLOW error */;
 601                 reply_md.user_ptr  = &request->rq_reply_cbid;
 602                 reply_md.eq_handle = ptlrpc_eq_h;
 603
 604                 /* We must see the unlink callback to unset rq_must_unlink,
 605                    so we can't auto-unlink */
 606                 rc = LNetMDAttach(reply_me_h, reply_md, LNET_RETAIN,
 607                                   &request->rq_reply_md_h);
 608                 if (rc != 0) {
 609                         CERROR("LNetMDAttach failed: %d\n", rc);
 610                         LASSERT (rc == -ENOMEM);
 611                         cfs_spin_lock(&request->rq_lock);
 612                         /* ...but the MD attach didn't succeed... */
 613                         request->rq_receiving_reply = 0;
 614                         cfs_spin_unlock(&request->rq_lock);
 615                         GOTO(cleanup_me, rc = -ENOMEM);
 616                 }
 617
 618                 CDEBUG(D_NET, "Setup reply buffer: %u bytes, xid "LPU64
 619                        ", portal %u\n",
 620                        request->rq_repbuf_len, request->rq_xid,
 621                        request->rq_reply_portal);
 622         }
 623
 624         /* add references on request for request_out_callback */
 625         ptlrpc_request_addref(request);
 626         if (obd->obd_svc_stats != NULL)
 627                 lprocfs_counter_add(obd->obd_svc_stats, PTLRPC_REQACTIVE_CNTR,
 628                         cfs_atomic_read(&request->rq_import->imp_inflight));
 629
 630         OBD_FAIL_TIMEOUT(OBD_FAIL_PTLRPC_DELAY_SEND, request->rq_timeout + 5);
 631
 632         cfs_gettimeofday(&request->rq_arrival_time);
 633         request->rq_sent = cfs_time_current_sec();
 634         /* We give the server rq_timeout secs to process the req, and
 635            add the network latency for our local timeout. */
 636         request->rq_deadline = request->rq_sent + request->rq_timeout +
 637                 ptlrpc_at_get_net_latency(request);
 638
 639         ptlrpc_pinger_sending_on_import(request->rq_import);
 640
 641         DEBUG_REQ(D_INFO, request, "send flg=%x",
 642                   lustre_msg_get_flags(request->rq_reqmsg));
 643         rc = ptl_send_buf(&request->rq_req_md_h,
 644                           request->rq_reqbuf, request->rq_reqdata_len,
 645                           LNET_NOACK_REQ, &request->rq_req_cbid,
 646                           connection,
 647                           request->rq_request_portal,
 648                           request->rq_xid, 0);
 649         if (rc == 0)
 650                 RETURN(rc);
 651
 652         ptlrpc_req_finished(request);
 653         if (noreply)
 654                 RETURN(rc);
 655
 656  cleanup_me:
 657         /* MEUnlink is safe; the PUT didn't even get off the ground, and
 658          * nobody apart from the PUT's target has the right nid+XID to
 659          * access the reply buffer. */
 660         rc2 = LNetMEUnlink(reply_me_h);
 661         LASSERT (rc2 == 0);
 662         /* UNLINKED callback called synchronously */
 663         LASSERT(!request->rq_receiving_reply);
 664
 665  cleanup_bulk:
 666         /* We do sync unlink here as there was no real transfer here so
 667          * the chance to have long unlink to sluggish net is smaller here. */
 668         ptlrpc_unregister_bulk(request, 0);
 669         return rc;
 670 }
 671
 672 int ptlrpc_register_rqbd(struct ptlrpc_request_buffer_desc *rqbd)
 673 {
 674         struct ptlrpc_service   *service = rqbd->rqbd_service;
 675         static lnet_process_id_t  match_id = {LNET_NID_ANY, LNET_PID_ANY};
 676         int                      rc;
 677         lnet_md_t                 md;
 678         lnet_handle_me_t          me_h;
 679
 680         CDEBUG(D_NET, "LNetMEAttach: portal %d\n",
 681                service->srv_req_portal);
 682
 683         if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_RQBD))
 684                 return (-ENOMEM);
 685
 686         rc = LNetMEAttach(service->srv_req_portal,
 687                           match_id, 0, ~0, LNET_UNLINK, LNET_INS_AFTER, &me_h);
 688         if (rc != 0) {
 689                 CERROR("LNetMEAttach failed: %d\n", rc);
 690                 return (-ENOMEM);
 691         }
 692
 693         LASSERT(rqbd->rqbd_refcount == 0);
 694         rqbd->rqbd_refcount = 1;
 695
 696         md.start     = rqbd->rqbd_buffer;
 697         md.length    = service->srv_buf_size;
 698         md.max_size  = service->srv_max_req_size;
 699         md.threshold = LNET_MD_THRESH_INF;
 700         md.options   = PTLRPC_MD_OPTIONS | LNET_MD_OP_PUT | LNET_MD_MAX_SIZE;
 701         md.user_ptr  = &rqbd->rqbd_cbid;
 702         md.eq_handle = ptlrpc_eq_h;
 703
 704         rc = LNetMDAttach(me_h, md, LNET_UNLINK, &rqbd->rqbd_md_h);
 705         if (rc == 0)
 706                 return (0);
 707
 708         CERROR("LNetMDAttach failed: %d; \n", rc);
 709         LASSERT (rc == -ENOMEM);
 710         rc = LNetMEUnlink (me_h);
 711         LASSERT (rc == 0);
 712         rqbd->rqbd_refcount = 0;
 713
 714         return (-ENOMEM);
 715 }