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 (c) 2002, 2010, Oracle and/or its affiliates. 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         /* We used to poison the pages with 0xab here because we did not want to
 180          * send any meaningful data over the wire for evicted clients (bug 9297)
 181          * However, this is no longer safe now that we use the page cache on the
 182          * OSS (bug 20560) */
 183
 184         /* The unlink ensures the callback happens ASAP and is the last
 185          * one.  If it fails, it must be because completion just happened,
 186          * but we must still l_wait_event() in this case, to give liblustre
 187          * a chance to run server_bulk_callback()*/
 188
 189         LNetMDUnlink(desc->bd_md_h);
 190
 191         for (;;) {
 192                 /* Network access will complete in finite time but the HUGE
 193                  * timeout lets us CWARN for visibility of sluggish NALs */
 194                 lwi = LWI_TIMEOUT_INTERVAL(cfs_time_seconds(LONG_UNLINK),
 195                                            cfs_time_seconds(1), NULL, NULL);
 196                 rc = l_wait_event(desc->bd_waitq,
 197                                   !ptlrpc_server_bulk_active(desc), &lwi);
 198                 if (rc == 0)
 199                         return;
 200
 201                 LASSERT(rc == -ETIMEDOUT);
 202                 CWARN("Unexpectedly long timeout: desc %p\n", desc);
 203         }
 204 }
 205
 206 int ptlrpc_register_bulk(struct ptlrpc_request *req)
 207 {
 208         struct ptlrpc_bulk_desc *desc = req->rq_bulk;
 209         lnet_process_id_t peer;
 210         int rc;
 211         int rc2;
 212         lnet_handle_me_t  me_h;
 213         lnet_md_t         md;
 214         ENTRY;
 215
 216         if (OBD_FAIL_CHECK_ONCE(OBD_FAIL_PTLRPC_BULK_GET_NET))
 217                 RETURN(0);
 218
 219         /* NB no locking required until desc is on the network */
 220         LASSERT (desc->bd_nob > 0);
 221         LASSERT (!desc->bd_network_rw);
 222         LASSERT (desc->bd_iov_count <= PTLRPC_MAX_BRW_PAGES);
 223         LASSERT (desc->bd_req != NULL);
 224         LASSERT (desc->bd_type == BULK_PUT_SINK ||
 225                  desc->bd_type == BULK_GET_SOURCE);
 226
 227         desc->bd_success = 0;
 228         desc->bd_sender = LNET_NID_ANY;
 229
 230         peer = desc->bd_import->imp_connection->c_peer;
 231
 232         md.user_ptr = &desc->bd_cbid;
 233         md.eq_handle = ptlrpc_eq_h;
 234         md.threshold = 1;                       /* PUT or GET */
 235         md.options = PTLRPC_MD_OPTIONS |
 236                      ((desc->bd_type == BULK_GET_SOURCE) ?
 237                       LNET_MD_OP_GET : LNET_MD_OP_PUT);
 238         ptlrpc_fill_bulk_md(&md, desc);
 239
 240         LASSERT (desc->bd_cbid.cbid_fn == client_bulk_callback);
 241         LASSERT (desc->bd_cbid.cbid_arg == desc);
 242
 243         /* XXX Registering the same xid on retried bulk makes my head
 244          * explode trying to understand how the original request's bulk
 245          * might interfere with the retried request -eeb
 246          * On the other hand replaying with the same xid is fine, since
 247          * we are guaranteed old request have completed. -green */
 248         LASSERTF(!(desc->bd_registered &&
 249                  req->rq_send_state != LUSTRE_IMP_REPLAY) ||
 250                  req->rq_xid != desc->bd_last_xid,
 251                  "registered: %d  rq_xid: "LPU64" bd_last_xid: "LPU64"\n",
 252                  desc->bd_registered, req->rq_xid, desc->bd_last_xid);
 253         desc->bd_registered = 1;
 254         desc->bd_last_xid = req->rq_xid;
 255
 256         rc = LNetMEAttach(desc->bd_portal, peer,
 257                          req->rq_xid, 0, LNET_UNLINK, LNET_INS_AFTER, &me_h);
 258         if (rc != 0) {
 259                 CERROR("LNetMEAttach failed: %d\n", rc);
 260                 LASSERT (rc == -ENOMEM);
 261                 RETURN (-ENOMEM);
 262         }
 263
 264         /* About to let the network at it... */
 265         desc->bd_network_rw = 1;
 266         rc = LNetMDAttach(me_h, md, LNET_UNLINK, &desc->bd_md_h);
 267         if (rc != 0) {
 268                 CERROR("LNetMDAttach failed: %d\n", rc);
 269                 LASSERT (rc == -ENOMEM);
 270                 desc->bd_network_rw = 0;
 271                 rc2 = LNetMEUnlink (me_h);
 272                 LASSERT (rc2 == 0);
 273                 RETURN (-ENOMEM);
 274         }
 275
 276         CDEBUG(D_NET, "Setup bulk %s buffers: %u pages %u bytes, xid "LPU64", "
 277                "portal %u\n",
 278                desc->bd_type == BULK_GET_SOURCE ? "get-source" : "put-sink",
 279                desc->bd_iov_count, desc->bd_nob,
 280                req->rq_xid, desc->bd_portal);
 281         RETURN(0);
 282 }
 283
 284 /* Disconnect a bulk desc from the network. Idempotent. Not
 285  * thread-safe (i.e. only interlocks with completion callback). */
 286 int ptlrpc_unregister_bulk(struct ptlrpc_request *req, int async)
 287 {
 288         struct ptlrpc_bulk_desc *desc = req->rq_bulk;
 289         cfs_waitq_t             *wq;
 290         struct l_wait_info       lwi;
 291         int                      rc;
 292         ENTRY;
 293
 294         LASSERT(!in_interrupt());     /* might sleep */
 295
 296         /* Let's setup deadline for reply unlink. */
 297         if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_BULK_UNLINK) &&
 298             async && req->rq_bulk_deadline == 0)
 299                 req->rq_bulk_deadline = cfs_time_current_sec() + LONG_UNLINK;
 300
 301         if (!ptlrpc_client_bulk_active(req))  /* completed or */
 302                 RETURN(1);                    /* never registered */
 303
 304         LASSERT(desc->bd_req == req);  /* bd_req NULL until registered */
 305
 306         /* the unlink ensures the callback happens ASAP and is the last
 307          * one.  If it fails, it must be because completion just happened,
 308          * but we must still l_wait_event() in this case to give liblustre
 309          * a chance to run client_bulk_callback() */
 310
 311         LNetMDUnlink(desc->bd_md_h);
 312
 313         if (!ptlrpc_client_bulk_active(req))  /* completed or */
 314                 RETURN(1);                    /* never registered */
 315
 316         /* Move to "Unregistering" phase as bulk was not unlinked yet. */
 317         ptlrpc_rqphase_move(req, RQ_PHASE_UNREGISTERING);
 318
 319         /* Do not wait for unlink to finish. */
 320         if (async)
 321                 RETURN(0);
 322
 323         if (req->rq_set != NULL)
 324                 wq = &req->rq_set->set_waitq;
 325         else
 326                 wq = &req->rq_reply_waitq;
 327
 328         for (;;) {
 329                 /* Network access will complete in finite time but the HUGE
 330                  * timeout lets us CWARN for visibility of sluggish NALs */
 331                 lwi = LWI_TIMEOUT_INTERVAL(cfs_time_seconds(LONG_UNLINK),
 332                                            cfs_time_seconds(1), NULL, NULL);
 333                 rc = l_wait_event(*wq, !ptlrpc_client_bulk_active(req), &lwi);
 334                 if (rc == 0) {
 335                         ptlrpc_rqphase_move(req, req->rq_next_phase);
 336                         RETURN(1);
 337                 }
 338
 339                 LASSERT(rc == -ETIMEDOUT);
 340                 DEBUG_REQ(D_WARNING, req, "Unexpectedly long timeout: desc %p",
 341                           desc);
 342         }
 343         RETURN(0);
 344 }
 345
 346 int ptlrpc_send_reply(struct ptlrpc_request *req, int flags)
 347 {
 348         struct ptlrpc_service     *svc = req->rq_rqbd->rqbd_service;
 349         struct ptlrpc_reply_state *rs = req->rq_reply_state;
 350         struct ptlrpc_connection  *conn;
 351         int                        service_time;
 352         unsigned int               offset = 0;
 353         int                        rc;
 354
 355         /* We must already have a reply buffer (only ptlrpc_error() may be
 356          * called without one).  We must also have a request buffer which
 357          * is either the actual (swabbed) incoming request, or a saved copy
 358          * if this is a req saved in target_queue_final_reply(). */
 359         LASSERT (req->rq_reqmsg != NULL);
 360         LASSERT (req->rq_repmsg != NULL);
 361         LASSERT (rs != NULL);
 362         LASSERT (req->rq_repmsg == rs->rs_msg);
 363         LASSERT ((flags & PTLRPC_REPLY_MAYBE_DIFFICULT) || !rs->rs_difficult);
 364         LASSERT (rs->rs_cb_id.cbid_fn == reply_out_callback);
 365         LASSERT (rs->rs_cb_id.cbid_arg == rs);
 366
 367         /* There may be no rq_export during failover */
 368
 369         if (req->rq_export && req->rq_export->exp_obd &&
 370             req->rq_export->exp_obd->obd_fail) {
 371                 /* Failed obd's only send ENODEV */
 372                 req->rq_type = PTL_RPC_MSG_ERR;
 373                 req->rq_status = -ENODEV;
 374                 CDEBUG(D_HA, "sending ENODEV from failed obd %d\n",
 375                        req->rq_export->exp_obd->obd_minor);
 376         }
 377
 378         if (req->rq_type != PTL_RPC_MSG_ERR)
 379                 req->rq_type = PTL_RPC_MSG_REPLY;
 380
 381         lustre_msg_set_type(req->rq_repmsg, req->rq_type);
 382         lustre_msg_set_status(req->rq_repmsg, req->rq_status);
 383         lustre_msg_set_opc(req->rq_repmsg, lustre_msg_get_opc(req->rq_reqmsg));
 384
 385         service_time = max_t(int, cfs_time_current_sec() -
 386                              req->rq_arrival_time.tv_sec, 1);
 387         if (!(flags & PTLRPC_REPLY_EARLY) &&
 388             (req->rq_type != PTL_RPC_MSG_ERR) &&
 389             !(lustre_msg_get_flags(req->rq_reqmsg) &
 390               (MSG_RESENT | MSG_REPLAY | MSG_LAST_REPLAY))) {
 391                 /* early replies, errors and recovery requests don't count
 392                  * toward our service time estimate */
 393                 int oldse = at_measured(&svc->srv_at_estimate, service_time);
 394                 if (oldse != 0)
 395                         DEBUG_REQ(D_ADAPTTO, req,
 396                                   "svc %s changed estimate from %d to %d",
 397                                   svc->srv_name, oldse,
 398                                   at_get(&svc->srv_at_estimate));
 399         }
 400         /* Report actual service time for client latency calc */
 401         lustre_msg_set_service_time(req->rq_repmsg, service_time);
 402         /* Report service time estimate for future client reqs, but report 0
 403          * (to be ignored by client) if it's a error reply during recovery.
 404          * (bz15815) */
 405         if (req->rq_type == PTL_RPC_MSG_ERR &&
 406             (req->rq_export == NULL || req->rq_export->exp_obd->obd_recovering))
 407                 lustre_msg_set_timeout(req->rq_repmsg, 0);
 408         else
 409                 lustre_msg_set_timeout(req->rq_repmsg,
 410                                        at_get(&svc->srv_at_estimate));
 411
 412         target_pack_pool_reply(req);
 413
 414         if (lustre_msghdr_get_flags(req->rq_reqmsg) & MSGHDR_AT_SUPPORT) {
 415                 /* early replies go to offset 0, regular replies go after that*/
 416                 if (flags & PTLRPC_REPLY_EARLY) {
 417                         offset = 0;
 418                         /* Include a checksum on early replies - must be done
 419                            after all other lustre_msg_set_* */
 420                         lustre_msg_set_cksum(req->rq_repmsg,
 421                                          lustre_msg_calc_cksum(req->rq_repmsg));
 422                 } else {
 423                         offset = lustre_msg_early_size(req);
 424                 }
 425         } else {
 426                 CDEBUG(D_ADAPTTO, "No early reply support: flags=%#x "
 427                        "req_flags=%#x magic=%d:%x/%x len=%d\n",
 428                       flags, lustre_msg_get_flags(req->rq_reqmsg),
 429                       lustre_msg_is_v1(req->rq_reqmsg),
 430                       lustre_msg_get_magic(req->rq_reqmsg),
 431                       lustre_msg_get_magic(req->rq_repmsg), req->rq_replen);
 432         }
 433
 434         if (req->rq_export == NULL || req->rq_export->exp_connection == NULL)
 435                 conn = ptlrpc_connection_get(req->rq_peer, req->rq_self, NULL);
 436         else
 437                 conn = ptlrpc_connection_addref(req->rq_export->exp_connection);
 438
 439         if (conn == NULL) {
 440                 CERROR("not replying on NULL connection\n"); /* bug 9635 */
 441                 return -ENOTCONN;
 442         }
 443
 444         atomic_inc (&svc->srv_outstanding_replies);
 445         ptlrpc_rs_addref(rs);                   /* +1 ref for the network */
 446         req->rq_sent = cfs_time_current_sec();
 447
 448         rc = ptl_send_buf (&rs->rs_md_h, req->rq_repmsg, req->rq_replen,
 449                            rs->rs_difficult ? LNET_ACK_REQ : LNET_NOACK_REQ,
 450                            &rs->rs_cb_id, conn, svc->srv_rep_portal,
 451                            req->rq_xid, offset);
 452         if (rc != 0) {
 453                 atomic_dec (&svc->srv_outstanding_replies);
 454                 ptlrpc_req_drop_rs(req);
 455         }
 456         ptlrpc_connection_put(conn);
 457         return rc;
 458 }
 459
 460 int ptlrpc_reply (struct ptlrpc_request *req)
 461 {
 462         return (ptlrpc_send_reply (req, 0));
 463 }
 464
 465 int ptlrpc_send_error(struct ptlrpc_request *req, int may_be_difficult)
 466 {
 467         int rc;
 468         ENTRY;
 469
 470         if (!req->rq_repmsg) {
 471                 rc = lustre_pack_reply(req, 1, NULL, NULL);
 472                 if (rc)
 473                         RETURN(rc);
 474         }
 475
 476         req->rq_type = PTL_RPC_MSG_ERR;
 477
 478         rc = ptlrpc_send_reply(req, may_be_difficult);
 479         RETURN(rc);
 480 }
 481
 482 int ptlrpc_error(struct ptlrpc_request *req)
 483 {
 484         return ptlrpc_send_error(req, 0);
 485 }
 486
 487 int ptl_send_rpc(struct ptlrpc_request *request, int noreply)
 488 {
 489         int rc;
 490         int rc2;
 491         struct ptlrpc_connection *connection;
 492         lnet_handle_me_t  reply_me_h;
 493         lnet_md_t         reply_md;
 494         struct obd_device *obd = request->rq_import->imp_obd;
 495         ENTRY;
 496
 497         OBD_FAIL_RETURN(OBD_FAIL_PTLRPC_DROP_RPC, 0);
 498
 499         LASSERT (request->rq_type == PTL_RPC_MSG_REQUEST);
 500
 501         /* If this is a re-transmit, we're required to have disengaged
 502          * cleanly from the previous attempt */
 503         LASSERT(!request->rq_receiving_reply);
 504
 505         if (request->rq_import->imp_obd &&
 506             request->rq_import->imp_obd->obd_fail) {
 507                 CDEBUG(D_HA, "muting rpc for failed imp obd %s\n",
 508                        request->rq_import->imp_obd->obd_name);
 509                 /* this prevents us from waiting in ptlrpc_queue_wait */
 510                 spin_lock(&request->rq_lock);
 511                 request->rq_err = 1;
 512                 spin_unlock(&request->rq_lock);
 513
 514                 request->rq_status = -ENODEV;
 515                 RETURN(-ENODEV);
 516         }
 517
 518         connection = request->rq_import->imp_connection;
 519
 520         if (request->rq_bulk != NULL) {
 521                 rc = ptlrpc_register_bulk (request);
 522                 if (rc != 0)
 523                         RETURN(rc);
 524         }
 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         if (!noreply) {
 538                 LASSERT (request->rq_replen != 0);
 539                 if (request->rq_repbuf == NULL)
 540                         OBD_ALLOC(request->rq_repbuf, request->rq_replen);
 541                 if (request->rq_repbuf == NULL) {
 542                         /* this prevents us from looping in ptlrpc_queue_wait */
 543                         spin_lock(&request->rq_lock);
 544                         request->rq_err = 1;
 545                         spin_unlock(&request->rq_lock);
 546
 547                         request->rq_status = -ENOMEM;
 548                         GOTO(cleanup_bulk, rc = -ENOMEM);
 549                 }
 550                 request->rq_repmsg = NULL;
 551
 552                 rc = LNetMEAttach(request->rq_reply_portal,/*XXX FIXME bug 249*/
 553                                   connection->c_peer, request->rq_xid, 0,
 554                                   LNET_UNLINK, LNET_INS_AFTER, &reply_me_h);
 555                 if (rc != 0) {
 556                         CERROR("LNetMEAttach failed: %d\n", rc);
 557                         LASSERT (rc == -ENOMEM);
 558                         GOTO(cleanup_repmsg, rc = -ENOMEM);
 559                 }
 560         }
 561
 562         spin_lock(&request->rq_lock);
 563         /* If the MD attach succeeds, there _will_ be a reply_in callback */
 564         request->rq_receiving_reply = !noreply;
 565         /* We are responsible for unlinking the reply buffer */
 566         request->rq_must_unlink = !noreply;
 567         /* Clear any flags that may be present from previous sends. */
 568         request->rq_replied = 0;
 569         request->rq_err = 0;
 570         request->rq_timedout = 0;
 571         request->rq_net_err = 0;
 572         request->rq_resend = 0;
 573         request->rq_restart = 0;
 574         request->rq_rep_swab_mask = 0;
 575         request->rq_reply_truncate = 0;
 576         spin_unlock(&request->rq_lock);
 577
 578         if (!noreply) {
 579                 reply_md.start     = request->rq_repbuf;
 580                 reply_md.length    = request->rq_replen;
 581                 /* Allow multiple early replies */
 582                 reply_md.threshold = LNET_MD_THRESH_INF;
 583                 /* Manage remote for early replies */
 584                 reply_md.options   = PTLRPC_MD_OPTIONS | LNET_MD_OP_PUT |
 585                         LNET_MD_MANAGE_REMOTE |
 586                         LNET_MD_TRUNCATE; /* allow to make EBIG error */
 587                 reply_md.user_ptr  = &request->rq_reply_cbid;
 588                 reply_md.eq_handle = ptlrpc_eq_h;
 589
 590                 /* We must see the unlink callback to unset rq_must_unlink,
 591                    so we can't auto-unlink */
 592                 rc = LNetMDAttach(reply_me_h, reply_md, LNET_RETAIN,
 593                                  &request->rq_reply_md_h);
 594                 if (rc != 0) {
 595                         CERROR("LNetMDAttach failed: %d\n", rc);
 596                         LASSERT (rc == -ENOMEM);
 597                         spin_lock(&request->rq_lock);
 598                         /* ...but the MD attach didn't succeed... */
 599                         request->rq_receiving_reply = 0;
 600                         spin_unlock(&request->rq_lock);
 601                         GOTO(cleanup_me, rc = -ENOMEM);
 602                 }
 603
 604                 CDEBUG(D_NET, "Setup reply buffer: %u bytes, xid "LPU64
 605                        ", portal %u\n",
 606                        request->rq_replen, request->rq_xid,
 607                        request->rq_reply_portal);
 608         }
 609
 610         /* add references on request for request_out_callback */
 611         ptlrpc_request_addref(request);
 612         if (obd->obd_svc_stats != NULL)
 613                 lprocfs_counter_add(obd->obd_svc_stats, PTLRPC_REQACTIVE_CNTR,
 614                                 atomic_read(&request->rq_import->imp_inflight));
 615
 616         OBD_FAIL_TIMEOUT(OBD_FAIL_PTLRPC_DELAY_SEND, request->rq_timeout + 5);
 617
 618         request->rq_sent = cfs_time_current_sec();
 619         do_gettimeofday(&request->rq_arrival_time);
 620         /* We give the server rq_timeout secs to process the req, and
 621            add the network latency for our local timeout. */
 622         request->rq_deadline = request->rq_sent + request->rq_timeout +
 623                     ptlrpc_at_get_net_latency(request);
 624
 625         ptlrpc_pinger_sending_on_import(request->rq_import);
 626
 627         DEBUG_REQ(D_INFO, request, "send flg=%x",
 628                   lustre_msg_get_flags(request->rq_reqmsg));
 629         rc = ptl_send_buf(&request->rq_req_md_h,
 630                           request->rq_reqmsg, request->rq_reqlen,
 631                           LNET_NOACK_REQ, &request->rq_req_cbid,
 632                           connection,
 633                           request->rq_request_portal,
 634                           request->rq_xid, 0);
 635         if (rc == 0)
 636                 RETURN(rc);
 637
 638         ptlrpc_req_finished(request);
 639         if (noreply)
 640                 RETURN(rc);
 641
 642  cleanup_me:
 643         /* MEUnlink is safe; the PUT didn't even get off the ground, and
 644          * nobody apart from the PUT's target has the right nid+XID to
 645          * access the reply buffer. */
 646         rc2 = LNetMEUnlink(reply_me_h);
 647         LASSERT (rc2 == 0);
 648         /* UNLINKED callback called synchronously */
 649         LASSERT(!request->rq_receiving_reply);
 650
 651  cleanup_repmsg:
 652         OBD_FREE(request->rq_repbuf, request->rq_replen);
 653         request->rq_repbuf = NULL;
 654         request->rq_repmsg = NULL; //remove
 655
 656  cleanup_bulk:
 657         /* We do sync unlink here as there was no real transfer here so
 658          * the chance to have long unlink to sluggish net is smaller here. */
 659         ptlrpc_unregister_bulk(request, 0);
 660         return rc;
 661 }
 662
 663 int ptlrpc_register_rqbd(struct ptlrpc_request_buffer_desc *rqbd)
 664 {
 665         struct ptlrpc_service   *service = rqbd->rqbd_service;
 666         static lnet_process_id_t  match_id = {LNET_NID_ANY, LNET_PID_ANY};
 667         int                      rc;
 668         lnet_md_t                 md;
 669         lnet_handle_me_t          me_h;
 670
 671         CDEBUG(D_NET, "LNetMEAttach: portal %d\n",
 672                service->srv_req_portal);
 673
 674         if (OBD_FAIL_CHECK_ONCE(OBD_FAIL_PTLRPC_RQBD))
 675                 return (-ENOMEM);
 676
 677         rc = LNetMEAttach(service->srv_req_portal,
 678                           match_id, 0, ~0, LNET_UNLINK, LNET_INS_AFTER, &me_h);
 679         if (rc != 0) {
 680                 CERROR("LNetMEAttach failed: %d\n", rc);
 681                 return (-ENOMEM);
 682         }
 683
 684         LASSERT(rqbd->rqbd_refcount == 0);
 685         rqbd->rqbd_refcount = 1;
 686
 687         md.start     = rqbd->rqbd_buffer;
 688         md.length    = service->srv_buf_size;
 689         md.max_size  = service->srv_max_req_size;
 690         md.threshold = LNET_MD_THRESH_INF;
 691         md.options   = PTLRPC_MD_OPTIONS | LNET_MD_OP_PUT | LNET_MD_MAX_SIZE;
 692         md.user_ptr  = &rqbd->rqbd_cbid;
 693         md.eq_handle = ptlrpc_eq_h;
 694
 695         rc = LNetMDAttach(me_h, md, LNET_UNLINK, &rqbd->rqbd_md_h);
 696         if (rc == 0)
 697                 return (0);
 698
 699         CERROR("LNetMDAttach failed: %d; \n", rc);
 700         LASSERT (rc == -ENOMEM);
 701         rc = LNetMEUnlink (me_h);
 702         LASSERT (rc == 0);
 703         rqbd->rqbd_refcount = 0;
 704
 705         return (-ENOMEM);
 706 }