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 /**
  48  * Helper function. Sends \a len bytes from \a base at offset \a offset
  49  * over \a conn connection to portal \a portal.
  50  * Returns 0 on success or error code.
  51  */
  52 static int ptl_send_buf (lnet_handle_md_t *mdh, void *base, int len,
  53                          lnet_ack_req_t ack, struct ptlrpc_cb_id *cbid,
  54                          struct ptlrpc_connection *conn, int portal, __u64 xid,
  55                          unsigned int offset)
  56 {
  57         int              rc;
  58         lnet_md_t         md;
  59         ENTRY;
  60
  61         LASSERT (portal != 0);
  62         LASSERT (conn != NULL);
  63         CDEBUG (D_INFO, "conn=%p id %s\n", conn, libcfs_id2str(conn->c_peer));
  64         md.start     = base;
  65         md.length    = len;
  66         md.threshold = (ack == LNET_ACK_REQ) ? 2 : 1;
  67         md.options   = PTLRPC_MD_OPTIONS;
  68         md.user_ptr  = cbid;
  69         md.eq_handle = ptlrpc_eq_h;
  70
  71         if (unlikely(ack == LNET_ACK_REQ &&
  72                      OBD_FAIL_CHECK_ORSET(OBD_FAIL_PTLRPC_ACK, OBD_FAIL_ONCE))){
  73                 /* don't ask for the ack to simulate failing client */
  74                 ack = LNET_NOACK_REQ;
  75         }
  76
  77         rc = LNetMDBind (md, LNET_UNLINK, mdh);
  78         if (unlikely(rc != 0)) {
  79                 CERROR ("LNetMDBind failed: %d\n", rc);
  80                 LASSERT (rc == -ENOMEM);
  81                 RETURN (-ENOMEM);
  82         }
  83
  84         CDEBUG(D_NET, "Sending %d bytes to portal %d, xid "LPD64", offset %u\n",
  85                len, portal, xid, offset);
  86
  87         rc = LNetPut (conn->c_self, *mdh, ack,
  88                       conn->c_peer, portal, xid, offset, 0);
  89         if (unlikely(rc != 0)) {
  90                 int rc2;
  91                 /* We're going to get an UNLINK event when I unlink below,
  92                  * which will complete just like any other failed send, so
  93                  * I fall through and return success here! */
  94                 CERROR("LNetPut(%s, %d, "LPD64") failed: %d\n",
  95                        libcfs_id2str(conn->c_peer), portal, xid, rc);
  96                 rc2 = LNetMDUnlink(*mdh);
  97                 LASSERTF(rc2 == 0, "rc2 = %d\n", rc2);
  98         }
  99
 100         RETURN (0);
 101 }
 102
 103 /**
 104  * Starts bulk transfer for descriptor \a desc
 105  * Returns 0 on success or error code.
 106  */
 107 int ptlrpc_start_bulk_transfer(struct ptlrpc_bulk_desc *desc)
 108 {
 109         struct ptlrpc_connection *conn = desc->bd_export->exp_connection;
 110         int                       rc;
 111         int                       rc2;
 112         lnet_md_t                 md;
 113         __u64                     xid;
 114         ENTRY;
 115
 116         if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_BULK_PUT_NET))
 117                 RETURN(0);
 118
 119         /* NB no locking required until desc is on the network */
 120         LASSERT (!desc->bd_network_rw);
 121         LASSERT (desc->bd_type == BULK_PUT_SOURCE ||
 122                  desc->bd_type == BULK_GET_SINK);
 123         desc->bd_success = 0;
 124
 125         md.user_ptr = &desc->bd_cbid;
 126         md.eq_handle = ptlrpc_eq_h;
 127         md.threshold = 2; /* SENT and ACK/REPLY */
 128         md.options = PTLRPC_MD_OPTIONS;
 129         ptlrpc_fill_bulk_md(&md, desc);
 130
 131         LASSERT (desc->bd_cbid.cbid_fn == server_bulk_callback);
 132         LASSERT (desc->bd_cbid.cbid_arg == desc);
 133
 134         /* NB total length may be 0 for a read past EOF, so we send a 0
 135          * length bulk, since the client expects a bulk event. */
 136
 137         rc = LNetMDBind(md, LNET_UNLINK, &desc->bd_md_h);
 138         if (rc != 0) {
 139                 CERROR("LNetMDBind failed: %d\n", rc);
 140                 LASSERT (rc == -ENOMEM);
 141                 RETURN(-ENOMEM);
 142         }
 143
 144         /* Client's bulk and reply matchbits are the same */
 145         xid = desc->bd_req->rq_xid;
 146         CDEBUG(D_NET, "Transferring %u pages %u bytes via portal %d "
 147                "id %s xid "LPX64"\n", desc->bd_iov_count,
 148                desc->bd_nob, desc->bd_portal,
 149                libcfs_id2str(conn->c_peer), xid);
 150
 151         /* Network is about to get at the memory */
 152         desc->bd_network_rw = 1;
 153
 154         if (desc->bd_type == BULK_PUT_SOURCE)
 155                 rc = LNetPut (conn->c_self, desc->bd_md_h, LNET_ACK_REQ,
 156                               conn->c_peer, desc->bd_portal, xid, 0, 0);
 157         else
 158                 rc = LNetGet (conn->c_self, desc->bd_md_h,
 159                               conn->c_peer, desc->bd_portal, xid, 0);
 160
 161         if (rc != 0) {
 162                 /* Can't send, so we unlink the MD bound above.  The UNLINK
 163                  * event this creates will signal completion with failure,
 164                  * so we return SUCCESS here! */
 165                 CERROR("Transfer(%s, %d, "LPX64") failed: %d\n",
 166                        libcfs_id2str(conn->c_peer), desc->bd_portal, xid, rc);
 167                 rc2 = LNetMDUnlink(desc->bd_md_h);
 168                 LASSERT (rc2 == 0);
 169         }
 170
 171         RETURN(0);
 172 }
 173
 174 /**
 175  * Server side bulk abort. Idempotent. Not thread-safe (i.e. only
 176  * serialises with completion callback)
 177  */
 178 void ptlrpc_abort_bulk(struct ptlrpc_bulk_desc *desc)
 179 {
 180         struct l_wait_info       lwi;
 181         int                      rc;
 182
 183         LASSERT(!cfs_in_interrupt());           /* might sleep */
 184
 185         if (!ptlrpc_server_bulk_active(desc))   /* completed or */
 186                 return;                         /* never started */
 187
 188         /* We used to poison the pages with 0xab here because we did not want to
 189          * send any meaningful data over the wire for evicted clients (bug 9297)
 190          * However, this is no longer safe now that we use the page cache on the
 191          * OSS (bug 20560) */
 192
 193         /* The unlink ensures the callback happens ASAP and is the last
 194          * one.  If it fails, it must be because completion just happened,
 195          * but we must still l_wait_event() in this case, to give liblustre
 196          * a chance to run server_bulk_callback()*/
 197
 198         LNetMDUnlink(desc->bd_md_h);
 199
 200         for (;;) {
 201                 /* Network access will complete in finite time but the HUGE
 202                  * timeout lets us CWARN for visibility of sluggish NALs */
 203                 lwi = LWI_TIMEOUT_INTERVAL(cfs_time_seconds(LONG_UNLINK),
 204                                            cfs_time_seconds(1), NULL, NULL);
 205                 rc = l_wait_event(desc->bd_waitq,
 206                                   !ptlrpc_server_bulk_active(desc), &lwi);
 207                 if (rc == 0)
 208                         return;
 209
 210                 LASSERT(rc == -ETIMEDOUT);
 211                 CWARN("Unexpectedly long timeout: desc %p\n", desc);
 212         }
 213 }
 214
 215 /**
 216  * Register bulk for later transfer
 217  * Returns 0 on success or error code.
 218  */
 219 int ptlrpc_register_bulk(struct ptlrpc_request *req)
 220 {
 221         struct ptlrpc_bulk_desc *desc = req->rq_bulk;
 222         lnet_process_id_t peer;
 223         int rc;
 224         int rc2;
 225         lnet_handle_me_t  me_h;
 226         lnet_md_t         md;
 227         ENTRY;
 228
 229         if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_BULK_GET_NET))
 230                 RETURN(0);
 231
 232         /* NB no locking required until desc is on the network */
 233         LASSERT (desc->bd_nob > 0);
 234         LASSERT (!desc->bd_network_rw);
 235         LASSERT (desc->bd_iov_count <= PTLRPC_MAX_BRW_PAGES);
 236         LASSERT (desc->bd_req != NULL);
 237         LASSERT (desc->bd_type == BULK_PUT_SINK ||
 238                  desc->bd_type == BULK_GET_SOURCE);
 239
 240         desc->bd_success = 0;
 241
 242         peer = desc->bd_import->imp_connection->c_peer;
 243
 244         md.user_ptr = &desc->bd_cbid;
 245         md.eq_handle = ptlrpc_eq_h;
 246         md.threshold = 1;                       /* PUT or GET */
 247         md.options = PTLRPC_MD_OPTIONS |
 248                      ((desc->bd_type == BULK_GET_SOURCE) ?
 249                       LNET_MD_OP_GET : LNET_MD_OP_PUT);
 250         ptlrpc_fill_bulk_md(&md, desc);
 251
 252         LASSERT (desc->bd_cbid.cbid_fn == client_bulk_callback);
 253         LASSERT (desc->bd_cbid.cbid_arg == desc);
 254
 255         /* XXX Registering the same xid on retried bulk makes my head
 256          * explode trying to understand how the original request's bulk
 257          * might interfere with the retried request -eeb
 258          * On the other hand replaying with the same xid is fine, since
 259          * we are guaranteed old request have completed. -green */
 260         LASSERTF(!(desc->bd_registered &&
 261                  req->rq_send_state != LUSTRE_IMP_REPLAY) ||
 262                  req->rq_xid != desc->bd_last_xid,
 263                  "registered: %d  rq_xid: "LPU64" bd_last_xid: "LPU64"\n",
 264                  desc->bd_registered, req->rq_xid, desc->bd_last_xid);
 265         desc->bd_registered = 1;
 266         desc->bd_last_xid = req->rq_xid;
 267
 268         rc = LNetMEAttach(desc->bd_portal, peer,
 269                          req->rq_xid, 0, LNET_UNLINK, LNET_INS_AFTER, &me_h);
 270         if (rc != 0) {
 271                 CERROR("LNetMEAttach failed: %d\n", rc);
 272                 LASSERT (rc == -ENOMEM);
 273                 RETURN (-ENOMEM);
 274         }
 275
 276         /* About to let the network at it... */
 277         desc->bd_network_rw = 1;
 278         rc = LNetMDAttach(me_h, md, LNET_UNLINK, &desc->bd_md_h);
 279         if (rc != 0) {
 280                 CERROR("LNetMDAttach failed: %d\n", rc);
 281                 LASSERT (rc == -ENOMEM);
 282                 desc->bd_network_rw = 0;
 283                 rc2 = LNetMEUnlink (me_h);
 284                 LASSERT (rc2 == 0);
 285                 RETURN (-ENOMEM);
 286         }
 287
 288         CDEBUG(D_NET, "Setup bulk %s buffers: %u pages %u bytes, xid "LPU64", "
 289                "portal %u\n",
 290                desc->bd_type == BULK_GET_SOURCE ? "get-source" : "put-sink",
 291                desc->bd_iov_count, desc->bd_nob,
 292                req->rq_xid, desc->bd_portal);
 293         RETURN(0);
 294 }
 295
 296 /**
 297  * Disconnect a bulk desc from the network. Idempotent. Not
 298  * thread-safe (i.e. only interlocks with completion callback).
 299  * Returns 1 on success or 0 if network unregistration failed for whatever
 300  * reason.
 301  */
 302 int ptlrpc_unregister_bulk(struct ptlrpc_request *req, int async)
 303 {
 304         struct ptlrpc_bulk_desc *desc = req->rq_bulk;
 305         cfs_waitq_t             *wq;
 306         struct l_wait_info       lwi;
 307         int                      rc;
 308         ENTRY;
 309
 310         LASSERT(!cfs_in_interrupt());     /* might sleep */
 311
 312         /* Let's setup deadline for reply unlink. */
 313         if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_BULK_UNLINK) &&
 314             async && req->rq_bulk_deadline == 0)
 315                 req->rq_bulk_deadline = cfs_time_current_sec() + LONG_UNLINK;
 316
 317         if (!ptlrpc_client_bulk_active(req))  /* completed or */
 318                 RETURN(1);                    /* never registered */
 319
 320         LASSERT(desc->bd_req == req);  /* bd_req NULL until registered */
 321
 322         /* the unlink ensures the callback happens ASAP and is the last
 323          * one.  If it fails, it must be because completion just happened,
 324          * but we must still l_wait_event() in this case to give liblustre
 325          * a chance to run client_bulk_callback() */
 326
 327         LNetMDUnlink(desc->bd_md_h);
 328
 329         if (!ptlrpc_client_bulk_active(req))  /* completed or */
 330                 RETURN(1);                    /* never registered */
 331
 332         /* Move to "Unregistering" phase as bulk was not unlinked yet. */
 333         ptlrpc_rqphase_move(req, RQ_PHASE_UNREGISTERING);
 334
 335         /* Do not wait for unlink to finish. */
 336         if (async)
 337                 RETURN(0);
 338
 339         if (req->rq_set != NULL)
 340                 wq = &req->rq_set->set_waitq;
 341         else
 342                 wq = &req->rq_reply_waitq;
 343
 344         for (;;) {
 345                 /* Network access will complete in finite time but the HUGE
 346                  * timeout lets us CWARN for visibility of sluggish NALs */
 347                 lwi = LWI_TIMEOUT_INTERVAL(cfs_time_seconds(LONG_UNLINK),
 348                                            cfs_time_seconds(1), NULL, NULL);
 349                 rc = l_wait_event(*wq, !ptlrpc_client_bulk_active(req), &lwi);
 350                 if (rc == 0) {
 351                         ptlrpc_rqphase_move(req, req->rq_next_phase);
 352                         RETURN(1);
 353                 }
 354
 355                 LASSERT(rc == -ETIMEDOUT);
 356                 DEBUG_REQ(D_WARNING, req, "Unexpectedly long timeout: desc %p",
 357                           desc);
 358         }
 359         RETURN(0);
 360 }
 361
 362 static void ptlrpc_at_set_reply(struct ptlrpc_request *req, int flags)
 363 {
 364         struct ptlrpc_service *svc = req->rq_rqbd->rqbd_service;
 365         int service_time = max_t(int, cfs_time_current_sec() -
 366                                  req->rq_arrival_time.tv_sec, 1);
 367
 368         if (!(flags & PTLRPC_REPLY_EARLY) &&
 369             (req->rq_type != PTL_RPC_MSG_ERR) &&
 370             (req->rq_reqmsg != NULL) &&
 371             !(lustre_msg_get_flags(req->rq_reqmsg) &
 372               (MSG_RESENT | MSG_REPLAY |
 373                MSG_REQ_REPLAY_DONE | MSG_LOCK_REPLAY_DONE))) {
 374                 /* early replies, errors and recovery requests don't count
 375                  * toward our service time estimate */
 376                 int oldse = at_measured(&svc->srv_at_estimate, service_time);
 377                 if (oldse != 0)
 378                         DEBUG_REQ(D_ADAPTTO, req,
 379                                   "svc %s changed estimate from %d to %d",
 380                                   svc->srv_name, oldse,
 381                                   at_get(&svc->srv_at_estimate));
 382         }
 383         /* Report actual service time for client latency calc */
 384         lustre_msg_set_service_time(req->rq_repmsg, service_time);
 385         /* Report service time estimate for future client reqs, but report 0
 386          * (to be ignored by client) if it's a error reply during recovery.
 387          * (bz15815) */
 388         if (req->rq_type == PTL_RPC_MSG_ERR &&
 389             (req->rq_export == NULL || req->rq_export->exp_obd->obd_recovering))
 390                 lustre_msg_set_timeout(req->rq_repmsg, 0);
 391         else
 392                 lustre_msg_set_timeout(req->rq_repmsg,
 393                                        at_get(&svc->srv_at_estimate));
 394
 395         if (req->rq_reqmsg &&
 396             !(lustre_msghdr_get_flags(req->rq_reqmsg) & MSGHDR_AT_SUPPORT)) {
 397                 CDEBUG(D_ADAPTTO, "No early reply support: flags=%#x "
 398                        "req_flags=%#x magic=%d:%x/%x len=%d\n",
 399                        flags, lustre_msg_get_flags(req->rq_reqmsg),
 400                        lustre_msg_is_v1(req->rq_reqmsg),
 401                        lustre_msg_get_magic(req->rq_reqmsg),
 402                        lustre_msg_get_magic(req->rq_repmsg), req->rq_replen);
 403         }
 404 }
 405
 406 /**
 407  * Send request reply from request \a req reply buffer.
 408  * \a flags defines reply types
 409  * Returns 0 on sucess or error code
 410  */
 411 int ptlrpc_send_reply(struct ptlrpc_request *req, int flags)
 412 {
 413         struct ptlrpc_service     *svc = req->rq_rqbd->rqbd_service;
 414         struct ptlrpc_reply_state *rs = req->rq_reply_state;
 415         struct ptlrpc_connection  *conn;
 416         int                        rc;
 417
 418         /* We must already have a reply buffer (only ptlrpc_error() may be
 419          * called without one). The reply generated by sptlrpc layer (e.g.
 420          * error notify, etc.) might have NULL rq->reqmsg; Otherwise we must
 421          * have a request buffer which is either the actual (swabbed) incoming
 422          * request, or a saved copy if this is a req saved in
 423          * target_queue_final_reply().
 424          */
 425         LASSERT (req->rq_no_reply == 0);
 426         LASSERT (req->rq_reqbuf != NULL);
 427         LASSERT (rs != NULL);
 428         LASSERT ((flags & PTLRPC_REPLY_MAYBE_DIFFICULT) || !rs->rs_difficult);
 429         LASSERT (req->rq_repmsg != NULL);
 430         LASSERT (req->rq_repmsg == rs->rs_msg);
 431         LASSERT (rs->rs_cb_id.cbid_fn == reply_out_callback);
 432         LASSERT (rs->rs_cb_id.cbid_arg == rs);
 433
 434         /* There may be no rq_export during failover */
 435
 436         if (unlikely(req->rq_export && req->rq_export->exp_obd &&
 437                      req->rq_export->exp_obd->obd_fail)) {
 438                 /* Failed obd's only send ENODEV */
 439                 req->rq_type = PTL_RPC_MSG_ERR;
 440                 req->rq_status = -ENODEV;
 441                 CDEBUG(D_HA, "sending ENODEV from failed obd %d\n",
 442                        req->rq_export->exp_obd->obd_minor);
 443         }
 444
 445         if (req->rq_type != PTL_RPC_MSG_ERR)
 446                 req->rq_type = PTL_RPC_MSG_REPLY;
 447
 448         lustre_msg_set_type(req->rq_repmsg, req->rq_type);
 449         lustre_msg_set_status(req->rq_repmsg, req->rq_status);
 450         lustre_msg_set_opc(req->rq_repmsg,
 451                 req->rq_reqmsg ? lustre_msg_get_opc(req->rq_reqmsg) : 0);
 452
 453         target_pack_pool_reply(req);
 454
 455         ptlrpc_at_set_reply(req, flags);
 456
 457         if (req->rq_export == NULL || req->rq_export->exp_connection == NULL)
 458                 conn = ptlrpc_connection_get(req->rq_peer, req->rq_self, NULL);
 459         else
 460                 conn = ptlrpc_connection_addref(req->rq_export->exp_connection);
 461
 462         if (unlikely(conn == NULL)) {
 463                 CERROR("not replying on NULL connection\n"); /* bug 9635 */
 464                 return -ENOTCONN;
 465         }
 466         ptlrpc_rs_addref(rs);                   /* +1 ref for the network */
 467
 468         rc = sptlrpc_svc_wrap_reply(req);
 469         if (unlikely(rc))
 470                 goto out;
 471
 472         req->rq_sent = cfs_time_current_sec();
 473
 474         rc = ptl_send_buf (&rs->rs_md_h, rs->rs_repbuf, rs->rs_repdata_len,
 475                            (rs->rs_difficult && !rs->rs_no_ack) ?
 476                            LNET_ACK_REQ : LNET_NOACK_REQ,
 477                            &rs->rs_cb_id, conn, svc->srv_rep_portal,
 478                            req->rq_xid, req->rq_reply_off);
 479 out:
 480         if (unlikely(rc != 0))
 481                 ptlrpc_req_drop_rs(req);
 482         ptlrpc_connection_put(conn);
 483         return rc;
 484 }
 485
 486 int ptlrpc_reply (struct ptlrpc_request *req)
 487 {
 488         if (req->rq_no_reply)
 489                 return 0;
 490         else
 491                 return (ptlrpc_send_reply(req, 0));
 492 }
 493
 494 /**
 495  * For request \a req send an error reply back. Create empty
 496  * reply buffers if necessary.
 497  */
 498 int ptlrpc_send_error(struct ptlrpc_request *req, int may_be_difficult)
 499 {
 500         int rc;
 501         ENTRY;
 502
 503         if (req->rq_no_reply)
 504                 RETURN(0);
 505
 506         if (!req->rq_repmsg) {
 507                 rc = lustre_pack_reply(req, 1, NULL, NULL);
 508                 if (rc)
 509                         RETURN(rc);
 510         }
 511
 512         req->rq_type = PTL_RPC_MSG_ERR;
 513
 514         rc = ptlrpc_send_reply(req, may_be_difficult);
 515         RETURN(rc);
 516 }
 517
 518 int ptlrpc_error(struct ptlrpc_request *req)
 519 {
 520         return ptlrpc_send_error(req, 0);
 521 }
 522
 523 /**
 524  * Send request \a request.
 525  * if \a noreply is set, don't expect any reply back and don't set up
 526  * reply buffers.
 527  * Returns 0 on success or error code.
 528  */
 529 int ptl_send_rpc(struct ptlrpc_request *request, int noreply)
 530 {
 531         int rc;
 532         int rc2;
 533         int mpflag = 0;
 534         struct ptlrpc_connection *connection;
 535         lnet_handle_me_t  reply_me_h;
 536         lnet_md_t         reply_md;
 537         struct obd_device *obd = request->rq_import->imp_obd;
 538         ENTRY;
 539
 540         if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_DROP_RPC))
 541                 RETURN(0);
 542
 543         LASSERT(request->rq_type == PTL_RPC_MSG_REQUEST);
 544         LASSERT(request->rq_wait_ctx == 0);
 545
 546         /* If this is a re-transmit, we're required to have disengaged
 547          * cleanly from the previous attempt */
 548         LASSERT(!request->rq_receiving_reply);
 549
 550         if (request->rq_import->imp_obd &&
 551             request->rq_import->imp_obd->obd_fail) {
 552                 CDEBUG(D_HA, "muting rpc for failed imp obd %s\n",
 553                        request->rq_import->imp_obd->obd_name);
 554                 /* this prevents us from waiting in ptlrpc_queue_wait */
 555                 request->rq_err = 1;
 556                 request->rq_status = -ENODEV;
 557                 RETURN(-ENODEV);
 558         }
 559
 560         connection = request->rq_import->imp_connection;
 561
 562         lustre_msg_set_handle(request->rq_reqmsg,
 563                               &request->rq_import->imp_remote_handle);
 564         lustre_msg_set_type(request->rq_reqmsg, PTL_RPC_MSG_REQUEST);
 565         lustre_msg_set_conn_cnt(request->rq_reqmsg,
 566                                 request->rq_import->imp_conn_cnt);
 567         lustre_msghdr_set_flags(request->rq_reqmsg,
 568                                 request->rq_import->imp_msghdr_flags);
 569
 570         if (request->rq_resend)
 571                 lustre_msg_add_flags(request->rq_reqmsg, MSG_RESENT);
 572
 573         if (request->rq_memalloc)
 574                 mpflag = cfs_memory_pressure_get_and_set();
 575
 576         rc = sptlrpc_cli_wrap_request(request);
 577         if (rc)
 578                 GOTO(out, rc);
 579
 580         /* bulk register should be done after wrap_request() */
 581         if (request->rq_bulk != NULL) {
 582                 rc = ptlrpc_register_bulk (request);
 583                 if (rc != 0)
 584                         GOTO(out, rc);
 585         }
 586
 587         if (!noreply) {
 588                 LASSERT (request->rq_replen != 0);
 589                 if (request->rq_repbuf == NULL) {
 590                         LASSERT(request->rq_repdata == NULL);
 591                         LASSERT(request->rq_repmsg == NULL);
 592                         rc = sptlrpc_cli_alloc_repbuf(request,
 593                                                       request->rq_replen);
 594                         if (rc) {
 595                                 /* this prevents us from looping in
 596                                  * ptlrpc_queue_wait */
 597                                 request->rq_err = 1;
 598                                 request->rq_status = rc;
 599                                 GOTO(cleanup_bulk, rc);
 600                         }
 601                 } else {
 602                         request->rq_repdata = NULL;
 603                         request->rq_repmsg = NULL;
 604                 }
 605
 606                 rc = LNetMEAttach(request->rq_reply_portal,/*XXX FIXME bug 249*/
 607                                   connection->c_peer, request->rq_xid, 0,
 608                                   LNET_UNLINK, LNET_INS_AFTER, &reply_me_h);
 609                 if (rc != 0) {
 610                         CERROR("LNetMEAttach failed: %d\n", rc);
 611                         LASSERT (rc == -ENOMEM);
 612                         GOTO(cleanup_bulk, rc = -ENOMEM);
 613                 }
 614         }
 615
 616         cfs_spin_lock(&request->rq_lock);
 617         /* If the MD attach succeeds, there _will_ be a reply_in callback */
 618         request->rq_receiving_reply = !noreply;
 619         /* We are responsible for unlinking the reply buffer */
 620         request->rq_must_unlink = !noreply;
 621         /* Clear any flags that may be present from previous sends. */
 622         request->rq_replied = 0;
 623         request->rq_err = 0;
 624         request->rq_timedout = 0;
 625         request->rq_net_err = 0;
 626         request->rq_resend = 0;
 627         request->rq_restart = 0;
 628         request->rq_reply_truncate = 0;
 629         cfs_spin_unlock(&request->rq_lock);
 630
 631         if (!noreply) {
 632                 reply_md.start     = request->rq_repbuf;
 633                 reply_md.length    = request->rq_repbuf_len;
 634                 /* Allow multiple early replies */
 635                 reply_md.threshold = LNET_MD_THRESH_INF;
 636                 /* Manage remote for early replies */
 637                 reply_md.options   = PTLRPC_MD_OPTIONS | LNET_MD_OP_PUT |
 638                         LNET_MD_MANAGE_REMOTE |
 639                         LNET_MD_TRUNCATE; /* allow to make EOVERFLOW error */;
 640                 reply_md.user_ptr  = &request->rq_reply_cbid;
 641                 reply_md.eq_handle = ptlrpc_eq_h;
 642
 643                 /* We must see the unlink callback to unset rq_must_unlink,
 644                    so we can't auto-unlink */
 645                 rc = LNetMDAttach(reply_me_h, reply_md, LNET_RETAIN,
 646                                   &request->rq_reply_md_h);
 647                 if (rc != 0) {
 648                         CERROR("LNetMDAttach failed: %d\n", rc);
 649                         LASSERT (rc == -ENOMEM);
 650                         cfs_spin_lock(&request->rq_lock);
 651                         /* ...but the MD attach didn't succeed... */
 652                         request->rq_receiving_reply = 0;
 653                         cfs_spin_unlock(&request->rq_lock);
 654                         GOTO(cleanup_me, rc = -ENOMEM);
 655                 }
 656
 657                 CDEBUG(D_NET, "Setup reply buffer: %u bytes, xid "LPU64
 658                        ", portal %u\n",
 659                        request->rq_repbuf_len, request->rq_xid,
 660                        request->rq_reply_portal);
 661         }
 662
 663         /* add references on request for request_out_callback */
 664         ptlrpc_request_addref(request);
 665         if (obd->obd_svc_stats != NULL)
 666                 lprocfs_counter_add(obd->obd_svc_stats, PTLRPC_REQACTIVE_CNTR,
 667                         cfs_atomic_read(&request->rq_import->imp_inflight));
 668
 669         OBD_FAIL_TIMEOUT(OBD_FAIL_PTLRPC_DELAY_SEND, request->rq_timeout + 5);
 670
 671         cfs_gettimeofday(&request->rq_arrival_time);
 672         request->rq_sent = cfs_time_current_sec();
 673         /* We give the server rq_timeout secs to process the req, and
 674            add the network latency for our local timeout. */
 675         request->rq_deadline = request->rq_sent + request->rq_timeout +
 676                 ptlrpc_at_get_net_latency(request);
 677
 678         ptlrpc_pinger_sending_on_import(request->rq_import);
 679
 680         DEBUG_REQ(D_INFO, request, "send flg=%x",
 681                   lustre_msg_get_flags(request->rq_reqmsg));
 682         rc = ptl_send_buf(&request->rq_req_md_h,
 683                           request->rq_reqbuf, request->rq_reqdata_len,
 684                           LNET_NOACK_REQ, &request->rq_req_cbid,
 685                           connection,
 686                           request->rq_request_portal,
 687                           request->rq_xid, 0);
 688         if (rc == 0)
 689                 GOTO(out, rc);
 690
 691         ptlrpc_req_finished(request);
 692         if (noreply)
 693                 GOTO(out, rc);
 694
 695  cleanup_me:
 696         /* MEUnlink is safe; the PUT didn't even get off the ground, and
 697          * nobody apart from the PUT's target has the right nid+XID to
 698          * access the reply buffer. */
 699         rc2 = LNetMEUnlink(reply_me_h);
 700         LASSERT (rc2 == 0);
 701         /* UNLINKED callback called synchronously */
 702         LASSERT(!request->rq_receiving_reply);
 703
 704  cleanup_bulk:
 705         /* We do sync unlink here as there was no real transfer here so
 706          * the chance to have long unlink to sluggish net is smaller here. */
 707         ptlrpc_unregister_bulk(request, 0);
 708  out:
 709         if (request->rq_memalloc)
 710                 cfs_memory_pressure_restore(mpflag);
 711         return rc;
 712 }
 713
 714 /**
 715  * Register request buffer descriptor for request receiving.
 716  */
 717 int ptlrpc_register_rqbd(struct ptlrpc_request_buffer_desc *rqbd)
 718 {
 719         struct ptlrpc_service   *service = rqbd->rqbd_service;
 720         static lnet_process_id_t  match_id = {LNET_NID_ANY, LNET_PID_ANY};
 721         int                      rc;
 722         lnet_md_t                 md;
 723         lnet_handle_me_t          me_h;
 724
 725         CDEBUG(D_NET, "LNetMEAttach: portal %d\n",
 726                service->srv_req_portal);
 727
 728         if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_RQBD))
 729                 return (-ENOMEM);
 730
 731         rc = LNetMEAttach(service->srv_req_portal,
 732                           match_id, 0, ~0, LNET_UNLINK, LNET_INS_AFTER, &me_h);
 733         if (rc != 0) {
 734                 CERROR("LNetMEAttach failed: %d\n", rc);
 735                 return (-ENOMEM);
 736         }
 737
 738         LASSERT(rqbd->rqbd_refcount == 0);
 739         rqbd->rqbd_refcount = 1;
 740
 741         md.start     = rqbd->rqbd_buffer;
 742         md.length    = service->srv_buf_size;
 743         md.max_size  = service->srv_max_req_size;
 744         md.threshold = LNET_MD_THRESH_INF;
 745         md.options   = PTLRPC_MD_OPTIONS | LNET_MD_OP_PUT | LNET_MD_MAX_SIZE;
 746         md.user_ptr  = &rqbd->rqbd_cbid;
 747         md.eq_handle = ptlrpc_eq_h;
 748
 749         rc = LNetMDAttach(me_h, md, LNET_UNLINK, &rqbd->rqbd_md_h);
 750         if (rc == 0)
 751                 return (0);
 752
 753         CERROR("LNetMDAttach failed: %d; \n", rc);
 754         LASSERT (rc == -ENOMEM);
 755         rc = LNetMEUnlink (me_h);
 756         LASSERT (rc == 0);
 757         rqbd->rqbd_refcount = 0;
 758
 759         return (-ENOMEM);
 760 }