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 /**
  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         LASSERTF (!desc->bd_registered || req->rq_xid != desc->bd_last_xid,
 259                   "registered: %d  rq_xid: "LPU64" bd_last_xid: "LPU64"\n",
 260                   desc->bd_registered, req->rq_xid, desc->bd_last_xid);
 261         desc->bd_registered = 1;
 262         desc->bd_last_xid = req->rq_xid;
 263
 264         rc = LNetMEAttach(desc->bd_portal, peer,
 265                          req->rq_xid, 0, LNET_UNLINK, LNET_INS_AFTER, &me_h);
 266         if (rc != 0) {
 267                 CERROR("LNetMEAttach failed: %d\n", rc);
 268                 LASSERT (rc == -ENOMEM);
 269                 RETURN (-ENOMEM);
 270         }
 271
 272         /* About to let the network at it... */
 273         desc->bd_network_rw = 1;
 274         rc = LNetMDAttach(me_h, md, LNET_UNLINK, &desc->bd_md_h);
 275         if (rc != 0) {
 276                 CERROR("LNetMDAttach failed: %d\n", rc);
 277                 LASSERT (rc == -ENOMEM);
 278                 desc->bd_network_rw = 0;
 279                 rc2 = LNetMEUnlink (me_h);
 280                 LASSERT (rc2 == 0);
 281                 RETURN (-ENOMEM);
 282         }
 283
 284         CDEBUG(D_NET, "Setup bulk %s buffers: %u pages %u bytes, xid "LPU64", "
 285                "portal %u\n",
 286                desc->bd_type == BULK_GET_SOURCE ? "get-source" : "put-sink",
 287                desc->bd_iov_count, desc->bd_nob,
 288                req->rq_xid, desc->bd_portal);
 289         RETURN(0);
 290 }
 291
 292 /**
 293  * Disconnect a bulk desc from the network. Idempotent. Not
 294  * thread-safe (i.e. only interlocks with completion callback).
 295  * Returns 1 on success or 0 if network unregistration failed for whatever
 296  * reason.
 297  */
 298 int ptlrpc_unregister_bulk(struct ptlrpc_request *req, int async)
 299 {
 300         struct ptlrpc_bulk_desc *desc = req->rq_bulk;
 301         cfs_waitq_t             *wq;
 302         struct l_wait_info       lwi;
 303         int                      rc;
 304         ENTRY;
 305
 306         LASSERT(!cfs_in_interrupt());     /* might sleep */
 307
 308         /* Let's setup deadline for reply unlink. */
 309         if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_BULK_UNLINK) &&
 310             async && req->rq_bulk_deadline == 0)
 311                 req->rq_bulk_deadline = cfs_time_current_sec() + LONG_UNLINK;
 312
 313         if (!ptlrpc_client_bulk_active(req))  /* completed or */
 314                 RETURN(1);                    /* never registered */
 315
 316         LASSERT(desc->bd_req == req);  /* bd_req NULL until registered */
 317
 318         /* the unlink ensures the callback happens ASAP and is the last
 319          * one.  If it fails, it must be because completion just happened,
 320          * but we must still l_wait_event() in this case to give liblustre
 321          * a chance to run client_bulk_callback() */
 322
 323         LNetMDUnlink(desc->bd_md_h);
 324
 325         if (!ptlrpc_client_bulk_active(req))  /* completed or */
 326                 RETURN(1);                    /* never registered */
 327
 328         /* Move to "Unregistering" phase as bulk was not unlinked yet. */
 329         ptlrpc_rqphase_move(req, RQ_PHASE_UNREGISTERING);
 330
 331         /* Do not wait for unlink to finish. */
 332         if (async)
 333                 RETURN(0);
 334
 335         if (req->rq_set != NULL)
 336                 wq = &req->rq_set->set_waitq;
 337         else
 338                 wq = &req->rq_reply_waitq;
 339
 340         for (;;) {
 341                 /* Network access will complete in finite time but the HUGE
 342                  * timeout lets us CWARN for visibility of sluggish NALs */
 343                 lwi = LWI_TIMEOUT_INTERVAL(cfs_time_seconds(LONG_UNLINK),
 344                                            cfs_time_seconds(1), NULL, NULL);
 345                 rc = l_wait_event(*wq, !ptlrpc_client_bulk_active(req), &lwi);
 346                 if (rc == 0) {
 347                         ptlrpc_rqphase_move(req, req->rq_next_phase);
 348                         RETURN(1);
 349                 }
 350
 351                 LASSERT(rc == -ETIMEDOUT);
 352                 DEBUG_REQ(D_WARNING, req, "Unexpectedly long timeout: desc %p",
 353                           desc);
 354         }
 355         RETURN(0);
 356 }
 357
 358 static void ptlrpc_at_set_reply(struct ptlrpc_request *req, int flags)
 359 {
 360         struct ptlrpc_service *svc = req->rq_rqbd->rqbd_service;
 361         int service_time = max_t(int, cfs_time_current_sec() -
 362                                  req->rq_arrival_time.tv_sec, 1);
 363
 364         if (!(flags & PTLRPC_REPLY_EARLY) &&
 365             (req->rq_type != PTL_RPC_MSG_ERR) &&
 366             (req->rq_reqmsg != NULL) &&
 367             !(lustre_msg_get_flags(req->rq_reqmsg) &
 368               (MSG_RESENT | MSG_REPLAY |
 369                MSG_REQ_REPLAY_DONE | MSG_LOCK_REPLAY_DONE))) {
 370                 /* early replies, errors and recovery requests don't count
 371                  * toward our service time estimate */
 372                 int oldse = at_measured(&svc->srv_at_estimate, service_time);
 373                 if (oldse != 0)
 374                         DEBUG_REQ(D_ADAPTTO, req,
 375                                   "svc %s changed estimate from %d to %d",
 376                                   svc->srv_name, oldse,
 377                                   at_get(&svc->srv_at_estimate));
 378         }
 379         /* Report actual service time for client latency calc */
 380         lustre_msg_set_service_time(req->rq_repmsg, service_time);
 381         /* Report service time estimate for future client reqs, but report 0
 382          * (to be ignored by client) if it's a error reply during recovery.
 383          * (bz15815) */
 384         if (req->rq_type == PTL_RPC_MSG_ERR &&
 385             (req->rq_export == NULL || req->rq_export->exp_obd->obd_recovering))
 386                 lustre_msg_set_timeout(req->rq_repmsg, 0);
 387         else
 388                 lustre_msg_set_timeout(req->rq_repmsg,
 389                                        at_get(&svc->srv_at_estimate));
 390
 391         if (req->rq_reqmsg &&
 392             !(lustre_msghdr_get_flags(req->rq_reqmsg) & MSGHDR_AT_SUPPORT)) {
 393                 CDEBUG(D_ADAPTTO, "No early reply support: flags=%#x "
 394                        "req_flags=%#x magic=%d:%x/%x len=%d\n",
 395                        flags, lustre_msg_get_flags(req->rq_reqmsg),
 396                        lustre_msg_is_v1(req->rq_reqmsg),
 397                        lustre_msg_get_magic(req->rq_reqmsg),
 398                        lustre_msg_get_magic(req->rq_repmsg), req->rq_replen);
 399         }
 400 }
 401
 402 /**
 403  * Send request reply from request \a req reply buffer.
 404  * \a flags defines reply types
 405  * Returns 0 on sucess or error code
 406  */
 407 int ptlrpc_send_reply(struct ptlrpc_request *req, int flags)
 408 {
 409         struct ptlrpc_service     *svc = req->rq_rqbd->rqbd_service;
 410         struct ptlrpc_reply_state *rs = req->rq_reply_state;
 411         struct ptlrpc_connection  *conn;
 412         int                        rc;
 413
 414         /* We must already have a reply buffer (only ptlrpc_error() may be
 415          * called without one). The reply generated by sptlrpc layer (e.g.
 416          * error notify, etc.) might have NULL rq->reqmsg; Otherwise we must
 417          * have a request buffer which is either the actual (swabbed) incoming
 418          * request, or a saved copy if this is a req saved in
 419          * target_queue_final_reply().
 420          */
 421         LASSERT (req->rq_no_reply == 0);
 422         LASSERT (req->rq_reqbuf != NULL);
 423         LASSERT (rs != NULL);
 424         LASSERT ((flags & PTLRPC_REPLY_MAYBE_DIFFICULT) || !rs->rs_difficult);
 425         LASSERT (req->rq_repmsg != NULL);
 426         LASSERT (req->rq_repmsg == rs->rs_msg);
 427         LASSERT (rs->rs_cb_id.cbid_fn == reply_out_callback);
 428         LASSERT (rs->rs_cb_id.cbid_arg == rs);
 429
 430         /* There may be no rq_export during failover */
 431
 432         if (unlikely(req->rq_export && req->rq_export->exp_obd &&
 433                      req->rq_export->exp_obd->obd_fail)) {
 434                 /* Failed obd's only send ENODEV */
 435                 req->rq_type = PTL_RPC_MSG_ERR;
 436                 req->rq_status = -ENODEV;
 437                 CDEBUG(D_HA, "sending ENODEV from failed obd %d\n",
 438                        req->rq_export->exp_obd->obd_minor);
 439         }
 440
 441         if (req->rq_type != PTL_RPC_MSG_ERR)
 442                 req->rq_type = PTL_RPC_MSG_REPLY;
 443
 444         lustre_msg_set_type(req->rq_repmsg, req->rq_type);
 445         lustre_msg_set_status(req->rq_repmsg, req->rq_status);
 446         lustre_msg_set_opc(req->rq_repmsg,
 447                 req->rq_reqmsg ? lustre_msg_get_opc(req->rq_reqmsg) : 0);
 448
 449         target_pack_pool_reply(req);
 450
 451         ptlrpc_at_set_reply(req, flags);
 452
 453         if (req->rq_export == NULL || req->rq_export->exp_connection == NULL)
 454                 conn = ptlrpc_connection_get(req->rq_peer, req->rq_self, NULL);
 455         else
 456                 conn = ptlrpc_connection_addref(req->rq_export->exp_connection);
 457
 458         if (unlikely(conn == NULL)) {
 459                 CERROR("not replying on NULL connection\n"); /* bug 9635 */
 460                 return -ENOTCONN;
 461         }
 462         cfs_atomic_inc (&svc->srv_outstanding_replies);
 463         ptlrpc_rs_addref(rs);                   /* +1 ref for the network */
 464
 465         rc = sptlrpc_svc_wrap_reply(req);
 466         if (unlikely(rc))
 467                 goto out;
 468
 469         req->rq_sent = cfs_time_current_sec();
 470
 471         rc = ptl_send_buf (&rs->rs_md_h, rs->rs_repbuf, rs->rs_repdata_len,
 472                            (rs->rs_difficult && !rs->rs_no_ack) ?
 473                            LNET_ACK_REQ : LNET_NOACK_REQ,
 474                            &rs->rs_cb_id, conn, svc->srv_rep_portal,
 475                            req->rq_xid, req->rq_reply_off);
 476 out:
 477         if (unlikely(rc != 0)) {
 478                 cfs_atomic_dec (&svc->srv_outstanding_replies);
 479                 ptlrpc_req_drop_rs(req);
 480         }
 481         ptlrpc_connection_put(conn);
 482         return rc;
 483 }
 484
 485 int ptlrpc_reply (struct ptlrpc_request *req)
 486 {
 487         if (req->rq_no_reply)
 488                 return 0;
 489         else
 490                 return (ptlrpc_send_reply(req, 0));
 491 }
 492
 493 /**
 494  * For request \a req send an error reply back. Create empty
 495  * reply buffers if necessary.
 496  */
 497 int ptlrpc_send_error(struct ptlrpc_request *req, int may_be_difficult)
 498 {
 499         int rc;
 500         ENTRY;
 501
 502         if (req->rq_no_reply)
 503                 RETURN(0);
 504
 505         if (!req->rq_repmsg) {
 506                 rc = lustre_pack_reply(req, 1, NULL, NULL);
 507                 if (rc)
 508                         RETURN(rc);
 509         }
 510
 511         req->rq_type = PTL_RPC_MSG_ERR;
 512
 513         rc = ptlrpc_send_reply(req, may_be_difficult);
 514         RETURN(rc);
 515 }
 516
 517 int ptlrpc_error(struct ptlrpc_request *req)
 518 {
 519         return ptlrpc_send_error(req, 0);
 520 }
 521
 522 /**
 523  * Send request \a request.
 524  * if \a noreply is set, don't expect any reply back and don't set up
 525  * reply buffers.
 526  * Returns 0 on success or error code.
 527  */
 528 int ptl_send_rpc(struct ptlrpc_request *request, int noreply)
 529 {
 530         int rc;
 531         int rc2;
 532         struct ptlrpc_connection *connection;
 533         lnet_handle_me_t  reply_me_h;
 534         lnet_md_t         reply_md;
 535         struct obd_device *obd = request->rq_import->imp_obd;
 536         ENTRY;
 537
 538         if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_DROP_RPC))
 539                 RETURN(0);
 540
 541         LASSERT(request->rq_type == PTL_RPC_MSG_REQUEST);
 542         LASSERT(request->rq_wait_ctx == 0);
 543
 544         /* If this is a re-transmit, we're required to have disengaged
 545          * cleanly from the previous attempt */
 546         LASSERT(!request->rq_receiving_reply);
 547
 548         if (request->rq_import->imp_obd &&
 549             request->rq_import->imp_obd->obd_fail) {
 550                 CDEBUG(D_HA, "muting rpc for failed imp obd %s\n",
 551                        request->rq_import->imp_obd->obd_name);
 552                 /* this prevents us from waiting in ptlrpc_queue_wait */
 553                 request->rq_err = 1;
 554                 request->rq_status = -ENODEV;
 555                 RETURN(-ENODEV);
 556         }
 557
 558         connection = request->rq_import->imp_connection;
 559
 560         lustre_msg_set_handle(request->rq_reqmsg,
 561                               &request->rq_import->imp_remote_handle);
 562         lustre_msg_set_type(request->rq_reqmsg, PTL_RPC_MSG_REQUEST);
 563         lustre_msg_set_conn_cnt(request->rq_reqmsg,
 564                                 request->rq_import->imp_conn_cnt);
 565         lustre_msghdr_set_flags(request->rq_reqmsg,
 566                                 request->rq_import->imp_msghdr_flags);
 567
 568         if (request->rq_resend)
 569                 lustre_msg_add_flags(request->rq_reqmsg, MSG_RESENT);
 570
 571         rc = sptlrpc_cli_wrap_request(request);
 572         if (rc)
 573                 RETURN(rc);
 574
 575         /* bulk register should be done after wrap_request() */
 576         if (request->rq_bulk != NULL) {
 577                 rc = ptlrpc_register_bulk (request);
 578                 if (rc != 0)
 579                         RETURN(rc);
 580         }
 581
 582         if (!noreply) {
 583                 LASSERT (request->rq_replen != 0);
 584                 if (request->rq_repbuf == NULL) {
 585                         LASSERT(request->rq_repdata == NULL);
 586                         LASSERT(request->rq_repmsg == NULL);
 587                         rc = sptlrpc_cli_alloc_repbuf(request,
 588                                                       request->rq_replen);
 589                         if (rc) {
 590                                 /* this prevents us from looping in
 591                                  * ptlrpc_queue_wait */
 592                                 request->rq_err = 1;
 593                                 request->rq_status = rc;
 594                                 GOTO(cleanup_bulk, rc);
 595                         }
 596                 } else {
 597                         request->rq_repdata = NULL;
 598                         request->rq_repmsg = NULL;
 599                 }
 600
 601                 rc = LNetMEAttach(request->rq_reply_portal,/*XXX FIXME bug 249*/
 602                                   connection->c_peer, request->rq_xid, 0,
 603                                   LNET_UNLINK, LNET_INS_AFTER, &reply_me_h);
 604                 if (rc != 0) {
 605                         CERROR("LNetMEAttach failed: %d\n", rc);
 606                         LASSERT (rc == -ENOMEM);
 607                         GOTO(cleanup_bulk, rc = -ENOMEM);
 608                 }
 609         }
 610
 611         cfs_spin_lock(&request->rq_lock);
 612         /* If the MD attach succeeds, there _will_ be a reply_in callback */
 613         request->rq_receiving_reply = !noreply;
 614         /* We are responsible for unlinking the reply buffer */
 615         request->rq_must_unlink = !noreply;
 616         /* Clear any flags that may be present from previous sends. */
 617         request->rq_replied = 0;
 618         request->rq_err = 0;
 619         request->rq_timedout = 0;
 620         request->rq_net_err = 0;
 621         request->rq_resend = 0;
 622         request->rq_restart = 0;
 623         request->rq_reply_truncate = 0;
 624         cfs_spin_unlock(&request->rq_lock);
 625
 626         if (!noreply) {
 627                 reply_md.start     = request->rq_repbuf;
 628                 reply_md.length    = request->rq_repbuf_len;
 629                 /* Allow multiple early replies */
 630                 reply_md.threshold = LNET_MD_THRESH_INF;
 631                 /* Manage remote for early replies */
 632                 reply_md.options   = PTLRPC_MD_OPTIONS | LNET_MD_OP_PUT |
 633                         LNET_MD_MANAGE_REMOTE |
 634                         LNET_MD_TRUNCATE; /* allow to make EOVERFLOW error */;
 635                 reply_md.user_ptr  = &request->rq_reply_cbid;
 636                 reply_md.eq_handle = ptlrpc_eq_h;
 637
 638                 /* We must see the unlink callback to unset rq_must_unlink,
 639                    so we can't auto-unlink */
 640                 rc = LNetMDAttach(reply_me_h, reply_md, LNET_RETAIN,
 641                                   &request->rq_reply_md_h);
 642                 if (rc != 0) {
 643                         CERROR("LNetMDAttach failed: %d\n", rc);
 644                         LASSERT (rc == -ENOMEM);
 645                         cfs_spin_lock(&request->rq_lock);
 646                         /* ...but the MD attach didn't succeed... */
 647                         request->rq_receiving_reply = 0;
 648                         cfs_spin_unlock(&request->rq_lock);
 649                         GOTO(cleanup_me, rc = -ENOMEM);
 650                 }
 651
 652                 CDEBUG(D_NET, "Setup reply buffer: %u bytes, xid "LPU64
 653                        ", portal %u\n",
 654                        request->rq_repbuf_len, request->rq_xid,
 655                        request->rq_reply_portal);
 656         }
 657
 658         /* add references on request for request_out_callback */
 659         ptlrpc_request_addref(request);
 660         if (obd->obd_svc_stats != NULL)
 661                 lprocfs_counter_add(obd->obd_svc_stats, PTLRPC_REQACTIVE_CNTR,
 662                         cfs_atomic_read(&request->rq_import->imp_inflight));
 663
 664         OBD_FAIL_TIMEOUT(OBD_FAIL_PTLRPC_DELAY_SEND, request->rq_timeout + 5);
 665
 666         cfs_gettimeofday(&request->rq_arrival_time);
 667         request->rq_sent = cfs_time_current_sec();
 668         /* We give the server rq_timeout secs to process the req, and
 669            add the network latency for our local timeout. */
 670         request->rq_deadline = request->rq_sent + request->rq_timeout +
 671                 ptlrpc_at_get_net_latency(request);
 672
 673         ptlrpc_pinger_sending_on_import(request->rq_import);
 674
 675         DEBUG_REQ(D_INFO, request, "send flg=%x",
 676                   lustre_msg_get_flags(request->rq_reqmsg));
 677         rc = ptl_send_buf(&request->rq_req_md_h,
 678                           request->rq_reqbuf, request->rq_reqdata_len,
 679                           LNET_NOACK_REQ, &request->rq_req_cbid,
 680                           connection,
 681                           request->rq_request_portal,
 682                           request->rq_xid, 0);
 683         if (rc == 0)
 684                 RETURN(rc);
 685
 686         ptlrpc_req_finished(request);
 687         if (noreply)
 688                 RETURN(rc);
 689
 690  cleanup_me:
 691         /* MEUnlink is safe; the PUT didn't even get off the ground, and
 692          * nobody apart from the PUT's target has the right nid+XID to
 693          * access the reply buffer. */
 694         rc2 = LNetMEUnlink(reply_me_h);
 695         LASSERT (rc2 == 0);
 696         /* UNLINKED callback called synchronously */
 697         LASSERT(!request->rq_receiving_reply);
 698
 699  cleanup_bulk:
 700         /* We do sync unlink here as there was no real transfer here so
 701          * the chance to have long unlink to sluggish net is smaller here. */
 702         ptlrpc_unregister_bulk(request, 0);
 703         return rc;
 704 }
 705
 706 /**
 707  * Register request buffer descriptor for request receiving.
 708  */
 709 int ptlrpc_register_rqbd(struct ptlrpc_request_buffer_desc *rqbd)
 710 {
 711         struct ptlrpc_service   *service = rqbd->rqbd_service;
 712         static lnet_process_id_t  match_id = {LNET_NID_ANY, LNET_PID_ANY};
 713         int                      rc;
 714         lnet_md_t                 md;
 715         lnet_handle_me_t          me_h;
 716
 717         CDEBUG(D_NET, "LNetMEAttach: portal %d\n",
 718                service->srv_req_portal);
 719
 720         if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_RQBD))
 721                 return (-ENOMEM);
 722
 723         rc = LNetMEAttach(service->srv_req_portal,
 724                           match_id, 0, ~0, LNET_UNLINK, LNET_INS_AFTER, &me_h);
 725         if (rc != 0) {
 726                 CERROR("LNetMEAttach failed: %d\n", rc);
 727                 return (-ENOMEM);
 728         }
 729
 730         LASSERT(rqbd->rqbd_refcount == 0);
 731         rqbd->rqbd_refcount = 1;
 732
 733         md.start     = rqbd->rqbd_buffer;
 734         md.length    = service->srv_buf_size;
 735         md.max_size  = service->srv_max_req_size;
 736         md.threshold = LNET_MD_THRESH_INF;
 737         md.options   = PTLRPC_MD_OPTIONS | LNET_MD_OP_PUT | LNET_MD_MAX_SIZE;
 738         md.user_ptr  = &rqbd->rqbd_cbid;
 739         md.eq_handle = ptlrpc_eq_h;
 740
 741         rc = LNetMDAttach(me_h, md, LNET_UNLINK, &rqbd->rqbd_md_h);
 742         if (rc == 0)
 743                 return (0);
 744
 745         CERROR("LNetMDAttach failed: %d; \n", rc);
 746         LASSERT (rc == -ENOMEM);
 747         rc = LNetMEUnlink (me_h);
 748         LASSERT (rc == 0);
 749         rqbd->rqbd_refcount = 0;
 750
 751         return (-ENOMEM);
 752 }