kernel/drivers/staging/lustre/lustre/ptlrpc/niobuf.c

   1 /*
   2  * GPL HEADER START
   3  *
   4  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
   5  *
   6  * This program is free software; you can redistribute it and/or modify
   7  * it under the terms of the GNU General Public License version 2 only,
   8  * as published by the Free Software Foundation.
   9  *
  10  * This program is distributed in the hope that it will be useful, but
  11  * WITHOUT ANY WARRANTY; without even the implied warranty of
  12  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  13  * General Public License version 2 for more details (a copy is included
  14  * in the LICENSE file that accompanied this code).
  15  *
  16  * You should have received a copy of the GNU General Public License
  17  * version 2 along with this program; If not, see
  18  * http://www.sun.com/software/products/lustre/docs/GPLv2.pdf
  19  *
  20  * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
  21  * CA 95054 USA or visit www.sun.com if you need additional information or
  22  * have any questions.
  23  *
  24  * GPL HEADER END
  25  */
  26 /*
  27  * Copyright (c) 2002, 2010, Oracle and/or its affiliates. All rights reserved.
  28  * Use is subject to license terms.
  29  *
  30  * Copyright (c) 2011, 2012, Intel Corporation.
  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 #include "../include/obd_support.h"
  39 #include "../include/lustre_net.h"
  40 #include "../include/lustre_lib.h"
  41 #include "../include/obd.h"
  42 #include "../include/obd_class.h"
  43 #include "ptlrpc_internal.h"
  44
  45 /**
  46  * Helper function. Sends \a len bytes from \a base at offset \a offset
  47  * over \a conn connection to portal \a portal.
  48  * Returns 0 on success or error code.
  49  */
  50 static int ptl_send_buf(lnet_handle_md_t *mdh, void *base, int len,
  51                         lnet_ack_req_t ack, struct ptlrpc_cb_id *cbid,
  52                         struct ptlrpc_connection *conn, int portal, __u64 xid,
  53                         unsigned int offset)
  54 {
  55         int rc;
  56         lnet_md_t md;
  57
  58         LASSERT(portal != 0);
  59         LASSERT(conn != NULL);
  60         CDEBUG(D_INFO, "conn=%p id %s\n", conn, libcfs_id2str(conn->c_peer));
  61         md.start = base;
  62         md.length = len;
  63         md.threshold = (ack == LNET_ACK_REQ) ? 2 : 1;
  64         md.options = PTLRPC_MD_OPTIONS;
  65         md.user_ptr = cbid;
  66         md.eq_handle = ptlrpc_eq_h;
  67
  68         if (unlikely(ack == LNET_ACK_REQ &&
  69                      OBD_FAIL_CHECK_ORSET(OBD_FAIL_PTLRPC_ACK,
  70                                           OBD_FAIL_ONCE))) {
  71                 /* don't ask for the ack to simulate failing client */
  72                 ack = LNET_NOACK_REQ;
  73         }
  74
  75         rc = LNetMDBind(md, LNET_UNLINK, mdh);
  76         if (unlikely(rc != 0)) {
  77                 CERROR("LNetMDBind failed: %d\n", rc);
  78                 LASSERT(rc == -ENOMEM);
  79                 return -ENOMEM;
  80         }
  81
  82         CDEBUG(D_NET, "Sending %d bytes to portal %d, xid %lld, offset %u\n",
  83                len, portal, xid, offset);
  84
  85         rc = LNetPut(conn->c_self, *mdh, ack,
  86                      conn->c_peer, portal, xid, offset, 0);
  87         if (unlikely(rc != 0)) {
  88                 int rc2;
  89                 /* We're going to get an UNLINK event when I unlink below,
  90                  * which will complete just like any other failed send, so
  91                  * I fall through and return success here! */
  92                 CERROR("LNetPut(%s, %d, %lld) failed: %d\n",
  93                        libcfs_id2str(conn->c_peer), portal, xid, rc);
  94                 rc2 = LNetMDUnlink(*mdh);
  95                 LASSERTF(rc2 == 0, "rc2 = %d\n", rc2);
  96         }
  97
  98         return 0;
  99 }
 100
 101 static void mdunlink_iterate_helper(lnet_handle_md_t *bd_mds, int count)
 102 {
 103         int i;
 104
 105         for (i = 0; i < count; i++)
 106                 LNetMDUnlink(bd_mds[i]);
 107 }
 108
 109 /**
 110  * Register bulk at the sender for later transfer.
 111  * Returns 0 on success or error code.
 112  */
 113 static int ptlrpc_register_bulk(struct ptlrpc_request *req)
 114 {
 115         struct ptlrpc_bulk_desc *desc = req->rq_bulk;
 116         lnet_process_id_t peer;
 117         int rc = 0;
 118         int rc2;
 119         int posted_md;
 120         int total_md;
 121         __u64 xid;
 122         lnet_handle_me_t me_h;
 123         lnet_md_t md;
 124
 125         if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_BULK_GET_NET))
 126                 return 0;
 127
 128         /* NB no locking required until desc is on the network */
 129         LASSERT(desc->bd_nob > 0);
 130         LASSERT(desc->bd_md_count == 0);
 131         LASSERT(desc->bd_md_max_brw <= PTLRPC_BULK_OPS_COUNT);
 132         LASSERT(desc->bd_iov_count <= PTLRPC_MAX_BRW_PAGES);
 133         LASSERT(desc->bd_req != NULL);
 134         LASSERT(desc->bd_type == BULK_PUT_SINK ||
 135                 desc->bd_type == BULK_GET_SOURCE);
 136
 137         /* cleanup the state of the bulk for it will be reused */
 138         if (req->rq_resend || req->rq_send_state == LUSTRE_IMP_REPLAY)
 139                 desc->bd_nob_transferred = 0;
 140         else
 141                 LASSERT(desc->bd_nob_transferred == 0);
 142
 143         desc->bd_failure = 0;
 144
 145         peer = desc->bd_import->imp_connection->c_peer;
 146
 147         LASSERT(desc->bd_cbid.cbid_fn == client_bulk_callback);
 148         LASSERT(desc->bd_cbid.cbid_arg == desc);
 149
 150         /* An XID is only used for a single request from the client.
 151          * For retried bulk transfers, a new XID will be allocated in
 152          * in ptlrpc_check_set() if it needs to be resent, so it is not
 153          * using the same RDMA match bits after an error.
 154          *
 155          * For multi-bulk RPCs, rq_xid is the last XID needed for bulks. The
 156          * first bulk XID is power-of-two aligned before rq_xid. LU-1431 */
 157         xid = req->rq_xid & ~((__u64)desc->bd_md_max_brw - 1);
 158         LASSERTF(!(desc->bd_registered &&
 159                    req->rq_send_state != LUSTRE_IMP_REPLAY) ||
 160                  xid != desc->bd_last_xid,
 161                  "registered: %d  rq_xid: %llu bd_last_xid: %llu\n",
 162                  desc->bd_registered, xid, desc->bd_last_xid);
 163
 164         total_md = (desc->bd_iov_count + LNET_MAX_IOV - 1) / LNET_MAX_IOV;
 165         desc->bd_registered = 1;
 166         desc->bd_last_xid = xid;
 167         desc->bd_md_count = total_md;
 168         md.user_ptr = &desc->bd_cbid;
 169         md.eq_handle = ptlrpc_eq_h;
 170         md.threshold = 1;                      /* PUT or GET */
 171
 172         for (posted_md = 0; posted_md < total_md; posted_md++, xid++) {
 173                 md.options = PTLRPC_MD_OPTIONS |
 174                              ((desc->bd_type == BULK_GET_SOURCE) ?
 175                               LNET_MD_OP_GET : LNET_MD_OP_PUT);
 176                 ptlrpc_fill_bulk_md(&md, desc, posted_md);
 177
 178                 rc = LNetMEAttach(desc->bd_portal, peer, xid, 0,
 179                                   LNET_UNLINK, LNET_INS_AFTER, &me_h);
 180                 if (rc != 0) {
 181                         CERROR("%s: LNetMEAttach failed x%llu/%d: rc = %d\n",
 182                                desc->bd_import->imp_obd->obd_name, xid,
 183                                posted_md, rc);
 184                         break;
 185                 }
 186
 187                 /* About to let the network at it... */
 188                 rc = LNetMDAttach(me_h, md, LNET_UNLINK,
 189                                   &desc->bd_mds[posted_md]);
 190                 if (rc != 0) {
 191                         CERROR("%s: LNetMDAttach failed x%llu/%d: rc = %d\n",
 192                                desc->bd_import->imp_obd->obd_name, xid,
 193                                posted_md, rc);
 194                         rc2 = LNetMEUnlink(me_h);
 195                         LASSERT(rc2 == 0);
 196                         break;
 197                 }
 198         }
 199
 200         if (rc != 0) {
 201                 LASSERT(rc == -ENOMEM);
 202                 spin_lock(&desc->bd_lock);
 203                 desc->bd_md_count -= total_md - posted_md;
 204                 spin_unlock(&desc->bd_lock);
 205                 LASSERT(desc->bd_md_count >= 0);
 206                 mdunlink_iterate_helper(desc->bd_mds, desc->bd_md_max_brw);
 207                 req->rq_status = -ENOMEM;
 208                 return -ENOMEM;
 209         }
 210
 211         /* Set rq_xid to matchbits of the final bulk so that server can
 212          * infer the number of bulks that were prepared */
 213         req->rq_xid = --xid;
 214         LASSERTF(desc->bd_last_xid == (req->rq_xid & PTLRPC_BULK_OPS_MASK),
 215                  "bd_last_xid = x%llu, rq_xid = x%llu\n",
 216                  desc->bd_last_xid, req->rq_xid);
 217
 218         spin_lock(&desc->bd_lock);
 219         /* Holler if peer manages to touch buffers before he knows the xid */
 220         if (desc->bd_md_count != total_md)
 221                 CWARN("%s: Peer %s touched %d buffers while I registered\n",
 222                       desc->bd_import->imp_obd->obd_name, libcfs_id2str(peer),
 223                       total_md - desc->bd_md_count);
 224         spin_unlock(&desc->bd_lock);
 225
 226         CDEBUG(D_NET, "Setup %u bulk %s buffers: %u pages %u bytes, xid x%#llx-%#llx, portal %u\n",
 227                desc->bd_md_count,
 228                desc->bd_type == BULK_GET_SOURCE ? "get-source" : "put-sink",
 229                desc->bd_iov_count, desc->bd_nob,
 230                desc->bd_last_xid, req->rq_xid, desc->bd_portal);
 231
 232         return 0;
 233 }
 234
 235 /**
 236  * Disconnect a bulk desc from the network. Idempotent. Not
 237  * thread-safe (i.e. only interlocks with completion callback).
 238  * Returns 1 on success or 0 if network unregistration failed for whatever
 239  * reason.
 240  */
 241 int ptlrpc_unregister_bulk(struct ptlrpc_request *req, int async)
 242 {
 243         struct ptlrpc_bulk_desc *desc = req->rq_bulk;
 244         wait_queue_head_t *wq;
 245         struct l_wait_info lwi;
 246         int rc;
 247
 248         LASSERT(!in_interrupt());     /* might sleep */
 249
 250         /* Let's setup deadline for reply unlink. */
 251         if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_BULK_UNLINK) &&
 252             async && req->rq_bulk_deadline == 0)
 253                 req->rq_bulk_deadline = ktime_get_real_seconds() + LONG_UNLINK;
 254
 255         if (ptlrpc_client_bulk_active(req) == 0)        /* completed or */
 256                 return 1;                               /* never registered */
 257
 258         LASSERT(desc->bd_req == req);  /* bd_req NULL until registered */
 259
 260         /* the unlink ensures the callback happens ASAP and is the last
 261          * one.  If it fails, it must be because completion just happened,
 262          * but we must still l_wait_event() in this case to give liblustre
 263          * a chance to run client_bulk_callback() */
 264         mdunlink_iterate_helper(desc->bd_mds, desc->bd_md_max_brw);
 265
 266         if (ptlrpc_client_bulk_active(req) == 0)        /* completed or */
 267                 return 1;                               /* never registered */
 268
 269         /* Move to "Unregistering" phase as bulk was not unlinked yet. */
 270         ptlrpc_rqphase_move(req, RQ_PHASE_UNREGISTERING);
 271
 272         /* Do not wait for unlink to finish. */
 273         if (async)
 274                 return 0;
 275
 276         if (req->rq_set != NULL)
 277                 wq = &req->rq_set->set_waitq;
 278         else
 279                 wq = &req->rq_reply_waitq;
 280
 281         for (;;) {
 282                 /* Network access will complete in finite time but the HUGE
 283                  * timeout lets us CWARN for visibility of sluggish NALs */
 284                 lwi = LWI_TIMEOUT_INTERVAL(cfs_time_seconds(LONG_UNLINK),
 285                                            cfs_time_seconds(1), NULL, NULL);
 286                 rc = l_wait_event(*wq, !ptlrpc_client_bulk_active(req), &lwi);
 287                 if (rc == 0) {
 288                         ptlrpc_rqphase_move(req, req->rq_next_phase);
 289                         return 1;
 290                 }
 291
 292                 LASSERT(rc == -ETIMEDOUT);
 293                 DEBUG_REQ(D_WARNING, req, "Unexpectedly long timeout: desc %p",
 294                           desc);
 295         }
 296         return 0;
 297 }
 298 EXPORT_SYMBOL(ptlrpc_unregister_bulk);
 299
 300 static void ptlrpc_at_set_reply(struct ptlrpc_request *req, int flags)
 301 {
 302         struct ptlrpc_service_part *svcpt = req->rq_rqbd->rqbd_svcpt;
 303         struct ptlrpc_service *svc = svcpt->scp_service;
 304         int service_time = max_t(int, ktime_get_real_seconds() -
 305                                  req->rq_arrival_time.tv_sec, 1);
 306
 307         if (!(flags & PTLRPC_REPLY_EARLY) &&
 308             (req->rq_type != PTL_RPC_MSG_ERR) &&
 309             (req->rq_reqmsg != NULL) &&
 310             !(lustre_msg_get_flags(req->rq_reqmsg) &
 311               (MSG_RESENT | MSG_REPLAY |
 312                MSG_REQ_REPLAY_DONE | MSG_LOCK_REPLAY_DONE))) {
 313                 /* early replies, errors and recovery requests don't count
 314                  * toward our service time estimate */
 315                 int oldse = at_measured(&svcpt->scp_at_estimate, service_time);
 316
 317                 if (oldse != 0) {
 318                         DEBUG_REQ(D_ADAPTTO, req,
 319                                   "svc %s changed estimate from %d to %d",
 320                                   svc->srv_name, oldse,
 321                                   at_get(&svcpt->scp_at_estimate));
 322                 }
 323         }
 324         /* Report actual service time for client latency calc */
 325         lustre_msg_set_service_time(req->rq_repmsg, service_time);
 326         /* Report service time estimate for future client reqs, but report 0
 327          * (to be ignored by client) if it's a error reply during recovery.
 328          * (bz15815) */
 329         if (req->rq_type == PTL_RPC_MSG_ERR && !req->rq_export)
 330                 lustre_msg_set_timeout(req->rq_repmsg, 0);
 331         else
 332                 lustre_msg_set_timeout(req->rq_repmsg,
 333                                        at_get(&svcpt->scp_at_estimate));
 334
 335         if (req->rq_reqmsg &&
 336             !(lustre_msghdr_get_flags(req->rq_reqmsg) & MSGHDR_AT_SUPPORT)) {
 337                 CDEBUG(D_ADAPTTO, "No early reply support: flags=%#x req_flags=%#x magic=%x/%x len=%d\n",
 338                        flags, lustre_msg_get_flags(req->rq_reqmsg),
 339                        lustre_msg_get_magic(req->rq_reqmsg),
 340                        lustre_msg_get_magic(req->rq_repmsg), req->rq_replen);
 341         }
 342 }
 343
 344 /**
 345  * Send request reply from request \a req reply buffer.
 346  * \a flags defines reply types
 347  * Returns 0 on success or error code
 348  */
 349 int ptlrpc_send_reply(struct ptlrpc_request *req, int flags)
 350 {
 351         struct ptlrpc_reply_state *rs = req->rq_reply_state;
 352         struct ptlrpc_connection *conn;
 353         int rc;
 354
 355         /* We must already have a reply buffer (only ptlrpc_error() may be
 356          * called without one). The reply generated by sptlrpc layer (e.g.
 357          * error notify, etc.) might have NULL rq->reqmsg; Otherwise we must
 358          * have a request buffer which is either the actual (swabbed) incoming
 359          * request, or a saved copy if this is a req saved in
 360          * target_queue_final_reply().
 361          */
 362         LASSERT(req->rq_no_reply == 0);
 363         LASSERT(req->rq_reqbuf != NULL);
 364         LASSERT(rs != NULL);
 365         LASSERT((flags & PTLRPC_REPLY_MAYBE_DIFFICULT) || !rs->rs_difficult);
 366         LASSERT(req->rq_repmsg != NULL);
 367         LASSERT(req->rq_repmsg == rs->rs_msg);
 368         LASSERT(rs->rs_cb_id.cbid_fn == reply_out_callback);
 369         LASSERT(rs->rs_cb_id.cbid_arg == rs);
 370
 371         /* There may be no rq_export during failover */
 372
 373         if (unlikely(req->rq_export && req->rq_export->exp_obd &&
 374                      req->rq_export->exp_obd->obd_fail)) {
 375                 /* Failed obd's only send ENODEV */
 376                 req->rq_type = PTL_RPC_MSG_ERR;
 377                 req->rq_status = -ENODEV;
 378                 CDEBUG(D_HA, "sending ENODEV from failed obd %d\n",
 379                        req->rq_export->exp_obd->obd_minor);
 380         }
 381
 382         /* In order to keep interoperability with the client (< 2.3) which
 383          * doesn't have pb_jobid in ptlrpc_body, We have to shrink the
 384          * ptlrpc_body in reply buffer to ptlrpc_body_v2, otherwise, the
 385          * reply buffer on client will be overflow.
 386          *
 387          * XXX Remove this whenever we drop the interoperability with
 388          * such client.
 389          */
 390         req->rq_replen = lustre_shrink_msg(req->rq_repmsg, 0,
 391                                            sizeof(struct ptlrpc_body_v2), 1);
 392
 393         if (req->rq_type != PTL_RPC_MSG_ERR)
 394                 req->rq_type = PTL_RPC_MSG_REPLY;
 395
 396         lustre_msg_set_type(req->rq_repmsg, req->rq_type);
 397         lustre_msg_set_status(req->rq_repmsg,
 398                               ptlrpc_status_hton(req->rq_status));
 399         lustre_msg_set_opc(req->rq_repmsg,
 400                 req->rq_reqmsg ? lustre_msg_get_opc(req->rq_reqmsg) : 0);
 401
 402         target_pack_pool_reply(req);
 403
 404         ptlrpc_at_set_reply(req, flags);
 405
 406         if (req->rq_export == NULL || req->rq_export->exp_connection == NULL)
 407                 conn = ptlrpc_connection_get(req->rq_peer, req->rq_self, NULL);
 408         else
 409                 conn = ptlrpc_connection_addref(req->rq_export->exp_connection);
 410
 411         if (unlikely(conn == NULL)) {
 412                 CERROR("not replying on NULL connection\n"); /* bug 9635 */
 413                 return -ENOTCONN;
 414         }
 415         ptlrpc_rs_addref(rs);              /* +1 ref for the network */
 416
 417         rc = sptlrpc_svc_wrap_reply(req);
 418         if (unlikely(rc))
 419                 goto out;
 420
 421         req->rq_sent = ktime_get_real_seconds();
 422
 423         rc = ptl_send_buf(&rs->rs_md_h, rs->rs_repbuf, rs->rs_repdata_len,
 424                           (rs->rs_difficult && !rs->rs_no_ack) ?
 425                           LNET_ACK_REQ : LNET_NOACK_REQ,
 426                           &rs->rs_cb_id, conn,
 427                           ptlrpc_req2svc(req)->srv_rep_portal,
 428                           req->rq_xid, req->rq_reply_off);
 429 out:
 430         if (unlikely(rc != 0))
 431                 ptlrpc_req_drop_rs(req);
 432         ptlrpc_connection_put(conn);
 433         return rc;
 434 }
 435 EXPORT_SYMBOL(ptlrpc_send_reply);
 436
 437 int ptlrpc_reply(struct ptlrpc_request *req)
 438 {
 439         if (req->rq_no_reply)
 440                 return 0;
 441         return ptlrpc_send_reply(req, 0);
 442 }
 443 EXPORT_SYMBOL(ptlrpc_reply);
 444
 445 /**
 446  * For request \a req send an error reply back. Create empty
 447  * reply buffers if necessary.
 448  */
 449 int ptlrpc_send_error(struct ptlrpc_request *req, int may_be_difficult)
 450 {
 451         int rc;
 452
 453         if (req->rq_no_reply)
 454                 return 0;
 455
 456         if (!req->rq_repmsg) {
 457                 rc = lustre_pack_reply(req, 1, NULL, NULL);
 458                 if (rc)
 459                         return rc;
 460         }
 461
 462         if (req->rq_status != -ENOSPC && req->rq_status != -EACCES &&
 463             req->rq_status != -EPERM && req->rq_status != -ENOENT &&
 464             req->rq_status != -EINPROGRESS && req->rq_status != -EDQUOT)
 465                 req->rq_type = PTL_RPC_MSG_ERR;
 466
 467         rc = ptlrpc_send_reply(req, may_be_difficult);
 468         return rc;
 469 }
 470 EXPORT_SYMBOL(ptlrpc_send_error);
 471
 472 int ptlrpc_error(struct ptlrpc_request *req)
 473 {
 474         return ptlrpc_send_error(req, 0);
 475 }
 476 EXPORT_SYMBOL(ptlrpc_error);
 477
 478 /**
 479  * Send request \a request.
 480  * if \a noreply is set, don't expect any reply back and don't set up
 481  * reply buffers.
 482  * Returns 0 on success or error code.
 483  */
 484 int ptl_send_rpc(struct ptlrpc_request *request, int noreply)
 485 {
 486         int rc;
 487         int rc2;
 488         int mpflag = 0;
 489         struct ptlrpc_connection *connection;
 490         lnet_handle_me_t reply_me_h;
 491         lnet_md_t reply_md;
 492         struct obd_device *obd = request->rq_import->imp_obd;
 493
 494         if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_DROP_RPC))
 495                 return 0;
 496
 497         LASSERT(request->rq_type == PTL_RPC_MSG_REQUEST);
 498         LASSERT(request->rq_wait_ctx == 0);
 499
 500         /* If this is a re-transmit, we're required to have disengaged
 501          * cleanly from the previous attempt */
 502         LASSERT(!request->rq_receiving_reply);
 503         LASSERT(!((lustre_msg_get_flags(request->rq_reqmsg) & MSG_REPLAY) &&
 504                 (request->rq_import->imp_state == LUSTRE_IMP_FULL)));
 505
 506         if (unlikely(obd != NULL && obd->obd_fail)) {
 507                 CDEBUG(D_HA, "muting rpc for failed imp obd %s\n",
 508                         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                 request->rq_status = -ENODEV;
 514                 return -ENODEV;
 515         }
 516
 517         connection = request->rq_import->imp_connection;
 518
 519         lustre_msg_set_handle(request->rq_reqmsg,
 520                               &request->rq_import->imp_remote_handle);
 521         lustre_msg_set_type(request->rq_reqmsg, PTL_RPC_MSG_REQUEST);
 522         lustre_msg_set_conn_cnt(request->rq_reqmsg,
 523                                 request->rq_import->imp_conn_cnt);
 524         lustre_msghdr_set_flags(request->rq_reqmsg,
 525                                 request->rq_import->imp_msghdr_flags);
 526
 527         if (request->rq_resend)
 528                 lustre_msg_add_flags(request->rq_reqmsg, MSG_RESENT);
 529
 530         if (request->rq_memalloc)
 531                 mpflag = cfs_memory_pressure_get_and_set();
 532
 533         rc = sptlrpc_cli_wrap_request(request);
 534         if (rc)
 535                 goto out;
 536
 537         /* bulk register should be done after wrap_request() */
 538         if (request->rq_bulk != NULL) {
 539                 rc = ptlrpc_register_bulk(request);
 540                 if (rc != 0)
 541                         goto out;
 542         }
 543
 544         if (!noreply) {
 545                 LASSERT(request->rq_replen != 0);
 546                 if (request->rq_repbuf == NULL) {
 547                         LASSERT(request->rq_repdata == NULL);
 548                         LASSERT(request->rq_repmsg == NULL);
 549                         rc = sptlrpc_cli_alloc_repbuf(request,
 550                                                       request->rq_replen);
 551                         if (rc) {
 552                                 /* this prevents us from looping in
 553                                  * ptlrpc_queue_wait */
 554                                 spin_lock(&request->rq_lock);
 555                                 request->rq_err = 1;
 556                                 spin_unlock(&request->rq_lock);
 557                                 request->rq_status = rc;
 558                                 goto cleanup_bulk;
 559                         }
 560                 } else {
 561                         request->rq_repdata = NULL;
 562                         request->rq_repmsg = NULL;
 563                 }
 564
 565                 rc = LNetMEAttach(request->rq_reply_portal,/*XXX FIXME bug 249*/
 566                                   connection->c_peer, request->rq_xid, 0,
 567                                   LNET_UNLINK, LNET_INS_AFTER, &reply_me_h);
 568                 if (rc != 0) {
 569                         CERROR("LNetMEAttach failed: %d\n", rc);
 570                         LASSERT(rc == -ENOMEM);
 571                         rc = -ENOMEM;
 572                         goto cleanup_bulk;
 573                 }
 574         }
 575
 576         spin_lock(&request->rq_lock);
 577         /* If the MD attach succeeds, there _will_ be a reply_in callback */
 578         request->rq_receiving_reply = !noreply;
 579         request->rq_req_unlink = 1;
 580         /* We are responsible for unlinking the reply buffer */
 581         request->rq_reply_unlink = !noreply;
 582         /* Clear any flags that may be present from previous sends. */
 583         request->rq_replied = 0;
 584         request->rq_err = 0;
 585         request->rq_timedout = 0;
 586         request->rq_net_err = 0;
 587         request->rq_resend = 0;
 588         request->rq_restart = 0;
 589         request->rq_reply_truncate = 0;
 590         spin_unlock(&request->rq_lock);
 591
 592         if (!noreply) {
 593                 reply_md.start = request->rq_repbuf;
 594                 reply_md.length = request->rq_repbuf_len;
 595                 /* Allow multiple early replies */
 596                 reply_md.threshold = LNET_MD_THRESH_INF;
 597                 /* Manage remote for early replies */
 598                 reply_md.options = PTLRPC_MD_OPTIONS | LNET_MD_OP_PUT |
 599                         LNET_MD_MANAGE_REMOTE |
 600                         LNET_MD_TRUNCATE; /* allow to make EOVERFLOW error */
 601                 reply_md.user_ptr = &request->rq_reply_cbid;
 602                 reply_md.eq_handle = ptlrpc_eq_h;
 603
 604                 /* We must see the unlink callback to unset rq_reply_unlink,
 605                    so we can't auto-unlink */
 606                 rc = LNetMDAttach(reply_me_h, reply_md, LNET_RETAIN,
 607                                   &request->rq_reply_md_h);
 608                 if (rc != 0) {
 609                         CERROR("LNetMDAttach failed: %d\n", rc);
 610                         LASSERT(rc == -ENOMEM);
 611                         spin_lock(&request->rq_lock);
 612                         /* ...but the MD attach didn't succeed... */
 613                         request->rq_receiving_reply = 0;
 614                         spin_unlock(&request->rq_lock);
 615                         rc = -ENOMEM;
 616                         goto cleanup_me;
 617                 }
 618
 619                 CDEBUG(D_NET, "Setup reply buffer: %u bytes, xid %llu, portal %u\n",
 620                        request->rq_repbuf_len, request->rq_xid,
 621                        request->rq_reply_portal);
 622         }
 623
 624         /* add references on request for request_out_callback */
 625         ptlrpc_request_addref(request);
 626         if (obd != NULL && obd->obd_svc_stats != NULL)
 627                 lprocfs_counter_add(obd->obd_svc_stats, PTLRPC_REQACTIVE_CNTR,
 628                         atomic_read(&request->rq_import->imp_inflight));
 629
 630         OBD_FAIL_TIMEOUT(OBD_FAIL_PTLRPC_DELAY_SEND, request->rq_timeout + 5);
 631
 632         ktime_get_real_ts64(&request->rq_arrival_time);
 633         request->rq_sent = ktime_get_real_seconds();
 634         /* We give the server rq_timeout secs to process the req, and
 635            add the network latency for our local timeout. */
 636         request->rq_deadline = request->rq_sent + request->rq_timeout +
 637                 ptlrpc_at_get_net_latency(request);
 638
 639         ptlrpc_pinger_sending_on_import(request->rq_import);
 640
 641         DEBUG_REQ(D_INFO, request, "send flg=%x",
 642                   lustre_msg_get_flags(request->rq_reqmsg));
 643         rc = ptl_send_buf(&request->rq_req_md_h,
 644                           request->rq_reqbuf, request->rq_reqdata_len,
 645                           LNET_NOACK_REQ, &request->rq_req_cbid,
 646                           connection,
 647                           request->rq_request_portal,
 648                           request->rq_xid, 0);
 649         if (rc == 0)
 650                 goto out;
 651
 652         ptlrpc_req_finished(request);
 653         if (noreply)
 654                 goto out;
 655
 656  cleanup_me:
 657         /* MEUnlink is safe; the PUT didn't even get off the ground, and
 658          * nobody apart from the PUT's target has the right nid+XID to
 659          * access the reply buffer. */
 660         rc2 = LNetMEUnlink(reply_me_h);
 661         LASSERT(rc2 == 0);
 662         /* UNLINKED callback called synchronously */
 663         LASSERT(!request->rq_receiving_reply);
 664
 665  cleanup_bulk:
 666         /* We do sync unlink here as there was no real transfer here so
 667          * the chance to have long unlink to sluggish net is smaller here. */
 668         ptlrpc_unregister_bulk(request, 0);
 669  out:
 670         if (request->rq_memalloc)
 671                 cfs_memory_pressure_restore(mpflag);
 672         return rc;
 673 }
 674 EXPORT_SYMBOL(ptl_send_rpc);
 675
 676 /**
 677  * Register request buffer descriptor for request receiving.
 678  */
 679 int ptlrpc_register_rqbd(struct ptlrpc_request_buffer_desc *rqbd)
 680 {
 681         struct ptlrpc_service *service = rqbd->rqbd_svcpt->scp_service;
 682         static lnet_process_id_t match_id = {LNET_NID_ANY, LNET_PID_ANY};
 683         int rc;
 684         lnet_md_t md;
 685         lnet_handle_me_t me_h;
 686
 687         CDEBUG(D_NET, "LNetMEAttach: portal %d\n",
 688                service->srv_req_portal);
 689
 690         if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_RQBD))
 691                 return -ENOMEM;
 692
 693         /* NB: CPT affinity service should use new LNet flag LNET_INS_LOCAL,
 694          * which means buffer can only be attached on local CPT, and LND
 695          * threads can find it by grabbing a local lock */
 696         rc = LNetMEAttach(service->srv_req_portal,
 697                           match_id, 0, ~0, LNET_UNLINK,
 698                           rqbd->rqbd_svcpt->scp_cpt >= 0 ?
 699                           LNET_INS_LOCAL : LNET_INS_AFTER, &me_h);
 700         if (rc != 0) {
 701                 CERROR("LNetMEAttach failed: %d\n", rc);
 702                 return -ENOMEM;
 703         }
 704
 705         LASSERT(rqbd->rqbd_refcount == 0);
 706         rqbd->rqbd_refcount = 1;
 707
 708         md.start = rqbd->rqbd_buffer;
 709         md.length = service->srv_buf_size;
 710         md.max_size = service->srv_max_req_size;
 711         md.threshold = LNET_MD_THRESH_INF;
 712         md.options = PTLRPC_MD_OPTIONS | LNET_MD_OP_PUT | LNET_MD_MAX_SIZE;
 713         md.user_ptr = &rqbd->rqbd_cbid;
 714         md.eq_handle = ptlrpc_eq_h;
 715
 716         rc = LNetMDAttach(me_h, md, LNET_UNLINK, &rqbd->rqbd_md_h);
 717         if (rc == 0)
 718                 return 0;
 719
 720         CERROR("LNetMDAttach failed: %d;\n", rc);
 721         LASSERT(rc == -ENOMEM);
 722         rc = LNetMEUnlink(me_h);
 723         LASSERT(rc == 0);
 724         rqbd->rqbd_refcount = 0;
 725
 726         return -ENOMEM;
 727 }