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, 2015, 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         CDEBUG(D_INFO, "conn=%p id %s\n", conn, libcfs_id2str(conn->c_peer));
  60         md.start = base;
  61         md.length = len;
  62         md.threshold = (ack == LNET_ACK_REQ) ? 2 : 1;
  63         md.options = PTLRPC_MD_OPTIONS;
  64         md.user_ptr = cbid;
  65         md.eq_handle = ptlrpc_eq_h;
  66
  67         if (unlikely(ack == LNET_ACK_REQ &&
  68                      OBD_FAIL_CHECK_ORSET(OBD_FAIL_PTLRPC_ACK,
  69                                           OBD_FAIL_ONCE))) {
  70                 /* don't ask for the ack to simulate failing client */
  71                 ack = LNET_NOACK_REQ;
  72         }
  73
  74         rc = LNetMDBind(md, LNET_UNLINK, mdh);
  75         if (unlikely(rc != 0)) {
  76                 CERROR("LNetMDBind failed: %d\n", rc);
  77                 LASSERT(rc == -ENOMEM);
  78                 return -ENOMEM;
  79         }
  80
  81         CDEBUG(D_NET, "Sending %d bytes to portal %d, xid %lld, offset %u\n",
  82                len, portal, xid, offset);
  83
  84         rc = LNetPut(conn->c_self, *mdh, ack,
  85                      conn->c_peer, portal, xid, offset, 0);
  86         if (unlikely(rc != 0)) {
  87                 int rc2;
  88                 /* We're going to get an UNLINK event when I unlink below,
  89                  * which will complete just like any other failed send, so
  90                  * I fall through and return success here! */
  91                 CERROR("LNetPut(%s, %d, %lld) failed: %d\n",
  92                        libcfs_id2str(conn->c_peer), portal, xid, rc);
  93                 rc2 = LNetMDUnlink(*mdh);
  94                 LASSERTF(rc2 == 0, "rc2 = %d\n", rc2);
  95         }
  96
  97         return 0;
  98 }
  99
 100 static void mdunlink_iterate_helper(lnet_handle_md_t *bd_mds, int count)
 101 {
 102         int i;
 103
 104         for (i = 0; i < count; i++)
 105                 LNetMDUnlink(bd_mds[i]);
 106 }
 107
 108 /**
 109  * Register bulk at the sender for later transfer.
 110  * Returns 0 on success or error code.
 111  */
 112 static int ptlrpc_register_bulk(struct ptlrpc_request *req)
 113 {
 114         struct ptlrpc_bulk_desc *desc = req->rq_bulk;
 115         lnet_process_id_t peer;
 116         int rc = 0;
 117         int rc2;
 118         int posted_md;
 119         int total_md;
 120         __u64 xid;
 121         lnet_handle_me_t me_h;
 122         lnet_md_t md;
 123
 124         if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_BULK_GET_NET))
 125                 return 0;
 126
 127         /* NB no locking required until desc is on the network */
 128         LASSERT(desc->bd_nob > 0);
 129         LASSERT(desc->bd_md_count == 0);
 130         LASSERT(desc->bd_md_max_brw <= PTLRPC_BULK_OPS_COUNT);
 131         LASSERT(desc->bd_iov_count <= PTLRPC_MAX_BRW_PAGES);
 132         LASSERT(desc->bd_req);
 133         LASSERT(desc->bd_type == BULK_PUT_SINK ||
 134                 desc->bd_type == BULK_GET_SOURCE);
 135
 136         /* cleanup the state of the bulk for it will be reused */
 137         if (req->rq_resend || req->rq_send_state == LUSTRE_IMP_REPLAY)
 138                 desc->bd_nob_transferred = 0;
 139         else
 140                 LASSERT(desc->bd_nob_transferred == 0);
 141
 142         desc->bd_failure = 0;
 143
 144         peer = desc->bd_import->imp_connection->c_peer;
 145
 146         LASSERT(desc->bd_cbid.cbid_fn == client_bulk_callback);
 147         LASSERT(desc->bd_cbid.cbid_arg == desc);
 148
 149         /* An XID is only used for a single request from the client.
 150          * For retried bulk transfers, a new XID will be allocated in
 151          * in ptlrpc_check_set() if it needs to be resent, so it is not
 152          * using the same RDMA match bits after an error.
 153          *
 154          * For multi-bulk RPCs, rq_xid is the last XID needed for bulks. The
 155          * first bulk XID is power-of-two aligned before rq_xid. LU-1431 */
 156         xid = req->rq_xid & ~((__u64)desc->bd_md_max_brw - 1);
 157         LASSERTF(!(desc->bd_registered &&
 158                    req->rq_send_state != LUSTRE_IMP_REPLAY) ||
 159                  xid != desc->bd_last_xid,
 160                  "registered: %d  rq_xid: %llu bd_last_xid: %llu\n",
 161                  desc->bd_registered, xid, desc->bd_last_xid);
 162
 163         total_md = (desc->bd_iov_count + LNET_MAX_IOV - 1) / LNET_MAX_IOV;
 164         desc->bd_registered = 1;
 165         desc->bd_last_xid = xid;
 166         desc->bd_md_count = total_md;
 167         md.user_ptr = &desc->bd_cbid;
 168         md.eq_handle = ptlrpc_eq_h;
 169         md.threshold = 1;                      /* PUT or GET */
 170
 171         for (posted_md = 0; posted_md < total_md; posted_md++, xid++) {
 172                 md.options = PTLRPC_MD_OPTIONS |
 173                              ((desc->bd_type == BULK_GET_SOURCE) ?
 174                               LNET_MD_OP_GET : LNET_MD_OP_PUT);
 175                 ptlrpc_fill_bulk_md(&md, desc, posted_md);
 176
 177                 rc = LNetMEAttach(desc->bd_portal, peer, xid, 0,
 178                                   LNET_UNLINK, LNET_INS_AFTER, &me_h);
 179                 if (rc != 0) {
 180                         CERROR("%s: LNetMEAttach failed x%llu/%d: rc = %d\n",
 181                                desc->bd_import->imp_obd->obd_name, xid,
 182                                posted_md, rc);
 183                         break;
 184                 }
 185
 186                 /* About to let the network at it... */
 187                 rc = LNetMDAttach(me_h, md, LNET_UNLINK,
 188                                   &desc->bd_mds[posted_md]);
 189                 if (rc != 0) {
 190                         CERROR("%s: LNetMDAttach failed x%llu/%d: rc = %d\n",
 191                                desc->bd_import->imp_obd->obd_name, xid,
 192                                posted_md, rc);
 193                         rc2 = LNetMEUnlink(me_h);
 194                         LASSERT(rc2 == 0);
 195                         break;
 196                 }
 197         }
 198
 199         if (rc != 0) {
 200                 LASSERT(rc == -ENOMEM);
 201                 spin_lock(&desc->bd_lock);
 202                 desc->bd_md_count -= total_md - posted_md;
 203                 spin_unlock(&desc->bd_lock);
 204                 LASSERT(desc->bd_md_count >= 0);
 205                 mdunlink_iterate_helper(desc->bd_mds, desc->bd_md_max_brw);
 206                 req->rq_status = -ENOMEM;
 207                 return -ENOMEM;
 208         }
 209
 210         /* Set rq_xid to matchbits of the final bulk so that server can
 211          * infer the number of bulks that were prepared */
 212         req->rq_xid = --xid;
 213         LASSERTF(desc->bd_last_xid == (req->rq_xid & PTLRPC_BULK_OPS_MASK),
 214                  "bd_last_xid = x%llu, rq_xid = x%llu\n",
 215                  desc->bd_last_xid, req->rq_xid);
 216
 217         spin_lock(&desc->bd_lock);
 218         /* Holler if peer manages to touch buffers before he knows the xid */
 219         if (desc->bd_md_count != total_md)
 220                 CWARN("%s: Peer %s touched %d buffers while I registered\n",
 221                       desc->bd_import->imp_obd->obd_name, libcfs_id2str(peer),
 222                       total_md - desc->bd_md_count);
 223         spin_unlock(&desc->bd_lock);
 224
 225         CDEBUG(D_NET, "Setup %u bulk %s buffers: %u pages %u bytes, xid x%#llx-%#llx, portal %u\n",
 226                desc->bd_md_count,
 227                desc->bd_type == BULK_GET_SOURCE ? "get-source" : "put-sink",
 228                desc->bd_iov_count, desc->bd_nob,
 229                desc->bd_last_xid, req->rq_xid, desc->bd_portal);
 230
 231         return 0;
 232 }
 233
 234 /**
 235  * Disconnect a bulk desc from the network. Idempotent. Not
 236  * thread-safe (i.e. only interlocks with completion callback).
 237  * Returns 1 on success or 0 if network unregistration failed for whatever
 238  * reason.
 239  */
 240 int ptlrpc_unregister_bulk(struct ptlrpc_request *req, int async)
 241 {
 242         struct ptlrpc_bulk_desc *desc = req->rq_bulk;
 243         wait_queue_head_t *wq;
 244         struct l_wait_info lwi;
 245         int rc;
 246
 247         LASSERT(!in_interrupt());     /* might sleep */
 248
 249         /* Let's setup deadline for reply unlink. */
 250         if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_BULK_UNLINK) &&
 251             async && req->rq_bulk_deadline == 0)
 252                 req->rq_bulk_deadline = ktime_get_real_seconds() + LONG_UNLINK;
 253
 254         if (ptlrpc_client_bulk_active(req) == 0)        /* completed or */
 255                 return 1;                               /* never registered */
 256
 257         LASSERT(desc->bd_req == req);  /* bd_req NULL until registered */
 258
 259         /* the unlink ensures the callback happens ASAP and is the last
 260          * one.  If it fails, it must be because completion just happened,
 261          * but we must still l_wait_event() in this case to give liblustre
 262          * a chance to run client_bulk_callback() */
 263         mdunlink_iterate_helper(desc->bd_mds, desc->bd_md_max_brw);
 264
 265         if (ptlrpc_client_bulk_active(req) == 0)        /* completed or */
 266                 return 1;                               /* never registered */
 267
 268         /* Move to "Unregistering" phase as bulk was not unlinked yet. */
 269         ptlrpc_rqphase_move(req, RQ_PHASE_UNREGISTERING);
 270
 271         /* Do not wait for unlink to finish. */
 272         if (async)
 273                 return 0;
 274
 275         if (req->rq_set)
 276                 wq = &req->rq_set->set_waitq;
 277         else
 278                 wq = &req->rq_reply_waitq;
 279
 280         for (;;) {
 281                 /* Network access will complete in finite time but the HUGE
 282                  * timeout lets us CWARN for visibility of sluggish NALs */
 283                 lwi = LWI_TIMEOUT_INTERVAL(cfs_time_seconds(LONG_UNLINK),
 284                                            cfs_time_seconds(1), NULL, NULL);
 285                 rc = l_wait_event(*wq, !ptlrpc_client_bulk_active(req), &lwi);
 286                 if (rc == 0) {
 287                         ptlrpc_rqphase_move(req, req->rq_next_phase);
 288                         return 1;
 289                 }
 290
 291                 LASSERT(rc == -ETIMEDOUT);
 292                 DEBUG_REQ(D_WARNING, req, "Unexpectedly long timeout: desc %p",
 293                           desc);
 294         }
 295         return 0;
 296 }
 297 EXPORT_SYMBOL(ptlrpc_unregister_bulk);
 298
 299 static void ptlrpc_at_set_reply(struct ptlrpc_request *req, int flags)
 300 {
 301         struct ptlrpc_service_part *svcpt = req->rq_rqbd->rqbd_svcpt;
 302         struct ptlrpc_service *svc = svcpt->scp_service;
 303         int service_time = max_t(int, ktime_get_real_seconds() -
 304                                  req->rq_arrival_time.tv_sec, 1);
 305
 306         if (!(flags & PTLRPC_REPLY_EARLY) &&
 307             (req->rq_type != PTL_RPC_MSG_ERR) && req->rq_reqmsg &&
 308             !(lustre_msg_get_flags(req->rq_reqmsg) &
 309               (MSG_RESENT | MSG_REPLAY |
 310                MSG_REQ_REPLAY_DONE | MSG_LOCK_REPLAY_DONE))) {
 311                 /* early replies, errors and recovery requests don't count
 312                  * toward our service time estimate */
 313                 int oldse = at_measured(&svcpt->scp_at_estimate, service_time);
 314
 315                 if (oldse != 0) {
 316                         DEBUG_REQ(D_ADAPTTO, req,
 317                                   "svc %s changed estimate from %d to %d",
 318                                   svc->srv_name, oldse,
 319                                   at_get(&svcpt->scp_at_estimate));
 320                 }
 321         }
 322         /* Report actual service time for client latency calc */
 323         lustre_msg_set_service_time(req->rq_repmsg, service_time);
 324         /* Report service time estimate for future client reqs, but report 0
 325          * (to be ignored by client) if it's a error reply during recovery.
 326          * (bz15815) */
 327         if (req->rq_type == PTL_RPC_MSG_ERR && !req->rq_export)
 328                 lustre_msg_set_timeout(req->rq_repmsg, 0);
 329         else
 330                 lustre_msg_set_timeout(req->rq_repmsg,
 331                                        at_get(&svcpt->scp_at_estimate));
 332
 333         if (req->rq_reqmsg &&
 334             !(lustre_msghdr_get_flags(req->rq_reqmsg) & MSGHDR_AT_SUPPORT)) {
 335                 CDEBUG(D_ADAPTTO, "No early reply support: flags=%#x req_flags=%#x magic=%x/%x len=%d\n",
 336                        flags, lustre_msg_get_flags(req->rq_reqmsg),
 337                        lustre_msg_get_magic(req->rq_reqmsg),
 338                        lustre_msg_get_magic(req->rq_repmsg), req->rq_replen);
 339         }
 340 }
 341
 342 /**
 343  * Send request reply from request \a req reply buffer.
 344  * \a flags defines reply types
 345  * Returns 0 on success or error code
 346  */
 347 int ptlrpc_send_reply(struct ptlrpc_request *req, int flags)
 348 {
 349         struct ptlrpc_reply_state *rs = req->rq_reply_state;
 350         struct ptlrpc_connection *conn;
 351         int rc;
 352
 353         /* We must already have a reply buffer (only ptlrpc_error() may be
 354          * called without one). The reply generated by sptlrpc layer (e.g.
 355          * error notify, etc.) might have NULL rq->reqmsg; Otherwise we must
 356          * have a request buffer which is either the actual (swabbed) incoming
 357          * request, or a saved copy if this is a req saved in
 358          * target_queue_final_reply().
 359          */
 360         LASSERT(req->rq_no_reply == 0);
 361         LASSERT(req->rq_reqbuf);
 362         LASSERT(rs);
 363         LASSERT((flags & PTLRPC_REPLY_MAYBE_DIFFICULT) || !rs->rs_difficult);
 364         LASSERT(req->rq_repmsg);
 365         LASSERT(req->rq_repmsg == rs->rs_msg);
 366         LASSERT(rs->rs_cb_id.cbid_fn == reply_out_callback);
 367         LASSERT(rs->rs_cb_id.cbid_arg == rs);
 368
 369         /* There may be no rq_export during failover */
 370
 371         if (unlikely(req->rq_export && req->rq_export->exp_obd &&
 372                      req->rq_export->exp_obd->obd_fail)) {
 373                 /* Failed obd's only send ENODEV */
 374                 req->rq_type = PTL_RPC_MSG_ERR;
 375                 req->rq_status = -ENODEV;
 376                 CDEBUG(D_HA, "sending ENODEV from failed obd %d\n",
 377                        req->rq_export->exp_obd->obd_minor);
 378         }
 379
 380         /* In order to keep interoperability with the client (< 2.3) which
 381          * doesn't have pb_jobid in ptlrpc_body, We have to shrink the
 382          * ptlrpc_body in reply buffer to ptlrpc_body_v2, otherwise, the
 383          * reply buffer on client will be overflow.
 384          *
 385          * XXX Remove this whenever we drop the interoperability with
 386          * such client.
 387          */
 388         req->rq_replen = lustre_shrink_msg(req->rq_repmsg, 0,
 389                                            sizeof(struct ptlrpc_body_v2), 1);
 390
 391         if (req->rq_type != PTL_RPC_MSG_ERR)
 392                 req->rq_type = PTL_RPC_MSG_REPLY;
 393
 394         lustre_msg_set_type(req->rq_repmsg, req->rq_type);
 395         lustre_msg_set_status(req->rq_repmsg,
 396                               ptlrpc_status_hton(req->rq_status));
 397         lustre_msg_set_opc(req->rq_repmsg,
 398                 req->rq_reqmsg ? lustre_msg_get_opc(req->rq_reqmsg) : 0);
 399
 400         target_pack_pool_reply(req);
 401
 402         ptlrpc_at_set_reply(req, flags);
 403
 404         if (!req->rq_export || !req->rq_export->exp_connection)
 405                 conn = ptlrpc_connection_get(req->rq_peer, req->rq_self, NULL);
 406         else
 407                 conn = ptlrpc_connection_addref(req->rq_export->exp_connection);
 408
 409         if (unlikely(!conn)) {
 410                 CERROR("not replying on NULL connection\n"); /* bug 9635 */
 411                 return -ENOTCONN;
 412         }
 413         ptlrpc_rs_addref(rs);              /* +1 ref for the network */
 414
 415         rc = sptlrpc_svc_wrap_reply(req);
 416         if (unlikely(rc))
 417                 goto out;
 418
 419         req->rq_sent = ktime_get_real_seconds();
 420
 421         rc = ptl_send_buf(&rs->rs_md_h, rs->rs_repbuf, rs->rs_repdata_len,
 422                           (rs->rs_difficult && !rs->rs_no_ack) ?
 423                           LNET_ACK_REQ : LNET_NOACK_REQ,
 424                           &rs->rs_cb_id, conn,
 425                           ptlrpc_req2svc(req)->srv_rep_portal,
 426                           req->rq_xid, req->rq_reply_off);
 427 out:
 428         if (unlikely(rc != 0))
 429                 ptlrpc_req_drop_rs(req);
 430         ptlrpc_connection_put(conn);
 431         return rc;
 432 }
 433 EXPORT_SYMBOL(ptlrpc_send_reply);
 434
 435 int ptlrpc_reply(struct ptlrpc_request *req)
 436 {
 437         if (req->rq_no_reply)
 438                 return 0;
 439         return ptlrpc_send_reply(req, 0);
 440 }
 441 EXPORT_SYMBOL(ptlrpc_reply);
 442
 443 /**
 444  * For request \a req send an error reply back. Create empty
 445  * reply buffers if necessary.
 446  */
 447 int ptlrpc_send_error(struct ptlrpc_request *req, int may_be_difficult)
 448 {
 449         int rc;
 450
 451         if (req->rq_no_reply)
 452                 return 0;
 453
 454         if (!req->rq_repmsg) {
 455                 rc = lustre_pack_reply(req, 1, NULL, NULL);
 456                 if (rc)
 457                         return rc;
 458         }
 459
 460         if (req->rq_status != -ENOSPC && req->rq_status != -EACCES &&
 461             req->rq_status != -EPERM && req->rq_status != -ENOENT &&
 462             req->rq_status != -EINPROGRESS && req->rq_status != -EDQUOT)
 463                 req->rq_type = PTL_RPC_MSG_ERR;
 464
 465         rc = ptlrpc_send_reply(req, may_be_difficult);
 466         return rc;
 467 }
 468 EXPORT_SYMBOL(ptlrpc_send_error);
 469
 470 int ptlrpc_error(struct ptlrpc_request *req)
 471 {
 472         return ptlrpc_send_error(req, 0);
 473 }
 474 EXPORT_SYMBOL(ptlrpc_error);
 475
 476 /**
 477  * Send request \a request.
 478  * if \a noreply is set, don't expect any reply back and don't set up
 479  * reply buffers.
 480  * Returns 0 on success or error code.
 481  */
 482 int ptl_send_rpc(struct ptlrpc_request *request, int noreply)
 483 {
 484         int rc;
 485         int rc2;
 486         int mpflag = 0;
 487         struct ptlrpc_connection *connection;
 488         lnet_handle_me_t reply_me_h;
 489         lnet_md_t reply_md;
 490         struct obd_device *obd = request->rq_import->imp_obd;
 491
 492         if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_DROP_RPC))
 493                 return 0;
 494
 495         LASSERT(request->rq_type == PTL_RPC_MSG_REQUEST);
 496         LASSERT(request->rq_wait_ctx == 0);
 497
 498         /* If this is a re-transmit, we're required to have disengaged
 499          * cleanly from the previous attempt */
 500         LASSERT(!request->rq_receiving_reply);
 501         LASSERT(!((lustre_msg_get_flags(request->rq_reqmsg) & MSG_REPLAY) &&
 502                 (request->rq_import->imp_state == LUSTRE_IMP_FULL)));
 503
 504         if (unlikely(obd && obd->obd_fail)) {
 505                 CDEBUG(D_HA, "muting rpc for failed imp obd %s\n",
 506                         obd->obd_name);
 507                 /* this prevents us from waiting in ptlrpc_queue_wait */
 508                 spin_lock(&request->rq_lock);
 509                 request->rq_err = 1;
 510                 spin_unlock(&request->rq_lock);
 511                 request->rq_status = -ENODEV;
 512                 return -ENODEV;
 513         }
 514
 515         connection = request->rq_import->imp_connection;
 516
 517         lustre_msg_set_handle(request->rq_reqmsg,
 518                               &request->rq_import->imp_remote_handle);
 519         lustre_msg_set_type(request->rq_reqmsg, PTL_RPC_MSG_REQUEST);
 520         lustre_msg_set_conn_cnt(request->rq_reqmsg,
 521                                 request->rq_import->imp_conn_cnt);
 522         lustre_msghdr_set_flags(request->rq_reqmsg,
 523                                 request->rq_import->imp_msghdr_flags);
 524
 525         if (request->rq_resend)
 526                 lustre_msg_add_flags(request->rq_reqmsg, MSG_RESENT);
 527
 528         if (request->rq_memalloc)
 529                 mpflag = cfs_memory_pressure_get_and_set();
 530
 531         rc = sptlrpc_cli_wrap_request(request);
 532         if (rc)
 533                 goto out;
 534
 535         /* bulk register should be done after wrap_request() */
 536         if (request->rq_bulk) {
 537                 rc = ptlrpc_register_bulk(request);
 538                 if (rc != 0)
 539                         goto out;
 540         }
 541
 542         if (!noreply) {
 543                 LASSERT(request->rq_replen != 0);
 544                 if (!request->rq_repbuf) {
 545                         LASSERT(!request->rq_repdata);
 546                         LASSERT(!request->rq_repmsg);
 547                         rc = sptlrpc_cli_alloc_repbuf(request,
 548                                                       request->rq_replen);
 549                         if (rc) {
 550                                 /* this prevents us from looping in
 551                                  * ptlrpc_queue_wait */
 552                                 spin_lock(&request->rq_lock);
 553                                 request->rq_err = 1;
 554                                 spin_unlock(&request->rq_lock);
 555                                 request->rq_status = rc;
 556                                 goto cleanup_bulk;
 557                         }
 558                 } else {
 559                         request->rq_repdata = NULL;
 560                         request->rq_repmsg = NULL;
 561                 }
 562
 563                 rc = LNetMEAttach(request->rq_reply_portal,/*XXX FIXME bug 249*/
 564                                   connection->c_peer, request->rq_xid, 0,
 565                                   LNET_UNLINK, LNET_INS_AFTER, &reply_me_h);
 566                 if (rc != 0) {
 567                         CERROR("LNetMEAttach failed: %d\n", rc);
 568                         LASSERT(rc == -ENOMEM);
 569                         rc = -ENOMEM;
 570                         goto cleanup_bulk;
 571                 }
 572         }
 573
 574         spin_lock(&request->rq_lock);
 575         /* If the MD attach succeeds, there _will_ be a reply_in callback */
 576         request->rq_receiving_reply = !noreply;
 577         request->rq_req_unlink = 1;
 578         /* We are responsible for unlinking the reply buffer */
 579         request->rq_reply_unlink = !noreply;
 580         /* Clear any flags that may be present from previous sends. */
 581         request->rq_replied = 0;
 582         request->rq_err = 0;
 583         request->rq_timedout = 0;
 584         request->rq_net_err = 0;
 585         request->rq_resend = 0;
 586         request->rq_restart = 0;
 587         request->rq_reply_truncate = 0;
 588         spin_unlock(&request->rq_lock);
 589
 590         if (!noreply) {
 591                 reply_md.start = request->rq_repbuf;
 592                 reply_md.length = request->rq_repbuf_len;
 593                 /* Allow multiple early replies */
 594                 reply_md.threshold = LNET_MD_THRESH_INF;
 595                 /* Manage remote for early replies */
 596                 reply_md.options = PTLRPC_MD_OPTIONS | LNET_MD_OP_PUT |
 597                         LNET_MD_MANAGE_REMOTE |
 598                         LNET_MD_TRUNCATE; /* allow to make EOVERFLOW error */
 599                 reply_md.user_ptr = &request->rq_reply_cbid;
 600                 reply_md.eq_handle = ptlrpc_eq_h;
 601
 602                 /* We must see the unlink callback to unset rq_reply_unlink,
 603                    so we can't auto-unlink */
 604                 rc = LNetMDAttach(reply_me_h, reply_md, LNET_RETAIN,
 605                                   &request->rq_reply_md_h);
 606                 if (rc != 0) {
 607                         CERROR("LNetMDAttach failed: %d\n", rc);
 608                         LASSERT(rc == -ENOMEM);
 609                         spin_lock(&request->rq_lock);
 610                         /* ...but the MD attach didn't succeed... */
 611                         request->rq_receiving_reply = 0;
 612                         spin_unlock(&request->rq_lock);
 613                         rc = -ENOMEM;
 614                         goto cleanup_me;
 615                 }
 616
 617                 CDEBUG(D_NET, "Setup reply buffer: %u bytes, xid %llu, portal %u\n",
 618                        request->rq_repbuf_len, request->rq_xid,
 619                        request->rq_reply_portal);
 620         }
 621
 622         /* add references on request for request_out_callback */
 623         ptlrpc_request_addref(request);
 624         if (obd && obd->obd_svc_stats)
 625                 lprocfs_counter_add(obd->obd_svc_stats, PTLRPC_REQACTIVE_CNTR,
 626                         atomic_read(&request->rq_import->imp_inflight));
 627
 628         OBD_FAIL_TIMEOUT(OBD_FAIL_PTLRPC_DELAY_SEND, request->rq_timeout + 5);
 629
 630         ktime_get_real_ts64(&request->rq_arrival_time);
 631         request->rq_sent = ktime_get_real_seconds();
 632         /* We give the server rq_timeout secs to process the req, and
 633            add the network latency for our local timeout. */
 634         request->rq_deadline = request->rq_sent + request->rq_timeout +
 635                 ptlrpc_at_get_net_latency(request);
 636
 637         ptlrpc_pinger_sending_on_import(request->rq_import);
 638
 639         DEBUG_REQ(D_INFO, request, "send flg=%x",
 640                   lustre_msg_get_flags(request->rq_reqmsg));
 641         rc = ptl_send_buf(&request->rq_req_md_h,
 642                           request->rq_reqbuf, request->rq_reqdata_len,
 643                           LNET_NOACK_REQ, &request->rq_req_cbid,
 644                           connection,
 645                           request->rq_request_portal,
 646                           request->rq_xid, 0);
 647         if (rc == 0)
 648                 goto out;
 649
 650         ptlrpc_req_finished(request);
 651         if (noreply)
 652                 goto out;
 653
 654  cleanup_me:
 655         /* MEUnlink is safe; the PUT didn't even get off the ground, and
 656          * nobody apart from the PUT's target has the right nid+XID to
 657          * access the reply buffer. */
 658         rc2 = LNetMEUnlink(reply_me_h);
 659         LASSERT(rc2 == 0);
 660         /* UNLINKED callback called synchronously */
 661         LASSERT(!request->rq_receiving_reply);
 662
 663  cleanup_bulk:
 664         /* We do sync unlink here as there was no real transfer here so
 665          * the chance to have long unlink to sluggish net is smaller here. */
 666         ptlrpc_unregister_bulk(request, 0);
 667  out:
 668         if (request->rq_memalloc)
 669                 cfs_memory_pressure_restore(mpflag);
 670         return rc;
 671 }
 672 EXPORT_SYMBOL(ptl_send_rpc);
 673
 674 /**
 675  * Register request buffer descriptor for request receiving.
 676  */
 677 int ptlrpc_register_rqbd(struct ptlrpc_request_buffer_desc *rqbd)
 678 {
 679         struct ptlrpc_service *service = rqbd->rqbd_svcpt->scp_service;
 680         static lnet_process_id_t match_id = {LNET_NID_ANY, LNET_PID_ANY};
 681         int rc;
 682         lnet_md_t md;
 683         lnet_handle_me_t me_h;
 684
 685         CDEBUG(D_NET, "LNetMEAttach: portal %d\n",
 686                service->srv_req_portal);
 687
 688         if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_RQBD))
 689                 return -ENOMEM;
 690
 691         /* NB: CPT affinity service should use new LNet flag LNET_INS_LOCAL,
 692          * which means buffer can only be attached on local CPT, and LND
 693          * threads can find it by grabbing a local lock */
 694         rc = LNetMEAttach(service->srv_req_portal,
 695                           match_id, 0, ~0, LNET_UNLINK,
 696                           rqbd->rqbd_svcpt->scp_cpt >= 0 ?
 697                           LNET_INS_LOCAL : LNET_INS_AFTER, &me_h);
 698         if (rc != 0) {
 699                 CERROR("LNetMEAttach failed: %d\n", rc);
 700                 return -ENOMEM;
 701         }
 702
 703         LASSERT(rqbd->rqbd_refcount == 0);
 704         rqbd->rqbd_refcount = 1;
 705
 706         md.start = rqbd->rqbd_buffer;
 707         md.length = service->srv_buf_size;
 708         md.max_size = service->srv_max_req_size;
 709         md.threshold = LNET_MD_THRESH_INF;
 710         md.options = PTLRPC_MD_OPTIONS | LNET_MD_OP_PUT | LNET_MD_MAX_SIZE;
 711         md.user_ptr = &rqbd->rqbd_cbid;
 712         md.eq_handle = ptlrpc_eq_h;
 713
 714         rc = LNetMDAttach(me_h, md, LNET_UNLINK, &rqbd->rqbd_md_h);
 715         if (rc == 0)
 716                 return 0;
 717
 718         CERROR("LNetMDAttach failed: %d;\n", rc);
 719         LASSERT(rc == -ENOMEM);
 720         rc = LNetMEUnlink(me_h);
 721         LASSERT(rc == 0);
 722         rqbd->rqbd_refcount = 0;
 723
 724         return -ENOMEM;
 725 }