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.gnu.org/licenses/gpl-2.0.html
  19  *
  20  * GPL HEADER END
  21  */
  22 /*
  23  * Copyright (c) 2002, 2010, Oracle and/or its affiliates. All rights reserved.
  24  * Use is subject to license terms.
  25  *
  26  * Copyright (c) 2011, 2015, Intel Corporation.
  27  */
  28 /*
  29  * This file is part of Lustre, http://www.lustre.org/
  30  * Lustre is a trademark of Sun Microsystems, Inc.
  31  */
  32
  33 #define DEBUG_SUBSYSTEM S_RPC
  34 #include <obd_support.h>
  35 #include <lustre_net.h>
  36 #include <lustre_lib.h>
  37 #include <obd.h>
  38 #include <obd_class.h>
  39 #include "ptlrpc_internal.h"
  40
  41 /**
  42  * Helper function. Sends \a len bytes from \a base at offset \a offset
  43  * over \a conn connection to portal \a portal.
  44  * Returns 0 on success or error code.
  45  */
  46 static int ptl_send_buf(struct lnet_handle_md *mdh, void *base, int len,
  47                         enum lnet_ack_req ack, struct ptlrpc_cb_id *cbid,
  48                         struct ptlrpc_connection *conn, int portal, __u64 xid,
  49                         unsigned int offset)
  50 {
  51         int rc;
  52         struct lnet_md md;
  53
  54         LASSERT(portal != 0);
  55         CDEBUG(D_INFO, "conn=%p id %s\n", conn, libcfs_id2str(conn->c_peer));
  56         md.start = base;
  57         md.length = len;
  58         md.threshold = (ack == LNET_ACK_REQ) ? 2 : 1;
  59         md.options = PTLRPC_MD_OPTIONS;
  60         md.user_ptr = cbid;
  61         md.eq_handle = ptlrpc_eq_h;
  62
  63         if (unlikely(ack == LNET_ACK_REQ &&
  64                      OBD_FAIL_CHECK_ORSET(OBD_FAIL_PTLRPC_ACK,
  65                                           OBD_FAIL_ONCE))) {
  66                 /* don't ask for the ack to simulate failing client */
  67                 ack = LNET_NOACK_REQ;
  68         }
  69
  70         rc = LNetMDBind(md, LNET_UNLINK, mdh);
  71         if (unlikely(rc != 0)) {
  72                 CERROR("LNetMDBind failed: %d\n", rc);
  73                 LASSERT(rc == -ENOMEM);
  74                 return -ENOMEM;
  75         }
  76
  77         CDEBUG(D_NET, "Sending %d bytes to portal %d, xid %lld, offset %u\n",
  78                len, portal, xid, offset);
  79
  80         rc = LNetPut(conn->c_self, *mdh, ack,
  81                      conn->c_peer, portal, xid, offset, 0);
  82         if (unlikely(rc != 0)) {
  83                 int rc2;
  84                 /* We're going to get an UNLINK event when I unlink below,
  85                  * which will complete just like any other failed send, so
  86                  * I fall through and return success here!
  87                  */
  88                 CERROR("LNetPut(%s, %d, %lld) failed: %d\n",
  89                        libcfs_id2str(conn->c_peer), portal, xid, rc);
  90                 rc2 = LNetMDUnlink(*mdh);
  91                 LASSERTF(rc2 == 0, "rc2 = %d\n", rc2);
  92         }
  93
  94         return 0;
  95 }
  96
  97 static void mdunlink_iterate_helper(struct lnet_handle_md *bd_mds, int count)
  98 {
  99         int i;
 100
 101         for (i = 0; i < count; i++)
 102                 LNetMDUnlink(bd_mds[i]);
 103 }
 104
 105 /**
 106  * Register bulk at the sender for later transfer.
 107  * Returns 0 on success or error code.
 108  */
 109 static int ptlrpc_register_bulk(struct ptlrpc_request *req)
 110 {
 111         struct ptlrpc_bulk_desc *desc = req->rq_bulk;
 112         struct lnet_process_id peer;
 113         int rc = 0;
 114         int rc2;
 115         int posted_md;
 116         int total_md;
 117         u64 mbits;
 118         struct lnet_handle_me me_h;
 119         struct lnet_md md;
 120
 121         if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_BULK_GET_NET))
 122                 return 0;
 123
 124         /* NB no locking required until desc is on the network */
 125         LASSERT(desc->bd_nob > 0);
 126         LASSERT(desc->bd_md_count == 0);
 127         LASSERT(desc->bd_md_max_brw <= PTLRPC_BULK_OPS_COUNT);
 128         LASSERT(desc->bd_iov_count <= PTLRPC_MAX_BRW_PAGES);
 129         LASSERT(desc->bd_req);
 130         LASSERT(ptlrpc_is_bulk_op_passive(desc->bd_type));
 131
 132         /* cleanup the state of the bulk for it will be reused */
 133         if (req->rq_resend || req->rq_send_state == LUSTRE_IMP_REPLAY)
 134                 desc->bd_nob_transferred = 0;
 135         else
 136                 LASSERT(desc->bd_nob_transferred == 0);
 137
 138         desc->bd_failure = 0;
 139
 140         peer = desc->bd_import->imp_connection->c_peer;
 141
 142         LASSERT(desc->bd_cbid.cbid_fn == client_bulk_callback);
 143         LASSERT(desc->bd_cbid.cbid_arg == desc);
 144
 145         total_md = DIV_ROUND_UP(desc->bd_iov_count, LNET_MAX_IOV);
 146         /* rq_mbits is matchbits of the final bulk */
 147         mbits = req->rq_mbits - total_md + 1;
 148
 149         LASSERTF(mbits == (req->rq_mbits & PTLRPC_BULK_OPS_MASK),
 150                  "first mbits = x%llu, last mbits = x%llu\n",
 151                  mbits, req->rq_mbits);
 152         LASSERTF(!(desc->bd_registered &&
 153                    req->rq_send_state != LUSTRE_IMP_REPLAY) ||
 154                  mbits != desc->bd_last_mbits,
 155                  "registered: %d  rq_mbits: %llu bd_last_mbits: %llu\n",
 156                  desc->bd_registered, mbits, desc->bd_last_mbits);
 157
 158         desc->bd_registered = 1;
 159         desc->bd_last_mbits = mbits;
 160         desc->bd_md_count = total_md;
 161         md.user_ptr = &desc->bd_cbid;
 162         md.eq_handle = ptlrpc_eq_h;
 163         md.threshold = 1;                      /* PUT or GET */
 164
 165         for (posted_md = 0; posted_md < total_md; posted_md++, mbits++) {
 166                 md.options = PTLRPC_MD_OPTIONS |
 167                              (ptlrpc_is_bulk_op_get(desc->bd_type) ?
 168                               LNET_MD_OP_GET : LNET_MD_OP_PUT);
 169                 ptlrpc_fill_bulk_md(&md, desc, posted_md);
 170
 171                 rc = LNetMEAttach(desc->bd_portal, peer, mbits, 0,
 172                                   LNET_UNLINK, LNET_INS_AFTER, &me_h);
 173                 if (rc != 0) {
 174                         CERROR("%s: LNetMEAttach failed x%llu/%d: rc = %d\n",
 175                                desc->bd_import->imp_obd->obd_name, mbits,
 176                                posted_md, rc);
 177                         break;
 178                 }
 179
 180                 /* About to let the network at it... */
 181                 rc = LNetMDAttach(me_h, md, LNET_UNLINK,
 182                                   &desc->bd_mds[posted_md]);
 183                 if (rc != 0) {
 184                         CERROR("%s: LNetMDAttach failed x%llu/%d: rc = %d\n",
 185                                desc->bd_import->imp_obd->obd_name, mbits,
 186                                posted_md, rc);
 187                         rc2 = LNetMEUnlink(me_h);
 188                         LASSERT(rc2 == 0);
 189                         break;
 190                 }
 191         }
 192
 193         if (rc != 0) {
 194                 LASSERT(rc == -ENOMEM);
 195                 spin_lock(&desc->bd_lock);
 196                 desc->bd_md_count -= total_md - posted_md;
 197                 spin_unlock(&desc->bd_lock);
 198                 LASSERT(desc->bd_md_count >= 0);
 199                 mdunlink_iterate_helper(desc->bd_mds, desc->bd_md_max_brw);
 200                 req->rq_status = -ENOMEM;
 201                 return -ENOMEM;
 202         }
 203
 204         spin_lock(&desc->bd_lock);
 205         /* Holler if peer manages to touch buffers before he knows the mbits */
 206         if (desc->bd_md_count != total_md)
 207                 CWARN("%s: Peer %s touched %d buffers while I registered\n",
 208                       desc->bd_import->imp_obd->obd_name, libcfs_id2str(peer),
 209                       total_md - desc->bd_md_count);
 210         spin_unlock(&desc->bd_lock);
 211
 212         CDEBUG(D_NET, "Setup %u bulk %s buffers: %u pages %u bytes, mbits x%#llx-%#llx, portal %u\n",
 213                desc->bd_md_count,
 214                ptlrpc_is_bulk_op_get(desc->bd_type) ? "get-source" : "put-sink",
 215                desc->bd_iov_count, desc->bd_nob,
 216                desc->bd_last_mbits, req->rq_mbits, desc->bd_portal);
 217
 218         return 0;
 219 }
 220
 221 /**
 222  * Disconnect a bulk desc from the network. Idempotent. Not
 223  * thread-safe (i.e. only interlocks with completion callback).
 224  * Returns 1 on success or 0 if network unregistration failed for whatever
 225  * reason.
 226  */
 227 int ptlrpc_unregister_bulk(struct ptlrpc_request *req, int async)
 228 {
 229         struct ptlrpc_bulk_desc *desc = req->rq_bulk;
 230         wait_queue_head_t *wq;
 231         struct l_wait_info lwi;
 232         int rc;
 233
 234         LASSERT(!in_interrupt());     /* might sleep */
 235
 236         /* Let's setup deadline for reply unlink. */
 237         if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_BULK_UNLINK) &&
 238             async && req->rq_bulk_deadline == 0 && cfs_fail_val == 0)
 239                 req->rq_bulk_deadline = ktime_get_real_seconds() + LONG_UNLINK;
 240
 241         if (ptlrpc_client_bulk_active(req) == 0)        /* completed or */
 242                 return 1;                               /* never registered */
 243
 244         LASSERT(desc->bd_req == req);  /* bd_req NULL until registered */
 245
 246         /* the unlink ensures the callback happens ASAP and is the last
 247          * one.  If it fails, it must be because completion just happened,
 248          * but we must still l_wait_event() in this case to give liblustre
 249          * a chance to run client_bulk_callback()
 250          */
 251         mdunlink_iterate_helper(desc->bd_mds, desc->bd_md_max_brw);
 252
 253         if (ptlrpc_client_bulk_active(req) == 0)        /* completed or */
 254                 return 1;                               /* never registered */
 255
 256         /* Move to "Unregistering" phase as bulk was not unlinked yet. */
 257         ptlrpc_rqphase_move(req, RQ_PHASE_UNREG_BULK);
 258
 259         /* Do not wait for unlink to finish. */
 260         if (async)
 261                 return 0;
 262
 263         if (req->rq_set)
 264                 wq = &req->rq_set->set_waitq;
 265         else
 266                 wq = &req->rq_reply_waitq;
 267
 268         for (;;) {
 269                 /* Network access will complete in finite time but the HUGE
 270                  * timeout lets us CWARN for visibility of sluggish LNDs
 271                  */
 272                 lwi = LWI_TIMEOUT_INTERVAL(cfs_time_seconds(LONG_UNLINK),
 273                                            cfs_time_seconds(1), NULL, NULL);
 274                 rc = l_wait_event(*wq, !ptlrpc_client_bulk_active(req), &lwi);
 275                 if (rc == 0) {
 276                         ptlrpc_rqphase_move(req, req->rq_next_phase);
 277                         return 1;
 278                 }
 279
 280                 LASSERT(rc == -ETIMEDOUT);
 281                 DEBUG_REQ(D_WARNING, req, "Unexpectedly long timeout: desc %p",
 282                           desc);
 283         }
 284         return 0;
 285 }
 286
 287 static void ptlrpc_at_set_reply(struct ptlrpc_request *req, int flags)
 288 {
 289         struct ptlrpc_service_part *svcpt = req->rq_rqbd->rqbd_svcpt;
 290         struct ptlrpc_service *svc = svcpt->scp_service;
 291         int service_time = max_t(int, ktime_get_real_seconds() -
 292                                  req->rq_arrival_time.tv_sec, 1);
 293
 294         if (!(flags & PTLRPC_REPLY_EARLY) &&
 295             (req->rq_type != PTL_RPC_MSG_ERR) && req->rq_reqmsg &&
 296             !(lustre_msg_get_flags(req->rq_reqmsg) &
 297               (MSG_RESENT | MSG_REPLAY |
 298                MSG_REQ_REPLAY_DONE | MSG_LOCK_REPLAY_DONE))) {
 299                 /* early replies, errors and recovery requests don't count
 300                  * toward our service time estimate
 301                  */
 302                 int oldse = at_measured(&svcpt->scp_at_estimate, service_time);
 303
 304                 if (oldse != 0) {
 305                         DEBUG_REQ(D_ADAPTTO, req,
 306                                   "svc %s changed estimate from %d to %d",
 307                                   svc->srv_name, oldse,
 308                                   at_get(&svcpt->scp_at_estimate));
 309                 }
 310         }
 311         /* Report actual service time for client latency calc */
 312         lustre_msg_set_service_time(req->rq_repmsg, service_time);
 313         /* Report service time estimate for future client reqs, but report 0
 314          * (to be ignored by client) if it's a error reply during recovery.
 315          * (bz15815)
 316          */
 317         if (req->rq_type == PTL_RPC_MSG_ERR && !req->rq_export)
 318                 lustre_msg_set_timeout(req->rq_repmsg, 0);
 319         else
 320                 lustre_msg_set_timeout(req->rq_repmsg,
 321                                        at_get(&svcpt->scp_at_estimate));
 322
 323         if (req->rq_reqmsg &&
 324             !(lustre_msghdr_get_flags(req->rq_reqmsg) & MSGHDR_AT_SUPPORT)) {
 325                 CDEBUG(D_ADAPTTO, "No early reply support: flags=%#x req_flags=%#x magic=%x/%x len=%d\n",
 326                        flags, lustre_msg_get_flags(req->rq_reqmsg),
 327                        lustre_msg_get_magic(req->rq_reqmsg),
 328                        lustre_msg_get_magic(req->rq_repmsg), req->rq_replen);
 329         }
 330 }
 331
 332 /**
 333  * Send request reply from request \a req reply buffer.
 334  * \a flags defines reply types
 335  * Returns 0 on success or error code
 336  */
 337 int ptlrpc_send_reply(struct ptlrpc_request *req, int flags)
 338 {
 339         struct ptlrpc_reply_state *rs = req->rq_reply_state;
 340         struct ptlrpc_connection *conn;
 341         int rc;
 342
 343         /* We must already have a reply buffer (only ptlrpc_error() may be
 344          * called without one). The reply generated by sptlrpc layer (e.g.
 345          * error notify, etc.) might have NULL rq->reqmsg; Otherwise we must
 346          * have a request buffer which is either the actual (swabbed) incoming
 347          * request, or a saved copy if this is a req saved in
 348          * target_queue_final_reply().
 349          */
 350         LASSERT(req->rq_no_reply == 0);
 351         LASSERT(req->rq_reqbuf);
 352         LASSERT(rs);
 353         LASSERT((flags & PTLRPC_REPLY_MAYBE_DIFFICULT) || !rs->rs_difficult);
 354         LASSERT(req->rq_repmsg);
 355         LASSERT(req->rq_repmsg == rs->rs_msg);
 356         LASSERT(rs->rs_cb_id.cbid_fn == reply_out_callback);
 357         LASSERT(rs->rs_cb_id.cbid_arg == rs);
 358
 359         /* There may be no rq_export during failover */
 360
 361         if (unlikely(req->rq_export && req->rq_export->exp_obd &&
 362                      req->rq_export->exp_obd->obd_fail)) {
 363                 /* Failed obd's only send ENODEV */
 364                 req->rq_type = PTL_RPC_MSG_ERR;
 365                 req->rq_status = -ENODEV;
 366                 CDEBUG(D_HA, "sending ENODEV from failed obd %d\n",
 367                        req->rq_export->exp_obd->obd_minor);
 368         }
 369
 370         /* In order to keep interoperability with the client (< 2.3) which
 371          * doesn't have pb_jobid in ptlrpc_body, We have to shrink the
 372          * ptlrpc_body in reply buffer to ptlrpc_body_v2, otherwise, the
 373          * reply buffer on client will be overflow.
 374          *
 375          * XXX Remove this whenever we drop the interoperability with
 376          * such client.
 377          */
 378         req->rq_replen = lustre_shrink_msg(req->rq_repmsg, 0,
 379                                            sizeof(struct ptlrpc_body_v2), 1);
 380
 381         if (req->rq_type != PTL_RPC_MSG_ERR)
 382                 req->rq_type = PTL_RPC_MSG_REPLY;
 383
 384         lustre_msg_set_type(req->rq_repmsg, req->rq_type);
 385         lustre_msg_set_status(req->rq_repmsg,
 386                               ptlrpc_status_hton(req->rq_status));
 387         lustre_msg_set_opc(req->rq_repmsg,
 388                            req->rq_reqmsg ?
 389                            lustre_msg_get_opc(req->rq_reqmsg) : 0);
 390
 391         target_pack_pool_reply(req);
 392
 393         ptlrpc_at_set_reply(req, flags);
 394
 395         if (!req->rq_export || !req->rq_export->exp_connection)
 396                 conn = ptlrpc_connection_get(req->rq_peer, req->rq_self, NULL);
 397         else
 398                 conn = ptlrpc_connection_addref(req->rq_export->exp_connection);
 399
 400         if (unlikely(!conn)) {
 401                 CERROR("not replying on NULL connection\n"); /* bug 9635 */
 402                 return -ENOTCONN;
 403         }
 404         ptlrpc_rs_addref(rs);              /* +1 ref for the network */
 405
 406         rc = sptlrpc_svc_wrap_reply(req);
 407         if (unlikely(rc))
 408                 goto out;
 409
 410         req->rq_sent = ktime_get_real_seconds();
 411
 412         rc = ptl_send_buf(&rs->rs_md_h, rs->rs_repbuf, rs->rs_repdata_len,
 413                           (rs->rs_difficult && !rs->rs_no_ack) ?
 414                           LNET_ACK_REQ : LNET_NOACK_REQ,
 415                           &rs->rs_cb_id, conn,
 416                           ptlrpc_req2svc(req)->srv_rep_portal,
 417                           req->rq_xid, req->rq_reply_off);
 418 out:
 419         if (unlikely(rc != 0))
 420                 ptlrpc_req_drop_rs(req);
 421         ptlrpc_connection_put(conn);
 422         return rc;
 423 }
 424
 425 int ptlrpc_reply(struct ptlrpc_request *req)
 426 {
 427         if (req->rq_no_reply)
 428                 return 0;
 429         return ptlrpc_send_reply(req, 0);
 430 }
 431
 432 /**
 433  * For request \a req send an error reply back. Create empty
 434  * reply buffers if necessary.
 435  */
 436 int ptlrpc_send_error(struct ptlrpc_request *req, int may_be_difficult)
 437 {
 438         int rc;
 439
 440         if (req->rq_no_reply)
 441                 return 0;
 442
 443         if (!req->rq_repmsg) {
 444                 rc = lustre_pack_reply(req, 1, NULL, NULL);
 445                 if (rc)
 446                         return rc;
 447         }
 448
 449         if (req->rq_status != -ENOSPC && req->rq_status != -EACCES &&
 450             req->rq_status != -EPERM && req->rq_status != -ENOENT &&
 451             req->rq_status != -EINPROGRESS && req->rq_status != -EDQUOT)
 452                 req->rq_type = PTL_RPC_MSG_ERR;
 453
 454         rc = ptlrpc_send_reply(req, may_be_difficult);
 455         return rc;
 456 }
 457
 458 int ptlrpc_error(struct ptlrpc_request *req)
 459 {
 460         return ptlrpc_send_error(req, 0);
 461 }
 462
 463 /**
 464  * Send request \a request.
 465  * if \a noreply is set, don't expect any reply back and don't set up
 466  * reply buffers.
 467  * Returns 0 on success or error code.
 468  */
 469 int ptl_send_rpc(struct ptlrpc_request *request, int noreply)
 470 {
 471         int rc;
 472         int rc2;
 473         int mpflag = 0;
 474         struct ptlrpc_connection *connection;
 475         struct lnet_handle_me reply_me_h;
 476         struct lnet_md reply_md;
 477         struct obd_import *imp = request->rq_import;
 478         struct obd_device *obd = imp->imp_obd;
 479
 480         if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_DROP_RPC))
 481                 return 0;
 482
 483         LASSERT(request->rq_type == PTL_RPC_MSG_REQUEST);
 484         LASSERT(request->rq_wait_ctx == 0);
 485
 486         /* If this is a re-transmit, we're required to have disengaged
 487          * cleanly from the previous attempt
 488          */
 489         LASSERT(!request->rq_receiving_reply);
 490         LASSERT(!((lustre_msg_get_flags(request->rq_reqmsg) & MSG_REPLAY) &&
 491                   (imp->imp_state == LUSTRE_IMP_FULL)));
 492
 493         if (unlikely(obd && obd->obd_fail)) {
 494                 CDEBUG(D_HA, "muting rpc for failed imp obd %s\n",
 495                        obd->obd_name);
 496                 /* this prevents us from waiting in ptlrpc_queue_wait */
 497                 spin_lock(&request->rq_lock);
 498                 request->rq_err = 1;
 499                 spin_unlock(&request->rq_lock);
 500                 request->rq_status = -ENODEV;
 501                 return -ENODEV;
 502         }
 503
 504         connection = imp->imp_connection;
 505
 506         lustre_msg_set_handle(request->rq_reqmsg,
 507                               &imp->imp_remote_handle);
 508         lustre_msg_set_type(request->rq_reqmsg, PTL_RPC_MSG_REQUEST);
 509         lustre_msg_set_conn_cnt(request->rq_reqmsg, imp->imp_conn_cnt);
 510         lustre_msghdr_set_flags(request->rq_reqmsg, imp->imp_msghdr_flags);
 511
 512         /*
 513          * If it's the first time to resend the request for EINPROGRESS,
 514          * we need to allocate a new XID (see after_reply()), it's different
 515          * from the resend for reply timeout.
 516          */
 517         if (request->rq_nr_resend && list_empty(&request->rq_unreplied_list)) {
 518                 __u64 min_xid = 0;
 519                 /*
 520                  * resend for EINPROGRESS, allocate new xid to avoid reply
 521                  * reconstruction
 522                  */
 523                 spin_lock(&imp->imp_lock);
 524                 ptlrpc_assign_next_xid_nolock(request);
 525                 min_xid = ptlrpc_known_replied_xid(imp);
 526                 spin_unlock(&imp->imp_lock);
 527
 528                 lustre_msg_set_last_xid(request->rq_reqmsg, min_xid);
 529                 DEBUG_REQ(D_RPCTRACE, request, "Allocating new xid for resend on EINPROGRESS");
 530         }
 531
 532         if (request->rq_bulk) {
 533                 ptlrpc_set_bulk_mbits(request);
 534                 lustre_msg_set_mbits(request->rq_reqmsg, request->rq_mbits);
 535         }
 536
 537         if (list_empty(&request->rq_unreplied_list) ||
 538             request->rq_xid <= imp->imp_known_replied_xid) {
 539                 DEBUG_REQ(D_ERROR, request,
 540                           "xid: %llu, replied: %llu, list_empty:%d\n",
 541                           request->rq_xid, imp->imp_known_replied_xid,
 542                           list_empty(&request->rq_unreplied_list));
 543                 LBUG();
 544         }
 545
 546         /**
 547          * For enabled AT all request should have AT_SUPPORT in the
 548          * FULL import state when OBD_CONNECT_AT is set
 549          */
 550         LASSERT(AT_OFF || imp->imp_state != LUSTRE_IMP_FULL ||
 551                 (imp->imp_msghdr_flags & MSGHDR_AT_SUPPORT) ||
 552                 !(imp->imp_connect_data.ocd_connect_flags &
 553                 OBD_CONNECT_AT));
 554
 555         if (request->rq_resend)
 556                 lustre_msg_add_flags(request->rq_reqmsg, MSG_RESENT);
 557
 558         if (request->rq_memalloc)
 559                 mpflag = cfs_memory_pressure_get_and_set();
 560
 561         rc = sptlrpc_cli_wrap_request(request);
 562         if (rc) {
 563                 /*
 564                  * set rq_sent so that this request is treated
 565                  * as a delayed send in the upper layers
 566                  */
 567                 if (rc == -ENOMEM)
 568                         request->rq_sent = ktime_get_seconds();
 569                 goto out;
 570         }
 571
 572         /* bulk register should be done after wrap_request() */
 573         if (request->rq_bulk) {
 574                 rc = ptlrpc_register_bulk(request);
 575                 if (rc != 0)
 576                         goto out;
 577         }
 578
 579         if (!noreply) {
 580                 LASSERT(request->rq_replen != 0);
 581                 if (!request->rq_repbuf) {
 582                         LASSERT(!request->rq_repdata);
 583                         LASSERT(!request->rq_repmsg);
 584                         rc = sptlrpc_cli_alloc_repbuf(request,
 585                                                       request->rq_replen);
 586                         if (rc) {
 587                                 /* this prevents us from looping in
 588                                  * ptlrpc_queue_wait
 589                                  */
 590                                 spin_lock(&request->rq_lock);
 591                                 request->rq_err = 1;
 592                                 spin_unlock(&request->rq_lock);
 593                                 request->rq_status = rc;
 594                                 goto cleanup_bulk;
 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                         rc = -ENOMEM;
 608                         goto cleanup_bulk;
 609                 }
 610         }
 611
 612         spin_lock(&request->rq_lock);
 613         /* We are responsible for unlinking the reply buffer */
 614         request->rq_reply_unlinked = noreply;
 615         request->rq_receiving_reply = !noreply;
 616         /* Clear any flags that may be present from previous sends. */
 617         request->rq_req_unlinked = 0;
 618         request->rq_replied = 0;
 619         request->rq_err = 0;
 620         request->rq_timedout = 0;
 621         request->rq_net_err = 0;
 622         request->rq_resend = 0;
 623         request->rq_restart = 0;
 624         request->rq_reply_truncated = 0;
 625         spin_unlock(&request->rq_lock);
 626
 627         if (!noreply) {
 628                 reply_md.start = request->rq_repbuf;
 629                 reply_md.length = request->rq_repbuf_len;
 630                 /* Allow multiple early replies */
 631                 reply_md.threshold = LNET_MD_THRESH_INF;
 632                 /* Manage remote for early replies */
 633                 reply_md.options = PTLRPC_MD_OPTIONS | LNET_MD_OP_PUT |
 634                         LNET_MD_MANAGE_REMOTE |
 635                         LNET_MD_TRUNCATE; /* allow to make EOVERFLOW error */
 636                 reply_md.user_ptr = &request->rq_reply_cbid;
 637                 reply_md.eq_handle = ptlrpc_eq_h;
 638
 639                 /* We must see the unlink callback to set rq_reply_unlinked,
 640                  * so we can't auto-unlink
 641                  */
 642                 rc = LNetMDAttach(reply_me_h, reply_md, LNET_RETAIN,
 643                                   &request->rq_reply_md_h);
 644                 if (rc != 0) {
 645                         CERROR("LNetMDAttach failed: %d\n", rc);
 646                         LASSERT(rc == -ENOMEM);
 647                         spin_lock(&request->rq_lock);
 648                         /* ...but the MD attach didn't succeed... */
 649                         request->rq_receiving_reply = 0;
 650                         spin_unlock(&request->rq_lock);
 651                         rc = -ENOMEM;
 652                         goto cleanup_me;
 653                 }
 654
 655                 CDEBUG(D_NET, "Setup reply buffer: %u bytes, xid %llu, portal %u\n",
 656                        request->rq_repbuf_len, request->rq_xid,
 657                        request->rq_reply_portal);
 658         }
 659
 660         /* add references on request for request_out_callback */
 661         ptlrpc_request_addref(request);
 662         if (obd && obd->obd_svc_stats)
 663                 lprocfs_counter_add(obd->obd_svc_stats, PTLRPC_REQACTIVE_CNTR,
 664                         atomic_read(&imp->imp_inflight));
 665
 666         OBD_FAIL_TIMEOUT(OBD_FAIL_PTLRPC_DELAY_SEND, request->rq_timeout + 5);
 667
 668         ktime_get_real_ts64(&request->rq_sent_tv);
 669         request->rq_sent = ktime_get_real_seconds();
 670         /* We give the server rq_timeout secs to process the req, and
 671          * add the network latency for our local timeout.
 672          */
 673         request->rq_deadline = request->rq_sent + request->rq_timeout +
 674                 ptlrpc_at_get_net_latency(request);
 675
 676         ptlrpc_pinger_sending_on_import(imp);
 677
 678         DEBUG_REQ(D_INFO, request, "send flg=%x",
 679                   lustre_msg_get_flags(request->rq_reqmsg));
 680         rc = ptl_send_buf(&request->rq_req_md_h,
 681                           request->rq_reqbuf, request->rq_reqdata_len,
 682                           LNET_NOACK_REQ, &request->rq_req_cbid,
 683                           connection,
 684                           request->rq_request_portal,
 685                           request->rq_xid, 0);
 686         if (likely(rc == 0))
 687                 goto out;
 688
 689         request->rq_req_unlinked = 1;
 690         ptlrpc_req_finished(request);
 691         if (noreply)
 692                 goto out;
 693
 694  cleanup_me:
 695         /* MEUnlink is safe; the PUT didn't even get off the ground, and
 696          * nobody apart from the PUT's target has the right nid+XID to
 697          * access the reply buffer.
 698          */
 699         rc2 = LNetMEUnlink(reply_me_h);
 700         LASSERT(rc2 == 0);
 701         /* UNLINKED callback called synchronously */
 702         LASSERT(!request->rq_receiving_reply);
 703
 704  cleanup_bulk:
 705         /* We do sync unlink here as there was no real transfer here so
 706          * the chance to have long unlink to sluggish net is smaller here.
 707          */
 708         ptlrpc_unregister_bulk(request, 0);
 709  out:
 710         if (request->rq_memalloc)
 711                 cfs_memory_pressure_restore(mpflag);
 712         return rc;
 713 }
 714 EXPORT_SYMBOL(ptl_send_rpc);
 715
 716 /**
 717  * Register request buffer descriptor for request receiving.
 718  */
 719 int ptlrpc_register_rqbd(struct ptlrpc_request_buffer_desc *rqbd)
 720 {
 721         struct ptlrpc_service *service = rqbd->rqbd_svcpt->scp_service;
 722         static struct lnet_process_id match_id = {LNET_NID_ANY, LNET_PID_ANY};
 723         int rc;
 724         struct lnet_md md;
 725         struct lnet_handle_me me_h;
 726
 727         CDEBUG(D_NET, "LNetMEAttach: portal %d\n",
 728                service->srv_req_portal);
 729
 730         if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_RQBD))
 731                 return -ENOMEM;
 732
 733         /* NB: CPT affinity service should use new LNet flag LNET_INS_LOCAL,
 734          * which means buffer can only be attached on local CPT, and LND
 735          * threads can find it by grabbing a local lock
 736          */
 737         rc = LNetMEAttach(service->srv_req_portal,
 738                           match_id, 0, ~0, LNET_UNLINK,
 739                           rqbd->rqbd_svcpt->scp_cpt >= 0 ?
 740                           LNET_INS_LOCAL : LNET_INS_AFTER, &me_h);
 741         if (rc != 0) {
 742                 CERROR("LNetMEAttach failed: %d\n", rc);
 743                 return -ENOMEM;
 744         }
 745
 746         LASSERT(rqbd->rqbd_refcount == 0);
 747         rqbd->rqbd_refcount = 1;
 748
 749         md.start = rqbd->rqbd_buffer;
 750         md.length = service->srv_buf_size;
 751         md.max_size = service->srv_max_req_size;
 752         md.threshold = LNET_MD_THRESH_INF;
 753         md.options = PTLRPC_MD_OPTIONS | LNET_MD_OP_PUT | LNET_MD_MAX_SIZE;
 754         md.user_ptr = &rqbd->rqbd_cbid;
 755         md.eq_handle = ptlrpc_eq_h;
 756
 757         rc = LNetMDAttach(me_h, md, LNET_UNLINK, &rqbd->rqbd_md_h);
 758         if (rc == 0)
 759                 return 0;
 760
 761         CERROR("LNetMDAttach failed: %d;\n", rc);
 762         LASSERT(rc == -ENOMEM);
 763         rc = LNetMEUnlink(me_h);
 764         LASSERT(rc == 0);
 765         rqbd->rqbd_refcount = 0;
 766
 767         return -ENOMEM;
 768 }