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 "../include/obd_support.h"
  35 #include "../include/lustre_net.h"
  36 #include "../include/lustre_lib.h"
  37 #include "../include/obd.h"
  38 #include "../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(lnet_handle_md_t *mdh, void *base, int len,
  47                         lnet_ack_req_t ack, struct ptlrpc_cb_id *cbid,
  48                         struct ptlrpc_connection *conn, int portal, __u64 xid,
  49                         unsigned int offset)
  50 {
  51         int rc;
  52         lnet_md_t 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(lnet_handle_md_t *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         lnet_process_id_t peer;
 113         int rc = 0;
 114         int rc2;
 115         int posted_md;
 116         int total_md;
 117         __u64 xid;
 118         lnet_handle_me_t me_h;
 119         lnet_md_t 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(desc->bd_type == BULK_PUT_SINK ||
 131                 desc->bd_type == BULK_GET_SOURCE);
 132
 133         /* cleanup the state of the bulk for it will be reused */
 134         if (req->rq_resend || req->rq_send_state == LUSTRE_IMP_REPLAY)
 135                 desc->bd_nob_transferred = 0;
 136         else
 137                 LASSERT(desc->bd_nob_transferred == 0);
 138
 139         desc->bd_failure = 0;
 140
 141         peer = desc->bd_import->imp_connection->c_peer;
 142
 143         LASSERT(desc->bd_cbid.cbid_fn == client_bulk_callback);
 144         LASSERT(desc->bd_cbid.cbid_arg == desc);
 145
 146         /* An XID is only used for a single request from the client.
 147          * For retried bulk transfers, a new XID will be allocated in
 148          * in ptlrpc_check_set() if it needs to be resent, so it is not
 149          * using the same RDMA match bits after an error.
 150          *
 151          * For multi-bulk RPCs, rq_xid is the last XID needed for bulks. The
 152          * first bulk XID is power-of-two aligned before rq_xid. LU-1431
 153          */
 154         xid = req->rq_xid & ~((__u64)desc->bd_md_max_brw - 1);
 155         LASSERTF(!(desc->bd_registered &&
 156                    req->rq_send_state != LUSTRE_IMP_REPLAY) ||
 157                  xid != desc->bd_last_xid,
 158                  "registered: %d  rq_xid: %llu bd_last_xid: %llu\n",
 159                  desc->bd_registered, xid, desc->bd_last_xid);
 160
 161         total_md = (desc->bd_iov_count + LNET_MAX_IOV - 1) / LNET_MAX_IOV;
 162         desc->bd_registered = 1;
 163         desc->bd_last_xid = xid;
 164         desc->bd_md_count = total_md;
 165         md.user_ptr = &desc->bd_cbid;
 166         md.eq_handle = ptlrpc_eq_h;
 167         md.threshold = 1;                      /* PUT or GET */
 168
 169         for (posted_md = 0; posted_md < total_md; posted_md++, xid++) {
 170                 md.options = PTLRPC_MD_OPTIONS |
 171                              ((desc->bd_type == BULK_GET_SOURCE) ?
 172                               LNET_MD_OP_GET : LNET_MD_OP_PUT);
 173                 ptlrpc_fill_bulk_md(&md, desc, posted_md);
 174
 175                 rc = LNetMEAttach(desc->bd_portal, peer, xid, 0,
 176                                   LNET_UNLINK, LNET_INS_AFTER, &me_h);
 177                 if (rc != 0) {
 178                         CERROR("%s: LNetMEAttach failed x%llu/%d: rc = %d\n",
 179                                desc->bd_import->imp_obd->obd_name, xid,
 180                                posted_md, rc);
 181                         break;
 182                 }
 183
 184                 /* About to let the network at it... */
 185                 rc = LNetMDAttach(me_h, md, LNET_UNLINK,
 186                                   &desc->bd_mds[posted_md]);
 187                 if (rc != 0) {
 188                         CERROR("%s: LNetMDAttach failed x%llu/%d: rc = %d\n",
 189                                desc->bd_import->imp_obd->obd_name, xid,
 190                                posted_md, rc);
 191                         rc2 = LNetMEUnlink(me_h);
 192                         LASSERT(rc2 == 0);
 193                         break;
 194                 }
 195         }
 196
 197         if (rc != 0) {
 198                 LASSERT(rc == -ENOMEM);
 199                 spin_lock(&desc->bd_lock);
 200                 desc->bd_md_count -= total_md - posted_md;
 201                 spin_unlock(&desc->bd_lock);
 202                 LASSERT(desc->bd_md_count >= 0);
 203                 mdunlink_iterate_helper(desc->bd_mds, desc->bd_md_max_brw);
 204                 req->rq_status = -ENOMEM;
 205                 return -ENOMEM;
 206         }
 207
 208         /* Set rq_xid to matchbits of the final bulk so that server can
 209          * infer the number of bulks that were prepared
 210          */
 211         req->rq_xid = --xid;
 212         LASSERTF(desc->bd_last_xid == (req->rq_xid & PTLRPC_BULK_OPS_MASK),
 213                  "bd_last_xid = x%llu, rq_xid = x%llu\n",
 214                  desc->bd_last_xid, req->rq_xid);
 215
 216         spin_lock(&desc->bd_lock);
 217         /* Holler if peer manages to touch buffers before he knows the xid */
 218         if (desc->bd_md_count != total_md)
 219                 CWARN("%s: Peer %s touched %d buffers while I registered\n",
 220                       desc->bd_import->imp_obd->obd_name, libcfs_id2str(peer),
 221                       total_md - desc->bd_md_count);
 222         spin_unlock(&desc->bd_lock);
 223
 224         CDEBUG(D_NET, "Setup %u bulk %s buffers: %u pages %u bytes, xid x%#llx-%#llx, portal %u\n",
 225                desc->bd_md_count,
 226                desc->bd_type == BULK_GET_SOURCE ? "get-source" : "put-sink",
 227                desc->bd_iov_count, desc->bd_nob,
 228                desc->bd_last_xid, req->rq_xid, desc->bd_portal);
 229
 230         return 0;
 231 }
 232
 233 /**
 234  * Disconnect a bulk desc from the network. Idempotent. Not
 235  * thread-safe (i.e. only interlocks with completion callback).
 236  * Returns 1 on success or 0 if network unregistration failed for whatever
 237  * reason.
 238  */
 239 int ptlrpc_unregister_bulk(struct ptlrpc_request *req, int async)
 240 {
 241         struct ptlrpc_bulk_desc *desc = req->rq_bulk;
 242         wait_queue_head_t *wq;
 243         struct l_wait_info lwi;
 244         int rc;
 245
 246         LASSERT(!in_interrupt());     /* might sleep */
 247
 248         /* Let's setup deadline for reply unlink. */
 249         if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_BULK_UNLINK) &&
 250             async && req->rq_bulk_deadline == 0 && cfs_fail_val == 0)
 251                 req->rq_bulk_deadline = ktime_get_real_seconds() + LONG_UNLINK;
 252
 253         if (ptlrpc_client_bulk_active(req) == 0)        /* completed or */
 254                 return 1;                               /* never registered */
 255
 256         LASSERT(desc->bd_req == req);  /* bd_req NULL until registered */
 257
 258         /* the unlink ensures the callback happens ASAP and is the last
 259          * one.  If it fails, it must be because completion just happened,
 260          * but we must still l_wait_event() in this case to give liblustre
 261          * a chance to run client_bulk_callback()
 262          */
 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_UNREG_BULK);
 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 LNDs
 283                  */
 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
 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                  */
 314                 int oldse = at_measured(&svcpt->scp_at_estimate, service_time);
 315
 316                 if (oldse != 0) {
 317                         DEBUG_REQ(D_ADAPTTO, req,
 318                                   "svc %s changed estimate from %d to %d",
 319                                   svc->srv_name, oldse,
 320                                   at_get(&svcpt->scp_at_estimate));
 321                 }
 322         }
 323         /* Report actual service time for client latency calc */
 324         lustre_msg_set_service_time(req->rq_repmsg, service_time);
 325         /* Report service time estimate for future client reqs, but report 0
 326          * (to be ignored by client) if it's a error reply during recovery.
 327          * (bz15815)
 328          */
 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);
 364         LASSERT(rs);
 365         LASSERT((flags & PTLRPC_REPLY_MAYBE_DIFFICULT) || !rs->rs_difficult);
 366         LASSERT(req->rq_repmsg);
 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 ?
 401                            lustre_msg_get_opc(req->rq_reqmsg) : 0);
 402
 403         target_pack_pool_reply(req);
 404
 405         ptlrpc_at_set_reply(req, flags);
 406
 407         if (!req->rq_export || !req->rq_export->exp_connection)
 408                 conn = ptlrpc_connection_get(req->rq_peer, req->rq_self, NULL);
 409         else
 410                 conn = ptlrpc_connection_addref(req->rq_export->exp_connection);
 411
 412         if (unlikely(!conn)) {
 413                 CERROR("not replying on NULL connection\n"); /* bug 9635 */
 414                 return -ENOTCONN;
 415         }
 416         ptlrpc_rs_addref(rs);              /* +1 ref for the network */
 417
 418         rc = sptlrpc_svc_wrap_reply(req);
 419         if (unlikely(rc))
 420                 goto out;
 421
 422         req->rq_sent = ktime_get_real_seconds();
 423
 424         rc = ptl_send_buf(&rs->rs_md_h, rs->rs_repbuf, rs->rs_repdata_len,
 425                           (rs->rs_difficult && !rs->rs_no_ack) ?
 426                           LNET_ACK_REQ : LNET_NOACK_REQ,
 427                           &rs->rs_cb_id, conn,
 428                           ptlrpc_req2svc(req)->srv_rep_portal,
 429                           req->rq_xid, req->rq_reply_off);
 430 out:
 431         if (unlikely(rc != 0))
 432                 ptlrpc_req_drop_rs(req);
 433         ptlrpc_connection_put(conn);
 434         return rc;
 435 }
 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
 444 /**
 445  * For request \a req send an error reply back. Create empty
 446  * reply buffers if necessary.
 447  */
 448 int ptlrpc_send_error(struct ptlrpc_request *req, int may_be_difficult)
 449 {
 450         int rc;
 451
 452         if (req->rq_no_reply)
 453                 return 0;
 454
 455         if (!req->rq_repmsg) {
 456                 rc = lustre_pack_reply(req, 1, NULL, NULL);
 457                 if (rc)
 458                         return rc;
 459         }
 460
 461         if (req->rq_status != -ENOSPC && req->rq_status != -EACCES &&
 462             req->rq_status != -EPERM && req->rq_status != -ENOENT &&
 463             req->rq_status != -EINPROGRESS && req->rq_status != -EDQUOT)
 464                 req->rq_type = PTL_RPC_MSG_ERR;
 465
 466         rc = ptlrpc_send_reply(req, may_be_difficult);
 467         return rc;
 468 }
 469
 470 int ptlrpc_error(struct ptlrpc_request *req)
 471 {
 472         return ptlrpc_send_error(req, 0);
 473 }
 474
 475 /**
 476  * Send request \a request.
 477  * if \a noreply is set, don't expect any reply back and don't set up
 478  * reply buffers.
 479  * Returns 0 on success or error code.
 480  */
 481 int ptl_send_rpc(struct ptlrpc_request *request, int noreply)
 482 {
 483         int rc;
 484         int rc2;
 485         int mpflag = 0;
 486         struct ptlrpc_connection *connection;
 487         lnet_handle_me_t reply_me_h;
 488         lnet_md_t reply_md;
 489         struct obd_import *imp = request->rq_import;
 490         struct obd_device *obd = imp->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          */
 501         LASSERT(!request->rq_receiving_reply);
 502         LASSERT(!((lustre_msg_get_flags(request->rq_reqmsg) & MSG_REPLAY) &&
 503                   (imp->imp_state == LUSTRE_IMP_FULL)));
 504
 505         if (unlikely(obd && obd->obd_fail)) {
 506                 CDEBUG(D_HA, "muting rpc for failed imp obd %s\n",
 507                        obd->obd_name);
 508                 /* this prevents us from waiting in ptlrpc_queue_wait */
 509                 spin_lock(&request->rq_lock);
 510                 request->rq_err = 1;
 511                 spin_unlock(&request->rq_lock);
 512                 request->rq_status = -ENODEV;
 513                 return -ENODEV;
 514         }
 515
 516         connection = imp->imp_connection;
 517
 518         lustre_msg_set_handle(request->rq_reqmsg,
 519                               &imp->imp_remote_handle);
 520         lustre_msg_set_type(request->rq_reqmsg, PTL_RPC_MSG_REQUEST);
 521         lustre_msg_set_conn_cnt(request->rq_reqmsg, imp->imp_conn_cnt);
 522         lustre_msghdr_set_flags(request->rq_reqmsg, imp->imp_msghdr_flags);
 523
 524         /**
 525          * For enabled AT all request should have AT_SUPPORT in the
 526          * FULL import state when OBD_CONNECT_AT is set
 527          */
 528         LASSERT(AT_OFF || imp->imp_state != LUSTRE_IMP_FULL ||
 529                 (imp->imp_msghdr_flags & MSGHDR_AT_SUPPORT) ||
 530                 !(imp->imp_connect_data.ocd_connect_flags &
 531                 OBD_CONNECT_AT));
 532
 533         if (request->rq_resend)
 534                 lustre_msg_add_flags(request->rq_reqmsg, MSG_RESENT);
 535
 536         if (request->rq_memalloc)
 537                 mpflag = cfs_memory_pressure_get_and_set();
 538
 539         rc = sptlrpc_cli_wrap_request(request);
 540         if (rc)
 541                 goto out;
 542
 543         /* bulk register should be done after wrap_request() */
 544         if (request->rq_bulk) {
 545                 rc = ptlrpc_register_bulk(request);
 546                 if (rc != 0)
 547                         goto out;
 548         }
 549
 550         if (!noreply) {
 551                 LASSERT(request->rq_replen != 0);
 552                 if (!request->rq_repbuf) {
 553                         LASSERT(!request->rq_repdata);
 554                         LASSERT(!request->rq_repmsg);
 555                         rc = sptlrpc_cli_alloc_repbuf(request,
 556                                                       request->rq_replen);
 557                         if (rc) {
 558                                 /* this prevents us from looping in
 559                                  * ptlrpc_queue_wait
 560                                  */
 561                                 spin_lock(&request->rq_lock);
 562                                 request->rq_err = 1;
 563                                 spin_unlock(&request->rq_lock);
 564                                 request->rq_status = rc;
 565                                 goto cleanup_bulk;
 566                         }
 567                 } else {
 568                         request->rq_repdata = NULL;
 569                         request->rq_repmsg = NULL;
 570                 }
 571
 572                 rc = LNetMEAttach(request->rq_reply_portal,/*XXX FIXME bug 249*/
 573                                   connection->c_peer, request->rq_xid, 0,
 574                                   LNET_UNLINK, LNET_INS_AFTER, &reply_me_h);
 575                 if (rc != 0) {
 576                         CERROR("LNetMEAttach failed: %d\n", rc);
 577                         LASSERT(rc == -ENOMEM);
 578                         rc = -ENOMEM;
 579                         goto cleanup_bulk;
 580                 }
 581         }
 582
 583         spin_lock(&request->rq_lock);
 584         /* We are responsible for unlinking the reply buffer */
 585         request->rq_reply_unlinked = noreply;
 586         request->rq_receiving_reply = !noreply;
 587         /* Clear any flags that may be present from previous sends. */
 588         request->rq_req_unlinked = 0;
 589         request->rq_replied = 0;
 590         request->rq_err = 0;
 591         request->rq_timedout = 0;
 592         request->rq_net_err = 0;
 593         request->rq_resend = 0;
 594         request->rq_restart = 0;
 595         request->rq_reply_truncated = 0;
 596         spin_unlock(&request->rq_lock);
 597
 598         if (!noreply) {
 599                 reply_md.start = request->rq_repbuf;
 600                 reply_md.length = request->rq_repbuf_len;
 601                 /* Allow multiple early replies */
 602                 reply_md.threshold = LNET_MD_THRESH_INF;
 603                 /* Manage remote for early replies */
 604                 reply_md.options = PTLRPC_MD_OPTIONS | LNET_MD_OP_PUT |
 605                         LNET_MD_MANAGE_REMOTE |
 606                         LNET_MD_TRUNCATE; /* allow to make EOVERFLOW error */
 607                 reply_md.user_ptr = &request->rq_reply_cbid;
 608                 reply_md.eq_handle = ptlrpc_eq_h;
 609
 610                 /* We must see the unlink callback to set rq_reply_unlinked,
 611                  * so we can't auto-unlink
 612                  */
 613                 rc = LNetMDAttach(reply_me_h, reply_md, LNET_RETAIN,
 614                                   &request->rq_reply_md_h);
 615                 if (rc != 0) {
 616                         CERROR("LNetMDAttach failed: %d\n", rc);
 617                         LASSERT(rc == -ENOMEM);
 618                         spin_lock(&request->rq_lock);
 619                         /* ...but the MD attach didn't succeed... */
 620                         request->rq_receiving_reply = 0;
 621                         spin_unlock(&request->rq_lock);
 622                         rc = -ENOMEM;
 623                         goto cleanup_me;
 624                 }
 625
 626                 CDEBUG(D_NET, "Setup reply buffer: %u bytes, xid %llu, portal %u\n",
 627                        request->rq_repbuf_len, request->rq_xid,
 628                        request->rq_reply_portal);
 629         }
 630
 631         /* add references on request for request_out_callback */
 632         ptlrpc_request_addref(request);
 633         if (obd && obd->obd_svc_stats)
 634                 lprocfs_counter_add(obd->obd_svc_stats, PTLRPC_REQACTIVE_CNTR,
 635                         atomic_read(&imp->imp_inflight));
 636
 637         OBD_FAIL_TIMEOUT(OBD_FAIL_PTLRPC_DELAY_SEND, request->rq_timeout + 5);
 638
 639         ktime_get_real_ts64(&request->rq_sent_tv);
 640         request->rq_sent = ktime_get_real_seconds();
 641         /* We give the server rq_timeout secs to process the req, and
 642          * add the network latency for our local timeout.
 643          */
 644         request->rq_deadline = request->rq_sent + request->rq_timeout +
 645                 ptlrpc_at_get_net_latency(request);
 646
 647         ptlrpc_pinger_sending_on_import(imp);
 648
 649         DEBUG_REQ(D_INFO, request, "send flg=%x",
 650                   lustre_msg_get_flags(request->rq_reqmsg));
 651         rc = ptl_send_buf(&request->rq_req_md_h,
 652                           request->rq_reqbuf, request->rq_reqdata_len,
 653                           LNET_NOACK_REQ, &request->rq_req_cbid,
 654                           connection,
 655                           request->rq_request_portal,
 656                           request->rq_xid, 0);
 657         if (likely(rc == 0))
 658                 goto out;
 659
 660         request->rq_req_unlinked = 1;
 661         ptlrpc_req_finished(request);
 662         if (noreply)
 663                 goto out;
 664
 665  cleanup_me:
 666         /* MEUnlink is safe; the PUT didn't even get off the ground, and
 667          * nobody apart from the PUT's target has the right nid+XID to
 668          * access the reply buffer.
 669          */
 670         rc2 = LNetMEUnlink(reply_me_h);
 671         LASSERT(rc2 == 0);
 672         /* UNLINKED callback called synchronously */
 673         LASSERT(!request->rq_receiving_reply);
 674
 675  cleanup_bulk:
 676         /* We do sync unlink here as there was no real transfer here so
 677          * the chance to have long unlink to sluggish net is smaller here.
 678          */
 679         ptlrpc_unregister_bulk(request, 0);
 680  out:
 681         if (request->rq_memalloc)
 682                 cfs_memory_pressure_restore(mpflag);
 683         return rc;
 684 }
 685 EXPORT_SYMBOL(ptl_send_rpc);
 686
 687 /**
 688  * Register request buffer descriptor for request receiving.
 689  */
 690 int ptlrpc_register_rqbd(struct ptlrpc_request_buffer_desc *rqbd)
 691 {
 692         struct ptlrpc_service *service = rqbd->rqbd_svcpt->scp_service;
 693         static lnet_process_id_t match_id = {LNET_NID_ANY, LNET_PID_ANY};
 694         int rc;
 695         lnet_md_t md;
 696         lnet_handle_me_t me_h;
 697
 698         CDEBUG(D_NET, "LNetMEAttach: portal %d\n",
 699                service->srv_req_portal);
 700
 701         if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_RQBD))
 702                 return -ENOMEM;
 703
 704         /* NB: CPT affinity service should use new LNet flag LNET_INS_LOCAL,
 705          * which means buffer can only be attached on local CPT, and LND
 706          * threads can find it by grabbing a local lock
 707          */
 708         rc = LNetMEAttach(service->srv_req_portal,
 709                           match_id, 0, ~0, LNET_UNLINK,
 710                           rqbd->rqbd_svcpt->scp_cpt >= 0 ?
 711                           LNET_INS_LOCAL : LNET_INS_AFTER, &me_h);
 712         if (rc != 0) {
 713                 CERROR("LNetMEAttach failed: %d\n", rc);
 714                 return -ENOMEM;
 715         }
 716
 717         LASSERT(rqbd->rqbd_refcount == 0);
 718         rqbd->rqbd_refcount = 1;
 719
 720         md.start = rqbd->rqbd_buffer;
 721         md.length = service->srv_buf_size;
 722         md.max_size = service->srv_max_req_size;
 723         md.threshold = LNET_MD_THRESH_INF;
 724         md.options = PTLRPC_MD_OPTIONS | LNET_MD_OP_PUT | LNET_MD_MAX_SIZE;
 725         md.user_ptr = &rqbd->rqbd_cbid;
 726         md.eq_handle = ptlrpc_eq_h;
 727
 728         rc = LNetMDAttach(me_h, md, LNET_UNLINK, &rqbd->rqbd_md_h);
 729         if (rc == 0)
 730                 return 0;
 731
 732         CERROR("LNetMDAttach failed: %d;\n", rc);
 733         LASSERT(rc == -ENOMEM);
 734         rc = LNetMEUnlink(me_h);
 735         LASSERT(rc == 0);
 736         rqbd->rqbd_refcount = 0;
 737
 738         return -ENOMEM;
 739 }