lustre/ptlrpc/niobuf.c

   1 /* -*- mode: c; c-basic-offset: 8; indent-tabs-mode: nil; -*-
   2  * vim:expandtab:shiftwidth=8:tabstop=8:
   3  *
   4  *  Copyright (c) 2002, 2003 Cluster File Systems, Inc.
   5  *
   6  *   This file is part of Lustre, http://www.lustre.org.
   7  *
   8  *   Lustre is free software; you can redistribute it and/or
   9  *   modify it under the terms of version 2 of the GNU General Public
  10  *   License as published by the Free Software Foundation.
  11  *
  12  *   Lustre is distributed in the hope that it will be useful,
  13  *   but WITHOUT ANY WARRANTY; without even the implied warranty of
  14  *   MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
  15  *   GNU General Public License for more details.
  16  *
  17  *   You should have received a copy of the GNU General Public License
  18  *   along with Lustre; if not, write to the Free Software
  19  *   Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA.
  20  *
  21  */
  22
  23 #define DEBUG_SUBSYSTEM S_RPC
  24 #ifndef __KERNEL__
  25 #include <liblustre.h>
  26 #endif
  27 #include <linux/obd_class.h>
  28 #include <linux/obd_support.h>
  29 #include <linux/lustre_net.h>
  30 #include <linux/lustre_lib.h>
  31 #include <linux/obd.h>
  32 #include <linux/lustre_sec.h>
  33 #include "ptlrpc_internal.h"
  34
  35 static int ptl_send_buf (ptl_handle_md_t *mdh, void *base, int len,
  36                          ptl_ack_req_t ack, struct ptlrpc_cb_id *cbid,
  37                          struct ptlrpc_connection *conn, int portal, __u64 xid)
  38 {
  39         int              rc;
  40         ptl_md_t         md;
  41         char str[PTL_NALFMT_SIZE];
  42         ENTRY;
  43
  44         LASSERT (portal != 0);
  45         LASSERT (conn != NULL);
  46         CDEBUG (D_INFO, "conn=%p ni %s id %s on %s\n",
  47                 conn, conn->c_peer.peer_ni->pni_name,
  48                 ptlrpc_id2str(&conn->c_peer, str),
  49                 conn->c_peer.peer_ni->pni_name);
  50         md.start     = base;
  51         md.length    = len;
  52         md.threshold = (ack == PTL_ACK_REQ) ? 2 : 1;
  53         md.options   = PTLRPC_MD_OPTIONS;
  54         md.user_ptr  = cbid;
  55         md.eq_handle    = conn->c_peer.peer_ni->pni_eq_h;
  56
  57         if (ack == PTL_ACK_REQ &&
  58             OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_ACK | OBD_FAIL_ONCE)) {
  59                 /* don't ask for the ack to simulate failing client */
  60                 ack = PTL_NOACK_REQ;
  61                 obd_fail_loc |= OBD_FAIL_ONCE | OBD_FAILED;
  62         }
  63
  64         rc = PtlMDBind (conn->c_peer.peer_ni->pni_ni_h, md,
  65                         PTL_UNLINK, mdh);
  66         if (rc != PTL_OK) {
  67                 CERROR ("PtlMDBind failed: %d\n", rc);
  68                 LASSERT (rc == PTL_NO_SPACE);
  69                 RETURN (-ENOMEM);
  70         }
  71
  72         CDEBUG(D_NET, "Sending %d bytes to portal %d, xid "LPD64"\n",
  73                len, portal, xid);
  74
  75         rc = PtlPut (*mdh, ack, conn->c_peer.peer_id, portal, 0, xid, 0, 0);
  76         if (rc != PTL_OK) {
  77                 int rc2;
  78                 /* We're going to get an UNLINK event when I unlink below,
  79                  * which will complete just like any other failed send, so
  80                  * I fall through and return success here! */
  81                 CERROR("PtlPut(%s, %d, "LPD64") failed: %d\n",
  82                        ptlrpc_id2str(&conn->c_peer, str),
  83                        portal, xid, rc);
  84                 rc2 = PtlMDUnlink(*mdh);
  85                 LASSERTF(rc2 == PTL_OK, "rc2 = %d\n", rc2);
  86         }
  87
  88         RETURN (0);
  89 }
  90
  91 int ptlrpc_start_bulk_transfer (struct ptlrpc_bulk_desc *desc)
  92 {
  93         int                 rc;
  94         int                 rc2;
  95         struct ptlrpc_peer *peer;
  96         ptl_md_t            md;
  97         __u64               xid;
  98         char                str[PTL_NALFMT_SIZE];
  99         ENTRY;
 100
 101         if (OBD_FAIL_CHECK_ONCE(OBD_FAIL_PTLRPC_BULK_PUT_NET))
 102                 RETURN(0);
 103
 104         /* NB no locking required until desc is on the network */
 105         LASSERT (!desc->bd_network_rw);
 106         LASSERT (desc->bd_type == BULK_PUT_SOURCE ||
 107                  desc->bd_type == BULK_GET_SINK);
 108         desc->bd_success = 0;
 109         peer = &desc->bd_export->exp_connection->c_peer;
 110
 111         md.user_ptr = &desc->bd_cbid;
 112         md.eq_handle = peer->peer_ni->pni_eq_h;
 113         md.threshold = 2; /* SENT and ACK/REPLY */
 114         md.options = PTLRPC_MD_OPTIONS;
 115         ptlrpc_fill_bulk_md(&md, desc);
 116
 117         LASSERT (desc->bd_cbid.cbid_fn == server_bulk_callback);
 118         LASSERT (desc->bd_cbid.cbid_arg == desc);
 119
 120         /* NB total length may be 0 for a read past EOF, so we send a 0
 121          * length bulk, since the client expects a bulk event. */
 122
 123         rc = PtlMDBind(peer->peer_ni->pni_ni_h, md,
 124                        PTL_UNLINK, &desc->bd_md_h);
 125         if (rc != PTL_OK) {
 126                 CERROR("PtlMDBind failed: %d\n", rc);
 127                 LASSERT (rc == PTL_NO_SPACE);
 128                 RETURN(-ENOMEM);
 129         }
 130
 131         /* Client's bulk and reply matchbits are the same */
 132         xid = desc->bd_req->rq_xid;
 133         CDEBUG(D_NET, "Transferring %u pages %u bytes via portal %d on %s "
 134                "nid %s pid %d xid "LPX64"\n", desc->bd_iov_count,
 135                desc->bd_nob, desc->bd_portal, peer->peer_ni->pni_name,
 136                ptlrpc_id2str(peer, str), peer->peer_id.pid, xid);
 137
 138         /* Network is about to get at the memory */
 139         desc->bd_network_rw = 1;
 140
 141         if (desc->bd_type == BULK_PUT_SOURCE)
 142                 rc = PtlPut (desc->bd_md_h, PTL_ACK_REQ, peer->peer_id,
 143                              desc->bd_portal, 0, xid, 0, 0);
 144         else
 145                 rc = PtlGet (desc->bd_md_h, peer->peer_id,
 146                              desc->bd_portal, 0, xid, 0);
 147
 148         if (rc != PTL_OK) {
 149                 /* Can't send, so we unlink the MD bound above.  The UNLINK
 150                  * event this creates will signal completion with failure,
 151                  * so we return SUCCESS here! */
 152                 CERROR("Transfer(%s, %d, "LPX64") failed: %d\n",
 153                        ptlrpc_id2str(peer, str),
 154                        desc->bd_portal, xid, rc);
 155                 rc2 = PtlMDUnlink(desc->bd_md_h);
 156                 LASSERT (rc2 == PTL_OK);
 157         }
 158
 159         RETURN(0);
 160 }
 161
 162 void ptlrpc_abort_bulk (struct ptlrpc_bulk_desc *desc)
 163 {
 164         /* Server side bulk abort. Idempotent. Not thread-safe (i.e. only
 165          * serialises with completion callback) */
 166         struct l_wait_info lwi;
 167         int                rc;
 168
 169         LASSERT (!in_interrupt ());             /* might sleep */
 170
 171         if (!ptlrpc_bulk_active(desc))          /* completed or */
 172                 return;                         /* never started */
 173
 174         /* The unlink ensures the callback happens ASAP and is the last
 175          * one.  If it fails, it must be because completion just happened,
 176          * but we must still l_wait_event() in this case, to give liblustre
 177          * a chance to run server_bulk_callback()*/
 178
 179         PtlMDUnlink (desc->bd_md_h);
 180
 181         for (;;) {
 182                 /* Network access will complete in finite time but the HUGE
 183                  * timeout lets us CWARN for visibility of sluggish NALs */
 184                 lwi = LWI_TIMEOUT (300 * HZ, NULL, NULL);
 185                 rc = l_wait_event(desc->bd_waitq,
 186                                   !ptlrpc_bulk_active(desc), &lwi);
 187                 if (rc == 0)
 188                         return;
 189
 190                 LASSERT(rc == -ETIMEDOUT);
 191                 CWARN("Unexpectedly long timeout: desc %p\n", desc);
 192         }
 193 }
 194
 195 int ptlrpc_register_bulk (struct ptlrpc_request *req)
 196 {
 197         struct ptlrpc_bulk_desc *desc = req->rq_bulk;
 198         struct ptlrpc_peer *peer;
 199         int rc;
 200         int rc2;
 201         ptl_handle_me_t  me_h;
 202         ptl_md_t         md;
 203         ENTRY;
 204
 205         if (OBD_FAIL_CHECK_ONCE(OBD_FAIL_PTLRPC_BULK_GET_NET))
 206                 RETURN(0);
 207
 208         /* NB no locking required until desc is on the network */
 209         LASSERT (desc->bd_nob > 0);
 210         LASSERT (!desc->bd_network_rw);
 211         LASSERT (desc->bd_iov_count <= PTLRPC_MAX_BRW_PAGES);
 212         LASSERT (desc->bd_req != NULL);
 213         LASSERT (desc->bd_type == BULK_PUT_SINK ||
 214                  desc->bd_type == BULK_GET_SOURCE);
 215
 216         desc->bd_success = 0;
 217
 218         peer = &desc->bd_import->imp_connection->c_peer;
 219
 220         md.user_ptr = &desc->bd_cbid;
 221         md.eq_handle = peer->peer_ni->pni_eq_h;
 222         md.threshold = 1;                       /* PUT or GET */
 223         md.options = PTLRPC_MD_OPTIONS |
 224                      ((desc->bd_type == BULK_GET_SOURCE) ?
 225                       PTL_MD_OP_GET : PTL_MD_OP_PUT);
 226         ptlrpc_fill_bulk_md(&md, desc);
 227
 228         LASSERT (desc->bd_cbid.cbid_fn == client_bulk_callback);
 229         LASSERT (desc->bd_cbid.cbid_arg == desc);
 230
 231         /* XXX Registering the same xid on retried bulk makes my head
 232          * explode trying to understand how the original request's bulk
 233          * might interfere with the retried request -eeb */
 234         LASSERT (!desc->bd_registered || req->rq_xid != desc->bd_last_xid);
 235         desc->bd_registered = 1;
 236         desc->bd_last_xid = req->rq_xid;
 237
 238         rc = PtlMEAttach(peer->peer_ni->pni_ni_h,
 239                          desc->bd_portal, desc->bd_import->imp_connection->c_peer.peer_id,
 240                          req->rq_xid, 0, PTL_UNLINK, PTL_INS_AFTER, &me_h);
 241         if (rc != PTL_OK) {
 242                 CERROR("PtlMEAttach failed: %d\n", rc);
 243                 LASSERT (rc == PTL_NO_SPACE);
 244                 RETURN (-ENOMEM);
 245         }
 246
 247         /* About to let the network at it... */
 248         desc->bd_network_rw = 1;
 249         rc = PtlMDAttach(me_h, md, PTL_UNLINK, &desc->bd_md_h);
 250         if (rc != PTL_OK) {
 251                 CERROR("PtlMDAttach failed: %d\n", rc);
 252                 LASSERT (rc == PTL_NO_SPACE);
 253                 desc->bd_network_rw = 0;
 254                 rc2 = PtlMEUnlink (me_h);
 255                 LASSERT (rc2 == PTL_OK);
 256                 RETURN (-ENOMEM);
 257         }
 258
 259         CDEBUG(D_NET, "Setup bulk %s buffers: %u pages %u bytes, xid "LPX64", "
 260                "portal %u on %s\n",
 261                desc->bd_type == BULK_GET_SOURCE ? "get-source" : "put-sink",
 262                desc->bd_iov_count, desc->bd_nob,
 263                req->rq_xid, desc->bd_portal, peer->peer_ni->pni_name);
 264         RETURN(0);
 265 }
 266
 267 void ptlrpc_unregister_bulk (struct ptlrpc_request *req)
 268 {
 269         /* Disconnect a bulk desc from the network. Idempotent. Not
 270          * thread-safe (i.e. only interlocks with completion callback). */
 271         struct ptlrpc_bulk_desc *desc = req->rq_bulk;
 272         wait_queue_head_t       *wq;
 273         struct l_wait_info       lwi;
 274         int                      rc;
 275
 276         LASSERT (!in_interrupt ());             /* might sleep */
 277
 278         if (!ptlrpc_bulk_active(desc))          /* completed or */
 279                 return;                         /* never registered */
 280
 281         LASSERT (desc->bd_req == req);          /* bd_req NULL until registered */
 282
 283         /* the unlink ensures the callback happens ASAP and is the last
 284          * one.  If it fails, it must be because completion just happened,
 285          * but we must still l_wait_event() in this case to give liblustre
 286          * a chance to run client_bulk_callback() */
 287
 288         PtlMDUnlink (desc->bd_md_h);
 289
 290         if (req->rq_set != NULL)
 291                 wq = &req->rq_set->set_waitq;
 292         else
 293                 wq = &req->rq_reply_waitq;
 294
 295         for (;;) {
 296                 /* Network access will complete in finite time but the HUGE
 297                  * timeout lets us CWARN for visibility of sluggish NALs */
 298                 lwi = LWI_TIMEOUT (300 * HZ, NULL, NULL);
 299                 rc = l_wait_event(*wq, !ptlrpc_bulk_active(desc), &lwi);
 300                 if (rc == 0)
 301                         return;
 302
 303                 LASSERT (rc == -ETIMEDOUT);
 304                 CWARN("Unexpectedly long timeout: desc %p\n", desc);
 305         }
 306 }
 307
 308 int ptlrpc_send_reply (struct ptlrpc_request *req, int may_be_difficult)
 309 {
 310         struct ptlrpc_service     *svc = req->rq_rqbd->rqbd_srv_ni->sni_service;
 311         struct ptlrpc_reply_state *rs = req->rq_reply_state;
 312         struct ptlrpc_connection  *conn;
 313         int                        rc;
 314
 315         /* We must already have a reply buffer (only ptlrpc_error() may be
 316          * called without one).  We usually also have a request buffer which
 317          * is either the actual (swabbed) incoming request, or a saved copy
 318          * if this is a req saved in target_queue_final_reply(). but this
 319          * will not be true since some security handling may skip the reqmsg
 320          * setting and prepare reply under normal ptlrpc layer */
 321         LASSERT (rs != NULL);
 322         LASSERT (req->rq_repmsg != NULL);
 323         LASSERT (may_be_difficult || !rs->rs_difficult);
 324         LASSERT (req->rq_repmsg == rs->rs_msg);
 325         LASSERT (rs->rs_cb_id.cbid_fn == reply_out_callback);
 326         LASSERT (rs->rs_cb_id.cbid_arg == rs);
 327
 328         LASSERT (req->rq_repmsg != NULL);
 329         if (req->rq_type != PTL_RPC_MSG_ERR)
 330                 req->rq_type = PTL_RPC_MSG_REPLY;
 331
 332         req->rq_repmsg->type   = req->rq_type;
 333         req->rq_repmsg->status = req->rq_status;
 334         req->rq_repmsg->opc    = req->rq_reqmsg ? req->rq_reqmsg->opc : 0;
 335
 336         if (req->rq_export == NULL)
 337                 conn = ptlrpc_get_connection(&req->rq_peer, NULL);
 338         else
 339                 conn = ptlrpc_connection_addref(req->rq_export->exp_connection);
 340
 341         atomic_inc (&svc->srv_outstanding_replies);
 342
 343         rc = svcsec_authorize(req);
 344         if (rc) {
 345                 CERROR("Error wrap reply message "LPX64"\n", req->rq_xid);
 346                 goto out;
 347         }
 348
 349         rc = ptl_send_buf (&rs->rs_md_h, rs->rs_repbuf, rs->rs_repdata_len,
 350                            rs->rs_difficult ? PTL_ACK_REQ : PTL_NOACK_REQ,
 351                            &rs->rs_cb_id, conn,
 352                            svc->srv_rep_portal, req->rq_xid);
 353 out:
 354         if (rc != 0) {
 355                 atomic_dec (&svc->srv_outstanding_replies);
 356
 357                 if (!rs->rs_difficult) {
 358                         /* Callers other than target_send_reply() expect me
 359                          * to clean up on a comms error */
 360                         lustre_free_reply_state (rs);
 361                         req->rq_reply_state = NULL;
 362                         req->rq_repmsg = NULL;
 363                 }
 364         }
 365         ptlrpc_put_connection(conn);
 366         return rc;
 367 }
 368
 369 int ptlrpc_reply (struct ptlrpc_request *req)
 370 {
 371         return (ptlrpc_send_reply (req, 0));
 372 }
 373
 374 int ptlrpc_error(struct ptlrpc_request *req)
 375 {
 376         int rc;
 377         ENTRY;
 378
 379         if (!req->rq_repmsg) {
 380                 rc = lustre_pack_reply(req, 0, NULL, NULL);
 381                 if (rc)
 382                         RETURN(rc);
 383         }
 384
 385         req->rq_type = PTL_RPC_MSG_ERR;
 386
 387         rc = ptlrpc_send_reply (req, 0);
 388         RETURN(rc);
 389 }
 390
 391 int ptl_send_rpc(struct ptlrpc_request *request)
 392 {
 393         int rc;
 394         int rc2;
 395         struct ptlrpc_connection *connection;
 396         unsigned long flags;
 397         ptl_handle_me_t  reply_me_h;
 398         ptl_md_t         reply_md;
 399         ENTRY;
 400
 401         LASSERT (request->rq_type == PTL_RPC_MSG_REQUEST);
 402
 403         /* If this is a re-transmit, we're required to have disengaged
 404          * cleanly from the previous attempt */
 405         LASSERT (!request->rq_receiving_reply);
 406
 407         connection = request->rq_import->imp_connection;
 408
 409         if (request->rq_bulk != NULL) {
 410                 rc = ptlrpc_register_bulk (request);
 411                 if (rc != 0)
 412                         RETURN(rc);
 413         }
 414
 415         request->rq_reqmsg->handle = request->rq_import->imp_remote_handle;
 416         request->rq_reqmsg->type = PTL_RPC_MSG_REQUEST;
 417         request->rq_reqmsg->conn_cnt = request->rq_import->imp_conn_cnt;
 418
 419         /* wrap_request might need to refresh gss cred, if this is called
 420          * in ptlrpcd then the whole daemon thread will be waiting on
 421          * gss negotiate rpc. FIXME
 422          */
 423         rc = ptlrpcs_cli_wrap_request(request);
 424         if (rc)
 425                 GOTO(cleanup_bulk, rc);
 426
 427         LASSERT (request->rq_replen != 0);
 428         if (request->rq_repbuf == NULL) {
 429                 rc = ptlrpcs_cli_alloc_repbuf(request, request->rq_replen);
 430                 if (rc)
 431                         GOTO(cleanup_bulk, rc);
 432         }
 433
 434         rc = PtlMEAttach(connection->c_peer.peer_ni->pni_ni_h,
 435                          request->rq_reply_portal, /* XXX FIXME bug 249 */
 436                          connection->c_peer.peer_id, request->rq_xid, 0, PTL_UNLINK,
 437                          PTL_INS_AFTER, &reply_me_h);
 438         if (rc != PTL_OK) {
 439                 CERROR("PtlMEAttach failed: %d\n", rc);
 440                 LASSERT (rc == PTL_NO_SPACE);
 441                 GOTO(cleanup_repmsg, rc = -ENOMEM);
 442         }
 443
 444         spin_lock_irqsave (&request->rq_lock, flags);
 445         /* If the MD attach succeeds, there _will_ be a reply_in callback */
 446         request->rq_receiving_reply = 1;
 447         /* Clear any flags that may be present from previous sends. */
 448         request->rq_replied = 0;
 449         request->rq_err = 0;
 450         request->rq_timedout = 0;
 451         request->rq_net_err = 0;
 452         request->rq_resend = 0;
 453         request->rq_ptlrpcs_restart = 0;
 454         request->rq_restart = 0;
 455         spin_unlock_irqrestore (&request->rq_lock, flags);
 456
 457         reply_md.start     = request->rq_repbuf;
 458         reply_md.length    = request->rq_repbuf_len;
 459         reply_md.threshold = 1;
 460         reply_md.options   = PTLRPC_MD_OPTIONS | PTL_MD_OP_PUT;
 461         reply_md.user_ptr  = &request->rq_reply_cbid;
 462         reply_md.eq_handle    = connection->c_peer.peer_ni->pni_eq_h;
 463
 464         rc = PtlMDAttach(reply_me_h, reply_md, PTL_UNLINK,
 465                          &request->rq_reply_md_h);
 466         if (rc != PTL_OK) {
 467                 CERROR("PtlMDAttach failed: %d\n", rc);
 468                 LASSERT (rc == PTL_NO_SPACE);
 469                 GOTO(cleanup_me, rc -ENOMEM);
 470         }
 471
 472         CDEBUG(D_NET, "Setup reply buffer: %u bytes, xid "LPU64
 473                ", portal %u on %s\n",
 474                request->rq_replen, request->rq_xid,
 475                request->rq_reply_portal,
 476                connection->c_peer.peer_ni->pni_name);
 477
 478         ptlrpc_request_addref(request);        /* +1 ref for the SENT callback */
 479
 480         request->rq_sent = LTIME_S(CURRENT_TIME);
 481         ptlrpc_pinger_sending_on_import(request->rq_import);
 482         rc = ptl_send_buf(&request->rq_req_md_h,
 483                           request->rq_reqbuf, request->rq_reqdata_len,
 484                           PTL_NOACK_REQ, &request->rq_req_cbid,
 485                           connection,
 486                           request->rq_request_portal,
 487                           request->rq_xid);
 488         if (rc == 0) {
 489                 ptlrpc_lprocfs_rpc_sent(request);
 490                 RETURN(rc);
 491         }
 492
 493         ptlrpc_req_finished (request);          /* drop callback ref */
 494
 495  cleanup_me:
 496         /* MEUnlink is safe; the PUT didn't even get off the ground, and
 497          * nobody apart from the PUT's target has the right nid+XID to
 498          * access the reply buffer. */
 499         rc2 = PtlMEUnlink(reply_me_h);
 500         LASSERT (rc2 == PTL_OK);
 501         /* UNLINKED callback called synchronously */
 502         LASSERT (!request->rq_receiving_reply);
 503
 504  cleanup_repmsg:
 505         ptlrpcs_cli_free_repbuf(request);
 506
 507  cleanup_bulk:
 508         if (request->rq_bulk != NULL)
 509                 ptlrpc_unregister_bulk(request);
 510
 511         return rc;
 512 }
 513
 514 int ptlrpc_register_rqbd (struct ptlrpc_request_buffer_desc *rqbd)
 515 {
 516         struct ptlrpc_srv_ni    *srv_ni = rqbd->rqbd_srv_ni;
 517         struct ptlrpc_service   *service = srv_ni->sni_service;
 518         static ptl_process_id_t  match_id = {PTL_NID_ANY, PTL_PID_ANY};
 519         int                      rc;
 520         ptl_md_t                 md;
 521         ptl_handle_me_t          me_h;
 522
 523         CDEBUG(D_NET, "PtlMEAttach: portal %d on %s\n",
 524                service->srv_req_portal, srv_ni->sni_ni->pni_name);
 525
 526         if (OBD_FAIL_CHECK_ONCE(OBD_FAIL_PTLRPC_RQBD))
 527                 return (-ENOMEM);
 528
 529         rc = PtlMEAttach(srv_ni->sni_ni->pni_ni_h, service->srv_req_portal,
 530                          match_id, 0, ~0, PTL_UNLINK, PTL_INS_AFTER, &me_h);
 531         if (rc != PTL_OK) {
 532                 CERROR("PtlMEAttach failed: %d\n", rc);
 533                 return (-ENOMEM);
 534         }
 535
 536         LASSERT(rqbd->rqbd_refcount == 0);
 537         rqbd->rqbd_refcount = 1;
 538
 539         md.start     = rqbd->rqbd_buffer;
 540         md.length    = service->srv_buf_size;
 541         md.max_size  = service->srv_max_req_size;
 542         md.threshold = PTL_MD_THRESH_INF;
 543         md.options   = PTLRPC_MD_OPTIONS | PTL_MD_OP_PUT | PTL_MD_MAX_SIZE;
 544         md.user_ptr  = &rqbd->rqbd_cbid;
 545         md.eq_handle = srv_ni->sni_ni->pni_eq_h;
 546
 547         rc = PtlMDAttach(me_h, md, PTL_UNLINK, &rqbd->rqbd_md_h);
 548         if (rc == PTL_OK)
 549                 return (0);
 550
 551         CERROR("PtlMDAttach failed: %d; \n", rc);
 552         LASSERT (rc == PTL_NO_SPACE);
 553         rc = PtlMEUnlink (me_h);
 554         LASSERT (rc == PTL_OK);
 555         rqbd->rqbd_refcount = 0;
 556
 557         return (-ENOMEM);
 558 }
 559
 560 static int rawrpc_timedout(void *data)
 561 {
 562         struct ptlrpc_request *req = (struct ptlrpc_request *) data;
 563         unsigned long flags;
 564
 565         spin_lock_irqsave(&req->rq_lock, flags);
 566         if (!req->rq_replied)
 567                 req->rq_timedout = 1;
 568         spin_unlock_irqrestore(&req->rq_lock, flags);
 569
 570         return 1;
 571 }
 572
 573 /* to make things as simple as possible */
 574 static int rawrpc_check_reply(struct ptlrpc_request *req)
 575 {
 576         unsigned long flags;
 577         int rc;
 578
 579         spin_lock_irqsave (&req->rq_lock, flags);
 580         rc = req->rq_replied || req->rq_net_err || req->rq_err ||
 581              req->rq_resend || req->rq_restart;
 582         spin_unlock_irqrestore(&req->rq_lock, flags);
 583         return rc;
 584 }
 585
 586 /*
 587  * Construct a fake ptlrpc_request to do the work, in order to
 588  * user the existing callback/wakeup facilities
 589  */
 590 int ptlrpc_do_rawrpc(struct obd_import *imp,
 591                      char *reqbuf, int reqlen,
 592                      char *repbuf, int *replenp,
 593                      int timeout)
 594 {
 595         struct ptlrpc_connection *conn;
 596         struct ptlrpc_request request; /* just a fake one */
 597         ptl_handle_me_t reply_me_h;
 598         ptl_md_t reply_md, req_md;
 599         struct l_wait_info lwi;
 600         unsigned long irq_flags;
 601         int rc;
 602         ENTRY;
 603
 604         LASSERT(imp);
 605         class_import_get(imp);
 606         if (imp->imp_state == LUSTRE_IMP_CLOSED) {
 607                 CWARN("raw rpc on closed imp(=>%s)? send anyway\n",
 608                        imp->imp_target_uuid.uuid);
 609         }
 610
 611         conn = imp->imp_connection;
 612
 613         /* initialize request */
 614         memset(&request, 0, sizeof(request));
 615         request.rq_req_cbid.cbid_fn = request_out_callback;
 616         request.rq_req_cbid.cbid_arg = &request;
 617         request.rq_reply_cbid.cbid_fn  = reply_in_callback;
 618         request.rq_reply_cbid.cbid_arg = &request;
 619         request.rq_reqbuf = reqbuf;
 620         request.rq_reqbuf_len = reqlen;
 621         request.rq_repbuf = repbuf;
 622         request.rq_repbuf_len = *replenp;
 623         request.rq_set = NULL;
 624         spin_lock_init(&request.rq_lock);
 625         init_waitqueue_head(&request.rq_reply_waitq);
 626         atomic_set(&request.rq_refcount, 1000000); /* never be droped */
 627         request.rq_xid = ptlrpc_next_xid();
 628
 629         /* add into sending list */
 630         spin_lock_irqsave(&imp->imp_lock, irq_flags);
 631         list_add_tail(&request.rq_list, &imp->imp_rawrpc_list);
 632         spin_unlock_irqrestore(&imp->imp_lock, irq_flags);
 633
 634         /* prepare reply buffer */
 635         rc = PtlMEAttach(conn->c_peer.peer_ni->pni_ni_h,
 636                          imp->imp_client->cli_reply_portal,
 637                          conn->c_peer.peer_id, request.rq_xid, 0, PTL_UNLINK,
 638                          PTL_INS_AFTER, &reply_me_h);
 639         if (rc != PTL_OK) {
 640                 CERROR("PtlMEAttach failed: %d\n", rc);
 641                 LASSERT (rc == PTL_NO_SPACE);
 642                 GOTO(cleanup_imp, rc = -ENOMEM);
 643         }
 644
 645         spin_lock_irqsave(&request.rq_lock, irq_flags);
 646         request.rq_receiving_reply = 1;
 647         spin_unlock_irqrestore(&request.rq_lock, irq_flags);
 648
 649         reply_md.start          = repbuf;
 650         reply_md.length         = *replenp;
 651         reply_md.threshold      = 1;
 652         reply_md.options        = PTLRPC_MD_OPTIONS | PTL_MD_OP_PUT;
 653         reply_md.user_ptr       = &request.rq_reply_cbid;
 654         reply_md.eq_handle      = conn->c_peer.peer_ni->pni_eq_h;
 655
 656         rc = PtlMDAttach(reply_me_h, reply_md, PTL_UNLINK,
 657                          &request.rq_reply_md_h);
 658         if (rc != PTL_OK) {
 659                 CERROR("PtlMDAttach failed: %d\n", rc);
 660                 LASSERT (rc == PTL_NO_SPACE);
 661                 GOTO(cleanup_me, rc = -ENOMEM);
 662         }
 663
 664         /* prepare request buffer */
 665         req_md.start            = reqbuf;
 666         req_md.length           = reqlen;
 667         req_md.threshold        = 1;
 668         req_md.options          = PTLRPC_MD_OPTIONS;
 669         req_md.user_ptr         = &request.rq_req_cbid;
 670         req_md.eq_handle        = conn->c_peer.peer_ni->pni_eq_h;
 671
 672         rc = PtlMDBind(conn->c_peer.peer_ni->pni_ni_h,
 673                        req_md, PTL_UNLINK, &request.rq_req_md_h);
 674         if (rc != PTL_OK) {
 675                 CERROR("PtlMDBind failed %d\n", rc);
 676                 LASSERT (rc == PTL_NO_SPACE);
 677                 GOTO(cleanup_me, rc = -ENOMEM);
 678         }
 679
 680         rc = PtlPut(request.rq_req_md_h, PTL_NOACK_REQ, conn->c_peer.peer_id,
 681                     imp->imp_client->cli_request_portal,
 682                     0, request.rq_xid, 0, 0);
 683         if (rc != PTL_OK) {
 684                 CERROR("PtlPut failed %d\n", rc);
 685                 GOTO(cleanup_md, rc);
 686         }
 687
 688         lwi = LWI_TIMEOUT(timeout * HZ, rawrpc_timedout, &request);
 689         l_wait_event(request.rq_reply_waitq,
 690                      rawrpc_check_reply(&request), &lwi);
 691
 692         ptlrpc_unregister_reply(&request);
 693
 694         if (request.rq_err || request.rq_resend || request.rq_intr ||
 695             request.rq_timedout || !request.rq_replied) {
 696                 CERROR("secinit rpc error: err %d, resend %d, "
 697                        "intr %d, timeout %d, replied %d\n",
 698                         request.rq_err, request.rq_resend, request.rq_intr,
 699                         request.rq_timedout, request.rq_replied);
 700                 rc = -EINVAL;
 701         } else {
 702                 *replenp = request.rq_nob_received;
 703                 rc = 0;
 704         }
 705         GOTO(cleanup_imp, rc);
 706
 707 cleanup_md:
 708         PtlMDUnlink(request.rq_req_md_h);
 709 cleanup_me:
 710         PtlMEUnlink(reply_me_h);
 711 cleanup_imp:
 712         spin_lock_irqsave(&imp->imp_lock, irq_flags);
 713         list_del_init(&request.rq_list);
 714         spin_unlock_irqrestore(&imp->imp_lock, irq_flags);
 715
 716         class_import_put(imp);
 717         RETURN(rc);
 718 }