lustre/ptlrpc/niobuf.c

   1 /* -*- mode: c; c-basic-offset: 8; indent-tabs-mode: nil; -*-
   2  * vim:expandtab:shiftwidth=8:tabstop=8:
   3  *
   4  * GPL HEADER START
   5  *
   6  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
   7  *
   8  * This program is free software; you can redistribute it and/or modify
   9  * it under the terms of the GNU General Public License version 2 only,
  10  * as published by the Free Software Foundation.
  11  *
  12  * This program is distributed in the hope that it will be useful, but
  13  * WITHOUT ANY WARRANTY; without even the implied warranty of
  14  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  15  * General Public License version 2 for more details (a copy is included
  16  * in the LICENSE file that accompanied this code).
  17  *
  18  * You should have received a copy of the GNU General Public License
  19  * version 2 along with this program; If not, see
  20  * http://www.sun.com/software/products/lustre/docs/GPLv2.pdf
  21  *
  22  * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
  23  * CA 95054 USA or visit www.sun.com if you need additional information or
  24  * have any questions.
  25  *
  26  * GPL HEADER END
  27  */
  28 /*
  29  * Copyright (c) 2002, 2010, Oracle and/or its affiliates. All rights reserved.
  30  * Use is subject to license terms.
  31  */
  32 /*
  33  * This file is part of Lustre, http://www.lustre.org/
  34  * Lustre is a trademark of Sun Microsystems, Inc.
  35  */
  36
  37 #define DEBUG_SUBSYSTEM S_RPC
  38 #ifndef __KERNEL__
  39 #include <liblustre.h>
  40 #endif
  41 #include <obd_support.h>
  42 #include <lustre_net.h>
  43 #include <lustre_lib.h>
  44 #include <obd.h>
  45 #include <obd_class.h>
  46 #include "ptlrpc_internal.h"
  47
  48 /**
  49  * Helper function. Sends \a len bytes from \a base at offset \a offset
  50  * over \a conn connection to portal \a portal.
  51  * Returns 0 on success or error code.
  52  */
  53 static int ptl_send_buf (lnet_handle_md_t *mdh, void *base, int len,
  54                          lnet_ack_req_t ack, struct ptlrpc_cb_id *cbid,
  55                          struct ptlrpc_connection *conn, int portal, __u64 xid,
  56                          unsigned int offset)
  57 {
  58         int              rc;
  59         lnet_md_t         md;
  60         ENTRY;
  61
  62         LASSERT (portal != 0);
  63         LASSERT (conn != NULL);
  64         CDEBUG (D_INFO, "conn=%p id %s\n", conn, libcfs_id2str(conn->c_peer));
  65         md.start     = base;
  66         md.length    = len;
  67         md.threshold = (ack == LNET_ACK_REQ) ? 2 : 1;
  68         md.options   = PTLRPC_MD_OPTIONS;
  69         md.user_ptr  = cbid;
  70         md.eq_handle = ptlrpc_eq_h;
  71
  72         if (unlikely(ack == LNET_ACK_REQ &&
  73                      OBD_FAIL_CHECK_ORSET(OBD_FAIL_PTLRPC_ACK, OBD_FAIL_ONCE))){
  74                 /* don't ask for the ack to simulate failing client */
  75                 ack = LNET_NOACK_REQ;
  76         }
  77
  78         rc = LNetMDBind (md, LNET_UNLINK, mdh);
  79         if (unlikely(rc != 0)) {
  80                 CERROR ("LNetMDBind failed: %d\n", rc);
  81                 LASSERT (rc == -ENOMEM);
  82                 RETURN (-ENOMEM);
  83         }
  84
  85         CDEBUG(D_NET, "Sending %d bytes to portal %d, xid "LPD64", offset %u\n",
  86                len, portal, xid, offset);
  87
  88         rc = LNetPut (conn->c_self, *mdh, ack,
  89                       conn->c_peer, portal, xid, offset, 0);
  90         if (unlikely(rc != 0)) {
  91                 int rc2;
  92                 /* We're going to get an UNLINK event when I unlink below,
  93                  * which will complete just like any other failed send, so
  94                  * I fall through and return success here! */
  95                 CERROR("LNetPut(%s, %d, "LPD64") failed: %d\n",
  96                        libcfs_id2str(conn->c_peer), portal, xid, rc);
  97                 rc2 = LNetMDUnlink(*mdh);
  98                 LASSERTF(rc2 == 0, "rc2 = %d\n", rc2);
  99         }
 100
 101         RETURN (0);
 102 }
 103
 104 #ifdef HAVE_SERVER_SUPPORT
 105 /**
 106  * Prepare bulk descriptor for specified incoming request \a req that
 107  * can fit \a npages * pages. \a type is bulk type. \a portal is where
 108  * the bulk to be sent. Used on server-side after request was already
 109  * received.
 110  * Returns pointer to newly allocatrd initialized bulk descriptor or NULL on
 111  * error.
 112  */
 113 struct ptlrpc_bulk_desc *ptlrpc_prep_bulk_exp(struct ptlrpc_request *req,
 114                                               int npages, int type, int portal)
 115 {
 116         struct obd_export *exp = req->rq_export;
 117         struct ptlrpc_bulk_desc *desc;
 118
 119         ENTRY;
 120         LASSERT(type == BULK_PUT_SOURCE || type == BULK_GET_SINK);
 121
 122         desc = new_bulk(npages, type, portal);
 123         if (desc == NULL)
 124                 RETURN(NULL);
 125
 126         desc->bd_export = class_export_get(exp);
 127         desc->bd_req = req;
 128
 129         desc->bd_cbid.cbid_fn  = server_bulk_callback;
 130         desc->bd_cbid.cbid_arg = desc;
 131
 132         /* NB we don't assign rq_bulk here; server-side requests are
 133          * re-used, and the handler frees the bulk desc explicitly. */
 134
 135         return desc;
 136 }
 137
 138 /**
 139  * Starts bulk transfer for descriptor \a desc
 140  * Returns 0 on success or error code.
 141  */
 142 int ptlrpc_start_bulk_transfer(struct ptlrpc_bulk_desc *desc)
 143 {
 144         struct ptlrpc_connection *conn = desc->bd_export->exp_connection;
 145         int                       rc;
 146         int                       rc2;
 147         lnet_md_t                 md;
 148         __u64                     xid;
 149         ENTRY;
 150
 151         if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_BULK_PUT_NET))
 152                 RETURN(0);
 153
 154         /* NB no locking required until desc is on the network */
 155         LASSERT (!desc->bd_network_rw);
 156         LASSERT (desc->bd_type == BULK_PUT_SOURCE ||
 157                  desc->bd_type == BULK_GET_SINK);
 158         desc->bd_success = 0;
 159
 160         md.user_ptr = &desc->bd_cbid;
 161         md.eq_handle = ptlrpc_eq_h;
 162         md.threshold = 2; /* SENT and ACK/REPLY */
 163         md.options = PTLRPC_MD_OPTIONS;
 164         ptlrpc_fill_bulk_md(&md, desc);
 165
 166         LASSERT (desc->bd_cbid.cbid_fn == server_bulk_callback);
 167         LASSERT (desc->bd_cbid.cbid_arg == desc);
 168
 169         /* NB total length may be 0 for a read past EOF, so we send a 0
 170          * length bulk, since the client expects a bulk event. */
 171
 172         rc = LNetMDBind(md, LNET_UNLINK, &desc->bd_md_h);
 173         if (rc != 0) {
 174                 CERROR("LNetMDBind failed: %d\n", rc);
 175                 LASSERT (rc == -ENOMEM);
 176                 RETURN(-ENOMEM);
 177         }
 178
 179         /* Client's bulk and reply matchbits are the same */
 180         xid = desc->bd_req->rq_xid;
 181         CDEBUG(D_NET, "Transferring %u pages %u bytes via portal %d "
 182                "id %s xid "LPX64"\n", desc->bd_iov_count,
 183                desc->bd_nob, desc->bd_portal,
 184                libcfs_id2str(conn->c_peer), xid);
 185
 186         /* Network is about to get at the memory */
 187         desc->bd_network_rw = 1;
 188
 189         if (desc->bd_type == BULK_PUT_SOURCE)
 190                 rc = LNetPut (conn->c_self, desc->bd_md_h, LNET_ACK_REQ,
 191                               conn->c_peer, desc->bd_portal, xid, 0, 0);
 192         else
 193                 rc = LNetGet (conn->c_self, desc->bd_md_h,
 194                               conn->c_peer, desc->bd_portal, xid, 0);
 195
 196         if (rc != 0) {
 197                 /* Can't send, so we unlink the MD bound above.  The UNLINK
 198                  * event this creates will signal completion with failure,
 199                  * so we return SUCCESS here! */
 200                 CERROR("Transfer(%s, %d, "LPX64") failed: %d\n",
 201                        libcfs_id2str(conn->c_peer), desc->bd_portal, xid, rc);
 202                 rc2 = LNetMDUnlink(desc->bd_md_h);
 203                 LASSERT (rc2 == 0);
 204         }
 205
 206         RETURN(0);
 207 }
 208
 209 /**
 210  * Server side bulk abort. Idempotent. Not thread-safe (i.e. only
 211  * serialises with completion callback)
 212  */
 213 void ptlrpc_abort_bulk(struct ptlrpc_bulk_desc *desc)
 214 {
 215         struct l_wait_info       lwi;
 216         int                      rc;
 217
 218         LASSERT(!cfs_in_interrupt());           /* might sleep */
 219
 220         if (!ptlrpc_server_bulk_active(desc))   /* completed or */
 221                 return;                         /* never started */
 222
 223         /* We used to poison the pages with 0xab here because we did not want to
 224          * send any meaningful data over the wire for evicted clients (bug 9297)
 225          * However, this is no longer safe now that we use the page cache on the
 226          * OSS (bug 20560) */
 227
 228         /* The unlink ensures the callback happens ASAP and is the last
 229          * one.  If it fails, it must be because completion just happened,
 230          * but we must still l_wait_event() in this case, to give liblustre
 231          * a chance to run server_bulk_callback()*/
 232
 233         LNetMDUnlink(desc->bd_md_h);
 234
 235         for (;;) {
 236                 /* Network access will complete in finite time but the HUGE
 237                  * timeout lets us CWARN for visibility of sluggish NALs */
 238                 lwi = LWI_TIMEOUT_INTERVAL(cfs_time_seconds(LONG_UNLINK),
 239                                            cfs_time_seconds(1), NULL, NULL);
 240                 rc = l_wait_event(desc->bd_waitq,
 241                                   !ptlrpc_server_bulk_active(desc), &lwi);
 242                 if (rc == 0)
 243                         return;
 244
 245                 LASSERT(rc == -ETIMEDOUT);
 246                 CWARN("Unexpectedly long timeout: desc %p\n", desc);
 247         }
 248 }
 249 #endif /* HAVE_SERVER_SUPPORT */
 250
 251 /**
 252  * Register bulk for later transfer
 253  * Returns 0 on success or error code.
 254  */
 255 int ptlrpc_register_bulk(struct ptlrpc_request *req)
 256 {
 257         struct ptlrpc_bulk_desc *desc = req->rq_bulk;
 258         lnet_process_id_t peer;
 259         int rc;
 260         int rc2;
 261         lnet_handle_me_t  me_h;
 262         lnet_md_t         md;
 263         ENTRY;
 264
 265         if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_BULK_GET_NET))
 266                 RETURN(0);
 267
 268         /* NB no locking required until desc is on the network */
 269         LASSERT (desc->bd_nob > 0);
 270         LASSERT (!desc->bd_network_rw);
 271         LASSERT (desc->bd_iov_count <= PTLRPC_MAX_BRW_PAGES);
 272         LASSERT (desc->bd_req != NULL);
 273         LASSERT (desc->bd_type == BULK_PUT_SINK ||
 274                  desc->bd_type == BULK_GET_SOURCE);
 275
 276         desc->bd_success = 0;
 277
 278         peer = desc->bd_import->imp_connection->c_peer;
 279
 280         md.user_ptr = &desc->bd_cbid;
 281         md.eq_handle = ptlrpc_eq_h;
 282         md.threshold = 1;                       /* PUT or GET */
 283         md.options = PTLRPC_MD_OPTIONS |
 284                      ((desc->bd_type == BULK_GET_SOURCE) ?
 285                       LNET_MD_OP_GET : LNET_MD_OP_PUT);
 286         ptlrpc_fill_bulk_md(&md, desc);
 287
 288         LASSERT (desc->bd_cbid.cbid_fn == client_bulk_callback);
 289         LASSERT (desc->bd_cbid.cbid_arg == desc);
 290
 291         /* XXX Registering the same xid on retried bulk makes my head
 292          * explode trying to understand how the original request's bulk
 293          * might interfere with the retried request -eeb
 294          * On the other hand replaying with the same xid is fine, since
 295          * we are guaranteed old request have completed. -green */
 296         LASSERTF(!(desc->bd_registered &&
 297                  req->rq_send_state != LUSTRE_IMP_REPLAY) ||
 298                  req->rq_xid != desc->bd_last_xid,
 299                  "registered: %d  rq_xid: "LPU64" bd_last_xid: "LPU64"\n",
 300                  desc->bd_registered, req->rq_xid, desc->bd_last_xid);
 301         desc->bd_registered = 1;
 302         desc->bd_last_xid = req->rq_xid;
 303
 304         rc = LNetMEAttach(desc->bd_portal, peer,
 305                          req->rq_xid, 0, LNET_UNLINK, LNET_INS_AFTER, &me_h);
 306         if (rc != 0) {
 307                 CERROR("LNetMEAttach failed: %d\n", rc);
 308                 LASSERT (rc == -ENOMEM);
 309                 RETURN (-ENOMEM);
 310         }
 311
 312         /* About to let the network at it... */
 313         desc->bd_network_rw = 1;
 314         rc = LNetMDAttach(me_h, md, LNET_UNLINK, &desc->bd_md_h);
 315         if (rc != 0) {
 316                 CERROR("LNetMDAttach failed: %d\n", rc);
 317                 LASSERT (rc == -ENOMEM);
 318                 desc->bd_network_rw = 0;
 319                 rc2 = LNetMEUnlink (me_h);
 320                 LASSERT (rc2 == 0);
 321                 RETURN (-ENOMEM);
 322         }
 323
 324         CDEBUG(D_NET, "Setup bulk %s buffers: %u pages %u bytes, xid "LPU64", "
 325                "portal %u\n",
 326                desc->bd_type == BULK_GET_SOURCE ? "get-source" : "put-sink",
 327                desc->bd_iov_count, desc->bd_nob,
 328                req->rq_xid, desc->bd_portal);
 329         RETURN(0);
 330 }
 331
 332 /**
 333  * Disconnect a bulk desc from the network. Idempotent. Not
 334  * thread-safe (i.e. only interlocks with completion callback).
 335  * Returns 1 on success or 0 if network unregistration failed for whatever
 336  * reason.
 337  */
 338 int ptlrpc_unregister_bulk(struct ptlrpc_request *req, int async)
 339 {
 340         struct ptlrpc_bulk_desc *desc = req->rq_bulk;
 341         cfs_waitq_t             *wq;
 342         struct l_wait_info       lwi;
 343         int                      rc;
 344         ENTRY;
 345
 346         LASSERT(!cfs_in_interrupt());     /* might sleep */
 347
 348         /* Let's setup deadline for reply unlink. */
 349         if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_BULK_UNLINK) &&
 350             async && req->rq_bulk_deadline == 0)
 351                 req->rq_bulk_deadline = cfs_time_current_sec() + LONG_UNLINK;
 352
 353         if (!ptlrpc_client_bulk_active(req))  /* completed or */
 354                 RETURN(1);                    /* never registered */
 355
 356         LASSERT(desc->bd_req == req);  /* bd_req NULL until registered */
 357
 358         /* the unlink ensures the callback happens ASAP and is the last
 359          * one.  If it fails, it must be because completion just happened,
 360          * but we must still l_wait_event() in this case to give liblustre
 361          * a chance to run client_bulk_callback() */
 362
 363         LNetMDUnlink(desc->bd_md_h);
 364
 365         if (!ptlrpc_client_bulk_active(req))  /* completed or */
 366                 RETURN(1);                    /* never registered */
 367
 368         /* Move to "Unregistering" phase as bulk was not unlinked yet. */
 369         ptlrpc_rqphase_move(req, RQ_PHASE_UNREGISTERING);
 370
 371         /* Do not wait for unlink to finish. */
 372         if (async)
 373                 RETURN(0);
 374
 375         if (req->rq_set != NULL)
 376                 wq = &req->rq_set->set_waitq;
 377         else
 378                 wq = &req->rq_reply_waitq;
 379
 380         for (;;) {
 381                 /* Network access will complete in finite time but the HUGE
 382                  * timeout lets us CWARN for visibility of sluggish NALs */
 383                 lwi = LWI_TIMEOUT_INTERVAL(cfs_time_seconds(LONG_UNLINK),
 384                                            cfs_time_seconds(1), NULL, NULL);
 385                 rc = l_wait_event(*wq, !ptlrpc_client_bulk_active(req), &lwi);
 386                 if (rc == 0) {
 387                         ptlrpc_rqphase_move(req, req->rq_next_phase);
 388                         RETURN(1);
 389                 }
 390
 391                 LASSERT(rc == -ETIMEDOUT);
 392                 DEBUG_REQ(D_WARNING, req, "Unexpectedly long timeout: desc %p",
 393                           desc);
 394         }
 395         RETURN(0);
 396 }
 397
 398 static void ptlrpc_at_set_reply(struct ptlrpc_request *req, int flags)
 399 {
 400         struct ptlrpc_service *svc = req->rq_rqbd->rqbd_service;
 401         int service_time = max_t(int, cfs_time_current_sec() -
 402                                  req->rq_arrival_time.tv_sec, 1);
 403
 404         if (!(flags & PTLRPC_REPLY_EARLY) &&
 405             (req->rq_type != PTL_RPC_MSG_ERR) &&
 406             (req->rq_reqmsg != NULL) &&
 407             !(lustre_msg_get_flags(req->rq_reqmsg) &
 408               (MSG_RESENT | MSG_REPLAY |
 409                MSG_REQ_REPLAY_DONE | MSG_LOCK_REPLAY_DONE))) {
 410                 /* early replies, errors and recovery requests don't count
 411                  * toward our service time estimate */
 412                 int oldse = at_measured(&svc->srv_at_estimate, service_time);
 413                 if (oldse != 0)
 414                         DEBUG_REQ(D_ADAPTTO, req,
 415                                   "svc %s changed estimate from %d to %d",
 416                                   svc->srv_name, oldse,
 417                                   at_get(&svc->srv_at_estimate));
 418         }
 419         /* Report actual service time for client latency calc */
 420         lustre_msg_set_service_time(req->rq_repmsg, service_time);
 421         /* Report service time estimate for future client reqs, but report 0
 422          * (to be ignored by client) if it's a error reply during recovery.
 423          * (bz15815) */
 424         if (req->rq_type == PTL_RPC_MSG_ERR &&
 425             (req->rq_export == NULL || req->rq_export->exp_obd->obd_recovering))
 426                 lustre_msg_set_timeout(req->rq_repmsg, 0);
 427         else
 428                 lustre_msg_set_timeout(req->rq_repmsg,
 429                                        at_get(&svc->srv_at_estimate));
 430
 431         if (req->rq_reqmsg &&
 432             !(lustre_msghdr_get_flags(req->rq_reqmsg) & MSGHDR_AT_SUPPORT)) {
 433                 CDEBUG(D_ADAPTTO, "No early reply support: flags=%#x "
 434                        "req_flags=%#x magic=%d:%x/%x len=%d\n",
 435                        flags, lustre_msg_get_flags(req->rq_reqmsg),
 436                        lustre_msg_is_v1(req->rq_reqmsg),
 437                        lustre_msg_get_magic(req->rq_reqmsg),
 438                        lustre_msg_get_magic(req->rq_repmsg), req->rq_replen);
 439         }
 440 }
 441
 442 /**
 443  * Send request reply from request \a req reply buffer.
 444  * \a flags defines reply types
 445  * Returns 0 on sucess or error code
 446  */
 447 int ptlrpc_send_reply(struct ptlrpc_request *req, int flags)
 448 {
 449         struct ptlrpc_service     *svc = req->rq_rqbd->rqbd_service;
 450         struct ptlrpc_reply_state *rs = req->rq_reply_state;
 451         struct ptlrpc_connection  *conn;
 452         int                        rc;
 453
 454         /* We must already have a reply buffer (only ptlrpc_error() may be
 455          * called without one). The reply generated by sptlrpc layer (e.g.
 456          * error notify, etc.) might have NULL rq->reqmsg; Otherwise we must
 457          * have a request buffer which is either the actual (swabbed) incoming
 458          * request, or a saved copy if this is a req saved in
 459          * target_queue_final_reply().
 460          */
 461         LASSERT (req->rq_no_reply == 0);
 462         LASSERT (req->rq_reqbuf != NULL);
 463         LASSERT (rs != NULL);
 464         LASSERT ((flags & PTLRPC_REPLY_MAYBE_DIFFICULT) || !rs->rs_difficult);
 465         LASSERT (req->rq_repmsg != NULL);
 466         LASSERT (req->rq_repmsg == rs->rs_msg);
 467         LASSERT (rs->rs_cb_id.cbid_fn == reply_out_callback);
 468         LASSERT (rs->rs_cb_id.cbid_arg == rs);
 469
 470         /* There may be no rq_export during failover */
 471
 472         if (unlikely(req->rq_export && req->rq_export->exp_obd &&
 473                      req->rq_export->exp_obd->obd_fail)) {
 474                 /* Failed obd's only send ENODEV */
 475                 req->rq_type = PTL_RPC_MSG_ERR;
 476                 req->rq_status = -ENODEV;
 477                 CDEBUG(D_HA, "sending ENODEV from failed obd %d\n",
 478                        req->rq_export->exp_obd->obd_minor);
 479         }
 480
 481         if (req->rq_type != PTL_RPC_MSG_ERR)
 482                 req->rq_type = PTL_RPC_MSG_REPLY;
 483
 484         lustre_msg_set_type(req->rq_repmsg, req->rq_type);
 485         lustre_msg_set_status(req->rq_repmsg, req->rq_status);
 486         lustre_msg_set_opc(req->rq_repmsg,
 487                 req->rq_reqmsg ? lustre_msg_get_opc(req->rq_reqmsg) : 0);
 488
 489         target_pack_pool_reply(req);
 490
 491         ptlrpc_at_set_reply(req, flags);
 492
 493         if (req->rq_export == NULL || req->rq_export->exp_connection == NULL)
 494                 conn = ptlrpc_connection_get(req->rq_peer, req->rq_self, NULL);
 495         else
 496                 conn = ptlrpc_connection_addref(req->rq_export->exp_connection);
 497
 498         if (unlikely(conn == NULL)) {
 499                 CERROR("not replying on NULL connection\n"); /* bug 9635 */
 500                 return -ENOTCONN;
 501         }
 502         ptlrpc_rs_addref(rs);                   /* +1 ref for the network */
 503
 504         rc = sptlrpc_svc_wrap_reply(req);
 505         if (unlikely(rc))
 506                 goto out;
 507
 508         req->rq_sent = cfs_time_current_sec();
 509
 510         rc = ptl_send_buf (&rs->rs_md_h, rs->rs_repbuf, rs->rs_repdata_len,
 511                            (rs->rs_difficult && !rs->rs_no_ack) ?
 512                            LNET_ACK_REQ : LNET_NOACK_REQ,
 513                            &rs->rs_cb_id, conn, svc->srv_rep_portal,
 514                            req->rq_xid, req->rq_reply_off);
 515 out:
 516         if (unlikely(rc != 0))
 517                 ptlrpc_req_drop_rs(req);
 518         ptlrpc_connection_put(conn);
 519         return rc;
 520 }
 521
 522 int ptlrpc_reply (struct ptlrpc_request *req)
 523 {
 524         if (req->rq_no_reply)
 525                 return 0;
 526         else
 527                 return (ptlrpc_send_reply(req, 0));
 528 }
 529
 530 /**
 531  * For request \a req send an error reply back. Create empty
 532  * reply buffers if necessary.
 533  */
 534 int ptlrpc_send_error(struct ptlrpc_request *req, int may_be_difficult)
 535 {
 536         int rc;
 537         ENTRY;
 538
 539         if (req->rq_no_reply)
 540                 RETURN(0);
 541
 542         if (!req->rq_repmsg) {
 543                 rc = lustre_pack_reply(req, 1, NULL, NULL);
 544                 if (rc)
 545                         RETURN(rc);
 546         }
 547
 548         req->rq_type = PTL_RPC_MSG_ERR;
 549
 550         rc = ptlrpc_send_reply(req, may_be_difficult);
 551         RETURN(rc);
 552 }
 553
 554 int ptlrpc_error(struct ptlrpc_request *req)
 555 {
 556         return ptlrpc_send_error(req, 0);
 557 }
 558
 559 /**
 560  * Send request \a request.
 561  * if \a noreply is set, don't expect any reply back and don't set up
 562  * reply buffers.
 563  * Returns 0 on success or error code.
 564  */
 565 int ptl_send_rpc(struct ptlrpc_request *request, int noreply)
 566 {
 567         int rc;
 568         int rc2;
 569         int mpflag = 0;
 570         struct ptlrpc_connection *connection;
 571         lnet_handle_me_t  reply_me_h;
 572         lnet_md_t         reply_md;
 573         struct obd_device *obd = request->rq_import->imp_obd;
 574         ENTRY;
 575
 576         if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_DROP_RPC))
 577                 RETURN(0);
 578
 579         LASSERT(request->rq_type == PTL_RPC_MSG_REQUEST);
 580         LASSERT(request->rq_wait_ctx == 0);
 581
 582         /* If this is a re-transmit, we're required to have disengaged
 583          * cleanly from the previous attempt */
 584         LASSERT(!request->rq_receiving_reply);
 585
 586         if (request->rq_import->imp_obd &&
 587             request->rq_import->imp_obd->obd_fail) {
 588                 CDEBUG(D_HA, "muting rpc for failed imp obd %s\n",
 589                        request->rq_import->imp_obd->obd_name);
 590                 /* this prevents us from waiting in ptlrpc_queue_wait */
 591                 request->rq_err = 1;
 592                 request->rq_status = -ENODEV;
 593                 RETURN(-ENODEV);
 594         }
 595
 596         connection = request->rq_import->imp_connection;
 597
 598         lustre_msg_set_handle(request->rq_reqmsg,
 599                               &request->rq_import->imp_remote_handle);
 600         lustre_msg_set_type(request->rq_reqmsg, PTL_RPC_MSG_REQUEST);
 601         lustre_msg_set_conn_cnt(request->rq_reqmsg,
 602                                 request->rq_import->imp_conn_cnt);
 603         lustre_msghdr_set_flags(request->rq_reqmsg,
 604                                 request->rq_import->imp_msghdr_flags);
 605
 606         if (request->rq_resend)
 607                 lustre_msg_add_flags(request->rq_reqmsg, MSG_RESENT);
 608
 609         if (request->rq_memalloc)
 610                 mpflag = cfs_memory_pressure_get_and_set();
 611
 612         rc = sptlrpc_cli_wrap_request(request);
 613         if (rc)
 614                 GOTO(out, rc);
 615
 616         /* bulk register should be done after wrap_request() */
 617         if (request->rq_bulk != NULL) {
 618                 rc = ptlrpc_register_bulk (request);
 619                 if (rc != 0)
 620                         GOTO(out, rc);
 621         }
 622
 623         if (!noreply) {
 624                 LASSERT (request->rq_replen != 0);
 625                 if (request->rq_repbuf == NULL) {
 626                         LASSERT(request->rq_repdata == NULL);
 627                         LASSERT(request->rq_repmsg == NULL);
 628                         rc = sptlrpc_cli_alloc_repbuf(request,
 629                                                       request->rq_replen);
 630                         if (rc) {
 631                                 /* this prevents us from looping in
 632                                  * ptlrpc_queue_wait */
 633                                 request->rq_err = 1;
 634                                 request->rq_status = rc;
 635                                 GOTO(cleanup_bulk, rc);
 636                         }
 637                 } else {
 638                         request->rq_repdata = NULL;
 639                         request->rq_repmsg = NULL;
 640                 }
 641
 642                 rc = LNetMEAttach(request->rq_reply_portal,/*XXX FIXME bug 249*/
 643                                   connection->c_peer, request->rq_xid, 0,
 644                                   LNET_UNLINK, LNET_INS_AFTER, &reply_me_h);
 645                 if (rc != 0) {
 646                         CERROR("LNetMEAttach failed: %d\n", rc);
 647                         LASSERT (rc == -ENOMEM);
 648                         GOTO(cleanup_bulk, rc = -ENOMEM);
 649                 }
 650         }
 651
 652         cfs_spin_lock(&request->rq_lock);
 653         /* If the MD attach succeeds, there _will_ be a reply_in callback */
 654         request->rq_receiving_reply = !noreply;
 655         /* We are responsible for unlinking the reply buffer */
 656         request->rq_must_unlink = !noreply;
 657         /* Clear any flags that may be present from previous sends. */
 658         request->rq_replied = 0;
 659         request->rq_err = 0;
 660         request->rq_timedout = 0;
 661         request->rq_net_err = 0;
 662         request->rq_resend = 0;
 663         request->rq_restart = 0;
 664         request->rq_reply_truncate = 0;
 665         cfs_spin_unlock(&request->rq_lock);
 666
 667         if (!noreply) {
 668                 reply_md.start     = request->rq_repbuf;
 669                 reply_md.length    = request->rq_repbuf_len;
 670                 /* Allow multiple early replies */
 671                 reply_md.threshold = LNET_MD_THRESH_INF;
 672                 /* Manage remote for early replies */
 673                 reply_md.options   = PTLRPC_MD_OPTIONS | LNET_MD_OP_PUT |
 674                         LNET_MD_MANAGE_REMOTE |
 675                         LNET_MD_TRUNCATE; /* allow to make EOVERFLOW error */;
 676                 reply_md.user_ptr  = &request->rq_reply_cbid;
 677                 reply_md.eq_handle = ptlrpc_eq_h;
 678
 679                 /* We must see the unlink callback to unset rq_must_unlink,
 680                    so we can't auto-unlink */
 681                 rc = LNetMDAttach(reply_me_h, reply_md, LNET_RETAIN,
 682                                   &request->rq_reply_md_h);
 683                 if (rc != 0) {
 684                         CERROR("LNetMDAttach failed: %d\n", rc);
 685                         LASSERT (rc == -ENOMEM);
 686                         cfs_spin_lock(&request->rq_lock);
 687                         /* ...but the MD attach didn't succeed... */
 688                         request->rq_receiving_reply = 0;
 689                         cfs_spin_unlock(&request->rq_lock);
 690                         GOTO(cleanup_me, rc = -ENOMEM);
 691                 }
 692
 693                 CDEBUG(D_NET, "Setup reply buffer: %u bytes, xid "LPU64
 694                        ", portal %u\n",
 695                        request->rq_repbuf_len, request->rq_xid,
 696                        request->rq_reply_portal);
 697         }
 698
 699         /* add references on request for request_out_callback */
 700         ptlrpc_request_addref(request);
 701         if (obd->obd_svc_stats != NULL)
 702                 lprocfs_counter_add(obd->obd_svc_stats, PTLRPC_REQACTIVE_CNTR,
 703                         cfs_atomic_read(&request->rq_import->imp_inflight));
 704
 705         OBD_FAIL_TIMEOUT(OBD_FAIL_PTLRPC_DELAY_SEND, request->rq_timeout + 5);
 706
 707         cfs_gettimeofday(&request->rq_arrival_time);
 708         request->rq_sent = cfs_time_current_sec();
 709         /* We give the server rq_timeout secs to process the req, and
 710            add the network latency for our local timeout. */
 711         request->rq_deadline = request->rq_sent + request->rq_timeout +
 712                 ptlrpc_at_get_net_latency(request);
 713
 714         ptlrpc_pinger_sending_on_import(request->rq_import);
 715
 716         DEBUG_REQ(D_INFO, request, "send flg=%x",
 717                   lustre_msg_get_flags(request->rq_reqmsg));
 718         rc = ptl_send_buf(&request->rq_req_md_h,
 719                           request->rq_reqbuf, request->rq_reqdata_len,
 720                           LNET_NOACK_REQ, &request->rq_req_cbid,
 721                           connection,
 722                           request->rq_request_portal,
 723                           request->rq_xid, 0);
 724         if (rc == 0)
 725                 GOTO(out, rc);
 726
 727         ptlrpc_req_finished(request);
 728         if (noreply)
 729                 GOTO(out, rc);
 730
 731  cleanup_me:
 732         /* MEUnlink is safe; the PUT didn't even get off the ground, and
 733          * nobody apart from the PUT's target has the right nid+XID to
 734          * access the reply buffer. */
 735         rc2 = LNetMEUnlink(reply_me_h);
 736         LASSERT (rc2 == 0);
 737         /* UNLINKED callback called synchronously */
 738         LASSERT(!request->rq_receiving_reply);
 739
 740  cleanup_bulk:
 741         /* We do sync unlink here as there was no real transfer here so
 742          * the chance to have long unlink to sluggish net is smaller here. */
 743         ptlrpc_unregister_bulk(request, 0);
 744  out:
 745         if (request->rq_memalloc)
 746                 cfs_memory_pressure_restore(mpflag);
 747         return rc;
 748 }
 749
 750 /**
 751  * Register request buffer descriptor for request receiving.
 752  */
 753 int ptlrpc_register_rqbd(struct ptlrpc_request_buffer_desc *rqbd)
 754 {
 755         struct ptlrpc_service   *service = rqbd->rqbd_service;
 756         static lnet_process_id_t  match_id = {LNET_NID_ANY, LNET_PID_ANY};
 757         int                      rc;
 758         lnet_md_t                 md;
 759         lnet_handle_me_t          me_h;
 760
 761         CDEBUG(D_NET, "LNetMEAttach: portal %d\n",
 762                service->srv_req_portal);
 763
 764         if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_RQBD))
 765                 return (-ENOMEM);
 766
 767         rc = LNetMEAttach(service->srv_req_portal,
 768                           match_id, 0, ~0, LNET_UNLINK, LNET_INS_AFTER, &me_h);
 769         if (rc != 0) {
 770                 CERROR("LNetMEAttach failed: %d\n", rc);
 771                 return (-ENOMEM);
 772         }
 773
 774         LASSERT(rqbd->rqbd_refcount == 0);
 775         rqbd->rqbd_refcount = 1;
 776
 777         md.start     = rqbd->rqbd_buffer;
 778         md.length    = service->srv_buf_size;
 779         md.max_size  = service->srv_max_req_size;
 780         md.threshold = LNET_MD_THRESH_INF;
 781         md.options   = PTLRPC_MD_OPTIONS | LNET_MD_OP_PUT | LNET_MD_MAX_SIZE;
 782         md.user_ptr  = &rqbd->rqbd_cbid;
 783         md.eq_handle = ptlrpc_eq_h;
 784
 785         rc = LNetMDAttach(me_h, md, LNET_UNLINK, &rqbd->rqbd_md_h);
 786         if (rc == 0)
 787                 return (0);
 788
 789         CERROR("LNetMDAttach failed: %d; \n", rc);
 790         LASSERT (rc == -ENOMEM);
 791         rc = LNetMEUnlink (me_h);
 792         LASSERT (rc == 0);
 793         rqbd->rqbd_refcount = 0;
 794
 795         return (-ENOMEM);
 796 }