lustre/ptlrpc/client.c

   1 /* -*- mode: c; c-basic-offset: 8; indent-tabs-mode: nil; -*-
   2  * vim:expandtab:shiftwidth=8:tabstop=8:
   3  *
   4  *  Copyright (C) 2002 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
  25 #include <linux/obd_support.h>
  26 #include <linux/obd_class.h>
  27 #include <linux/lustre_lib.h>
  28 #include <linux/lustre_ha.h>
  29 #include <linux/lustre_import.h>
  30
  31 void ptlrpc_init_client(int req_portal, int rep_portal, char *name,
  32                         struct ptlrpc_client *cl)
  33 {
  34         cl->cli_request_portal = req_portal;
  35         cl->cli_reply_portal   = rep_portal;
  36         cl->cli_name           = name;
  37 }
  38
  39 __u8 *ptlrpc_req_to_uuid(struct ptlrpc_request *req)
  40 {
  41         return req->rq_connection->c_remote_uuid;
  42 }
  43
  44 struct ptlrpc_connection *ptlrpc_uuid_to_connection(obd_uuid_t uuid)
  45 {
  46         struct ptlrpc_connection *c;
  47         struct lustre_peer peer;
  48         int err;
  49
  50         err = kportal_uuid_to_peer(uuid, &peer);
  51         if (err != 0) {
  52                 CERROR("cannot find peer %s!\n", uuid);
  53                 return NULL;
  54         }
  55
  56         c = ptlrpc_get_connection(&peer, uuid);
  57         if (c) {
  58                 memcpy(c->c_remote_uuid, uuid, sizeof(c->c_remote_uuid));
  59                 c->c_epoch++;
  60         }
  61
  62         CDEBUG(D_INFO, "%s -> %p\n", uuid, c);
  63
  64         return c;
  65 }
  66
  67 void ptlrpc_readdress_connection(struct ptlrpc_connection *conn,obd_uuid_t uuid)
  68 {
  69         struct lustre_peer peer;
  70         int err;
  71
  72         err = kportal_uuid_to_peer(uuid, &peer);
  73         if (err != 0) {
  74                 CERROR("cannot find peer %s!\n", uuid);
  75                 return;
  76         }
  77
  78         memcpy(&conn->c_peer, &peer, sizeof(peer));
  79         return;
  80 }
  81
  82 struct ptlrpc_bulk_desc *ptlrpc_prep_bulk(struct ptlrpc_connection *conn)
  83 {
  84         struct ptlrpc_bulk_desc *desc;
  85
  86         OBD_ALLOC(desc, sizeof(*desc));
  87         if (desc != NULL) {
  88                 desc->bd_connection = ptlrpc_connection_addref(conn);
  89                 atomic_set(&desc->bd_refcount, 1);
  90                 init_waitqueue_head(&desc->bd_waitq);
  91                 INIT_LIST_HEAD(&desc->bd_page_list);
  92                 ptl_set_inv_handle(&desc->bd_md_h);
  93                 ptl_set_inv_handle(&desc->bd_me_h);
  94         }
  95
  96         return desc;
  97 }
  98
  99 struct ptlrpc_bulk_page *ptlrpc_prep_bulk_page(struct ptlrpc_bulk_desc *desc)
 100 {
 101         struct ptlrpc_bulk_page *bulk;
 102
 103         OBD_ALLOC(bulk, sizeof(*bulk));
 104         if (bulk != NULL) {
 105                 bulk->bp_desc = desc;
 106                 list_add_tail(&bulk->bp_link, &desc->bd_page_list);
 107                 desc->bd_page_count++;
 108         }
 109         return bulk;
 110 }
 111
 112 void ptlrpc_free_bulk(struct ptlrpc_bulk_desc *desc)
 113 {
 114         struct list_head *tmp, *next;
 115         ENTRY;
 116         if (desc == NULL) {
 117                 EXIT;
 118                 return;
 119         }
 120
 121         list_for_each_safe(tmp, next, &desc->bd_page_list) {
 122                 struct ptlrpc_bulk_page *bulk;
 123                 bulk = list_entry(tmp, struct ptlrpc_bulk_page, bp_link);
 124                 ptlrpc_free_bulk_page(bulk);
 125         }
 126
 127         ptlrpc_put_connection(desc->bd_connection);
 128
 129         OBD_FREE(desc, sizeof(*desc));
 130         EXIT;
 131 }
 132
 133 void ptlrpc_free_bulk_page(struct ptlrpc_bulk_page *bulk)
 134 {
 135         ENTRY;
 136         if (bulk == NULL) {
 137                 EXIT;
 138                 return;
 139         }
 140
 141         list_del(&bulk->bp_link);
 142         bulk->bp_desc->bd_page_count--;
 143         OBD_FREE(bulk, sizeof(*bulk));
 144         EXIT;
 145 }
 146
 147 struct ptlrpc_request *ptlrpc_prep_req(struct obd_import *imp, int opcode,
 148                                        int count, int *lengths, char **bufs)
 149 {
 150         struct ptlrpc_connection *conn = imp->imp_connection;
 151         struct ptlrpc_request *request;
 152         int rc;
 153         ENTRY;
 154
 155         OBD_ALLOC(request, sizeof(*request));
 156         if (!request) {
 157                 CERROR("request allocation out of memory\n");
 158                 RETURN(NULL);
 159         }
 160
 161         rc = lustre_pack_msg(count, lengths, bufs,
 162                              &request->rq_reqlen, &request->rq_reqmsg);
 163         if (rc) {
 164                 CERROR("cannot pack request %d\n", rc);
 165                 OBD_FREE(request, sizeof(*request));
 166                 RETURN(NULL);
 167         }
 168
 169         request->rq_level = LUSTRE_CONN_FULL;
 170         request->rq_type = PTL_RPC_MSG_REQUEST;
 171         request->rq_import = imp;
 172
 173         /* XXX FIXME bug 625069 */
 174         request->rq_request_portal = imp->imp_client->cli_request_portal;
 175         request->rq_reply_portal = imp->imp_client->cli_reply_portal;
 176
 177         request->rq_connection = ptlrpc_connection_addref(conn);
 178
 179         INIT_LIST_HEAD(&request->rq_list);
 180         /*
 181          * This will be reduced once when the sender is finished (waiting for
 182          * reply, f.e.), and once when the request has been committed and is
 183          * removed from the to-be-committed list.
 184          *
 185          * Also, the refcount will be increased in ptl_send_rpc immediately
 186          * before we hand it off to portals, and there will be a corresponding
 187          * decrease in request_out_cb (which is called to indicate that portals
 188          * is finished with the request, and it can be safely freed).
 189          *
 190          * (Except in the DLM server case, where it will be dropped twice
 191          * by the sender, and then the last time by request_out_callback.)
 192          */
 193         atomic_set(&request->rq_refcount, 2);
 194
 195         spin_lock(&conn->c_lock);
 196         request->rq_xid = HTON__u32(++conn->c_xid_out);
 197         spin_unlock(&conn->c_lock);
 198
 199         request->rq_reqmsg->magic = PTLRPC_MSG_MAGIC;
 200         request->rq_reqmsg->version = PTLRPC_MSG_VERSION;
 201         request->rq_reqmsg->opc = HTON__u32(opcode);
 202         request->rq_reqmsg->flags = 0;
 203
 204         ptlrpc_hdl2req(request, &imp->imp_handle);
 205         RETURN(request);
 206 }
 207
 208 void ptlrpc_req_finished(struct ptlrpc_request *request)
 209 {
 210         if (request == NULL)
 211                 return;
 212
 213         if (atomic_dec_and_test(&request->rq_refcount))
 214                 ptlrpc_free_req(request);
 215         else
 216                 DEBUG_REQ(D_INFO, request, "refcount now %u",
 217                           atomic_read(&request->rq_refcount));
 218 }
 219
 220 void ptlrpc_free_req(struct ptlrpc_request *request)
 221 {
 222         ENTRY;
 223         if (request == NULL) {
 224                 EXIT;
 225                 return;
 226         }
 227
 228         if (atomic_read(&request->rq_refcount) != 0) {
 229                 CERROR("freeing request %p (%d->%s:%d) with refcount %d\n",
 230                        request, request->rq_reqmsg->opc,
 231                        request->rq_connection->c_remote_uuid,
 232                        request->rq_import->imp_client->cli_request_portal,
 233                        request->rq_refcount);
 234                 /* LBUG(); */
 235         }
 236
 237         if (request->rq_repmsg != NULL) {
 238                 OBD_FREE(request->rq_repmsg, request->rq_replen);
 239                 request->rq_repmsg = NULL;
 240                 request->rq_reply_md.start = NULL;
 241         }
 242         if (request->rq_reqmsg != NULL) {
 243                 OBD_FREE(request->rq_reqmsg, request->rq_reqlen);
 244                 request->rq_reqmsg = NULL;
 245         }
 246
 247         if (request->rq_connection) {
 248                 spin_lock(&request->rq_connection->c_lock);
 249                 list_del_init(&request->rq_list);
 250                 spin_unlock(&request->rq_connection->c_lock);
 251         }
 252
 253         ptlrpc_put_connection(request->rq_connection);
 254         OBD_FREE(request, sizeof(*request));
 255         EXIT;
 256 }
 257
 258 static int ptlrpc_check_reply(struct ptlrpc_request *req)
 259 {
 260         int rc = 0;
 261
 262         if (req->rq_repmsg != NULL) {
 263                 struct ptlrpc_connection *conn = req->rq_import->imp_connection;
 264                 if (req->rq_level > conn->c_level) {
 265                         CDEBUG(D_HA,
 266                                "rep to xid "LPD64" op %d to %s:%d: "
 267                                "recovery started, ignoring (%d > %d)\n",
 268                                (unsigned long long)req->rq_xid,
 269                                req->rq_reqmsg->opc, conn->c_remote_uuid,
 270                                req->rq_import->imp_client->cli_request_portal,
 271                                req->rq_level, conn->c_level);
 272                         req->rq_repmsg = NULL;
 273                         GOTO(out, rc = 0);
 274                 }
 275                 req->rq_transno = NTOH__u64(req->rq_repmsg->transno);
 276                 req->rq_flags |= PTL_RPC_FL_REPLIED;
 277                 GOTO(out, rc = 1);
 278         }
 279
 280         if (req->rq_flags & PTL_RPC_FL_RESEND) {
 281                 CERROR("-- RESTART --\n");
 282                 GOTO(out, rc = 1);
 283         }
 284
 285         if (req->rq_flags & PTL_RPC_FL_ERR) {
 286                 CERROR("-- ABORTED --\n");
 287                 GOTO(out, rc = 1);
 288         }
 289
 290  out:
 291         CDEBUG(D_NET, "req = %p, rc = %d\n", req, rc);
 292         return rc;
 293 }
 294
 295 int ptlrpc_check_status(struct ptlrpc_request *req, int err)
 296 {
 297         ENTRY;
 298
 299         if (err != 0) {
 300                 CERROR("err is %d\n", err);
 301                 RETURN(err);
 302         }
 303
 304         if (req == NULL) {
 305                 CERROR("req == NULL\n");
 306                 RETURN(-ENOMEM);
 307         }
 308
 309         if (req->rq_repmsg == NULL) {
 310                 CERROR("req->rq_repmsg == NULL\n");
 311                 RETURN(-ENOMEM);
 312         }
 313
 314         err = req->rq_repmsg->status;
 315         if (req->rq_repmsg->type == NTOH__u32(PTL_RPC_MSG_ERR)) {
 316                 CERROR("req->rq_repmsg->type == PTL_RPC_MSG_ERR\n");
 317                 RETURN(err ? err : -EINVAL);
 318         }
 319
 320         if (err != 0) {
 321                 if (err < 0)
 322                         CERROR("req->rq_repmsg->status is %d\n", err);
 323                 else
 324                         CDEBUG(D_INFO, "req->rq_repmsg->status is %d\n", err);
 325                 /* XXX: translate this error from net to host */
 326                 RETURN(err);
 327         }
 328
 329         RETURN(0);
 330 }
 331
 332 static void ptlrpc_cleanup_request_buf(struct ptlrpc_request *request)
 333 {
 334         OBD_FREE(request->rq_reqmsg, request->rq_reqlen);
 335         request->rq_reqmsg = NULL;
 336         request->rq_reqlen = 0;
 337 }
 338
 339 /* Abort this request and cleanup any resources associated with it. */
 340 static int ptlrpc_abort(struct ptlrpc_request *request)
 341 {
 342         /* First remove the ME for the reply; in theory, this means
 343          * that we can tear down the buffer safely. */
 344         PtlMEUnlink(request->rq_reply_me_h);
 345         OBD_FREE(request->rq_reply_md.start, request->rq_replen);
 346         request->rq_repmsg = NULL;
 347         request->rq_replen = 0;
 348         return 0;
 349 }
 350
 351 /* caller must hold conn->c_lock */
 352 void ptlrpc_free_committed(struct ptlrpc_connection *conn)
 353 {
 354         struct list_head *tmp, *saved;
 355         struct ptlrpc_request *req;
 356
 357 restart:
 358         list_for_each_safe(tmp, saved, &conn->c_sending_head) {
 359                 req = list_entry(tmp, struct ptlrpc_request, rq_list);
 360
 361                 if (req->rq_flags & PTL_RPC_FL_REPLAY) {
 362                         DEBUG_REQ(D_HA, req, "keeping (FL_REPLAY)");
 363                         continue;
 364                 }
 365
 366                 if (!(req->rq_flags & PTL_RPC_FL_REPLIED)) {
 367                         DEBUG_REQ(D_HA, req, "keeping (in-flight)");
 368                         continue;
 369                 }
 370
 371                 /* not yet committed */
 372                 if (req->rq_transno > conn->c_last_committed)
 373                         break;
 374
 375                 DEBUG_REQ(D_HA, req, "committing (last_committed %Lu)",
 376                           (long long)conn->c_last_committed);
 377                 if (atomic_dec_and_test(&req->rq_refcount)) {
 378                         /* We do this to prevent free_req deadlock.
 379                          * Restarting after each removal is not so bad, as we are
 380                          * almost always deleting the first item in the list.
 381                          *
 382                          * If we use a recursive lock here, we can skip the
 383                          * unlock/lock/restart sequence.
 384                          */
 385                         spin_unlock(&conn->c_lock);
 386                         ptlrpc_free_req(req);
 387                         spin_lock(&conn->c_lock);
 388                         goto restart;
 389                 } else {
 390                         list_del(&req->rq_list);
 391                         list_add(&req->rq_list, &conn->c_dying_head);
 392                 }
 393         }
 394
 395         EXIT;
 396         return;
 397 }
 398
 399 void ptlrpc_cleanup_client(struct obd_import *imp)
 400 {
 401         struct list_head *tmp, *saved;
 402         struct ptlrpc_request *req;
 403         struct ptlrpc_connection *conn = imp->imp_connection;
 404         ENTRY;
 405
 406         LASSERT(conn);
 407
 408 restart1:
 409         spin_lock(&conn->c_lock);
 410         list_for_each_safe(tmp, saved, &conn->c_sending_head) {
 411                 req = list_entry(tmp, struct ptlrpc_request, rq_list);
 412                 if (req->rq_import != imp)
 413                         continue;
 414                 /* XXX we should make sure that nobody's sleeping on these! */
 415                 DEBUG_REQ(D_HA, req, "cleaning up from sending list");
 416                 list_del_init(&req->rq_list);
 417                 req->rq_import = NULL;
 418                 spin_unlock(&conn->c_lock);
 419                 ptlrpc_req_finished(req);
 420                 goto restart1;
 421         }
 422 restart2:
 423         list_for_each_safe(tmp, saved, &conn->c_dying_head) {
 424                 req = list_entry(tmp, struct ptlrpc_request, rq_list);
 425                 if (req->rq_import != imp)
 426                         continue;
 427                 DEBUG_REQ(D_ERROR, req, "on dying list at cleanup");
 428                 list_del_init(&req->rq_list);
 429                 req->rq_import = NULL;
 430                 spin_unlock(&conn->c_lock);
 431                 ptlrpc_req_finished(req);
 432                 spin_lock(&conn->c_lock);
 433                 goto restart2;
 434         }
 435         spin_unlock(&conn->c_lock);
 436
 437         EXIT;
 438         return;
 439 }
 440
 441 void ptlrpc_continue_req(struct ptlrpc_request *req)
 442 {
 443         ENTRY;
 444         CDEBUG(D_HA, "continue delayed request "LPD64" opc %d\n",
 445                req->rq_xid, req->rq_reqmsg->opc);
 446         req->rq_reqmsg->addr = req->rq_import->imp_handle.addr;
 447         req->rq_reqmsg->cookie = req->rq_import->imp_handle.cookie;
 448         wake_up(&req->rq_wait_for_rep);
 449         EXIT;
 450 }
 451
 452 void ptlrpc_resend_req(struct ptlrpc_request *req)
 453 {
 454         ENTRY;
 455         CDEBUG(D_HA, "resend request "LPD64", opc %d\n",
 456                req->rq_xid, req->rq_reqmsg->opc);
 457         req->rq_reqmsg->addr = req->rq_import->imp_handle.addr;
 458         req->rq_reqmsg->cookie = req->rq_import->imp_handle.cookie;
 459         req->rq_status = -EAGAIN;
 460         req->rq_level = LUSTRE_CONN_RECOVD;
 461         req->rq_flags |= PTL_RPC_FL_RESEND;
 462         req->rq_flags &= ~PTL_RPC_FL_TIMEOUT;
 463         wake_up(&req->rq_wait_for_rep);
 464         EXIT;
 465 }
 466
 467 void ptlrpc_restart_req(struct ptlrpc_request *req)
 468 {
 469         ENTRY;
 470         CDEBUG(D_HA, "restart completed request "LPD64", opc %d\n",
 471                req->rq_xid, req->rq_reqmsg->opc);
 472         req->rq_status = -ERESTARTSYS;
 473         req->rq_flags |= PTL_RPC_FL_RECOVERY;
 474         req->rq_flags &= ~PTL_RPC_FL_TIMEOUT;
 475         wake_up(&req->rq_wait_for_rep);
 476         EXIT;
 477 }
 478
 479 static int expired_request(void *data)
 480 {
 481         struct ptlrpc_request *req = data;
 482
 483         ENTRY;
 484         if (!req) {
 485                 CERROR("NULL req!");
 486                 LBUG();
 487                 RETURN(0);
 488         }
 489
 490         DEBUG_REQ(D_ERROR, req, "timeout");
 491         req->rq_flags |= PTL_RPC_FL_TIMEOUT;
 492
 493         if (!req->rq_import) {
 494                 DEBUG_REQ(D_ERROR, req, "NULL import");
 495                 LBUG();
 496                 RETURN(0);
 497         }
 498
 499         if (!req->rq_import->imp_connection) {
 500                 DEBUG_REQ(D_ERROR, req, "NULL connection");
 501                 LBUG();
 502                 RETURN(0);
 503         }
 504
 505         if (!req->rq_import->imp_connection->c_recovd_data.rd_recovd)
 506                 RETURN(1);
 507
 508         req->rq_timeout = 0;
 509         req->rq_connection->c_level = LUSTRE_CONN_RECOVD;
 510         recovd_conn_fail(req->rq_import->imp_connection);
 511
 512         /* If this request is for recovery or other primordial tasks,
 513          * don't go back to sleep.
 514          */
 515         if (req->rq_level < LUSTRE_CONN_FULL)
 516                 RETURN(1);
 517         RETURN(0);
 518 }
 519
 520 static int interrupted_request(void *data)
 521 {
 522         struct ptlrpc_request *req = data;
 523         ENTRY;
 524         req->rq_flags |= PTL_RPC_FL_INTR;
 525         RETURN(1); /* ignored, as of this writing */
 526 }
 527
 528 /* If we're being torn down by umount -f, or the import has been
 529  * invalidated (such as by an OST failure), the request must fail with
 530  * -EIO.
 531  *
 532  * Must be called with conn->c_lock held, will drop it if it returns -EIO.
 533  *
 534  * XXX this should just be testing the import, and umount_begin shouldn't touch
 535  * XXX the connection.
 536  */
 537 #define EIO_IF_INVALID(conn, req)                                             \
 538 if ((conn->c_flags & CONN_INVALID) ||                                         \
 539     (req->rq_import->imp_flags & IMP_INVALID)) {                              \
 540         DEBUG_REQ(D_ERROR, req, "%s_INVALID:",                                \
 541                   (conn->c_flags & CONN_INVALID) ? "CONN" : "IMP");           \
 542         spin_unlock(&conn->c_lock);                                           \
 543         RETURN(-EIO);                                                         \
 544 }
 545
 546 int ptlrpc_queue_wait(struct ptlrpc_request *req)
 547 {
 548         int rc = 0;
 549         struct l_wait_info lwi;
 550         struct ptlrpc_client *cli = req->rq_import->imp_client;
 551         struct ptlrpc_connection *conn = req->rq_import->imp_connection;
 552         ENTRY;
 553
 554         init_waitqueue_head(&req->rq_wait_for_rep);
 555         //        DEBUG_REQ(D_HA, req, "subsys: %s:", cli->cli_name);
 556
 557         /* XXX probably both an import and connection level are needed */
 558         if (req->rq_level > conn->c_level) {
 559                 spin_lock(&conn->c_lock);
 560                 EIO_IF_INVALID(conn, req);
 561                 list_del(&req->rq_list);
 562                 list_add_tail(&req->rq_list, &conn->c_delayed_head);
 563                 spin_unlock(&conn->c_lock);
 564
 565                 DEBUG_REQ(D_HA, req, "waiting for recovery: (%d < %d)",
 566                           req->rq_level, conn->c_level);
 567                 lwi = LWI_INTR(NULL, NULL);
 568                 rc = l_wait_event(req->rq_wait_for_rep,
 569                                   (req->rq_level <= conn->c_level) ||
 570                                   (req->rq_flags & PTL_RPC_FL_ERR), &lwi);
 571
 572                 spin_lock(&conn->c_lock);
 573                 list_del_init(&req->rq_list);
 574                 spin_unlock(&conn->c_lock);
 575
 576                 if (req->rq_flags & PTL_RPC_FL_ERR)
 577                         RETURN(-EIO);
 578
 579                 if (rc)
 580                         RETURN(rc);
 581
 582                 CERROR("process %d resumed\n", current->pid);
 583         }
 584  resend:
 585         req->rq_timeout = obd_timeout;
 586         spin_lock(&conn->c_lock);
 587         EIO_IF_INVALID(conn, req);
 588
 589         list_del(&req->rq_list);
 590         list_add_tail(&req->rq_list, &conn->c_sending_head);
 591         spin_unlock(&conn->c_lock);
 592         rc = ptl_send_rpc(req);
 593         if (rc) {
 594                 CDEBUG(D_HA, "error %d, opcode %d, need recovery\n", rc,
 595                        req->rq_reqmsg->opc);
 596                 /* the sleep below will time out, triggering recovery */
 597         }
 598
 599         DEBUG_REQ(D_NET, req, "-- sleeping");
 600         lwi = LWI_TIMEOUT_INTR(req->rq_timeout * HZ, expired_request,
 601                                interrupted_request, req);
 602         l_wait_event(req->rq_wait_for_rep, ptlrpc_check_reply(req), &lwi);
 603         DEBUG_REQ(D_NET, req, "-- done sleeping");
 604
 605         if (req->rq_flags & PTL_RPC_FL_ERR) {
 606                 ptlrpc_abort(req);
 607                 GOTO(out, rc = -EIO);
 608         }
 609
 610         /* Don't resend if we were interrupted. */
 611         if ((req->rq_flags & (PTL_RPC_FL_RESEND | PTL_RPC_FL_INTR)) ==
 612             PTL_RPC_FL_RESEND) {
 613                 req->rq_flags &= ~PTL_RPC_FL_RESEND;
 614                 DEBUG_REQ(D_HA, req, "resending: ");
 615                 goto resend;
 616         }
 617
 618         // up(&cli->cli_rpc_sem);
 619         if (req->rq_flags & PTL_RPC_FL_INTR) {
 620                 if (!(req->rq_flags & PTL_RPC_FL_TIMEOUT))
 621                         LBUG(); /* should only be interrupted if we timed out */
 622                 /* Clean up the dangling reply buffers */
 623                 ptlrpc_abort(req);
 624                 GOTO(out, rc = -EINTR);
 625         }
 626
 627         if (req->rq_flags & PTL_RPC_FL_TIMEOUT)
 628                 GOTO(out, rc = -ETIMEDOUT);
 629
 630         if (!(req->rq_flags & PTL_RPC_FL_REPLIED))
 631                 GOTO(out, rc = req->rq_status);
 632
 633         rc = lustre_unpack_msg(req->rq_repmsg, req->rq_replen);
 634         if (rc) {
 635                 CERROR("unpack_rep failed: %d\n", rc);
 636                 GOTO(out, rc);
 637         }
 638 #if 0
 639         /* FIXME: Enable when BlueArc makes new release */
 640         if (req->rq_repmsg->type != PTL_RPC_MSG_REPLY &&
 641             req->rq_repmsg->type != PTL_RPC_MSG_ERR) {
 642                 CERROR("invalid packet type received (type=%u)\n",
 643                        req->rq_repmsg->type);
 644                 LBUG();
 645                 GOTO(out, rc = -EINVAL);
 646         }
 647 #endif
 648         CDEBUG(D_NET, "got rep "LPD64"\n", req->rq_xid);
 649         if (req->rq_repmsg->status == 0)
 650                 CDEBUG(D_NET, "--> buf %p len %d status %d\n", req->rq_repmsg,
 651                        req->rq_replen, req->rq_repmsg->status);
 652
 653         spin_lock(&conn->c_lock);
 654
 655         /* Requests that aren't from replayable imports, or which don't have
 656          * transno information, can be "committed" early.
 657          */
 658
 659         if ((req->rq_import->imp_flags & IMP_REPLAYABLE) == 0 ||
 660             req->rq_repmsg->transno == 0) {
 661                 /* This import doesn't support replay, so we can just "commit"
 662                  * this request now.
 663                  */
 664                 DEBUG_REQ(D_HA, req, "not replayable, committing:");
 665                 list_del_init(&req->rq_list);
 666                 spin_unlock(&conn->c_lock);
 667                 ptlrpc_req_finished(req); /* Must be called unlocked. */
 668                 spin_lock(&conn->c_lock);
 669         }
 670
 671         /* Replay-enabled imports return commit-status information. */
 672         if (req->rq_import->imp_flags & IMP_REPLAYABLE) {
 673                 /* XXX this needs to be per-import, or multiple MDS services on
 674                  * XXX the same system are going to interfere messily with each
 675                  * XXX others' transno spaces.
 676                  */
 677                 conn->c_last_xid = req->rq_repmsg->last_xid;
 678                 conn->c_last_committed = req->rq_repmsg->last_committed;
 679                 ptlrpc_free_committed(conn);
 680         }
 681
 682         spin_unlock(&conn->c_lock);
 683
 684         EXIT;
 685  out:
 686         return rc;
 687 }
 688
 689 #undef EIO_IF_INVALID
 690
 691 int ptlrpc_replay_req(struct ptlrpc_request *req)
 692 {
 693         int rc = 0, old_level, old_status = 0;
 694         // struct ptlrpc_client *cli = req->rq_import->imp_client;
 695         struct l_wait_info lwi;
 696         ENTRY;
 697
 698         init_waitqueue_head(&req->rq_wait_for_rep);
 699         DEBUG_REQ(D_NET, req, "");
 700
 701         req->rq_timeout = obd_timeout;
 702         req->rq_reqmsg->addr = req->rq_import->imp_handle.addr;
 703         req->rq_reqmsg->cookie = req->rq_import->imp_handle.cookie;
 704
 705         /* temporarily set request to RECOVD level (reset at out:) */
 706         old_level = req->rq_level;
 707         if (req->rq_flags & PTL_RPC_FL_REPLIED)
 708                 old_status = req->rq_repmsg->status;
 709         req->rq_level = LUSTRE_CONN_RECOVD;
 710         rc = ptl_send_rpc(req);
 711         if (rc) {
 712                 CERROR("error %d, opcode %d\n", rc, req->rq_reqmsg->opc);
 713                 ptlrpc_cleanup_request_buf(req);
 714                 // up(&cli->cli_rpc_sem);
 715                 GOTO(out, rc = -rc);
 716         }
 717
 718         CDEBUG(D_OTHER, "-- sleeping\n");
 719         lwi = LWI_INTR(NULL, NULL); /* XXX needs timeout, nested recovery */
 720         l_wait_event(req->rq_wait_for_rep, ptlrpc_check_reply(req), &lwi);
 721         CDEBUG(D_OTHER, "-- done\n");
 722
 723         // up(&cli->cli_rpc_sem);
 724
 725         if (!(req->rq_flags & PTL_RPC_FL_REPLIED)) {
 726                 CERROR("Unknown reason for wakeup\n");
 727                 /* XXX Phil - I end up here when I kill obdctl */
 728                 ptlrpc_abort(req);
 729                 GOTO(out, rc = -EINTR);
 730         }
 731
 732         rc = lustre_unpack_msg(req->rq_repmsg, req->rq_replen);
 733         if (rc) {
 734                 CERROR("unpack_rep failed: %d\n", rc);
 735                 GOTO(out, rc);
 736         }
 737
 738         CDEBUG(D_NET, "got rep "LPD64"\n", req->rq_xid);
 739
 740         /* let the callback do fixups, possibly including in the request */
 741         if (req->rq_replay_cb)
 742                 req->rq_replay_cb(req);
 743
 744         if ((req->rq_flags & PTL_RPC_FL_REPLIED) &&
 745             req->rq_repmsg->status != old_status) {
 746                 DEBUG_REQ(D_HA, req, "status %d, old was %d",
 747                           req->rq_repmsg->status, old_status);
 748         }
 749
 750  out:
 751         req->rq_level = old_level;
 752         RETURN(rc);
 753 }