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                 INIT_LIST_HEAD(&desc->bd_set_chain);
  93                 ptl_set_inv_handle(&desc->bd_md_h);
  94                 ptl_set_inv_handle(&desc->bd_me_h);
  95         }
  96
  97         return desc;
  98 }
  99
 100 int ptlrpc_bulk_error(struct ptlrpc_bulk_desc *desc)
 101 {
 102         int rc = 0;
 103         if (desc->bd_flags & PTL_RPC_FL_TIMEOUT) {
 104                 rc = (desc->bd_flags & PTL_RPC_FL_INTR ? -ERESTARTSYS :
 105                       -ETIMEDOUT);
 106         }
 107         return rc;
 108 }
 109
 110 struct ptlrpc_bulk_page *ptlrpc_prep_bulk_page(struct ptlrpc_bulk_desc *desc)
 111 {
 112         struct ptlrpc_bulk_page *bulk;
 113
 114         OBD_ALLOC(bulk, sizeof(*bulk));
 115         if (bulk != NULL) {
 116                 bulk->bp_desc = desc;
 117                 list_add_tail(&bulk->bp_link, &desc->bd_page_list);
 118                 desc->bd_page_count++;
 119         }
 120         return bulk;
 121 }
 122
 123 void ptlrpc_free_bulk(struct ptlrpc_bulk_desc *desc)
 124 {
 125         struct list_head *tmp, *next;
 126         ENTRY;
 127         if (desc == NULL) {
 128                 EXIT;
 129                 return;
 130         }
 131
 132         LASSERT(list_empty(&desc->bd_set_chain));
 133
 134         list_for_each_safe(tmp, next, &desc->bd_page_list) {
 135                 struct ptlrpc_bulk_page *bulk;
 136                 bulk = list_entry(tmp, struct ptlrpc_bulk_page, bp_link);
 137                 ptlrpc_free_bulk_page(bulk);
 138         }
 139
 140         ptlrpc_put_connection(desc->bd_connection);
 141
 142         OBD_FREE(desc, sizeof(*desc));
 143         EXIT;
 144 }
 145
 146 void ptlrpc_free_bulk_page(struct ptlrpc_bulk_page *bulk)
 147 {
 148         ENTRY;
 149         if (bulk == NULL) {
 150                 EXIT;
 151                 return;
 152         }
 153
 154         list_del(&bulk->bp_link);
 155         bulk->bp_desc->bd_page_count--;
 156         OBD_FREE(bulk, sizeof(*bulk));
 157         EXIT;
 158 }
 159
 160 static int ll_sync_brw_timeout(void *data)
 161 {
 162         struct obd_brw_set *set = data;
 163         struct list_head *tmp;
 164         int failed = 0;
 165         ENTRY;
 166
 167         LASSERT(set);
 168
 169         set->brw_flags |= PTL_RPC_FL_TIMEOUT;
 170
 171         list_for_each(tmp, &set->brw_desc_head) {
 172                 struct ptlrpc_bulk_desc *desc =
 173                         list_entry(tmp, struct ptlrpc_bulk_desc, bd_set_chain);
 174
 175                 /* Skip descriptors that were completed successfully. */
 176                 if (desc->bd_flags & (PTL_BULK_FL_RCVD | PTL_BULK_FL_SENT))
 177                         continue;
 178
 179                 LASSERT(desc->bd_connection);
 180
 181                 /* If PtlMDUnlink succeeds, then it hasn't completed yet.  If it
 182                  * fails, the bulk finished _just_ in time (after the timeout
 183                  * fired but before we got this far) and we'll let it live.
 184                  */
 185                 if (PtlMDUnlink(desc->bd_md_h) != 0) {
 186                         CERROR("Near-miss on OST %s -- need to adjust "
 187                                "obd_timeout?\n",
 188                                desc->bd_connection->c_remote_uuid);
 189                         continue;
 190                 }
 191
 192                 CERROR("IO of %d pages to/from %s:%d (conn %p) timed out\n",
 193                        desc->bd_page_count, desc->bd_connection->c_remote_uuid,
 194                        desc->bd_portal, desc->bd_connection);
 195
 196                 /* This one will "never" arrive, don't wait for it. */
 197                 if (atomic_dec_and_test(&set->brw_refcount))
 198                         wake_up(&set->brw_waitq);
 199
 200                 if (class_signal_connection_failure)
 201                         class_signal_connection_failure(desc->bd_connection);
 202                 else
 203                         failed = 1;
 204         }
 205
 206         /* 0 = We go back to sleep, until we're resumed or interrupted */
 207         /* 1 = We can't be recovered, just abort the syscall with -ETIMEDOUT */
 208         RETURN(failed);
 209 }
 210
 211 static int ll_sync_brw_intr(void *data)
 212 {
 213         struct obd_brw_set *set = data;
 214
 215         ENTRY;
 216         set->brw_flags |= PTL_RPC_FL_INTR;
 217         RETURN(1); /* ignored, as of this writing */
 218 }
 219
 220 int ll_brw_sync_wait(struct obd_brw_set *set, int phase)
 221 {
 222         struct l_wait_info lwi;
 223         struct list_head *tmp, *next;
 224         int rc = 0;
 225         ENTRY;
 226
 227         switch(phase) {
 228         case CB_PHASE_START:
 229                 lwi = LWI_TIMEOUT_INTR(obd_timeout * HZ, ll_sync_brw_timeout,
 230                                        ll_sync_brw_intr, set);
 231                 rc = l_wait_event(set->brw_waitq,
 232                                   atomic_read(&set->brw_refcount) == 0, &lwi);
 233
 234                 list_for_each_safe(tmp, next, &set->brw_desc_head) {
 235                         struct ptlrpc_bulk_desc *desc =
 236                                 list_entry(tmp, struct ptlrpc_bulk_desc,
 237                                            bd_set_chain);
 238                         list_del_init(&desc->bd_set_chain);
 239                         ptlrpc_bulk_decref(desc);
 240                 }
 241                 break;
 242         case CB_PHASE_FINISH:
 243                 if (atomic_dec_and_test(&set->brw_refcount))
 244                         wake_up(&set->brw_waitq);
 245                 break;
 246         default:
 247                 LBUG();
 248         }
 249
 250         RETURN(rc);
 251 }
 252
 253 struct ptlrpc_request *ptlrpc_prep_req(struct obd_import *imp, int opcode,
 254                                        int count, int *lengths, char **bufs)
 255 {
 256         struct ptlrpc_connection *conn = imp->imp_connection;
 257         struct ptlrpc_request *request;
 258         int rc;
 259         ENTRY;
 260
 261         OBD_ALLOC(request, sizeof(*request));
 262         if (!request) {
 263                 CERROR("request allocation out of memory\n");
 264                 RETURN(NULL);
 265         }
 266
 267         rc = lustre_pack_msg(count, lengths, bufs,
 268                              &request->rq_reqlen, &request->rq_reqmsg);
 269         if (rc) {
 270                 CERROR("cannot pack request %d\n", rc);
 271                 OBD_FREE(request, sizeof(*request));
 272                 RETURN(NULL);
 273         }
 274
 275         request->rq_level = LUSTRE_CONN_FULL;
 276         request->rq_type = PTL_RPC_MSG_REQUEST;
 277         request->rq_import = imp;
 278
 279         /* XXX FIXME bug 625069 */
 280         request->rq_request_portal = imp->imp_client->cli_request_portal;
 281         request->rq_reply_portal = imp->imp_client->cli_reply_portal;
 282
 283         request->rq_connection = ptlrpc_connection_addref(conn);
 284
 285         INIT_LIST_HEAD(&request->rq_list);
 286         atomic_set(&request->rq_refcount, 1);
 287
 288         spin_lock(&imp->imp_lock);
 289         request->rq_xid = HTON__u32(++imp->imp_last_xid);
 290         spin_unlock(&imp->imp_lock);
 291
 292         request->rq_reqmsg->magic = PTLRPC_MSG_MAGIC;
 293         request->rq_reqmsg->version = PTLRPC_MSG_VERSION;
 294         request->rq_reqmsg->opc = HTON__u32(opcode);
 295         request->rq_reqmsg->flags = 0;
 296
 297         ptlrpc_hdl2req(request, &imp->imp_handle);
 298         RETURN(request);
 299 }
 300
 301 static void __ptlrpc_free_req(struct ptlrpc_request *request, int locked)
 302 {
 303         ENTRY;
 304         if (request == NULL) {
 305                 EXIT;
 306                 return;
 307         }
 308
 309         if (atomic_read(&request->rq_refcount) != 0) {
 310                 CERROR("freeing request %p (%d->%s:%d) with refcount %d\n",
 311                        request, request->rq_reqmsg->opc,
 312                        request->rq_connection->c_remote_uuid,
 313                        request->rq_import->imp_client->cli_request_portal,
 314                        request->rq_refcount);
 315                 /* LBUG(); */
 316         }
 317
 318         if (request->rq_repmsg != NULL) {
 319                 OBD_FREE(request->rq_repmsg, request->rq_replen);
 320                 request->rq_repmsg = NULL;
 321                 request->rq_reply_md.start = NULL;
 322         }
 323         if (request->rq_reqmsg != NULL) {
 324                 OBD_FREE(request->rq_reqmsg, request->rq_reqlen);
 325                 request->rq_reqmsg = NULL;
 326         }
 327
 328         if (request->rq_import) {
 329                 if (!locked)
 330                         spin_lock(&request->rq_import->imp_lock);
 331                 list_del_init(&request->rq_list);
 332                 if (!locked)
 333                         spin_unlock(&request->rq_import->imp_lock);
 334         }
 335
 336         ptlrpc_put_connection(request->rq_connection);
 337         OBD_FREE(request, sizeof(*request));
 338         EXIT;
 339 }
 340
 341 void ptlrpc_free_req(struct ptlrpc_request *request)
 342 {
 343         __ptlrpc_free_req(request, 0);
 344 }
 345
 346 static int __ptlrpc_req_finished(struct ptlrpc_request *request, int locked)
 347 {
 348         ENTRY;
 349         if (request == NULL)
 350                 RETURN(1);
 351
 352         if (atomic_dec_and_test(&request->rq_refcount)) {
 353                 __ptlrpc_free_req(request, locked);
 354                 RETURN(1);
 355         }
 356
 357         DEBUG_REQ(D_INFO, request, "refcount now %u",
 358                   atomic_read(&request->rq_refcount));
 359         RETURN(0);
 360 }
 361
 362 void ptlrpc_req_finished(struct ptlrpc_request *request)
 363 {
 364         __ptlrpc_req_finished(request, 0);
 365 }
 366
 367 static int ptlrpc_check_reply(struct ptlrpc_request *req)
 368 {
 369         int rc = 0;
 370
 371         ENTRY;
 372         if (req->rq_repmsg != NULL) {
 373                 req->rq_transno = NTOH__u64(req->rq_repmsg->transno);
 374                 req->rq_flags |= PTL_RPC_FL_REPLIED;
 375                 GOTO(out, rc = 1);
 376         }
 377
 378         if (req->rq_flags & PTL_RPC_FL_RESEND) {
 379                 ENTRY;
 380                 DEBUG_REQ(D_ERROR, req, "RESEND:");
 381                 GOTO(out, rc = 1);
 382         }
 383
 384         if (req->rq_flags & PTL_RPC_FL_ERR) {
 385                 ENTRY;
 386                 DEBUG_REQ(D_ERROR, req, "ABORTED:");
 387                 GOTO(out, rc = 1);
 388         }
 389
 390         if (req->rq_flags & PTL_RPC_FL_RESTART) {
 391                 DEBUG_REQ(D_ERROR, req, "RESTART:");
 392                 GOTO(out, rc = 1);
 393         }
 394         EXIT;
 395  out:
 396         DEBUG_REQ(D_NET, req, "rc = %d for", rc);
 397         return rc;
 398 }
 399
 400 static int ptlrpc_check_status(struct ptlrpc_request *req)
 401 {
 402         int err;
 403         ENTRY;
 404
 405         err = req->rq_repmsg->status;
 406         if (req->rq_repmsg->type == NTOH__u32(PTL_RPC_MSG_ERR)) {
 407                 DEBUG_REQ(D_ERROR, req, "type == PTL_RPC_MSG_ERR");
 408                 RETURN(err ? err : -EINVAL);
 409         }
 410
 411         if (err < 0) {
 412                 DEBUG_REQ(D_ERROR, req, "status is %d", err);
 413         } else if (err > 0) {
 414                 /* XXX: translate this error from net to host */
 415                 DEBUG_REQ(D_INFO, req, "status is %d", err);
 416         }
 417
 418         RETURN(err);
 419 }
 420
 421 static void ptlrpc_cleanup_request_buf(struct ptlrpc_request *request)
 422 {
 423         OBD_FREE(request->rq_reqmsg, request->rq_reqlen);
 424         request->rq_reqmsg = NULL;
 425         request->rq_reqlen = 0;
 426 }
 427
 428 /* Abort this request and cleanup any resources associated with it. */
 429 static int ptlrpc_abort(struct ptlrpc_request *request)
 430 {
 431         /* First remove the ME for the reply; in theory, this means
 432          * that we can tear down the buffer safely. */
 433         PtlMEUnlink(request->rq_reply_me_h);
 434         OBD_FREE(request->rq_reply_md.start, request->rq_replen);
 435         request->rq_repmsg = NULL;
 436         request->rq_replen = 0;
 437         return 0;
 438 }
 439
 440 /* caller must hold imp->imp_lock */
 441 void ptlrpc_free_committed(struct obd_import *imp)
 442 {
 443         struct list_head *tmp, *saved;
 444         struct ptlrpc_request *req;
 445         ENTRY;
 446
 447 #ifndef __arch_um__
 448         LASSERT(spin_is_locked(&imp->imp_lock));
 449 #endif
 450
 451         CDEBUG(D_HA, "committing for xid "LPU64", last_committed "LPU64"\n",
 452                imp->imp_peer_last_xid, imp->imp_peer_committed_transno);
 453
 454         list_for_each_safe(tmp, saved, &imp->imp_replay_list) {
 455                 req = list_entry(tmp, struct ptlrpc_request, rq_list);
 456
 457                 if (req->rq_flags & PTL_RPC_FL_REPLAY) {
 458                         DEBUG_REQ(D_HA, req, "keeping (FL_REPLAY)");
 459                         continue;
 460                 }
 461
 462                 /* not yet committed */
 463                 if (req->rq_transno > imp->imp_peer_committed_transno) {
 464                         DEBUG_REQ(D_HA, req, "stopping search");
 465                         break;
 466                 }
 467
 468                 DEBUG_REQ(D_HA, req, "committing (last_committed "LPU64")",
 469                           imp->imp_peer_committed_transno);
 470                 __ptlrpc_req_finished(req, 1);
 471         }
 472
 473         EXIT;
 474         return;
 475 }
 476
 477 void ptlrpc_cleanup_client(struct obd_import *imp)
 478 {
 479         struct list_head *tmp, *saved;
 480         struct ptlrpc_request *req;
 481         struct ptlrpc_connection *conn = imp->imp_connection;
 482         ENTRY;
 483
 484         LASSERT(conn);
 485
 486         spin_lock(&imp->imp_lock);
 487         list_for_each_safe(tmp, saved, &imp->imp_replay_list) {
 488                 req = list_entry(tmp, struct ptlrpc_request, rq_list);
 489
 490                 /* XXX we should make sure that nobody's sleeping on these! */
 491                 DEBUG_REQ(D_HA, req, "cleaning up from sending list");
 492                 list_del_init(&req->rq_list);
 493                 req->rq_import = NULL;
 494                 __ptlrpc_req_finished(req, 0);
 495         }
 496         spin_unlock(&imp->imp_lock);
 497
 498         EXIT;
 499         return;
 500 }
 501
 502 void ptlrpc_continue_req(struct ptlrpc_request *req)
 503 {
 504         ENTRY;
 505         DEBUG_REQ(D_HA, req, "continuing delayed request");
 506         req->rq_reqmsg->addr = req->rq_import->imp_handle.addr;
 507         req->rq_reqmsg->cookie = req->rq_import->imp_handle.cookie;
 508         wake_up(&req->rq_wait_for_rep);
 509         EXIT;
 510 }
 511
 512 void ptlrpc_resend_req(struct ptlrpc_request *req)
 513 {
 514         ENTRY;
 515         DEBUG_REQ(D_HA, req, "resending");
 516         req->rq_reqmsg->addr = req->rq_import->imp_handle.addr;
 517         req->rq_reqmsg->cookie = req->rq_import->imp_handle.cookie;
 518         req->rq_status = -EAGAIN;
 519         req->rq_level = LUSTRE_CONN_RECOVD;
 520         req->rq_flags |= PTL_RPC_FL_RESEND;
 521         req->rq_flags &= ~PTL_RPC_FL_TIMEOUT;
 522         wake_up(&req->rq_wait_for_rep);
 523         EXIT;
 524 }
 525
 526 void ptlrpc_restart_req(struct ptlrpc_request *req)
 527 {
 528         ENTRY;
 529         DEBUG_REQ(D_HA, req, "restarting (possibly-)completed request");
 530         req->rq_status = -ERESTARTSYS;
 531         req->rq_flags |= PTL_RPC_FL_RESTART;
 532         req->rq_flags &= ~PTL_RPC_FL_TIMEOUT;
 533         wake_up(&req->rq_wait_for_rep);
 534         EXIT;
 535 }
 536
 537 static int expired_request(void *data)
 538 {
 539         struct ptlrpc_request *req = data;
 540
 541         ENTRY;
 542         if (!req) {
 543                 CERROR("NULL req!");
 544                 LBUG();
 545                 RETURN(0);
 546         }
 547
 548         DEBUG_REQ(D_ERROR, req, "timeout");
 549         req->rq_flags |= PTL_RPC_FL_TIMEOUT;
 550
 551         if (!req->rq_import) {
 552                 DEBUG_REQ(D_ERROR, req, "NULL import");
 553                 LBUG();
 554                 RETURN(0);
 555         }
 556
 557         if (!req->rq_import->imp_connection) {
 558                 DEBUG_REQ(D_ERROR, req, "NULL connection");
 559                 LBUG();
 560                 RETURN(0);
 561         }
 562
 563         if (!req->rq_import->imp_connection->c_recovd_data.rd_recovd)
 564                 RETURN(1);
 565
 566         req->rq_timeout = 0;
 567         recovd_conn_fail(req->rq_import->imp_connection);
 568
 569 #if 0
 570         /* If this request is for recovery or other primordial tasks,
 571          * don't go back to sleep.
 572          */
 573         if (req->rq_level < LUSTRE_CONN_FULL)
 574                 RETURN(1);
 575 #endif
 576         RETURN(0);
 577 }
 578
 579 static int interrupted_request(void *data)
 580 {
 581         struct ptlrpc_request *req = data;
 582         ENTRY;
 583         req->rq_flags |= PTL_RPC_FL_INTR;
 584         RETURN(1); /* ignored, as of this writing */
 585 }
 586
 587 /* If the import has been invalidated (such as by an OST failure), the
 588  * request must fail with -EIO.
 589  *
 590  * Must be called with imp_lock held, will drop it if it returns -EIO.
 591  */
 592 #define EIO_IF_INVALID(req)                                                   \
 593 if (req->rq_import->imp_flags & IMP_INVALID) {                                \
 594         DEBUG_REQ(D_ERROR, req, "IMP_INVALID:");                              \
 595         spin_unlock(&imp->imp_lock);                                          \
 596         RETURN(-EIO);                                                         \
 597 }
 598
 599 int ptlrpc_queue_wait(struct ptlrpc_request *req)
 600 {
 601         int rc = 0;
 602         struct l_wait_info lwi;
 603         struct obd_import *imp = req->rq_import;
 604         struct ptlrpc_connection *conn = imp->imp_connection;
 605         ENTRY;
 606
 607         init_waitqueue_head(&req->rq_wait_for_rep);
 608
 609         /* for distributed debugging */
 610         req->rq_reqmsg->status = HTON__u32(current->pid);
 611         CDEBUG(D_RPCTRACE, "Sending RPC pid:xid:nid:opc %d:"LPU64":%x:%d\n",
 612                NTOH__u32(req->rq_reqmsg->status), req->rq_xid,
 613                conn->c_peer.peer_nid, NTOH__u32(req->rq_reqmsg->opc));
 614
 615         if (req->rq_level > imp->imp_level) {
 616                 spin_lock(&imp->imp_lock);
 617                 EIO_IF_INVALID(req);
 618                 list_del(&req->rq_list);
 619                 list_add_tail(&req->rq_list, &imp->imp_delayed_list);
 620                 spin_unlock(&imp->imp_lock);
 621
 622                 DEBUG_REQ(D_HA, req, "\"%s\" waiting for recovery: (%d < %d)",
 623                           current->comm, req->rq_level, imp->imp_level);
 624                 lwi = LWI_INTR(NULL, NULL);
 625                 rc = l_wait_event(req->rq_wait_for_rep,
 626                                   (req->rq_level <= imp->imp_level) ||
 627                                   (req->rq_flags & PTL_RPC_FL_ERR), &lwi);
 628
 629                 spin_lock(&imp->imp_lock);
 630                 list_del_init(&req->rq_list);
 631                 spin_unlock(&imp->imp_lock);
 632
 633                 if (req->rq_flags & PTL_RPC_FL_ERR)
 634                         RETURN(-EIO);
 635
 636                 if (rc)
 637                         RETURN(rc);
 638
 639                 CERROR("process %d resumed\n", current->pid);
 640         }
 641  resend:
 642         req->rq_timeout = obd_timeout;
 643         spin_lock(&imp->imp_lock);
 644         EIO_IF_INVALID(req);
 645
 646         LASSERT(list_empty(&req->rq_list));
 647         list_add_tail(&req->rq_list, &imp->imp_sending_list);
 648         spin_unlock(&imp->imp_lock);
 649         rc = ptl_send_rpc(req);
 650         if (rc) {
 651                 CDEBUG(D_HA, "error %d, opcode %d, need recovery\n", rc,
 652                        req->rq_reqmsg->opc);
 653                 /* sleep for a jiffy, then trigger recovery */
 654                 lwi = LWI_TIMEOUT_INTR(1, expired_request,
 655                                        interrupted_request, req);
 656         } else {
 657                 DEBUG_REQ(D_NET, req, "-- sleeping");
 658                 lwi = LWI_TIMEOUT_INTR(req->rq_timeout * HZ, expired_request,
 659                                        interrupted_request, req);
 660         }
 661         l_wait_event(req->rq_wait_for_rep, ptlrpc_check_reply(req), &lwi);
 662         DEBUG_REQ(D_NET, req, "-- done sleeping");
 663
 664         spin_lock(&imp->imp_lock);
 665         list_del_init(&req->rq_list);
 666         spin_unlock(&imp->imp_lock);
 667
 668         if (req->rq_flags & PTL_RPC_FL_ERR) {
 669                 ptlrpc_abort(req);
 670                 GOTO(out, rc = -EIO);
 671         }
 672
 673         /* Don't resend if we were interrupted. */
 674         if ((req->rq_flags & (PTL_RPC_FL_RESEND | PTL_RPC_FL_INTR)) ==
 675             PTL_RPC_FL_RESEND) {
 676                 req->rq_flags &= ~PTL_RPC_FL_RESEND;
 677                 DEBUG_REQ(D_HA, req, "resending: ");
 678                 goto resend;
 679         }
 680
 681         if (req->rq_flags & PTL_RPC_FL_INTR) {
 682                 if (!(req->rq_flags & PTL_RPC_FL_TIMEOUT))
 683                         LBUG(); /* should only be interrupted if we timed out */
 684                 /* Clean up the dangling reply buffers */
 685                 ptlrpc_abort(req);
 686                 GOTO(out, rc = -EINTR);
 687         }
 688
 689         if (req->rq_flags & PTL_RPC_FL_TIMEOUT)
 690                 GOTO(out, rc = -ETIMEDOUT);
 691
 692         if (!(req->rq_flags & PTL_RPC_FL_REPLIED))
 693                 GOTO(out, rc = req->rq_status);
 694
 695         rc = lustre_unpack_msg(req->rq_repmsg, req->rq_replen);
 696         if (rc) {
 697                 CERROR("unpack_rep failed: %d\n", rc);
 698                 GOTO(out, rc);
 699         }
 700 #if 0
 701         /* FIXME: Enable when BlueArc makes new release */
 702         if (req->rq_repmsg->type != PTL_RPC_MSG_REPLY &&
 703             req->rq_repmsg->type != PTL_RPC_MSG_ERR) {
 704                 CERROR("invalid packet type received (type=%u)\n",
 705                        req->rq_repmsg->type);
 706                 LBUG();
 707                 GOTO(out, rc = -EINVAL);
 708         }
 709 #endif
 710         CDEBUG(D_NET, "got rep "LPU64"\n", req->rq_xid);
 711         if (req->rq_repmsg->status == 0)
 712                 CDEBUG(D_NET, "--> buf %p len %d status %d\n", req->rq_repmsg,
 713                        req->rq_replen, req->rq_repmsg->status);
 714
 715
 716         if (req->rq_import->imp_flags & IMP_REPLAYABLE) {
 717                 spin_lock(&imp->imp_lock);
 718                 if (req->rq_flags & PTL_RPC_FL_REPLAY || req->rq_transno != 0) {
 719                         /* Balanced in ptlrpc_free_committed, usually. */
 720                         atomic_inc(&req->rq_refcount);
 721                         list_add_tail(&req->rq_list, &imp->imp_replay_list);
 722                 }
 723
 724                 if (req->rq_transno > imp->imp_max_transno) {
 725                         imp->imp_max_transno = req->rq_transno;
 726                 } else if (req->rq_transno != 0 &&
 727                            imp->imp_level == LUSTRE_CONN_FULL) {
 728                         CERROR("got transno "LPD64" after "LPD64": recovery "
 729                                "may not work\n", req->rq_transno,
 730                                imp->imp_max_transno);
 731                 }
 732
 733                 /* Replay-enabled imports return commit-status information. */
 734                 imp->imp_peer_last_xid = req->rq_repmsg->last_xid;
 735                 imp->imp_peer_committed_transno =
 736                         req->rq_repmsg->last_committed;
 737                 ptlrpc_free_committed(imp);
 738                 spin_unlock(&imp->imp_lock);
 739         }
 740
 741         rc = ptlrpc_check_status(req);
 742
 743         EXIT;
 744  out:
 745         return rc;
 746 }
 747
 748 #undef EIO_IF_INVALID
 749
 750 int ptlrpc_replay_req(struct ptlrpc_request *req)
 751 {
 752         int rc = 0, old_level, old_status = 0;
 753         // struct ptlrpc_client *cli = req->rq_import->imp_client;
 754         struct l_wait_info lwi;
 755         ENTRY;
 756
 757         init_waitqueue_head(&req->rq_wait_for_rep);
 758         DEBUG_REQ(D_NET, req, "");
 759
 760         req->rq_timeout = obd_timeout;
 761         req->rq_reqmsg->addr = req->rq_import->imp_handle.addr;
 762         req->rq_reqmsg->cookie = req->rq_import->imp_handle.cookie;
 763
 764         /* temporarily set request to RECOVD level (reset at out:) */
 765         old_level = req->rq_level;
 766         if (req->rq_flags & PTL_RPC_FL_REPLIED)
 767                 old_status = req->rq_repmsg->status;
 768         req->rq_level = LUSTRE_CONN_RECOVD;
 769         rc = ptl_send_rpc(req);
 770         if (rc) {
 771                 CERROR("error %d, opcode %d\n", rc, req->rq_reqmsg->opc);
 772                 ptlrpc_cleanup_request_buf(req);
 773                 // up(&cli->cli_rpc_sem);
 774                 GOTO(out, rc = -rc);
 775         }
 776
 777         CDEBUG(D_OTHER, "-- sleeping\n");
 778         lwi = LWI_INTR(NULL, NULL); /* XXX needs timeout, nested recovery */
 779         l_wait_event(req->rq_wait_for_rep, ptlrpc_check_reply(req), &lwi);
 780         CDEBUG(D_OTHER, "-- done\n");
 781
 782         // up(&cli->cli_rpc_sem);
 783
 784         if (!(req->rq_flags & PTL_RPC_FL_REPLIED)) {
 785                 CERROR("Unknown reason for wakeup\n");
 786                 /* XXX Phil - I end up here when I kill obdctl */
 787                 ptlrpc_abort(req);
 788                 GOTO(out, rc = -EINTR);
 789         }
 790
 791         rc = lustre_unpack_msg(req->rq_repmsg, req->rq_replen);
 792         if (rc) {
 793                 CERROR("unpack_rep failed: %d\n", rc);
 794                 GOTO(out, rc);
 795         }
 796
 797         CDEBUG(D_NET, "got rep "LPD64"\n", req->rq_xid);
 798
 799         /* let the callback do fixups, possibly including in the request */
 800         if (req->rq_replay_cb)
 801                 req->rq_replay_cb(req);
 802
 803         if ((req->rq_flags & PTL_RPC_FL_REPLIED) &&
 804             req->rq_repmsg->status != old_status) {
 805                 DEBUG_REQ(D_HA, req, "status %d, old was %d",
 806                           req->rq_repmsg->status, old_status);
 807         }
 808
 809  out:
 810         req->rq_level = old_level;
 811         RETURN(rc);
 812 }