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                 desc->bd_connection->c_level = LUSTRE_CONN_RECOVD;
 196
 197                 /* This one will "never" arrive, don't wait for it. */
 198                 if (atomic_dec_and_test(&set->brw_refcount))
 199                         wake_up(&set->brw_waitq);
 200
 201                 if (class_signal_connection_failure)
 202                         class_signal_connection_failure(desc->bd_connection);
 203                 else
 204                         failed = 1;
 205         }
 206
 207         /* 0 = We go back to sleep, until we're resumed or interrupted */
 208         /* 1 = We can't be recovered, just abort the syscall with -ETIMEDOUT */
 209         RETURN(failed);
 210 }
 211
 212 static int ll_sync_brw_intr(void *data)
 213 {
 214         struct obd_brw_set *set = data;
 215
 216         ENTRY;
 217         set->brw_flags |= PTL_RPC_FL_INTR;
 218         RETURN(1); /* ignored, as of this writing */
 219 }
 220
 221 int ll_brw_sync_wait(struct obd_brw_set *set, int phase)
 222 {
 223         struct l_wait_info lwi;
 224         struct list_head *tmp, *next;
 225         int rc = 0;
 226         ENTRY;
 227
 228         switch(phase) {
 229         case CB_PHASE_START:
 230                 lwi = LWI_TIMEOUT_INTR(obd_timeout * HZ, ll_sync_brw_timeout,
 231                                        ll_sync_brw_intr, set);
 232                 rc = l_wait_event(set->brw_waitq,
 233                                   atomic_read(&set->brw_refcount) == 0, &lwi);
 234
 235                 list_for_each_safe(tmp, next, &set->brw_desc_head) {
 236                         struct ptlrpc_bulk_desc *desc =
 237                                 list_entry(tmp, struct ptlrpc_bulk_desc,
 238                                            bd_set_chain);
 239                         list_del_init(&desc->bd_set_chain);
 240                         ptlrpc_bulk_decref(desc);
 241                 }
 242                 break;
 243         case CB_PHASE_FINISH:
 244                 if (atomic_dec_and_test(&set->brw_refcount))
 245                         wake_up(&set->brw_waitq);
 246                 break;
 247         default:
 248                 LBUG();
 249         }
 250
 251         RETURN(rc);
 252 }
 253
 254 struct ptlrpc_request *ptlrpc_prep_req(struct obd_import *imp, int opcode,
 255                                        int count, int *lengths, char **bufs)
 256 {
 257         struct ptlrpc_connection *conn = imp->imp_connection;
 258         struct ptlrpc_request *request;
 259         int rc;
 260         ENTRY;
 261
 262         OBD_ALLOC(request, sizeof(*request));
 263         if (!request) {
 264                 CERROR("request allocation out of memory\n");
 265                 RETURN(NULL);
 266         }
 267
 268         rc = lustre_pack_msg(count, lengths, bufs,
 269                              &request->rq_reqlen, &request->rq_reqmsg);
 270         if (rc) {
 271                 CERROR("cannot pack request %d\n", rc);
 272                 OBD_FREE(request, sizeof(*request));
 273                 RETURN(NULL);
 274         }
 275
 276         request->rq_level = LUSTRE_CONN_FULL;
 277         request->rq_type = PTL_RPC_MSG_REQUEST;
 278         request->rq_import = imp;
 279
 280         /* XXX FIXME bug 625069 */
 281         request->rq_request_portal = imp->imp_client->cli_request_portal;
 282         request->rq_reply_portal = imp->imp_client->cli_reply_portal;
 283
 284         request->rq_connection = ptlrpc_connection_addref(conn);
 285
 286         INIT_LIST_HEAD(&request->rq_list);
 287         /*
 288          * This will be reduced once when the sender is finished (waiting for
 289          * reply, f.e.), and once when the request has been committed and is
 290          * removed from the to-be-committed list.
 291          *
 292          * Also, the refcount will be increased in ptl_send_rpc immediately
 293          * before we hand it off to portals, and there will be a corresponding
 294          * decrease in request_out_cb (which is called to indicate that portals
 295          * is finished with the request, and it can be safely freed).
 296          *
 297          * (Except in the DLM server case, where it will be dropped twice
 298          * by the sender, and then the last time by request_out_callback.)
 299          */
 300         atomic_set(&request->rq_refcount, 2);
 301
 302         spin_lock(&imp->imp_lock);
 303         request->rq_xid = HTON__u32(++imp->imp_last_xid);
 304         spin_unlock(&imp->imp_lock);
 305
 306         request->rq_reqmsg->magic = PTLRPC_MSG_MAGIC;
 307         request->rq_reqmsg->version = PTLRPC_MSG_VERSION;
 308         request->rq_reqmsg->opc = HTON__u32(opcode);
 309         request->rq_reqmsg->flags = 0;
 310
 311         ptlrpc_hdl2req(request, &imp->imp_handle);
 312         RETURN(request);
 313 }
 314
 315 static void __ptlrpc_free_req(struct ptlrpc_request *request, int locked)
 316 {
 317         ENTRY;
 318         if (request == NULL) {
 319                 EXIT;
 320                 return;
 321         }
 322
 323         if (atomic_read(&request->rq_refcount) != 0) {
 324                 CERROR("freeing request %p (%d->%s:%d) with refcount %d\n",
 325                        request, request->rq_reqmsg->opc,
 326                        request->rq_connection->c_remote_uuid,
 327                        request->rq_import->imp_client->cli_request_portal,
 328                        request->rq_refcount);
 329                 /* LBUG(); */
 330         }
 331
 332         if (request->rq_repmsg != NULL) {
 333                 OBD_FREE(request->rq_repmsg, request->rq_replen);
 334                 request->rq_repmsg = NULL;
 335                 request->rq_reply_md.start = NULL;
 336         }
 337         if (request->rq_reqmsg != NULL) {
 338                 OBD_FREE(request->rq_reqmsg, request->rq_reqlen);
 339                 request->rq_reqmsg = NULL;
 340         }
 341
 342         if (request->rq_import) {
 343                 if (!locked)
 344                         spin_lock(&request->rq_import->imp_lock);
 345                 list_del_init(&request->rq_list);
 346                 if (!locked)
 347                         spin_unlock(&request->rq_import->imp_lock);
 348         }
 349
 350         ptlrpc_put_connection(request->rq_connection);
 351         OBD_FREE(request, sizeof(*request));
 352         EXIT;
 353 }
 354
 355 void ptlrpc_free_req(struct ptlrpc_request *request)
 356 {
 357         __ptlrpc_free_req(request, 0);
 358 }
 359
 360 static int __ptlrpc_req_finished(struct ptlrpc_request *request, int locked)
 361 {
 362         ENTRY;
 363         if (request == NULL)
 364                 RETURN(1);
 365
 366         if (atomic_dec_and_test(&request->rq_refcount)) {
 367                 __ptlrpc_free_req(request, locked);
 368                 RETURN(1);
 369         }
 370
 371         DEBUG_REQ(D_INFO, request, "refcount now %u",
 372                   atomic_read(&request->rq_refcount));
 373         RETURN(0);
 374 }
 375
 376 void ptlrpc_req_finished(struct ptlrpc_request *request)
 377 {
 378         __ptlrpc_req_finished(request, 0);
 379 }
 380
 381 static int ptlrpc_check_reply(struct ptlrpc_request *req)
 382 {
 383         int rc = 0;
 384
 385         if (req->rq_repmsg != NULL) {
 386                 struct obd_import *imp = req->rq_import;
 387                 struct ptlrpc_connection *conn = imp->imp_connection;
 388                 ENTRY;
 389                 if (req->rq_level > conn->c_level) {
 390                         DEBUG_REQ(D_HA, req,
 391                                "recovery started, ignoring (%d > %d)",
 392                                req->rq_level, conn->c_level);
 393                         req->rq_repmsg = NULL;
 394                         GOTO(out, rc = 0);
 395                 }
 396                 req->rq_transno = NTOH__u64(req->rq_repmsg->transno);
 397                 spin_lock(&imp->imp_lock);
 398                 if (req->rq_transno > imp->imp_max_transno) {
 399                         imp->imp_max_transno = req->rq_transno;
 400                 } else if (req->rq_transno != 0) {
 401                         if (conn->c_level == LUSTRE_CONN_FULL) {
 402                                 CERROR("got transno "LPD64" after "
 403                                        LPD64": recovery may not work\n",
 404                                        req->rq_transno, imp->imp_max_transno);
 405                         }
 406                 }
 407                 spin_unlock(&imp->imp_lock);
 408                 req->rq_flags |= PTL_RPC_FL_REPLIED;
 409                 GOTO(out, rc = 1);
 410         }
 411
 412         if (req->rq_flags & PTL_RPC_FL_RESEND) {
 413                 DEBUG_REQ(D_ERROR, req, "RESEND:");
 414                 GOTO(out, rc = 1);
 415         }
 416
 417         if (req->rq_flags & PTL_RPC_FL_ERR) {
 418                 DEBUG_REQ(D_ERROR, req, "ABORTED:");
 419                 GOTO(out, rc = 1);
 420         }
 421
 422         if (req->rq_flags & PTL_RPC_FL_RESTART) {
 423                 DEBUG_REQ(D_ERROR, req, "RESTART:");
 424                 GOTO(out, rc = 1);
 425         }
 426  out:
 427         DEBUG_REQ(D_NET, req, "rc = %d for", rc);
 428         return rc;
 429 }
 430
 431 static int ptlrpc_check_status(struct ptlrpc_request *req)
 432 {
 433         int err;
 434         ENTRY;
 435
 436         err = req->rq_repmsg->status;
 437         if (req->rq_repmsg->type == NTOH__u32(PTL_RPC_MSG_ERR)) {
 438                 CERROR("req->rq_repmsg->type == PTL_RPC_MSG_ERR\n");
 439                 RETURN(err ? err : -EINVAL);
 440         }
 441
 442         if (err != 0) {
 443                 if (err < 0)
 444                         CDEBUG(D_INFO, "req->rq_repmsg->status is %d\n", err);
 445                 else
 446                         CDEBUG(D_INFO, "req->rq_repmsg->status is %d\n", err);
 447         }
 448
 449         RETURN(err);
 450 }
 451
 452 static void ptlrpc_cleanup_request_buf(struct ptlrpc_request *request)
 453 {
 454         OBD_FREE(request->rq_reqmsg, request->rq_reqlen);
 455         request->rq_reqmsg = NULL;
 456         request->rq_reqlen = 0;
 457 }
 458
 459 /* Abort this request and cleanup any resources associated with it. */
 460 static int ptlrpc_abort(struct ptlrpc_request *request)
 461 {
 462         /* First remove the ME for the reply; in theory, this means
 463          * that we can tear down the buffer safely. */
 464         PtlMEUnlink(request->rq_reply_me_h);
 465         OBD_FREE(request->rq_reply_md.start, request->rq_replen);
 466         request->rq_repmsg = NULL;
 467         request->rq_replen = 0;
 468         return 0;
 469 }
 470
 471 /* caller must hold imp->imp_lock */
 472 void ptlrpc_free_committed(struct obd_import *imp)
 473 {
 474         struct list_head *tmp, *saved;
 475         struct ptlrpc_request *req;
 476
 477         list_for_each_safe(tmp, saved, &imp->imp_request_list) {
 478                 req = list_entry(tmp, struct ptlrpc_request, rq_list);
 479
 480                 if (req->rq_flags & PTL_RPC_FL_REPLAY) {
 481                         DEBUG_REQ(D_HA, req, "keeping (FL_REPLAY)");
 482                         continue;
 483                 }
 484
 485                 /* If neither replied-to nor restarted, keep it. */
 486                 if (!(req->rq_flags &
 487                       (PTL_RPC_FL_REPLIED | PTL_RPC_FL_RESTART))) {
 488                         DEBUG_REQ(D_HA, req, "keeping (in-flight)");
 489                         continue;
 490                 }
 491
 492                 /* This needs to match the commit test in ptlrpc_queue_wait() */
 493                 if (!(req->rq_import->imp_flags & IMP_REPLAYABLE) ||
 494                     req->rq_transno == 0) {
 495                         DEBUG_REQ(D_HA, req, "keeping (queue_wait will free)");
 496                         continue;
 497                 }
 498
 499                 /* not yet committed */
 500                 if (req->rq_transno > imp->imp_peer_committed_transno)
 501                         break;
 502
 503                 DEBUG_REQ(D_HA, req, "committing (last_committed %Lu)",
 504                           imp->imp_peer_committed_transno);
 505                 __ptlrpc_req_finished(req, 1);
 506         }
 507
 508         EXIT;
 509         return;
 510 }
 511
 512 void ptlrpc_cleanup_client(struct obd_import *imp)
 513 {
 514         struct list_head *tmp, *saved;
 515         struct ptlrpc_request *req;
 516         struct ptlrpc_connection *conn = imp->imp_connection;
 517         ENTRY;
 518
 519         LASSERT(conn);
 520
 521         spin_lock(&imp->imp_lock);
 522         list_for_each_safe(tmp, saved, &imp->imp_request_list) {
 523                 req = list_entry(tmp, struct ptlrpc_request, rq_list);
 524
 525                 /* XXX we should make sure that nobody's sleeping on these! */
 526                 DEBUG_REQ(D_HA, req, "cleaning up from sending list");
 527                 list_del_init(&req->rq_list);
 528                 req->rq_import = NULL;
 529                 __ptlrpc_req_finished(req, 0);
 530         }
 531         spin_unlock(&imp->imp_lock);
 532
 533         EXIT;
 534         return;
 535 }
 536
 537 void ptlrpc_continue_req(struct ptlrpc_request *req)
 538 {
 539         ENTRY;
 540         DEBUG_REQ(D_HA, req, "continuing delayed request");
 541         req->rq_reqmsg->addr = req->rq_import->imp_handle.addr;
 542         req->rq_reqmsg->cookie = req->rq_import->imp_handle.cookie;
 543         wake_up(&req->rq_wait_for_rep);
 544         EXIT;
 545 }
 546
 547 void ptlrpc_resend_req(struct ptlrpc_request *req)
 548 {
 549         ENTRY;
 550         DEBUG_REQ(D_HA, req, "resending");
 551         req->rq_reqmsg->addr = req->rq_import->imp_handle.addr;
 552         req->rq_reqmsg->cookie = req->rq_import->imp_handle.cookie;
 553         req->rq_status = -EAGAIN;
 554         req->rq_level = LUSTRE_CONN_RECOVD;
 555         req->rq_flags |= PTL_RPC_FL_RESEND;
 556         req->rq_flags &= ~PTL_RPC_FL_TIMEOUT;
 557         wake_up(&req->rq_wait_for_rep);
 558         EXIT;
 559 }
 560
 561 void ptlrpc_restart_req(struct ptlrpc_request *req)
 562 {
 563         ENTRY;
 564         DEBUG_REQ(D_HA, req, "restarting (possibly-)completed request");
 565         req->rq_status = -ERESTARTSYS;
 566         req->rq_flags |= PTL_RPC_FL_RESTART;
 567         req->rq_flags &= ~PTL_RPC_FL_TIMEOUT;
 568         wake_up(&req->rq_wait_for_rep);
 569         EXIT;
 570 }
 571
 572 static int expired_request(void *data)
 573 {
 574         struct ptlrpc_request *req = data;
 575
 576         ENTRY;
 577         if (!req) {
 578                 CERROR("NULL req!");
 579                 LBUG();
 580                 RETURN(0);
 581         }
 582
 583         DEBUG_REQ(D_ERROR, req, "timeout");
 584         req->rq_flags |= PTL_RPC_FL_TIMEOUT;
 585
 586         if (!req->rq_import) {
 587                 DEBUG_REQ(D_ERROR, req, "NULL import");
 588                 LBUG();
 589                 RETURN(0);
 590         }
 591
 592         if (!req->rq_import->imp_connection) {
 593                 DEBUG_REQ(D_ERROR, req, "NULL connection");
 594                 LBUG();
 595                 RETURN(0);
 596         }
 597
 598         if (!req->rq_import->imp_connection->c_recovd_data.rd_recovd)
 599                 RETURN(1);
 600
 601         req->rq_timeout = 0;
 602         req->rq_connection->c_level = LUSTRE_CONN_RECOVD;
 603         recovd_conn_fail(req->rq_import->imp_connection);
 604
 605         /* If this request is for recovery or other primordial tasks,
 606          * don't go back to sleep.
 607          */
 608         if (req->rq_level < LUSTRE_CONN_FULL)
 609                 RETURN(1);
 610         RETURN(0);
 611 }
 612
 613 static int interrupted_request(void *data)
 614 {
 615         struct ptlrpc_request *req = data;
 616         ENTRY;
 617         req->rq_flags |= PTL_RPC_FL_INTR;
 618         RETURN(1); /* ignored, as of this writing */
 619 }
 620
 621 /* If we're being torn down by umount -f, or the import has been
 622  * invalidated (such as by an OST failure), the request must fail with
 623  * -EIO.
 624  *
 625  * Must be called with conn->c_lock held, will drop it if it returns -EIO.
 626  *
 627  * XXX this should just be testing the import, and umount_begin shouldn't touch
 628  * XXX the connection.
 629  */
 630 #define EIO_IF_INVALID(conn, req)                                             \
 631 if ((conn->c_flags & CONN_INVALID) ||                                         \
 632     (req->rq_import->imp_flags & IMP_INVALID)) {                              \
 633         DEBUG_REQ(D_ERROR, req, "%s_INVALID:",                                \
 634                   (conn->c_flags & CONN_INVALID) ? "CONN" : "IMP");           \
 635         spin_unlock(&conn->c_lock);                                           \
 636         RETURN(-EIO);                                                         \
 637 }
 638
 639 int ptlrpc_queue_wait(struct ptlrpc_request *req)
 640 {
 641         int rc = 0;
 642         struct l_wait_info lwi;
 643         struct obd_import *imp = req->rq_import;
 644         struct ptlrpc_connection *conn = imp->imp_connection;
 645         ENTRY;
 646
 647         init_waitqueue_head(&req->rq_wait_for_rep);
 648         req->rq_reqmsg->status = HTON__u32(current->pid); /* for distributed debugging */
 649         CDEBUG(D_RPCTRACE, "Sending RPC pid:xid:nid:opc %d:"LPU64":%x:%d\n",
 650                NTOH__u32(req->rq_reqmsg->status), req->rq_xid,
 651                conn->c_peer.peer_nid, NTOH__u32(req->rq_reqmsg->opc));
 652
 653
 654         /* XXX probably both an import and connection level are needed */
 655         if (req->rq_level > conn->c_level) {
 656                 spin_lock(&conn->c_lock);
 657                 EIO_IF_INVALID(conn, req);
 658                 list_del(&req->rq_list);
 659                 list_add_tail(&req->rq_list, &conn->c_delayed_head);
 660                 spin_unlock(&conn->c_lock);
 661
 662                 DEBUG_REQ(D_HA, req, "waiting for recovery: (%d < %d)",
 663                           req->rq_level, conn->c_level);
 664                 lwi = LWI_INTR(NULL, NULL);
 665                 rc = l_wait_event(req->rq_wait_for_rep,
 666                                   (req->rq_level <= conn->c_level) ||
 667                                   (req->rq_flags & PTL_RPC_FL_ERR), &lwi);
 668
 669                 spin_lock(&conn->c_lock);
 670                 list_del_init(&req->rq_list);
 671                 spin_unlock(&conn->c_lock);
 672
 673                 if (req->rq_flags & PTL_RPC_FL_ERR)
 674                         RETURN(-EIO);
 675
 676                 if (rc)
 677                         RETURN(rc);
 678
 679                 CERROR("process %d resumed\n", current->pid);
 680         }
 681  resend:
 682         req->rq_timeout = obd_timeout;
 683         spin_lock(&conn->c_lock);
 684         EIO_IF_INVALID(conn, req);
 685
 686         list_del(&req->rq_list);
 687         list_add_tail(&req->rq_list, &imp->imp_request_list);
 688         spin_unlock(&conn->c_lock);
 689         rc = ptl_send_rpc(req);
 690         if (rc) {
 691                 CDEBUG(D_HA, "error %d, opcode %d, need recovery\n", rc,
 692                        req->rq_reqmsg->opc);
 693                 /* sleep for a jiffy, then trigger recovery */
 694                 lwi = LWI_TIMEOUT_INTR(1, expired_request,
 695                                        interrupted_request, req);
 696         } else {
 697                 DEBUG_REQ(D_NET, req, "-- sleeping");
 698                 lwi = LWI_TIMEOUT_INTR(req->rq_timeout * HZ, expired_request,
 699                                        interrupted_request, req);
 700         }
 701         l_wait_event(req->rq_wait_for_rep, ptlrpc_check_reply(req), &lwi);
 702         DEBUG_REQ(D_NET, req, "-- done sleeping");
 703
 704         if (req->rq_flags & PTL_RPC_FL_ERR) {
 705                 ptlrpc_abort(req);
 706                 GOTO(out, rc = -EIO);
 707         }
 708
 709         /* Don't resend if we were interrupted. */
 710         if ((req->rq_flags & (PTL_RPC_FL_RESEND | PTL_RPC_FL_INTR)) ==
 711             PTL_RPC_FL_RESEND) {
 712                 req->rq_flags &= ~PTL_RPC_FL_RESEND;
 713                 DEBUG_REQ(D_HA, req, "resending: ");
 714                 goto resend;
 715         }
 716
 717         // up(&cli->cli_rpc_sem);
 718         if (req->rq_flags & PTL_RPC_FL_INTR) {
 719                 if (!(req->rq_flags & PTL_RPC_FL_TIMEOUT))
 720                         LBUG(); /* should only be interrupted if we timed out */
 721                 /* Clean up the dangling reply buffers */
 722                 ptlrpc_abort(req);
 723                 GOTO(out, rc = -EINTR);
 724         }
 725
 726         if (req->rq_flags & PTL_RPC_FL_TIMEOUT)
 727                 GOTO(out, rc = -ETIMEDOUT);
 728
 729         if (!(req->rq_flags & PTL_RPC_FL_REPLIED))
 730                 GOTO(out, rc = req->rq_status);
 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 #if 0
 738         /* FIXME: Enable when BlueArc makes new release */
 739         if (req->rq_repmsg->type != PTL_RPC_MSG_REPLY &&
 740             req->rq_repmsg->type != PTL_RPC_MSG_ERR) {
 741                 CERROR("invalid packet type received (type=%u)\n",
 742                        req->rq_repmsg->type);
 743                 LBUG();
 744                 GOTO(out, rc = -EINVAL);
 745         }
 746 #endif
 747         CDEBUG(D_NET, "got rep "LPU64"\n", req->rq_xid);
 748         if (req->rq_repmsg->status == 0)
 749                 CDEBUG(D_NET, "--> buf %p len %d status %d\n", req->rq_repmsg,
 750                        req->rq_replen, req->rq_repmsg->status);
 751
 752         spin_lock(&conn->c_lock);
 753
 754         /* Requests that aren't from replayable imports, or which don't have
 755          * transno information, can be "committed" early.
 756          *
 757          * But don't commit anything that's kept indefinitely for replay (has
 758          * the PTL_RPC_FL_REPLAY flag set), such as open requests.
 759          *
 760          * This needs to match the commit test in ptlrpc_free_committed().
 761          */
 762         if (!(req->rq_import->imp_flags & IMP_REPLAYABLE) ||
 763             (req->rq_repmsg->transno == 0 &&
 764              (req->rq_flags & PTL_RPC_FL_REPLAY) == 0)) {
 765                 /* This import doesn't support replay, so we can just "commit"
 766                  * this request now.
 767                  */
 768                 DEBUG_REQ(D_HA, req, "not replayable, committing:");
 769                 list_del_init(&req->rq_list);
 770                 __ptlrpc_req_finished(req, 1);
 771         }
 772         if (req->rq_import->imp_flags & IMP_REPLAYABLE) {
 773                 /* Replay-enabled imports return commit-status information. */
 774                 imp->imp_peer_last_xid = req->rq_repmsg->last_xid;
 775                 imp->imp_peer_committed_transno =
 776                         req->rq_repmsg->last_committed;
 777                 ptlrpc_free_committed(imp);
 778         }
 779
 780         rc = ptlrpc_check_status(req);
 781         spin_unlock(&conn->c_lock);
 782
 783         EXIT;
 784  out:
 785         return rc;
 786 }
 787
 788 #undef EIO_IF_INVALID
 789
 790 int ptlrpc_replay_req(struct ptlrpc_request *req)
 791 {
 792         int rc = 0, old_level, old_status = 0;
 793         // struct ptlrpc_client *cli = req->rq_import->imp_client;
 794         struct l_wait_info lwi;
 795         ENTRY;
 796
 797         init_waitqueue_head(&req->rq_wait_for_rep);
 798         DEBUG_REQ(D_NET, req, "");
 799
 800         req->rq_timeout = obd_timeout;
 801         req->rq_reqmsg->addr = req->rq_import->imp_handle.addr;
 802         req->rq_reqmsg->cookie = req->rq_import->imp_handle.cookie;
 803
 804         /* temporarily set request to RECOVD level (reset at out:) */
 805         old_level = req->rq_level;
 806         if (req->rq_flags & PTL_RPC_FL_REPLIED)
 807                 old_status = req->rq_repmsg->status;
 808         req->rq_level = LUSTRE_CONN_RECOVD;
 809         rc = ptl_send_rpc(req);
 810         if (rc) {
 811                 CERROR("error %d, opcode %d\n", rc, req->rq_reqmsg->opc);
 812                 ptlrpc_cleanup_request_buf(req);
 813                 // up(&cli->cli_rpc_sem);
 814                 GOTO(out, rc = -rc);
 815         }
 816
 817         CDEBUG(D_OTHER, "-- sleeping\n");
 818         lwi = LWI_INTR(NULL, NULL); /* XXX needs timeout, nested recovery */
 819         l_wait_event(req->rq_wait_for_rep, ptlrpc_check_reply(req), &lwi);
 820         CDEBUG(D_OTHER, "-- done\n");
 821
 822         // up(&cli->cli_rpc_sem);
 823
 824         if (!(req->rq_flags & PTL_RPC_FL_REPLIED)) {
 825                 CERROR("Unknown reason for wakeup\n");
 826                 /* XXX Phil - I end up here when I kill obdctl */
 827                 ptlrpc_abort(req);
 828                 GOTO(out, rc = -EINTR);
 829         }
 830
 831         rc = lustre_unpack_msg(req->rq_repmsg, req->rq_replen);
 832         if (rc) {
 833                 CERROR("unpack_rep failed: %d\n", rc);
 834                 GOTO(out, rc);
 835         }
 836
 837         CDEBUG(D_NET, "got rep "LPD64"\n", req->rq_xid);
 838
 839         /* let the callback do fixups, possibly including in the request */
 840         if (req->rq_replay_cb)
 841                 req->rq_replay_cb(req);
 842
 843         if ((req->rq_flags & PTL_RPC_FL_REPLIED) &&
 844             req->rq_repmsg->status != old_status) {
 845                 DEBUG_REQ(D_HA, req, "status %d, old was %d",
 846                           req->rq_repmsg->status, old_status);
 847         }
 848
 849  out:
 850         req->rq_level = old_level;
 851         RETURN(rc);
 852 }