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         if (atomic_read(&desc->bd_refcount) != 0)
 135                 CERROR("freeing desc %p with refcount %d!\n", desc,
 136                        atomic_read(&desc->bd_refcount));
 137
 138         list_for_each_safe(tmp, next, &desc->bd_page_list) {
 139                 struct ptlrpc_bulk_page *bulk;
 140                 bulk = list_entry(tmp, struct ptlrpc_bulk_page, bp_link);
 141                 ptlrpc_free_bulk_page(bulk);
 142         }
 143
 144         ptlrpc_put_connection(desc->bd_connection);
 145
 146         OBD_FREE(desc, sizeof(*desc));
 147         EXIT;
 148 }
 149
 150 void ptlrpc_free_bulk_page(struct ptlrpc_bulk_page *bulk)
 151 {
 152         ENTRY;
 153         if (bulk == NULL) {
 154                 EXIT;
 155                 return;
 156         }
 157
 158         list_del(&bulk->bp_link);
 159         bulk->bp_desc->bd_page_count--;
 160         OBD_FREE(bulk, sizeof(*bulk));
 161         EXIT;
 162 }
 163
 164 static int ll_sync_brw_timeout(void *data)
 165 {
 166         struct obd_brw_set *set = data;
 167         struct list_head *tmp;
 168         int failed = 0;
 169         ENTRY;
 170
 171         LASSERT(set);
 172
 173         set->brw_flags |= PTL_RPC_FL_TIMEOUT;
 174
 175         list_for_each(tmp, &set->brw_desc_head) {
 176                 struct ptlrpc_bulk_desc *desc =
 177                         list_entry(tmp, struct ptlrpc_bulk_desc, bd_set_chain);
 178
 179                 /* Skip descriptors that were completed successfully. */
 180                 if (desc->bd_flags & (PTL_BULK_FL_RCVD | PTL_BULK_FL_SENT))
 181                         continue;
 182
 183                 LASSERT(desc->bd_connection);
 184
 185                 /* If PtlMDUnlink succeeds, then it hasn't completed yet.  If it
 186                  * fails, the bulk finished _just_ in time (after the timeout
 187                  * fired but before we got this far) and we'll let it live.
 188                  */
 189                 if (PtlMDUnlink(desc->bd_md_h) != 0) {
 190                         CERROR("Near-miss on OST %s -- need to adjust "
 191                                "obd_timeout?\n",
 192                                desc->bd_connection->c_remote_uuid);
 193                         continue;
 194                 }
 195
 196                 CERROR("IO of %d pages to/from %s:%d (conn %p) timed out\n",
 197                        desc->bd_page_count, desc->bd_connection->c_remote_uuid,
 198                        desc->bd_portal, desc->bd_connection);
 199
 200                 /* This one will "never" arrive, don't wait for it. */
 201                 if (atomic_dec_and_test(&set->brw_refcount))
 202                         wake_up(&set->brw_waitq);
 203
 204                 if (class_signal_connection_failure)
 205                         class_signal_connection_failure(desc->bd_connection);
 206                 else
 207                         failed = 1;
 208         }
 209
 210         /* 0 = We go back to sleep, until we're resumed or interrupted */
 211         /* 1 = We can't be recovered, just abort the syscall with -ETIMEDOUT */
 212         RETURN(failed);
 213 }
 214
 215 static int ll_sync_brw_intr(void *data)
 216 {
 217         struct obd_brw_set *set = data;
 218
 219         ENTRY;
 220         set->brw_flags |= PTL_RPC_FL_INTR;
 221         RETURN(1); /* ignored, as of this writing */
 222 }
 223
 224 int ll_brw_sync_wait(struct obd_brw_set *set, int phase)
 225 {
 226         struct l_wait_info lwi;
 227         struct list_head *tmp, *next;
 228         int rc = 0;
 229         ENTRY;
 230
 231         switch(phase) {
 232         case CB_PHASE_START:
 233                 lwi = LWI_TIMEOUT_INTR(obd_timeout * HZ, ll_sync_brw_timeout,
 234                                        ll_sync_brw_intr, set);
 235                 rc = l_wait_event(set->brw_waitq,
 236                                   atomic_read(&set->brw_refcount) == 0, &lwi);
 237
 238                 list_for_each_safe(tmp, next, &set->brw_desc_head) {
 239                         struct ptlrpc_bulk_desc *desc =
 240                                 list_entry(tmp, struct ptlrpc_bulk_desc,
 241                                            bd_set_chain);
 242                         list_del_init(&desc->bd_set_chain);
 243                         ptlrpc_bulk_decref(desc);
 244                 }
 245                 break;
 246         case CB_PHASE_FINISH:
 247                 if (atomic_dec_and_test(&set->brw_refcount))
 248                         wake_up(&set->brw_waitq);
 249                 break;
 250         default:
 251                 LBUG();
 252         }
 253
 254         RETURN(rc);
 255 }
 256
 257 struct ptlrpc_request *ptlrpc_prep_req(struct obd_import *imp, int opcode,
 258                                        int count, int *lengths, char **bufs)
 259 {
 260         struct ptlrpc_connection *conn;
 261         struct ptlrpc_request *request;
 262         unsigned long flags;
 263         int rc;
 264         ENTRY;
 265
 266         LASSERT((unsigned long)imp > 0x1000);
 267         conn = imp->imp_connection;
 268
 269         OBD_ALLOC(request, sizeof(*request));
 270         if (!request) {
 271                 CERROR("request allocation out of memory\n");
 272                 RETURN(NULL);
 273         }
 274
 275         rc = lustre_pack_msg(count, lengths, bufs,
 276                              &request->rq_reqlen, &request->rq_reqmsg);
 277         if (rc) {
 278                 CERROR("cannot pack request %d\n", rc);
 279                 OBD_FREE(request, sizeof(*request));
 280                 RETURN(NULL);
 281         }
 282
 283         request->rq_level = LUSTRE_CONN_FULL;
 284         request->rq_type = PTL_RPC_MSG_REQUEST;
 285         request->rq_import = imp;
 286
 287         /* XXX FIXME bug 625069 */
 288         request->rq_request_portal = imp->imp_client->cli_request_portal;
 289         request->rq_reply_portal = imp->imp_client->cli_reply_portal;
 290
 291         request->rq_connection = ptlrpc_connection_addref(conn);
 292
 293         INIT_LIST_HEAD(&request->rq_list);
 294         atomic_set(&request->rq_refcount, 1);
 295
 296         spin_lock_irqsave(&imp->imp_lock, flags);
 297         request->rq_xid = HTON__u32(++imp->imp_last_xid);
 298         spin_unlock_irqrestore(&imp->imp_lock, flags);
 299
 300         request->rq_reqmsg->magic = PTLRPC_MSG_MAGIC;
 301         request->rq_reqmsg->version = PTLRPC_MSG_VERSION;
 302         request->rq_reqmsg->opc = HTON__u32(opcode);
 303         request->rq_reqmsg->flags = 0;
 304
 305         ptlrpc_hdl2req(request, &imp->imp_handle);
 306         RETURN(request);
 307 }
 308
 309 static void __ptlrpc_free_req(struct ptlrpc_request *request, int locked)
 310 {
 311         ENTRY;
 312         if (request == NULL) {
 313                 EXIT;
 314                 return;
 315         }
 316
 317         if (atomic_read(&request->rq_refcount) != 0) {
 318                 CERROR("freeing request %p (%d->%s:%d) with refcount %d\n",
 319                        request, request->rq_reqmsg->opc,
 320                        request->rq_connection->c_remote_uuid,
 321                        request->rq_import->imp_client->cli_request_portal,
 322                        atomic_read (&request->rq_refcount));
 323                 /* LBUG(); */
 324         }
 325
 326         if (request->rq_repmsg != NULL) {
 327                 OBD_FREE(request->rq_repmsg, request->rq_replen);
 328                 request->rq_repmsg = NULL;
 329                 request->rq_reply_md.start = NULL;
 330         }
 331         if (request->rq_reqmsg != NULL) {
 332                 OBD_FREE(request->rq_reqmsg, request->rq_reqlen);
 333                 request->rq_reqmsg = NULL;
 334         }
 335
 336         if (request->rq_import) {
 337                 unsigned long flags = 0;
 338                 if (!locked)
 339                         spin_lock_irqsave(&request->rq_import->imp_lock, flags);
 340                 list_del_init(&request->rq_list);
 341                 if (!locked)
 342                         spin_unlock_irqrestore(&request->rq_import->imp_lock,
 343                                                flags);
 344         }
 345
 346         ptlrpc_put_connection(request->rq_connection);
 347         OBD_FREE(request, sizeof(*request));
 348         EXIT;
 349 }
 350
 351 void ptlrpc_free_req(struct ptlrpc_request *request)
 352 {
 353         __ptlrpc_free_req(request, 0);
 354 }
 355
 356 static int __ptlrpc_req_finished(struct ptlrpc_request *request, int locked)
 357 {
 358         ENTRY;
 359         if (request == NULL)
 360                 RETURN(1);
 361
 362         if (request == (void *)(long)(0x5a5a5a5a5a5a5a5a)) {
 363                 CERROR("dereferencing freed request (bug 575)\n");
 364                 LBUG();
 365                 RETURN(1);
 366         }
 367
 368         DEBUG_REQ(D_INFO, request, "refcount now %u",
 369                   atomic_read(&request->rq_refcount) - 1);
 370
 371         if (atomic_dec_and_test(&request->rq_refcount)) {
 372                 __ptlrpc_free_req(request, locked);
 373                 RETURN(1);
 374         }
 375
 376         RETURN(0);
 377 }
 378
 379 void ptlrpc_req_finished(struct ptlrpc_request *request)
 380 {
 381         __ptlrpc_req_finished(request, 0);
 382 }
 383
 384 static int ptlrpc_check_reply(struct ptlrpc_request *req)
 385 {
 386         int rc = 0;
 387
 388         ENTRY;
 389         if (req->rq_repmsg != NULL) {
 390                 req->rq_transno = NTOH__u64(req->rq_repmsg->transno);
 391                 /* Store transno in reqmsg for replay. */
 392                 req->rq_reqmsg->transno = req->rq_repmsg->transno;
 393                 req->rq_flags |= PTL_RPC_FL_REPLIED;
 394                 GOTO(out, rc = 1);
 395         }
 396
 397         if (req->rq_flags & PTL_RPC_FL_RESEND) {
 398                 ENTRY;
 399                 DEBUG_REQ(D_ERROR, req, "RESEND:");
 400                 GOTO(out, rc = 1);
 401         }
 402
 403         if (req->rq_flags & PTL_RPC_FL_ERR) {
 404                 ENTRY;
 405                 DEBUG_REQ(D_ERROR, req, "ABORTED:");
 406                 GOTO(out, rc = 1);
 407         }
 408
 409         if (req->rq_flags & PTL_RPC_FL_RESTART) {
 410                 DEBUG_REQ(D_ERROR, req, "RESTART:");
 411                 GOTO(out, rc = 1);
 412         }
 413         EXIT;
 414  out:
 415         DEBUG_REQ(D_NET, req, "rc = %d for", rc);
 416         return rc;
 417 }
 418
 419 static int ptlrpc_check_status(struct ptlrpc_request *req)
 420 {
 421         int err;
 422         ENTRY;
 423
 424         err = req->rq_repmsg->status;
 425         if (req->rq_repmsg->type == NTOH__u32(PTL_RPC_MSG_ERR)) {
 426                 DEBUG_REQ(D_ERROR, req, "type == PTL_RPC_MSG_ERR (%d)", err);
 427                 RETURN(err ? err : -EINVAL);
 428         }
 429
 430         if (err < 0) {
 431                 DEBUG_REQ(D_ERROR, req, "status is %d", err);
 432         } else if (err > 0) {
 433                 /* XXX: translate this error from net to host */
 434                 DEBUG_REQ(D_INFO, req, "status is %d", err);
 435         }
 436
 437         RETURN(err);
 438 }
 439
 440 static void ptlrpc_cleanup_request_buf(struct ptlrpc_request *request)
 441 {
 442         OBD_FREE(request->rq_reqmsg, request->rq_reqlen);
 443         request->rq_reqmsg = NULL;
 444         request->rq_reqlen = 0;
 445 }
 446
 447 /* Abort this request and cleanup any resources associated with it. */
 448 static int ptlrpc_abort(struct ptlrpc_request *request)
 449 {
 450         /* First remove the ME for the reply; in theory, this means
 451          * that we can tear down the buffer safely. */
 452         if (PtlMEUnlink(request->rq_reply_me_h) != PTL_OK)
 453                 RETURN(0);
 454         OBD_FREE(request->rq_reply_md.start, request->rq_replen);
 455
 456         memset(&request->rq_reply_me_h, 0, sizeof(request->rq_reply_me_h));
 457         request->rq_reply_md.start = NULL;
 458         request->rq_repmsg = NULL;
 459         return 0;
 460 }
 461
 462 /* caller must hold imp->imp_lock */
 463 void ptlrpc_free_committed(struct obd_import *imp)
 464 {
 465         struct list_head *tmp, *saved;
 466         struct ptlrpc_request *req;
 467         ENTRY;
 468
 469 #ifdef CONFIG_SMP
 470         LASSERT(spin_is_locked(&imp->imp_lock));
 471 #endif
 472
 473         CDEBUG(D_HA, "committing for xid "LPU64", last_committed "LPU64"\n",
 474                imp->imp_peer_last_xid, imp->imp_peer_committed_transno);
 475
 476         list_for_each_safe(tmp, saved, &imp->imp_replay_list) {
 477                 req = list_entry(tmp, struct ptlrpc_request, rq_list);
 478
 479                 if (req->rq_flags & PTL_RPC_FL_REPLAY) {
 480                         DEBUG_REQ(D_HA, req, "keeping (FL_REPLAY)");
 481                         continue;
 482                 }
 483
 484                 /* not yet committed */
 485                 if (req->rq_transno > imp->imp_peer_committed_transno) {
 486                         DEBUG_REQ(D_HA, req, "stopping search");
 487                         break;
 488                 }
 489
 490                 DEBUG_REQ(D_HA, req, "committing (last_committed "LPU64")",
 491                           imp->imp_peer_committed_transno);
 492                 __ptlrpc_req_finished(req, 1);
 493         }
 494
 495         EXIT;
 496         return;
 497 }
 498
 499 void ptlrpc_cleanup_client(struct obd_import *imp)
 500 {
 501         struct list_head *tmp, *saved;
 502         struct ptlrpc_request *req;
 503         struct ptlrpc_connection *conn = imp->imp_connection;
 504         unsigned long flags;
 505         ENTRY;
 506
 507         LASSERT(conn);
 508
 509         spin_lock_irqsave(&imp->imp_lock, flags);
 510         list_for_each_safe(tmp, saved, &imp->imp_replay_list) {
 511                 req = list_entry(tmp, struct ptlrpc_request, rq_list);
 512
 513                 /* XXX we should make sure that nobody's sleeping on these! */
 514                 DEBUG_REQ(D_HA, req, "cleaning up from sending list");
 515                 list_del_init(&req->rq_list);
 516                 req->rq_import = NULL;
 517                 __ptlrpc_req_finished(req, 0);
 518         }
 519         spin_unlock_irqrestore(&imp->imp_lock, flags);
 520
 521         EXIT;
 522         return;
 523 }
 524
 525 void ptlrpc_continue_req(struct ptlrpc_request *req)
 526 {
 527         ENTRY;
 528         DEBUG_REQ(D_HA, req, "continuing delayed request");
 529         req->rq_reqmsg->addr = req->rq_import->imp_handle.addr;
 530         req->rq_reqmsg->cookie = req->rq_import->imp_handle.cookie;
 531         wake_up(&req->rq_wait_for_rep);
 532         EXIT;
 533 }
 534
 535 void ptlrpc_resend_req(struct ptlrpc_request *req)
 536 {
 537         ENTRY;
 538         DEBUG_REQ(D_HA, req, "resending");
 539         req->rq_reqmsg->addr = req->rq_import->imp_handle.addr;
 540         req->rq_reqmsg->cookie = req->rq_import->imp_handle.cookie;
 541         req->rq_status = -EAGAIN;
 542         req->rq_level = LUSTRE_CONN_RECOVD;
 543         req->rq_flags |= PTL_RPC_FL_RESEND;
 544         req->rq_flags &= ~PTL_RPC_FL_TIMEOUT;
 545         wake_up(&req->rq_wait_for_rep);
 546         EXIT;
 547 }
 548
 549 void ptlrpc_restart_req(struct ptlrpc_request *req)
 550 {
 551         ENTRY;
 552         DEBUG_REQ(D_HA, req, "restarting (possibly-)completed request");
 553         req->rq_status = -ERESTARTSYS;
 554         req->rq_flags |= PTL_RPC_FL_RESTART;
 555         req->rq_flags &= ~PTL_RPC_FL_TIMEOUT;
 556         wake_up(&req->rq_wait_for_rep);
 557         EXIT;
 558 }
 559
 560 static int expired_request(void *data)
 561 {
 562         struct ptlrpc_request *req = data;
 563
 564         ENTRY;
 565         if (!req) {
 566                 CERROR("NULL req!");
 567                 LBUG();
 568                 RETURN(0);
 569         }
 570
 571         DEBUG_REQ(D_ERROR, req, "timeout");
 572         ptlrpc_abort(req);
 573         req->rq_flags |= PTL_RPC_FL_TIMEOUT;
 574
 575         if (!req->rq_import) {
 576                 DEBUG_REQ(D_ERROR, req, "NULL import");
 577                 LBUG();
 578                 RETURN(0);
 579         }
 580
 581         if (!req->rq_import->imp_connection) {
 582                 DEBUG_REQ(D_ERROR, req, "NULL connection");
 583                 LBUG();
 584                 RETURN(0);
 585         }
 586
 587         if (!req->rq_import->imp_connection->c_recovd_data.rd_recovd)
 588                 RETURN(1);
 589
 590         recovd_conn_fail(req->rq_import->imp_connection);
 591
 592         /* If this request is for recovery or other primordial tasks,
 593          * don't go back to sleep.
 594          */
 595         if (req->rq_level < LUSTRE_CONN_FULL)
 596                 RETURN(1);
 597         RETURN(0);
 598 }
 599
 600 static int interrupted_request(void *data)
 601 {
 602         struct ptlrpc_request *req = data;
 603         ENTRY;
 604         req->rq_flags |= PTL_RPC_FL_INTR;
 605         RETURN(1); /* ignored, as of this writing */
 606 }
 607
 608 int ptlrpc_queue_wait(struct ptlrpc_request *req)
 609 {
 610         int rc = 0;
 611         struct l_wait_info lwi;
 612         struct obd_import *imp = req->rq_import;
 613         struct ptlrpc_connection *conn = imp->imp_connection;
 614         unsigned int flags;
 615         ENTRY;
 616
 617         init_waitqueue_head(&req->rq_wait_for_rep);
 618
 619         /* for distributed debugging */
 620         req->rq_reqmsg->status = HTON__u32(current->pid);
 621         CDEBUG(D_RPCTRACE, "Sending RPC pid:xid:nid:opc %d:"LPU64":%x:%d\n",
 622                NTOH__u32(req->rq_reqmsg->status), req->rq_xid,
 623                conn->c_peer.peer_nid, NTOH__u32(req->rq_reqmsg->opc));
 624
 625         spin_lock_irqsave(&imp->imp_lock, flags);
 626
 627         /*
 628          * If the import has been invalidated (such as by an OST failure), the
 629          * request must fail with -EIO.
 630          */
 631         if (req->rq_import->imp_flags & IMP_INVALID) {
 632                 DEBUG_REQ(D_ERROR, req, "IMP_INVALID:");
 633                 spin_unlock_irqrestore(&imp->imp_lock, flags);
 634                 RETURN(-EIO);
 635         }
 636
 637         if (req->rq_level > imp->imp_level) {
 638                 list_del(&req->rq_list);
 639                 list_add_tail(&req->rq_list, &imp->imp_delayed_list);
 640                 spin_unlock_irqrestore(&imp->imp_lock, flags);
 641
 642                 DEBUG_REQ(D_HA, req, "\"%s\" waiting for recovery: (%d < %d)",
 643                           current->comm, req->rq_level, imp->imp_level);
 644                 lwi = LWI_INTR(NULL, NULL);
 645                 rc = l_wait_event(req->rq_wait_for_rep,
 646                                   (req->rq_level <= imp->imp_level) ||
 647                                   (req->rq_flags & PTL_RPC_FL_ERR), &lwi);
 648
 649                 spin_lock_irqsave(&imp->imp_lock, flags);
 650                 list_del_init(&req->rq_list);
 651
 652                 if (req->rq_flags & PTL_RPC_FL_ERR)
 653                         rc = -EIO;
 654
 655                 if (rc) {
 656                         spin_unlock_irqrestore(&imp->imp_lock, flags);
 657                         RETURN(rc);
 658                 }
 659
 660                 CERROR("process %d resumed\n", current->pid);
 661         }
 662  resend:
 663
 664         LASSERT(list_empty(&req->rq_list));
 665         list_add_tail(&req->rq_list, &imp->imp_sending_list);
 666         spin_unlock_irqrestore(&imp->imp_lock, flags);
 667         rc = ptl_send_rpc(req);
 668         if (rc) {
 669                 CDEBUG(D_HA, "error %d, opcode %d, need recovery\n", rc,
 670                        req->rq_reqmsg->opc);
 671                 /* sleep for a jiffy, then trigger recovery */
 672                 lwi = LWI_TIMEOUT_INTR(1, expired_request,
 673                                        interrupted_request, req);
 674         } else {
 675                 DEBUG_REQ(D_NET, req, "-- sleeping");
 676                 lwi = LWI_TIMEOUT_INTR(obd_timeout * HZ, expired_request,
 677                                        interrupted_request, req);
 678         }
 679         l_wait_event(req->rq_wait_for_rep, ptlrpc_check_reply(req), &lwi);
 680         DEBUG_REQ(D_NET, req, "-- done sleeping");
 681
 682         spin_lock_irqsave(&imp->imp_lock, flags);
 683         list_del_init(&req->rq_list);
 684         spin_unlock_irqrestore(&imp->imp_lock, flags);
 685
 686         if (req->rq_flags & PTL_RPC_FL_ERR) {
 687                 ptlrpc_abort(req);
 688                 GOTO(out, rc = -EIO);
 689         }
 690
 691         /* Don't resend if we were interrupted. */
 692         if ((req->rq_flags & (PTL_RPC_FL_RESEND | PTL_RPC_FL_INTR)) ==
 693             PTL_RPC_FL_RESEND) {
 694                 req->rq_flags &= ~PTL_RPC_FL_RESEND;
 695                 lustre_msg_add_flags(req->rq_reqmsg, MSG_RESENT);
 696                 DEBUG_REQ(D_HA, req, "resending: ");
 697                 spin_lock_irqsave(&imp->imp_lock, flags);
 698                 goto resend;
 699         }
 700
 701         if (req->rq_flags & PTL_RPC_FL_INTR) {
 702                 if (!(req->rq_flags & PTL_RPC_FL_TIMEOUT))
 703                         LBUG(); /* should only be interrupted if we timed out */
 704                 /* Clean up the dangling reply buffers */
 705                 ptlrpc_abort(req);
 706                 GOTO(out, rc = -EINTR);
 707         }
 708
 709         if (req->rq_flags & PTL_RPC_FL_TIMEOUT)
 710                 GOTO(out, rc = -ETIMEDOUT);
 711
 712         if (!(req->rq_flags & PTL_RPC_FL_REPLIED))
 713                 GOTO(out, rc = req->rq_status);
 714
 715         rc = lustre_unpack_msg(req->rq_repmsg, req->rq_replen);
 716         if (rc) {
 717                 CERROR("unpack_rep failed: %d\n", rc);
 718                 GOTO(out, rc);
 719         }
 720 #if 0
 721         /* FIXME: Enable when BlueArc makes new release */
 722         if (req->rq_repmsg->type != PTL_RPC_MSG_REPLY &&
 723             req->rq_repmsg->type != PTL_RPC_MSG_ERR) {
 724                 CERROR("invalid packet type received (type=%u)\n",
 725                        req->rq_repmsg->type);
 726                 LBUG();
 727                 GOTO(out, rc = -EINVAL);
 728         }
 729 #endif
 730         DEBUG_REQ(D_NET, req, "status %d", req->rq_repmsg->status);
 731
 732         /* We're a rejected connection, need to invalidate and rebuild. */
 733         if (req->rq_repmsg->status == -ENOTCONN) {
 734                 spin_lock_irqsave(&imp->imp_lock, flags);
 735                 /* If someone else is reconnecting us (CONN_RECOVD) or has
 736                  * already completed it (handle mismatch), then we just need
 737                  * to get out.
 738                  */
 739                 if (imp->imp_level == LUSTRE_CONN_RECOVD ||
 740                     imp->imp_handle.addr != req->rq_reqmsg->addr ||
 741                     imp->imp_handle.cookie != req->rq_reqmsg->cookie) {
 742                         spin_unlock_irqrestore(&imp->imp_lock, flags);
 743                         GOTO(out, rc = -EIO);
 744                 }
 745                 imp->imp_level = LUSTRE_CONN_RECOVD;
 746                 spin_unlock_irqrestore(&imp->imp_lock, flags);
 747                 rc = imp->imp_recover(imp, PTLRPC_RECOVD_PHASE_NOTCONN);
 748                 if (rc)
 749                         LBUG();
 750                 GOTO(out, rc = -EIO);
 751         }
 752
 753         rc = ptlrpc_check_status(req);
 754
 755         if (req->rq_import->imp_flags & IMP_REPLAYABLE) {
 756                 spin_lock_irqsave(&imp->imp_lock, flags);
 757                 if ((req->rq_flags & PTL_RPC_FL_REPLAY || req->rq_transno != 0)
 758                     && rc >= 0) {
 759                         /* Balanced in ptlrpc_free_committed, usually. */
 760                         atomic_inc(&req->rq_refcount);
 761                         list_add_tail(&req->rq_list, &imp->imp_replay_list);
 762                 }
 763
 764                 if (req->rq_transno > imp->imp_max_transno) {
 765                         imp->imp_max_transno = req->rq_transno;
 766                 } else if (req->rq_transno != 0 &&
 767                            imp->imp_level == LUSTRE_CONN_FULL) {
 768                         CDEBUG(D_HA, "got transno "LPD64" after "LPD64
 769                                ": recovery may not work\n", req->rq_transno,
 770                                imp->imp_max_transno);
 771                 }
 772
 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                 spin_unlock_irqrestore(&imp->imp_lock, flags);
 779         }
 780
 781         EXIT;
 782  out:
 783         return rc;
 784 }
 785
 786 int ptlrpc_replay_req(struct ptlrpc_request *req)
 787 {
 788         int rc = 0, old_level, old_status = 0;
 789         // struct ptlrpc_client *cli = req->rq_import->imp_client;
 790         struct l_wait_info lwi;
 791         ENTRY;
 792
 793         init_waitqueue_head(&req->rq_wait_for_rep);
 794         DEBUG_REQ(D_NET, req, "");
 795
 796         req->rq_reqmsg->addr = req->rq_import->imp_handle.addr;
 797         req->rq_reqmsg->cookie = req->rq_import->imp_handle.cookie;
 798
 799         /* temporarily set request to RECOVD level (reset at out:) */
 800         old_level = req->rq_level;
 801         if (req->rq_flags & PTL_RPC_FL_REPLIED)
 802                 old_status = req->rq_repmsg->status;
 803         req->rq_level = LUSTRE_CONN_RECOVD;
 804         rc = ptl_send_rpc(req);
 805         if (rc) {
 806                 CERROR("error %d, opcode %d\n", rc, req->rq_reqmsg->opc);
 807                 ptlrpc_cleanup_request_buf(req);
 808                 // up(&cli->cli_rpc_sem);
 809                 GOTO(out, rc = -rc);
 810         }
 811
 812         CDEBUG(D_OTHER, "-- sleeping\n");
 813         lwi = LWI_INTR(NULL, NULL); /* XXX needs timeout, nested recovery */
 814         l_wait_event(req->rq_wait_for_rep, ptlrpc_check_reply(req), &lwi);
 815         CDEBUG(D_OTHER, "-- done\n");
 816
 817         // up(&cli->cli_rpc_sem);
 818
 819         if (!(req->rq_flags & PTL_RPC_FL_REPLIED)) {
 820                 CERROR("Unknown reason for wakeup\n");
 821                 /* XXX Phil - I end up here when I kill obdctl */
 822                 ptlrpc_abort(req);
 823                 GOTO(out, rc = -EINTR);
 824         }
 825
 826         rc = lustre_unpack_msg(req->rq_repmsg, req->rq_replen);
 827         if (rc) {
 828                 CERROR("unpack_rep failed: %d\n", rc);
 829                 GOTO(out, rc);
 830         }
 831
 832         CDEBUG(D_NET, "got rep "LPD64"\n", req->rq_xid);
 833
 834         /* let the callback do fixups, possibly including in the request */
 835         if (req->rq_replay_cb)
 836                 req->rq_replay_cb(req);
 837
 838         if ((req->rq_flags & PTL_RPC_FL_REPLIED) &&
 839             req->rq_repmsg->status != old_status) {
 840                 DEBUG_REQ(D_HA, req, "status %d, old was %d",
 841                           req->rq_repmsg->status, old_status);
 842         }
 843
 844  out:
 845         req->rq_level = old_level;
 846         RETURN(rc);
 847 }
 848
 849 /* XXX looks a lot like super.c:invalidate_request_list, don't it? */
 850 void ptlrpc_abort_inflight(struct obd_import *imp)
 851 {
 852         unsigned long flags;
 853         struct list_head *tmp, *n;
 854
 855         /* Make sure that no new requests get processed for this import.
 856          * ptlrpc_queue_wait must (and does) hold imp_lock while testing this
 857          * flag and then putting requests on sending_list or delayed_list.
 858          */
 859         spin_lock_irqsave(&imp->imp_lock, flags);
 860         imp->imp_flags |= IMP_INVALID;
 861         spin_unlock_irqrestore(&imp->imp_lock, flags);
 862
 863         list_for_each_safe(tmp, n, &imp->imp_sending_list) {
 864                 struct ptlrpc_request *req =
 865                         list_entry(tmp, struct ptlrpc_request, rq_list);
 866
 867                 DEBUG_REQ(D_HA, req, "inflight");
 868                 req->rq_flags |= PTL_RPC_FL_ERR;
 869                 wake_up(&req->rq_wait_for_rep);
 870         }
 871
 872         list_for_each_safe(tmp, n, &imp->imp_delayed_list) {
 873                 struct ptlrpc_request *req =
 874                         list_entry(tmp, struct ptlrpc_request, rq_list);
 875
 876                 DEBUG_REQ(D_HA, req, "aborting waiting req");
 877                 req->rq_flags |= PTL_RPC_FL_ERR;
 878                 wake_up(&req->rq_wait_for_rep);
 879         }
 880 }