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, 2003 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 #ifndef __KERNEL__
  25 #include <errno.h>
  26 #include <signal.h>
  27 #include <liblustre.h>
  28 #endif
  29
  30 #include <linux/obd_support.h>
  31 #include <linux/obd_class.h>
  32 #include <linux/lustre_lib.h>
  33 #include <linux/lustre_ha.h>
  34 #include <linux/lustre_import.h>
  35
  36 void ptlrpc_init_client(int req_portal, int rep_portal, char *name,
  37                         struct ptlrpc_client *cl)
  38 {
  39         cl->cli_request_portal = req_portal;
  40         cl->cli_reply_portal   = rep_portal;
  41         cl->cli_name           = name;
  42 }
  43
  44 struct obd_uuid *ptlrpc_req_to_uuid(struct ptlrpc_request *req)
  45 {
  46         return &req->rq_connection->c_remote_uuid;
  47 }
  48
  49 struct ptlrpc_connection *ptlrpc_uuid_to_connection(struct obd_uuid *uuid)
  50 {
  51         struct ptlrpc_connection *c;
  52         struct ptlrpc_peer peer;
  53         int err;
  54
  55         err = ptlrpc_uuid_to_peer(uuid, &peer);
  56         if (err != 0) {
  57                 CERROR("cannot find peer %s!\n", uuid->uuid);
  58                 return NULL;
  59         }
  60
  61         c = ptlrpc_get_connection(&peer, uuid);
  62         if (c) {
  63                 memcpy(c->c_remote_uuid.uuid,
  64                        uuid->uuid, sizeof(c->c_remote_uuid.uuid));
  65                 c->c_epoch++;
  66         }
  67
  68         CDEBUG(D_INFO, "%s -> %p\n", uuid->uuid, c);
  69
  70         return c;
  71 }
  72
  73 void ptlrpc_readdress_connection(struct ptlrpc_connection *conn,struct obd_uuid *uuid)
  74 {
  75         struct ptlrpc_peer peer;
  76         int err;
  77
  78         err = ptlrpc_uuid_to_peer (uuid, &peer);
  79         if (err != 0) {
  80                 CERROR("cannot find peer %s!\n", uuid->uuid);
  81                 return;
  82         }
  83
  84         memcpy (&conn->c_peer, &peer, sizeof (peer));
  85         return;
  86 }
  87
  88 struct ptlrpc_bulk_desc *ptlrpc_prep_bulk(struct ptlrpc_connection *conn)
  89 {
  90         struct ptlrpc_bulk_desc *desc;
  91
  92         OBD_ALLOC(desc, sizeof(*desc));
  93         if (desc != NULL) {
  94                 desc->bd_connection = ptlrpc_connection_addref(conn);
  95                 atomic_set(&desc->bd_refcount, 1);
  96                 init_waitqueue_head(&desc->bd_waitq);
  97                 INIT_LIST_HEAD(&desc->bd_page_list);
  98                 INIT_LIST_HEAD(&desc->bd_set_chain);
  99                 ptl_set_inv_handle(&desc->bd_md_h);
 100                 ptl_set_inv_handle(&desc->bd_me_h);
 101         }
 102
 103         return desc;
 104 }
 105
 106 int ptlrpc_bulk_error(struct ptlrpc_bulk_desc *desc)
 107 {
 108         int rc = 0;
 109         if (desc->bd_flags & PTL_RPC_FL_TIMEOUT) {
 110                 rc = (desc->bd_flags & PTL_RPC_FL_INTR ? -ERESTARTSYS :
 111                       -ETIMEDOUT);
 112         }
 113         return rc;
 114 }
 115
 116 struct ptlrpc_bulk_page *ptlrpc_prep_bulk_page(struct ptlrpc_bulk_desc *desc)
 117 {
 118         struct ptlrpc_bulk_page *bulk;
 119
 120         OBD_ALLOC(bulk, sizeof(*bulk));
 121         if (bulk != NULL) {
 122                 bulk->bp_desc = desc;
 123                 list_add_tail(&bulk->bp_link, &desc->bd_page_list);
 124                 desc->bd_page_count++;
 125         }
 126         return bulk;
 127 }
 128
 129 void ptlrpc_free_bulk(struct ptlrpc_bulk_desc *desc)
 130 {
 131         struct list_head *tmp, *next;
 132         ENTRY;
 133         if (desc == NULL) {
 134                 EXIT;
 135                 return;
 136         }
 137
 138         LASSERT(list_empty(&desc->bd_set_chain));
 139
 140         if (atomic_read(&desc->bd_refcount) != 0)
 141                 CERROR("freeing desc %p with refcount %d!\n", desc,
 142                        atomic_read(&desc->bd_refcount));
 143
 144         list_for_each_safe(tmp, next, &desc->bd_page_list) {
 145                 struct ptlrpc_bulk_page *bulk;
 146                 bulk = list_entry(tmp, struct ptlrpc_bulk_page, bp_link);
 147                 ptlrpc_free_bulk_page(bulk);
 148         }
 149
 150         ptlrpc_put_connection(desc->bd_connection);
 151
 152         OBD_FREE(desc, sizeof(*desc));
 153         EXIT;
 154 }
 155
 156 void ptlrpc_free_bulk_page(struct ptlrpc_bulk_page *bulk)
 157 {
 158         ENTRY;
 159         if (bulk == NULL) {
 160                 EXIT;
 161                 return;
 162         }
 163
 164         list_del(&bulk->bp_link);
 165         bulk->bp_desc->bd_page_count--;
 166         OBD_FREE(bulk, sizeof(*bulk));
 167         EXIT;
 168 }
 169
 170 static int ll_sync_brw_timeout(void *data)
 171 {
 172         struct obd_brw_set *set = data;
 173         struct list_head *tmp;
 174         int failed = 0;
 175         ENTRY;
 176
 177         LASSERT(set);
 178
 179         set->brw_flags |= PTL_RPC_FL_TIMEOUT;
 180
 181         list_for_each(tmp, &set->brw_desc_head) {
 182                 struct ptlrpc_bulk_desc *desc =
 183                         list_entry(tmp, struct ptlrpc_bulk_desc, bd_set_chain);
 184
 185                 /* Skip descriptors that were completed successfully. */
 186                 if (desc->bd_flags & (PTL_BULK_FL_RCVD | PTL_BULK_FL_SENT))
 187                         continue;
 188
 189                 LASSERT(desc->bd_connection);
 190
 191                 /* If PtlMDUnlink succeeds, then bulk I/O on the MD hasn't
 192                  * even started yet.  XXX where do we kunmup the thing?
 193                  *
 194                  * If it fail with PTL_MD_BUSY, then the network is still
 195                  * reading/writing the buffers and we must wait for it to
 196                  * complete (which it will within finite time, most
 197                  * probably with failure; we really need portals error
 198                  * events to detect that).
 199                  *
 200                  * Otherwise (PTL_INV_MD) it completed after the bd_flags
 201                  * test above!
 202                  */
 203                 if (PtlMDUnlink(desc->bd_md_h) != PTL_OK) {
 204                         CERROR("Near-miss on OST %s -- need to adjust "
 205                                "obd_timeout?\n",
 206                                desc->bd_connection->c_remote_uuid.uuid);
 207                         continue;
 208                 }
 209
 210                 CERROR("IO of %d pages to/from %s:%d (conn %p) timed out\n",
 211                        desc->bd_page_count,
 212                        desc->bd_connection->c_remote_uuid.uuid,
 213                        desc->bd_portal, desc->bd_connection);
 214
 215                 /* This one will "never" arrive, don't wait for it. */
 216                 if (atomic_dec_and_test(&set->brw_refcount))
 217                         wake_up(&set->brw_waitq);
 218
 219                 if (class_signal_connection_failure)
 220                         class_signal_connection_failure(desc->bd_connection);
 221                 else
 222                         failed = 1;
 223         }
 224
 225         /* 0 = We go back to sleep, until we're resumed or interrupted */
 226         /* 1 = We can't be recovered, just abort the syscall with -ETIMEDOUT */
 227         RETURN(failed);
 228 }
 229
 230 static int ll_sync_brw_intr(void *data)
 231 {
 232         struct obd_brw_set *set = data;
 233
 234         ENTRY;
 235         set->brw_flags |= PTL_RPC_FL_INTR;
 236         RETURN(1); /* ignored, as of this writing */
 237 }
 238
 239 int ll_brw_sync_wait(struct obd_brw_set *set, int phase)
 240 {
 241         struct l_wait_info lwi;
 242         struct list_head *tmp, *next;
 243         int rc = 0;
 244         ENTRY;
 245
 246         switch(phase) {
 247         case CB_PHASE_START:
 248                 lwi = LWI_TIMEOUT_INTR(obd_timeout * HZ, ll_sync_brw_timeout,
 249                                        ll_sync_brw_intr, set);
 250                 rc = l_wait_event(set->brw_waitq,
 251                                   atomic_read(&set->brw_refcount) == 0, &lwi);
 252
 253                 list_for_each_safe(tmp, next, &set->brw_desc_head) {
 254                         struct ptlrpc_bulk_desc *desc =
 255                                 list_entry(tmp, struct ptlrpc_bulk_desc,
 256                                            bd_set_chain);
 257                         list_del_init(&desc->bd_set_chain);
 258                         ptlrpc_bulk_decref(desc);
 259                 }
 260                 break;
 261         case CB_PHASE_FINISH:
 262                 if (atomic_dec_and_test(&set->brw_refcount))
 263                         wake_up(&set->brw_waitq);
 264                 break;
 265         default:
 266                 LBUG();
 267         }
 268
 269         RETURN(rc);
 270 }
 271
 272 struct ptlrpc_request *ptlrpc_prep_req(struct obd_import *imp, int opcode,
 273                                        int count, int *lengths, char **bufs)
 274 {
 275         struct ptlrpc_connection *conn;
 276         struct ptlrpc_request *request;
 277         int rc;
 278         ENTRY;
 279
 280         LASSERT((unsigned long)imp > 0x1000);
 281         conn = imp->imp_connection;
 282
 283         OBD_ALLOC(request, sizeof(*request));
 284         if (!request) {
 285                 CERROR("request allocation out of memory\n");
 286                 RETURN(NULL);
 287         }
 288
 289         rc = lustre_pack_msg(count, lengths, bufs,
 290                              &request->rq_reqlen, &request->rq_reqmsg);
 291         if (rc) {
 292                 CERROR("cannot pack request %d\n", rc);
 293                 OBD_FREE(request, sizeof(*request));
 294                 RETURN(NULL);
 295         }
 296
 297         request->rq_level = LUSTRE_CONN_FULL;
 298         request->rq_type = PTL_RPC_MSG_REQUEST;
 299         request->rq_import = imp;
 300
 301         /* XXX FIXME bug 625069, now 249 */
 302         request->rq_request_portal = imp->imp_client->cli_request_portal;
 303         request->rq_reply_portal = imp->imp_client->cli_reply_portal;
 304
 305         request->rq_connection = ptlrpc_connection_addref(conn);
 306
 307         INIT_LIST_HEAD(&request->rq_list);
 308         atomic_set(&request->rq_refcount, 1);
 309
 310         request->rq_reqmsg->magic = PTLRPC_MSG_MAGIC;
 311         request->rq_reqmsg->version = PTLRPC_MSG_VERSION;
 312         request->rq_reqmsg->opc = HTON__u32(opcode);
 313         request->rq_reqmsg->flags = 0;
 314
 315         ptlrpc_hdl2req(request, &imp->imp_handle);
 316         RETURN(request);
 317 }
 318
 319 static void __ptlrpc_free_req(struct ptlrpc_request *request, int locked)
 320 {
 321         ENTRY;
 322         if (request == NULL) {
 323                 EXIT;
 324                 return;
 325         }
 326
 327         /* We must take it off the imp_replay_list first.  Otherwise, we'll set
 328          * request->rq_reqmsg to NULL while osc_close is dereferencing it. */
 329         if (request->rq_import) {
 330                 unsigned long flags = 0;
 331                 if (!locked)
 332                         spin_lock_irqsave(&request->rq_import->imp_lock, flags);
 333                 list_del_init(&request->rq_list);
 334                 if (!locked)
 335                         spin_unlock_irqrestore(&request->rq_import->imp_lock,
 336                                                flags);
 337         }
 338
 339         if (atomic_read(&request->rq_refcount) != 0) {
 340                 CERROR("freeing request %p (%d->%s:%d) with refcount %d\n",
 341                        request, request->rq_reqmsg->opc,
 342                        request->rq_connection->c_remote_uuid.uuid,
 343                        request->rq_import->imp_client->cli_request_portal,
 344                        atomic_read (&request->rq_refcount));
 345                 LBUG();
 346         }
 347
 348         if (request->rq_repmsg != NULL) {
 349                 OBD_FREE(request->rq_repmsg, request->rq_replen);
 350                 request->rq_repmsg = NULL;
 351                 request->rq_reply_md.start = NULL;
 352         }
 353         if (request->rq_reqmsg != NULL) {
 354                 OBD_FREE(request->rq_reqmsg, request->rq_reqlen);
 355                 request->rq_reqmsg = NULL;
 356         }
 357
 358         ptlrpc_put_connection(request->rq_connection);
 359         OBD_FREE(request, sizeof(*request));
 360         EXIT;
 361 }
 362
 363 void ptlrpc_free_req(struct ptlrpc_request *request)
 364 {
 365         __ptlrpc_free_req(request, 0);
 366 }
 367
 368 static int __ptlrpc_req_finished(struct ptlrpc_request *request, int locked)
 369 {
 370         ENTRY;
 371         if (request == NULL)
 372                 RETURN(1);
 373
 374         if (request == (void *)(long)(0x5a5a5a5a5a5a5a5a)) {
 375                 CERROR("dereferencing freed request (bug 575)\n");
 376                 LBUG();
 377                 RETURN(1);
 378         }
 379
 380         DEBUG_REQ(D_INFO, request, "refcount now %u",
 381                   atomic_read(&request->rq_refcount) - 1);
 382
 383         if (atomic_dec_and_test(&request->rq_refcount)) {
 384                 __ptlrpc_free_req(request, locked);
 385                 RETURN(1);
 386         }
 387
 388         RETURN(0);
 389 }
 390
 391 void ptlrpc_req_finished(struct ptlrpc_request *request)
 392 {
 393         __ptlrpc_req_finished(request, 0);
 394 }
 395
 396 static int ptlrpc_check_reply(struct ptlrpc_request *req)
 397 {
 398         int rc = 0;
 399
 400         ENTRY;
 401         if (req->rq_repmsg != NULL) {
 402                 req->rq_transno = NTOH__u64(req->rq_repmsg->transno);
 403                 /* Store transno in reqmsg for replay. */
 404                 req->rq_reqmsg->transno = req->rq_repmsg->transno;
 405                 req->rq_flags |= PTL_RPC_FL_REPLIED;
 406                 GOTO(out, rc = 1);
 407         }
 408
 409         if (req->rq_flags & PTL_RPC_FL_RESEND) {
 410                 DEBUG_REQ(D_ERROR, req, "RESEND:");
 411                 GOTO(out, rc = 1);
 412         }
 413
 414         if (req->rq_flags & PTL_RPC_FL_ERR) {
 415                 ENTRY;
 416                 DEBUG_REQ(D_ERROR, req, "ABORTED:");
 417                 GOTO(out, rc = 1);
 418         }
 419
 420         if (req->rq_flags & PTL_RPC_FL_RESTART) {
 421                 DEBUG_REQ(D_ERROR, req, "RESTART:");
 422                 GOTO(out, rc = 1);
 423         }
 424         EXIT;
 425  out:
 426         DEBUG_REQ(D_NET, req, "rc = %d for", rc);
 427         return rc;
 428 }
 429
 430 static int ptlrpc_check_status(struct ptlrpc_request *req)
 431 {
 432         int err;
 433         ENTRY;
 434
 435         err = req->rq_repmsg->status;
 436         if (req->rq_repmsg->type == NTOH__u32(PTL_RPC_MSG_ERR)) {
 437                 DEBUG_REQ(D_ERROR, req, "type == PTL_RPC_MSG_ERR (%d)", err);
 438                 RETURN(err ? err : -EINVAL);
 439         }
 440
 441         if (err < 0) {
 442                 DEBUG_REQ(D_INFO, req, "status is %d", err);
 443         } else if (err > 0) {
 444                 /* XXX: translate this error from net to host */
 445                 DEBUG_REQ(D_INFO, req, "status is %d", err);
 446         }
 447
 448         RETURN(err);
 449 }
 450
 451 static void ptlrpc_cleanup_request_buf(struct ptlrpc_request *request)
 452 {
 453         OBD_FREE(request->rq_reqmsg, request->rq_reqlen);
 454         request->rq_reqmsg = NULL;
 455         request->rq_reqlen = 0;
 456 }
 457
 458 /* Abort this request and cleanup any resources associated with it. */
 459 int ptlrpc_abort(struct ptlrpc_request *request)
 460 {
 461         /* First remove the ME for the reply; in theory, this means
 462          * that we can tear down the buffer safely. */
 463         if (PtlMEUnlink(request->rq_reply_me_h) != PTL_OK)
 464                 RETURN(0);
 465         OBD_FREE(request->rq_reply_md.start, request->rq_replen);
 466
 467         memset(&request->rq_reply_me_h, 0, sizeof(request->rq_reply_me_h));
 468         request->rq_reply_md.start = NULL;
 469         request->rq_repmsg = NULL;
 470         return 0;
 471 }
 472
 473 /* caller must hold imp->imp_lock */
 474 void ptlrpc_free_committed(struct obd_import *imp)
 475 {
 476         struct list_head *tmp, *saved;
 477         struct ptlrpc_request *req;
 478         ENTRY;
 479
 480         LASSERT(imp != NULL);
 481
 482 #ifdef CONFIG_SMP
 483         LASSERT(spin_is_locked(&imp->imp_lock));
 484 #endif
 485
 486         CDEBUG(D_HA, "%s: committing for last_committed "LPU64"\n",
 487                imp->imp_obd->obd_name, imp->imp_peer_committed_transno);
 488
 489         list_for_each_safe(tmp, saved, &imp->imp_replay_list) {
 490                 req = list_entry(tmp, struct ptlrpc_request, rq_list);
 491
 492                 if (req->rq_flags & PTL_RPC_FL_REPLAY) {
 493                         DEBUG_REQ(D_HA, req, "keeping (FL_REPLAY)");
 494                         continue;
 495                 }
 496
 497                 /* not yet committed */
 498                 if (req->rq_transno > imp->imp_peer_committed_transno) {
 499                         DEBUG_REQ(D_HA, req, "stopping search");
 500                         break;
 501                 }
 502
 503                 DEBUG_REQ(D_HA, req, "committing (last_committed "LPU64")",
 504                           imp->imp_peer_committed_transno);
 505                 list_del_init(&req->rq_list);
 506                 __ptlrpc_req_finished(req, 1);
 507         }
 508
 509         EXIT;
 510         return;
 511 }
 512
 513 void ptlrpc_cleanup_client(struct obd_import *imp)
 514 {
 515         struct list_head *tmp, *saved;
 516         struct ptlrpc_request *req;
 517         struct ptlrpc_connection *conn = imp->imp_connection;
 518         unsigned long flags;
 519         ENTRY;
 520
 521         LASSERT(conn);
 522
 523         spin_lock_irqsave(&imp->imp_lock, flags);
 524         list_for_each_safe(tmp, saved, &imp->imp_replay_list) {
 525                 req = list_entry(tmp, struct ptlrpc_request, rq_list);
 526
 527                 /* XXX we should make sure that nobody's sleeping on these! */
 528                 DEBUG_REQ(D_HA, req, "cleaning up from sending list");
 529                 list_del_init(&req->rq_list);
 530                 req->rq_import = NULL;
 531                 __ptlrpc_req_finished(req, 0);
 532         }
 533         spin_unlock_irqrestore(&imp->imp_lock, flags);
 534
 535         EXIT;
 536         return;
 537 }
 538
 539 void ptlrpc_continue_req(struct ptlrpc_request *req)
 540 {
 541         DEBUG_REQ(D_HA, req, "continuing delayed request");
 542         req->rq_reqmsg->addr = req->rq_import->imp_handle.addr;
 543         req->rq_reqmsg->cookie = req->rq_import->imp_handle.cookie;
 544         wake_up(&req->rq_wait_for_rep);
 545 }
 546
 547 void ptlrpc_resend_req(struct ptlrpc_request *req)
 548 {
 549         DEBUG_REQ(D_HA, req, "resending");
 550         req->rq_reqmsg->addr = req->rq_import->imp_handle.addr;
 551         req->rq_reqmsg->cookie = req->rq_import->imp_handle.cookie;
 552         req->rq_status = -EAGAIN;
 553         req->rq_level = LUSTRE_CONN_RECOVD;
 554         req->rq_flags |= PTL_RPC_FL_RESEND;
 555         req->rq_flags &= ~PTL_RPC_FL_TIMEOUT;
 556         wake_up(&req->rq_wait_for_rep);
 557 }
 558
 559 void ptlrpc_restart_req(struct ptlrpc_request *req)
 560 {
 561         DEBUG_REQ(D_HA, req, "restarting (possibly-)completed request");
 562         req->rq_status = -ERESTARTSYS;
 563         req->rq_flags |= PTL_RPC_FL_RESTART;
 564         req->rq_flags &= ~PTL_RPC_FL_TIMEOUT;
 565         wake_up(&req->rq_wait_for_rep);
 566 }
 567
 568 static int expired_request(void *data)
 569 {
 570         struct ptlrpc_request *req = data;
 571
 572         ENTRY;
 573         if (!req) {
 574                 CERROR("NULL req!");
 575                 LBUG();
 576                 RETURN(0);
 577         }
 578
 579         DEBUG_REQ(D_ERROR, req, "timeout");
 580         ptlrpc_abort(req);
 581         req->rq_flags |= PTL_RPC_FL_TIMEOUT;
 582
 583         if (!req->rq_import) {
 584                 DEBUG_REQ(D_HA, req, "NULL import; already cleaned up?");
 585                 RETURN(1);
 586         }
 587
 588         if (!req->rq_import->imp_connection) {
 589                 DEBUG_REQ(D_ERROR, req, "NULL connection");
 590                 LBUG();
 591                 RETURN(0);
 592         }
 593
 594         if (!req->rq_import->imp_connection->c_recovd_data.rd_recovd)
 595                 RETURN(1);
 596
 597         recovd_conn_fail(req->rq_import->imp_connection);
 598
 599         /* If this request is for recovery or other primordial tasks,
 600          * don't go back to sleep.
 601          */
 602         if (req->rq_level < LUSTRE_CONN_FULL)
 603                 RETURN(1);
 604         RETURN(0);
 605 }
 606
 607 static int interrupted_request(void *data)
 608 {
 609         struct ptlrpc_request *req = data;
 610         ENTRY;
 611         req->rq_flags |= PTL_RPC_FL_INTR;
 612         RETURN(1); /* ignored, as of this writing */
 613 }
 614
 615 struct ptlrpc_request *ptlrpc_request_addref(struct ptlrpc_request *req)
 616 {
 617         ENTRY;
 618         atomic_inc(&req->rq_refcount);
 619         RETURN(req);
 620 }
 621
 622 void ptlrpc_retain_replayable_request(struct ptlrpc_request *req,
 623                                       struct obd_import *imp)
 624 {
 625         struct list_head *tmp;
 626
 627 #ifdef CONFIG_SMP
 628         LASSERT(spin_is_locked(&imp->imp_lock));
 629 #endif
 630
 631         LASSERT(imp->imp_flags & IMP_REPLAYABLE);
 632         /* Balanced in ptlrpc_free_committed, usually. */
 633         ptlrpc_request_addref(req);
 634         list_for_each_prev(tmp, &imp->imp_replay_list) {
 635                 struct ptlrpc_request *iter =
 636                         list_entry(tmp, struct ptlrpc_request, rq_list);
 637
 638                 /* We may have duplicate transnos if we create and then
 639                  * open a file, or for closes retained if to match creating
 640                  * opens, so use req->rq_xid as a secondary key.
 641                  * (See bugs 684, 685, and 428.)
 642                  */
 643                 if (iter->rq_transno > req->rq_transno)
 644                         continue;
 645
 646                 if (iter->rq_transno == req->rq_transno) {
 647                         LASSERT(iter->rq_xid != req->rq_xid);
 648                         if (iter->rq_xid > req->rq_xid)
 649                                 continue;
 650                 }
 651
 652                 list_add(&req->rq_list, &iter->rq_list);
 653                 return;
 654         }
 655
 656         list_add_tail(&req->rq_list, &imp->imp_replay_list);
 657 }
 658
 659 int ptlrpc_queue_wait(struct ptlrpc_request *req)
 660 {
 661         int rc = 0;
 662         struct l_wait_info lwi;
 663         struct obd_import *imp = req->rq_import;
 664         struct ptlrpc_connection *conn = imp->imp_connection;
 665         unsigned int flags;
 666         ENTRY;
 667
 668         init_waitqueue_head(&req->rq_wait_for_rep);
 669
 670         req->rq_xid = HTON__u32(ptlrpc_next_xid());
 671
 672         /* for distributed debugging */
 673         req->rq_reqmsg->status = HTON__u32(current->pid);
 674         CDEBUG(D_RPCTRACE, "Sending RPC pid:xid:nid:opc %d:"LPU64":%s:"LPX64
 675                ":%d\n", NTOH__u32(req->rq_reqmsg->status), req->rq_xid,
 676                conn->c_peer.peer_ni->pni_name, conn->c_peer.peer_nid,
 677                NTOH__u32(req->rq_reqmsg->opc));
 678
 679         spin_lock_irqsave(&imp->imp_lock, flags);
 680
 681         /*
 682          * If the import has been invalidated (such as by an OST failure), the
 683          * request must fail with -EIO.
 684          */
 685         if (req->rq_import->imp_flags & IMP_INVALID) {
 686                 DEBUG_REQ(D_ERROR, req, "IMP_INVALID:");
 687                 spin_unlock_irqrestore(&imp->imp_lock, flags);
 688                 RETURN(-EIO);
 689         }
 690
 691         if (req->rq_level > imp->imp_level) {
 692                 list_del(&req->rq_list);
 693                 list_add_tail(&req->rq_list, &imp->imp_delayed_list);
 694                 spin_unlock_irqrestore(&imp->imp_lock, flags);
 695
 696                 DEBUG_REQ(D_HA, req, "\"%s\" waiting for recovery: (%d < %d)",
 697                           current->comm, req->rq_level, imp->imp_level);
 698                 lwi = LWI_INTR(NULL, NULL);
 699                 rc = l_wait_event(req->rq_wait_for_rep,
 700                                   (req->rq_level <= imp->imp_level) ||
 701                                   (req->rq_flags & PTL_RPC_FL_ERR), &lwi);
 702
 703                 if (req->rq_flags & PTL_RPC_FL_ERR)
 704                         rc = -EIO;
 705
 706                 if (!req->rq_import)
 707                         RETURN(rc);
 708
 709                 spin_lock_irqsave(&imp->imp_lock, flags);
 710                 list_del_init(&req->rq_list);
 711
 712                 if (rc) {
 713                         spin_unlock_irqrestore(&imp->imp_lock, flags);
 714                         RETURN(rc);
 715                 }
 716
 717                 CERROR("process %d resumed\n", current->pid);
 718         }
 719  resend:
 720
 721         LASSERT(list_empty(&req->rq_list));
 722         list_add_tail(&req->rq_list, &imp->imp_sending_list);
 723         spin_unlock_irqrestore(&imp->imp_lock, flags);
 724         rc = ptl_send_rpc(req);
 725         if (rc) {
 726                 CDEBUG(D_HA, "error %d, opcode %d, need recovery\n", rc,
 727                        req->rq_reqmsg->opc);
 728                 /* sleep for a jiffy, then trigger recovery */
 729                 lwi = LWI_TIMEOUT_INTR(1, expired_request,
 730                                        interrupted_request, req);
 731         } else {
 732                 DEBUG_REQ(D_NET, req, "-- sleeping");
 733                 lwi = LWI_TIMEOUT_INTR(obd_timeout * HZ, expired_request,
 734                                        interrupted_request, req);
 735         }
 736 #ifdef __KERNEL__
 737         l_wait_event(req->rq_wait_for_rep, ptlrpc_check_reply(req), &lwi);
 738 #else
 739         {
 740                 extern int reply_in_callback(ptl_event_t *ev);
 741                 ptl_event_t reply_ev;
 742                 PtlEQWait(req->rq_connection->c_peer.peer_ni->pni_reply_in_eq_h, &reply_ev);
 743                 reply_in_callback(&reply_ev);
 744         }
 745 #endif
 746
 747         DEBUG_REQ(D_NET, req, "-- done sleeping");
 748
 749         spin_lock_irqsave(&imp->imp_lock, flags);
 750         list_del_init(&req->rq_list);
 751         spin_unlock_irqrestore(&imp->imp_lock, flags);
 752
 753         if (req->rq_flags & PTL_RPC_FL_ERR) {
 754                 ptlrpc_abort(req);
 755                 GOTO(out, rc = -EIO);
 756         }
 757
 758         /* Don't resend if we were interrupted. */
 759         if ((req->rq_flags & (PTL_RPC_FL_RESEND | PTL_RPC_FL_INTR)) ==
 760             PTL_RPC_FL_RESEND) {
 761                 if (req->rq_flags & PTL_RPC_FL_NO_RESEND) {
 762                         ptlrpc_abort(req); /* clean up reply buffers */
 763                         req->rq_flags &= ~PTL_RPC_FL_NO_RESEND;
 764                         GOTO(out, rc = -ETIMEDOUT);
 765                 }
 766                 req->rq_flags &= ~PTL_RPC_FL_RESEND;
 767                 lustre_msg_add_flags(req->rq_reqmsg, MSG_RESENT);
 768                 DEBUG_REQ(D_HA, req, "resending: ");
 769                 spin_lock_irqsave(&imp->imp_lock, flags);
 770                 goto resend;
 771         }
 772
 773         if (req->rq_flags & PTL_RPC_FL_INTR) {
 774                 if (!(req->rq_flags & PTL_RPC_FL_TIMEOUT))
 775                         LBUG(); /* should only be interrupted if we timed out */
 776                 /* Clean up the dangling reply buffers */
 777                 ptlrpc_abort(req);
 778                 GOTO(out, rc = -EINTR);
 779         }
 780
 781         if (req->rq_flags & PTL_RPC_FL_TIMEOUT)
 782                 GOTO(out, rc = -ETIMEDOUT);
 783
 784         if (!(req->rq_flags & PTL_RPC_FL_REPLIED))
 785                 GOTO(out, rc = req->rq_status);
 786
 787         rc = lustre_unpack_msg(req->rq_repmsg, req->rq_replen);
 788         if (rc) {
 789                 CERROR("unpack_rep failed: %d\n", rc);
 790                 GOTO(out, rc);
 791         }
 792 #if 0
 793         /* FIXME: Enable when BlueArc makes new release */
 794         if (req->rq_repmsg->type != PTL_RPC_MSG_REPLY &&
 795             req->rq_repmsg->type != PTL_RPC_MSG_ERR) {
 796                 CERROR("invalid packet type received (type=%u)\n",
 797                        req->rq_repmsg->type);
 798                 LBUG();
 799                 GOTO(out, rc = -EINVAL);
 800         }
 801 #endif
 802         DEBUG_REQ(D_NET, req, "status %d", req->rq_repmsg->status);
 803
 804         /* We're a rejected connection, need to invalidate and rebuild. */
 805         if (req->rq_repmsg->status == -ENOTCONN) {
 806                 spin_lock_irqsave(&imp->imp_lock, flags);
 807                 /* If someone else is reconnecting us (CONN_RECOVD) or has
 808                  * already completed it (handle mismatch), then we just need
 809                  * to get out.
 810                  */
 811                 if (imp->imp_level == LUSTRE_CONN_RECOVD ||
 812                     imp->imp_handle.addr != req->rq_reqmsg->addr ||
 813                     imp->imp_handle.cookie != req->rq_reqmsg->cookie) {
 814                         spin_unlock_irqrestore(&imp->imp_lock, flags);
 815                         GOTO(out, rc = -EIO);
 816                 }
 817                 imp->imp_level = LUSTRE_CONN_RECOVD;
 818                 spin_unlock_irqrestore(&imp->imp_lock, flags);
 819                 rc = imp->imp_recover(imp, PTLRPC_RECOVD_PHASE_NOTCONN);
 820                 if (rc)
 821                         LBUG();
 822                 GOTO(out, rc = -EIO);
 823         }
 824
 825         rc = ptlrpc_check_status(req);
 826
 827         if (req->rq_import->imp_flags & IMP_REPLAYABLE) {
 828                 spin_lock_irqsave(&imp->imp_lock, flags);
 829                 if ((req->rq_flags & PTL_RPC_FL_REPLAY || req->rq_transno != 0)
 830                     && rc >= 0) {
 831                         ptlrpc_retain_replayable_request(req, imp);
 832                 }
 833
 834                 if (req->rq_transno > imp->imp_max_transno) {
 835                         imp->imp_max_transno = req->rq_transno;
 836                 }
 837
 838                 /* Replay-enabled imports return commit-status information. */
 839                 if (req->rq_repmsg->last_committed) {
 840                         imp->imp_peer_committed_transno =
 841                                 req->rq_repmsg->last_committed;
 842                 }
 843                 ptlrpc_free_committed(imp);
 844                 spin_unlock_irqrestore(&imp->imp_lock, flags);
 845         }
 846
 847         EXIT;
 848  out:
 849         return rc;
 850 }
 851
 852 int ptlrpc_replay_req(struct ptlrpc_request *req)
 853 {
 854         int rc = 0, old_level, old_status = 0;
 855         // struct ptlrpc_client *cli = req->rq_import->imp_client;
 856         struct l_wait_info lwi;
 857         ENTRY;
 858
 859         init_waitqueue_head(&req->rq_wait_for_rep);
 860         DEBUG_REQ(D_NET, req, "");
 861
 862         req->rq_reqmsg->addr = req->rq_import->imp_handle.addr;
 863         req->rq_reqmsg->cookie = req->rq_import->imp_handle.cookie;
 864
 865         /* temporarily set request to RECOVD level (reset at out:) */
 866         old_level = req->rq_level;
 867         if (req->rq_flags & PTL_RPC_FL_REPLIED)
 868                 old_status = req->rq_repmsg->status;
 869         req->rq_level = LUSTRE_CONN_RECOVD;
 870         rc = ptl_send_rpc(req);
 871         if (rc) {
 872                 CERROR("error %d, opcode %d\n", rc, req->rq_reqmsg->opc);
 873                 ptlrpc_cleanup_request_buf(req);
 874                 // up(&cli->cli_rpc_sem);
 875                 GOTO(out, rc = -rc);
 876         }
 877
 878         CDEBUG(D_OTHER, "-- sleeping\n");
 879         lwi = LWI_INTR(NULL, NULL); /* XXX needs timeout, nested recovery */
 880         l_wait_event(req->rq_wait_for_rep, ptlrpc_check_reply(req), &lwi);
 881         CDEBUG(D_OTHER, "-- done\n");
 882
 883         // up(&cli->cli_rpc_sem);
 884
 885         if (!(req->rq_flags & PTL_RPC_FL_REPLIED)) {
 886                 CERROR("Unknown reason for wakeup\n");
 887                 /* XXX Phil - I end up here when I kill obdctl */
 888                 ptlrpc_abort(req);
 889                 GOTO(out, rc = -EINTR);
 890         }
 891
 892         rc = lustre_unpack_msg(req->rq_repmsg, req->rq_replen);
 893         if (rc) {
 894                 CERROR("unpack_rep failed: %d\n", rc);
 895                 GOTO(out, rc);
 896         }
 897
 898         CDEBUG(D_NET, "got rep "LPD64"\n", req->rq_xid);
 899
 900         /* let the callback do fixups, possibly including in the request */
 901         if (req->rq_replay_cb)
 902                 req->rq_replay_cb(req);
 903
 904         if ((req->rq_flags & PTL_RPC_FL_REPLIED) &&
 905             req->rq_repmsg->status != old_status) {
 906                 DEBUG_REQ(D_HA, req, "status %d, old was %d",
 907                           req->rq_repmsg->status, old_status);
 908         }
 909
 910  out:
 911         req->rq_level = old_level;
 912         RETURN(rc);
 913 }
 914
 915 /* XXX looks a lot like super.c:invalidate_request_list, don't it? */
 916 void ptlrpc_abort_inflight(struct obd_import *imp, int dying_import)
 917 {
 918         unsigned long flags;
 919         struct list_head *tmp, *n;
 920
 921         /* Make sure that no new requests get processed for this import.
 922          * ptlrpc_queue_wait must (and does) hold imp_lock while testing this
 923          * flag and then putting requests on sending_list or delayed_list.
 924          */
 925         if ((imp->imp_flags & IMP_REPLAYABLE) == 0) {
 926                 spin_lock_irqsave(&imp->imp_lock, flags);
 927                 imp->imp_flags |= IMP_INVALID;
 928                 spin_unlock_irqrestore(&imp->imp_lock, flags);
 929         }
 930
 931         list_for_each_safe(tmp, n, &imp->imp_sending_list) {
 932                 struct ptlrpc_request *req =
 933                         list_entry(tmp, struct ptlrpc_request, rq_list);
 934
 935                 DEBUG_REQ(D_HA, req, "inflight");
 936                 req->rq_flags |= PTL_RPC_FL_ERR;
 937                 if (dying_import)
 938                         req->rq_import = NULL;
 939                 wake_up(&req->rq_wait_for_rep);
 940         }
 941
 942         list_for_each_safe(tmp, n, &imp->imp_delayed_list) {
 943                 struct ptlrpc_request *req =
 944                         list_entry(tmp, struct ptlrpc_request, rq_list);
 945
 946                 DEBUG_REQ(D_HA, req, "aborting waiting req");
 947                 req->rq_flags |= PTL_RPC_FL_ERR;
 948                 if (dying_import)
 949                         req->rq_import = NULL;
 950                 wake_up(&req->rq_wait_for_rep);
 951         }
 952 }