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         obd_brw_set_addref(set);
 247         switch(phase) {
 248         case CB_PHASE_START:
 249                 lwi = LWI_TIMEOUT_INTR(obd_timeout * HZ, ll_sync_brw_timeout,
 250                                        ll_sync_brw_intr, set);
 251                 rc = l_wait_event(set->brw_waitq,
 252                                   atomic_read(&set->brw_desc_count) == 0, &lwi);
 253
 254                 list_for_each_safe(tmp, next, &set->brw_desc_head) {
 255                         struct ptlrpc_bulk_desc *desc =
 256                                 list_entry(tmp, struct ptlrpc_bulk_desc,
 257                                            bd_set_chain);
 258                         list_del_init(&desc->bd_set_chain);
 259                         ptlrpc_bulk_decref(desc);
 260                 }
 261                 break;
 262         case CB_PHASE_FINISH:
 263                 if (atomic_dec_and_test(&set->brw_desc_count))
 264                         wake_up(&set->brw_waitq);
 265                 break;
 266         default:
 267                 LBUG();
 268         }
 269         obd_brw_set_decref(set);
 270
 271         RETURN(rc);
 272 }
 273
 274 struct ptlrpc_request *ptlrpc_prep_req(struct obd_import *imp, int opcode,
 275                                        int count, int *lengths, char **bufs)
 276 {
 277         struct ptlrpc_connection *conn;
 278         struct ptlrpc_request *request;
 279         int rc;
 280         ENTRY;
 281
 282         LASSERT((unsigned long)imp > 0x1000);
 283         conn = imp->imp_connection;
 284
 285         OBD_ALLOC(request, sizeof(*request));
 286         if (!request) {
 287                 CERROR("request allocation out of memory\n");
 288                 RETURN(NULL);
 289         }
 290
 291         rc = lustre_pack_msg(count, lengths, bufs,
 292                              &request->rq_reqlen, &request->rq_reqmsg);
 293         if (rc) {
 294                 CERROR("cannot pack request %d\n", rc);
 295                 OBD_FREE(request, sizeof(*request));
 296                 RETURN(NULL);
 297         }
 298
 299         request->rq_timeout = obd_timeout;
 300         request->rq_level = LUSTRE_CONN_FULL;
 301         request->rq_type = PTL_RPC_MSG_REQUEST;
 302         request->rq_import = imp;
 303
 304         /* XXX FIXME bug 625069, now 249 */
 305         request->rq_request_portal = imp->imp_client->cli_request_portal;
 306         request->rq_reply_portal = imp->imp_client->cli_reply_portal;
 307
 308         request->rq_connection = ptlrpc_connection_addref(conn);
 309
 310         INIT_LIST_HEAD(&request->rq_list);
 311         atomic_set(&request->rq_refcount, 1);
 312
 313         request->rq_reqmsg->magic = PTLRPC_MSG_MAGIC;
 314         request->rq_reqmsg->version = PTLRPC_MSG_VERSION;
 315         request->rq_reqmsg->opc = HTON__u32(opcode);
 316         request->rq_reqmsg->flags = 0;
 317
 318         ptlrpc_hdl2req(request, &imp->imp_handle);
 319         RETURN(request);
 320 }
 321
 322 static void __ptlrpc_free_req(struct ptlrpc_request *request, int locked)
 323 {
 324         ENTRY;
 325         if (request == NULL) {
 326                 EXIT;
 327                 return;
 328         }
 329
 330         /* We must take it off the imp_replay_list first.  Otherwise, we'll set
 331          * request->rq_reqmsg to NULL while osc_close is dereferencing it. */
 332         if (request->rq_import) {
 333                 unsigned long flags = 0;
 334                 if (!locked)
 335                         spin_lock_irqsave(&request->rq_import->imp_lock, flags);
 336                 list_del_init(&request->rq_list);
 337                 if (!locked)
 338                         spin_unlock_irqrestore(&request->rq_import->imp_lock,
 339                                                flags);
 340         }
 341
 342         if (atomic_read(&request->rq_refcount) != 0) {
 343                 CERROR("freeing request %p (%d->%s:%d) with refcount %d\n",
 344                        request, request->rq_reqmsg->opc,
 345                        request->rq_connection->c_remote_uuid.uuid,
 346                        request->rq_import->imp_client->cli_request_portal,
 347                        atomic_read (&request->rq_refcount));
 348                 LBUG();
 349         }
 350
 351         if (request->rq_repmsg != NULL) {
 352                 OBD_FREE(request->rq_repmsg, request->rq_replen);
 353                 request->rq_repmsg = NULL;
 354                 request->rq_reply_md.start = NULL;
 355         }
 356         if (request->rq_reqmsg != NULL) {
 357                 OBD_FREE(request->rq_reqmsg, request->rq_reqlen);
 358                 request->rq_reqmsg = NULL;
 359         }
 360
 361         ptlrpc_put_connection(request->rq_connection);
 362         OBD_FREE(request, sizeof(*request));
 363         EXIT;
 364 }
 365
 366 void ptlrpc_free_req(struct ptlrpc_request *request)
 367 {
 368         __ptlrpc_free_req(request, 0);
 369 }
 370
 371 static int __ptlrpc_req_finished(struct ptlrpc_request *request, int locked)
 372 {
 373         ENTRY;
 374         if (request == NULL)
 375                 RETURN(1);
 376
 377         if (request == (void *)(long)(0x5a5a5a5a5a5a5a5a)) {
 378                 CERROR("dereferencing freed request (bug 575)\n");
 379                 LBUG();
 380                 RETURN(1);
 381         }
 382
 383         DEBUG_REQ(D_INFO, request, "refcount now %u",
 384                   atomic_read(&request->rq_refcount) - 1);
 385
 386         if (atomic_dec_and_test(&request->rq_refcount)) {
 387                 __ptlrpc_free_req(request, locked);
 388                 RETURN(1);
 389         }
 390
 391         RETURN(0);
 392 }
 393
 394 void ptlrpc_req_finished(struct ptlrpc_request *request)
 395 {
 396         __ptlrpc_req_finished(request, 0);
 397 }
 398
 399 static int ptlrpc_check_reply(struct ptlrpc_request *req)
 400 {
 401         int rc = 0;
 402
 403         ENTRY;
 404         if (req->rq_repmsg != NULL) {
 405                 req->rq_transno = NTOH__u64(req->rq_repmsg->transno);
 406                 /* Store transno in reqmsg for replay. */
 407                 req->rq_reqmsg->transno = req->rq_repmsg->transno;
 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                 ENTRY;
 419                 DEBUG_REQ(D_ERROR, req, "ABORTED:");
 420                 GOTO(out, rc = 1);
 421         }
 422
 423         if (req->rq_flags & PTL_RPC_FL_RESTART) {
 424                 DEBUG_REQ(D_ERROR, req, "RESTART:");
 425                 GOTO(out, rc = 1);
 426         }
 427         EXIT;
 428  out:
 429         DEBUG_REQ(D_NET, req, "rc = %d for", rc);
 430         return rc;
 431 }
 432
 433 static int ptlrpc_check_status(struct ptlrpc_request *req)
 434 {
 435         int err;
 436         ENTRY;
 437
 438         err = req->rq_repmsg->status;
 439         if (req->rq_repmsg->type == NTOH__u32(PTL_RPC_MSG_ERR)) {
 440                 DEBUG_REQ(D_ERROR, req, "type == PTL_RPC_MSG_ERR (%d)", err);
 441                 RETURN(err ? err : -EINVAL);
 442         }
 443
 444         if (err < 0) {
 445                 DEBUG_REQ(D_INFO, req, "status is %d", err);
 446         } else if (err > 0) {
 447                 /* XXX: translate this error from net to host */
 448                 DEBUG_REQ(D_INFO, req, "status is %d", err);
 449         }
 450
 451         RETURN(err);
 452 }
 453
 454 static void ptlrpc_cleanup_request_buf(struct ptlrpc_request *request)
 455 {
 456         OBD_FREE(request->rq_reqmsg, request->rq_reqlen);
 457         request->rq_reqmsg = NULL;
 458         request->rq_reqlen = 0;
 459 }
 460
 461 /* Abort this request and cleanup any resources associated with it. */
 462 int ptlrpc_abort(struct ptlrpc_request *request)
 463 {
 464         /* First remove the ME for the reply; in theory, this means
 465          * that we can tear down the buffer safely. */
 466         if (PtlMEUnlink(request->rq_reply_me_h) != PTL_OK)
 467                 RETURN(0);
 468         OBD_FREE(request->rq_reply_md.start, request->rq_replen);
 469
 470         memset(&request->rq_reply_me_h, 0, sizeof(request->rq_reply_me_h));
 471         request->rq_reply_md.start = NULL;
 472         request->rq_repmsg = NULL;
 473         return 0;
 474 }
 475
 476 /* caller must hold imp->imp_lock */
 477 void ptlrpc_free_committed(struct obd_import *imp)
 478 {
 479         struct list_head *tmp, *saved;
 480         struct ptlrpc_request *req;
 481         ENTRY;
 482
 483         LASSERT(imp != NULL);
 484
 485 #ifdef CONFIG_SMP
 486         LASSERT(spin_is_locked(&imp->imp_lock));
 487 #endif
 488
 489         CDEBUG(D_HA, "%s: committing for last_committed "LPU64"\n",
 490                imp->imp_obd->obd_name, imp->imp_peer_committed_transno);
 491
 492         list_for_each_safe(tmp, saved, &imp->imp_replay_list) {
 493                 req = list_entry(tmp, struct ptlrpc_request, rq_list);
 494
 495                 if (req->rq_flags & PTL_RPC_FL_REPLAY) {
 496                         DEBUG_REQ(D_HA, req, "keeping (FL_REPLAY)");
 497                         continue;
 498                 }
 499
 500                 /* not yet committed */
 501                 if (req->rq_transno > imp->imp_peer_committed_transno) {
 502                         DEBUG_REQ(D_HA, req, "stopping search");
 503                         break;
 504                 }
 505
 506                 DEBUG_REQ(D_HA, req, "committing (last_committed "LPU64")",
 507                           imp->imp_peer_committed_transno);
 508                 list_del_init(&req->rq_list);
 509                 __ptlrpc_req_finished(req, 1);
 510         }
 511
 512         EXIT;
 513         return;
 514 }
 515
 516 void ptlrpc_cleanup_client(struct obd_import *imp)
 517 {
 518         struct list_head *tmp, *saved;
 519         struct ptlrpc_request *req;
 520         struct ptlrpc_connection *conn = imp->imp_connection;
 521         unsigned long flags;
 522         ENTRY;
 523
 524         LASSERT(conn);
 525
 526         spin_lock_irqsave(&imp->imp_lock, flags);
 527         list_for_each_safe(tmp, saved, &imp->imp_replay_list) {
 528                 req = list_entry(tmp, struct ptlrpc_request, rq_list);
 529
 530                 /* XXX we should make sure that nobody's sleeping on these! */
 531                 DEBUG_REQ(D_HA, req, "cleaning up from sending list");
 532                 list_del_init(&req->rq_list);
 533                 req->rq_import = NULL;
 534                 __ptlrpc_req_finished(req, 0);
 535         }
 536         spin_unlock_irqrestore(&imp->imp_lock, flags);
 537
 538         EXIT;
 539         return;
 540 }
 541
 542 void ptlrpc_continue_req(struct ptlrpc_request *req)
 543 {
 544         DEBUG_REQ(D_HA, req, "continuing delayed request");
 545         req->rq_reqmsg->addr = req->rq_import->imp_handle.addr;
 546         req->rq_reqmsg->cookie = req->rq_import->imp_handle.cookie;
 547         wake_up(&req->rq_wait_for_rep);
 548 }
 549
 550 void ptlrpc_resend_req(struct ptlrpc_request *req)
 551 {
 552         DEBUG_REQ(D_HA, req, "resending");
 553         req->rq_reqmsg->addr = req->rq_import->imp_handle.addr;
 554         req->rq_reqmsg->cookie = req->rq_import->imp_handle.cookie;
 555         req->rq_status = -EAGAIN;
 556         req->rq_level = LUSTRE_CONN_RECOVD;
 557         req->rq_flags |= PTL_RPC_FL_RESEND;
 558         req->rq_flags &= ~PTL_RPC_FL_TIMEOUT;
 559         wake_up(&req->rq_wait_for_rep);
 560 }
 561
 562 void ptlrpc_restart_req(struct ptlrpc_request *req)
 563 {
 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 }
 570
 571 static int expired_request(void *data)
 572 {
 573         struct ptlrpc_request *req = data;
 574
 575         ENTRY;
 576         if (!req) {
 577                 CERROR("NULL req!");
 578                 LBUG();
 579                 RETURN(0);
 580         }
 581
 582         DEBUG_REQ(D_ERROR, req, "timeout");
 583         ptlrpc_abort(req);
 584         req->rq_flags |= PTL_RPC_FL_TIMEOUT;
 585
 586         if (!req->rq_import) {
 587                 DEBUG_REQ(D_HA, req, "NULL import; already cleaned up?");
 588                 RETURN(1);
 589         }
 590
 591         if (!req->rq_import->imp_connection) {
 592                 DEBUG_REQ(D_ERROR, req, "NULL connection");
 593                 LBUG();
 594                 RETURN(0);
 595         }
 596
 597         if (!req->rq_import->imp_connection->c_recovd_data.rd_recovd)
 598                 RETURN(1);
 599
 600         recovd_conn_fail(req->rq_import->imp_connection);
 601
 602         /* If this request is for recovery or other primordial tasks,
 603          * don't go back to sleep.
 604          */
 605         if (req->rq_level < LUSTRE_CONN_FULL)
 606                 RETURN(1);
 607         RETURN(0);
 608 }
 609
 610 static int interrupted_request(void *data)
 611 {
 612         struct ptlrpc_request *req = data;
 613         ENTRY;
 614         req->rq_flags |= PTL_RPC_FL_INTR;
 615         RETURN(1); /* ignored, as of this writing */
 616 }
 617
 618 struct ptlrpc_request *ptlrpc_request_addref(struct ptlrpc_request *req)
 619 {
 620         ENTRY;
 621         atomic_inc(&req->rq_refcount);
 622         RETURN(req);
 623 }
 624
 625 void ptlrpc_retain_replayable_request(struct ptlrpc_request *req,
 626                                       struct obd_import *imp)
 627 {
 628         struct list_head *tmp;
 629
 630 #ifdef CONFIG_SMP
 631         LASSERT(spin_is_locked(&imp->imp_lock));
 632 #endif
 633
 634         LASSERT(imp->imp_flags & IMP_REPLAYABLE);
 635         /* Balanced in ptlrpc_free_committed, usually. */
 636         ptlrpc_request_addref(req);
 637         list_for_each_prev(tmp, &imp->imp_replay_list) {
 638                 struct ptlrpc_request *iter =
 639                         list_entry(tmp, struct ptlrpc_request, rq_list);
 640
 641                 /* We may have duplicate transnos if we create and then
 642                  * open a file, or for closes retained if to match creating
 643                  * opens, so use req->rq_xid as a secondary key.
 644                  * (See bugs 684, 685, and 428.)
 645                  */
 646                 if (iter->rq_transno > req->rq_transno)
 647                         continue;
 648
 649                 if (iter->rq_transno == req->rq_transno) {
 650                         LASSERT(iter->rq_xid != req->rq_xid);
 651                         if (iter->rq_xid > req->rq_xid)
 652                                 continue;
 653                 }
 654
 655                 list_add(&req->rq_list, &iter->rq_list);
 656                 return;
 657         }
 658
 659         list_add_tail(&req->rq_list, &imp->imp_replay_list);
 660 }
 661
 662 int ptlrpc_queue_wait(struct ptlrpc_request *req)
 663 {
 664         int rc = 0;
 665         struct l_wait_info lwi;
 666         struct obd_import *imp = req->rq_import;
 667         struct ptlrpc_connection *conn = imp->imp_connection;
 668         unsigned int flags;
 669         ENTRY;
 670
 671         init_waitqueue_head(&req->rq_wait_for_rep);
 672
 673         req->rq_xid = HTON__u32(ptlrpc_next_xid());
 674
 675         /* for distributed debugging */
 676         req->rq_reqmsg->status = HTON__u32(current->pid);
 677         CDEBUG(D_RPCTRACE, "Sending RPC pid:xid:nid:opc %d:"LPU64":%s:"LPX64
 678                ":%d\n", NTOH__u32(req->rq_reqmsg->status), req->rq_xid,
 679                conn->c_peer.peer_ni->pni_name, conn->c_peer.peer_nid,
 680                NTOH__u32(req->rq_reqmsg->opc));
 681
 682         spin_lock_irqsave(&imp->imp_lock, flags);
 683
 684         /*
 685          * If the import has been invalidated (such as by an OST failure), the
 686          * request must fail with -EIO.
 687          */
 688         if (req->rq_import->imp_flags & IMP_INVALID) {
 689                 DEBUG_REQ(D_ERROR, req, "IMP_INVALID:");
 690                 spin_unlock_irqrestore(&imp->imp_lock, flags);
 691                 RETURN(-EIO);
 692         }
 693
 694         if (req->rq_level > imp->imp_level) {
 695                 list_del(&req->rq_list);
 696                 list_add_tail(&req->rq_list, &imp->imp_delayed_list);
 697                 spin_unlock_irqrestore(&imp->imp_lock, flags);
 698
 699                 DEBUG_REQ(D_HA, req, "\"%s\" waiting for recovery: (%d < %d)",
 700                           current->comm, req->rq_level, imp->imp_level);
 701                 lwi = LWI_INTR(NULL, NULL);
 702                 rc = l_wait_event(req->rq_wait_for_rep,
 703                                   (req->rq_level <= imp->imp_level) ||
 704                                   (req->rq_flags & PTL_RPC_FL_ERR), &lwi);
 705
 706                 if (req->rq_flags & PTL_RPC_FL_ERR)
 707                         rc = -EIO;
 708
 709                 if (!req->rq_import)
 710                         RETURN(rc);
 711
 712                 spin_lock_irqsave(&imp->imp_lock, flags);
 713                 list_del_init(&req->rq_list);
 714
 715                 if (rc) {
 716                         spin_unlock_irqrestore(&imp->imp_lock, flags);
 717                         RETURN(rc);
 718                 }
 719
 720                 CERROR("process %d resumed\n", current->pid);
 721         }
 722  resend:
 723
 724         LASSERT(list_empty(&req->rq_list));
 725         list_add_tail(&req->rq_list, &imp->imp_sending_list);
 726         spin_unlock_irqrestore(&imp->imp_lock, flags);
 727         rc = ptl_send_rpc(req);
 728         if (rc) {
 729                 CDEBUG(D_HA, "error %d, opcode %d, need recovery\n", rc,
 730                        req->rq_reqmsg->opc);
 731                 /* sleep for a jiffy, then trigger recovery */
 732                 lwi = LWI_TIMEOUT_INTR(1, expired_request,
 733                                        interrupted_request, req);
 734         } else {
 735                 DEBUG_REQ(D_NET, req, "-- sleeping");
 736                 lwi = LWI_TIMEOUT_INTR(req->rq_timeout * HZ, expired_request,
 737                                        interrupted_request, req);
 738         }
 739 #ifdef __KERNEL__
 740         l_wait_event(req->rq_wait_for_rep, ptlrpc_check_reply(req), &lwi);
 741 #else
 742         {
 743                 extern int reply_in_callback(ptl_event_t *ev);
 744                 ptl_event_t reply_ev;
 745                 PtlEQWait(req->rq_connection->c_peer.peer_ni->pni_reply_in_eq_h, &reply_ev);
 746                 reply_in_callback(&reply_ev);
 747         }
 748 #endif
 749
 750         DEBUG_REQ(D_NET, req, "-- done sleeping");
 751
 752         spin_lock_irqsave(&imp->imp_lock, flags);
 753         list_del_init(&req->rq_list);
 754         spin_unlock_irqrestore(&imp->imp_lock, flags);
 755
 756         if (req->rq_flags & PTL_RPC_FL_ERR) {
 757                 ptlrpc_abort(req);
 758                 GOTO(out, rc = -EIO);
 759         }
 760
 761         /* Don't resend if we were interrupted. */
 762         if ((req->rq_flags & (PTL_RPC_FL_RESEND | PTL_RPC_FL_INTR)) ==
 763             PTL_RPC_FL_RESEND) {
 764                 if (req->rq_flags & PTL_RPC_FL_NO_RESEND) {
 765                         ptlrpc_abort(req); /* clean up reply buffers */
 766                         req->rq_flags &= ~PTL_RPC_FL_NO_RESEND;
 767                         GOTO(out, rc = -ETIMEDOUT);
 768                 }
 769                 req->rq_flags &= ~PTL_RPC_FL_RESEND;
 770                 lustre_msg_add_flags(req->rq_reqmsg, MSG_RESENT);
 771                 DEBUG_REQ(D_HA, req, "resending: ");
 772                 spin_lock_irqsave(&imp->imp_lock, flags);
 773                 goto resend;
 774         }
 775
 776         if (req->rq_flags & PTL_RPC_FL_INTR) {
 777                 if (!(req->rq_flags & PTL_RPC_FL_TIMEOUT))
 778                         LBUG(); /* should only be interrupted if we timed out */
 779                 /* Clean up the dangling reply buffers */
 780                 ptlrpc_abort(req);
 781                 GOTO(out, rc = -EINTR);
 782         }
 783
 784         if (req->rq_flags & PTL_RPC_FL_TIMEOUT)
 785                 GOTO(out, rc = -ETIMEDOUT);
 786
 787         if (!(req->rq_flags & PTL_RPC_FL_REPLIED))
 788                 GOTO(out, rc = req->rq_status);
 789
 790         rc = lustre_unpack_msg(req->rq_repmsg, req->rq_replen);
 791         if (rc) {
 792                 CERROR("unpack_rep failed: %d\n", rc);
 793                 GOTO(out, rc);
 794         }
 795 #if 0
 796         /* FIXME: Enable when BlueArc makes new release */
 797         if (req->rq_repmsg->type != PTL_RPC_MSG_REPLY &&
 798             req->rq_repmsg->type != PTL_RPC_MSG_ERR) {
 799                 CERROR("invalid packet type received (type=%u)\n",
 800                        req->rq_repmsg->type);
 801                 LBUG();
 802                 GOTO(out, rc = -EINVAL);
 803         }
 804 #endif
 805         DEBUG_REQ(D_NET, req, "status %d", req->rq_repmsg->status);
 806
 807         /* We're a rejected connection, need to invalidate and rebuild. */
 808         if (req->rq_repmsg->status == -ENOTCONN) {
 809                 spin_lock_irqsave(&imp->imp_lock, flags);
 810                 /* If someone else is reconnecting us (CONN_RECOVD) or has
 811                  * already completed it (handle mismatch), then we just need
 812                  * to get out.
 813                  */
 814                 if (imp->imp_level == LUSTRE_CONN_RECOVD ||
 815                     imp->imp_handle.addr != req->rq_reqmsg->addr ||
 816                     imp->imp_handle.cookie != req->rq_reqmsg->cookie) {
 817                         spin_unlock_irqrestore(&imp->imp_lock, flags);
 818                         GOTO(out, rc = -EIO);
 819                 }
 820                 imp->imp_level = LUSTRE_CONN_RECOVD;
 821                 spin_unlock_irqrestore(&imp->imp_lock, flags);
 822                 if (imp->imp_recover != NULL) {
 823                         rc = imp->imp_recover(imp, PTLRPC_RECOVD_PHASE_NOTCONN);
 824                         if (rc)
 825                                 LBUG();
 826                 }
 827                 GOTO(out, rc = -EIO);
 828         }
 829
 830         rc = ptlrpc_check_status(req);
 831
 832         if (req->rq_import->imp_flags & IMP_REPLAYABLE) {
 833                 spin_lock_irqsave(&imp->imp_lock, flags);
 834                 if ((req->rq_flags & PTL_RPC_FL_REPLAY || req->rq_transno != 0)
 835                     && rc >= 0) {
 836                         ptlrpc_retain_replayable_request(req, imp);
 837                 }
 838
 839                 if (req->rq_transno > imp->imp_max_transno) {
 840                         imp->imp_max_transno = req->rq_transno;
 841                 }
 842
 843                 /* Replay-enabled imports return commit-status information. */
 844                 if (req->rq_repmsg->last_committed) {
 845                         imp->imp_peer_committed_transno =
 846                                 req->rq_repmsg->last_committed;
 847                 }
 848                 ptlrpc_free_committed(imp);
 849                 spin_unlock_irqrestore(&imp->imp_lock, flags);
 850         }
 851
 852         EXIT;
 853  out:
 854         return rc;
 855 }
 856
 857 int ptlrpc_replay_req(struct ptlrpc_request *req)
 858 {
 859         int rc = 0, old_level, old_status = 0;
 860         // struct ptlrpc_client *cli = req->rq_import->imp_client;
 861         struct l_wait_info lwi;
 862         ENTRY;
 863
 864         init_waitqueue_head(&req->rq_wait_for_rep);
 865         DEBUG_REQ(D_NET, req, "");
 866
 867         req->rq_reqmsg->addr = req->rq_import->imp_handle.addr;
 868         req->rq_reqmsg->cookie = req->rq_import->imp_handle.cookie;
 869
 870         /* temporarily set request to RECOVD level (reset at out:) */
 871         old_level = req->rq_level;
 872         if (req->rq_flags & PTL_RPC_FL_REPLIED)
 873                 old_status = req->rq_repmsg->status;
 874         req->rq_level = LUSTRE_CONN_RECOVD;
 875         rc = ptl_send_rpc(req);
 876         if (rc) {
 877                 CERROR("error %d, opcode %d\n", rc, req->rq_reqmsg->opc);
 878                 ptlrpc_cleanup_request_buf(req);
 879                 // up(&cli->cli_rpc_sem);
 880                 GOTO(out, rc = -rc);
 881         }
 882
 883         CDEBUG(D_OTHER, "-- sleeping\n");
 884         lwi = LWI_INTR(NULL, NULL); /* XXX needs timeout, nested recovery */
 885         l_wait_event(req->rq_wait_for_rep, ptlrpc_check_reply(req), &lwi);
 886         CDEBUG(D_OTHER, "-- done\n");
 887
 888         // up(&cli->cli_rpc_sem);
 889
 890         if (!(req->rq_flags & PTL_RPC_FL_REPLIED)) {
 891                 CERROR("Unknown reason for wakeup\n");
 892                 /* XXX Phil - I end up here when I kill obdctl */
 893                 ptlrpc_abort(req);
 894                 GOTO(out, rc = -EINTR);
 895         }
 896
 897         rc = lustre_unpack_msg(req->rq_repmsg, req->rq_replen);
 898         if (rc) {
 899                 CERROR("unpack_rep failed: %d\n", rc);
 900                 GOTO(out, rc);
 901         }
 902
 903         CDEBUG(D_NET, "got rep "LPD64"\n", req->rq_xid);
 904
 905         /* let the callback do fixups, possibly including in the request */
 906         if (req->rq_replay_cb)
 907                 req->rq_replay_cb(req);
 908
 909         if ((req->rq_flags & PTL_RPC_FL_REPLIED) &&
 910             req->rq_repmsg->status != old_status) {
 911                 DEBUG_REQ(D_HA, req, "status %d, old was %d",
 912                           req->rq_repmsg->status, old_status);
 913         }
 914
 915  out:
 916         req->rq_level = old_level;
 917         RETURN(rc);
 918 }
 919
 920 /* XXX looks a lot like super.c:invalidate_request_list, don't it? */
 921 void ptlrpc_abort_inflight(struct obd_import *imp, int dying_import)
 922 {
 923         unsigned long flags;
 924         struct list_head *tmp, *n;
 925         ENTRY;
 926
 927         /* Make sure that no new requests get processed for this import.
 928          * ptlrpc_queue_wait must (and does) hold imp_lock while testing this
 929          * flag and then putting requests on sending_list or delayed_list.
 930          */
 931         if ((imp->imp_flags & IMP_REPLAYABLE) == 0) {
 932                 spin_lock_irqsave(&imp->imp_lock, flags);
 933                 imp->imp_flags |= IMP_INVALID;
 934                 spin_unlock_irqrestore(&imp->imp_lock, flags);
 935         }
 936
 937         list_for_each_safe(tmp, n, &imp->imp_sending_list) {
 938                 struct ptlrpc_request *req =
 939                         list_entry(tmp, struct ptlrpc_request, rq_list);
 940
 941                 DEBUG_REQ(D_HA, req, "inflight");
 942                 req->rq_flags |= PTL_RPC_FL_ERR;
 943                 if (dying_import)
 944                         req->rq_import = NULL;
 945                 wake_up(&req->rq_wait_for_rep);
 946         }
 947
 948         list_for_each_safe(tmp, n, &imp->imp_delayed_list) {
 949                 struct ptlrpc_request *req =
 950                         list_entry(tmp, struct ptlrpc_request, rq_list);
 951
 952                 DEBUG_REQ(D_HA, req, "aborting waiting req");
 953                 req->rq_flags |= PTL_RPC_FL_ERR;
 954                 if (dying_import)
 955                         req->rq_import = NULL;
 956                 wake_up(&req->rq_wait_for_rep);
 957         }
 958         EXIT;
 959 }