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 #include "ptlrpc_internal.h"
  37
  38 void ptlrpc_init_client(int req_portal, int rep_portal, char *name,
  39                         struct ptlrpc_client *cl)
  40 {
  41         cl->cli_request_portal = req_portal;
  42         cl->cli_reply_portal   = rep_portal;
  43         cl->cli_name           = name;
  44 }
  45
  46 struct ptlrpc_connection *ptlrpc_uuid_to_connection(struct obd_uuid *uuid)
  47 {
  48         struct ptlrpc_connection *c;
  49         struct ptlrpc_peer peer;
  50         int err;
  51
  52         err = ptlrpc_uuid_to_peer(uuid, &peer);
  53         if (err != 0) {
  54                 CERROR("cannot find peer %s!\n", uuid->uuid);
  55                 return NULL;
  56         }
  57
  58         c = ptlrpc_get_connection(&peer, uuid);
  59         if (c) {
  60                 memcpy(c->c_remote_uuid.uuid,
  61                        uuid->uuid, sizeof(c->c_remote_uuid.uuid));
  62         }
  63
  64         CDEBUG(D_INFO, "%s -> %p\n", uuid->uuid, c);
  65
  66         return c;
  67 }
  68
  69 void ptlrpc_readdress_connection(struct ptlrpc_connection *conn,
  70                                  struct obd_uuid *uuid)
  71 {
  72         struct ptlrpc_peer peer;
  73         int err;
  74
  75         err = ptlrpc_uuid_to_peer(uuid, &peer);
  76         if (err != 0) {
  77                 CERROR("cannot find peer %s!\n", uuid->uuid);
  78                 return;
  79         }
  80
  81         memcpy(&conn->c_peer, &peer, sizeof (peer));
  82         return;
  83 }
  84
  85 static inline struct ptlrpc_bulk_desc *new_bulk(int npages, int type, int portal)
  86 {
  87         struct ptlrpc_bulk_desc *desc;
  88
  89         OBD_ALLOC(desc, offsetof (struct ptlrpc_bulk_desc, bd_iov[npages]));
  90         if (!desc)
  91                 return NULL;
  92
  93         spin_lock_init(&desc->bd_lock);
  94         init_waitqueue_head(&desc->bd_waitq);
  95         desc->bd_max_iov = npages;
  96         desc->bd_iov_count = 0;
  97         desc->bd_md_h = PTL_INVALID_HANDLE;
  98         desc->bd_portal = portal;
  99         desc->bd_type = type;
 100
 101         return desc;
 102 }
 103
 104 struct ptlrpc_bulk_desc *ptlrpc_prep_bulk_imp (struct ptlrpc_request *req,
 105                                                int npages, int type, int portal)
 106 {
 107         struct obd_import *imp = req->rq_import;
 108         struct ptlrpc_bulk_desc *desc;
 109
 110         LASSERT(type == BULK_PUT_SINK || type == BULK_GET_SOURCE);
 111         desc = new_bulk(npages, type, portal);
 112         if (desc == NULL)
 113                 RETURN(NULL);
 114
 115         desc->bd_import_generation = req->rq_import_generation;
 116         desc->bd_import = class_import_get(imp);
 117         desc->bd_req = req;
 118
 119         desc->bd_cbid.cbid_fn  = client_bulk_callback;
 120         desc->bd_cbid.cbid_arg = desc;
 121
 122         /* This makes req own desc, and free it when she frees herself */
 123         req->rq_bulk = desc;
 124
 125         return desc;
 126 }
 127
 128 struct ptlrpc_bulk_desc *ptlrpc_prep_bulk_exp (struct ptlrpc_request *req,
 129                                                int npages, int type, int portal)
 130 {
 131         struct obd_export *exp = req->rq_export;
 132         struct ptlrpc_bulk_desc *desc;
 133
 134         LASSERT(type == BULK_PUT_SOURCE || type == BULK_GET_SINK);
 135
 136         desc = new_bulk(npages, type, portal);
 137         if (desc == NULL)
 138                 RETURN(NULL);
 139
 140         desc->bd_export = class_export_get(exp);
 141         desc->bd_req = req;
 142
 143         desc->bd_cbid.cbid_fn  = server_bulk_callback;
 144         desc->bd_cbid.cbid_arg = desc;
 145
 146         /* NB we don't assign rq_bulk here; server-side requests are
 147          * re-used, and the handler frees the bulk desc explicitly. */
 148
 149         return desc;
 150 }
 151
 152 void ptlrpc_prep_bulk_page(struct ptlrpc_bulk_desc *desc,
 153                            struct page *page, int pageoffset, int len)
 154 {
 155         LASSERT(desc->bd_iov_count < desc->bd_max_iov);
 156         LASSERT(page != NULL);
 157         LASSERT(pageoffset >= 0);
 158         LASSERT(len > 0);
 159         LASSERT(pageoffset + len <= PAGE_SIZE);
 160
 161         desc->bd_nob += len;
 162
 163         pers_bulk_add_page(desc, page, pageoffset, len);
 164 }
 165
 166 void ptlrpc_free_bulk(struct ptlrpc_bulk_desc *desc)
 167 {
 168         ENTRY;
 169
 170         LASSERT(desc != NULL);
 171         LASSERT(desc->bd_iov_count != 0x5a5a5a5a); /* not freed already */
 172         LASSERT(!desc->bd_network_rw);         /* network hands off or */
 173         LASSERT((desc->bd_export != NULL) ^ (desc->bd_import != NULL));
 174         if (desc->bd_export)
 175                 class_export_put(desc->bd_export);
 176         else
 177                 class_import_put(desc->bd_import);
 178
 179         OBD_FREE(desc, offsetof(struct ptlrpc_bulk_desc,
 180                                 bd_iov[desc->bd_max_iov]));
 181         EXIT;
 182 }
 183
 184 struct ptlrpc_request *ptlrpc_prep_req(struct obd_import *imp, int opcode,
 185                                        int count, int *lengths, char **bufs)
 186 {
 187         struct ptlrpc_request *request;
 188         int rc;
 189         ENTRY;
 190
 191         LASSERT((unsigned long)imp > 0x1000);
 192
 193         OBD_ALLOC(request, sizeof(*request));
 194         if (!request) {
 195                 CERROR("request allocation out of memory\n");
 196                 RETURN(NULL);
 197         }
 198
 199         rc = lustre_pack_request(request, count, lengths, bufs);
 200         if (rc) {
 201                 CERROR("cannot pack request %d\n", rc);
 202                 OBD_FREE(request, sizeof(*request));
 203                 RETURN(NULL);
 204         }
 205
 206         if (imp->imp_server_timeout)
 207                 request->rq_timeout = obd_timeout / 2;
 208         else
 209                 request->rq_timeout = obd_timeout;
 210         request->rq_send_state = LUSTRE_IMP_FULL;
 211         request->rq_type = PTL_RPC_MSG_REQUEST;
 212         request->rq_import = class_import_get(imp);
 213
 214         request->rq_req_cbid.cbid_fn  = request_out_callback;
 215         request->rq_req_cbid.cbid_arg = request;
 216
 217         request->rq_reply_cbid.cbid_fn  = reply_in_callback;
 218         request->rq_reply_cbid.cbid_arg = request;
 219
 220         request->rq_phase = RQ_PHASE_NEW;
 221
 222         /* XXX FIXME bug 249 */
 223         request->rq_request_portal = imp->imp_client->cli_request_portal;
 224         request->rq_reply_portal = imp->imp_client->cli_reply_portal;
 225
 226         spin_lock_init(&request->rq_lock);
 227         INIT_LIST_HEAD(&request->rq_list);
 228         INIT_LIST_HEAD(&request->rq_replay_list);
 229         INIT_LIST_HEAD(&request->rq_set_chain);
 230         init_waitqueue_head(&request->rq_reply_waitq);
 231         request->rq_xid = ptlrpc_next_xid();
 232         atomic_set(&request->rq_refcount, 1);
 233
 234         request->rq_reqmsg->opc = opcode;
 235         request->rq_reqmsg->flags = 0;
 236
 237         RETURN(request);
 238 }
 239
 240 struct ptlrpc_request_set *ptlrpc_prep_set(void)
 241 {
 242         struct ptlrpc_request_set *set;
 243
 244         OBD_ALLOC(set, sizeof *set);
 245         if (!set)
 246                 RETURN(NULL);
 247         INIT_LIST_HEAD(&set->set_requests);
 248         init_waitqueue_head(&set->set_waitq);
 249         set->set_remaining = 0;
 250         spin_lock_init(&set->set_new_req_lock);
 251         INIT_LIST_HEAD(&set->set_new_requests);
 252
 253         RETURN(set);
 254 }
 255
 256 /* Finish with this set; opposite of prep_set. */
 257 void ptlrpc_set_destroy(struct ptlrpc_request_set *set)
 258 {
 259         struct list_head *tmp;
 260         struct list_head *next;
 261         int               expected_phase;
 262         int               n = 0;
 263         ENTRY;
 264
 265         /* Requests on the set should either all be completed, or all be new */
 266         expected_phase = (set->set_remaining == 0) ?
 267                          RQ_PHASE_COMPLETE : RQ_PHASE_NEW;
 268         list_for_each (tmp, &set->set_requests) {
 269                 struct ptlrpc_request *req =
 270                         list_entry(tmp, struct ptlrpc_request, rq_set_chain);
 271
 272                 LASSERT(req->rq_phase == expected_phase);
 273                 n++;
 274         }
 275
 276         LASSERT(set->set_remaining == 0 || set->set_remaining == n);
 277
 278         list_for_each_safe(tmp, next, &set->set_requests) {
 279                 struct ptlrpc_request *req =
 280                         list_entry(tmp, struct ptlrpc_request, rq_set_chain);
 281                 list_del_init(&req->rq_set_chain);
 282
 283                 LASSERT(req->rq_phase == expected_phase);
 284
 285                 if (req->rq_phase == RQ_PHASE_NEW) {
 286
 287                         if (req->rq_interpret_reply != NULL) {
 288                                 int (*interpreter)(struct ptlrpc_request *,
 289                                                    void *, int) =
 290                                         req->rq_interpret_reply;
 291
 292                                 /* higher level (i.e. LOV) failed;
 293                                  * let the sub reqs clean up */
 294                                 req->rq_status = -EBADR;
 295                                 interpreter(req, &req->rq_async_args,
 296                                             req->rq_status);
 297                         }
 298                         set->set_remaining--;
 299                 }
 300
 301                 req->rq_set = NULL;
 302                 ptlrpc_req_finished (req);
 303         }
 304
 305         LASSERT(set->set_remaining == 0);
 306
 307         OBD_FREE(set, sizeof(*set));
 308         EXIT;
 309 }
 310
 311 void ptlrpc_set_add_req(struct ptlrpc_request_set *set,
 312                         struct ptlrpc_request *req)
 313 {
 314         /* The set takes over the caller's request reference */
 315         list_add_tail(&req->rq_set_chain, &set->set_requests);
 316         req->rq_set = set;
 317         set->set_remaining++;
 318         atomic_inc(&req->rq_import->imp_inflight);
 319 }
 320
 321 /* lock so many callers can add things, the context that owns the set
 322  * is supposed to notice these and move them into the set proper. */
 323 void ptlrpc_set_add_new_req(struct ptlrpc_request_set *set,
 324                             struct ptlrpc_request *req)
 325 {
 326         unsigned long flags;
 327         spin_lock_irqsave(&set->set_new_req_lock, flags);
 328         /* The set takes over the caller's request reference */
 329         list_add_tail(&req->rq_set_chain, &set->set_new_requests);
 330         req->rq_set = set;
 331         spin_unlock_irqrestore(&set->set_new_req_lock, flags);
 332 }
 333
 334 /*
 335  * Based on the current state of the import, determine if the request
 336  * can be sent, is an error, or should be delayed.
 337  *
 338  * Returns true if this request should be delayed. If false, and
 339  * *status is set, then the request can not be sent and *status is the
 340  * error code.  If false and status is 0, then request can be sent.
 341  *
 342  * The imp->imp_lock must be held.
 343  */
 344 static int ptlrpc_import_delay_req(struct obd_import *imp,
 345                                    struct ptlrpc_request *req, int *status)
 346 {
 347         int delay = 0;
 348         ENTRY;
 349
 350         LASSERT (status != NULL);
 351         *status = 0;
 352
 353         if (imp->imp_state == LUSTRE_IMP_NEW) {
 354                 DEBUG_REQ(D_ERROR, req, "Uninitialized import.");
 355                 *status = -EIO;
 356                 LBUG();
 357         }
 358         else if (imp->imp_state == LUSTRE_IMP_CLOSED) {
 359                 DEBUG_REQ(D_ERROR, req, "IMP_CLOSED ");
 360                 *status = -EIO;
 361         }
 362         /* allow CONNECT even if import is invalid */
 363         else if (req->rq_send_state == LUSTRE_IMP_CONNECTING &&
 364                  imp->imp_state == LUSTRE_IMP_CONNECTING) {
 365                 ;
 366         }
 367         /*
 368          * If the import has been invalidated (such as by an OST failure), the
 369          * request must fail with -EIO.
 370          */
 371         else if (imp->imp_invalid) {
 372                 DEBUG_REQ(D_ERROR, req, "IMP_INVALID");
 373                 *status = -EIO;
 374         }
 375         else if (req->rq_import_generation != imp->imp_generation) {
 376                 DEBUG_REQ(D_ERROR, req, "req wrong generation:");
 377                 *status = -EIO;
 378         }
 379         else if (req->rq_send_state != imp->imp_state) {
 380                 if (imp->imp_obd->obd_no_recov || imp->imp_dlm_fake
 381                     || req->rq_no_delay)
 382                         *status = -EWOULDBLOCK;
 383                 else
 384                         delay = 1;
 385         }
 386
 387         RETURN(delay);
 388 }
 389
 390 static int ptlrpc_check_reply(struct ptlrpc_request *req)
 391 {
 392         unsigned long flags;
 393         int rc = 0;
 394         ENTRY;
 395
 396         /* serialise with network callback */
 397         spin_lock_irqsave (&req->rq_lock, flags);
 398
 399         if (req->rq_replied) {
 400                 DEBUG_REQ(D_NET, req, "REPLIED:");
 401                 GOTO(out, rc = 1);
 402         }
 403
 404         if (req->rq_err) {
 405                 DEBUG_REQ(D_ERROR, req, "ABORTED:");
 406                 GOTO(out, rc = 1);
 407         }
 408
 409         if (req->rq_resend) {
 410                 DEBUG_REQ(D_ERROR, req, "RESEND:");
 411                 GOTO(out, rc = 1);
 412         }
 413
 414         if (req->rq_restart) {
 415                 DEBUG_REQ(D_ERROR, req, "RESTART:");
 416                 GOTO(out, rc = 1);
 417         }
 418         EXIT;
 419  out:
 420         spin_unlock_irqrestore (&req->rq_lock, flags);
 421         DEBUG_REQ(D_NET, req, "rc = %d for", rc);
 422         return rc;
 423 }
 424
 425 static int ptlrpc_check_status(struct ptlrpc_request *req)
 426 {
 427         int err;
 428         ENTRY;
 429
 430         err = req->rq_repmsg->status;
 431         if (req->rq_repmsg->type == PTL_RPC_MSG_ERR) {
 432                 DEBUG_REQ(D_ERROR, req, "type == PTL_RPC_MSG_ERR, err == %d",
 433                           err);
 434                 RETURN(err < 0 ? err : -EINVAL);
 435         }
 436
 437         if (err < 0) {
 438                 DEBUG_REQ(D_INFO, req, "status is %d", err);
 439         } else if (err > 0) {
 440                 /* XXX: translate this error from net to host */
 441                 DEBUG_REQ(D_INFO, req, "status is %d", err);
 442         }
 443
 444         RETURN(err);
 445 }
 446
 447 static int after_reply(struct ptlrpc_request *req)
 448 {
 449         unsigned long flags;
 450         struct obd_import *imp = req->rq_import;
 451         int rc;
 452         ENTRY;
 453
 454         LASSERT(!req->rq_receiving_reply);
 455
 456         /* NB Until this point, the whole of the incoming message,
 457          * including buflens, status etc is in the sender's byte order. */
 458
 459 #if SWAB_PARANOIA
 460         /* Clear reply swab mask; this is a new reply in sender's byte order */
 461         req->rq_rep_swab_mask = 0;
 462 #endif
 463         LASSERT (req->rq_nob_received <= req->rq_replen);
 464         rc = lustre_unpack_msg(req->rq_repmsg, req->rq_nob_received);
 465         if (rc) {
 466                 CERROR("unpack_rep failed: %d\n", rc);
 467                 RETURN(-EPROTO);
 468         }
 469
 470         if (req->rq_repmsg->type != PTL_RPC_MSG_REPLY &&
 471             req->rq_repmsg->type != PTL_RPC_MSG_ERR) {
 472                 CERROR("invalid packet type received (type=%u)\n",
 473                        req->rq_repmsg->type);
 474                 RETURN(-EPROTO);
 475         }
 476
 477         /* Store transno in reqmsg for replay. */
 478         req->rq_reqmsg->transno = req->rq_transno = req->rq_repmsg->transno;
 479
 480         rc = ptlrpc_check_status(req);
 481
 482         /* Either we've been evicted, or the server has failed for
 483          * some reason. Try to reconnect, and if that fails, punt to the
 484          * upcall. */
 485         if (rc == -ENOTCONN) {
 486                 if (req->rq_send_state != LUSTRE_IMP_FULL ||
 487                     imp->imp_obd->obd_no_recov || imp->imp_dlm_fake) {
 488                         RETURN(-ENOTCONN);
 489                 }
 490
 491                 ptlrpc_request_handle_notconn(req);
 492
 493                 RETURN(rc);
 494         }
 495
 496         if (req->rq_import->imp_replayable) {
 497                 spin_lock_irqsave(&imp->imp_lock, flags);
 498                 if (req->rq_replay || req->rq_transno != 0)
 499                         ptlrpc_retain_replayable_request(req, imp);
 500                 else if (req->rq_commit_cb != NULL)
 501                         req->rq_commit_cb(req);
 502
 503                 if (req->rq_transno > imp->imp_max_transno)
 504                         imp->imp_max_transno = req->rq_transno;
 505
 506                 /* Replay-enabled imports return commit-status information. */
 507                 if (req->rq_repmsg->last_committed)
 508                         imp->imp_peer_committed_transno =
 509                                 req->rq_repmsg->last_committed;
 510                 ptlrpc_free_committed(imp);
 511                 spin_unlock_irqrestore(&imp->imp_lock, flags);
 512         }
 513
 514         RETURN(rc);
 515 }
 516
 517 static int ptlrpc_send_new_req(struct ptlrpc_request *req)
 518 {
 519         struct obd_import     *imp;
 520         unsigned long          flags;
 521         int rc;
 522         ENTRY;
 523
 524         LASSERT(req->rq_phase == RQ_PHASE_NEW);
 525         req->rq_phase = RQ_PHASE_RPC;
 526
 527         imp = req->rq_import;
 528         spin_lock_irqsave(&imp->imp_lock, flags);
 529
 530         req->rq_import_generation = imp->imp_generation;
 531
 532         if (ptlrpc_import_delay_req(imp, req, &rc)) {
 533                 spin_lock (&req->rq_lock);
 534                 req->rq_waiting = 1;
 535                 spin_unlock (&req->rq_lock);
 536
 537                 DEBUG_REQ(D_HA, req, "req from PID %d waiting for recovery: "
 538                           "(%s != %s)",
 539                           req->rq_reqmsg->status,
 540                           ptlrpc_import_state_name(req->rq_send_state),
 541                           ptlrpc_import_state_name(imp->imp_state));
 542                 LASSERT(list_empty (&req->rq_list));
 543
 544                 list_add_tail(&req->rq_list, &imp->imp_delayed_list);
 545                 spin_unlock_irqrestore(&imp->imp_lock, flags);
 546                 RETURN(0);
 547         }
 548
 549         if (rc != 0) {
 550                 spin_unlock_irqrestore(&imp->imp_lock, flags);
 551                 req->rq_status = rc;
 552                 req->rq_phase = RQ_PHASE_INTERPRET;
 553                 RETURN(rc);
 554         }
 555
 556         /* XXX this is the same as ptlrpc_queue_wait */
 557         LASSERT(list_empty(&req->rq_list));
 558         list_add_tail(&req->rq_list, &imp->imp_sending_list);
 559         spin_unlock_irqrestore(&imp->imp_lock, flags);
 560
 561         req->rq_reqmsg->status = current->pid;
 562         CDEBUG(D_RPCTRACE, "Sending RPC pname:cluuid:pid:xid:ni:nid:opc"
 563                " %s:%s:%d:"LPU64":%s:"LPX64":%d\n", current->comm,
 564                imp->imp_obd->obd_uuid.uuid, req->rq_reqmsg->status,
 565                req->rq_xid,
 566                imp->imp_connection->c_peer.peer_ni->pni_name,
 567                imp->imp_connection->c_peer.peer_nid,
 568                req->rq_reqmsg->opc);
 569
 570         rc = ptl_send_rpc(req);
 571         if (rc) {
 572                 DEBUG_REQ(D_HA, req, "send failed (%d); expect timeout", rc);
 573                 req->rq_timeout = 1;
 574                 RETURN(rc);
 575         }
 576         RETURN(0);
 577 }
 578
 579 int ptlrpc_check_set(struct ptlrpc_request_set *set)
 580 {
 581         unsigned long flags;
 582         struct list_head *tmp;
 583         int force_timer_recalc = 0;
 584         ENTRY;
 585
 586         if (set->set_remaining == 0)
 587                 RETURN(1);
 588
 589         list_for_each(tmp, &set->set_requests) {
 590                 struct ptlrpc_request *req =
 591                         list_entry(tmp, struct ptlrpc_request, rq_set_chain);
 592                 struct obd_import *imp = req->rq_import;
 593                 int rc = 0;
 594
 595                 if (req->rq_phase == RQ_PHASE_NEW &&
 596                     ptlrpc_send_new_req(req)) {
 597                         force_timer_recalc = 1;
 598                 }
 599
 600                 if (!(req->rq_phase == RQ_PHASE_RPC ||
 601                       req->rq_phase == RQ_PHASE_BULK ||
 602                       req->rq_phase == RQ_PHASE_INTERPRET ||
 603                       req->rq_phase == RQ_PHASE_COMPLETE)) {
 604                         DEBUG_REQ(D_ERROR, req, "bad phase %x", req->rq_phase);
 605                         LBUG();
 606                 }
 607
 608                 if (req->rq_phase == RQ_PHASE_COMPLETE)
 609                         continue;
 610
 611                 if (req->rq_phase == RQ_PHASE_INTERPRET)
 612                         GOTO(interpret, req->rq_status);
 613
 614                 if (req->rq_err) {
 615                         ptlrpc_unregister_reply(req);
 616                         if (req->rq_status == 0)
 617                                 req->rq_status = -EIO;
 618                         req->rq_phase = RQ_PHASE_INTERPRET;
 619
 620                         spin_lock_irqsave(&imp->imp_lock, flags);
 621                         list_del_init(&req->rq_list);
 622                         spin_unlock_irqrestore(&imp->imp_lock, flags);
 623
 624                         GOTO(interpret, req->rq_status);
 625                 }
 626
 627                 /* ptlrpc_queue_wait->l_wait_event guarantees that rq_intr
 628                  * will only be set after rq_timedout, but the oig waiting
 629                  * path sets rq_intr irrespective of whether ptlrpcd has
 630                  * seen a timeout.  our policy is to only interpret
 631                  * interrupted rpcs after they have timed out */
 632                 if (req->rq_intr && (req->rq_timedout || req->rq_waiting)) {
 633                         /* NB could be on delayed list */
 634                         ptlrpc_unregister_reply(req);
 635                         req->rq_status = -EINTR;
 636                         req->rq_phase = RQ_PHASE_INTERPRET;
 637
 638                         spin_lock_irqsave(&imp->imp_lock, flags);
 639                         list_del_init(&req->rq_list);
 640                         spin_unlock_irqrestore(&imp->imp_lock, flags);
 641
 642                         GOTO(interpret, req->rq_status);
 643                 }
 644
 645                 if (req->rq_phase == RQ_PHASE_RPC) {
 646                         if (req->rq_waiting || req->rq_resend) {
 647                                 int status;
 648
 649                                 spin_lock_irqsave(&imp->imp_lock, flags);
 650
 651                                 if (ptlrpc_import_delay_req(imp, req, &status)){
 652                                         spin_unlock_irqrestore(&imp->imp_lock,
 653                                                                flags);
 654                                         continue;
 655                                 }
 656
 657                                 list_del_init(&req->rq_list);
 658                                 if (status != 0)  {
 659                                         req->rq_status = status;
 660                                         req->rq_phase = RQ_PHASE_INTERPRET;
 661                                         spin_unlock_irqrestore(&imp->imp_lock,
 662                                                                flags);
 663                                         GOTO(interpret, req->rq_status);
 664                                 }
 665                                 if (req->rq_no_resend) {
 666                                         req->rq_status = -ENOTCONN;
 667                                         req->rq_phase = RQ_PHASE_INTERPRET;
 668                                         spin_unlock_irqrestore(&imp->imp_lock,
 669                                                                flags);
 670                                         GOTO(interpret, req->rq_status);
 671                                 }
 672                                 list_add_tail(&req->rq_list,
 673                                               &imp->imp_sending_list);
 674
 675                                 spin_unlock_irqrestore(&imp->imp_lock, flags);
 676
 677                                 req->rq_waiting = 0;
 678                                 if (req->rq_resend) {
 679                                         lustre_msg_add_flags(req->rq_reqmsg,
 680                                                              MSG_RESENT);
 681                                         ptlrpc_unregister_reply(req);
 682                                         if (req->rq_bulk) {
 683                                                 __u64 old_xid = req->rq_xid;
 684
 685                                                 ptlrpc_unregister_bulk (req);
 686
 687                                                 /* ensure previous bulk fails */
 688                                                 req->rq_xid = ptlrpc_next_xid();
 689                                                 CDEBUG(D_HA, "resend bulk "
 690                                                        "old x"LPU64
 691                                                        " new x"LPU64"\n",
 692                                                        old_xid, req->rq_xid);
 693                                         }
 694                                 }
 695
 696                                 rc = ptl_send_rpc(req);
 697                                 if (rc) {
 698                                         DEBUG_REQ(D_HA, req, "send failed (%d)",
 699                                                   rc);
 700                                         force_timer_recalc = 1;
 701                                         req->rq_timeout = 0;
 702                                 }
 703                                 /* need to reset the timeout */
 704                                 force_timer_recalc = 1;
 705                         }
 706
 707                         /* Still waiting for a reply? */
 708                         if (ptlrpc_client_receiving_reply(req))
 709                                 continue;
 710
 711                         /* Did we actually receive a reply? */
 712                         if (!ptlrpc_client_replied(req))
 713                                 continue;
 714
 715                         spin_lock_irqsave(&imp->imp_lock, flags);
 716                         list_del_init(&req->rq_list);
 717                         spin_unlock_irqrestore(&imp->imp_lock, flags);
 718
 719                         req->rq_status = after_reply(req);
 720                         if (req->rq_resend) {
 721                                 /* Add this req to the delayed list so
 722                                    it can be errored if the import is
 723                                    evicted after recovery. */
 724                                 spin_lock_irqsave (&req->rq_lock, flags);
 725                                 list_add_tail(&req->rq_list,
 726                                               &imp->imp_delayed_list);
 727                                 spin_unlock_irqrestore(&req->rq_lock, flags);
 728                                 continue;
 729                         }
 730
 731                         /* If there is no bulk associated with this request,
 732                          * then we're done and should let the interpreter
 733                          * process the reply.  Similarly if the RPC returned
 734                          * an error, and therefore the bulk will never arrive.
 735                          */
 736                         if (req->rq_bulk == NULL || req->rq_status != 0) {
 737                                 req->rq_phase = RQ_PHASE_INTERPRET;
 738                                 GOTO(interpret, req->rq_status);
 739                         }
 740
 741                         req->rq_phase = RQ_PHASE_BULK;
 742                 }
 743
 744                 LASSERT(req->rq_phase == RQ_PHASE_BULK);
 745                 if (ptlrpc_bulk_active(req->rq_bulk))
 746                         continue;
 747
 748                 if (!req->rq_bulk->bd_success) {
 749                         /* The RPC reply arrived OK, but the bulk screwed
 750                          * up!  Dead wierd since the server told us the RPC
 751                          * was good after getting the REPLY for her GET or
 752                          * the ACK for her PUT. */
 753                         DEBUG_REQ(D_ERROR, req, "bulk transfer failed");
 754                         LBUG();
 755                 }
 756
 757                 req->rq_phase = RQ_PHASE_INTERPRET;
 758
 759         interpret:
 760                 LASSERT(req->rq_phase == RQ_PHASE_INTERPRET);
 761                 LASSERT(!req->rq_receiving_reply);
 762
 763                 ptlrpc_unregister_reply(req);
 764                 if (req->rq_bulk != NULL)
 765                         ptlrpc_unregister_bulk (req);
 766
 767                 req->rq_phase = RQ_PHASE_COMPLETE;
 768
 769                 if (req->rq_interpret_reply != NULL) {
 770                         int (*interpreter)(struct ptlrpc_request *,void *,int) =
 771                                 req->rq_interpret_reply;
 772                         req->rq_status = interpreter(req, &req->rq_async_args,
 773                                                      req->rq_status);
 774                 }
 775
 776                 CDEBUG(D_RPCTRACE, "Completed RPC pname:cluuid:pid:xid:ni:nid:"
 777                        "opc %s:%s:%d:"LPU64":%s:"LPX64":%d\n", current->comm,
 778                        imp->imp_obd->obd_uuid.uuid, req->rq_reqmsg->status,
 779                        req->rq_xid,
 780                        imp->imp_connection->c_peer.peer_ni->pni_name,
 781                        imp->imp_connection->c_peer.peer_nid,
 782                        req->rq_reqmsg->opc);
 783
 784                 set->set_remaining--;
 785
 786                 atomic_dec(&imp->imp_inflight);
 787                 wake_up(&imp->imp_recovery_waitq);
 788         }
 789
 790         /* If we hit an error, we want to recover promptly. */
 791         RETURN(set->set_remaining == 0 || force_timer_recalc);
 792 }
 793
 794 int ptlrpc_expire_one_request(struct ptlrpc_request *req)
 795 {
 796         unsigned long      flags;
 797         struct obd_import *imp = req->rq_import;
 798         ENTRY;
 799
 800         DEBUG_REQ(D_ERROR, req, "timeout");
 801
 802         spin_lock_irqsave (&req->rq_lock, flags);
 803         req->rq_timedout = 1;
 804         spin_unlock_irqrestore (&req->rq_lock, flags);
 805
 806         ptlrpc_unregister_reply (req);
 807
 808         if (req->rq_bulk != NULL)
 809                 ptlrpc_unregister_bulk (req);
 810
 811         if (imp == NULL) {
 812                 DEBUG_REQ(D_HA, req, "NULL import: already cleaned up?");
 813                 RETURN(1);
 814         }
 815
 816         /* The DLM server doesn't want recovery run on its imports. */
 817         if (imp->imp_dlm_fake)
 818                 RETURN(1);
 819
 820         /* If this request is for recovery or other primordial tasks,
 821          * then error it out here. */
 822         if (req->rq_send_state != LUSTRE_IMP_FULL ||
 823             imp->imp_obd->obd_no_recov) {
 824                 spin_lock_irqsave (&req->rq_lock, flags);
 825                 req->rq_status = -ETIMEDOUT;
 826                 req->rq_err = 1;
 827                 spin_unlock_irqrestore (&req->rq_lock, flags);
 828                 RETURN(1);
 829         }
 830
 831         ptlrpc_fail_import(imp, req->rq_import_generation);
 832
 833         RETURN(0);
 834 }
 835
 836 int ptlrpc_expired_set(void *data)
 837 {
 838         struct ptlrpc_request_set *set = data;
 839         struct list_head          *tmp;
 840         time_t                     now = LTIME_S (CURRENT_TIME);
 841         ENTRY;
 842
 843         LASSERT(set != NULL);
 844
 845         /* A timeout expired; see which reqs it applies to... */
 846         list_for_each (tmp, &set->set_requests) {
 847                 struct ptlrpc_request *req =
 848                         list_entry(tmp, struct ptlrpc_request, rq_set_chain);
 849
 850                 /* request in-flight? */
 851                 if (!((req->rq_phase == RQ_PHASE_RPC && !req->rq_waiting
 852                        && !req->rq_resend) ||
 853                       (req->rq_phase == RQ_PHASE_BULK)))
 854                         continue;
 855
 856                 if (req->rq_timedout ||           /* already dealt with */
 857                     req->rq_sent + req->rq_timeout > now) /* not expired */
 858                         continue;
 859
 860                 /* deal with this guy */
 861                 ptlrpc_expire_one_request (req);
 862         }
 863
 864         /* When waiting for a whole set, we always to break out of the
 865          * sleep so we can recalculate the timeout, or enable interrupts
 866          * iff everyone's timed out.
 867          */
 868         RETURN(1);
 869 }
 870
 871 void ptlrpc_mark_interrupted(struct ptlrpc_request *req)
 872 {
 873         unsigned long flags;
 874         spin_lock_irqsave(&req->rq_lock, flags);
 875         req->rq_intr = 1;
 876         spin_unlock_irqrestore(&req->rq_lock, flags);
 877 }
 878
 879 void ptlrpc_interrupted_set(void *data)
 880 {
 881         struct ptlrpc_request_set *set = data;
 882         struct list_head *tmp;
 883
 884         LASSERT(set != NULL);
 885         CERROR("INTERRUPTED SET %p\n", set);
 886
 887         list_for_each(tmp, &set->set_requests) {
 888                 struct ptlrpc_request *req =
 889                         list_entry(tmp, struct ptlrpc_request, rq_set_chain);
 890
 891                 if (req->rq_phase != RQ_PHASE_RPC)
 892                         continue;
 893
 894                 ptlrpc_mark_interrupted(req);
 895         }
 896 }
 897
 898 int ptlrpc_set_next_timeout(struct ptlrpc_request_set *set)
 899 {
 900         struct list_head      *tmp;
 901         time_t                 now = LTIME_S(CURRENT_TIME);
 902         time_t                 deadline;
 903         int                    timeout = 0;
 904         struct ptlrpc_request *req;
 905         ENTRY;
 906
 907         SIGNAL_MASK_ASSERT(); /* XXX BUG 1511 */
 908
 909         list_for_each(tmp, &set->set_requests) {
 910                 req = list_entry(tmp, struct ptlrpc_request, rq_set_chain);
 911
 912                 /* request in-flight? */
 913                 if (!((req->rq_phase == RQ_PHASE_RPC && !req->rq_waiting) ||
 914                       (req->rq_phase == RQ_PHASE_BULK)))
 915                         continue;
 916
 917                 if (req->rq_timedout)   /* already timed out */
 918                         continue;
 919
 920                 deadline = req->rq_sent + req->rq_timeout;
 921                 if (deadline <= now)    /* actually expired already */
 922                         timeout = 1;    /* ASAP */
 923                 else if (timeout == 0 || timeout > deadline - now)
 924                         timeout = deadline - now;
 925         }
 926         RETURN(timeout);
 927 }
 928
 929
 930 int ptlrpc_set_wait(struct ptlrpc_request_set *set)
 931 {
 932         struct list_head      *tmp;
 933         struct ptlrpc_request *req;
 934         struct l_wait_info     lwi;
 935         int                    rc, timeout;
 936         ENTRY;
 937
 938         LASSERT(!list_empty(&set->set_requests));
 939         list_for_each(tmp, &set->set_requests) {
 940                 req = list_entry(tmp, struct ptlrpc_request, rq_set_chain);
 941                 if (req->rq_phase == RQ_PHASE_NEW)
 942                         (void)ptlrpc_send_new_req(req);
 943         }
 944
 945         do {
 946                 timeout = ptlrpc_set_next_timeout(set);
 947
 948                 /* wait until all complete, interrupted, or an in-flight
 949                  * req times out */
 950                 CDEBUG(D_HA, "set %p going to sleep for %d seconds\n",
 951                        set, timeout);
 952                 lwi = LWI_TIMEOUT_INTR((timeout ? timeout : 1) * HZ,
 953                                        ptlrpc_expired_set,
 954                                        ptlrpc_interrupted_set, set);
 955                 rc = l_wait_event(set->set_waitq, ptlrpc_check_set(set), &lwi);
 956
 957                 LASSERT(rc == 0 || rc == -EINTR || rc == -ETIMEDOUT);
 958
 959                 /* -EINTR => all requests have been flagged rq_intr so next
 960                  * check completes.
 961                  * -ETIMEOUTD => someone timed out.  When all reqs have
 962                  * timed out, signals are enabled allowing completion with
 963                  * EINTR.
 964                  * I don't really care if we go once more round the loop in
 965                  * the error cases -eeb. */
 966         } while (rc != 0);
 967
 968         LASSERT(set->set_remaining == 0);
 969
 970         rc = 0;
 971         list_for_each(tmp, &set->set_requests) {
 972                 req = list_entry(tmp, struct ptlrpc_request, rq_set_chain);
 973
 974                 LASSERT(req->rq_phase == RQ_PHASE_COMPLETE);
 975                 if (req->rq_status != 0)
 976                         rc = req->rq_status;
 977         }
 978
 979         if (set->set_interpret != NULL) {
 980                 int (*interpreter)(struct ptlrpc_request_set *set,void *,int) =
 981                         set->set_interpret;
 982                 rc = interpreter (set, &set->set_args, rc);
 983         }
 984
 985         RETURN(rc);
 986 }
 987
 988 static void __ptlrpc_free_req(struct ptlrpc_request *request, int locked)
 989 {
 990         ENTRY;
 991         if (request == NULL) {
 992                 EXIT;
 993                 return;
 994         }
 995
 996         LASSERTF(!request->rq_receiving_reply, "req %p\n", request);
 997         LASSERTF(request->rq_rqbd == NULL, "req %p\n",request);/* client-side */
 998         LASSERTF(list_empty(&request->rq_list), "req %p\n", request);
 999         LASSERTF(list_empty(&request->rq_set_chain), "req %p\n", request);
1000
1001         /* We must take it off the imp_replay_list first.  Otherwise, we'll set
1002          * request->rq_reqmsg to NULL while osc_close is dereferencing it. */
1003         if (request->rq_import != NULL) {
1004                 unsigned long flags = 0;
1005                 if (!locked)
1006                         spin_lock_irqsave(&request->rq_import->imp_lock, flags);
1007                 list_del_init(&request->rq_replay_list);
1008                 if (!locked)
1009                         spin_unlock_irqrestore(&request->rq_import->imp_lock,
1010                                                flags);
1011         }
1012         LASSERTF(list_empty(&request->rq_replay_list), "req %p\n", request);
1013
1014         if (atomic_read(&request->rq_refcount) != 0) {
1015                 DEBUG_REQ(D_ERROR, request,
1016                           "freeing request with nonzero refcount");
1017                 LBUG();
1018         }
1019
1020         if (request->rq_repmsg != NULL) {
1021                 OBD_FREE(request->rq_repmsg, request->rq_replen);
1022                 request->rq_repmsg = NULL;
1023         }
1024         if (request->rq_reqmsg != NULL) {
1025                 OBD_FREE(request->rq_reqmsg, request->rq_reqlen);
1026                 request->rq_reqmsg = NULL;
1027         }
1028         if (request->rq_export != NULL) {
1029                 class_export_put(request->rq_export);
1030                 request->rq_export = NULL;
1031         }
1032         if (request->rq_import != NULL) {
1033                 class_import_put(request->rq_import);
1034                 request->rq_import = NULL;
1035         }
1036         if (request->rq_bulk != NULL)
1037                 ptlrpc_free_bulk(request->rq_bulk);
1038
1039         OBD_FREE(request, sizeof(*request));
1040         EXIT;
1041 }
1042
1043 void ptlrpc_free_req(struct ptlrpc_request *request)
1044 {
1045         __ptlrpc_free_req(request, 0);
1046 }
1047
1048 static int __ptlrpc_req_finished(struct ptlrpc_request *request, int locked);
1049 void ptlrpc_req_finished_with_imp_lock(struct ptlrpc_request *request)
1050 {
1051         LASSERT_SPIN_LOCKED(&request->rq_import->imp_lock);
1052         (void)__ptlrpc_req_finished(request, 1);
1053 }
1054
1055 static int __ptlrpc_req_finished(struct ptlrpc_request *request, int locked)
1056 {
1057         ENTRY;
1058         if (request == NULL)
1059                 RETURN(1);
1060
1061         if (request == (void *)(unsigned long)(0x5a5a5a5a5a5a5a5a) ||
1062             request->rq_reqmsg == (void *)(unsigned long)(0x5a5a5a5a5a5a5a5a)) {
1063                 CERROR("dereferencing freed request (bug 575)\n");
1064                 LBUG();
1065                 RETURN(1);
1066         }
1067
1068         DEBUG_REQ(D_INFO, request, "refcount now %u",
1069                   atomic_read(&request->rq_refcount) - 1);
1070
1071         if (atomic_dec_and_test(&request->rq_refcount)) {
1072                 __ptlrpc_free_req(request, locked);
1073                 RETURN(1);
1074         }
1075
1076         RETURN(0);
1077 }
1078
1079 void ptlrpc_req_finished(struct ptlrpc_request *request)
1080 {
1081         __ptlrpc_req_finished(request, 0);
1082 }
1083
1084 /* Disengage the client's reply buffer from the network
1085  * NB does _NOT_ unregister any client-side bulk.
1086  * IDEMPOTENT, but _not_ safe against concurrent callers.
1087  * The request owner (i.e. the thread doing the I/O) must call...
1088  */
1089 void ptlrpc_unregister_reply (struct ptlrpc_request *request)
1090 {
1091         int                rc;
1092         wait_queue_head_t *wq;
1093         struct l_wait_info lwi;
1094
1095         LASSERT(!in_interrupt ());             /* might sleep */
1096
1097         if (!ptlrpc_client_receiving_reply(request))
1098                 return;
1099
1100         rc = PtlMDUnlink (request->rq_reply_md_h);
1101         if (rc == PTL_MD_INVALID) {
1102                 LASSERT (!ptlrpc_client_receiving_reply(request));
1103                 return;
1104         }
1105
1106         LASSERT (rc == PTL_OK);
1107
1108         if (request->rq_set == NULL)
1109                 wq = &request->rq_set->set_waitq;
1110         else
1111                 wq = &request->rq_reply_waitq;
1112
1113         for (;;) {
1114                 /* Network access will complete in finite time but the HUGE
1115                  * timeout lets us CWARN for visibility of sluggish NALs */
1116                 lwi = LWI_TIMEOUT(300 * HZ, NULL, NULL);
1117                 rc = l_wait_event (*wq, !ptlrpc_client_receiving_reply(request), &lwi);
1118                 if (rc == 0)
1119                         return;
1120
1121                 LASSERT (rc == -ETIMEDOUT);
1122                 DEBUG_REQ(D_WARNING, request, "Unexpectedly long timeout");
1123         }
1124 }
1125
1126 /* caller must hold imp->imp_lock */
1127 void ptlrpc_free_committed(struct obd_import *imp)
1128 {
1129         struct list_head *tmp, *saved;
1130         struct ptlrpc_request *req;
1131         struct ptlrpc_request *last_req = NULL; /* temporary fire escape */
1132         ENTRY;
1133
1134         LASSERT(imp != NULL);
1135
1136         LASSERT_SPIN_LOCKED(&imp->imp_lock);
1137
1138         CDEBUG(D_HA, "%s: committing for last_committed "LPU64"\n",
1139                imp->imp_obd->obd_name, imp->imp_peer_committed_transno);
1140
1141         list_for_each_safe(tmp, saved, &imp->imp_replay_list) {
1142                 req = list_entry(tmp, struct ptlrpc_request, rq_replay_list);
1143
1144                 /* XXX ok to remove when 1357 resolved - rread 05/29/03  */
1145                 LASSERT(req != last_req);
1146                 last_req = req;
1147
1148                 if (req->rq_import_generation < imp->imp_generation) {
1149                         DEBUG_REQ(D_HA, req, "freeing request with old gen");
1150                         GOTO(free_req, 0);
1151                 }
1152
1153                 if (req->rq_replay) {
1154                         DEBUG_REQ(D_HA, req, "keeping (FL_REPLAY)");
1155                         continue;
1156                 }
1157
1158                 /* not yet committed */
1159                 if (req->rq_transno > imp->imp_peer_committed_transno) {
1160                         DEBUG_REQ(D_HA, req, "stopping search");
1161                         break;
1162                 }
1163
1164                 DEBUG_REQ(D_HA, req, "committing (last_committed "LPU64")",
1165                           imp->imp_peer_committed_transno);
1166 free_req:
1167                 if (req->rq_commit_cb != NULL)
1168                         req->rq_commit_cb(req);
1169                 list_del_init(&req->rq_replay_list);
1170                 __ptlrpc_req_finished(req, 1);
1171         }
1172
1173         EXIT;
1174         return;
1175 }
1176
1177 void ptlrpc_cleanup_client(struct obd_import *imp)
1178 {
1179         ENTRY;
1180         EXIT;
1181         return;
1182 }
1183
1184 void ptlrpc_resend_req(struct ptlrpc_request *req)
1185 {
1186         unsigned long flags;
1187
1188         DEBUG_REQ(D_HA, req, "going to resend");
1189         req->rq_reqmsg->handle.cookie = 0;
1190         req->rq_status = -EAGAIN;
1191
1192         spin_lock_irqsave (&req->rq_lock, flags);
1193         req->rq_resend = 1;
1194         req->rq_timedout = 0;
1195         if (req->rq_bulk) {
1196                 __u64 old_xid = req->rq_xid;
1197
1198                 /* ensure previous bulk fails */
1199                 req->rq_xid = ptlrpc_next_xid();
1200                 CDEBUG(D_HA, "resend bulk old x"LPU64" new x"LPU64"\n",
1201                        old_xid, req->rq_xid);
1202         }
1203         ptlrpc_wake_client_req(req);
1204         spin_unlock_irqrestore (&req->rq_lock, flags);
1205 }
1206
1207 /* XXX: this function and rq_status are currently unused */
1208 void ptlrpc_restart_req(struct ptlrpc_request *req)
1209 {
1210         unsigned long flags;
1211
1212         DEBUG_REQ(D_HA, req, "restarting (possibly-)completed request");
1213         req->rq_status = -ERESTARTSYS;
1214
1215         spin_lock_irqsave (&req->rq_lock, flags);
1216         req->rq_restart = 1;
1217         req->rq_timedout = 0;
1218         ptlrpc_wake_client_req(req);
1219         spin_unlock_irqrestore (&req->rq_lock, flags);
1220 }
1221
1222 static int expired_request(void *data)
1223 {
1224         struct ptlrpc_request *req = data;
1225         ENTRY;
1226
1227         RETURN(ptlrpc_expire_one_request(req));
1228 }
1229
1230 static void interrupted_request(void *data)
1231 {
1232         unsigned long flags;
1233
1234         struct ptlrpc_request *req = data;
1235         DEBUG_REQ(D_HA, req, "request interrupted");
1236         spin_lock_irqsave (&req->rq_lock, flags);
1237         req->rq_intr = 1;
1238         spin_unlock_irqrestore (&req->rq_lock, flags);
1239 }
1240
1241 struct ptlrpc_request *ptlrpc_request_addref(struct ptlrpc_request *req)
1242 {
1243         ENTRY;
1244         atomic_inc(&req->rq_refcount);
1245         RETURN(req);
1246 }
1247
1248 void ptlrpc_retain_replayable_request(struct ptlrpc_request *req,
1249                                       struct obd_import *imp)
1250 {
1251         struct list_head *tmp;
1252
1253         LASSERT_SPIN_LOCKED(&imp->imp_lock);
1254
1255         /* don't re-add requests that have been replayed */
1256         if (!list_empty(&req->rq_replay_list))
1257                 return;
1258
1259         lustre_msg_add_flags(req->rq_reqmsg,
1260                              MSG_REPLAY);
1261
1262         LASSERT(imp->imp_replayable);
1263         /* Balanced in ptlrpc_free_committed, usually. */
1264         ptlrpc_request_addref(req);
1265         list_for_each_prev(tmp, &imp->imp_replay_list) {
1266                 struct ptlrpc_request *iter =
1267                         list_entry(tmp, struct ptlrpc_request, rq_replay_list);
1268
1269                 /* We may have duplicate transnos if we create and then
1270                  * open a file, or for closes retained if to match creating
1271                  * opens, so use req->rq_xid as a secondary key.
1272                  * (See bugs 684, 685, and 428.)
1273                  * XXX no longer needed, but all opens need transnos!
1274                  */
1275                 if (iter->rq_transno > req->rq_transno)
1276                         continue;
1277
1278                 if (iter->rq_transno == req->rq_transno) {
1279                         LASSERT(iter->rq_xid != req->rq_xid);
1280                         if (iter->rq_xid > req->rq_xid)
1281                                 continue;
1282                 }
1283
1284                 list_add(&req->rq_replay_list, &iter->rq_replay_list);
1285                 return;
1286         }
1287
1288         list_add_tail(&req->rq_replay_list, &imp->imp_replay_list);
1289 }
1290
1291 int ptlrpc_queue_wait(struct ptlrpc_request *req)
1292 {
1293         int rc = 0;
1294         int brc;
1295         struct l_wait_info lwi;
1296         struct obd_import *imp = req->rq_import;
1297         unsigned long flags;
1298         int timeout = 0;
1299         ENTRY;
1300
1301         LASSERT(req->rq_set == NULL);
1302         LASSERT(!req->rq_receiving_reply);
1303         atomic_inc(&imp->imp_inflight);
1304
1305         /* for distributed debugging */
1306         req->rq_reqmsg->status = current->pid;
1307         LASSERT(imp->imp_obd != NULL);
1308         CDEBUG(D_RPCTRACE, "Sending RPC pname:cluuid:pid:xid:ni:nid:opc "
1309                "%s:%s:%d:"LPU64":%s:"LPX64":%d\n", current->comm,
1310                imp->imp_obd->obd_uuid.uuid,
1311                req->rq_reqmsg->status, req->rq_xid,
1312                imp->imp_connection->c_peer.peer_ni->pni_name,
1313                imp->imp_connection->c_peer.peer_nid,
1314                req->rq_reqmsg->opc);
1315
1316         /* Mark phase here for a little debug help */
1317         req->rq_phase = RQ_PHASE_RPC;
1318
1319         spin_lock_irqsave(&imp->imp_lock, flags);
1320         req->rq_import_generation = imp->imp_generation;
1321 restart:
1322         if (ptlrpc_import_delay_req(imp, req, &rc)) {
1323                 list_del(&req->rq_list);
1324
1325                 list_add_tail(&req->rq_list, &imp->imp_delayed_list);
1326                 spin_unlock_irqrestore(&imp->imp_lock, flags);
1327
1328                 DEBUG_REQ(D_HA, req, "\"%s\" waiting for recovery: (%s != %s)",
1329                           current->comm,
1330                           ptlrpc_import_state_name(req->rq_send_state),
1331                           ptlrpc_import_state_name(imp->imp_state));
1332                 lwi = LWI_INTR(interrupted_request, req);
1333                 rc = l_wait_event(req->rq_reply_waitq,
1334                                   (req->rq_send_state == imp->imp_state ||
1335                                    req->rq_err),
1336                                   &lwi);
1337                 DEBUG_REQ(D_HA, req, "\"%s\" awake: (%s == %s or %d == 1)",
1338                           current->comm,
1339                           ptlrpc_import_state_name(imp->imp_state),
1340                           ptlrpc_import_state_name(req->rq_send_state),
1341                           req->rq_err);
1342
1343                 spin_lock_irqsave(&imp->imp_lock, flags);
1344                 list_del_init(&req->rq_list);
1345
1346                 if (req->rq_err) {
1347                         rc = -EIO;
1348                 }
1349                 else if (req->rq_intr) {
1350                         rc = -EINTR;
1351                 }
1352                 else if (req->rq_no_resend) {
1353                         spin_unlock_irqrestore(&imp->imp_lock, flags);
1354                         GOTO(out, rc = -ETIMEDOUT);
1355                 }
1356                 else {
1357                         GOTO(restart, rc);
1358                 }
1359         }
1360
1361         if (rc != 0) {
1362                 list_del_init(&req->rq_list);
1363                 spin_unlock_irqrestore(&imp->imp_lock, flags);
1364                 req->rq_status = rc; // XXX this ok?
1365                 GOTO(out, rc);
1366         }
1367
1368         if (req->rq_resend) {
1369                 lustre_msg_add_flags(req->rq_reqmsg, MSG_RESENT);
1370
1371                 if (req->rq_bulk != NULL)
1372                         ptlrpc_unregister_bulk (req);
1373
1374                 DEBUG_REQ(D_HA, req, "resending: ");
1375         }
1376
1377         /* XXX this is the same as ptlrpc_set_wait */
1378         LASSERT(list_empty(&req->rq_list));
1379         list_add_tail(&req->rq_list, &imp->imp_sending_list);
1380         spin_unlock_irqrestore(&imp->imp_lock, flags);
1381
1382         rc = ptl_send_rpc(req);
1383         if (rc) {
1384                 DEBUG_REQ(D_HA, req, "send failed (%d); recovering", rc);
1385                 timeout = 1;
1386         } else {
1387                 timeout = MAX(req->rq_timeout * HZ, 1);
1388                 DEBUG_REQ(D_NET, req, "-- sleeping");
1389         }
1390         lwi = LWI_TIMEOUT_INTR(timeout, expired_request, interrupted_request,
1391                                req);
1392         l_wait_event(req->rq_reply_waitq, ptlrpc_check_reply(req), &lwi);
1393         DEBUG_REQ(D_NET, req, "-- done sleeping");
1394
1395         CDEBUG(D_RPCTRACE, "Completed RPC pname:cluuid:pid:xid:ni:nid:opc "
1396                "%s:%s:%d:"LPU64":%s:"LPX64":%d\n", current->comm,
1397                imp->imp_obd->obd_uuid.uuid,
1398                req->rq_reqmsg->status, req->rq_xid,
1399                imp->imp_connection->c_peer.peer_ni->pni_name,
1400                imp->imp_connection->c_peer.peer_nid,
1401                req->rq_reqmsg->opc);
1402
1403         spin_lock_irqsave(&imp->imp_lock, flags);
1404         list_del_init(&req->rq_list);
1405         spin_unlock_irqrestore(&imp->imp_lock, flags);
1406
1407         /* If the reply was received normally, this just grabs the spinlock
1408          * (ensuring the reply callback has returned), sees that
1409          * req->rq_receiving_reply is clear and returns. */
1410         ptlrpc_unregister_reply (req);
1411
1412         if (req->rq_err)
1413                 GOTO(out, rc = -EIO);
1414
1415         /* Resend if we need to, unless we were interrupted. */
1416         if (req->rq_resend && !req->rq_intr) {
1417                 /* ...unless we were specifically told otherwise. */
1418                 if (req->rq_no_resend)
1419                         GOTO(out, rc = -ETIMEDOUT);
1420                 spin_lock_irqsave(&imp->imp_lock, flags);
1421                 goto restart;
1422         }
1423
1424         if (req->rq_intr) {
1425                 /* Should only be interrupted if we timed out. */
1426                 if (!req->rq_timedout)
1427                         DEBUG_REQ(D_ERROR, req,
1428                                   "rq_intr set but rq_timedout not");
1429                 GOTO(out, rc = -EINTR);
1430         }
1431
1432         if (req->rq_timedout) {                 /* non-recoverable timeout */
1433                 GOTO(out, rc = -ETIMEDOUT);
1434         }
1435
1436         if (!req->rq_replied) {
1437                 /* How can this be? -eeb */
1438                 DEBUG_REQ(D_ERROR, req, "!rq_replied: ");
1439                 LBUG();
1440                 GOTO(out, rc = req->rq_status);
1441         }
1442
1443         rc = after_reply (req);
1444         /* NB may return +ve success rc */
1445         if (req->rq_resend) {
1446                 spin_lock_irqsave(&imp->imp_lock, flags);
1447                 goto restart;
1448         }
1449
1450  out:
1451         if (req->rq_bulk != NULL) {
1452                 if (rc >= 0) {
1453                         /* success so far.  Note that anything going wrong
1454                          * with bulk now, is EXTREMELY strange, since the
1455                          * server must have believed that the bulk
1456                          * tranferred OK before she replied with success to
1457                          * me. */
1458                         lwi = LWI_TIMEOUT(timeout, NULL, NULL);
1459                         brc = l_wait_event(req->rq_reply_waitq,
1460                                            !ptlrpc_bulk_active(req->rq_bulk),
1461                                            &lwi);
1462                         LASSERT(brc == 0 || brc == -ETIMEDOUT);
1463                         if (brc != 0) {
1464                                 LASSERT(brc == -ETIMEDOUT);
1465                                 DEBUG_REQ(D_ERROR, req, "bulk timed out");
1466                                 rc = brc;
1467                         } else if (!req->rq_bulk->bd_success) {
1468                                 DEBUG_REQ(D_ERROR, req, "bulk transfer failed");
1469                                 rc = -EIO;
1470                         }
1471                 }
1472                 if (rc < 0)
1473                         ptlrpc_unregister_bulk (req);
1474         }
1475
1476         LASSERT(!req->rq_receiving_reply);
1477         req->rq_phase = RQ_PHASE_INTERPRET;
1478
1479         atomic_dec(&imp->imp_inflight);
1480         wake_up(&imp->imp_recovery_waitq);
1481         RETURN(rc);
1482 }
1483
1484 struct ptlrpc_replay_async_args {
1485         int praa_old_state;
1486         int praa_old_status;
1487 };
1488
1489 static int ptlrpc_replay_interpret(struct ptlrpc_request *req,
1490                                     void * data, int rc)
1491 {
1492         struct ptlrpc_replay_async_args *aa = data;
1493         struct obd_import *imp = req->rq_import;
1494         unsigned long flags;
1495
1496         atomic_dec(&imp->imp_replay_inflight);
1497
1498         if (!req->rq_replied) {
1499                 CERROR("request replay timed out, restarting recovery\n");
1500                 GOTO(out, rc = -ETIMEDOUT);
1501         }
1502
1503 #if SWAB_PARANOIA
1504         /* Clear reply swab mask; this is a new reply in sender's byte order */
1505         req->rq_rep_swab_mask = 0;
1506 #endif
1507         LASSERT (req->rq_nob_received <= req->rq_replen);
1508         rc = lustre_unpack_msg(req->rq_repmsg, req->rq_nob_received);
1509         if (rc) {
1510                 CERROR("unpack_rep failed: %d\n", rc);
1511                 GOTO(out, rc = -EPROTO);
1512         }
1513
1514         if (req->rq_repmsg->type == PTL_RPC_MSG_ERR &&
1515             req->rq_repmsg->status == -ENOTCONN)
1516                 GOTO(out, rc = req->rq_repmsg->status);
1517
1518         /* The transno had better not change over replay. */
1519         LASSERT(req->rq_reqmsg->transno == req->rq_repmsg->transno);
1520
1521         DEBUG_REQ(D_HA, req, "got rep");
1522
1523         /* let the callback do fixups, possibly including in the request */
1524         if (req->rq_replay_cb)
1525                 req->rq_replay_cb(req);
1526
1527         if (req->rq_replied && req->rq_repmsg->status != aa->praa_old_status) {
1528                 DEBUG_REQ(D_ERROR, req, "status %d, old was %d",
1529                           req->rq_repmsg->status, aa->praa_old_status);
1530         } else {
1531                 /* Put it back for re-replay. */
1532                 req->rq_repmsg->status = aa->praa_old_status;
1533         }
1534
1535         spin_lock_irqsave(&imp->imp_lock, flags);
1536         imp->imp_last_replay_transno = req->rq_transno;
1537         spin_unlock_irqrestore(&imp->imp_lock, flags);
1538
1539         /* continue with recovery */
1540         rc = ptlrpc_import_recovery_state_machine(imp);
1541  out:
1542         req->rq_send_state = aa->praa_old_state;
1543
1544         if (rc != 0)
1545                 /* this replay failed, so restart recovery */
1546                 ptlrpc_connect_import(imp, NULL);
1547
1548         RETURN(rc);
1549 }
1550
1551
1552 int ptlrpc_replay_req(struct ptlrpc_request *req)
1553 {
1554         struct ptlrpc_replay_async_args *aa;
1555         ENTRY;
1556
1557         LASSERT(req->rq_import->imp_state == LUSTRE_IMP_REPLAY);
1558
1559         /* Not handling automatic bulk replay yet (or ever?) */
1560         LASSERT(req->rq_bulk == NULL);
1561
1562         DEBUG_REQ(D_HA, req, "REPLAY");
1563
1564         LASSERT (sizeof (*aa) <= sizeof (req->rq_async_args));
1565         aa = (struct ptlrpc_replay_async_args *)&req->rq_async_args;
1566         memset(aa, 0, sizeof *aa);
1567
1568         /* Prepare request to be resent with ptlrpcd */
1569         aa->praa_old_state = req->rq_send_state;
1570         req->rq_send_state = LUSTRE_IMP_REPLAY;
1571         req->rq_phase = RQ_PHASE_NEW;
1572         /*
1573          * Q: "How can a req get on the replay list if it wasn't replied?"
1574          * A: "If we failed during the replay of this request, it will still
1575          *     be on the list, but rq_replied will have been reset to 0."
1576          */
1577         if (req->rq_replied) {
1578                 aa->praa_old_status = req->rq_repmsg->status;
1579                 req->rq_status = 0;
1580                 req->rq_replied = 0;
1581         }
1582
1583         req->rq_interpret_reply = ptlrpc_replay_interpret;
1584         atomic_inc(&req->rq_import->imp_replay_inflight);
1585         ptlrpc_request_addref(req); /* ptlrpcd needs a ref */
1586
1587         ptlrpcd_add_req(req);
1588         RETURN(0);
1589 }
1590
1591 void ptlrpc_abort_inflight(struct obd_import *imp)
1592 {
1593         unsigned long flags;
1594         struct list_head *tmp, *n;
1595         ENTRY;
1596
1597         /* Make sure that no new requests get processed for this import.
1598          * ptlrpc_{queue,set}_wait must (and does) hold imp_lock while testing
1599          * this flag and then putting requests on sending_list or delayed_list.
1600          */
1601         spin_lock_irqsave(&imp->imp_lock, flags);
1602
1603         /* XXX locking?  Maybe we should remove each request with the list
1604          * locked?  Also, how do we know if the requests on the list are
1605          * being freed at this time?
1606          */
1607         list_for_each_safe(tmp, n, &imp->imp_sending_list) {
1608                 struct ptlrpc_request *req =
1609                         list_entry(tmp, struct ptlrpc_request, rq_list);
1610
1611                 DEBUG_REQ(D_HA, req, "inflight");
1612
1613                 spin_lock (&req->rq_lock);
1614                 if (req->rq_import_generation < imp->imp_generation) {
1615                         req->rq_err = 1;
1616                         ptlrpc_wake_client_req(req);
1617                 }
1618                 spin_unlock (&req->rq_lock);
1619         }
1620
1621         list_for_each_safe(tmp, n, &imp->imp_delayed_list) {
1622                 struct ptlrpc_request *req =
1623                         list_entry(tmp, struct ptlrpc_request, rq_list);
1624
1625                 DEBUG_REQ(D_HA, req, "aborting waiting req");
1626
1627                 spin_lock (&req->rq_lock);
1628                 if (req->rq_import_generation < imp->imp_generation) {
1629                         req->rq_err = 1;
1630                         ptlrpc_wake_client_req(req);
1631                 }
1632                 spin_unlock (&req->rq_lock);
1633         }
1634
1635         /* Last chance to free reqs left on the replay list, but we
1636          * will still leak reqs that haven't comitted.  */
1637         if (imp->imp_replayable)
1638                 ptlrpc_free_committed(imp);
1639
1640         spin_unlock_irqrestore(&imp->imp_lock, flags);
1641
1642         EXIT;
1643 }
1644
1645 static __u64 ptlrpc_last_xid = 0;
1646 static spinlock_t ptlrpc_last_xid_lock = SPIN_LOCK_UNLOCKED;
1647
1648 __u64 ptlrpc_next_xid(void)
1649 {
1650         __u64 tmp;
1651         spin_lock(&ptlrpc_last_xid_lock);
1652         tmp = ++ptlrpc_last_xid;
1653         spin_unlock(&ptlrpc_last_xid_lock);
1654         return tmp;
1655 }
1656
1657