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                                 LASSERT (!ptlrpc_client_receiving_reply(req));
 650                                 LASSERT (req->rq_bulk == NULL ||
 651                                          !ptlrpc_bulk_active(req->rq_bulk));
 652
 653                                 spin_lock_irqsave(&imp->imp_lock, flags);
 654
 655                                 if (ptlrpc_import_delay_req(imp, req, &status)) {
 656                                         spin_unlock_irqrestore(&imp->imp_lock,
 657                                                                flags);
 658                                         continue;
 659                                 }
 660
 661                                 list_del_init(&req->rq_list);
 662                                 if (status != 0)  {
 663                                         req->rq_status = status;
 664                                         req->rq_phase = RQ_PHASE_INTERPRET;
 665                                         spin_unlock_irqrestore(&imp->imp_lock,
 666                                                                flags);
 667                                         GOTO(interpret, req->rq_status);
 668                                 }
 669                                 if (req->rq_no_resend) {
 670                                         req->rq_status = -ENOTCONN;
 671                                         req->rq_phase = RQ_PHASE_INTERPRET;
 672                                         spin_unlock_irqrestore(&imp->imp_lock,
 673                                                                flags);
 674                                         GOTO(interpret, req->rq_status);
 675                                 }
 676                                 list_add_tail(&req->rq_list,
 677                                               &imp->imp_sending_list);
 678
 679                                 spin_unlock_irqrestore(&imp->imp_lock, flags);
 680
 681                                 req->rq_waiting = 0;
 682                                 if (req->rq_resend) {
 683                                         lustre_msg_add_flags(req->rq_reqmsg,
 684                                                              MSG_RESENT);
 685                                         ptlrpc_unregister_reply(req);
 686                                         if (req->rq_bulk) {
 687                                                 __u64 old_xid = req->rq_xid;
 688
 689                                                 /* ensure previous bulk fails */
 690                                                 req->rq_xid = ptlrpc_next_xid();
 691                                                 CDEBUG(D_HA, "resend bulk "
 692                                                        "old x"LPU64
 693                                                        " new x"LPU64"\n",
 694                                                        old_xid, req->rq_xid);
 695                                         }
 696                                 }
 697
 698                                 rc = ptl_send_rpc(req);
 699                                 if (rc) {
 700                                         DEBUG_REQ(D_HA, req, "send failed (%d)",
 701                                                   rc);
 702                                         force_timer_recalc = 1;
 703                                         req->rq_timeout = 0;
 704                                 }
 705                                 /* need to reset the timeout */
 706                                 force_timer_recalc = 1;
 707                         }
 708
 709                         /* Still waiting for a reply? */
 710                         if (ptlrpc_client_receiving_reply(req))
 711                                 continue;
 712
 713                         /* Did we actually receive a reply? */
 714                         if (!ptlrpc_client_replied(req))
 715                                 continue;
 716
 717                         spin_lock_irqsave(&imp->imp_lock, flags);
 718                         list_del_init(&req->rq_list);
 719                         spin_unlock_irqrestore(&imp->imp_lock, flags);
 720
 721                         req->rq_status = after_reply(req);
 722                         if (req->rq_resend) {
 723                                 /* Add this req to the delayed list so
 724                                    it can be errored if the import is
 725                                    evicted after recovery. */
 726                                 spin_lock_irqsave (&req->rq_lock, flags);
 727                                 list_add_tail(&req->rq_list,
 728                                               &imp->imp_delayed_list);
 729                                 spin_unlock_irqrestore(&req->rq_lock, flags);
 730                                 continue;
 731                         }
 732
 733                         /* If there is no bulk associated with this request,
 734                          * then we're done and should let the interpreter
 735                          * process the reply.  Similarly if the RPC returned
 736                          * an error, and therefore the bulk will never arrive.
 737                          */
 738                         if (req->rq_bulk == NULL || req->rq_status != 0) {
 739                                 req->rq_phase = RQ_PHASE_INTERPRET;
 740                                 GOTO(interpret, req->rq_status);
 741                         }
 742
 743                         req->rq_phase = RQ_PHASE_BULK;
 744                 }
 745
 746                 LASSERT(req->rq_phase == RQ_PHASE_BULK);
 747                 if (ptlrpc_bulk_active(req->rq_bulk))
 748                         continue;
 749
 750                 if (!req->rq_bulk->bd_success) {
 751                         /* The RPC reply arrived OK, but the bulk screwed
 752                          * up!  Dead wierd since the server told us the RPC
 753                          * was good after getting the REPLY for her GET or
 754                          * the ACK for her PUT. */
 755                         DEBUG_REQ(D_ERROR, req, "bulk transfer failed");
 756                         LBUG();
 757                 }
 758
 759                 req->rq_phase = RQ_PHASE_INTERPRET;
 760
 761         interpret:
 762                 LASSERT(req->rq_phase == RQ_PHASE_INTERPRET);
 763                 LASSERT(!req->rq_receiving_reply);
 764
 765                 ptlrpc_unregister_reply(req);
 766                 if (req->rq_bulk != NULL)
 767                         ptlrpc_unregister_bulk (req);
 768
 769                 req->rq_phase = RQ_PHASE_COMPLETE;
 770
 771                 if (req->rq_interpret_reply != NULL) {
 772                         int (*interpreter)(struct ptlrpc_request *,void *,int) =
 773                                 req->rq_interpret_reply;
 774                         req->rq_status = interpreter(req, &req->rq_async_args,
 775                                                      req->rq_status);
 776                 }
 777
 778                 CDEBUG(D_RPCTRACE, "Completed RPC pname:cluuid:pid:xid:ni:nid:"
 779                        "opc %s:%s:%d:"LPU64":%s:"LPX64":%d\n", current->comm,
 780                        imp->imp_obd->obd_uuid.uuid, req->rq_reqmsg->status,
 781                        req->rq_xid,
 782                        imp->imp_connection->c_peer.peer_ni->pni_name,
 783                        imp->imp_connection->c_peer.peer_nid,
 784                        req->rq_reqmsg->opc);
 785
 786                 set->set_remaining--;
 787
 788                 atomic_dec(&imp->imp_inflight);
 789                 wake_up(&imp->imp_recovery_waitq);
 790         }
 791
 792         /* If we hit an error, we want to recover promptly. */
 793         RETURN(set->set_remaining == 0 || force_timer_recalc);
 794 }
 795
 796 int ptlrpc_expire_one_request(struct ptlrpc_request *req)
 797 {
 798         unsigned long      flags;
 799         struct obd_import *imp = req->rq_import;
 800         ENTRY;
 801
 802         DEBUG_REQ(D_ERROR, req, "timeout");
 803
 804         spin_lock_irqsave (&req->rq_lock, flags);
 805         req->rq_timedout = 1;
 806         spin_unlock_irqrestore (&req->rq_lock, flags);
 807
 808         ptlrpc_unregister_reply (req);
 809
 810         if (req->rq_bulk != NULL)
 811                 ptlrpc_unregister_bulk (req);
 812
 813         if (imp == NULL) {
 814                 DEBUG_REQ(D_HA, req, "NULL import: already cleaned up?");
 815                 RETURN(1);
 816         }
 817
 818         /* The DLM server doesn't want recovery run on its imports. */
 819         if (imp->imp_dlm_fake)
 820                 RETURN(1);
 821
 822         /* If this request is for recovery or other primordial tasks,
 823          * then error it out here. */
 824         if (req->rq_send_state != LUSTRE_IMP_FULL ||
 825             imp->imp_obd->obd_no_recov) {
 826                 spin_lock_irqsave (&req->rq_lock, flags);
 827                 req->rq_status = -ETIMEDOUT;
 828                 req->rq_err = 1;
 829                 spin_unlock_irqrestore (&req->rq_lock, flags);
 830                 RETURN(1);
 831         }
 832
 833         ptlrpc_fail_import(imp, req->rq_import_generation);
 834
 835         RETURN(0);
 836 }
 837
 838 int ptlrpc_expired_set(void *data)
 839 {
 840         struct ptlrpc_request_set *set = data;
 841         struct list_head          *tmp;
 842         time_t                     now = LTIME_S (CURRENT_TIME);
 843         ENTRY;
 844
 845         LASSERT(set != NULL);
 846
 847         /* A timeout expired; see which reqs it applies to... */
 848         list_for_each (tmp, &set->set_requests) {
 849                 struct ptlrpc_request *req =
 850                         list_entry(tmp, struct ptlrpc_request, rq_set_chain);
 851
 852                 /* request in-flight? */
 853                 if (!((req->rq_phase == RQ_PHASE_RPC && !req->rq_waiting
 854                        && !req->rq_resend) ||
 855                       (req->rq_phase == RQ_PHASE_BULK)))
 856                         continue;
 857
 858                 if (req->rq_timedout ||           /* already dealt with */
 859                     req->rq_sent + req->rq_timeout > now) /* not expired */
 860                         continue;
 861
 862                 /* deal with this guy */
 863                 ptlrpc_expire_one_request (req);
 864         }
 865
 866         /* When waiting for a whole set, we always to break out of the
 867          * sleep so we can recalculate the timeout, or enable interrupts
 868          * iff everyone's timed out.
 869          */
 870         RETURN(1);
 871 }
 872
 873 void ptlrpc_mark_interrupted(struct ptlrpc_request *req)
 874 {
 875         unsigned long flags;
 876         spin_lock_irqsave(&req->rq_lock, flags);
 877         req->rq_intr = 1;
 878         spin_unlock_irqrestore(&req->rq_lock, flags);
 879 }
 880
 881 void ptlrpc_interrupted_set(void *data)
 882 {
 883         struct ptlrpc_request_set *set = data;
 884         struct list_head *tmp;
 885
 886         LASSERT(set != NULL);
 887         CERROR("INTERRUPTED SET %p\n", set);
 888
 889         list_for_each(tmp, &set->set_requests) {
 890                 struct ptlrpc_request *req =
 891                         list_entry(tmp, struct ptlrpc_request, rq_set_chain);
 892
 893                 if (req->rq_phase != RQ_PHASE_RPC)
 894                         continue;
 895
 896                 ptlrpc_mark_interrupted(req);
 897         }
 898 }
 899
 900 int ptlrpc_set_next_timeout(struct ptlrpc_request_set *set)
 901 {
 902         struct list_head      *tmp;
 903         time_t                 now = LTIME_S(CURRENT_TIME);
 904         time_t                 deadline;
 905         int                    timeout = 0;
 906         struct ptlrpc_request *req;
 907         ENTRY;
 908
 909         SIGNAL_MASK_ASSERT(); /* XXX BUG 1511 */
 910
 911         list_for_each(tmp, &set->set_requests) {
 912                 req = list_entry(tmp, struct ptlrpc_request, rq_set_chain);
 913
 914                 /* request in-flight? */
 915                 if (!((req->rq_phase == RQ_PHASE_RPC && !req->rq_waiting) ||
 916                       (req->rq_phase == RQ_PHASE_BULK)))
 917                         continue;
 918
 919                 if (req->rq_timedout)   /* already timed out */
 920                         continue;
 921
 922                 deadline = req->rq_sent + req->rq_timeout;
 923                 if (deadline <= now)    /* actually expired already */
 924                         timeout = 1;    /* ASAP */
 925                 else if (timeout == 0 || timeout > deadline - now)
 926                         timeout = deadline - now;
 927         }
 928         RETURN(timeout);
 929 }
 930
 931
 932 int ptlrpc_set_wait(struct ptlrpc_request_set *set)
 933 {
 934         struct list_head      *tmp;
 935         struct ptlrpc_request *req;
 936         struct l_wait_info     lwi;
 937         int                    rc, timeout;
 938         ENTRY;
 939
 940         LASSERT(!list_empty(&set->set_requests));
 941         list_for_each(tmp, &set->set_requests) {
 942                 req = list_entry(tmp, struct ptlrpc_request, rq_set_chain);
 943                 if (req->rq_phase == RQ_PHASE_NEW)
 944                         (void)ptlrpc_send_new_req(req);
 945         }
 946
 947         do {
 948                 timeout = ptlrpc_set_next_timeout(set);
 949
 950                 /* wait until all complete, interrupted, or an in-flight
 951                  * req times out */
 952                 CDEBUG(D_HA, "set %p going to sleep for %d seconds\n",
 953                        set, timeout);
 954                 lwi = LWI_TIMEOUT_INTR((timeout ? timeout : 1) * HZ,
 955                                        ptlrpc_expired_set,
 956                                        ptlrpc_interrupted_set, set);
 957                 rc = l_wait_event(set->set_waitq, ptlrpc_check_set(set), &lwi);
 958
 959                 LASSERT(rc == 0 || rc == -EINTR || rc == -ETIMEDOUT);
 960
 961                 /* -EINTR => all requests have been flagged rq_intr so next
 962                  * check completes.
 963                  * -ETIMEOUTD => someone timed out.  When all reqs have
 964                  * timed out, signals are enabled allowing completion with
 965                  * EINTR.
 966                  * I don't really care if we go once more round the loop in
 967                  * the error cases -eeb. */
 968         } while (rc != 0);
 969
 970         LASSERT(set->set_remaining == 0);
 971
 972         rc = 0;
 973         list_for_each(tmp, &set->set_requests) {
 974                 req = list_entry(tmp, struct ptlrpc_request, rq_set_chain);
 975
 976                 LASSERT(req->rq_phase == RQ_PHASE_COMPLETE);
 977                 if (req->rq_status != 0)
 978                         rc = req->rq_status;
 979         }
 980
 981         if (set->set_interpret != NULL) {
 982                 int (*interpreter)(struct ptlrpc_request_set *set,void *,int) =
 983                         set->set_interpret;
 984                 rc = interpreter (set, &set->set_args, rc);
 985         }
 986
 987         RETURN(rc);
 988 }
 989
 990 static void __ptlrpc_free_req(struct ptlrpc_request *request, int locked)
 991 {
 992         ENTRY;
 993         if (request == NULL) {
 994                 EXIT;
 995                 return;
 996         }
 997
 998         LASSERTF(!request->rq_receiving_reply, "req %p\n", request);
 999         LASSERTF(request->rq_rqbd == NULL, "req %p\n",request);/* client-side */
1000         LASSERTF(list_empty(&request->rq_list), "req %p\n", request);
1001         LASSERTF(list_empty(&request->rq_set_chain), "req %p\n", request);
1002
1003         /* We must take it off the imp_replay_list first.  Otherwise, we'll set
1004          * request->rq_reqmsg to NULL while osc_close is dereferencing it. */
1005         if (request->rq_import != NULL) {
1006                 unsigned long flags = 0;
1007                 if (!locked)
1008                         spin_lock_irqsave(&request->rq_import->imp_lock, flags);
1009                 list_del_init(&request->rq_replay_list);
1010                 if (!locked)
1011                         spin_unlock_irqrestore(&request->rq_import->imp_lock,
1012                                                flags);
1013         }
1014         LASSERTF(list_empty(&request->rq_replay_list), "req %p\n", request);
1015
1016         if (atomic_read(&request->rq_refcount) != 0) {
1017                 DEBUG_REQ(D_ERROR, request,
1018                           "freeing request with nonzero refcount");
1019                 LBUG();
1020         }
1021
1022         if (request->rq_repmsg != NULL) {
1023                 OBD_FREE(request->rq_repmsg, request->rq_replen);
1024                 request->rq_repmsg = NULL;
1025         }
1026         if (request->rq_reqmsg != NULL) {
1027                 OBD_FREE(request->rq_reqmsg, request->rq_reqlen);
1028                 request->rq_reqmsg = NULL;
1029         }
1030         if (request->rq_export != NULL) {
1031                 class_export_put(request->rq_export);
1032                 request->rq_export = NULL;
1033         }
1034         if (request->rq_import != NULL) {
1035                 class_import_put(request->rq_import);
1036                 request->rq_import = NULL;
1037         }
1038         if (request->rq_bulk != NULL)
1039                 ptlrpc_free_bulk(request->rq_bulk);
1040
1041         OBD_FREE(request, sizeof(*request));
1042         EXIT;
1043 }
1044
1045 void ptlrpc_free_req(struct ptlrpc_request *request)
1046 {
1047         __ptlrpc_free_req(request, 0);
1048 }
1049
1050 static int __ptlrpc_req_finished(struct ptlrpc_request *request, int locked);
1051 void ptlrpc_req_finished_with_imp_lock(struct ptlrpc_request *request)
1052 {
1053         LASSERT_SPIN_LOCKED(&request->rq_import->imp_lock);
1054         (void)__ptlrpc_req_finished(request, 1);
1055 }
1056
1057 static int __ptlrpc_req_finished(struct ptlrpc_request *request, int locked)
1058 {
1059         ENTRY;
1060         if (request == NULL)
1061                 RETURN(1);
1062
1063         if (request == (void *)(unsigned long)(0x5a5a5a5a5a5a5a5a) ||
1064             request->rq_reqmsg == (void *)(unsigned long)(0x5a5a5a5a5a5a5a5a)) {
1065                 CERROR("dereferencing freed request (bug 575)\n");
1066                 LBUG();
1067                 RETURN(1);
1068         }
1069
1070         DEBUG_REQ(D_INFO, request, "refcount now %u",
1071                   atomic_read(&request->rq_refcount) - 1);
1072
1073         if (atomic_dec_and_test(&request->rq_refcount)) {
1074                 __ptlrpc_free_req(request, locked);
1075                 RETURN(1);
1076         }
1077
1078         RETURN(0);
1079 }
1080
1081 void ptlrpc_req_finished(struct ptlrpc_request *request)
1082 {
1083         __ptlrpc_req_finished(request, 0);
1084 }
1085
1086 /* Disengage the client's reply buffer from the network
1087  * NB does _NOT_ unregister any client-side bulk.
1088  * IDEMPOTENT, but _not_ safe against concurrent callers.
1089  * The request owner (i.e. the thread doing the I/O) must call...
1090  */
1091 void ptlrpc_unregister_reply (struct ptlrpc_request *request)
1092 {
1093         int                rc;
1094         wait_queue_head_t *wq;
1095         struct l_wait_info lwi;
1096
1097         LASSERT(!in_interrupt ());             /* might sleep */
1098
1099         if (!ptlrpc_client_receiving_reply(request))
1100                 return;
1101
1102         rc = PtlMDUnlink (request->rq_reply_md_h);
1103         if (rc == PTL_MD_INVALID) {
1104                 LASSERT (!ptlrpc_client_receiving_reply(request));
1105                 return;
1106         }
1107
1108         LASSERT (rc == PTL_OK);
1109
1110         if (request->rq_set == NULL)
1111                 wq = &request->rq_set->set_waitq;
1112         else
1113                 wq = &request->rq_reply_waitq;
1114
1115         for (;;) {
1116                 /* Network access will complete in finite time but the HUGE
1117                  * timeout lets us CWARN for visibility of sluggish NALs */
1118                 lwi = LWI_TIMEOUT(300 * HZ, NULL, NULL);
1119                 rc = l_wait_event (*wq, !ptlrpc_client_receiving_reply(request), &lwi);
1120                 if (rc == 0)
1121                         return;
1122
1123                 LASSERT (rc == -ETIMEDOUT);
1124                 DEBUG_REQ(D_WARNING, request, "Unexpectedly long timeout");
1125         }
1126 }
1127
1128 /* caller must hold imp->imp_lock */
1129 void ptlrpc_free_committed(struct obd_import *imp)
1130 {
1131         struct list_head *tmp, *saved;
1132         struct ptlrpc_request *req;
1133         struct ptlrpc_request *last_req = NULL; /* temporary fire escape */
1134         ENTRY;
1135
1136         LASSERT(imp != NULL);
1137
1138         LASSERT_SPIN_LOCKED(&imp->imp_lock);
1139
1140         CDEBUG(D_HA, "%s: committing for last_committed "LPU64"\n",
1141                imp->imp_obd->obd_name, imp->imp_peer_committed_transno);
1142
1143         list_for_each_safe(tmp, saved, &imp->imp_replay_list) {
1144                 req = list_entry(tmp, struct ptlrpc_request, rq_replay_list);
1145
1146                 /* XXX ok to remove when 1357 resolved - rread 05/29/03  */
1147                 LASSERT(req != last_req);
1148                 last_req = req;
1149
1150                 if (req->rq_import_generation < imp->imp_generation) {
1151                         DEBUG_REQ(D_HA, req, "freeing request with old gen");
1152                         GOTO(free_req, 0);
1153                 }
1154
1155                 if (req->rq_replay) {
1156                         DEBUG_REQ(D_HA, req, "keeping (FL_REPLAY)");
1157                         continue;
1158                 }
1159
1160                 /* not yet committed */
1161                 if (req->rq_transno > imp->imp_peer_committed_transno) {
1162                         DEBUG_REQ(D_HA, req, "stopping search");
1163                         break;
1164                 }
1165
1166                 DEBUG_REQ(D_HA, req, "committing (last_committed "LPU64")",
1167                           imp->imp_peer_committed_transno);
1168 free_req:
1169                 if (req->rq_commit_cb != NULL)
1170                         req->rq_commit_cb(req);
1171                 list_del_init(&req->rq_replay_list);
1172                 __ptlrpc_req_finished(req, 1);
1173         }
1174
1175         EXIT;
1176         return;
1177 }
1178
1179 void ptlrpc_cleanup_client(struct obd_import *imp)
1180 {
1181         ENTRY;
1182         EXIT;
1183         return;
1184 }
1185
1186 void ptlrpc_resend_req(struct ptlrpc_request *req)
1187 {
1188         unsigned long flags;
1189
1190         DEBUG_REQ(D_HA, req, "going to resend");
1191         req->rq_reqmsg->handle.cookie = 0;
1192         req->rq_status = -EAGAIN;
1193
1194         spin_lock_irqsave (&req->rq_lock, flags);
1195         req->rq_resend = 1;
1196         req->rq_timedout = 0;
1197         if (req->rq_bulk) {
1198                 __u64 old_xid = req->rq_xid;
1199
1200                 /* ensure previous bulk fails */
1201                 req->rq_xid = ptlrpc_next_xid();
1202                 CDEBUG(D_HA, "resend bulk old x"LPU64" new x"LPU64"\n",
1203                        old_xid, req->rq_xid);
1204         }
1205         ptlrpc_wake_client_req(req);
1206         spin_unlock_irqrestore (&req->rq_lock, flags);
1207 }
1208
1209 /* XXX: this function and rq_status are currently unused */
1210 void ptlrpc_restart_req(struct ptlrpc_request *req)
1211 {
1212         unsigned long flags;
1213
1214         DEBUG_REQ(D_HA, req, "restarting (possibly-)completed request");
1215         req->rq_status = -ERESTARTSYS;
1216
1217         spin_lock_irqsave (&req->rq_lock, flags);
1218         req->rq_restart = 1;
1219         req->rq_timedout = 0;
1220         ptlrpc_wake_client_req(req);
1221         spin_unlock_irqrestore (&req->rq_lock, flags);
1222 }
1223
1224 static int expired_request(void *data)
1225 {
1226         struct ptlrpc_request *req = data;
1227         ENTRY;
1228
1229         RETURN(ptlrpc_expire_one_request(req));
1230 }
1231
1232 static void interrupted_request(void *data)
1233 {
1234         unsigned long flags;
1235
1236         struct ptlrpc_request *req = data;
1237         DEBUG_REQ(D_HA, req, "request interrupted");
1238         spin_lock_irqsave (&req->rq_lock, flags);
1239         req->rq_intr = 1;
1240         spin_unlock_irqrestore (&req->rq_lock, flags);
1241 }
1242
1243 struct ptlrpc_request *ptlrpc_request_addref(struct ptlrpc_request *req)
1244 {
1245         ENTRY;
1246         atomic_inc(&req->rq_refcount);
1247         RETURN(req);
1248 }
1249
1250 void ptlrpc_retain_replayable_request(struct ptlrpc_request *req,
1251                                       struct obd_import *imp)
1252 {
1253         struct list_head *tmp;
1254
1255         LASSERT_SPIN_LOCKED(&imp->imp_lock);
1256
1257         /* don't re-add requests that have been replayed */
1258         if (!list_empty(&req->rq_replay_list))
1259                 return;
1260
1261         lustre_msg_add_flags(req->rq_reqmsg,
1262                              MSG_REPLAY);
1263
1264         LASSERT(imp->imp_replayable);
1265         /* Balanced in ptlrpc_free_committed, usually. */
1266         ptlrpc_request_addref(req);
1267         list_for_each_prev(tmp, &imp->imp_replay_list) {
1268                 struct ptlrpc_request *iter =
1269                         list_entry(tmp, struct ptlrpc_request, rq_replay_list);
1270
1271                 /* We may have duplicate transnos if we create and then
1272                  * open a file, or for closes retained if to match creating
1273                  * opens, so use req->rq_xid as a secondary key.
1274                  * (See bugs 684, 685, and 428.)
1275                  * XXX no longer needed, but all opens need transnos!
1276                  */
1277                 if (iter->rq_transno > req->rq_transno)
1278                         continue;
1279
1280                 if (iter->rq_transno == req->rq_transno) {
1281                         LASSERT(iter->rq_xid != req->rq_xid);
1282                         if (iter->rq_xid > req->rq_xid)
1283                                 continue;
1284                 }
1285
1286                 list_add(&req->rq_replay_list, &iter->rq_replay_list);
1287                 return;
1288         }
1289
1290         list_add_tail(&req->rq_replay_list, &imp->imp_replay_list);
1291 }
1292
1293 int ptlrpc_queue_wait(struct ptlrpc_request *req)
1294 {
1295         int rc = 0;
1296         int brc;
1297         struct l_wait_info lwi;
1298         struct obd_import *imp = req->rq_import;
1299         unsigned long flags;
1300         int timeout = 0;
1301         ENTRY;
1302
1303         LASSERT(req->rq_set == NULL);
1304         LASSERT(!req->rq_receiving_reply);
1305         atomic_inc(&imp->imp_inflight);
1306
1307         /* for distributed debugging */
1308         req->rq_reqmsg->status = current->pid;
1309         LASSERT(imp->imp_obd != NULL);
1310         CDEBUG(D_RPCTRACE, "Sending RPC pname:cluuid:pid:xid:ni:nid:opc "
1311                "%s:%s:%d:"LPU64":%s:"LPX64":%d\n", current->comm,
1312                imp->imp_obd->obd_uuid.uuid,
1313                req->rq_reqmsg->status, req->rq_xid,
1314                imp->imp_connection->c_peer.peer_ni->pni_name,
1315                imp->imp_connection->c_peer.peer_nid,
1316                req->rq_reqmsg->opc);
1317
1318         /* Mark phase here for a little debug help */
1319         req->rq_phase = RQ_PHASE_RPC;
1320
1321         spin_lock_irqsave(&imp->imp_lock, flags);
1322         req->rq_import_generation = imp->imp_generation;
1323 restart:
1324         if (ptlrpc_import_delay_req(imp, req, &rc)) {
1325                 list_del(&req->rq_list);
1326
1327                 list_add_tail(&req->rq_list, &imp->imp_delayed_list);
1328                 spin_unlock_irqrestore(&imp->imp_lock, flags);
1329
1330                 DEBUG_REQ(D_HA, req, "\"%s\" waiting for recovery: (%s != %s)",
1331                           current->comm,
1332                           ptlrpc_import_state_name(req->rq_send_state),
1333                           ptlrpc_import_state_name(imp->imp_state));
1334                 lwi = LWI_INTR(interrupted_request, req);
1335                 rc = l_wait_event(req->rq_reply_waitq,
1336                                   (req->rq_send_state == imp->imp_state ||
1337                                    req->rq_err),
1338                                   &lwi);
1339                 DEBUG_REQ(D_HA, req, "\"%s\" awake: (%s == %s or %d == 1)",
1340                           current->comm,
1341                           ptlrpc_import_state_name(imp->imp_state),
1342                           ptlrpc_import_state_name(req->rq_send_state),
1343                           req->rq_err);
1344
1345                 spin_lock_irqsave(&imp->imp_lock, flags);
1346                 list_del_init(&req->rq_list);
1347
1348                 if (req->rq_err) {
1349                         rc = -EIO;
1350                 }
1351                 else if (req->rq_intr) {
1352                         rc = -EINTR;
1353                 }
1354                 else if (req->rq_no_resend) {
1355                         spin_unlock_irqrestore(&imp->imp_lock, flags);
1356                         GOTO(out, rc = -ETIMEDOUT);
1357                 }
1358                 else {
1359                         GOTO(restart, rc);
1360                 }
1361         }
1362
1363         if (rc != 0) {
1364                 list_del_init(&req->rq_list);
1365                 spin_unlock_irqrestore(&imp->imp_lock, flags);
1366                 req->rq_status = rc; // XXX this ok?
1367                 GOTO(out, rc);
1368         }
1369
1370         if (req->rq_resend) {
1371                 lustre_msg_add_flags(req->rq_reqmsg, MSG_RESENT);
1372
1373                 if (req->rq_bulk != NULL)
1374                         ptlrpc_unregister_bulk (req);
1375
1376                 DEBUG_REQ(D_HA, req, "resending: ");
1377         }
1378
1379         /* XXX this is the same as ptlrpc_set_wait */
1380         LASSERT(list_empty(&req->rq_list));
1381         list_add_tail(&req->rq_list, &imp->imp_sending_list);
1382         spin_unlock_irqrestore(&imp->imp_lock, flags);
1383
1384         rc = ptl_send_rpc(req);
1385         if (rc) {
1386                 DEBUG_REQ(D_HA, req, "send failed (%d); recovering", rc);
1387                 timeout = 1;
1388         } else {
1389                 timeout = MAX(req->rq_timeout * HZ, 1);
1390                 DEBUG_REQ(D_NET, req, "-- sleeping");
1391         }
1392         lwi = LWI_TIMEOUT_INTR(timeout, expired_request, interrupted_request,
1393                                req);
1394         l_wait_event(req->rq_reply_waitq, ptlrpc_check_reply(req), &lwi);
1395         DEBUG_REQ(D_NET, req, "-- done sleeping");
1396
1397         CDEBUG(D_RPCTRACE, "Completed RPC pname:cluuid:pid:xid:ni:nid:opc "
1398                "%s:%s:%d:"LPU64":%s:"LPX64":%d\n", current->comm,
1399                imp->imp_obd->obd_uuid.uuid,
1400                req->rq_reqmsg->status, req->rq_xid,
1401                imp->imp_connection->c_peer.peer_ni->pni_name,
1402                imp->imp_connection->c_peer.peer_nid,
1403                req->rq_reqmsg->opc);
1404
1405         spin_lock_irqsave(&imp->imp_lock, flags);
1406         list_del_init(&req->rq_list);
1407         spin_unlock_irqrestore(&imp->imp_lock, flags);
1408
1409         /* If the reply was received normally, this just grabs the spinlock
1410          * (ensuring the reply callback has returned), sees that
1411          * req->rq_receiving_reply is clear and returns. */
1412         ptlrpc_unregister_reply (req);
1413
1414         if (req->rq_err)
1415                 GOTO(out, rc = -EIO);
1416
1417         /* Resend if we need to, unless we were interrupted. */
1418         if (req->rq_resend && !req->rq_intr) {
1419                 /* ...unless we were specifically told otherwise. */
1420                 if (req->rq_no_resend)
1421                         GOTO(out, rc = -ETIMEDOUT);
1422                 spin_lock_irqsave(&imp->imp_lock, flags);
1423                 goto restart;
1424         }
1425
1426         if (req->rq_intr) {
1427                 /* Should only be interrupted if we timed out. */
1428                 if (!req->rq_timedout)
1429                         DEBUG_REQ(D_ERROR, req,
1430                                   "rq_intr set but rq_timedout not");
1431                 GOTO(out, rc = -EINTR);
1432         }
1433
1434         if (req->rq_timedout) {                 /* non-recoverable timeout */
1435                 GOTO(out, rc = -ETIMEDOUT);
1436         }
1437
1438         if (!req->rq_replied) {
1439                 /* How can this be? -eeb */
1440                 DEBUG_REQ(D_ERROR, req, "!rq_replied: ");
1441                 LBUG();
1442                 GOTO(out, rc = req->rq_status);
1443         }
1444
1445         rc = after_reply (req);
1446         /* NB may return +ve success rc */
1447         if (req->rq_resend) {
1448                 spin_lock_irqsave(&imp->imp_lock, flags);
1449                 goto restart;
1450         }
1451
1452  out:
1453         if (req->rq_bulk != NULL) {
1454                 if (rc >= 0) {
1455                         /* success so far.  Note that anything going wrong
1456                          * with bulk now, is EXTREMELY strange, since the
1457                          * server must have believed that the bulk
1458                          * tranferred OK before she replied with success to
1459                          * me. */
1460                         lwi = LWI_TIMEOUT(timeout, NULL, NULL);
1461                         brc = l_wait_event(req->rq_reply_waitq,
1462                                            !ptlrpc_bulk_active(req->rq_bulk),
1463                                            &lwi);
1464                         LASSERT(brc == 0 || brc == -ETIMEDOUT);
1465                         if (brc != 0) {
1466                                 LASSERT(brc == -ETIMEDOUT);
1467                                 DEBUG_REQ(D_ERROR, req, "bulk timed out");
1468                                 rc = brc;
1469                         } else if (!req->rq_bulk->bd_success) {
1470                                 DEBUG_REQ(D_ERROR, req, "bulk transfer failed");
1471                                 rc = -EIO;
1472                         }
1473                 }
1474                 if (rc < 0)
1475                         ptlrpc_unregister_bulk (req);
1476         }
1477
1478         LASSERT(!req->rq_receiving_reply);
1479         req->rq_phase = RQ_PHASE_INTERPRET;
1480
1481         atomic_dec(&imp->imp_inflight);
1482         wake_up(&imp->imp_recovery_waitq);
1483         RETURN(rc);
1484 }
1485
1486 struct ptlrpc_replay_async_args {
1487         int praa_old_state;
1488         int praa_old_status;
1489 };
1490
1491 static int ptlrpc_replay_interpret(struct ptlrpc_request *req,
1492                                     void * data, int rc)
1493 {
1494         struct ptlrpc_replay_async_args *aa = data;
1495         struct obd_import *imp = req->rq_import;
1496         unsigned long flags;
1497
1498         atomic_dec(&imp->imp_replay_inflight);
1499
1500         if (!req->rq_replied) {
1501                 CERROR("request replay timed out, restarting recovery\n");
1502                 GOTO(out, rc = -ETIMEDOUT);
1503         }
1504
1505 #if SWAB_PARANOIA
1506         /* Clear reply swab mask; this is a new reply in sender's byte order */
1507         req->rq_rep_swab_mask = 0;
1508 #endif
1509         LASSERT (req->rq_nob_received <= req->rq_replen);
1510         rc = lustre_unpack_msg(req->rq_repmsg, req->rq_nob_received);
1511         if (rc) {
1512                 CERROR("unpack_rep failed: %d\n", rc);
1513                 GOTO(out, rc = -EPROTO);
1514         }
1515
1516         if (req->rq_repmsg->type == PTL_RPC_MSG_ERR &&
1517             req->rq_repmsg->status == -ENOTCONN)
1518                 GOTO(out, rc = req->rq_repmsg->status);
1519
1520         /* The transno had better not change over replay. */
1521         LASSERT(req->rq_reqmsg->transno == req->rq_repmsg->transno);
1522
1523         DEBUG_REQ(D_HA, req, "got rep");
1524
1525         /* let the callback do fixups, possibly including in the request */
1526         if (req->rq_replay_cb)
1527                 req->rq_replay_cb(req);
1528
1529         if (req->rq_replied && req->rq_repmsg->status != aa->praa_old_status) {
1530                 DEBUG_REQ(D_ERROR, req, "status %d, old was %d",
1531                           req->rq_repmsg->status, aa->praa_old_status);
1532         } else {
1533                 /* Put it back for re-replay. */
1534                 req->rq_repmsg->status = aa->praa_old_status;
1535         }
1536
1537         spin_lock_irqsave(&imp->imp_lock, flags);
1538         imp->imp_last_replay_transno = req->rq_transno;
1539         spin_unlock_irqrestore(&imp->imp_lock, flags);
1540
1541         /* continue with recovery */
1542         rc = ptlrpc_import_recovery_state_machine(imp);
1543  out:
1544         req->rq_send_state = aa->praa_old_state;
1545
1546         if (rc != 0)
1547                 /* this replay failed, so restart recovery */
1548                 ptlrpc_connect_import(imp, NULL);
1549
1550         RETURN(rc);
1551 }
1552
1553
1554 int ptlrpc_replay_req(struct ptlrpc_request *req)
1555 {
1556         struct ptlrpc_replay_async_args *aa;
1557         ENTRY;
1558
1559         LASSERT(req->rq_import->imp_state == LUSTRE_IMP_REPLAY);
1560
1561         /* Not handling automatic bulk replay yet (or ever?) */
1562         LASSERT(req->rq_bulk == NULL);
1563
1564         DEBUG_REQ(D_HA, req, "REPLAY");
1565
1566         LASSERT (sizeof (*aa) <= sizeof (req->rq_async_args));
1567         aa = (struct ptlrpc_replay_async_args *)&req->rq_async_args;
1568         memset(aa, 0, sizeof *aa);
1569
1570         /* Prepare request to be resent with ptlrpcd */
1571         aa->praa_old_state = req->rq_send_state;
1572         req->rq_send_state = LUSTRE_IMP_REPLAY;
1573         req->rq_phase = RQ_PHASE_NEW;
1574         /*
1575          * Q: "How can a req get on the replay list if it wasn't replied?"
1576          * A: "If we failed during the replay of this request, it will still
1577          *     be on the list, but rq_replied will have been reset to 0."
1578          */
1579         if (req->rq_replied) {
1580                 aa->praa_old_status = req->rq_repmsg->status;
1581                 req->rq_status = 0;
1582                 req->rq_replied = 0;
1583         }
1584
1585         req->rq_interpret_reply = ptlrpc_replay_interpret;
1586         atomic_inc(&req->rq_import->imp_replay_inflight);
1587         ptlrpc_request_addref(req); /* ptlrpcd needs a ref */
1588
1589         ptlrpcd_add_req(req);
1590         RETURN(0);
1591 }
1592
1593 void ptlrpc_abort_inflight(struct obd_import *imp)
1594 {
1595         unsigned long flags;
1596         struct list_head *tmp, *n;
1597         ENTRY;
1598
1599         /* Make sure that no new requests get processed for this import.
1600          * ptlrpc_{queue,set}_wait must (and does) hold imp_lock while testing
1601          * this flag and then putting requests on sending_list or delayed_list.
1602          */
1603         spin_lock_irqsave(&imp->imp_lock, flags);
1604
1605         /* XXX locking?  Maybe we should remove each request with the list
1606          * locked?  Also, how do we know if the requests on the list are
1607          * being freed at this time?
1608          */
1609         list_for_each_safe(tmp, n, &imp->imp_sending_list) {
1610                 struct ptlrpc_request *req =
1611                         list_entry(tmp, struct ptlrpc_request, rq_list);
1612
1613                 DEBUG_REQ(D_HA, req, "inflight");
1614
1615                 spin_lock (&req->rq_lock);
1616                 if (req->rq_import_generation < imp->imp_generation) {
1617                         req->rq_err = 1;
1618                         ptlrpc_wake_client_req(req);
1619                 }
1620                 spin_unlock (&req->rq_lock);
1621         }
1622
1623         list_for_each_safe(tmp, n, &imp->imp_delayed_list) {
1624                 struct ptlrpc_request *req =
1625                         list_entry(tmp, struct ptlrpc_request, rq_list);
1626
1627                 DEBUG_REQ(D_HA, req, "aborting waiting req");
1628
1629                 spin_lock (&req->rq_lock);
1630                 if (req->rq_import_generation < imp->imp_generation) {
1631                         req->rq_err = 1;
1632                         ptlrpc_wake_client_req(req);
1633                 }
1634                 spin_unlock (&req->rq_lock);
1635         }
1636
1637         /* Last chance to free reqs left on the replay list, but we
1638          * will still leak reqs that haven't comitted.  */
1639         if (imp->imp_replayable)
1640                 ptlrpc_free_committed(imp);
1641
1642         spin_unlock_irqrestore(&imp->imp_lock, flags);
1643
1644         EXIT;
1645 }
1646
1647 static __u64 ptlrpc_last_xid = 0;
1648 static spinlock_t ptlrpc_last_xid_lock = SPIN_LOCK_UNLOCKED;
1649
1650 __u64 ptlrpc_next_xid(void)
1651 {
1652         __u64 tmp;
1653         spin_lock(&ptlrpc_last_xid_lock);
1654         tmp = ++ptlrpc_last_xid;
1655         spin_unlock(&ptlrpc_last_xid_lock);
1656         return tmp;
1657 }
1658
1659