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         ptlrpc_add_bulk_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                         spin_unlock_irqrestore(&imp->imp_lock, flags);
 502                         req->rq_commit_cb(req);
 503                         spin_lock_irqsave(&imp->imp_lock, flags);
 504                 }
 505
 506                 if (req->rq_transno > imp->imp_max_transno)
 507                         imp->imp_max_transno = req->rq_transno;
 508
 509                 /* Replay-enabled imports return commit-status information. */
 510                 if (req->rq_repmsg->last_committed)
 511                         imp->imp_peer_committed_transno =
 512                                 req->rq_repmsg->last_committed;
 513                 ptlrpc_free_committed(imp);
 514                 spin_unlock_irqrestore(&imp->imp_lock, flags);
 515         }
 516
 517         RETURN(rc);
 518 }
 519
 520 static int ptlrpc_send_new_req(struct ptlrpc_request *req)
 521 {
 522         char                   str[PTL_NALFMT_SIZE];
 523         struct obd_import     *imp;
 524         unsigned long          flags;
 525         int rc;
 526         ENTRY;
 527
 528         LASSERT(req->rq_phase == RQ_PHASE_NEW);
 529         req->rq_phase = RQ_PHASE_RPC;
 530
 531         imp = req->rq_import;
 532         spin_lock_irqsave(&imp->imp_lock, flags);
 533
 534         req->rq_import_generation = imp->imp_generation;
 535
 536         if (ptlrpc_import_delay_req(imp, req, &rc)) {
 537                 spin_lock (&req->rq_lock);
 538                 req->rq_waiting = 1;
 539                 spin_unlock (&req->rq_lock);
 540
 541                 DEBUG_REQ(D_HA, req, "req from PID %d waiting for recovery: "
 542                           "(%s != %s)",
 543                           req->rq_reqmsg->status,
 544                           ptlrpc_import_state_name(req->rq_send_state),
 545                           ptlrpc_import_state_name(imp->imp_state));
 546                 LASSERT(list_empty (&req->rq_list));
 547
 548                 list_add_tail(&req->rq_list, &imp->imp_delayed_list);
 549                 spin_unlock_irqrestore(&imp->imp_lock, flags);
 550                 RETURN(0);
 551         }
 552
 553         if (rc != 0) {
 554                 spin_unlock_irqrestore(&imp->imp_lock, flags);
 555                 req->rq_status = rc;
 556                 req->rq_phase = RQ_PHASE_INTERPRET;
 557                 RETURN(rc);
 558         }
 559
 560         /* XXX this is the same as ptlrpc_queue_wait */
 561         LASSERT(list_empty(&req->rq_list));
 562         list_add_tail(&req->rq_list, &imp->imp_sending_list);
 563         spin_unlock_irqrestore(&imp->imp_lock, flags);
 564
 565         req->rq_reqmsg->status = current->pid;
 566         CDEBUG(D_RPCTRACE, "Sending RPC pname:cluuid:pid:xid:ni:nid:opc"
 567                " %s:%s:%d:"LPU64":%s:%s:%d\n", current->comm,
 568                imp->imp_obd->obd_uuid.uuid, req->rq_reqmsg->status,
 569                req->rq_xid,
 570                imp->imp_connection->c_peer.peer_ni->pni_name,
 571                ptlrpc_peernid2str(&imp->imp_connection->c_peer, str),
 572                req->rq_reqmsg->opc);
 573
 574         rc = ptl_send_rpc(req);
 575         if (rc) {
 576                 DEBUG_REQ(D_HA, req, "send failed (%d); expect timeout", rc);
 577                 req->rq_timeout = 1;
 578                 RETURN(rc);
 579         }
 580         RETURN(0);
 581 }
 582
 583 int ptlrpc_check_set(struct ptlrpc_request_set *set)
 584 {
 585         char str[PTL_NALFMT_SIZE];
 586         unsigned long flags;
 587         struct list_head *tmp;
 588         int force_timer_recalc = 0;
 589         ENTRY;
 590
 591         if (set->set_remaining == 0)
 592                 RETURN(1);
 593
 594         list_for_each(tmp, &set->set_requests) {
 595                 struct ptlrpc_request *req =
 596                         list_entry(tmp, struct ptlrpc_request, rq_set_chain);
 597                 struct obd_import *imp = req->rq_import;
 598                 int rc = 0;
 599
 600                 if (req->rq_phase == RQ_PHASE_NEW &&
 601                     ptlrpc_send_new_req(req)) {
 602                         force_timer_recalc = 1;
 603                 }
 604
 605                 if (!(req->rq_phase == RQ_PHASE_RPC ||
 606                       req->rq_phase == RQ_PHASE_BULK ||
 607                       req->rq_phase == RQ_PHASE_INTERPRET ||
 608                       req->rq_phase == RQ_PHASE_COMPLETE)) {
 609                         DEBUG_REQ(D_ERROR, req, "bad phase %x", req->rq_phase);
 610                         LBUG();
 611                 }
 612
 613                 if (req->rq_phase == RQ_PHASE_COMPLETE)
 614                         continue;
 615
 616                 if (req->rq_phase == RQ_PHASE_INTERPRET)
 617                         GOTO(interpret, req->rq_status);
 618
 619                 if (req->rq_err) {
 620                         ptlrpc_unregister_reply(req);
 621                         if (req->rq_status == 0)
 622                                 req->rq_status = -EIO;
 623                         req->rq_phase = RQ_PHASE_INTERPRET;
 624
 625                         spin_lock_irqsave(&imp->imp_lock, flags);
 626                         list_del_init(&req->rq_list);
 627                         spin_unlock_irqrestore(&imp->imp_lock, flags);
 628
 629                         GOTO(interpret, req->rq_status);
 630                 }
 631
 632                 /* ptlrpc_queue_wait->l_wait_event guarantees that rq_intr
 633                  * will only be set after rq_timedout, but the oig waiting
 634                  * path sets rq_intr irrespective of whether ptlrpcd has
 635                  * seen a timeout.  our policy is to only interpret
 636                  * interrupted rpcs after they have timed out */
 637                 if (req->rq_intr && (req->rq_timedout || req->rq_waiting)) {
 638                         /* NB could be on delayed list */
 639                         ptlrpc_unregister_reply(req);
 640                         req->rq_status = -EINTR;
 641                         req->rq_phase = RQ_PHASE_INTERPRET;
 642
 643                         spin_lock_irqsave(&imp->imp_lock, flags);
 644                         list_del_init(&req->rq_list);
 645                         spin_unlock_irqrestore(&imp->imp_lock, flags);
 646
 647                         GOTO(interpret, req->rq_status);
 648                 }
 649
 650                 if (req->rq_phase == RQ_PHASE_RPC) {
 651                         if (req->rq_waiting || req->rq_resend) {
 652                                 int status;
 653
 654                                 LASSERT (!ptlrpc_client_receiving_reply(req));
 655                                 LASSERT (req->rq_bulk == NULL ||
 656                                          !ptlrpc_bulk_active(req->rq_bulk));
 657
 658                                 spin_lock_irqsave(&imp->imp_lock, flags);
 659
 660                                 if (ptlrpc_import_delay_req(imp, req, &status)) {
 661                                         spin_unlock_irqrestore(&imp->imp_lock,
 662                                                                flags);
 663                                         continue;
 664                                 }
 665
 666                                 list_del_init(&req->rq_list);
 667                                 if (status != 0)  {
 668                                         req->rq_status = status;
 669                                         req->rq_phase = RQ_PHASE_INTERPRET;
 670                                         spin_unlock_irqrestore(&imp->imp_lock,
 671                                                                flags);
 672                                         GOTO(interpret, req->rq_status);
 673                                 }
 674                                 if (req->rq_no_resend) {
 675                                         req->rq_status = -ENOTCONN;
 676                                         req->rq_phase = RQ_PHASE_INTERPRET;
 677                                         spin_unlock_irqrestore(&imp->imp_lock,
 678                                                                flags);
 679                                         GOTO(interpret, req->rq_status);
 680                                 }
 681                                 list_add_tail(&req->rq_list,
 682                                               &imp->imp_sending_list);
 683
 684                                 spin_unlock_irqrestore(&imp->imp_lock, flags);
 685
 686                                 req->rq_waiting = 0;
 687                                 if (req->rq_resend) {
 688                                         lustre_msg_add_flags(req->rq_reqmsg,
 689                                                              MSG_RESENT);
 690                                         ptlrpc_unregister_reply(req);
 691                                         if (req->rq_bulk) {
 692                                                 __u64 old_xid = req->rq_xid;
 693
 694                                                 /* ensure previous bulk fails */
 695                                                 req->rq_xid = ptlrpc_next_xid();
 696                                                 CDEBUG(D_HA, "resend bulk "
 697                                                        "old x"LPU64
 698                                                        " new x"LPU64"\n",
 699                                                        old_xid, req->rq_xid);
 700                                         }
 701                                 }
 702
 703                                 rc = ptl_send_rpc(req);
 704                                 if (rc) {
 705                                         DEBUG_REQ(D_HA, req, "send failed (%d)",
 706                                                   rc);
 707                                         force_timer_recalc = 1;
 708                                         req->rq_timeout = 0;
 709                                 }
 710                                 /* need to reset the timeout */
 711                                 force_timer_recalc = 1;
 712                         }
 713
 714                         /* Still waiting for a reply? */
 715                         if (ptlrpc_client_receiving_reply(req))
 716                                 continue;
 717
 718                         /* Did we actually receive a reply? */
 719                         if (!ptlrpc_client_replied(req))
 720                                 continue;
 721
 722                         spin_lock_irqsave(&imp->imp_lock, flags);
 723                         list_del_init(&req->rq_list);
 724                         spin_unlock_irqrestore(&imp->imp_lock, flags);
 725
 726                         req->rq_status = after_reply(req);
 727                         if (req->rq_resend) {
 728                                 /* Add this req to the delayed list so
 729                                    it can be errored if the import is
 730                                    evicted after recovery. */
 731                                 spin_lock_irqsave (&req->rq_lock, flags);
 732                                 list_add_tail(&req->rq_list,
 733                                               &imp->imp_delayed_list);
 734                                 spin_unlock_irqrestore(&req->rq_lock, flags);
 735                                 continue;
 736                         }
 737
 738                         /* If there is no bulk associated with this request,
 739                          * then we're done and should let the interpreter
 740                          * process the reply.  Similarly if the RPC returned
 741                          * an error, and therefore the bulk will never arrive.
 742                          */
 743                         if (req->rq_bulk == NULL || req->rq_status != 0) {
 744                                 req->rq_phase = RQ_PHASE_INTERPRET;
 745                                 GOTO(interpret, req->rq_status);
 746                         }
 747
 748                         req->rq_phase = RQ_PHASE_BULK;
 749                 }
 750
 751                 LASSERT(req->rq_phase == RQ_PHASE_BULK);
 752                 if (ptlrpc_bulk_active(req->rq_bulk))
 753                         continue;
 754
 755                 if (!req->rq_bulk->bd_success) {
 756                         /* The RPC reply arrived OK, but the bulk screwed
 757                          * up!  Dead wierd since the server told us the RPC
 758                          * was good after getting the REPLY for her GET or
 759                          * the ACK for her PUT. */
 760                         DEBUG_REQ(D_ERROR, req, "bulk transfer failed");
 761                         LBUG();
 762                 }
 763
 764                 req->rq_phase = RQ_PHASE_INTERPRET;
 765
 766         interpret:
 767                 LASSERT(req->rq_phase == RQ_PHASE_INTERPRET);
 768                 LASSERT(!req->rq_receiving_reply);
 769
 770                 ptlrpc_unregister_reply(req);
 771                 if (req->rq_bulk != NULL)
 772                         ptlrpc_unregister_bulk (req);
 773
 774                 req->rq_phase = RQ_PHASE_COMPLETE;
 775
 776                 if (req->rq_interpret_reply != NULL) {
 777                         int (*interpreter)(struct ptlrpc_request *,void *,int) =
 778                                 req->rq_interpret_reply;
 779                         req->rq_status = interpreter(req, &req->rq_async_args,
 780                                                      req->rq_status);
 781                 }
 782
 783                 CDEBUG(D_RPCTRACE, "Completed RPC pname:cluuid:pid:xid:ni:nid:"
 784                        "opc %s:%s:%d:"LPU64":%s:%s:%d\n", current->comm,
 785                        imp->imp_obd->obd_uuid.uuid, req->rq_reqmsg->status,
 786                        req->rq_xid,
 787                        imp->imp_connection->c_peer.peer_ni->pni_name,
 788                        ptlrpc_peernid2str(&imp->imp_connection->c_peer, str),
 789                        req->rq_reqmsg->opc);
 790
 791                 set->set_remaining--;
 792
 793                 atomic_dec(&imp->imp_inflight);
 794                 wake_up(&imp->imp_recovery_waitq);
 795         }
 796
 797         /* If we hit an error, we want to recover promptly. */
 798         RETURN(set->set_remaining == 0 || force_timer_recalc);
 799 }
 800
 801 int ptlrpc_expire_one_request(struct ptlrpc_request *req)
 802 {
 803         unsigned long      flags;
 804         struct obd_import *imp = req->rq_import;
 805         ENTRY;
 806
 807         DEBUG_REQ(D_ERROR, req, "timeout");
 808
 809         spin_lock_irqsave (&req->rq_lock, flags);
 810         req->rq_timedout = 1;
 811         spin_unlock_irqrestore (&req->rq_lock, flags);
 812
 813         ptlrpc_unregister_reply (req);
 814
 815         if (req->rq_bulk != NULL)
 816                 ptlrpc_unregister_bulk (req);
 817
 818         if (imp == NULL) {
 819                 DEBUG_REQ(D_HA, req, "NULL import: already cleaned up?");
 820                 RETURN(1);
 821         }
 822
 823         /* The DLM server doesn't want recovery run on its imports. */
 824         if (imp->imp_dlm_fake)
 825                 RETURN(1);
 826
 827         /* If this request is for recovery or other primordial tasks,
 828          * then error it out here. */
 829         if (req->rq_send_state != LUSTRE_IMP_FULL ||
 830             imp->imp_obd->obd_no_recov) {
 831                 spin_lock_irqsave (&req->rq_lock, flags);
 832                 req->rq_status = -ETIMEDOUT;
 833                 req->rq_err = 1;
 834                 spin_unlock_irqrestore (&req->rq_lock, flags);
 835                 RETURN(1);
 836         }
 837
 838         ptlrpc_fail_import(imp, req->rq_import_generation);
 839
 840         RETURN(0);
 841 }
 842
 843 int ptlrpc_expired_set(void *data)
 844 {
 845         struct ptlrpc_request_set *set = data;
 846         struct list_head          *tmp;
 847         time_t                     now = LTIME_S (CURRENT_TIME);
 848         ENTRY;
 849
 850         LASSERT(set != NULL);
 851
 852         /* A timeout expired; see which reqs it applies to... */
 853         list_for_each (tmp, &set->set_requests) {
 854                 struct ptlrpc_request *req =
 855                         list_entry(tmp, struct ptlrpc_request, rq_set_chain);
 856
 857                 /* request in-flight? */
 858                 if (!((req->rq_phase == RQ_PHASE_RPC && !req->rq_waiting
 859                        && !req->rq_resend) ||
 860                       (req->rq_phase == RQ_PHASE_BULK)))
 861                         continue;
 862
 863                 if (req->rq_timedout ||           /* already dealt with */
 864                     req->rq_sent + req->rq_timeout > now) /* not expired */
 865                         continue;
 866
 867                 /* deal with this guy */
 868                 ptlrpc_expire_one_request (req);
 869         }
 870
 871         /* When waiting for a whole set, we always to break out of the
 872          * sleep so we can recalculate the timeout, or enable interrupts
 873          * iff everyone's timed out.
 874          */
 875         RETURN(1);
 876 }
 877
 878 void ptlrpc_mark_interrupted(struct ptlrpc_request *req)
 879 {
 880         unsigned long flags;
 881         spin_lock_irqsave(&req->rq_lock, flags);
 882         req->rq_intr = 1;
 883         spin_unlock_irqrestore(&req->rq_lock, flags);
 884 }
 885
 886 void ptlrpc_interrupted_set(void *data)
 887 {
 888         struct ptlrpc_request_set *set = data;
 889         struct list_head *tmp;
 890
 891         LASSERT(set != NULL);
 892         CERROR("INTERRUPTED SET %p\n", set);
 893
 894         list_for_each(tmp, &set->set_requests) {
 895                 struct ptlrpc_request *req =
 896                         list_entry(tmp, struct ptlrpc_request, rq_set_chain);
 897
 898                 if (req->rq_phase != RQ_PHASE_RPC)
 899                         continue;
 900
 901                 ptlrpc_mark_interrupted(req);
 902         }
 903 }
 904
 905 int ptlrpc_set_next_timeout(struct ptlrpc_request_set *set)
 906 {
 907         struct list_head      *tmp;
 908         time_t                 now = LTIME_S(CURRENT_TIME);
 909         time_t                 deadline;
 910         int                    timeout = 0;
 911         struct ptlrpc_request *req;
 912         ENTRY;
 913
 914         SIGNAL_MASK_ASSERT(); /* XXX BUG 1511 */
 915
 916         list_for_each(tmp, &set->set_requests) {
 917                 req = list_entry(tmp, struct ptlrpc_request, rq_set_chain);
 918
 919                 /* request in-flight? */
 920                 if (!((req->rq_phase == RQ_PHASE_RPC && !req->rq_waiting) ||
 921                       (req->rq_phase == RQ_PHASE_BULK)))
 922                         continue;
 923
 924                 if (req->rq_timedout)   /* already timed out */
 925                         continue;
 926
 927                 deadline = req->rq_sent + req->rq_timeout;
 928                 if (deadline <= now)    /* actually expired already */
 929                         timeout = 1;    /* ASAP */
 930                 else if (timeout == 0 || timeout > deadline - now)
 931                         timeout = deadline - now;
 932         }
 933         RETURN(timeout);
 934 }
 935
 936
 937 int ptlrpc_set_wait(struct ptlrpc_request_set *set)
 938 {
 939         struct list_head      *tmp;
 940         struct ptlrpc_request *req;
 941         struct l_wait_info     lwi;
 942         int                    rc, timeout;
 943         ENTRY;
 944
 945         LASSERT(!list_empty(&set->set_requests));
 946         list_for_each(tmp, &set->set_requests) {
 947                 req = list_entry(tmp, struct ptlrpc_request, rq_set_chain);
 948                 if (req->rq_phase == RQ_PHASE_NEW)
 949                         (void)ptlrpc_send_new_req(req);
 950         }
 951
 952         do {
 953                 timeout = ptlrpc_set_next_timeout(set);
 954
 955                 /* wait until all complete, interrupted, or an in-flight
 956                  * req times out */
 957                 CDEBUG(D_HA, "set %p going to sleep for %d seconds\n",
 958                        set, timeout);
 959                 lwi = LWI_TIMEOUT_INTR((timeout ? timeout : 1) * HZ,
 960                                        ptlrpc_expired_set,
 961                                        ptlrpc_interrupted_set, set);
 962                 rc = l_wait_event(set->set_waitq, ptlrpc_check_set(set), &lwi);
 963
 964                 LASSERT(rc == 0 || rc == -EINTR || rc == -ETIMEDOUT);
 965
 966                 /* -EINTR => all requests have been flagged rq_intr so next
 967                  * check completes.
 968                  * -ETIMEOUTD => someone timed out.  When all reqs have
 969                  * timed out, signals are enabled allowing completion with
 970                  * EINTR.
 971                  * I don't really care if we go once more round the loop in
 972                  * the error cases -eeb. */
 973         } while (rc != 0);
 974
 975         LASSERT(set->set_remaining == 0);
 976
 977         rc = 0;
 978         list_for_each(tmp, &set->set_requests) {
 979                 req = list_entry(tmp, struct ptlrpc_request, rq_set_chain);
 980
 981                 LASSERT(req->rq_phase == RQ_PHASE_COMPLETE);
 982                 if (req->rq_status != 0)
 983                         rc = req->rq_status;
 984         }
 985
 986         if (set->set_interpret != NULL) {
 987                 int (*interpreter)(struct ptlrpc_request_set *set,void *,int) =
 988                         set->set_interpret;
 989                 rc = interpreter (set, &set->set_args, rc);
 990         }
 991
 992         RETURN(rc);
 993 }
 994
 995 static void __ptlrpc_free_req(struct ptlrpc_request *request, int locked)
 996 {
 997         ENTRY;
 998         if (request == NULL) {
 999                 EXIT;
1000                 return;
1001         }
1002
1003         LASSERTF(!request->rq_receiving_reply, "req %p\n", request);
1004         LASSERTF(request->rq_rqbd == NULL, "req %p\n",request);/* client-side */
1005         LASSERTF(list_empty(&request->rq_list), "req %p\n", request);
1006         LASSERTF(list_empty(&request->rq_set_chain), "req %p\n", request);
1007
1008         /* We must take it off the imp_replay_list first.  Otherwise, we'll set
1009          * request->rq_reqmsg to NULL while osc_close is dereferencing it. */
1010         if (request->rq_import != NULL) {
1011                 unsigned long flags = 0;
1012                 if (!locked)
1013                         spin_lock_irqsave(&request->rq_import->imp_lock, flags);
1014                 list_del_init(&request->rq_replay_list);
1015                 if (!locked)
1016                         spin_unlock_irqrestore(&request->rq_import->imp_lock,
1017                                                flags);
1018         }
1019         LASSERTF(list_empty(&request->rq_replay_list), "req %p\n", request);
1020
1021         if (atomic_read(&request->rq_refcount) != 0) {
1022                 DEBUG_REQ(D_ERROR, request,
1023                           "freeing request with nonzero refcount");
1024                 LBUG();
1025         }
1026
1027         if (request->rq_repmsg != NULL) {
1028                 OBD_FREE(request->rq_repmsg, request->rq_replen);
1029                 request->rq_repmsg = NULL;
1030         }
1031         if (request->rq_reqmsg != NULL) {
1032                 OBD_FREE(request->rq_reqmsg, request->rq_reqlen);
1033                 request->rq_reqmsg = NULL;
1034         }
1035         if (request->rq_export != NULL) {
1036                 class_export_put(request->rq_export);
1037                 request->rq_export = NULL;
1038         }
1039         if (request->rq_import != NULL) {
1040                 class_import_put(request->rq_import);
1041                 request->rq_import = NULL;
1042         }
1043         if (request->rq_bulk != NULL)
1044                 ptlrpc_free_bulk(request->rq_bulk);
1045
1046         OBD_FREE(request, sizeof(*request));
1047         EXIT;
1048 }
1049
1050 void ptlrpc_free_req(struct ptlrpc_request *request)
1051 {
1052         __ptlrpc_free_req(request, 0);
1053 }
1054
1055 static int __ptlrpc_req_finished(struct ptlrpc_request *request, int locked);
1056 void ptlrpc_req_finished_with_imp_lock(struct ptlrpc_request *request)
1057 {
1058         LASSERT_SPIN_LOCKED(&request->rq_import->imp_lock);
1059         (void)__ptlrpc_req_finished(request, 1);
1060 }
1061
1062 static int __ptlrpc_req_finished(struct ptlrpc_request *request, int locked)
1063 {
1064         ENTRY;
1065         if (request == NULL)
1066                 RETURN(1);
1067
1068         if (request == LP_POISON ||
1069             request->rq_reqmsg == LP_POISON) {
1070                 CERROR("dereferencing freed request (bug 575)\n");
1071                 LBUG();
1072                 RETURN(1);
1073         }
1074
1075         DEBUG_REQ(D_INFO, request, "refcount now %u",
1076                   atomic_read(&request->rq_refcount) - 1);
1077
1078         if (atomic_dec_and_test(&request->rq_refcount)) {
1079                 __ptlrpc_free_req(request, locked);
1080                 RETURN(1);
1081         }
1082
1083         RETURN(0);
1084 }
1085
1086 void ptlrpc_req_finished(struct ptlrpc_request *request)
1087 {
1088         __ptlrpc_req_finished(request, 0);
1089 }
1090
1091 /* Disengage the client's reply buffer from the network
1092  * NB does _NOT_ unregister any client-side bulk.
1093  * IDEMPOTENT, but _not_ safe against concurrent callers.
1094  * The request owner (i.e. the thread doing the I/O) must call...
1095  */
1096 void ptlrpc_unregister_reply (struct ptlrpc_request *request)
1097 {
1098         int                rc;
1099         wait_queue_head_t *wq;
1100         struct l_wait_info lwi;
1101
1102         LASSERT(!in_interrupt ());             /* might sleep */
1103
1104         if (!ptlrpc_client_receiving_reply(request))
1105                 return;
1106
1107         PtlMDUnlink (request->rq_reply_md_h);
1108
1109         /* We have to l_wait_event() whatever the result, to give liblustre
1110          * a chance to run reply_in_callback() */
1111
1112         if (request->rq_set == NULL)
1113                 wq = &request->rq_set->set_waitq;
1114         else
1115                 wq = &request->rq_reply_waitq;
1116
1117         for (;;) {
1118                 /* Network access will complete in finite time but the HUGE
1119                  * timeout lets us CWARN for visibility of sluggish NALs */
1120                 lwi = LWI_TIMEOUT(300 * HZ, NULL, NULL);
1121                 rc = l_wait_event (*wq, !ptlrpc_client_receiving_reply(request), &lwi);
1122                 if (rc == 0)
1123                         return;
1124
1125                 LASSERT (rc == -ETIMEDOUT);
1126                 DEBUG_REQ(D_WARNING, request, "Unexpectedly long timeout");
1127         }
1128 }
1129
1130 /* caller must hold imp->imp_lock */
1131 void ptlrpc_free_committed(struct obd_import *imp)
1132 {
1133         struct list_head *tmp, *saved;
1134         struct ptlrpc_request *req;
1135         struct ptlrpc_request *last_req = NULL; /* temporary fire escape */
1136         ENTRY;
1137
1138         LASSERT(imp != NULL);
1139
1140         LASSERT_SPIN_LOCKED(&imp->imp_lock);
1141
1142         CDEBUG(D_HA, "%s: committing for last_committed "LPU64"\n",
1143                imp->imp_obd->obd_name, imp->imp_peer_committed_transno);
1144
1145         list_for_each_safe(tmp, saved, &imp->imp_replay_list) {
1146                 req = list_entry(tmp, struct ptlrpc_request, rq_replay_list);
1147
1148                 /* XXX ok to remove when 1357 resolved - rread 05/29/03  */
1149                 LASSERT(req != last_req);
1150                 last_req = req;
1151
1152                 if (req->rq_import_generation < imp->imp_generation) {
1153                         DEBUG_REQ(D_HA, req, "freeing request with old gen");
1154                         GOTO(free_req, 0);
1155                 }
1156
1157                 if (req->rq_replay) {
1158                         DEBUG_REQ(D_HA, req, "keeping (FL_REPLAY)");
1159                         continue;
1160                 }
1161
1162                 /* not yet committed */
1163                 if (req->rq_transno > imp->imp_peer_committed_transno) {
1164                         DEBUG_REQ(D_HA, req, "stopping search");
1165                         break;
1166                 }
1167
1168                 DEBUG_REQ(D_HA, req, "committing (last_committed "LPU64")",
1169                           imp->imp_peer_committed_transno);
1170 free_req:
1171                 if (req->rq_commit_cb != NULL)
1172                         req->rq_commit_cb(req);
1173                 list_del_init(&req->rq_replay_list);
1174                 __ptlrpc_req_finished(req, 1);
1175         }
1176
1177         EXIT;
1178         return;
1179 }
1180
1181 void ptlrpc_cleanup_client(struct obd_import *imp)
1182 {
1183         ENTRY;
1184         EXIT;
1185         return;
1186 }
1187
1188 void ptlrpc_resend_req(struct ptlrpc_request *req)
1189 {
1190         unsigned long flags;
1191
1192         DEBUG_REQ(D_HA, req, "going to resend");
1193         req->rq_reqmsg->handle.cookie = 0;
1194         req->rq_status = -EAGAIN;
1195
1196         spin_lock_irqsave (&req->rq_lock, flags);
1197         req->rq_resend = 1;
1198         req->rq_timedout = 0;
1199         if (req->rq_bulk) {
1200                 __u64 old_xid = req->rq_xid;
1201
1202                 /* ensure previous bulk fails */
1203                 req->rq_xid = ptlrpc_next_xid();
1204                 CDEBUG(D_HA, "resend bulk old x"LPU64" new x"LPU64"\n",
1205                        old_xid, req->rq_xid);
1206         }
1207         ptlrpc_wake_client_req(req);
1208         spin_unlock_irqrestore (&req->rq_lock, flags);
1209 }
1210
1211 /* XXX: this function and rq_status are currently unused */
1212 void ptlrpc_restart_req(struct ptlrpc_request *req)
1213 {
1214         unsigned long flags;
1215
1216         DEBUG_REQ(D_HA, req, "restarting (possibly-)completed request");
1217         req->rq_status = -ERESTARTSYS;
1218
1219         spin_lock_irqsave (&req->rq_lock, flags);
1220         req->rq_restart = 1;
1221         req->rq_timedout = 0;
1222         ptlrpc_wake_client_req(req);
1223         spin_unlock_irqrestore (&req->rq_lock, flags);
1224 }
1225
1226 static int expired_request(void *data)
1227 {
1228         struct ptlrpc_request *req = data;
1229         ENTRY;
1230
1231         RETURN(ptlrpc_expire_one_request(req));
1232 }
1233
1234 static void interrupted_request(void *data)
1235 {
1236         unsigned long flags;
1237
1238         struct ptlrpc_request *req = data;
1239         DEBUG_REQ(D_HA, req, "request interrupted");
1240         spin_lock_irqsave (&req->rq_lock, flags);
1241         req->rq_intr = 1;
1242         spin_unlock_irqrestore (&req->rq_lock, flags);
1243 }
1244
1245 struct ptlrpc_request *ptlrpc_request_addref(struct ptlrpc_request *req)
1246 {
1247         ENTRY;
1248         atomic_inc(&req->rq_refcount);
1249         RETURN(req);
1250 }
1251
1252 void ptlrpc_retain_replayable_request(struct ptlrpc_request *req,
1253                                       struct obd_import *imp)
1254 {
1255         struct list_head *tmp;
1256
1257         LASSERT_SPIN_LOCKED(&imp->imp_lock);
1258
1259         /* don't re-add requests that have been replayed */
1260         if (!list_empty(&req->rq_replay_list))
1261                 return;
1262
1263         lustre_msg_add_flags(req->rq_reqmsg,
1264                              MSG_REPLAY);
1265
1266         LASSERT(imp->imp_replayable);
1267         /* Balanced in ptlrpc_free_committed, usually. */
1268         ptlrpc_request_addref(req);
1269         list_for_each_prev(tmp, &imp->imp_replay_list) {
1270                 struct ptlrpc_request *iter =
1271                         list_entry(tmp, struct ptlrpc_request, rq_replay_list);
1272
1273                 /* We may have duplicate transnos if we create and then
1274                  * open a file, or for closes retained if to match creating
1275                  * opens, so use req->rq_xid as a secondary key.
1276                  * (See bugs 684, 685, and 428.)
1277                  * XXX no longer needed, but all opens need transnos!
1278                  */
1279                 if (iter->rq_transno > req->rq_transno)
1280                         continue;
1281
1282                 if (iter->rq_transno == req->rq_transno) {
1283                         LASSERT(iter->rq_xid != req->rq_xid);
1284                         if (iter->rq_xid > req->rq_xid)
1285                                 continue;
1286                 }
1287
1288                 list_add(&req->rq_replay_list, &iter->rq_replay_list);
1289                 return;
1290         }
1291
1292         list_add_tail(&req->rq_replay_list, &imp->imp_replay_list);
1293 }
1294
1295 int ptlrpc_queue_wait(struct ptlrpc_request *req)
1296 {
1297         char str[PTL_NALFMT_SIZE];
1298         int rc = 0;
1299         int brc;
1300         struct l_wait_info lwi;
1301         struct obd_import *imp = req->rq_import;
1302         unsigned long flags;
1303         int timeout = 0;
1304         ENTRY;
1305
1306         LASSERT(req->rq_set == NULL);
1307         LASSERT(!req->rq_receiving_reply);
1308         atomic_inc(&imp->imp_inflight);
1309
1310         /* for distributed debugging */
1311         req->rq_reqmsg->status = current->pid;
1312         LASSERT(imp->imp_obd != NULL);
1313         CDEBUG(D_RPCTRACE, "Sending RPC pname:cluuid:pid:xid:ni:nid:opc "
1314                "%s:%s:%d:"LPU64":%s:%s:%d\n", current->comm,
1315                imp->imp_obd->obd_uuid.uuid,
1316                req->rq_reqmsg->status, req->rq_xid,
1317                imp->imp_connection->c_peer.peer_ni->pni_name,
1318                ptlrpc_peernid2str(&imp->imp_connection->c_peer, str),
1319                req->rq_reqmsg->opc);
1320
1321         /* Mark phase here for a little debug help */
1322         req->rq_phase = RQ_PHASE_RPC;
1323
1324         spin_lock_irqsave(&imp->imp_lock, flags);
1325         req->rq_import_generation = imp->imp_generation;
1326 restart:
1327         if (ptlrpc_import_delay_req(imp, req, &rc)) {
1328                 list_del(&req->rq_list);
1329
1330                 list_add_tail(&req->rq_list, &imp->imp_delayed_list);
1331                 spin_unlock_irqrestore(&imp->imp_lock, flags);
1332
1333                 DEBUG_REQ(D_HA, req, "\"%s\" waiting for recovery: (%s != %s)",
1334                           current->comm,
1335                           ptlrpc_import_state_name(req->rq_send_state),
1336                           ptlrpc_import_state_name(imp->imp_state));
1337                 lwi = LWI_INTR(interrupted_request, req);
1338                 rc = l_wait_event(req->rq_reply_waitq,
1339                                   (req->rq_send_state == imp->imp_state ||
1340                                    req->rq_err),
1341                                   &lwi);
1342                 DEBUG_REQ(D_HA, req, "\"%s\" awake: (%s == %s or %d == 1)",
1343                           current->comm,
1344                           ptlrpc_import_state_name(imp->imp_state),
1345                           ptlrpc_import_state_name(req->rq_send_state),
1346                           req->rq_err);
1347
1348                 spin_lock_irqsave(&imp->imp_lock, flags);
1349                 list_del_init(&req->rq_list);
1350
1351                 if (req->rq_err) {
1352                         rc = -EIO;
1353                 }
1354                 else if (req->rq_intr) {
1355                         rc = -EINTR;
1356                 }
1357                 else if (req->rq_no_resend) {
1358                         spin_unlock_irqrestore(&imp->imp_lock, flags);
1359                         GOTO(out, rc = -ETIMEDOUT);
1360                 }
1361                 else {
1362                         GOTO(restart, rc);
1363                 }
1364         }
1365
1366         if (rc != 0) {
1367                 list_del_init(&req->rq_list);
1368                 spin_unlock_irqrestore(&imp->imp_lock, flags);
1369                 req->rq_status = rc; // XXX this ok?
1370                 GOTO(out, rc);
1371         }
1372
1373         if (req->rq_resend) {
1374                 lustre_msg_add_flags(req->rq_reqmsg, MSG_RESENT);
1375
1376                 if (req->rq_bulk != NULL)
1377                         ptlrpc_unregister_bulk (req);
1378
1379                 DEBUG_REQ(D_HA, req, "resending: ");
1380         }
1381
1382         /* XXX this is the same as ptlrpc_set_wait */
1383         LASSERT(list_empty(&req->rq_list));
1384         list_add_tail(&req->rq_list, &imp->imp_sending_list);
1385         spin_unlock_irqrestore(&imp->imp_lock, flags);
1386
1387         rc = ptl_send_rpc(req);
1388         if (rc) {
1389                 DEBUG_REQ(D_HA, req, "send failed (%d); recovering", rc);
1390                 timeout = 1;
1391         } else {
1392                 timeout = MAX(req->rq_timeout * HZ, 1);
1393                 DEBUG_REQ(D_NET, req, "-- sleeping");
1394         }
1395         lwi = LWI_TIMEOUT_INTR(timeout, expired_request, interrupted_request,
1396                                req);
1397         l_wait_event(req->rq_reply_waitq, ptlrpc_check_reply(req), &lwi);
1398         DEBUG_REQ(D_NET, req, "-- done sleeping");
1399
1400         CDEBUG(D_RPCTRACE, "Completed RPC pname:cluuid:pid:xid:ni:nid:opc "
1401                "%s:%s:%d:"LPU64":%s:%s:%d\n", current->comm,
1402                imp->imp_obd->obd_uuid.uuid,
1403                req->rq_reqmsg->status, req->rq_xid,
1404                imp->imp_connection->c_peer.peer_ni->pni_name,
1405                ptlrpc_peernid2str(&imp->imp_connection->c_peer, str),
1406                req->rq_reqmsg->opc);
1407
1408         spin_lock_irqsave(&imp->imp_lock, flags);
1409         list_del_init(&req->rq_list);
1410         spin_unlock_irqrestore(&imp->imp_lock, flags);
1411
1412         /* If the reply was received normally, this just grabs the spinlock
1413          * (ensuring the reply callback has returned), sees that
1414          * req->rq_receiving_reply is clear and returns. */
1415         ptlrpc_unregister_reply (req);
1416
1417         if (req->rq_err)
1418                 GOTO(out, rc = -EIO);
1419
1420         /* Resend if we need to, unless we were interrupted. */
1421         if (req->rq_resend && !req->rq_intr) {
1422                 /* ...unless we were specifically told otherwise. */
1423                 if (req->rq_no_resend)
1424                         GOTO(out, rc = -ETIMEDOUT);
1425                 spin_lock_irqsave(&imp->imp_lock, flags);
1426                 goto restart;
1427         }
1428
1429         if (req->rq_intr) {
1430                 /* Should only be interrupted if we timed out. */
1431                 if (!req->rq_timedout)
1432                         DEBUG_REQ(D_ERROR, req,
1433                                   "rq_intr set but rq_timedout not");
1434                 GOTO(out, rc = -EINTR);
1435         }
1436
1437         if (req->rq_timedout) {                 /* non-recoverable timeout */
1438                 GOTO(out, rc = -ETIMEDOUT);
1439         }
1440
1441         if (!req->rq_replied) {
1442                 /* How can this be? -eeb */
1443                 DEBUG_REQ(D_ERROR, req, "!rq_replied: ");
1444                 LBUG();
1445                 GOTO(out, rc = req->rq_status);
1446         }
1447
1448         rc = after_reply (req);
1449         /* NB may return +ve success rc */
1450         if (req->rq_resend) {
1451                 spin_lock_irqsave(&imp->imp_lock, flags);
1452                 goto restart;
1453         }
1454
1455  out:
1456         if (req->rq_bulk != NULL) {
1457                 if (rc >= 0) {
1458                         /* success so far.  Note that anything going wrong
1459                          * with bulk now, is EXTREMELY strange, since the
1460                          * server must have believed that the bulk
1461                          * tranferred OK before she replied with success to
1462                          * me. */
1463                         lwi = LWI_TIMEOUT(timeout, NULL, NULL);
1464                         brc = l_wait_event(req->rq_reply_waitq,
1465                                            !ptlrpc_bulk_active(req->rq_bulk),
1466                                            &lwi);
1467                         LASSERT(brc == 0 || brc == -ETIMEDOUT);
1468                         if (brc != 0) {
1469                                 LASSERT(brc == -ETIMEDOUT);
1470                                 DEBUG_REQ(D_ERROR, req, "bulk timed out");
1471                                 rc = brc;
1472                         } else if (!req->rq_bulk->bd_success) {
1473                                 DEBUG_REQ(D_ERROR, req, "bulk transfer failed");
1474                                 rc = -EIO;
1475                         }
1476                 }
1477                 if (rc < 0)
1478                         ptlrpc_unregister_bulk (req);
1479         }
1480
1481         LASSERT(!req->rq_receiving_reply);
1482         req->rq_phase = RQ_PHASE_INTERPRET;
1483
1484         atomic_dec(&imp->imp_inflight);
1485         wake_up(&imp->imp_recovery_waitq);
1486         RETURN(rc);
1487 }
1488
1489 struct ptlrpc_replay_async_args {
1490         int praa_old_state;
1491         int praa_old_status;
1492 };
1493
1494 static int ptlrpc_replay_interpret(struct ptlrpc_request *req,
1495                                     void * data, int rc)
1496 {
1497         struct ptlrpc_replay_async_args *aa = data;
1498         struct obd_import *imp = req->rq_import;
1499         unsigned long flags;
1500
1501         atomic_dec(&imp->imp_replay_inflight);
1502
1503         if (!req->rq_replied) {
1504                 CERROR("request replay timed out, restarting recovery\n");
1505                 GOTO(out, rc = -ETIMEDOUT);
1506         }
1507
1508 #if SWAB_PARANOIA
1509         /* Clear reply swab mask; this is a new reply in sender's byte order */
1510         req->rq_rep_swab_mask = 0;
1511 #endif
1512         LASSERT (req->rq_nob_received <= req->rq_replen);
1513         rc = lustre_unpack_msg(req->rq_repmsg, req->rq_nob_received);
1514         if (rc) {
1515                 CERROR("unpack_rep failed: %d\n", rc);
1516                 GOTO(out, rc = -EPROTO);
1517         }
1518
1519         if (req->rq_repmsg->type == PTL_RPC_MSG_ERR &&
1520             req->rq_repmsg->status == -ENOTCONN)
1521                 GOTO(out, rc = req->rq_repmsg->status);
1522
1523         /* The transno had better not change over replay. */
1524         LASSERT(req->rq_reqmsg->transno == req->rq_repmsg->transno);
1525
1526         DEBUG_REQ(D_HA, req, "got rep");
1527
1528         /* let the callback do fixups, possibly including in the request */
1529         if (req->rq_replay_cb)
1530                 req->rq_replay_cb(req);
1531
1532         if (req->rq_replied && req->rq_repmsg->status != aa->praa_old_status) {
1533                 DEBUG_REQ(D_ERROR, req, "status %d, old was %d",
1534                           req->rq_repmsg->status, aa->praa_old_status);
1535         } else {
1536                 /* Put it back for re-replay. */
1537                 req->rq_repmsg->status = aa->praa_old_status;
1538         }
1539
1540         spin_lock_irqsave(&imp->imp_lock, flags);
1541         imp->imp_last_replay_transno = req->rq_transno;
1542         spin_unlock_irqrestore(&imp->imp_lock, flags);
1543
1544         /* continue with recovery */
1545         rc = ptlrpc_import_recovery_state_machine(imp);
1546  out:
1547         req->rq_send_state = aa->praa_old_state;
1548
1549         if (rc != 0)
1550                 /* this replay failed, so restart recovery */
1551                 ptlrpc_connect_import(imp, NULL);
1552
1553         RETURN(rc);
1554 }
1555
1556
1557 int ptlrpc_replay_req(struct ptlrpc_request *req)
1558 {
1559         struct ptlrpc_replay_async_args *aa;
1560         ENTRY;
1561
1562         LASSERT(req->rq_import->imp_state == LUSTRE_IMP_REPLAY);
1563
1564         /* Not handling automatic bulk replay yet (or ever?) */
1565         LASSERT(req->rq_bulk == NULL);
1566
1567         DEBUG_REQ(D_HA, req, "REPLAY");
1568
1569         LASSERT (sizeof (*aa) <= sizeof (req->rq_async_args));
1570         aa = (struct ptlrpc_replay_async_args *)&req->rq_async_args;
1571         memset(aa, 0, sizeof *aa);
1572
1573         /* Prepare request to be resent with ptlrpcd */
1574         aa->praa_old_state = req->rq_send_state;
1575         req->rq_send_state = LUSTRE_IMP_REPLAY;
1576         req->rq_phase = RQ_PHASE_NEW;
1577         /*
1578          * Q: "How can a req get on the replay list if it wasn't replied?"
1579          * A: "If we failed during the replay of this request, it will still
1580          *     be on the list, but rq_replied will have been reset to 0."
1581          */
1582         if (req->rq_replied) {
1583                 aa->praa_old_status = req->rq_repmsg->status;
1584                 req->rq_status = 0;
1585                 req->rq_replied = 0;
1586         }
1587
1588         req->rq_interpret_reply = ptlrpc_replay_interpret;
1589         atomic_inc(&req->rq_import->imp_replay_inflight);
1590         ptlrpc_request_addref(req); /* ptlrpcd needs a ref */
1591
1592         ptlrpcd_add_req(req);
1593         RETURN(0);
1594 }
1595
1596 void ptlrpc_abort_inflight(struct obd_import *imp)
1597 {
1598         unsigned long flags;
1599         struct list_head *tmp, *n;
1600         ENTRY;
1601
1602         /* Make sure that no new requests get processed for this import.
1603          * ptlrpc_{queue,set}_wait must (and does) hold imp_lock while testing
1604          * this flag and then putting requests on sending_list or delayed_list.
1605          */
1606         spin_lock_irqsave(&imp->imp_lock, flags);
1607
1608         /* XXX locking?  Maybe we should remove each request with the list
1609          * locked?  Also, how do we know if the requests on the list are
1610          * being freed at this time?
1611          */
1612         list_for_each_safe(tmp, n, &imp->imp_sending_list) {
1613                 struct ptlrpc_request *req =
1614                         list_entry(tmp, struct ptlrpc_request, rq_list);
1615
1616                 DEBUG_REQ(D_HA, req, "inflight");
1617
1618                 spin_lock (&req->rq_lock);
1619                 if (req->rq_import_generation < imp->imp_generation) {
1620                         req->rq_err = 1;
1621                         ptlrpc_wake_client_req(req);
1622                 }
1623                 spin_unlock (&req->rq_lock);
1624         }
1625
1626         list_for_each_safe(tmp, n, &imp->imp_delayed_list) {
1627                 struct ptlrpc_request *req =
1628                         list_entry(tmp, struct ptlrpc_request, rq_list);
1629
1630                 DEBUG_REQ(D_HA, req, "aborting waiting req");
1631
1632                 spin_lock (&req->rq_lock);
1633                 if (req->rq_import_generation < imp->imp_generation) {
1634                         req->rq_err = 1;
1635                         ptlrpc_wake_client_req(req);
1636                 }
1637                 spin_unlock (&req->rq_lock);
1638         }
1639
1640         /* Last chance to free reqs left on the replay list, but we
1641          * will still leak reqs that haven't comitted.  */
1642         if (imp->imp_replayable)
1643                 ptlrpc_free_committed(imp);
1644
1645         spin_unlock_irqrestore(&imp->imp_lock, flags);
1646
1647         EXIT;
1648 }
1649
1650 static __u64 ptlrpc_last_xid = 0;
1651 static spinlock_t ptlrpc_last_xid_lock = SPIN_LOCK_UNLOCKED;
1652
1653 __u64 ptlrpc_next_xid(void)
1654 {
1655         __u64 tmp;
1656         spin_lock(&ptlrpc_last_xid_lock);
1657         tmp = ++ptlrpc_last_xid;
1658         spin_unlock(&ptlrpc_last_xid_lock);
1659         return tmp;
1660 }
1661
1662