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 the Lustre file system, http://www.lustre.org
   7  *   Lustre is a trademark of Cluster File Systems, Inc.
   8  *
   9  *   You may have signed or agreed to another license before downloading
  10  *   this software.  If so, you are bound by the terms and conditions
  11  *   of that agreement, and the following does not apply to you.  See the
  12  *   LICENSE file included with this distribution for more information.
  13  *
  14  *   If you did not agree to a different license, then this copy of Lustre
  15  *   is open source software; you can redistribute it and/or modify it
  16  *   under the terms of version 2 of the GNU General Public License as
  17  *   published by the Free Software Foundation.
  18  *
  19  *   In either case, Lustre is distributed in the hope that it will be
  20  *   useful, but WITHOUT ANY WARRANTY; without even the implied warranty
  21  *   of MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
  22  *   license text for more details.
  23  *
  24  */
  25
  26 #define DEBUG_SUBSYSTEM S_RPC
  27 #ifndef __KERNEL__
  28 #include <errno.h>
  29 #include <signal.h>
  30 #include <liblustre.h>
  31 #endif
  32
  33 #include <obd_support.h>
  34 #include <obd_class.h>
  35 #include <lustre_lib.h>
  36 #include <lustre_ha.h>
  37 #include <lustre_import.h>
  38
  39 #include "ptlrpc_internal.h"
  40
  41 void ptlrpc_init_client(int req_portal, int rep_portal, char *name,
  42                         struct ptlrpc_client *cl)
  43 {
  44         cl->cli_request_portal = req_portal;
  45         cl->cli_reply_portal   = rep_portal;
  46         cl->cli_name           = name;
  47 }
  48
  49 struct ptlrpc_connection *ptlrpc_uuid_to_connection(struct obd_uuid *uuid)
  50 {
  51         struct ptlrpc_connection *c;
  52         lnet_nid_t                self;
  53         lnet_process_id_t         peer;
  54         int                       err;
  55
  56         err = ptlrpc_uuid_to_peer(uuid, &peer, &self);
  57         if (err != 0) {
  58                 CERROR("cannot find peer %s!\n", uuid->uuid);
  59                 return NULL;
  60         }
  61
  62         c = ptlrpc_get_connection(peer, self, uuid);
  63         if (c) {
  64                 memcpy(c->c_remote_uuid.uuid,
  65                        uuid->uuid, sizeof(c->c_remote_uuid.uuid));
  66         }
  67
  68         CDEBUG(D_INFO, "%s -> %p\n", uuid->uuid, c);
  69
  70         return c;
  71 }
  72
  73 void ptlrpc_readdress_connection(struct ptlrpc_connection *conn,
  74                                  struct obd_uuid *uuid)
  75 {
  76         lnet_nid_t        self;
  77         lnet_process_id_t peer;
  78         int               err;
  79
  80         err = ptlrpc_uuid_to_peer(uuid, &peer, &self);
  81         if (err != 0) {
  82                 CERROR("cannot find peer %s!\n", uuid->uuid);
  83                 return;
  84         }
  85
  86         conn->c_peer = peer;
  87         conn->c_self = self;
  88         return;
  89 }
  90
  91 static inline struct ptlrpc_bulk_desc *new_bulk(int npages, int type, int portal)
  92 {
  93         struct ptlrpc_bulk_desc *desc;
  94
  95         OBD_ALLOC(desc, offsetof (struct ptlrpc_bulk_desc, bd_iov[npages]));
  96         if (!desc)
  97                 return NULL;
  98
  99         spin_lock_init(&desc->bd_lock);
 100         cfs_waitq_init(&desc->bd_waitq);
 101         desc->bd_max_iov = npages;
 102         desc->bd_iov_count = 0;
 103         desc->bd_md_h = LNET_INVALID_HANDLE;
 104         desc->bd_portal = portal;
 105         desc->bd_type = type;
 106
 107         return desc;
 108 }
 109
 110 struct ptlrpc_bulk_desc *ptlrpc_prep_bulk_imp (struct ptlrpc_request *req,
 111                                                int npages, int type, int portal)
 112 {
 113         struct obd_import *imp = req->rq_import;
 114         struct ptlrpc_bulk_desc *desc;
 115
 116         ENTRY;
 117         LASSERT(type == BULK_PUT_SINK || type == BULK_GET_SOURCE);
 118         desc = new_bulk(npages, type, portal);
 119         if (desc == NULL)
 120                 RETURN(NULL);
 121
 122         desc->bd_import_generation = req->rq_import_generation;
 123         desc->bd_import = class_import_get(imp);
 124         desc->bd_req = req;
 125
 126         desc->bd_cbid.cbid_fn  = client_bulk_callback;
 127         desc->bd_cbid.cbid_arg = desc;
 128
 129         /* This makes req own desc, and free it when she frees herself */
 130         req->rq_bulk = desc;
 131
 132         return desc;
 133 }
 134
 135 struct ptlrpc_bulk_desc *ptlrpc_prep_bulk_exp (struct ptlrpc_request *req,
 136                                                int npages, int type, int portal)
 137 {
 138         struct obd_export *exp = req->rq_export;
 139         struct ptlrpc_bulk_desc *desc;
 140
 141         ENTRY;
 142         LASSERT(type == BULK_PUT_SOURCE || type == BULK_GET_SINK);
 143
 144         desc = new_bulk(npages, type, portal);
 145         if (desc == NULL)
 146                 RETURN(NULL);
 147
 148         desc->bd_export = class_export_get(exp);
 149         desc->bd_req = req;
 150
 151         desc->bd_cbid.cbid_fn  = server_bulk_callback;
 152         desc->bd_cbid.cbid_arg = desc;
 153
 154         /* NB we don't assign rq_bulk here; server-side requests are
 155          * re-used, and the handler frees the bulk desc explicitly. */
 156
 157         return desc;
 158 }
 159
 160 void ptlrpc_prep_bulk_page(struct ptlrpc_bulk_desc *desc,
 161                            cfs_page_t *page, int pageoffset, int len)
 162 {
 163         LASSERT(desc->bd_iov_count < desc->bd_max_iov);
 164         LASSERT(page != NULL);
 165         LASSERT(pageoffset >= 0);
 166         LASSERT(len > 0);
 167         LASSERT(pageoffset + len <= CFS_PAGE_SIZE);
 168
 169         desc->bd_nob += len;
 170
 171         ptlrpc_add_bulk_page(desc, page, pageoffset, len);
 172 }
 173
 174 void ptlrpc_free_bulk(struct ptlrpc_bulk_desc *desc)
 175 {
 176         ENTRY;
 177
 178         LASSERT(desc != NULL);
 179         LASSERT(desc->bd_iov_count != LI_POISON); /* not freed already */
 180         LASSERT(!desc->bd_network_rw);         /* network hands off or */
 181         LASSERT((desc->bd_export != NULL) ^ (desc->bd_import != NULL));
 182
 183         sptlrpc_enc_pool_put_pages(desc);
 184
 185         if (desc->bd_export)
 186                 class_export_put(desc->bd_export);
 187         else
 188                 class_import_put(desc->bd_import);
 189
 190         OBD_FREE(desc, offsetof(struct ptlrpc_bulk_desc,
 191                                 bd_iov[desc->bd_max_iov]));
 192         EXIT;
 193 }
 194
 195 void ptlrpc_free_rq_pool(struct ptlrpc_request_pool *pool)
 196 {
 197         struct list_head *l, *tmp;
 198         struct ptlrpc_request *req;
 199
 200         if (!pool)
 201                 return;
 202
 203         list_for_each_safe(l, tmp, &pool->prp_req_list) {
 204                 req = list_entry(l, struct ptlrpc_request, rq_list);
 205                 list_del(&req->rq_list);
 206                 LASSERT(req->rq_reqbuf);
 207                 LASSERT(req->rq_reqbuf_len == pool->prp_rq_size);
 208                 OBD_FREE(req->rq_reqbuf, pool->prp_rq_size);
 209                 OBD_FREE(req, sizeof(*req));
 210         }
 211         OBD_FREE(pool, sizeof(*pool));
 212 }
 213
 214 void ptlrpc_add_rqs_to_pool(struct ptlrpc_request_pool *pool, int num_rq)
 215 {
 216         int i;
 217         int size = 1;
 218
 219         while (size < pool->prp_rq_size + SPTLRPC_MAX_PAYLOAD)
 220                 size <<= 1;
 221
 222         LASSERTF(list_empty(&pool->prp_req_list) || size == pool->prp_rq_size,
 223                  "Trying to change pool size with nonempty pool "
 224                  "from %d to %d bytes\n", pool->prp_rq_size, size);
 225
 226         spin_lock(&pool->prp_lock);
 227         pool->prp_rq_size = size;
 228         for (i = 0; i < num_rq; i++) {
 229                 struct ptlrpc_request *req;
 230                 struct lustre_msg *msg;
 231
 232                 spin_unlock(&pool->prp_lock);
 233                 OBD_ALLOC(req, sizeof(struct ptlrpc_request));
 234                 if (!req)
 235                         return;
 236                 OBD_ALLOC_GFP(msg, size, CFS_ALLOC_STD);
 237                 if (!msg) {
 238                         OBD_FREE(req, sizeof(struct ptlrpc_request));
 239                         return;
 240                 }
 241                 req->rq_reqbuf = msg;
 242                 req->rq_reqbuf_len = size;
 243                 req->rq_pool = pool;
 244                 spin_lock(&pool->prp_lock);
 245                 list_add_tail(&req->rq_list, &pool->prp_req_list);
 246         }
 247         spin_unlock(&pool->prp_lock);
 248         return;
 249 }
 250
 251 struct ptlrpc_request_pool *ptlrpc_init_rq_pool(int num_rq, int msgsize,
 252                                                 void (*populate_pool)(struct ptlrpc_request_pool *, int))
 253 {
 254         struct ptlrpc_request_pool *pool;
 255
 256         OBD_ALLOC(pool, sizeof (struct ptlrpc_request_pool));
 257         if (!pool)
 258                 return NULL;
 259
 260         /* Request next power of two for the allocation, because internally
 261            kernel would do exactly this */
 262
 263         spin_lock_init(&pool->prp_lock);
 264         CFS_INIT_LIST_HEAD(&pool->prp_req_list);
 265         pool->prp_rq_size = msgsize;
 266         pool->prp_populate = populate_pool;
 267
 268         populate_pool(pool, num_rq);
 269
 270         if (list_empty(&pool->prp_req_list)) {
 271                 /* have not allocated a single request for the pool */
 272                 OBD_FREE(pool, sizeof (struct ptlrpc_request_pool));
 273                 pool = NULL;
 274         }
 275         return pool;
 276 }
 277
 278 static struct ptlrpc_request *ptlrpc_prep_req_from_pool(struct ptlrpc_request_pool *pool)
 279 {
 280         struct ptlrpc_request *request;
 281         struct lustre_msg *reqbuf;
 282
 283         if (!pool)
 284                 return NULL;
 285
 286         spin_lock(&pool->prp_lock);
 287
 288         /* See if we have anything in a pool, and bail out if nothing,
 289          * in writeout path, where this matters, this is safe to do, because
 290          * nothing is lost in this case, and when some in-flight requests
 291          * complete, this code will be called again. */
 292         if (unlikely(list_empty(&pool->prp_req_list))) {
 293                 spin_unlock(&pool->prp_lock);
 294                 return NULL;
 295         }
 296
 297         request = list_entry(pool->prp_req_list.next, struct ptlrpc_request,
 298                              rq_list);
 299         list_del(&request->rq_list);
 300         spin_unlock(&pool->prp_lock);
 301
 302         LASSERT(request->rq_reqbuf);
 303         LASSERT(request->rq_pool);
 304
 305         reqbuf = request->rq_reqbuf;
 306         memset(request, 0, sizeof(*request));
 307         request->rq_reqbuf = reqbuf;
 308         request->rq_reqbuf_len = pool->prp_rq_size;
 309         request->rq_pool = pool;
 310         return request;
 311 }
 312
 313 static void __ptlrpc_free_req_to_pool(struct ptlrpc_request *request)
 314 {
 315         struct ptlrpc_request_pool *pool = request->rq_pool;
 316
 317         spin_lock(&pool->prp_lock);
 318         LASSERT(list_empty(&request->rq_list));
 319         list_add_tail(&request->rq_list, &pool->prp_req_list);
 320         spin_unlock(&pool->prp_lock);
 321 }
 322
 323 struct ptlrpc_request *
 324 ptlrpc_prep_req_pool(struct obd_import *imp, __u32 version, int opcode,
 325                      int count, int *lengths, char **bufs,
 326                      struct ptlrpc_request_pool *pool,
 327                      struct ptlrpc_cli_ctx *ctx)
 328 {
 329         struct ptlrpc_request *request = NULL;
 330         int rc;
 331         ENTRY;
 332
 333         /* The obd disconnected */
 334         if (imp == NULL)
 335                 return NULL;
 336
 337         LASSERT(imp != LP_POISON);
 338         LASSERT((unsigned long)imp->imp_client > 0x1000);
 339         LASSERT(imp->imp_client != LP_POISON);
 340
 341         if (pool)
 342                 request = ptlrpc_prep_req_from_pool(pool);
 343
 344         if (!request)
 345                 OBD_ALLOC(request, sizeof(*request));
 346
 347         if (!request) {
 348                 CERROR("request allocation out of memory\n");
 349                 RETURN(NULL);
 350         }
 351
 352         request->rq_import = class_import_get(imp);
 353
 354         if (unlikely(ctx))
 355                 request->rq_cli_ctx = sptlrpc_cli_ctx_get(ctx);
 356         else {
 357                 rc = sptlrpc_req_get_ctx(request);
 358                 if (rc)
 359                         GOTO(out_free, rc);
 360         }
 361
 362         sptlrpc_req_set_flavor(request, opcode);
 363
 364         rc = lustre_pack_request(request, imp->imp_msg_magic, count, lengths,
 365                                  bufs);
 366         if (rc) {
 367                 LASSERT(!request->rq_pool);
 368                 GOTO(out_ctx, rc);
 369         }
 370
 371         lustre_msg_add_version(request->rq_reqmsg, version);
 372
 373         if (imp->imp_server_timeout)
 374                 request->rq_timeout = obd_timeout / 2;
 375         else
 376                 request->rq_timeout = obd_timeout;
 377         request->rq_send_state = LUSTRE_IMP_FULL;
 378         request->rq_type = PTL_RPC_MSG_REQUEST;
 379         request->rq_export = NULL;
 380
 381         request->rq_req_cbid.cbid_fn  = request_out_callback;
 382         request->rq_req_cbid.cbid_arg = request;
 383
 384         request->rq_reply_cbid.cbid_fn  = reply_in_callback;
 385         request->rq_reply_cbid.cbid_arg = request;
 386
 387         request->rq_phase = RQ_PHASE_NEW;
 388
 389         /* XXX FIXME bug 249 */
 390         request->rq_request_portal = imp->imp_client->cli_request_portal;
 391         request->rq_reply_portal = imp->imp_client->cli_reply_portal;
 392
 393         spin_lock_init(&request->rq_lock);
 394         CFS_INIT_LIST_HEAD(&request->rq_list);
 395         CFS_INIT_LIST_HEAD(&request->rq_replay_list);
 396         CFS_INIT_LIST_HEAD(&request->rq_mod_list);
 397         CFS_INIT_LIST_HEAD(&request->rq_ctx_chain);
 398         CFS_INIT_LIST_HEAD(&request->rq_set_chain);
 399         cfs_waitq_init(&request->rq_reply_waitq);
 400         request->rq_xid = ptlrpc_next_xid();
 401         atomic_set(&request->rq_refcount, 1);
 402
 403         lustre_msg_set_opc(request->rq_reqmsg, opcode);
 404         lustre_msg_set_flags(request->rq_reqmsg, 0);
 405
 406         RETURN(request);
 407 out_ctx:
 408         sptlrpc_cli_ctx_put(request->rq_cli_ctx, 1);
 409 out_free:
 410         class_import_put(imp);
 411         if (request->rq_pool)
 412                 __ptlrpc_free_req_to_pool(request);
 413         else
 414                 OBD_FREE(request, sizeof(*request));
 415         return NULL;
 416 }
 417
 418 struct ptlrpc_request *
 419 ptlrpc_prep_req(struct obd_import *imp, __u32 version, int opcode, int count,
 420                 int *lengths, char **bufs)
 421 {
 422         return ptlrpc_prep_req_pool(imp, version, opcode, count, lengths, bufs,
 423                                     NULL, NULL);
 424 }
 425
 426 struct ptlrpc_request_set *ptlrpc_prep_set(void)
 427 {
 428         struct ptlrpc_request_set *set;
 429
 430         ENTRY;
 431         OBD_ALLOC(set, sizeof *set);
 432         if (!set)
 433                 RETURN(NULL);
 434         CFS_INIT_LIST_HEAD(&set->set_requests);
 435         cfs_waitq_init(&set->set_waitq);
 436         set->set_remaining = 0;
 437         spin_lock_init(&set->set_new_req_lock);
 438         CFS_INIT_LIST_HEAD(&set->set_new_requests);
 439
 440         RETURN(set);
 441 }
 442
 443 /* Finish with this set; opposite of prep_set. */
 444 void ptlrpc_set_destroy(struct ptlrpc_request_set *set)
 445 {
 446         struct list_head *tmp;
 447         struct list_head *next;
 448         int               expected_phase;
 449         int               n = 0;
 450         ENTRY;
 451
 452         /* Requests on the set should either all be completed, or all be new */
 453         expected_phase = (set->set_remaining == 0) ?
 454                          RQ_PHASE_COMPLETE : RQ_PHASE_NEW;
 455         list_for_each (tmp, &set->set_requests) {
 456                 struct ptlrpc_request *req =
 457                         list_entry(tmp, struct ptlrpc_request, rq_set_chain);
 458
 459                 LASSERT(req->rq_phase == expected_phase);
 460                 n++;
 461         }
 462
 463         LASSERT(set->set_remaining == 0 || set->set_remaining == n);
 464
 465         list_for_each_safe(tmp, next, &set->set_requests) {
 466                 struct ptlrpc_request *req =
 467                         list_entry(tmp, struct ptlrpc_request, rq_set_chain);
 468                 list_del_init(&req->rq_set_chain);
 469
 470                 LASSERT(req->rq_phase == expected_phase);
 471
 472                 if (req->rq_phase == RQ_PHASE_NEW) {
 473
 474                         if (req->rq_interpret_reply != NULL) {
 475                                 int (*interpreter)(struct ptlrpc_request *,
 476                                                    void *, int) =
 477                                         req->rq_interpret_reply;
 478
 479                                 /* higher level (i.e. LOV) failed;
 480                                  * let the sub reqs clean up */
 481                                 req->rq_status = -EBADR;
 482                                 interpreter(req, &req->rq_async_args,
 483                                             req->rq_status);
 484                         }
 485                         set->set_remaining--;
 486                 }
 487
 488                 req->rq_set = NULL;
 489                 ptlrpc_req_finished (req);
 490         }
 491
 492         LASSERT(set->set_remaining == 0);
 493
 494         OBD_FREE(set, sizeof(*set));
 495         EXIT;
 496 }
 497
 498 void ptlrpc_set_add_req(struct ptlrpc_request_set *set,
 499                         struct ptlrpc_request *req)
 500 {
 501         /* The set takes over the caller's request reference */
 502         list_add_tail(&req->rq_set_chain, &set->set_requests);
 503         req->rq_set = set;
 504         set->set_remaining++;
 505
 506         atomic_inc(&req->rq_import->imp_inflight);
 507 }
 508
 509 /* lock so many callers can add things, the context that owns the set
 510  * is supposed to notice these and move them into the set proper. */
 511 void ptlrpc_set_add_new_req(struct ptlrpc_request_set *set,
 512                             struct ptlrpc_request *req)
 513 {
 514         spin_lock(&set->set_new_req_lock);
 515         /* The set takes over the caller's request reference */
 516         list_add_tail(&req->rq_set_chain, &set->set_new_requests);
 517         req->rq_set = set;
 518         spin_unlock(&set->set_new_req_lock);
 519 }
 520
 521 /*
 522  * Based on the current state of the import, determine if the request
 523  * can be sent, is an error, or should be delayed.
 524  *
 525  * Returns true if this request should be delayed. If false, and
 526  * *status is set, then the request can not be sent and *status is the
 527  * error code.  If false and status is 0, then request can be sent.
 528  *
 529  * The imp->imp_lock must be held.
 530  */
 531 static int ptlrpc_import_delay_req(struct obd_import *imp,
 532                                    struct ptlrpc_request *req, int *status)
 533 {
 534         int delay = 0;
 535         ENTRY;
 536
 537         LASSERT (status != NULL);
 538         *status = 0;
 539
 540         if (req->rq_ctx_init || req->rq_ctx_fini) {
 541                 /* always allow ctx init/fini rpc go through */
 542         } else if (imp->imp_state == LUSTRE_IMP_NEW) {
 543                 DEBUG_REQ(D_ERROR, req, "Uninitialized import.");
 544                 *status = -EIO;
 545                 LBUG();
 546         } else if (imp->imp_state == LUSTRE_IMP_CLOSED) {
 547                 DEBUG_REQ(D_ERROR, req, "IMP_CLOSED ");
 548                 *status = -EIO;
 549         } else if (req->rq_send_state == LUSTRE_IMP_CONNECTING &&
 550                  imp->imp_state == LUSTRE_IMP_CONNECTING) {
 551                 /* allow CONNECT even if import is invalid */ ;
 552         } else if (imp->imp_invalid) {
 553                 /* If the import has been invalidated (such as by an OST
 554                  * failure) the request must fail with -ESHUTDOWN.  This
 555                  * indicates the requests should be discarded; an -EIO
 556                  * may result in a resend of the request. */
 557                 if (!imp->imp_deactive)
 558                         DEBUG_REQ(D_ERROR, req, "IMP_INVALID");
 559                 *status = -ESHUTDOWN; /* bz 12940 */
 560         } else if (req->rq_import_generation != imp->imp_generation) {
 561                 DEBUG_REQ(D_ERROR, req, "req wrong generation:");
 562                 *status = -EIO;
 563         } else if (req->rq_send_state != imp->imp_state) {
 564                 if (imp->imp_obd->obd_no_recov || imp->imp_dlm_fake ||
 565                     req->rq_no_delay)
 566                         *status = -EWOULDBLOCK;
 567                 else
 568                         delay = 1;
 569         }
 570
 571         RETURN(delay);
 572 }
 573
 574 static int ptlrpc_check_reply(struct ptlrpc_request *req)
 575 {
 576         int rc = 0;
 577         ENTRY;
 578
 579         /* serialise with network callback */
 580         spin_lock(&req->rq_lock);
 581
 582         if (req->rq_replied)
 583                 GOTO(out, rc = 1);
 584
 585         if (req->rq_net_err && !req->rq_timedout) {
 586                 spin_unlock(&req->rq_lock);
 587                 rc = ptlrpc_expire_one_request(req);
 588                 spin_lock(&req->rq_lock);
 589                 GOTO(out, rc);
 590         }
 591
 592         if (req->rq_err)
 593                 GOTO(out, rc = 1);
 594
 595         if (req->rq_resend)
 596                 GOTO(out, rc = 1);
 597
 598         if (req->rq_restart)
 599                 GOTO(out, rc = 1);
 600         EXIT;
 601  out:
 602         spin_unlock(&req->rq_lock);
 603         DEBUG_REQ(D_NET, req, "rc = %d for", rc);
 604         return rc;
 605 }
 606
 607 static int ptlrpc_check_status(struct ptlrpc_request *req)
 608 {
 609         int err;
 610         ENTRY;
 611
 612         err = lustre_msg_get_status(req->rq_repmsg);
 613         if (lustre_msg_get_type(req->rq_repmsg) == PTL_RPC_MSG_ERR) {
 614                 struct obd_import *imp = req->rq_import;
 615                 __u32 opc = lustre_msg_get_opc(req->rq_reqmsg);
 616                 LCONSOLE_ERROR_MSG(0x011,"an error occurred while communicating"
 617                                 " with %s. The %s operation failed with %d\n",
 618                                 libcfs_nid2str(imp->imp_connection->c_peer.nid),
 619                                 ll_opcode2str(opc), err);
 620                 RETURN(err < 0 ? err : -EINVAL);
 621         }
 622
 623         if (err < 0) {
 624                 DEBUG_REQ(D_INFO, req, "status is %d", err);
 625         } else if (err > 0) {
 626                 /* XXX: translate this error from net to host */
 627                 DEBUG_REQ(D_INFO, req, "status is %d", err);
 628         }
 629
 630         RETURN(err);
 631 }
 632
 633 static int after_reply(struct ptlrpc_request *req)
 634 {
 635         struct obd_import *imp = req->rq_import;
 636         int rc;
 637         ENTRY;
 638
 639         LASSERT(!req->rq_receiving_reply);
 640         LASSERT(req->rq_nob_received <= req->rq_repbuf_len);
 641
 642         /* NB Until this point, the whole of the incoming message,
 643          * including buflens, status etc is in the sender's byte order. */
 644
 645         /* Clear reply swab mask; this is a new reply in sender's byte order */
 646         req->rq_rep_swab_mask = 0;
 647
 648         rc = sptlrpc_cli_unwrap_reply(req);
 649         if (rc) {
 650                 DEBUG_REQ(D_ERROR, req, "unwrap reply failed (%d):", rc);
 651                 RETURN(rc);
 652         }
 653
 654         /* security layer unwrap might ask resend this request */
 655         if (req->rq_resend)
 656                 RETURN(0);
 657
 658         rc = lustre_unpack_msg(req->rq_repmsg, req->rq_replen);
 659         if (rc) {
 660                 DEBUG_REQ(D_ERROR, req, "unpack_rep failed: %d", rc);
 661                 RETURN(-EPROTO);
 662         }
 663
 664         rc = lustre_unpack_rep_ptlrpc_body(req, MSG_PTLRPC_BODY_OFF);
 665         if (rc) {
 666                 DEBUG_REQ(D_ERROR, req, "unpack ptlrpc body failed: %d", rc);
 667                 RETURN(-EPROTO);
 668         }
 669
 670         if (lustre_msg_get_type(req->rq_repmsg) != PTL_RPC_MSG_REPLY &&
 671             lustre_msg_get_type(req->rq_repmsg) != PTL_RPC_MSG_ERR) {
 672                 DEBUG_REQ(D_ERROR, req, "invalid packet received (type=%u)",
 673                           lustre_msg_get_type(req->rq_repmsg));
 674                 RETURN(-EPROTO);
 675         }
 676
 677         rc = ptlrpc_check_status(req);
 678         imp->imp_connect_error = rc;
 679
 680         if (rc) {
 681                 /* Either we've been evicted, or the server has failed for
 682                  * some reason. Try to reconnect, and if that fails, punt to
 683                  * the upcall. */
 684                 if (ll_rpc_recoverable_error(rc)) {
 685                         if (req->rq_send_state != LUSTRE_IMP_FULL ||
 686                             imp->imp_obd->obd_no_recov || imp->imp_dlm_fake) {
 687                                 RETURN(rc);
 688                         }
 689                         ptlrpc_request_handle_notconn(req);
 690                         RETURN(rc);
 691                 }
 692         } else {
 693                 /* Let's look if server send slv. Do it only for RPC with
 694                  * rc == 0. */
 695                 if (imp->imp_obd->obd_namespace) {
 696                         /* Disconnect rpc is sent when namespace is already
 697                          * destroyed. Let's check this and will not try update
 698                          * pool. */
 699                         ldlm_cli_update_pool(req);
 700                 }
 701         }
 702
 703         /* Store transno in reqmsg for replay. */
 704         req->rq_transno = lustre_msg_get_transno(req->rq_repmsg);
 705         lustre_msg_set_transno(req->rq_reqmsg, req->rq_transno);
 706
 707         if (req->rq_import->imp_replayable) {
 708                 spin_lock(&imp->imp_lock);
 709                 /* no point in adding already-committed requests to the replay
 710                  * list, we will just remove them immediately. b=9829 */
 711                 if (req->rq_transno != 0 &&
 712                     (req->rq_transno >
 713                      lustre_msg_get_last_committed(req->rq_repmsg) ||
 714                      req->rq_replay))
 715                         ptlrpc_retain_replayable_request(req, imp);
 716                 else if (req->rq_commit_cb != NULL) {
 717                         spin_unlock(&imp->imp_lock);
 718                         req->rq_commit_cb(req);
 719                         spin_lock(&imp->imp_lock);
 720                 }
 721
 722                 /* Replay-enabled imports return commit-status information. */
 723                 if (lustre_msg_get_last_committed(req->rq_repmsg)) {
 724                         imp->imp_peer_committed_transno =
 725                                 lustre_msg_get_last_committed(req->rq_repmsg);
 726                 }
 727                 ptlrpc_free_committed(imp);
 728                 spin_unlock(&imp->imp_lock);
 729         }
 730
 731         RETURN(rc);
 732 }
 733
 734 static int ptlrpc_send_new_req(struct ptlrpc_request *req)
 735 {
 736         struct obd_import     *imp;
 737         int rc;
 738         ENTRY;
 739
 740         LASSERT(req->rq_phase == RQ_PHASE_NEW);
 741         if (req->rq_sent && (req->rq_sent > CURRENT_SECONDS))
 742                 RETURN (0);
 743
 744         req->rq_phase = RQ_PHASE_RPC;
 745
 746         imp = req->rq_import;
 747         spin_lock(&imp->imp_lock);
 748
 749         req->rq_import_generation = imp->imp_generation;
 750
 751         if (ptlrpc_import_delay_req(imp, req, &rc)) {
 752                 spin_lock (&req->rq_lock);
 753                 req->rq_waiting = 1;
 754                 spin_unlock (&req->rq_lock);
 755
 756                 DEBUG_REQ(D_HA, req, "req from PID %d waiting for recovery: "
 757                           "(%s != %s)",
 758                           lustre_msg_get_status(req->rq_reqmsg) ,
 759                           ptlrpc_import_state_name(req->rq_send_state),
 760                           ptlrpc_import_state_name(imp->imp_state));
 761                 LASSERT(list_empty (&req->rq_list));
 762
 763                 list_add_tail(&req->rq_list, &imp->imp_delayed_list);
 764                 spin_unlock(&imp->imp_lock);
 765                 RETURN(0);
 766         }
 767
 768         if (rc != 0) {
 769                 spin_unlock(&imp->imp_lock);
 770                 req->rq_status = rc;
 771                 req->rq_phase = RQ_PHASE_INTERPRET;
 772                 RETURN(rc);
 773         }
 774
 775         /* XXX this is the same as ptlrpc_queue_wait */
 776         LASSERT(list_empty(&req->rq_list));
 777         list_add_tail(&req->rq_list, &imp->imp_sending_list);
 778         spin_unlock(&imp->imp_lock);
 779
 780         lustre_msg_set_status(req->rq_reqmsg, cfs_curproc_pid());
 781
 782         rc = sptlrpc_req_refresh_ctx(req, -1);
 783         if (rc) {
 784                 if (req->rq_err) {
 785                         req->rq_status = rc;
 786                         RETURN(1);
 787                 } else {
 788                         /* here begins timeout counting */
 789                         req->rq_sent = CURRENT_SECONDS;
 790                         req->rq_wait_ctx = 1;
 791                         RETURN(0);
 792                 }
 793         }
 794
 795         CDEBUG(D_RPCTRACE, "Sending RPC pname:cluuid:pid:xid:nid:opc"
 796                " %s:%s:%d:"LPU64":%s:%d\n", cfs_curproc_comm(),
 797                imp->imp_obd->obd_uuid.uuid,
 798                lustre_msg_get_status(req->rq_reqmsg), req->rq_xid,
 799                libcfs_nid2str(imp->imp_connection->c_peer.nid),
 800                lustre_msg_get_opc(req->rq_reqmsg));
 801
 802         rc = ptl_send_rpc(req, 0);
 803         if (rc) {
 804                 DEBUG_REQ(D_HA, req, "send failed (%d); expect timeout", rc);
 805                 req->rq_net_err = 1;
 806                 RETURN(rc);
 807         }
 808         RETURN(0);
 809 }
 810
 811 /* this sends any unsent RPCs in @set and returns TRUE if all are sent */
 812 int ptlrpc_check_set(struct ptlrpc_request_set *set)
 813 {
 814         struct list_head *tmp;
 815         int force_timer_recalc = 0;
 816         ENTRY;
 817
 818         if (set->set_remaining == 0)
 819                 RETURN(1);
 820
 821         list_for_each(tmp, &set->set_requests) {
 822                 struct ptlrpc_request *req =
 823                         list_entry(tmp, struct ptlrpc_request, rq_set_chain);
 824                 struct obd_import *imp = req->rq_import;
 825                 int rc = 0;
 826
 827                 if (req->rq_phase == RQ_PHASE_NEW &&
 828                     ptlrpc_send_new_req(req)) {
 829                         force_timer_recalc = 1;
 830                 }
 831                 /* delayed send - skip */
 832                 if (req->rq_phase == RQ_PHASE_NEW && req->rq_sent)
 833                         continue;
 834
 835                 if (!(req->rq_phase == RQ_PHASE_RPC ||
 836                       req->rq_phase == RQ_PHASE_BULK ||
 837                       req->rq_phase == RQ_PHASE_INTERPRET ||
 838                       req->rq_phase == RQ_PHASE_COMPLETE)) {
 839                         DEBUG_REQ(D_ERROR, req, "bad phase %x", req->rq_phase);
 840                         LBUG();
 841                 }
 842
 843                 if (req->rq_phase == RQ_PHASE_COMPLETE)
 844                         continue;
 845
 846                 if (req->rq_phase == RQ_PHASE_INTERPRET)
 847                         GOTO(interpret, req->rq_status);
 848
 849                 if (req->rq_net_err && !req->rq_timedout)
 850                         ptlrpc_expire_one_request(req);
 851
 852                 if (req->rq_err) {
 853                         ptlrpc_unregister_reply(req);
 854                         req->rq_replied = 0;
 855                         if (req->rq_status == 0)
 856                                 req->rq_status = -EIO;
 857                         req->rq_phase = RQ_PHASE_INTERPRET;
 858
 859                         spin_lock(&imp->imp_lock);
 860                         list_del_init(&req->rq_list);
 861                         spin_unlock(&imp->imp_lock);
 862
 863                         GOTO(interpret, req->rq_status);
 864                 }
 865
 866                 /* ptlrpc_queue_wait->l_wait_event guarantees that rq_intr
 867                  * will only be set after rq_timedout, but the oig waiting
 868                  * path sets rq_intr irrespective of whether ptlrpcd has
 869                  * seen a timeout.  our policy is to only interpret
 870                  * interrupted rpcs after they have timed out */
 871                 if (req->rq_intr && (req->rq_timedout || req->rq_waiting ||
 872                                      req->rq_wait_ctx)) {
 873                         /* NB could be on delayed list */
 874                         ptlrpc_unregister_reply(req);
 875                         req->rq_status = -EINTR;
 876                         req->rq_phase = RQ_PHASE_INTERPRET;
 877
 878                         spin_lock(&imp->imp_lock);
 879                         list_del_init(&req->rq_list);
 880                         spin_unlock(&imp->imp_lock);
 881
 882                         GOTO(interpret, req->rq_status);
 883                 }
 884
 885                 if (req->rq_phase == RQ_PHASE_RPC) {
 886                         if (req->rq_timedout || req->rq_resend ||
 887                             req->rq_waiting || req->rq_wait_ctx) {
 888                                 int status;
 889
 890                                 /* rq_wait_ctx is only touched in ptlrpcd,
 891                                  * no lock needed here.
 892                                  */
 893                                 if (req->rq_wait_ctx)
 894                                         goto check_ctx;
 895
 896                                 ptlrpc_unregister_reply(req);
 897
 898                                 spin_lock(&imp->imp_lock);
 899
 900                                 if (ptlrpc_import_delay_req(imp, req, &status)){
 901                                         spin_unlock(&imp->imp_lock);
 902                                         continue;
 903                                 }
 904
 905                                 list_del_init(&req->rq_list);
 906                                 if (status != 0)  {
 907                                         req->rq_status = status;
 908                                         req->rq_phase = RQ_PHASE_INTERPRET;
 909                                         spin_unlock(&imp->imp_lock);
 910                                         GOTO(interpret, req->rq_status);
 911                                 }
 912                                 if (req->rq_no_resend && !req->rq_wait_ctx) {
 913                                         req->rq_status = -ENOTCONN;
 914                                         req->rq_phase = RQ_PHASE_INTERPRET;
 915                                         spin_unlock(&imp->imp_lock);
 916                                         GOTO(interpret, req->rq_status);
 917                                 }
 918                                 list_add_tail(&req->rq_list,
 919                                               &imp->imp_sending_list);
 920
 921                                 spin_unlock(&imp->imp_lock);
 922
 923                                 req->rq_waiting = 0;
 924                                 if (req->rq_resend) {
 925                                         lustre_msg_add_flags(req->rq_reqmsg,
 926                                                              MSG_RESENT);
 927                                         if (req->rq_bulk) {
 928                                                 __u64 old_xid = req->rq_xid;
 929
 930                                                 ptlrpc_unregister_bulk (req);
 931
 932                                                 /* ensure previous bulk fails */
 933                                                 req->rq_xid = ptlrpc_next_xid();
 934                                                 CDEBUG(D_HA, "resend bulk "
 935                                                        "old x"LPU64
 936                                                        " new x"LPU64"\n",
 937                                                        old_xid, req->rq_xid);
 938                                         }
 939                                 }
 940 check_ctx:
 941                                 status = sptlrpc_req_refresh_ctx(req, -1);
 942                                 if (status) {
 943                                         if (req->rq_err) {
 944                                                 req->rq_status = status;
 945                                                 force_timer_recalc = 1;
 946                                         }
 947                                         if (!req->rq_wait_ctx) {
 948                                                 /* begins timeout counting */
 949                                                 req->rq_sent = CURRENT_SECONDS;
 950                                                 req->rq_wait_ctx = 1;
 951                                         }
 952                                         continue;
 953                                 } else {
 954                                         req->rq_sent = 0;
 955                                         req->rq_wait_ctx = 0;
 956                                 }
 957
 958                                 rc = ptl_send_rpc(req, 0);
 959                                 if (rc) {
 960                                         DEBUG_REQ(D_HA, req, "send failed (%d)",
 961                                                   rc);
 962                                         force_timer_recalc = 1;
 963                                         req->rq_net_err = 1;
 964                                 }
 965                                 /* need to reset the timeout */
 966                                 force_timer_recalc = 1;
 967                         }
 968
 969                         /* Still waiting for a reply? */
 970                         if (ptlrpc_client_receiving_reply(req))
 971                                 continue;
 972
 973                         /* Did we actually receive a reply? */
 974                         if (!ptlrpc_client_replied(req))
 975                                 continue;
 976
 977                         spin_lock(&imp->imp_lock);
 978                         list_del_init(&req->rq_list);
 979                         spin_unlock(&imp->imp_lock);
 980
 981                         req->rq_status = after_reply(req);
 982                         if (req->rq_resend) {
 983                                 /* Add this req to the delayed list so
 984                                    it can be errored if the import is
 985                                    evicted after recovery. */
 986                                 spin_lock(&imp->imp_lock);
 987                                 list_add_tail(&req->rq_list,
 988                                               &imp->imp_delayed_list);
 989                                 spin_unlock(&imp->imp_lock);
 990                                 continue;
 991                         }
 992
 993                         /* If there is no bulk associated with this request,
 994                          * then we're done and should let the interpreter
 995                          * process the reply.  Similarly if the RPC returned
 996                          * an error, and therefore the bulk will never arrive.
 997                          */
 998                         if (req->rq_bulk == NULL || req->rq_status != 0) {
 999                                 req->rq_phase = RQ_PHASE_INTERPRET;
1000                                 GOTO(interpret, req->rq_status);
1001                         }
1002
1003                         req->rq_phase = RQ_PHASE_BULK;
1004                 }
1005
1006                 LASSERT(req->rq_phase == RQ_PHASE_BULK);
1007                 if (ptlrpc_bulk_active(req->rq_bulk))
1008                         continue;
1009
1010                 if (!req->rq_bulk->bd_success) {
1011                         /* The RPC reply arrived OK, but the bulk screwed
1012                          * up!  Dead wierd since the server told us the RPC
1013                          * was good after getting the REPLY for her GET or
1014                          * the ACK for her PUT. */
1015                         DEBUG_REQ(D_ERROR, req, "bulk transfer failed");
1016                         LBUG();
1017                 }
1018
1019                 req->rq_phase = RQ_PHASE_INTERPRET;
1020
1021         interpret:
1022                 LASSERT(req->rq_phase == RQ_PHASE_INTERPRET);
1023                 LASSERT(!req->rq_receiving_reply);
1024
1025                 ptlrpc_unregister_reply(req);
1026                 if (req->rq_bulk != NULL)
1027                         ptlrpc_unregister_bulk (req);
1028
1029                 if (req->rq_interpret_reply != NULL) {
1030                         int (*interpreter)(struct ptlrpc_request *,void *,int) =
1031                                 req->rq_interpret_reply;
1032                         req->rq_status = interpreter(req, &req->rq_async_args,
1033                                                      req->rq_status);
1034                 }
1035                 req->rq_phase = RQ_PHASE_COMPLETE;
1036
1037                 CDEBUG(D_RPCTRACE, "Completed RPC pname:cluuid:pid:xid:nid:"
1038                        "opc %s:%s:%d:"LPU64":%s:%d\n", cfs_curproc_comm(),
1039                        imp->imp_obd->obd_uuid.uuid,
1040                        lustre_msg_get_status(req->rq_reqmsg), req->rq_xid,
1041                        libcfs_nid2str(imp->imp_connection->c_peer.nid),
1042                        lustre_msg_get_opc(req->rq_reqmsg));
1043
1044                 set->set_remaining--;
1045
1046                 atomic_dec(&imp->imp_inflight);
1047                 cfs_waitq_signal(&imp->imp_recovery_waitq);
1048         }
1049
1050         /* If we hit an error, we want to recover promptly. */
1051         RETURN(set->set_remaining == 0 || force_timer_recalc);
1052 }
1053
1054 int ptlrpc_expire_one_request(struct ptlrpc_request *req)
1055 {
1056         struct obd_import *imp = req->rq_import;
1057         int rc = 0;
1058         ENTRY;
1059
1060         DEBUG_REQ(D_ERROR|D_NETERROR, req, "%s (sent at %lu, %lus ago)",
1061                   req->rq_net_err ? "network error" : "timeout",
1062                   (long)req->rq_sent, CURRENT_SECONDS - req->rq_sent);
1063
1064         if (imp != NULL && obd_debug_peer_on_timeout)
1065                 LNetCtl(IOC_LIBCFS_DEBUG_PEER, &imp->imp_connection->c_peer);
1066
1067         spin_lock(&req->rq_lock);
1068         req->rq_timedout = 1;
1069         req->rq_wait_ctx = 0;
1070         spin_unlock(&req->rq_lock);
1071
1072         ptlrpc_unregister_reply (req);
1073
1074         if (obd_dump_on_timeout)
1075                 libcfs_debug_dumplog();
1076
1077         if (req->rq_bulk != NULL)
1078                 ptlrpc_unregister_bulk (req);
1079
1080         if (imp == NULL) {
1081                 DEBUG_REQ(D_HA, req, "NULL import: already cleaned up?");
1082                 RETURN(1);
1083         }
1084
1085         /* The DLM server doesn't want recovery run on its imports. */
1086         if (imp->imp_dlm_fake)
1087                 RETURN(1);
1088
1089         /* If this request is for recovery or other primordial tasks,
1090          * then error it out here. */
1091         if (req->rq_ctx_init || req->rq_ctx_fini ||
1092             req->rq_send_state != LUSTRE_IMP_FULL ||
1093             imp->imp_obd->obd_no_recov) {
1094                 spin_lock(&req->rq_lock);
1095                 req->rq_status = -ETIMEDOUT;
1096                 req->rq_err = 1;
1097                 spin_unlock(&req->rq_lock);
1098                 RETURN(1);
1099         }
1100
1101         /* if request can't be resend we can't wait answer after timeout */
1102         if (req->rq_no_resend) {
1103                 DEBUG_REQ(D_RPCTRACE, req, "TIMEOUT-NORESEND:");
1104                 rc = 1;
1105         }
1106
1107         ptlrpc_fail_import(imp, lustre_msg_get_conn_cnt(req->rq_reqmsg));
1108
1109         RETURN(rc);
1110 }
1111
1112 int ptlrpc_expired_set(void *data)
1113 {
1114         struct ptlrpc_request_set *set = data;
1115         struct list_head          *tmp;
1116         time_t                     now = CURRENT_SECONDS;
1117         ENTRY;
1118
1119         LASSERT(set != NULL);
1120
1121         /* A timeout expired; see which reqs it applies to... */
1122         list_for_each (tmp, &set->set_requests) {
1123                 struct ptlrpc_request *req =
1124                         list_entry(tmp, struct ptlrpc_request, rq_set_chain);
1125
1126                 /* request in-flight? */
1127                 if (!((req->rq_phase == RQ_PHASE_RPC && !req->rq_waiting &&
1128                        !req->rq_resend) ||
1129                       (req->rq_phase == RQ_PHASE_BULK)))
1130                         continue;
1131
1132                 if (req->rq_timedout ||           /* already dealt with */
1133                     req->rq_sent + req->rq_timeout > now) /* not expired */
1134                         continue;
1135
1136                 /* deal with this guy */
1137                 ptlrpc_expire_one_request (req);
1138         }
1139
1140         /* When waiting for a whole set, we always to break out of the
1141          * sleep so we can recalculate the timeout, or enable interrupts
1142          * iff everyone's timed out.
1143          */
1144         RETURN(1);
1145 }
1146
1147 void ptlrpc_mark_interrupted(struct ptlrpc_request *req)
1148 {
1149         spin_lock(&req->rq_lock);
1150         req->rq_intr = 1;
1151         spin_unlock(&req->rq_lock);
1152 }
1153
1154 void ptlrpc_interrupted_set(void *data)
1155 {
1156         struct ptlrpc_request_set *set = data;
1157         struct list_head *tmp;
1158
1159         LASSERT(set != NULL);
1160         CERROR("INTERRUPTED SET %p\n", set);
1161
1162         list_for_each(tmp, &set->set_requests) {
1163                 struct ptlrpc_request *req =
1164                         list_entry(tmp, struct ptlrpc_request, rq_set_chain);
1165
1166                 if (req->rq_phase != RQ_PHASE_RPC)
1167                         continue;
1168
1169                 ptlrpc_mark_interrupted(req);
1170         }
1171 }
1172
1173 int ptlrpc_set_next_timeout(struct ptlrpc_request_set *set)
1174 {
1175         struct list_head      *tmp;
1176         time_t                 now = CURRENT_SECONDS;
1177         time_t                 deadline;
1178         int                    timeout = 0;
1179         struct ptlrpc_request *req;
1180         ENTRY;
1181
1182         SIGNAL_MASK_ASSERT(); /* XXX BUG 1511 */
1183
1184         list_for_each(tmp, &set->set_requests) {
1185                 req = list_entry(tmp, struct ptlrpc_request, rq_set_chain);
1186
1187                 /* request in-flight? */
1188                 if (!((req->rq_phase == RQ_PHASE_RPC && !req->rq_waiting) ||
1189                       (req->rq_phase == RQ_PHASE_BULK) ||
1190                       (req->rq_phase == RQ_PHASE_NEW)))
1191                         continue;
1192
1193                 if (req->rq_timedout)   /* already timed out */
1194                         continue;
1195
1196                 if (req->rq_phase == RQ_PHASE_NEW)
1197                         deadline = req->rq_sent;
1198                 else
1199                         deadline = req->rq_sent + req->rq_timeout;
1200
1201                 if (deadline <= now)    /* actually expired already */
1202                         timeout = 1;    /* ASAP */
1203                 else if (timeout == 0 || timeout > deadline - now)
1204                         timeout = deadline - now;
1205         }
1206         RETURN(timeout);
1207 }
1208
1209 int ptlrpc_set_wait(struct ptlrpc_request_set *set)
1210 {
1211         struct list_head      *tmp;
1212         struct ptlrpc_request *req;
1213         struct l_wait_info     lwi;
1214         int                    rc, timeout;
1215         ENTRY;
1216
1217         if (list_empty(&set->set_requests))
1218                 RETURN(0);
1219
1220         list_for_each(tmp, &set->set_requests) {
1221                 req = list_entry(tmp, struct ptlrpc_request, rq_set_chain);
1222                 if (req->rq_phase == RQ_PHASE_NEW)
1223                         (void)ptlrpc_send_new_req(req);
1224         }
1225
1226         do {
1227                 timeout = ptlrpc_set_next_timeout(set);
1228
1229                 /* wait until all complete, interrupted, or an in-flight
1230                  * req times out */
1231                 CDEBUG(D_RPCTRACE, "set %p going to sleep for %d seconds\n",
1232                        set, timeout);
1233                 lwi = LWI_TIMEOUT_INTR(cfs_time_seconds(timeout ? timeout : 1),
1234                                        ptlrpc_expired_set,
1235                                        ptlrpc_interrupted_set, set);
1236                 rc = l_wait_event(set->set_waitq, ptlrpc_check_set(set), &lwi);
1237
1238                 LASSERT(rc == 0 || rc == -EINTR || rc == -ETIMEDOUT);
1239
1240                 /* -EINTR => all requests have been flagged rq_intr so next
1241                  * check completes.
1242                  * -ETIMEOUTD => someone timed out.  When all reqs have
1243                  * timed out, signals are enabled allowing completion with
1244                  * EINTR.
1245                  * I don't really care if we go once more round the loop in
1246                  * the error cases -eeb. */
1247         } while (rc != 0 || set->set_remaining != 0);
1248
1249         LASSERT(set->set_remaining == 0);
1250
1251         rc = 0;
1252         list_for_each(tmp, &set->set_requests) {
1253                 req = list_entry(tmp, struct ptlrpc_request, rq_set_chain);
1254
1255                 LASSERT(req->rq_phase == RQ_PHASE_COMPLETE);
1256                 if (req->rq_status != 0)
1257                         rc = req->rq_status;
1258         }
1259
1260         if (set->set_interpret != NULL) {
1261                 int (*interpreter)(struct ptlrpc_request_set *set,void *,int) =
1262                         set->set_interpret;
1263                 rc = interpreter (set, set->set_arg, rc);
1264         }
1265
1266         RETURN(rc);
1267 }
1268
1269 static void __ptlrpc_free_req(struct ptlrpc_request *request, int locked)
1270 {
1271         ENTRY;
1272         if (request == NULL) {
1273                 EXIT;
1274                 return;
1275         }
1276
1277         LASSERTF(!request->rq_receiving_reply, "req %p\n", request);
1278         LASSERTF(request->rq_rqbd == NULL, "req %p\n",request);/* client-side */
1279         LASSERTF(list_empty(&request->rq_list), "req %p\n", request);
1280         LASSERTF(list_empty(&request->rq_set_chain), "req %p\n", request);
1281         LASSERT(request->rq_cli_ctx);
1282
1283         /* We must take it off the imp_replay_list first.  Otherwise, we'll set
1284          * request->rq_reqmsg to NULL while osc_close is dereferencing it. */
1285         if (request->rq_import != NULL) {
1286                 if (!locked)
1287                         spin_lock(&request->rq_import->imp_lock);
1288                 list_del_init(&request->rq_mod_list);
1289                 list_del_init(&request->rq_replay_list);
1290                 if (!locked)
1291                         spin_unlock(&request->rq_import->imp_lock);
1292         }
1293         LASSERTF(list_empty(&request->rq_replay_list), "req %p\n", request);
1294
1295         if (atomic_read(&request->rq_refcount) != 0) {
1296                 DEBUG_REQ(D_ERROR, request,
1297                           "freeing request with nonzero refcount");
1298                 LBUG();
1299         }
1300
1301         if (request->rq_repbuf != NULL)
1302                 sptlrpc_cli_free_repbuf(request);
1303         if (request->rq_export != NULL) {
1304                 class_export_put(request->rq_export);
1305                 request->rq_export = NULL;
1306         }
1307         if (request->rq_import != NULL) {
1308                 class_import_put(request->rq_import);
1309                 request->rq_import = NULL;
1310         }
1311         if (request->rq_bulk != NULL)
1312                 ptlrpc_free_bulk(request->rq_bulk);
1313
1314         if (request->rq_reqbuf != NULL || request->rq_clrbuf != NULL)
1315                 sptlrpc_cli_free_reqbuf(request);
1316
1317         sptlrpc_req_put_ctx(request, !locked);
1318
1319         if (request->rq_pool)
1320                 __ptlrpc_free_req_to_pool(request);
1321         else
1322                 OBD_FREE(request, sizeof(*request));
1323         EXIT;
1324 }
1325
1326 void ptlrpc_free_req(struct ptlrpc_request *request)
1327 {
1328         __ptlrpc_free_req(request, 0);
1329 }
1330
1331 static int __ptlrpc_req_finished(struct ptlrpc_request *request, int locked);
1332 void ptlrpc_req_finished_with_imp_lock(struct ptlrpc_request *request)
1333 {
1334         LASSERT_SPIN_LOCKED(&request->rq_import->imp_lock);
1335         (void)__ptlrpc_req_finished(request, 1);
1336 }
1337
1338 static int __ptlrpc_req_finished(struct ptlrpc_request *request, int locked)
1339 {
1340         ENTRY;
1341         if (request == NULL)
1342                 RETURN(1);
1343
1344         if (request == LP_POISON ||
1345             request->rq_reqmsg == LP_POISON) {
1346                 CERROR("dereferencing freed request (bug 575)\n");
1347                 LBUG();
1348                 RETURN(1);
1349         }
1350
1351         DEBUG_REQ(D_INFO, request, "refcount now %u",
1352                   atomic_read(&request->rq_refcount) - 1);
1353
1354         if (atomic_dec_and_test(&request->rq_refcount)) {
1355                 __ptlrpc_free_req(request, locked);
1356                 RETURN(1);
1357         }
1358
1359         RETURN(0);
1360 }
1361
1362 void ptlrpc_req_finished(struct ptlrpc_request *request)
1363 {
1364         __ptlrpc_req_finished(request, 0);
1365 }
1366
1367 __u64 ptlrpc_req_xid(struct ptlrpc_request *request)
1368 {
1369         return request->rq_xid;
1370 }
1371 EXPORT_SYMBOL(ptlrpc_req_xid);
1372
1373 /* Disengage the client's reply buffer from the network
1374  * NB does _NOT_ unregister any client-side bulk.
1375  * IDEMPOTENT, but _not_ safe against concurrent callers.
1376  * The request owner (i.e. the thread doing the I/O) must call...
1377  */
1378 void ptlrpc_unregister_reply (struct ptlrpc_request *request)
1379 {
1380         int                rc;
1381         cfs_waitq_t       *wq;
1382         struct l_wait_info lwi;
1383
1384         LASSERT(!in_interrupt ());             /* might sleep */
1385
1386         if (!ptlrpc_client_receiving_reply(request))
1387                 return;
1388
1389         LNetMDUnlink (request->rq_reply_md_h);
1390
1391         /* We have to l_wait_event() whatever the result, to give liblustre
1392          * a chance to run reply_in_callback() */
1393
1394         if (request->rq_set != NULL)
1395                 wq = &request->rq_set->set_waitq;
1396         else
1397                 wq = &request->rq_reply_waitq;
1398
1399         for (;;) {
1400                 /* Network access will complete in finite time but the HUGE
1401                  * timeout lets us CWARN for visibility of sluggish NALs */
1402                 lwi = LWI_TIMEOUT(cfs_time_seconds(300), NULL, NULL);
1403                 rc = l_wait_event (*wq, !ptlrpc_client_receiving_reply(request), &lwi);
1404                 if (rc == 0)
1405                         return;
1406
1407                 LASSERT (rc == -ETIMEDOUT);
1408                 DEBUG_REQ(D_WARNING, request, "Unexpectedly long timeout");
1409         }
1410 }
1411
1412 /* caller must hold imp->imp_lock */
1413 void ptlrpc_free_committed(struct obd_import *imp)
1414 {
1415         struct list_head *tmp, *saved;
1416         struct ptlrpc_request *req;
1417         struct ptlrpc_request *last_req = NULL; /* temporary fire escape */
1418         ENTRY;
1419
1420         LASSERT(imp != NULL);
1421
1422         LASSERT_SPIN_LOCKED(&imp->imp_lock);
1423
1424
1425         if (imp->imp_peer_committed_transno == imp->imp_last_transno_checked &&
1426             imp->imp_generation == imp->imp_last_generation_checked) {
1427                 CDEBUG(D_RPCTRACE, "%s: skip recheck: last_committed "LPU64"\n",
1428                        imp->imp_obd->obd_name, imp->imp_peer_committed_transno);
1429                 return;
1430         }
1431
1432         CDEBUG(D_RPCTRACE, "%s: committing for last_committed "LPU64" gen %d\n",
1433                imp->imp_obd->obd_name, imp->imp_peer_committed_transno,
1434                imp->imp_generation);
1435         imp->imp_last_transno_checked = imp->imp_peer_committed_transno;
1436         imp->imp_last_generation_checked = imp->imp_generation;
1437
1438         list_for_each_safe(tmp, saved, &imp->imp_replay_list) {
1439                 req = list_entry(tmp, struct ptlrpc_request, rq_replay_list);
1440
1441                 /* XXX ok to remove when 1357 resolved - rread 05/29/03  */
1442                 LASSERT(req != last_req);
1443                 last_req = req;
1444
1445                 if (req->rq_import_generation < imp->imp_generation) {
1446                         DEBUG_REQ(D_RPCTRACE, req, "free request with old gen");
1447                         GOTO(free_req, 0);
1448                 }
1449
1450                 if (req->rq_replay) {
1451                         DEBUG_REQ(D_RPCTRACE, req, "keeping (FL_REPLAY)");
1452                         continue;
1453                 }
1454
1455                 /* not yet committed */
1456                 if (req->rq_transno > imp->imp_peer_committed_transno) {
1457                         DEBUG_REQ(D_RPCTRACE, req, "stopping search");
1458                         break;
1459                 }
1460
1461                 DEBUG_REQ(D_RPCTRACE, req, "commit (last_committed "LPU64")",
1462                           imp->imp_peer_committed_transno);
1463 free_req:
1464                 spin_lock(&req->rq_lock);
1465                 req->rq_replay = 0;
1466                 spin_unlock(&req->rq_lock);
1467                 if (req->rq_commit_cb != NULL)
1468                         req->rq_commit_cb(req);
1469                 list_del_init(&req->rq_replay_list);
1470                 __ptlrpc_req_finished(req, 1);
1471         }
1472
1473         EXIT;
1474         return;
1475 }
1476
1477 void ptlrpc_cleanup_client(struct obd_import *imp)
1478 {
1479         ENTRY;
1480         EXIT;
1481         return;
1482 }
1483
1484 void ptlrpc_resend_req(struct ptlrpc_request *req)
1485 {
1486         DEBUG_REQ(D_HA, req, "going to resend");
1487         lustre_msg_set_handle(req->rq_reqmsg, &(struct lustre_handle){ 0 });
1488         req->rq_status = -EAGAIN;
1489
1490         spin_lock(&req->rq_lock);
1491         req->rq_resend = 1;
1492         req->rq_net_err = 0;
1493         req->rq_timedout = 0;
1494         if (req->rq_bulk) {
1495                 __u64 old_xid = req->rq_xid;
1496
1497                 /* ensure previous bulk fails */
1498                 req->rq_xid = ptlrpc_next_xid();
1499                 CDEBUG(D_HA, "resend bulk old x"LPU64" new x"LPU64"\n",
1500                        old_xid, req->rq_xid);
1501         }
1502         ptlrpc_wake_client_req(req);
1503         spin_unlock(&req->rq_lock);
1504 }
1505
1506 /* XXX: this function and rq_status are currently unused */
1507 void ptlrpc_restart_req(struct ptlrpc_request *req)
1508 {
1509         DEBUG_REQ(D_HA, req, "restarting (possibly-)completed request");
1510         req->rq_status = -ERESTARTSYS;
1511
1512         spin_lock(&req->rq_lock);
1513         req->rq_restart = 1;
1514         req->rq_timedout = 0;
1515         ptlrpc_wake_client_req(req);
1516         spin_unlock(&req->rq_lock);
1517 }
1518
1519 static int expired_request(void *data)
1520 {
1521         struct ptlrpc_request *req = data;
1522         ENTRY;
1523
1524         /* some failure can suspend regular timeouts */
1525         if (ptlrpc_check_suspend())
1526                 RETURN(1);
1527
1528         RETURN(ptlrpc_expire_one_request(req));
1529 }
1530
1531 static void interrupted_request(void *data)
1532 {
1533         struct ptlrpc_request *req = data;
1534         DEBUG_REQ(D_HA, req, "request interrupted");
1535         spin_lock(&req->rq_lock);
1536         req->rq_intr = 1;
1537         spin_unlock(&req->rq_lock);
1538 }
1539
1540 struct ptlrpc_request *ptlrpc_request_addref(struct ptlrpc_request *req)
1541 {
1542         ENTRY;
1543         atomic_inc(&req->rq_refcount);
1544         RETURN(req);
1545 }
1546
1547 void ptlrpc_retain_replayable_request(struct ptlrpc_request *req,
1548                                       struct obd_import *imp)
1549 {
1550         struct list_head *tmp;
1551
1552         LASSERT_SPIN_LOCKED(&imp->imp_lock);
1553
1554         /* clear this for new requests that were resent as well
1555            as resent replayed requests. */
1556         lustre_msg_clear_flags(req->rq_reqmsg, MSG_RESENT);
1557
1558         /* don't re-add requests that have been replayed */
1559         if (!list_empty(&req->rq_replay_list))
1560                 return;
1561
1562         lustre_msg_add_flags(req->rq_reqmsg, MSG_REPLAY);
1563
1564         LASSERT(imp->imp_replayable);
1565         /* Balanced in ptlrpc_free_committed, usually. */
1566         ptlrpc_request_addref(req);
1567         list_for_each_prev(tmp, &imp->imp_replay_list) {
1568                 struct ptlrpc_request *iter =
1569                         list_entry(tmp, struct ptlrpc_request, rq_replay_list);
1570
1571                 /* We may have duplicate transnos if we create and then
1572                  * open a file, or for closes retained if to match creating
1573                  * opens, so use req->rq_xid as a secondary key.
1574                  * (See bugs 684, 685, and 428.)
1575                  * XXX no longer needed, but all opens need transnos!
1576                  */
1577                 if (iter->rq_transno > req->rq_transno)
1578                         continue;
1579
1580                 if (iter->rq_transno == req->rq_transno) {
1581                         LASSERT(iter->rq_xid != req->rq_xid);
1582                         if (iter->rq_xid > req->rq_xid)
1583                                 continue;
1584                 }
1585
1586                 list_add(&req->rq_replay_list, &iter->rq_replay_list);
1587                 return;
1588         }
1589
1590         list_add_tail(&req->rq_replay_list, &imp->imp_replay_list);
1591 }
1592
1593 int ptlrpc_queue_wait(struct ptlrpc_request *req)
1594 {
1595         int rc = 0;
1596         int brc;
1597         struct l_wait_info lwi;
1598         struct obd_import *imp = req->rq_import;
1599         cfs_duration_t timeout = 0;
1600         ENTRY;
1601
1602         LASSERT(req->rq_set == NULL);
1603         LASSERT(!req->rq_receiving_reply);
1604         atomic_inc(&imp->imp_inflight);
1605
1606         /* for distributed debugging */
1607         lustre_msg_set_status(req->rq_reqmsg, cfs_curproc_pid());
1608         LASSERT(imp->imp_obd != NULL);
1609         CDEBUG(D_RPCTRACE, "Sending RPC pname:cluuid:pid:xid:nid:opc "
1610                "%s:%s:%d:"LPU64":%s:%d\n", cfs_curproc_comm(),
1611                imp->imp_obd->obd_uuid.uuid,
1612                lustre_msg_get_status(req->rq_reqmsg), req->rq_xid,
1613                libcfs_nid2str(imp->imp_connection->c_peer.nid),
1614                lustre_msg_get_opc(req->rq_reqmsg));
1615
1616         /* Mark phase here for a little debug help */
1617         req->rq_phase = RQ_PHASE_RPC;
1618
1619         spin_lock(&imp->imp_lock);
1620         req->rq_import_generation = imp->imp_generation;
1621 restart:
1622         if (ptlrpc_import_delay_req(imp, req, &rc)) {
1623                 list_del(&req->rq_list);
1624
1625                 list_add_tail(&req->rq_list, &imp->imp_delayed_list);
1626                 spin_unlock(&imp->imp_lock);
1627
1628                 DEBUG_REQ(D_HA, req, "\"%s\" waiting for recovery: (%s != %s)",
1629                           cfs_curproc_comm(),
1630                           ptlrpc_import_state_name(req->rq_send_state),
1631                           ptlrpc_import_state_name(imp->imp_state));
1632                 lwi = LWI_INTR(interrupted_request, req);
1633                 rc = l_wait_event(req->rq_reply_waitq,
1634                                   (req->rq_send_state == imp->imp_state ||
1635                                    req->rq_err || req->rq_intr),
1636                                   &lwi);
1637                 DEBUG_REQ(D_HA, req, "\"%s\" awake: (%s == %s or %d/%d == 1)",
1638                           cfs_curproc_comm(),
1639                           ptlrpc_import_state_name(imp->imp_state),
1640                           ptlrpc_import_state_name(req->rq_send_state),
1641                           req->rq_err, req->rq_intr);
1642
1643                 spin_lock(&imp->imp_lock);
1644                 list_del_init(&req->rq_list);
1645
1646                 if (req->rq_err) {
1647                         rc = -EIO;
1648                 }
1649                 else if (req->rq_intr) {
1650                         rc = -EINTR;
1651                 }
1652                 else if (req->rq_no_resend) {
1653                         spin_unlock(&imp->imp_lock);
1654                         GOTO(out, rc = -ETIMEDOUT);
1655                 }
1656                 else {
1657                         GOTO(restart, rc);
1658                 }
1659         }
1660
1661         if (rc != 0) {
1662                 list_del_init(&req->rq_list);
1663                 spin_unlock(&imp->imp_lock);
1664                 req->rq_status = rc; // XXX this ok?
1665                 GOTO(out, rc);
1666         }
1667
1668         if (req->rq_resend) {
1669                 lustre_msg_add_flags(req->rq_reqmsg, MSG_RESENT);
1670
1671                 if (req->rq_bulk != NULL) {
1672                         ptlrpc_unregister_bulk (req);
1673
1674                         /* bulk requests are supposed to be
1675                          * idempotent, so we are free to bump the xid
1676                          * here, which we need to do before
1677                          * registering the bulk again (bug 6371).
1678                          * print the old xid first for sanity.
1679                          */
1680                         DEBUG_REQ(D_HA, req, "bumping xid for bulk: ");
1681                         req->rq_xid = ptlrpc_next_xid();
1682                 }
1683
1684                 DEBUG_REQ(D_HA, req, "resending: ");
1685         }
1686
1687         /* XXX this is the same as ptlrpc_set_wait */
1688         LASSERT(list_empty(&req->rq_list));
1689         list_add_tail(&req->rq_list, &imp->imp_sending_list);
1690         spin_unlock(&imp->imp_lock);
1691
1692         rc = sptlrpc_req_refresh_ctx(req, 0);
1693         if (rc) {
1694                 if (req->rq_err) {
1695                         /* we got fatal ctx refresh error, directly jump out
1696                          * thus we can pass back the actual error code.
1697                          */
1698                         spin_lock(&imp->imp_lock);
1699                         list_del_init(&req->rq_list);
1700                         spin_unlock(&imp->imp_lock);
1701
1702                         CERROR("Failed to refresh ctx of req %p: %d\n", req, rc);
1703                         GOTO(out, rc);
1704                 }
1705                 /* simulating we got error during send rpc */
1706                 goto after_send;
1707         }
1708
1709         rc = ptl_send_rpc(req, 0);
1710         if (rc) {
1711                 DEBUG_REQ(D_HA, req, "send failed (%d); recovering", rc);
1712                 timeout = CFS_TICK;
1713         } else {
1714                 timeout = cfs_timeout_cap(cfs_time_seconds(req->rq_timeout));
1715                 DEBUG_REQ(D_NET, req,
1716                           "-- sleeping for "CFS_DURATION_T" jiffies", timeout);
1717         }
1718 repeat:
1719         lwi = LWI_TIMEOUT_INTR(timeout, expired_request, interrupted_request,
1720                                req);
1721         rc = l_wait_event(req->rq_reply_waitq, ptlrpc_check_reply(req), &lwi);
1722         if (rc == -ETIMEDOUT && ptlrpc_check_and_wait_suspend(req))
1723                 goto repeat;
1724
1725 after_send:
1726         CDEBUG(D_RPCTRACE, "Completed RPC pname:cluuid:pid:xid:nid:opc "
1727                "%s:%s:%d:"LPU64":%s:%d\n", cfs_curproc_comm(),
1728                imp->imp_obd->obd_uuid.uuid,
1729                lustre_msg_get_status(req->rq_reqmsg), req->rq_xid,
1730                libcfs_nid2str(imp->imp_connection->c_peer.nid),
1731                lustre_msg_get_opc(req->rq_reqmsg));
1732
1733         spin_lock(&imp->imp_lock);
1734         list_del_init(&req->rq_list);
1735         spin_unlock(&imp->imp_lock);
1736
1737         /* If the reply was received normally, this just grabs the spinlock
1738          * (ensuring the reply callback has returned), sees that
1739          * req->rq_receiving_reply is clear and returns. */
1740         ptlrpc_unregister_reply (req);
1741
1742         if (req->rq_err)
1743                 GOTO(out, rc = -EIO);
1744
1745         /* Resend if we need to, unless we were interrupted. */
1746         if (req->rq_resend && !req->rq_intr) {
1747                 /* ...unless we were specifically told otherwise. */
1748                 if (req->rq_no_resend)
1749                         GOTO(out, rc = -ETIMEDOUT);
1750                 spin_lock(&imp->imp_lock);
1751                 goto restart;
1752         }
1753
1754         if (req->rq_intr) {
1755                 /* Should only be interrupted if we timed out. */
1756                 if (!req->rq_timedout)
1757                         DEBUG_REQ(D_ERROR, req,
1758                                   "rq_intr set but rq_timedout not");
1759                 GOTO(out, rc = -EINTR);
1760         }
1761
1762         if (req->rq_timedout) {                 /* non-recoverable timeout */
1763                 GOTO(out, rc = -ETIMEDOUT);
1764         }
1765
1766         if (!req->rq_replied) {
1767                 /* How can this be? -eeb */
1768                 DEBUG_REQ(D_ERROR, req, "!rq_replied: ");
1769                 LBUG();
1770                 GOTO(out, rc = req->rq_status);
1771         }
1772
1773         rc = after_reply(req);
1774         /* NB may return +ve success rc */
1775         if (req->rq_resend) {
1776                 spin_lock(&imp->imp_lock);
1777                 goto restart;
1778         }
1779
1780  out:
1781         if (req->rq_bulk != NULL) {
1782                 if (rc >= 0) {
1783                         /* success so far.  Note that anything going wrong
1784                          * with bulk now, is EXTREMELY strange, since the
1785                          * server must have believed that the bulk
1786                          * tranferred OK before she replied with success to
1787                          * me. */
1788                         lwi = LWI_TIMEOUT(timeout, NULL, NULL);
1789                         brc = l_wait_event(req->rq_reply_waitq,
1790                                            !ptlrpc_bulk_active(req->rq_bulk),
1791                                            &lwi);
1792                         LASSERT(brc == 0 || brc == -ETIMEDOUT);
1793                         if (brc != 0) {
1794                                 LASSERT(brc == -ETIMEDOUT);
1795                                 DEBUG_REQ(D_ERROR, req, "bulk timed out");
1796                                 rc = brc;
1797                         } else if (!req->rq_bulk->bd_success) {
1798                                 DEBUG_REQ(D_ERROR, req, "bulk transfer failed");
1799                                 rc = -EIO;
1800                         }
1801                 }
1802                 if (rc < 0)
1803                         ptlrpc_unregister_bulk (req);
1804         }
1805
1806         LASSERT(!req->rq_receiving_reply);
1807         req->rq_phase = RQ_PHASE_INTERPRET;
1808
1809         atomic_dec(&imp->imp_inflight);
1810         cfs_waitq_signal(&imp->imp_recovery_waitq);
1811         RETURN(rc);
1812 }
1813
1814 struct ptlrpc_replay_async_args {
1815         int praa_old_state;
1816         int praa_old_status;
1817 };
1818
1819 static int ptlrpc_replay_interpret(struct ptlrpc_request *req,
1820                                     void * data, int rc)
1821 {
1822         struct ptlrpc_replay_async_args *aa = data;
1823         struct obd_import *imp = req->rq_import;
1824
1825         ENTRY;
1826         atomic_dec(&imp->imp_replay_inflight);
1827
1828         if (!req->rq_replied) {
1829                 CERROR("request replay timed out, restarting recovery\n");
1830                 GOTO(out, rc = -ETIMEDOUT);
1831         }
1832
1833         if (lustre_msg_get_type(req->rq_repmsg) == PTL_RPC_MSG_ERR &&
1834             (lustre_msg_get_status(req->rq_repmsg) == -ENOTCONN ||
1835              lustre_msg_get_status(req->rq_repmsg) == -ENODEV))
1836                 GOTO(out, rc = lustre_msg_get_status(req->rq_repmsg));
1837
1838         /* The transno had better not change over replay. */
1839         LASSERT(lustre_msg_get_transno(req->rq_reqmsg) ==
1840                 lustre_msg_get_transno(req->rq_repmsg));
1841
1842         DEBUG_REQ(D_HA, req, "got rep");
1843
1844         /* let the callback do fixups, possibly including in the request */
1845         if (req->rq_replay_cb)
1846                 req->rq_replay_cb(req);
1847
1848         if (req->rq_replied &&
1849             lustre_msg_get_status(req->rq_repmsg) != aa->praa_old_status) {
1850                 DEBUG_REQ(D_ERROR, req, "status %d, old was %d",
1851                           lustre_msg_get_status(req->rq_repmsg),
1852                           aa->praa_old_status);
1853         } else {
1854                 /* Put it back for re-replay. */
1855                 lustre_msg_set_status(req->rq_repmsg, aa->praa_old_status);
1856         }
1857
1858         /*
1859          * Errors while replay can set transno to 0, but
1860          * imp_last_replay_transno shouldn't be set to 0 anyway
1861          */
1862         if (req->rq_transno > 0) {
1863                 spin_lock(&imp->imp_lock);
1864                 LASSERT(req->rq_transno <= imp->imp_last_replay_transno);
1865                 imp->imp_last_replay_transno = req->rq_transno;
1866                 spin_unlock(&imp->imp_lock);
1867         } else
1868                 CERROR("Transno is 0 during replay!\n");
1869         /* continue with recovery */
1870         rc = ptlrpc_import_recovery_state_machine(imp);
1871  out:
1872         req->rq_send_state = aa->praa_old_state;
1873
1874         if (rc != 0)
1875                 /* this replay failed, so restart recovery */
1876                 ptlrpc_connect_import(imp, NULL);
1877
1878         RETURN(rc);
1879 }
1880
1881 int ptlrpc_replay_req(struct ptlrpc_request *req)
1882 {
1883         struct ptlrpc_replay_async_args *aa;
1884         ENTRY;
1885
1886         LASSERT(req->rq_import->imp_state == LUSTRE_IMP_REPLAY);
1887
1888         /* Not handling automatic bulk replay yet (or ever?) */
1889         LASSERT(req->rq_bulk == NULL);
1890
1891         DEBUG_REQ(D_HA, req, "REPLAY");
1892
1893         LASSERT (sizeof (*aa) <= sizeof (req->rq_async_args));
1894         aa = (struct ptlrpc_replay_async_args *)&req->rq_async_args;
1895         memset(aa, 0, sizeof *aa);
1896
1897         /* Prepare request to be resent with ptlrpcd */
1898         aa->praa_old_state = req->rq_send_state;
1899         req->rq_send_state = LUSTRE_IMP_REPLAY;
1900         req->rq_phase = RQ_PHASE_NEW;
1901         aa->praa_old_status = lustre_msg_get_status(req->rq_repmsg);
1902         req->rq_status = 0;
1903
1904         req->rq_interpret_reply = ptlrpc_replay_interpret;
1905         atomic_inc(&req->rq_import->imp_replay_inflight);
1906         ptlrpc_request_addref(req); /* ptlrpcd needs a ref */
1907
1908         ptlrpcd_add_req(req);
1909         RETURN(0);
1910 }
1911
1912 void ptlrpc_abort_inflight(struct obd_import *imp)
1913 {
1914         struct list_head *tmp, *n;
1915         ENTRY;
1916
1917         /* Make sure that no new requests get processed for this import.
1918          * ptlrpc_{queue,set}_wait must (and does) hold imp_lock while testing
1919          * this flag and then putting requests on sending_list or delayed_list.
1920          */
1921         spin_lock(&imp->imp_lock);
1922
1923         /* XXX locking?  Maybe we should remove each request with the list
1924          * locked?  Also, how do we know if the requests on the list are
1925          * being freed at this time?
1926          */
1927         list_for_each_safe(tmp, n, &imp->imp_sending_list) {
1928                 struct ptlrpc_request *req =
1929                         list_entry(tmp, struct ptlrpc_request, rq_list);
1930
1931                 DEBUG_REQ(D_RPCTRACE, req, "inflight");
1932
1933                 spin_lock (&req->rq_lock);
1934                 if (req->rq_import_generation < imp->imp_generation) {
1935                         req->rq_err = 1;
1936                         ptlrpc_wake_client_req(req);
1937                 }
1938                 spin_unlock (&req->rq_lock);
1939         }
1940
1941         list_for_each_safe(tmp, n, &imp->imp_delayed_list) {
1942                 struct ptlrpc_request *req =
1943                         list_entry(tmp, struct ptlrpc_request, rq_list);
1944
1945                 DEBUG_REQ(D_RPCTRACE, req, "aborting waiting req");
1946
1947                 spin_lock (&req->rq_lock);
1948                 if (req->rq_import_generation < imp->imp_generation) {
1949                         req->rq_err = 1;
1950                         ptlrpc_wake_client_req(req);
1951                 }
1952                 spin_unlock (&req->rq_lock);
1953         }
1954
1955         /* Last chance to free reqs left on the replay list, but we
1956          * will still leak reqs that haven't committed.  */
1957         if (imp->imp_replayable)
1958                 ptlrpc_free_committed(imp);
1959
1960         spin_unlock(&imp->imp_lock);
1961
1962         EXIT;
1963 }
1964
1965 static __u64 ptlrpc_last_xid = 0;
1966 spinlock_t ptlrpc_last_xid_lock;
1967
1968 __u64 ptlrpc_next_xid(void)
1969 {
1970         __u64 tmp;
1971         spin_lock(&ptlrpc_last_xid_lock);
1972         tmp = ++ptlrpc_last_xid;
1973         spin_unlock(&ptlrpc_last_xid_lock);
1974         return tmp;
1975 }
1976
1977 __u64 ptlrpc_sample_next_xid(void)
1978 {
1979         __u64 tmp;
1980         spin_lock(&ptlrpc_last_xid_lock);
1981         tmp = ptlrpc_last_xid + 1;
1982         spin_unlock(&ptlrpc_last_xid_lock);
1983         return tmp;
1984 }
1985 EXPORT_SYMBOL(ptlrpc_sample_next_xid);