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