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                 /* no point in adding already-committed requests to the replay
 646                  * list, we will just remove them immediately. b=9829 */
 647                 if (req->rq_transno != 0 &&
 648                     (req->rq_transno >
 649                      lustre_msg_get_last_committed(req->rq_repmsg) ||
 650                      req->rq_replay))
 651                         ptlrpc_retain_replayable_request(req, imp);
 652                 else if (req->rq_commit_cb != NULL) {
 653                         spin_unlock(&imp->imp_lock);
 654                         req->rq_commit_cb(req);
 655                         spin_lock(&imp->imp_lock);
 656                 }
 657
 658                 /* Replay-enabled imports return commit-status information. */
 659                 if (lustre_msg_get_last_committed(req->rq_repmsg))
 660                         imp->imp_peer_committed_transno =
 661                                 lustre_msg_get_last_committed(req->rq_repmsg);
 662                 ptlrpc_free_committed(imp);
 663                 spin_unlock(&imp->imp_lock);
 664         }
 665
 666         RETURN(rc);
 667 }
 668
 669 static int ptlrpc_send_new_req(struct ptlrpc_request *req)
 670 {
 671         struct obd_import     *imp;
 672         int rc;
 673         ENTRY;
 674
 675         LASSERT(req->rq_phase == RQ_PHASE_NEW);
 676         req->rq_phase = RQ_PHASE_RPC;
 677
 678         imp = req->rq_import;
 679         spin_lock(&imp->imp_lock);
 680
 681         req->rq_import_generation = imp->imp_generation;
 682
 683         if (ptlrpc_import_delay_req(imp, req, &rc)) {
 684                 spin_lock (&req->rq_lock);
 685                 req->rq_waiting = 1;
 686                 spin_unlock (&req->rq_lock);
 687
 688                 DEBUG_REQ(D_HA, req, "req from PID %d waiting for recovery: "
 689                           "(%s != %s)",
 690                           lustre_msg_get_status(req->rq_reqmsg) ,
 691                           ptlrpc_import_state_name(req->rq_send_state),
 692                           ptlrpc_import_state_name(imp->imp_state));
 693                 LASSERT(list_empty (&req->rq_list));
 694
 695                 list_add_tail(&req->rq_list, &imp->imp_delayed_list);
 696                 spin_unlock(&imp->imp_lock);
 697                 RETURN(0);
 698         }
 699
 700         if (rc != 0) {
 701                 spin_unlock(&imp->imp_lock);
 702                 req->rq_status = rc;
 703                 req->rq_phase = RQ_PHASE_INTERPRET;
 704                 RETURN(rc);
 705         }
 706
 707         /* XXX this is the same as ptlrpc_queue_wait */
 708         LASSERT(list_empty(&req->rq_list));
 709         list_add_tail(&req->rq_list, &imp->imp_sending_list);
 710         spin_unlock(&imp->imp_lock);
 711
 712         lustre_msg_set_status(req->rq_reqmsg, cfs_curproc_pid());
 713         CDEBUG(D_RPCTRACE, "Sending RPC pname:cluuid:pid:xid:nid:opc"
 714                " %s:%s:%d:"LPU64":%s:%d\n", cfs_curproc_comm(),
 715                imp->imp_obd->obd_uuid.uuid,
 716                lustre_msg_get_status(req->rq_reqmsg), req->rq_xid,
 717                libcfs_nid2str(imp->imp_connection->c_peer.nid),
 718                lustre_msg_get_opc(req->rq_reqmsg));
 719
 720         rc = ptl_send_rpc(req, 0);
 721         if (rc) {
 722                 DEBUG_REQ(D_HA, req, "send failed (%d); expect timeout", rc);
 723                 req->rq_net_err = 1;
 724                 RETURN(rc);
 725         }
 726         RETURN(0);
 727 }
 728
 729 /* this sends any unsent RPCs in @set and returns TRUE if all are sent */
 730 int ptlrpc_check_set(struct ptlrpc_request_set *set)
 731 {
 732         struct list_head *tmp;
 733         int force_timer_recalc = 0;
 734         ENTRY;
 735
 736         if (set->set_remaining == 0)
 737                 RETURN(1);
 738
 739         list_for_each(tmp, &set->set_requests) {
 740                 struct ptlrpc_request *req =
 741                         list_entry(tmp, struct ptlrpc_request, rq_set_chain);
 742                 struct obd_import *imp = req->rq_import;
 743                 int rc = 0;
 744
 745                 if (req->rq_phase == RQ_PHASE_NEW &&
 746                     ptlrpc_send_new_req(req)) {
 747                         force_timer_recalc = 1;
 748                 }
 749
 750                 if (!(req->rq_phase == RQ_PHASE_RPC ||
 751                       req->rq_phase == RQ_PHASE_BULK ||
 752                       req->rq_phase == RQ_PHASE_INTERPRET ||
 753                       req->rq_phase == RQ_PHASE_COMPLETE)) {
 754                         DEBUG_REQ(D_ERROR, req, "bad phase %x", req->rq_phase);
 755                         LBUG();
 756                 }
 757
 758                 if (req->rq_phase == RQ_PHASE_COMPLETE)
 759                         continue;
 760
 761                 if (req->rq_phase == RQ_PHASE_INTERPRET)
 762                         GOTO(interpret, req->rq_status);
 763
 764                 if (req->rq_net_err && !req->rq_timedout)
 765                         ptlrpc_expire_one_request(req);
 766
 767                 if (req->rq_err) {
 768                         ptlrpc_unregister_reply(req);
 769                         if (req->rq_status == 0)
 770                                 req->rq_status = -EIO;
 771                         req->rq_phase = RQ_PHASE_INTERPRET;
 772
 773                         spin_lock(&imp->imp_lock);
 774                         list_del_init(&req->rq_list);
 775                         spin_unlock(&imp->imp_lock);
 776
 777                         GOTO(interpret, req->rq_status);
 778                 }
 779
 780                 /* ptlrpc_queue_wait->l_wait_event guarantees that rq_intr
 781                  * will only be set after rq_timedout, but the oig waiting
 782                  * path sets rq_intr irrespective of whether ptlrpcd has
 783                  * seen a timeout.  our policy is to only interpret
 784                  * interrupted rpcs after they have timed out */
 785                 if (req->rq_intr && (req->rq_timedout || req->rq_waiting)) {
 786                         /* NB could be on delayed list */
 787                         ptlrpc_unregister_reply(req);
 788                         req->rq_status = -EINTR;
 789                         req->rq_phase = RQ_PHASE_INTERPRET;
 790
 791                         spin_lock(&imp->imp_lock);
 792                         list_del_init(&req->rq_list);
 793                         spin_unlock(&imp->imp_lock);
 794
 795                         GOTO(interpret, req->rq_status);
 796                 }
 797
 798                 if (req->rq_phase == RQ_PHASE_RPC) {
 799                         if (req->rq_timedout||req->rq_waiting||req->rq_resend) {
 800                                 int status;
 801
 802                                 ptlrpc_unregister_reply(req);
 803
 804                                 spin_lock(&imp->imp_lock);
 805
 806                                 if (ptlrpc_import_delay_req(imp, req, &status)){
 807                                         spin_unlock(&imp->imp_lock);
 808                                         continue;
 809                                 }
 810
 811                                 list_del_init(&req->rq_list);
 812                                 if (status != 0)  {
 813                                         req->rq_status = status;
 814                                         req->rq_phase = RQ_PHASE_INTERPRET;
 815                                         spin_unlock(&imp->imp_lock);
 816                                         GOTO(interpret, req->rq_status);
 817                                 }
 818                                 if (req->rq_no_resend) {
 819                                         req->rq_status = -ENOTCONN;
 820                                         req->rq_phase = RQ_PHASE_INTERPRET;
 821                                         spin_unlock(&imp->imp_lock);
 822                                         GOTO(interpret, req->rq_status);
 823                                 }
 824                                 list_add_tail(&req->rq_list,
 825                                               &imp->imp_sending_list);
 826
 827                                 spin_unlock(&imp->imp_lock);
 828
 829                                 req->rq_waiting = 0;
 830                                 if (req->rq_resend) {
 831                                         lustre_msg_add_flags(req->rq_reqmsg,
 832                                                              MSG_RESENT);
 833                                         if (req->rq_bulk) {
 834                                                 __u64 old_xid = req->rq_xid;
 835
 836                                                 ptlrpc_unregister_bulk (req);
 837
 838                                                 /* ensure previous bulk fails */
 839                                                 req->rq_xid = ptlrpc_next_xid();
 840                                                 CDEBUG(D_HA, "resend bulk "
 841                                                        "old x"LPU64
 842                                                        " new x"LPU64"\n",
 843                                                        old_xid, req->rq_xid);
 844                                         }
 845                                 }
 846
 847                                 rc = ptl_send_rpc(req, 0);
 848                                 if (rc) {
 849                                         DEBUG_REQ(D_HA, req, "send failed (%d)",
 850                                                   rc);
 851                                         force_timer_recalc = 1;
 852                                         req->rq_net_err = 1;
 853                                 }
 854                                 /* need to reset the timeout */
 855                                 force_timer_recalc = 1;
 856                         }
 857
 858                         /* Still waiting for a reply? */
 859                         if (ptlrpc_client_receiving_reply(req))
 860                                 continue;
 861
 862                         /* Did we actually receive a reply? */
 863                         if (!ptlrpc_client_replied(req))
 864                                 continue;
 865
 866                         spin_lock(&imp->imp_lock);
 867                         list_del_init(&req->rq_list);
 868                         spin_unlock(&imp->imp_lock);
 869
 870                         req->rq_status = after_reply(req);
 871                         if (req->rq_resend) {
 872                                 /* Add this req to the delayed list so
 873                                    it can be errored if the import is
 874                                    evicted after recovery. */
 875                                 spin_lock(&req->rq_lock);
 876                                 list_add_tail(&req->rq_list,
 877                                               &imp->imp_delayed_list);
 878                                 spin_unlock(&req->rq_lock);
 879                                 continue;
 880                         }
 881
 882                         /* If there is no bulk associated with this request,
 883                          * then we're done and should let the interpreter
 884                          * process the reply.  Similarly if the RPC returned
 885                          * an error, and therefore the bulk will never arrive.
 886                          */
 887                         if (req->rq_bulk == NULL || req->rq_status != 0) {
 888                                 req->rq_phase = RQ_PHASE_INTERPRET;
 889                                 GOTO(interpret, req->rq_status);
 890                         }
 891
 892                         req->rq_phase = RQ_PHASE_BULK;
 893                 }
 894
 895                 LASSERT(req->rq_phase == RQ_PHASE_BULK);
 896                 if (ptlrpc_bulk_active(req->rq_bulk))
 897                         continue;
 898
 899                 if (!req->rq_bulk->bd_success) {
 900                         /* The RPC reply arrived OK, but the bulk screwed
 901                          * up!  Dead wierd since the server told us the RPC
 902                          * was good after getting the REPLY for her GET or
 903                          * the ACK for her PUT. */
 904                         DEBUG_REQ(D_ERROR, req, "bulk transfer failed");
 905                         LBUG();
 906                 }
 907
 908                 req->rq_phase = RQ_PHASE_INTERPRET;
 909
 910         interpret:
 911                 LASSERT(req->rq_phase == RQ_PHASE_INTERPRET);
 912                 LASSERT(!req->rq_receiving_reply);
 913
 914                 ptlrpc_unregister_reply(req);
 915                 if (req->rq_bulk != NULL)
 916                         ptlrpc_unregister_bulk (req);
 917
 918                 req->rq_phase = RQ_PHASE_COMPLETE;
 919
 920                 if (req->rq_interpret_reply != NULL) {
 921                         int (*interpreter)(struct ptlrpc_request *,void *,int) =
 922                                 req->rq_interpret_reply;
 923                         req->rq_status = interpreter(req, &req->rq_async_args,
 924                                                      req->rq_status);
 925                 }
 926
 927                 CDEBUG(D_RPCTRACE, "Completed RPC pname:cluuid:pid:xid:nid:"
 928                        "opc %s:%s:%d:"LPU64":%s:%d\n", cfs_curproc_comm(),
 929                        imp->imp_obd->obd_uuid.uuid,
 930                        lustre_msg_get_status(req->rq_reqmsg), req->rq_xid,
 931                        libcfs_nid2str(imp->imp_connection->c_peer.nid),
 932                        lustre_msg_get_opc(req->rq_reqmsg));
 933
 934                 set->set_remaining--;
 935
 936                 atomic_dec(&imp->imp_inflight);
 937                 cfs_waitq_signal(&imp->imp_recovery_waitq);
 938         }
 939
 940         /* If we hit an error, we want to recover promptly. */
 941         RETURN(set->set_remaining == 0 || force_timer_recalc);
 942 }
 943
 944 int ptlrpc_expire_one_request(struct ptlrpc_request *req)
 945 {
 946         struct obd_import *imp = req->rq_import;
 947         int rc = 0;
 948         ENTRY;
 949
 950         DEBUG_REQ(D_ERROR|D_NETERROR, req, "timeout (sent at %lu, %lus ago)",
 951                   (long)req->rq_sent, CURRENT_SECONDS - req->rq_sent);
 952
 953         if (imp != NULL && obd_debug_peer_on_timeout)
 954                 LNetCtl(IOC_LIBCFS_DEBUG_PEER, &imp->imp_connection->c_peer);
 955
 956         spin_lock(&req->rq_lock);
 957         req->rq_timedout = 1;
 958         spin_unlock(&req->rq_lock);
 959
 960         ptlrpc_unregister_reply (req);
 961
 962         if (obd_dump_on_timeout)
 963                 libcfs_debug_dumplog();
 964
 965         if (req->rq_bulk != NULL)
 966                 ptlrpc_unregister_bulk (req);
 967
 968         if (imp == NULL) {
 969                 DEBUG_REQ(D_HA, req, "NULL import: already cleaned up?");
 970                 RETURN(1);
 971         }
 972
 973         /* The DLM server doesn't want recovery run on its imports. */
 974         if (imp->imp_dlm_fake)
 975                 RETURN(1);
 976
 977         /* If this request is for recovery or other primordial tasks,
 978          * then error it out here. */
 979         if (req->rq_send_state != LUSTRE_IMP_FULL ||
 980             imp->imp_obd->obd_no_recov) {
 981                 spin_lock(&req->rq_lock);
 982                 req->rq_status = -ETIMEDOUT;
 983                 req->rq_err = 1;
 984                 spin_unlock(&req->rq_lock);
 985                 RETURN(1);
 986         }
 987
 988         /* if request can't be resend we can't wait answer after timeout */
 989         if (req->rq_no_resend) {
 990                 DEBUG_REQ(D_RPCTRACE, req, "TIMEOUT-NORESEND:");
 991                 rc = 1;
 992         }
 993
 994         ptlrpc_fail_import(imp, lustre_msg_get_conn_cnt(req->rq_reqmsg));
 995
 996         RETURN(rc);
 997 }
 998
 999 int ptlrpc_expired_set(void *data)
1000 {
1001         struct ptlrpc_request_set *set = data;
1002         struct list_head          *tmp;
1003         time_t                     now = CURRENT_SECONDS;
1004         ENTRY;
1005
1006         LASSERT(set != NULL);
1007
1008         /* A timeout expired; see which reqs it applies to... */
1009         list_for_each (tmp, &set->set_requests) {
1010                 struct ptlrpc_request *req =
1011                         list_entry(tmp, struct ptlrpc_request, rq_set_chain);
1012
1013                 /* request in-flight? */
1014                 if (!((req->rq_phase == RQ_PHASE_RPC && !req->rq_waiting &&
1015                        !req->rq_resend) ||
1016                       (req->rq_phase == RQ_PHASE_BULK)))
1017                         continue;
1018
1019                 if (req->rq_timedout ||           /* already dealt with */
1020                     req->rq_sent + req->rq_timeout > now) /* not expired */
1021                         continue;
1022
1023                 /* deal with this guy */
1024                 ptlrpc_expire_one_request (req);
1025         }
1026
1027         /* When waiting for a whole set, we always to break out of the
1028          * sleep so we can recalculate the timeout, or enable interrupts
1029          * iff everyone's timed out.
1030          */
1031         RETURN(1);
1032 }
1033
1034 void ptlrpc_mark_interrupted(struct ptlrpc_request *req)
1035 {
1036         spin_lock(&req->rq_lock);
1037         req->rq_intr = 1;
1038         spin_unlock(&req->rq_lock);
1039 }
1040
1041 void ptlrpc_interrupted_set(void *data)
1042 {
1043         struct ptlrpc_request_set *set = data;
1044         struct list_head *tmp;
1045
1046         LASSERT(set != NULL);
1047         CERROR("INTERRUPTED SET %p\n", set);
1048
1049         list_for_each(tmp, &set->set_requests) {
1050                 struct ptlrpc_request *req =
1051                         list_entry(tmp, struct ptlrpc_request, rq_set_chain);
1052
1053                 if (req->rq_phase != RQ_PHASE_RPC)
1054                         continue;
1055
1056                 ptlrpc_mark_interrupted(req);
1057         }
1058 }
1059
1060 int ptlrpc_set_next_timeout(struct ptlrpc_request_set *set)
1061 {
1062         struct list_head      *tmp;
1063         time_t                 now = CURRENT_SECONDS;
1064         time_t                 deadline;
1065         int                    timeout = 0;
1066         struct ptlrpc_request *req;
1067         ENTRY;
1068
1069         SIGNAL_MASK_ASSERT(); /* XXX BUG 1511 */
1070
1071         list_for_each(tmp, &set->set_requests) {
1072                 req = list_entry(tmp, struct ptlrpc_request, rq_set_chain);
1073
1074                 /* request in-flight? */
1075                 if (!((req->rq_phase == RQ_PHASE_RPC && !req->rq_waiting) ||
1076                       (req->rq_phase == RQ_PHASE_BULK)))
1077                         continue;
1078
1079                 if (req->rq_timedout)   /* already timed out */
1080                         continue;
1081
1082                 deadline = req->rq_sent + req->rq_timeout;
1083                 if (deadline <= now)    /* actually expired already */
1084                         timeout = 1;    /* ASAP */
1085                 else if (timeout == 0 || timeout > deadline - now)
1086                         timeout = deadline - now;
1087         }
1088         RETURN(timeout);
1089 }
1090
1091 int ptlrpc_set_wait(struct ptlrpc_request_set *set)
1092 {
1093         struct list_head      *tmp;
1094         struct ptlrpc_request *req;
1095         struct l_wait_info     lwi;
1096         int                    rc, timeout;
1097         ENTRY;
1098
1099         if (list_empty(&set->set_requests))
1100                 RETURN(0);
1101
1102         list_for_each(tmp, &set->set_requests) {
1103                 req = list_entry(tmp, struct ptlrpc_request, rq_set_chain);
1104                 if (req->rq_phase == RQ_PHASE_NEW)
1105                         (void)ptlrpc_send_new_req(req);
1106         }
1107
1108         do {
1109                 timeout = ptlrpc_set_next_timeout(set);
1110
1111                 /* wait until all complete, interrupted, or an in-flight
1112                  * req times out */
1113                 CDEBUG(D_HA, "set %p going to sleep for %d seconds\n",
1114                        set, timeout);
1115                 lwi = LWI_TIMEOUT_INTR(cfs_time_seconds(timeout ? timeout : 1),
1116                                        ptlrpc_expired_set,
1117                                        ptlrpc_interrupted_set, set);
1118                 rc = l_wait_event(set->set_waitq, ptlrpc_check_set(set), &lwi);
1119
1120                 LASSERT(rc == 0 || rc == -EINTR || rc == -ETIMEDOUT);
1121
1122                 /* -EINTR => all requests have been flagged rq_intr so next
1123                  * check completes.
1124                  * -ETIMEOUTD => someone timed out.  When all reqs have
1125                  * timed out, signals are enabled allowing completion with
1126                  * EINTR.
1127                  * I don't really care if we go once more round the loop in
1128                  * the error cases -eeb. */
1129         } while (rc != 0 || set->set_remaining != 0);
1130
1131         LASSERT(set->set_remaining == 0);
1132
1133         rc = 0;
1134         list_for_each(tmp, &set->set_requests) {
1135                 req = list_entry(tmp, struct ptlrpc_request, rq_set_chain);
1136
1137                 LASSERT(req->rq_phase == RQ_PHASE_COMPLETE);
1138                 if (req->rq_status != 0)
1139                         rc = req->rq_status;
1140         }
1141
1142         if (set->set_interpret != NULL) {
1143                 int (*interpreter)(struct ptlrpc_request_set *set,void *,int) =
1144                         set->set_interpret;
1145                 rc = interpreter (set, set->set_arg, rc);
1146         }
1147
1148         RETURN(rc);
1149 }
1150
1151 static void __ptlrpc_free_req_to_pool(struct ptlrpc_request *request)
1152 {
1153         struct ptlrpc_request_pool *pool = request->rq_pool;
1154
1155         spin_lock(&pool->prp_lock);
1156         list_add_tail(&request->rq_list, &pool->prp_req_list);
1157         spin_unlock(&pool->prp_lock);
1158 }
1159
1160 static void __ptlrpc_free_req(struct ptlrpc_request *request, int locked)
1161 {
1162         ENTRY;
1163         if (request == NULL) {
1164                 EXIT;
1165                 return;
1166         }
1167
1168         LASSERTF(!request->rq_receiving_reply, "req %p\n", request);
1169         LASSERTF(request->rq_rqbd == NULL, "req %p\n",request);/* client-side */
1170         LASSERTF(list_empty(&request->rq_list), "req %p\n", request);
1171         LASSERTF(list_empty(&request->rq_set_chain), "req %p\n", request);
1172
1173         /* We must take it off the imp_replay_list first.  Otherwise, we'll set
1174          * request->rq_reqmsg to NULL while osc_close is dereferencing it. */
1175         if (request->rq_import != NULL) {
1176                 if (!locked)
1177                         spin_lock(&request->rq_import->imp_lock);
1178                 list_del_init(&request->rq_replay_list);
1179                 if (!locked)
1180                         spin_unlock(&request->rq_import->imp_lock);
1181         }
1182         LASSERTF(list_empty(&request->rq_replay_list), "req %p\n", request);
1183
1184         if (atomic_read(&request->rq_refcount) != 0) {
1185                 DEBUG_REQ(D_ERROR, request,
1186                           "freeing request with nonzero refcount");
1187                 LBUG();
1188         }
1189
1190         if (request->rq_repmsg != NULL) {
1191                 OBD_FREE(request->rq_repmsg, request->rq_replen);
1192                 request->rq_repmsg = NULL;
1193         }
1194         if (request->rq_export != NULL) {
1195                 class_export_put(request->rq_export);
1196                 request->rq_export = NULL;
1197         }
1198         if (request->rq_import != NULL) {
1199                 class_import_put(request->rq_import);
1200                 request->rq_import = NULL;
1201         }
1202         if (request->rq_bulk != NULL)
1203                 ptlrpc_free_bulk(request->rq_bulk);
1204
1205         if (request->rq_pool) {
1206                 __ptlrpc_free_req_to_pool(request);
1207         } else {
1208                 if (request->rq_reqmsg != NULL) {
1209                         OBD_FREE(request->rq_reqmsg, request->rq_reqlen);
1210                         request->rq_reqmsg = NULL;
1211                 }
1212                 OBD_FREE(request, sizeof(*request));
1213         }
1214         EXIT;
1215 }
1216
1217 void ptlrpc_free_req(struct ptlrpc_request *request)
1218 {
1219         __ptlrpc_free_req(request, 0);
1220 }
1221
1222 static int __ptlrpc_req_finished(struct ptlrpc_request *request, int locked);
1223 void ptlrpc_req_finished_with_imp_lock(struct ptlrpc_request *request)
1224 {
1225         LASSERT_SPIN_LOCKED(&request->rq_import->imp_lock);
1226         (void)__ptlrpc_req_finished(request, 1);
1227 }
1228
1229 static int __ptlrpc_req_finished(struct ptlrpc_request *request, int locked)
1230 {
1231         ENTRY;
1232         if (request == NULL)
1233                 RETURN(1);
1234
1235         if (request == LP_POISON ||
1236             request->rq_reqmsg == LP_POISON) {
1237                 CERROR("dereferencing freed request (bug 575)\n");
1238                 LBUG();
1239                 RETURN(1);
1240         }
1241
1242         DEBUG_REQ(D_INFO, request, "refcount now %u",
1243                   atomic_read(&request->rq_refcount) - 1);
1244
1245         if (atomic_dec_and_test(&request->rq_refcount)) {
1246                 __ptlrpc_free_req(request, locked);
1247                 RETURN(1);
1248         }
1249
1250         RETURN(0);
1251 }
1252
1253 void ptlrpc_req_finished(struct ptlrpc_request *request)
1254 {
1255         __ptlrpc_req_finished(request, 0);
1256 }
1257
1258 __u64 ptlrpc_req_xid(struct ptlrpc_request *request)
1259 {
1260         return request->rq_xid;
1261 }
1262 EXPORT_SYMBOL(ptlrpc_req_xid);
1263
1264 /* Disengage the client's reply buffer from the network
1265  * NB does _NOT_ unregister any client-side bulk.
1266  * IDEMPOTENT, but _not_ safe against concurrent callers.
1267  * The request owner (i.e. the thread doing the I/O) must call...
1268  */
1269 void ptlrpc_unregister_reply (struct ptlrpc_request *request)
1270 {
1271         int                rc;
1272         cfs_waitq_t       *wq;
1273         struct l_wait_info lwi;
1274         ENTRY;
1275
1276         LASSERT(!in_interrupt ());             /* might sleep */
1277
1278         if (!ptlrpc_client_receiving_reply(request))
1279                 return;
1280
1281         LNetMDUnlink (request->rq_reply_md_h);
1282
1283         /* We have to l_wait_event() whatever the result, to give liblustre
1284          * a chance to run reply_in_callback() */
1285
1286         if (request->rq_set != NULL)
1287                 wq = &request->rq_set->set_waitq;
1288         else
1289                 wq = &request->rq_reply_waitq;
1290
1291         for (;;) {
1292                 /* Network access will complete in finite time but the HUGE
1293                  * timeout lets us CWARN for visibility of sluggish NALs */
1294                 lwi = LWI_TIMEOUT(cfs_time_seconds(300), NULL, NULL);
1295                 rc = l_wait_event (*wq, !ptlrpc_client_receiving_reply(request), &lwi);
1296                 if (rc == 0)
1297                         return;
1298
1299                 LASSERT (rc == -ETIMEDOUT);
1300                 DEBUG_REQ(D_WARNING, request, "Unexpectedly long timeout");
1301         }
1302         EXIT;
1303 }
1304
1305 /* caller must hold imp->imp_lock */
1306 void ptlrpc_free_committed(struct obd_import *imp)
1307 {
1308         struct list_head *tmp, *saved;
1309         struct ptlrpc_request *req;
1310         struct ptlrpc_request *last_req = NULL; /* temporary fire escape */
1311         ENTRY;
1312
1313         LASSERT(imp != NULL);
1314
1315         LASSERT_SPIN_LOCKED(&imp->imp_lock);
1316
1317
1318         if (imp->imp_peer_committed_transno == imp->imp_last_transno_checked &&
1319             imp->imp_generation == imp->imp_last_generation_checked) {
1320                 CDEBUG(D_HA, "%s: skip recheck for last_committed "LPU64"\n",
1321                        imp->imp_obd->obd_name, imp->imp_peer_committed_transno);
1322                 return;
1323         }
1324
1325         CDEBUG(D_HA, "%s: committing for last_committed "LPU64" gen %d\n",
1326                imp->imp_obd->obd_name, imp->imp_peer_committed_transno,
1327                imp->imp_generation);
1328         imp->imp_last_transno_checked = imp->imp_peer_committed_transno;
1329         imp->imp_last_generation_checked = imp->imp_generation;
1330
1331         list_for_each_safe(tmp, saved, &imp->imp_replay_list) {
1332                 req = list_entry(tmp, struct ptlrpc_request, rq_replay_list);
1333
1334                 /* XXX ok to remove when 1357 resolved - rread 05/29/03  */
1335                 LASSERT(req != last_req);
1336                 last_req = req;
1337
1338                 if (req->rq_import_generation < imp->imp_generation) {
1339                         DEBUG_REQ(D_HA, req, "freeing request with old gen");
1340                         GOTO(free_req, 0);
1341                 }
1342
1343                 if (req->rq_replay) {
1344                         DEBUG_REQ(D_HA, req, "keeping (FL_REPLAY)");
1345                         continue;
1346                 }
1347
1348                 /* not yet committed */
1349                 if (req->rq_transno > imp->imp_peer_committed_transno) {
1350                         DEBUG_REQ(D_HA, req, "stopping search");
1351                         break;
1352                 }
1353
1354                 DEBUG_REQ(D_HA, req, "committing (last_committed "LPU64")",
1355                           imp->imp_peer_committed_transno);
1356 free_req:
1357                 if (req->rq_commit_cb != NULL)
1358                         req->rq_commit_cb(req);
1359                 list_del_init(&req->rq_replay_list);
1360                 __ptlrpc_req_finished(req, 1);
1361         }
1362
1363         EXIT;
1364         return;
1365 }
1366
1367 void ptlrpc_cleanup_client(struct obd_import *imp)
1368 {
1369         ENTRY;
1370         EXIT;
1371         return;
1372 }
1373
1374 void ptlrpc_resend_req(struct ptlrpc_request *req)
1375 {
1376         DEBUG_REQ(D_HA, req, "going to resend");
1377         lustre_msg_set_handle(req->rq_reqmsg, &(struct lustre_handle){ 0 });
1378         req->rq_status = -EAGAIN;
1379
1380         spin_lock(&req->rq_lock);
1381         req->rq_resend = 1;
1382         req->rq_net_err = 0;
1383         req->rq_timedout = 0;
1384         if (req->rq_bulk) {
1385                 __u64 old_xid = req->rq_xid;
1386
1387                 /* ensure previous bulk fails */
1388                 req->rq_xid = ptlrpc_next_xid();
1389                 CDEBUG(D_HA, "resend bulk old x"LPU64" new x"LPU64"\n",
1390                        old_xid, req->rq_xid);
1391         }
1392         ptlrpc_wake_client_req(req);
1393         spin_unlock(&req->rq_lock);
1394 }
1395
1396 /* XXX: this function and rq_status are currently unused */
1397 void ptlrpc_restart_req(struct ptlrpc_request *req)
1398 {
1399         DEBUG_REQ(D_HA, req, "restarting (possibly-)completed request");
1400         req->rq_status = -ERESTARTSYS;
1401
1402         spin_lock(&req->rq_lock);
1403         req->rq_restart = 1;
1404         req->rq_timedout = 0;
1405         ptlrpc_wake_client_req(req);
1406         spin_unlock(&req->rq_lock);
1407 }
1408
1409 static int expired_request(void *data)
1410 {
1411         struct ptlrpc_request *req = data;
1412         ENTRY;
1413
1414         RETURN(ptlrpc_expire_one_request(req));
1415 }
1416
1417 static void interrupted_request(void *data)
1418 {
1419         struct ptlrpc_request *req = data;
1420         DEBUG_REQ(D_HA, req, "request interrupted");
1421         spin_lock(&req->rq_lock);
1422         req->rq_intr = 1;
1423         spin_unlock(&req->rq_lock);
1424 }
1425
1426 struct ptlrpc_request *ptlrpc_request_addref(struct ptlrpc_request *req)
1427 {
1428         ENTRY;
1429         atomic_inc(&req->rq_refcount);
1430         RETURN(req);
1431 }
1432
1433 void ptlrpc_retain_replayable_request(struct ptlrpc_request *req,
1434                                       struct obd_import *imp)
1435 {
1436         struct list_head *tmp;
1437
1438         LASSERT_SPIN_LOCKED(&imp->imp_lock);
1439
1440         /* clear this for new requests that were resent as well
1441            as resent replayed requests. */
1442         lustre_msg_clear_flags(req->rq_reqmsg, MSG_RESENT);
1443
1444         /* don't re-add requests that have been replayed */
1445         if (!list_empty(&req->rq_replay_list))
1446                 return;
1447
1448         lustre_msg_add_flags(req->rq_reqmsg, MSG_REPLAY);
1449
1450         LASSERT(imp->imp_replayable);
1451         /* Balanced in ptlrpc_free_committed, usually. */
1452         ptlrpc_request_addref(req);
1453         list_for_each_prev(tmp, &imp->imp_replay_list) {
1454                 struct ptlrpc_request *iter =
1455                         list_entry(tmp, struct ptlrpc_request, rq_replay_list);
1456
1457                 /* We may have duplicate transnos if we create and then
1458                  * open a file, or for closes retained if to match creating
1459                  * opens, so use req->rq_xid as a secondary key.
1460                  * (See bugs 684, 685, and 428.)
1461                  * XXX no longer needed, but all opens need transnos!
1462                  */
1463                 if (iter->rq_transno > req->rq_transno)
1464                         continue;
1465
1466                 if (iter->rq_transno == req->rq_transno) {
1467                         LASSERT(iter->rq_xid != req->rq_xid);
1468                         if (iter->rq_xid > req->rq_xid)
1469                                 continue;
1470                 }
1471
1472                 list_add(&req->rq_replay_list, &iter->rq_replay_list);
1473                 return;
1474         }
1475
1476         list_add_tail(&req->rq_replay_list, &imp->imp_replay_list);
1477 }
1478
1479 int ptlrpc_queue_wait(struct ptlrpc_request *req)
1480 {
1481         int rc = 0;
1482         int brc;
1483         struct l_wait_info lwi;
1484         struct obd_import *imp = req->rq_import;
1485         cfs_duration_t timeout = 0;
1486         ENTRY;
1487
1488         LASSERT(req->rq_set == NULL);
1489         LASSERT(!req->rq_receiving_reply);
1490         atomic_inc(&imp->imp_inflight);
1491
1492         /* for distributed debugging */
1493         lustre_msg_set_status(req->rq_reqmsg, cfs_curproc_pid());
1494         LASSERT(imp->imp_obd != NULL);
1495         CDEBUG(D_RPCTRACE, "Sending RPC pname:cluuid:pid:xid:nid:opc "
1496                "%s:%s:%d:"LPU64":%s:%d\n", cfs_curproc_comm(),
1497                imp->imp_obd->obd_uuid.uuid,
1498                lustre_msg_get_status(req->rq_reqmsg), req->rq_xid,
1499                libcfs_nid2str(imp->imp_connection->c_peer.nid),
1500                lustre_msg_get_opc(req->rq_reqmsg));
1501
1502         /* Mark phase here for a little debug help */
1503         req->rq_phase = RQ_PHASE_RPC;
1504
1505         spin_lock(&imp->imp_lock);
1506         req->rq_import_generation = imp->imp_generation;
1507 restart:
1508         if (ptlrpc_import_delay_req(imp, req, &rc)) {
1509                 list_del(&req->rq_list);
1510
1511                 list_add_tail(&req->rq_list, &imp->imp_delayed_list);
1512                 spin_unlock(&imp->imp_lock);
1513
1514                 DEBUG_REQ(D_HA, req, "\"%s\" waiting for recovery: (%s != %s)",
1515                           cfs_curproc_comm(),
1516                           ptlrpc_import_state_name(req->rq_send_state),
1517                           ptlrpc_import_state_name(imp->imp_state));
1518                 lwi = LWI_INTR(interrupted_request, req);
1519                 rc = l_wait_event(req->rq_reply_waitq,
1520                                   (req->rq_send_state == imp->imp_state ||
1521                                    req->rq_err || req->rq_intr),
1522                                   &lwi);
1523                 DEBUG_REQ(D_HA, req, "\"%s\" awake: (%s == %s or %d/%d == 1)",
1524                           cfs_curproc_comm(),
1525                           ptlrpc_import_state_name(imp->imp_state),
1526                           ptlrpc_import_state_name(req->rq_send_state),
1527                           req->rq_err, req->rq_intr);
1528
1529                 spin_lock(&imp->imp_lock);
1530                 list_del_init(&req->rq_list);
1531
1532                 if (req->rq_err) {
1533                         rc = -EIO;
1534                 }
1535                 else if (req->rq_intr) {
1536                         rc = -EINTR;
1537                 }
1538                 else if (req->rq_no_resend) {
1539                         spin_unlock(&imp->imp_lock);
1540                         GOTO(out, rc = -ETIMEDOUT);
1541                 }
1542                 else {
1543                         GOTO(restart, rc);
1544                 }
1545         }
1546
1547         if (rc != 0) {
1548                 list_del_init(&req->rq_list);
1549                 spin_unlock(&imp->imp_lock);
1550                 req->rq_status = rc; // XXX this ok?
1551                 GOTO(out, rc);
1552         }
1553
1554         if (req->rq_resend) {
1555                 lustre_msg_add_flags(req->rq_reqmsg, MSG_RESENT);
1556
1557                 if (req->rq_bulk != NULL) {
1558                         ptlrpc_unregister_bulk (req);
1559
1560                         /* bulk requests are supposed to be
1561                          * idempotent, so we are free to bump the xid
1562                          * here, which we need to do before
1563                          * registering the bulk again (bug 6371).
1564                          * print the old xid first for sanity.
1565                          */
1566                         DEBUG_REQ(D_HA, req, "bumping xid for bulk: ");
1567                         req->rq_xid = ptlrpc_next_xid();
1568                 }
1569
1570                 DEBUG_REQ(D_HA, req, "resending: ");
1571         }
1572
1573         /* XXX this is the same as ptlrpc_set_wait */
1574         LASSERT(list_empty(&req->rq_list));
1575         list_add_tail(&req->rq_list, &imp->imp_sending_list);
1576         spin_unlock(&imp->imp_lock);
1577
1578         rc = ptl_send_rpc(req, 0);
1579         if (rc) {
1580                 DEBUG_REQ(D_HA, req, "send failed (%d); recovering", rc);
1581                 timeout = CFS_TICK;
1582         } else {
1583                 timeout = cfs_timeout_cap(cfs_time_seconds(req->rq_timeout));
1584                 DEBUG_REQ(D_NET, req,
1585                           "-- sleeping for "CFS_DURATION_T" jiffies", timeout);
1586         }
1587         lwi = LWI_TIMEOUT_INTR(timeout, expired_request, interrupted_request,
1588                                req);
1589         l_wait_event(req->rq_reply_waitq, ptlrpc_check_reply(req), &lwi);
1590
1591         CDEBUG(D_RPCTRACE, "Completed RPC pname:cluuid:pid:xid:nid:opc "
1592                "%s:%s:%d:"LPU64":%s:%d\n", cfs_curproc_comm(),
1593                imp->imp_obd->obd_uuid.uuid,
1594                lustre_msg_get_status(req->rq_reqmsg), req->rq_xid,
1595                libcfs_nid2str(imp->imp_connection->c_peer.nid),
1596                lustre_msg_get_opc(req->rq_reqmsg));
1597
1598         spin_lock(&imp->imp_lock);
1599         list_del_init(&req->rq_list);
1600         spin_unlock(&imp->imp_lock);
1601
1602         /* If the reply was received normally, this just grabs the spinlock
1603          * (ensuring the reply callback has returned), sees that
1604          * req->rq_receiving_reply is clear and returns. */
1605         ptlrpc_unregister_reply (req);
1606
1607         if (req->rq_err)
1608                 GOTO(out, rc = -EIO);
1609
1610         /* Resend if we need to, unless we were interrupted. */
1611         if (req->rq_resend && !req->rq_intr) {
1612                 /* ...unless we were specifically told otherwise. */
1613                 if (req->rq_no_resend)
1614                         GOTO(out, rc = -ETIMEDOUT);
1615                 spin_lock(&imp->imp_lock);
1616                 goto restart;
1617         }
1618
1619         if (req->rq_intr) {
1620                 /* Should only be interrupted if we timed out. */
1621                 if (!req->rq_timedout)
1622                         DEBUG_REQ(D_ERROR, req,
1623                                   "rq_intr set but rq_timedout not");
1624                 GOTO(out, rc = -EINTR);
1625         }
1626
1627         if (req->rq_timedout) {                 /* non-recoverable timeout */
1628                 GOTO(out, rc = -ETIMEDOUT);
1629         }
1630
1631         if (!req->rq_replied) {
1632                 /* How can this be? -eeb */
1633                 DEBUG_REQ(D_ERROR, req, "!rq_replied: ");
1634                 LBUG();
1635                 GOTO(out, rc = req->rq_status);
1636         }
1637
1638         rc = after_reply(req);
1639         /* NB may return +ve success rc */
1640         if (req->rq_resend) {
1641                 spin_lock(&imp->imp_lock);
1642                 goto restart;
1643         }
1644
1645  out:
1646         if (req->rq_bulk != NULL) {
1647                 if (rc >= 0) {
1648                         /* success so far.  Note that anything going wrong
1649                          * with bulk now, is EXTREMELY strange, since the
1650                          * server must have believed that the bulk
1651                          * tranferred OK before she replied with success to
1652                          * me. */
1653                         lwi = LWI_TIMEOUT(timeout, NULL, NULL);
1654                         brc = l_wait_event(req->rq_reply_waitq,
1655                                            !ptlrpc_bulk_active(req->rq_bulk),
1656                                            &lwi);
1657                         LASSERT(brc == 0 || brc == -ETIMEDOUT);
1658                         if (brc != 0) {
1659                                 LASSERT(brc == -ETIMEDOUT);
1660                                 DEBUG_REQ(D_ERROR, req, "bulk timed out");
1661                                 rc = brc;
1662                         } else if (!req->rq_bulk->bd_success) {
1663                                 DEBUG_REQ(D_ERROR, req, "bulk transfer failed");
1664                                 rc = -EIO;
1665                         }
1666                 }
1667                 if (rc < 0)
1668                         ptlrpc_unregister_bulk (req);
1669         }
1670
1671         LASSERT(!req->rq_receiving_reply);
1672         req->rq_phase = RQ_PHASE_INTERPRET;
1673
1674         atomic_dec(&imp->imp_inflight);
1675         cfs_waitq_signal(&imp->imp_recovery_waitq);
1676         RETURN(rc);
1677 }
1678
1679 struct ptlrpc_replay_async_args {
1680         int praa_old_state;
1681         int praa_old_status;
1682 };
1683
1684 static int ptlrpc_replay_interpret(struct ptlrpc_request *req,
1685                                     void * data, int rc)
1686 {
1687         struct ptlrpc_replay_async_args *aa = data;
1688         struct obd_import *imp = req->rq_import;
1689
1690         ENTRY;
1691         atomic_dec(&imp->imp_replay_inflight);
1692
1693         if (!req->rq_replied) {
1694                 CERROR("request replay timed out, restarting recovery\n");
1695                 GOTO(out, rc = -ETIMEDOUT);
1696         }
1697
1698         if (lustre_msg_get_type(req->rq_repmsg) == PTL_RPC_MSG_ERR &&
1699             lustre_msg_get_status(req->rq_repmsg) == -ENOTCONN)
1700                 GOTO(out, rc = lustre_msg_get_status(req->rq_repmsg));
1701
1702         /* The transno had better not change over replay. */
1703         LASSERT(lustre_msg_get_transno(req->rq_reqmsg) ==
1704                 lustre_msg_get_transno(req->rq_repmsg));
1705
1706         DEBUG_REQ(D_HA, req, "got rep");
1707
1708         /* let the callback do fixups, possibly including in the request */
1709         if (req->rq_replay_cb)
1710                 req->rq_replay_cb(req);
1711
1712         if (req->rq_replied &&
1713             lustre_msg_get_status(req->rq_repmsg) != aa->praa_old_status) {
1714                 DEBUG_REQ(D_ERROR, req, "status %d, old was %d",
1715                           lustre_msg_get_status(req->rq_repmsg),
1716                           aa->praa_old_status);
1717         } else {
1718                 /* Put it back for re-replay. */
1719                 lustre_msg_set_status(req->rq_repmsg, aa->praa_old_status);
1720         }
1721
1722         spin_lock(&imp->imp_lock);
1723         imp->imp_last_replay_transno = req->rq_transno;
1724         spin_unlock(&imp->imp_lock);
1725
1726         /* continue with recovery */
1727         rc = ptlrpc_import_recovery_state_machine(imp);
1728  out:
1729         req->rq_send_state = aa->praa_old_state;
1730
1731         if (rc != 0)
1732                 /* this replay failed, so restart recovery */
1733                 ptlrpc_connect_import(imp, NULL);
1734
1735         RETURN(rc);
1736 }
1737
1738
1739 int ptlrpc_replay_req(struct ptlrpc_request *req)
1740 {
1741         struct ptlrpc_replay_async_args *aa;
1742         ENTRY;
1743
1744         LASSERT(req->rq_import->imp_state == LUSTRE_IMP_REPLAY);
1745
1746         /* Not handling automatic bulk replay yet (or ever?) */
1747         LASSERT(req->rq_bulk == NULL);
1748
1749         DEBUG_REQ(D_HA, req, "REPLAY");
1750
1751         LASSERT (sizeof (*aa) <= sizeof (req->rq_async_args));
1752         aa = (struct ptlrpc_replay_async_args *)&req->rq_async_args;
1753         memset(aa, 0, sizeof *aa);
1754
1755         /* Prepare request to be resent with ptlrpcd */
1756         aa->praa_old_state = req->rq_send_state;
1757         req->rq_send_state = LUSTRE_IMP_REPLAY;
1758         req->rq_phase = RQ_PHASE_NEW;
1759         aa->praa_old_status = lustre_msg_get_status(req->rq_repmsg);
1760         req->rq_status = 0;
1761
1762         req->rq_interpret_reply = ptlrpc_replay_interpret;
1763         atomic_inc(&req->rq_import->imp_replay_inflight);
1764         ptlrpc_request_addref(req); /* ptlrpcd needs a ref */
1765
1766         ptlrpcd_add_req(req);
1767         RETURN(0);
1768 }
1769
1770 void ptlrpc_abort_inflight(struct obd_import *imp)
1771 {
1772         struct list_head *tmp, *n;
1773         ENTRY;
1774
1775         /* Make sure that no new requests get processed for this import.
1776          * ptlrpc_{queue,set}_wait must (and does) hold imp_lock while testing
1777          * this flag and then putting requests on sending_list or delayed_list.
1778          */
1779         spin_lock(&imp->imp_lock);
1780
1781         /* XXX locking?  Maybe we should remove each request with the list
1782          * locked?  Also, how do we know if the requests on the list are
1783          * being freed at this time?
1784          */
1785         list_for_each_safe(tmp, n, &imp->imp_sending_list) {
1786                 struct ptlrpc_request *req =
1787                         list_entry(tmp, struct ptlrpc_request, rq_list);
1788
1789                 DEBUG_REQ(D_HA, req, "inflight");
1790
1791                 spin_lock (&req->rq_lock);
1792                 if (req->rq_import_generation < imp->imp_generation) {
1793                         req->rq_err = 1;
1794                         ptlrpc_wake_client_req(req);
1795                 }
1796                 spin_unlock (&req->rq_lock);
1797         }
1798
1799         list_for_each_safe(tmp, n, &imp->imp_delayed_list) {
1800                 struct ptlrpc_request *req =
1801                         list_entry(tmp, struct ptlrpc_request, rq_list);
1802
1803                 DEBUG_REQ(D_HA, req, "aborting waiting req");
1804
1805                 spin_lock (&req->rq_lock);
1806                 if (req->rq_import_generation < imp->imp_generation) {
1807                         req->rq_err = 1;
1808                         ptlrpc_wake_client_req(req);
1809                 }
1810                 spin_unlock (&req->rq_lock);
1811         }
1812
1813         /* Last chance to free reqs left on the replay list, but we
1814          * will still leak reqs that haven't committed.  */
1815         if (imp->imp_replayable)
1816                 ptlrpc_free_committed(imp);
1817
1818         spin_unlock(&imp->imp_lock);
1819
1820         EXIT;
1821 }
1822
1823 static __u64 ptlrpc_last_xid = 0;
1824 spinlock_t ptlrpc_last_xid_lock;
1825
1826 __u64 ptlrpc_next_xid(void)
1827 {
1828         __u64 tmp;
1829         spin_lock(&ptlrpc_last_xid_lock);
1830         tmp = ++ptlrpc_last_xid;
1831         spin_unlock(&ptlrpc_last_xid_lock);
1832         return tmp;
1833 }
1834
1835 __u64 ptlrpc_sample_next_xid(void)
1836 {
1837         __u64 tmp;
1838         spin_lock(&ptlrpc_last_xid_lock);
1839         tmp = ptlrpc_last_xid + 1;
1840         spin_unlock(&ptlrpc_last_xid_lock);
1841         return tmp;
1842 }
1843 EXPORT_SYMBOL(ptlrpc_sample_next_xid);