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 #include <lustre_req_layout.h>
  39
  40 #include "ptlrpc_internal.h"
  41
  42 void ptlrpc_init_client(int req_portal, int rep_portal, char *name,
  43                         struct ptlrpc_client *cl)
  44 {
  45         cl->cli_request_portal = req_portal;
  46         cl->cli_reply_portal   = rep_portal;
  47         cl->cli_name           = name;
  48 }
  49
  50 struct ptlrpc_connection *ptlrpc_uuid_to_connection(struct obd_uuid *uuid)
  51 {
  52         struct ptlrpc_connection *c;
  53         lnet_nid_t                self;
  54         lnet_process_id_t         peer;
  55         int                       err;
  56
  57         err = ptlrpc_uuid_to_peer(uuid, &peer, &self);
  58         if (err != 0) {
  59                 CERROR("cannot find peer %s!\n", uuid->uuid);
  60                 return NULL;
  61         }
  62
  63         c = ptlrpc_get_connection(peer, self, uuid);
  64         if (c) {
  65                 memcpy(c->c_remote_uuid.uuid,
  66                        uuid->uuid, sizeof(c->c_remote_uuid.uuid));
  67         }
  68
  69         CDEBUG(D_INFO, "%s -> %p\n", uuid->uuid, c);
  70
  71         return c;
  72 }
  73
  74 void ptlrpc_readdress_connection(struct ptlrpc_connection *conn,
  75                                  struct obd_uuid *uuid)
  76 {
  77         lnet_nid_t        self;
  78         lnet_process_id_t peer;
  79         int               err;
  80
  81         err = ptlrpc_uuid_to_peer(uuid, &peer, &self);
  82         if (err != 0) {
  83                 CERROR("cannot find peer %s!\n", uuid->uuid);
  84                 return;
  85         }
  86
  87         conn->c_peer = peer;
  88         conn->c_self = self;
  89         return;
  90 }
  91
  92 static inline struct ptlrpc_bulk_desc *new_bulk(int npages, int type, int portal)
  93 {
  94         struct ptlrpc_bulk_desc *desc;
  95
  96         OBD_ALLOC(desc, offsetof (struct ptlrpc_bulk_desc, bd_iov[npages]));
  97         if (!desc)
  98                 return NULL;
  99
 100         spin_lock_init(&desc->bd_lock);
 101         cfs_waitq_init(&desc->bd_waitq);
 102         desc->bd_max_iov = npages;
 103         desc->bd_iov_count = 0;
 104         desc->bd_md_h = LNET_INVALID_HANDLE;
 105         desc->bd_portal = portal;
 106         desc->bd_type = type;
 107
 108         return desc;
 109 }
 110
 111 struct ptlrpc_bulk_desc *ptlrpc_prep_bulk_imp (struct ptlrpc_request *req,
 112                                                int npages, int type, int portal)
 113 {
 114         struct obd_import *imp = req->rq_import;
 115         struct ptlrpc_bulk_desc *desc;
 116
 117         ENTRY;
 118         LASSERT(type == BULK_PUT_SINK || type == BULK_GET_SOURCE);
 119         desc = new_bulk(npages, type, portal);
 120         if (desc == NULL)
 121                 RETURN(NULL);
 122
 123         desc->bd_import_generation = req->rq_import_generation;
 124         desc->bd_import = class_import_get(imp);
 125         desc->bd_req = req;
 126
 127         desc->bd_cbid.cbid_fn  = client_bulk_callback;
 128         desc->bd_cbid.cbid_arg = desc;
 129
 130         /* This makes req own desc, and free it when she frees herself */
 131         req->rq_bulk = desc;
 132
 133         return desc;
 134 }
 135
 136 struct ptlrpc_bulk_desc *ptlrpc_prep_bulk_exp(struct ptlrpc_request *req,
 137                                               int npages, int type, int portal)
 138 {
 139         struct obd_export *exp = req->rq_export;
 140         struct ptlrpc_bulk_desc *desc;
 141
 142         ENTRY;
 143         LASSERT(type == BULK_PUT_SOURCE || type == BULK_GET_SINK);
 144
 145         desc = new_bulk(npages, type, portal);
 146         if (desc == NULL)
 147                 RETURN(NULL);
 148
 149         desc->bd_export = class_export_get(exp);
 150         desc->bd_req = req;
 151
 152         desc->bd_cbid.cbid_fn  = server_bulk_callback;
 153         desc->bd_cbid.cbid_arg = desc;
 154
 155         /* NB we don't assign rq_bulk here; server-side requests are
 156          * re-used, and the handler frees the bulk desc explicitly. */
 157
 158         return desc;
 159 }
 160
 161 void ptlrpc_prep_bulk_page(struct ptlrpc_bulk_desc *desc,
 162                            cfs_page_t *page, int pageoffset, int len)
 163 {
 164         LASSERT(desc->bd_iov_count < desc->bd_max_iov);
 165         LASSERT(page != NULL);
 166         LASSERT(pageoffset >= 0);
 167         LASSERT(len > 0);
 168         LASSERT(pageoffset + len <= CFS_PAGE_SIZE);
 169
 170         desc->bd_nob += len;
 171
 172         ptlrpc_add_bulk_page(desc, page, pageoffset, len);
 173 }
 174
 175 void ptlrpc_free_bulk(struct ptlrpc_bulk_desc *desc)
 176 {
 177         ENTRY;
 178
 179         LASSERT(desc != NULL);
 180         LASSERT(desc->bd_iov_count != LI_POISON); /* not freed already */
 181         LASSERT(!desc->bd_network_rw);         /* network hands off or */
 182         LASSERT((desc->bd_export != NULL) ^ (desc->bd_import != NULL));
 183
 184         sptlrpc_enc_pool_put_pages(desc);
 185
 186         if (desc->bd_export)
 187                 class_export_put(desc->bd_export);
 188         else
 189                 class_import_put(desc->bd_import);
 190
 191         OBD_FREE(desc, offsetof(struct ptlrpc_bulk_desc,
 192                                 bd_iov[desc->bd_max_iov]));
 193         EXIT;
 194 }
 195
 196 void ptlrpc_free_rq_pool(struct ptlrpc_request_pool *pool)
 197 {
 198         struct list_head *l, *tmp;
 199         struct ptlrpc_request *req;
 200
 201         if (!pool)
 202                 return;
 203
 204         list_for_each_safe(l, tmp, &pool->prp_req_list) {
 205                 req = list_entry(l, struct ptlrpc_request, rq_list);
 206                 list_del(&req->rq_list);
 207                 LASSERT(req->rq_reqbuf);
 208                 LASSERT(req->rq_reqbuf_len == pool->prp_rq_size);
 209                 OBD_FREE(req->rq_reqbuf, pool->prp_rq_size);
 210                 OBD_FREE(req, sizeof(*req));
 211         }
 212         OBD_FREE(pool, sizeof(*pool));
 213 }
 214
 215 void ptlrpc_add_rqs_to_pool(struct ptlrpc_request_pool *pool, int num_rq)
 216 {
 217         int i;
 218         int size = 1;
 219
 220         while (size < pool->prp_rq_size + SPTLRPC_MAX_PAYLOAD)
 221                 size <<= 1;
 222
 223         LASSERTF(list_empty(&pool->prp_req_list) || size == pool->prp_rq_size,
 224                  "Trying to change pool size with nonempty pool "
 225                  "from %d to %d bytes\n", pool->prp_rq_size, size);
 226
 227         spin_lock(&pool->prp_lock);
 228         pool->prp_rq_size = size;
 229         for (i = 0; i < num_rq; i++) {
 230                 struct ptlrpc_request *req;
 231                 struct lustre_msg *msg;
 232
 233                 spin_unlock(&pool->prp_lock);
 234                 OBD_ALLOC(req, sizeof(struct ptlrpc_request));
 235                 if (!req)
 236                         return;
 237                 OBD_ALLOC_GFP(msg, size, CFS_ALLOC_STD);
 238                 if (!msg) {
 239                         OBD_FREE(req, sizeof(struct ptlrpc_request));
 240                         return;
 241                 }
 242                 req->rq_reqbuf = msg;
 243                 req->rq_reqbuf_len = size;
 244                 req->rq_pool = pool;
 245                 spin_lock(&pool->prp_lock);
 246                 list_add_tail(&req->rq_list, &pool->prp_req_list);
 247         }
 248         spin_unlock(&pool->prp_lock);
 249         return;
 250 }
 251
 252 struct ptlrpc_request_pool *ptlrpc_init_rq_pool(int num_rq, int msgsize,
 253                                                 void (*populate_pool)(struct ptlrpc_request_pool *, int))
 254 {
 255         struct ptlrpc_request_pool *pool;
 256
 257         OBD_ALLOC(pool, sizeof (struct ptlrpc_request_pool));
 258         if (!pool)
 259                 return NULL;
 260
 261         /* Request next power of two for the allocation, because internally
 262            kernel would do exactly this */
 263
 264         spin_lock_init(&pool->prp_lock);
 265         CFS_INIT_LIST_HEAD(&pool->prp_req_list);
 266         pool->prp_rq_size = msgsize;
 267         pool->prp_populate = populate_pool;
 268
 269         populate_pool(pool, num_rq);
 270
 271         if (list_empty(&pool->prp_req_list)) {
 272                 /* have not allocated a single request for the pool */
 273                 OBD_FREE(pool, sizeof (struct ptlrpc_request_pool));
 274                 pool = NULL;
 275         }
 276         return pool;
 277 }
 278
 279 static struct ptlrpc_request *ptlrpc_prep_req_from_pool(struct ptlrpc_request_pool *pool)
 280 {
 281         struct ptlrpc_request *request;
 282         struct lustre_msg *reqbuf;
 283
 284         if (!pool)
 285                 return NULL;
 286
 287         spin_lock(&pool->prp_lock);
 288
 289         /* See if we have anything in a pool, and bail out if nothing,
 290          * in writeout path, where this matters, this is safe to do, because
 291          * nothing is lost in this case, and when some in-flight requests
 292          * complete, this code will be called again. */
 293         if (unlikely(list_empty(&pool->prp_req_list))) {
 294                 spin_unlock(&pool->prp_lock);
 295                 return NULL;
 296         }
 297
 298         request = list_entry(pool->prp_req_list.next, struct ptlrpc_request,
 299                              rq_list);
 300         list_del(&request->rq_list);
 301         spin_unlock(&pool->prp_lock);
 302
 303         LASSERT(request->rq_reqbuf);
 304         LASSERT(request->rq_pool);
 305
 306         reqbuf = request->rq_reqbuf;
 307         memset(request, 0, sizeof(*request));
 308         request->rq_reqbuf = reqbuf;
 309         request->rq_reqbuf_len = pool->prp_rq_size;
 310         request->rq_pool = pool;
 311         return request;
 312 }
 313
 314 static void __ptlrpc_free_req_to_pool(struct ptlrpc_request *request)
 315 {
 316         struct ptlrpc_request_pool *pool = request->rq_pool;
 317
 318         spin_lock(&pool->prp_lock);
 319         LASSERT(list_empty(&request->rq_list));
 320         list_add_tail(&request->rq_list, &pool->prp_req_list);
 321         spin_unlock(&pool->prp_lock);
 322 }
 323
 324 static int __ptlrpc_request_bufs_pack(struct ptlrpc_request *request,
 325                                       __u32 version, int opcode,
 326                                       int count, int *lengths, char **bufs,
 327                                       struct ptlrpc_cli_ctx *ctx)
 328 {
 329         struct obd_import  *imp = request->rq_import;
 330         int                 rc;
 331         ENTRY;
 332
 333         if (unlikely(ctx))
 334                 request->rq_cli_ctx = sptlrpc_cli_ctx_get(ctx);
 335         else {
 336                 rc = sptlrpc_req_get_ctx(request);
 337                 if (rc)
 338                         GOTO(out_free, rc);
 339         }
 340
 341         sptlrpc_req_set_flavor(request, opcode);
 342
 343         rc = lustre_pack_request(request, imp->imp_msg_magic, count,
 344                                  lengths, bufs);
 345         if (rc) {
 346                 LASSERT(!request->rq_pool);
 347                 GOTO(out_ctx, rc);
 348         }
 349
 350         lustre_msg_add_version(request->rq_reqmsg, version);
 351
 352         if (imp->imp_server_timeout)
 353                 request->rq_timeout = obd_timeout / 2;
 354         else
 355                 request->rq_timeout = obd_timeout;
 356         request->rq_send_state = LUSTRE_IMP_FULL;
 357         request->rq_type = PTL_RPC_MSG_REQUEST;
 358         request->rq_export = NULL;
 359
 360         request->rq_req_cbid.cbid_fn  = request_out_callback;
 361         request->rq_req_cbid.cbid_arg = request;
 362
 363         request->rq_reply_cbid.cbid_fn  = reply_in_callback;
 364         request->rq_reply_cbid.cbid_arg = request;
 365
 366         request->rq_phase = RQ_PHASE_NEW;
 367
 368         /* XXX FIXME bug 249 */
 369         request->rq_request_portal = imp->imp_client->cli_request_portal;
 370         request->rq_reply_portal = imp->imp_client->cli_reply_portal;
 371
 372         spin_lock_init(&request->rq_lock);
 373         CFS_INIT_LIST_HEAD(&request->rq_list);
 374         CFS_INIT_LIST_HEAD(&request->rq_replay_list);
 375         CFS_INIT_LIST_HEAD(&request->rq_mod_list);
 376         CFS_INIT_LIST_HEAD(&request->rq_ctx_chain);
 377         CFS_INIT_LIST_HEAD(&request->rq_set_chain);
 378         CFS_INIT_LIST_HEAD(&request->rq_history_list);
 379         cfs_waitq_init(&request->rq_reply_waitq);
 380         request->rq_xid = ptlrpc_next_xid();
 381         atomic_set(&request->rq_refcount, 1);
 382
 383         lustre_msg_set_opc(request->rq_reqmsg, opcode);
 384         lustre_msg_set_flags(request->rq_reqmsg, 0);
 385
 386         RETURN(0);
 387 out_ctx:
 388         sptlrpc_cli_ctx_put(request->rq_cli_ctx, 1);
 389 out_free:
 390         class_import_put(imp);
 391         return rc;
 392 }
 393
 394 int ptlrpc_request_bufs_pack(struct ptlrpc_request *request,
 395                              __u32 version, int opcode, char **bufs,
 396                              struct ptlrpc_cli_ctx *ctx)
 397 {
 398         int count;
 399
 400         count = req_capsule_filled_sizes(&request->rq_pill, RCL_CLIENT);
 401         return __ptlrpc_request_bufs_pack(request, version, opcode, count,
 402                                           request->rq_pill.rc_area[RCL_CLIENT],
 403                                           bufs, ctx);
 404 }
 405 EXPORT_SYMBOL(ptlrpc_request_bufs_pack);
 406
 407 int ptlrpc_request_pack(struct ptlrpc_request *request,
 408                         __u32 version, int opcode)
 409 {
 410         return ptlrpc_request_bufs_pack(request, version, opcode, NULL, NULL);
 411 }
 412
 413 static inline
 414 struct ptlrpc_request *__ptlrpc_request_alloc(struct obd_import *imp,
 415                                               struct ptlrpc_request_pool *pool)
 416 {
 417         struct ptlrpc_request *request = NULL;
 418
 419         if (pool)
 420                 request = ptlrpc_prep_req_from_pool(pool);
 421
 422         if (!request)
 423                 OBD_ALLOC_PTR(request);
 424
 425         if (request) {
 426                 LASSERT((unsigned long)imp > 0x1000);
 427                 LASSERT(imp != LP_POISON);
 428                 LASSERT((unsigned long)imp->imp_client > 0x1000);
 429                 LASSERT(imp->imp_client != LP_POISON);
 430
 431                 request->rq_import = class_import_get(imp);
 432         } else {
 433                 CERROR("request allocation out of memory\n");
 434         }
 435
 436         return request;
 437 }
 438
 439 static struct ptlrpc_request *
 440 ptlrpc_request_alloc_internal(struct obd_import *imp,
 441                               struct ptlrpc_request_pool * pool,
 442                               const struct req_format *format)
 443 {
 444         struct ptlrpc_request *request;
 445
 446         request = __ptlrpc_request_alloc(imp, pool);
 447         if (request == NULL)
 448                 return NULL;
 449
 450         req_capsule_init(&request->rq_pill, request, RCL_CLIENT);
 451         req_capsule_set(&request->rq_pill, format);
 452         return request;
 453 }
 454
 455 struct ptlrpc_request *ptlrpc_request_alloc(struct obd_import *imp,
 456                                             const struct req_format *format)
 457 {
 458         return ptlrpc_request_alloc_internal(imp, NULL, format);
 459 }
 460
 461 struct ptlrpc_request *ptlrpc_request_alloc_pool(struct obd_import *imp,
 462                                             struct ptlrpc_request_pool * pool,
 463                                             const struct req_format *format)
 464 {
 465         return ptlrpc_request_alloc_internal(imp, pool, format);
 466 }
 467
 468 void ptlrpc_request_free(struct ptlrpc_request *request)
 469 {
 470         if (request->rq_pool)
 471                 __ptlrpc_free_req_to_pool(request);
 472         else
 473                 OBD_FREE_PTR(request);
 474 }
 475
 476 struct ptlrpc_request *ptlrpc_request_alloc_pack(struct obd_import *imp,
 477                                                 const struct req_format *format,
 478                                                 __u32 version, int opcode)
 479 {
 480         struct ptlrpc_request *req = ptlrpc_request_alloc(imp, format);
 481         int                    rc;
 482
 483         if (req) {
 484                 rc = ptlrpc_request_pack(req, version, opcode);
 485                 if (rc) {
 486                         ptlrpc_request_free(req);
 487                         req = NULL;
 488                 }
 489         }
 490         return req;
 491 }
 492
 493 struct ptlrpc_request *
 494 ptlrpc_prep_req_pool(struct obd_import *imp,
 495                      __u32 version, int opcode,
 496                      int count, int *lengths, char **bufs,
 497                      struct ptlrpc_request_pool *pool)
 498 {
 499         struct ptlrpc_request *request;
 500         int                    rc;
 501
 502         request = __ptlrpc_request_alloc(imp, pool);
 503         if (!request)
 504                 return NULL;
 505
 506         rc = __ptlrpc_request_bufs_pack(request, version, opcode, count,
 507                                         lengths, bufs, NULL);
 508         if (rc) {
 509                 ptlrpc_request_free(request);
 510                 request = NULL;
 511         }
 512         return request;
 513 }
 514
 515 struct ptlrpc_request *
 516 ptlrpc_prep_req(struct obd_import *imp, __u32 version, int opcode, int count,
 517                 int *lengths, char **bufs)
 518 {
 519         return ptlrpc_prep_req_pool(imp, version, opcode, count, lengths, bufs,
 520                                     NULL);
 521 }
 522
 523 struct ptlrpc_request_set *ptlrpc_prep_set(void)
 524 {
 525         struct ptlrpc_request_set *set;
 526
 527         ENTRY;
 528         OBD_ALLOC(set, sizeof *set);
 529         if (!set)
 530                 RETURN(NULL);
 531         CFS_INIT_LIST_HEAD(&set->set_requests);
 532         cfs_waitq_init(&set->set_waitq);
 533         set->set_remaining = 0;
 534         spin_lock_init(&set->set_new_req_lock);
 535         CFS_INIT_LIST_HEAD(&set->set_new_requests);
 536         CFS_INIT_LIST_HEAD(&set->set_cblist);
 537
 538         RETURN(set);
 539 }
 540
 541 /* Finish with this set; opposite of prep_set. */
 542 void ptlrpc_set_destroy(struct ptlrpc_request_set *set)
 543 {
 544         struct list_head *tmp;
 545         struct list_head *next;
 546         int               expected_phase;
 547         int               n = 0;
 548         ENTRY;
 549
 550         /* Requests on the set should either all be completed, or all be new */
 551         expected_phase = (set->set_remaining == 0) ?
 552                          RQ_PHASE_COMPLETE : RQ_PHASE_NEW;
 553         list_for_each (tmp, &set->set_requests) {
 554                 struct ptlrpc_request *req =
 555                         list_entry(tmp, struct ptlrpc_request, rq_set_chain);
 556
 557                 LASSERT(req->rq_phase == expected_phase);
 558                 n++;
 559         }
 560
 561         LASSERT(set->set_remaining == 0 || set->set_remaining == n);
 562
 563         list_for_each_safe(tmp, next, &set->set_requests) {
 564                 struct ptlrpc_request *req =
 565                         list_entry(tmp, struct ptlrpc_request, rq_set_chain);
 566                 list_del_init(&req->rq_set_chain);
 567
 568                 LASSERT(req->rq_phase == expected_phase);
 569
 570                 if (req->rq_phase == RQ_PHASE_NEW) {
 571
 572                         if (req->rq_interpret_reply != NULL) {
 573                                 int (*interpreter)(struct ptlrpc_request *,
 574                                                    void *, int) =
 575                                         req->rq_interpret_reply;
 576
 577                                 /* higher level (i.e. LOV) failed;
 578                                  * let the sub reqs clean up */
 579                                 req->rq_status = -EBADR;
 580                                 interpreter(req, &req->rq_async_args,
 581                                             req->rq_status);
 582                         }
 583                         set->set_remaining--;
 584                 }
 585
 586                 req->rq_set = NULL;
 587                 ptlrpc_req_finished (req);
 588         }
 589
 590         LASSERT(set->set_remaining == 0);
 591
 592         OBD_FREE(set, sizeof(*set));
 593         EXIT;
 594 }
 595
 596 int ptlrpc_set_add_cb(struct ptlrpc_request_set *set,
 597                       set_interpreter_func fn, void *data)
 598 {
 599         struct ptlrpc_set_cbdata *cbdata;
 600
 601         OBD_ALLOC_PTR(cbdata);
 602         if (cbdata == NULL)
 603                 RETURN(-ENOMEM);
 604
 605         cbdata->psc_interpret = fn;
 606         cbdata->psc_data = data;
 607         list_add_tail(&cbdata->psc_item, &set->set_cblist);
 608
 609         RETURN(0);
 610 }
 611
 612 void ptlrpc_set_add_req(struct ptlrpc_request_set *set,
 613                         struct ptlrpc_request *req)
 614 {
 615         /* The set takes over the caller's request reference */
 616         list_add_tail(&req->rq_set_chain, &set->set_requests);
 617         req->rq_set = set;
 618         set->set_remaining++;
 619
 620         atomic_inc(&req->rq_import->imp_inflight);
 621 }
 622
 623 /* lock so many callers can add things, the context that owns the set
 624  * is supposed to notice these and move them into the set proper. */
 625 void ptlrpc_set_add_new_req(struct ptlrpc_request_set *set,
 626                             struct ptlrpc_request *req)
 627 {
 628         spin_lock(&set->set_new_req_lock);
 629         /* The set takes over the caller's request reference */
 630         list_add_tail(&req->rq_set_chain, &set->set_new_requests);
 631         req->rq_set = set;
 632         spin_unlock(&set->set_new_req_lock);
 633 }
 634
 635 /*
 636  * Based on the current state of the import, determine if the request
 637  * can be sent, is an error, or should be delayed.
 638  *
 639  * Returns true if this request should be delayed. If false, and
 640  * *status is set, then the request can not be sent and *status is the
 641  * error code.  If false and status is 0, then request can be sent.
 642  *
 643  * The imp->imp_lock must be held.
 644  */
 645 static int ptlrpc_import_delay_req(struct obd_import *imp,
 646                                    struct ptlrpc_request *req, int *status)
 647 {
 648         int delay = 0;
 649         ENTRY;
 650
 651         LASSERT (status != NULL);
 652         *status = 0;
 653
 654         if (req->rq_ctx_init || req->rq_ctx_fini) {
 655                 /* always allow ctx init/fini rpc go through */
 656         } else if (imp->imp_state == LUSTRE_IMP_NEW) {
 657                 DEBUG_REQ(D_ERROR, req, "Uninitialized import.");
 658                 *status = -EIO;
 659                 LBUG();
 660         } else if (imp->imp_state == LUSTRE_IMP_CLOSED) {
 661                 DEBUG_REQ(D_ERROR, req, "IMP_CLOSED ");
 662                 *status = -EIO;
 663         } else if (req->rq_send_state == LUSTRE_IMP_CONNECTING &&
 664                    imp->imp_state == LUSTRE_IMP_CONNECTING) {
 665                 /* allow CONNECT even if import is invalid */ ;
 666                 if (atomic_read(&imp->imp_inval_count) != 0) {
 667                         DEBUG_REQ(D_ERROR, req, "invalidate in flight");
 668                         *status = -EIO;
 669                 }
 670
 671         } else if ((imp->imp_invalid && (!imp->imp_recon_bk)) ||
 672                                          imp->imp_obd->obd_no_recov) {
 673                 /* If the import has been invalidated (such as by an OST
 674                  * failure), and if the import(MGC) tried all of its connection
 675                  * list (Bug 13464), the request must fail with -ESHUTDOWN.
 676                  * This indicates the requests should be discarded; an -EIO
 677                  * may result in a resend of the request. */
 678                 if (!imp->imp_deactive)
 679                           DEBUG_REQ(D_ERROR, req, "IMP_INVALID");
 680                 *status = -ESHUTDOWN; /* bz 12940 */
 681         } else if (req->rq_import_generation != imp->imp_generation) {
 682                 DEBUG_REQ(D_ERROR, req, "req wrong generation:");
 683                 *status = -EIO;
 684         } else if (req->rq_send_state != imp->imp_state) {
 685                 /* invalidate in progress - any requests should be drop */
 686                 if (atomic_read(&imp->imp_inval_count) != 0) {
 687                         DEBUG_REQ(D_ERROR, req, "invalidate in flight");
 688                         *status = -EIO;
 689                 } else if (imp->imp_dlm_fake || req->rq_no_delay) {
 690                         *status = -EWOULDBLOCK;
 691                 } else {
 692                         delay = 1;
 693                 }
 694         }
 695
 696         RETURN(delay);
 697 }
 698
 699 static int ptlrpc_check_reply(struct ptlrpc_request *req)
 700 {
 701         int rc = 0;
 702         ENTRY;
 703
 704         /* serialise with network callback */
 705         spin_lock(&req->rq_lock);
 706
 707         if (req->rq_replied)
 708                 GOTO(out, rc = 1);
 709
 710         if (req->rq_net_err && !req->rq_timedout) {
 711                 spin_unlock(&req->rq_lock);
 712                 rc = ptlrpc_expire_one_request(req);
 713                 spin_lock(&req->rq_lock);
 714                 GOTO(out, rc);
 715         }
 716
 717         if (req->rq_err)
 718                 GOTO(out, rc = 1);
 719
 720         if (req->rq_resend)
 721                 GOTO(out, rc = 1);
 722
 723         if (req->rq_restart)
 724                 GOTO(out, rc = 1);
 725         EXIT;
 726  out:
 727         spin_unlock(&req->rq_lock);
 728         DEBUG_REQ(D_NET, req, "rc = %d for", rc);
 729         return rc;
 730 }
 731
 732 static int ptlrpc_check_status(struct ptlrpc_request *req)
 733 {
 734         int err;
 735         ENTRY;
 736
 737         err = lustre_msg_get_status(req->rq_repmsg);
 738         if (lustre_msg_get_type(req->rq_repmsg) == PTL_RPC_MSG_ERR) {
 739                 struct obd_import *imp = req->rq_import;
 740                 __u32 opc = lustre_msg_get_opc(req->rq_reqmsg);
 741                 LCONSOLE_ERROR_MSG(0x011,"an error occurred while communicating"
 742                                 " with %s. The %s operation failed with %d\n",
 743                                 libcfs_nid2str(imp->imp_connection->c_peer.nid),
 744                                 ll_opcode2str(opc), err);
 745                 RETURN(err < 0 ? err : -EINVAL);
 746         }
 747
 748         if (err < 0) {
 749                 DEBUG_REQ(D_INFO, req, "status is %d", err);
 750         } else if (err > 0) {
 751                 /* XXX: translate this error from net to host */
 752                 DEBUG_REQ(D_INFO, req, "status is %d", err);
 753         }
 754
 755         RETURN(err);
 756 }
 757
 758 static int after_reply(struct ptlrpc_request *req)
 759 {
 760         struct obd_import *imp = req->rq_import;
 761         struct obd_device *obd = req->rq_import->imp_obd;
 762         int rc;
 763         struct timeval work_start;
 764         long timediff;
 765         ENTRY;
 766
 767         LASSERT(!req->rq_receiving_reply);
 768         LASSERT(obd);
 769         LASSERT(req->rq_nob_received <= req->rq_repbuf_len);
 770
 771         /* NB Until this point, the whole of the incoming message,
 772          * including buflens, status etc is in the sender's byte order. */
 773
 774         /* Clear reply swab mask; this is a new reply in sender's byte order */
 775         req->rq_rep_swab_mask = 0;
 776
 777         rc = sptlrpc_cli_unwrap_reply(req);
 778         if (rc) {
 779                 DEBUG_REQ(D_ERROR, req, "unwrap reply failed (%d):", rc);
 780                 RETURN(rc);
 781         }
 782
 783         /* security layer unwrap might ask resend this request */
 784         if (req->rq_resend)
 785                 RETURN(0);
 786
 787         rc = lustre_unpack_msg(req->rq_repmsg, req->rq_replen);
 788         if (rc) {
 789                 DEBUG_REQ(D_ERROR, req, "unpack_rep failed: %d", rc);
 790                 RETURN(-EPROTO);
 791         }
 792
 793         rc = lustre_unpack_rep_ptlrpc_body(req, MSG_PTLRPC_BODY_OFF);
 794         if (rc) {
 795                 DEBUG_REQ(D_ERROR, req, "unpack ptlrpc body failed: %d", rc);
 796                 RETURN(-EPROTO);
 797         }
 798
 799         do_gettimeofday(&work_start);
 800         timediff = cfs_timeval_sub(&work_start, &req->rq_arrival_time, NULL);
 801         if (obd->obd_svc_stats != NULL)
 802                 lprocfs_counter_add(obd->obd_svc_stats, PTLRPC_REQWAIT_CNTR,
 803                                     timediff);
 804
 805         if (lustre_msg_get_type(req->rq_repmsg) != PTL_RPC_MSG_REPLY &&
 806             lustre_msg_get_type(req->rq_repmsg) != PTL_RPC_MSG_ERR) {
 807                 DEBUG_REQ(D_ERROR, req, "invalid packet received (type=%u)",
 808                           lustre_msg_get_type(req->rq_repmsg));
 809                 RETURN(-EPROTO);
 810         }
 811
 812         rc = ptlrpc_check_status(req);
 813         imp->imp_connect_error = rc;
 814
 815         if (rc) {
 816                 /* Either we've been evicted, or the server has failed for
 817                  * some reason. Try to reconnect, and if that fails, punt to
 818                  * the upcall. */
 819                 if (ll_rpc_recoverable_error(rc)) {
 820                         if (req->rq_send_state != LUSTRE_IMP_FULL ||
 821                             imp->imp_obd->obd_no_recov || imp->imp_dlm_fake) {
 822                                 RETURN(rc);
 823                         }
 824                         ptlrpc_request_handle_notconn(req);
 825                         RETURN(rc);
 826                 }
 827         } else {
 828                 /* Let's look if server send slv. Do it only for RPC with
 829                  * rc == 0. */
 830                 if (imp->imp_obd->obd_namespace) {
 831                         /* Disconnect rpc is sent when namespace is already
 832                          * destroyed. Let's check this and will not try update
 833                          * pool. */
 834                         ldlm_cli_update_pool(req);
 835                 }
 836         }
 837
 838         /* Store transno in reqmsg for replay. */
 839         req->rq_transno = lustre_msg_get_transno(req->rq_repmsg);
 840         lustre_msg_set_transno(req->rq_reqmsg, req->rq_transno);
 841
 842         if (req->rq_import->imp_replayable) {
 843                 spin_lock(&imp->imp_lock);
 844                 /* no point in adding already-committed requests to the replay
 845                  * list, we will just remove them immediately. b=9829 */
 846                 if (req->rq_transno != 0 &&
 847                     (req->rq_transno >
 848                      lustre_msg_get_last_committed(req->rq_repmsg) ||
 849                      req->rq_replay))
 850                         ptlrpc_retain_replayable_request(req, imp);
 851                 else if (req->rq_commit_cb != NULL) {
 852                         spin_unlock(&imp->imp_lock);
 853                         req->rq_commit_cb(req);
 854                         spin_lock(&imp->imp_lock);
 855                 }
 856
 857                 /* Replay-enabled imports return commit-status information. */
 858                 if (lustre_msg_get_last_committed(req->rq_repmsg)) {
 859                         imp->imp_peer_committed_transno =
 860                                 lustre_msg_get_last_committed(req->rq_repmsg);
 861                 }
 862                 ptlrpc_free_committed(imp);
 863                 spin_unlock(&imp->imp_lock);
 864         }
 865
 866         RETURN(rc);
 867 }
 868
 869 static int ptlrpc_send_new_req(struct ptlrpc_request *req)
 870 {
 871         struct obd_import     *imp;
 872         int rc;
 873         ENTRY;
 874
 875         LASSERT(req->rq_phase == RQ_PHASE_NEW);
 876         if (req->rq_sent && (req->rq_sent > CURRENT_SECONDS))
 877                 RETURN (0);
 878
 879         req->rq_phase = RQ_PHASE_RPC;
 880
 881         imp = req->rq_import;
 882         spin_lock(&imp->imp_lock);
 883
 884         req->rq_import_generation = imp->imp_generation;
 885
 886         if (ptlrpc_import_delay_req(imp, req, &rc)) {
 887                 spin_lock (&req->rq_lock);
 888                 req->rq_waiting = 1;
 889                 spin_unlock (&req->rq_lock);
 890
 891                 DEBUG_REQ(D_HA, req, "req from PID %d waiting for recovery: "
 892                           "(%s != %s)",
 893                           lustre_msg_get_status(req->rq_reqmsg) ,
 894                           ptlrpc_import_state_name(req->rq_send_state),
 895                           ptlrpc_import_state_name(imp->imp_state));
 896                 LASSERT(list_empty (&req->rq_list));
 897
 898                 list_add_tail(&req->rq_list, &imp->imp_delayed_list);
 899                 spin_unlock(&imp->imp_lock);
 900                 RETURN(0);
 901         }
 902
 903         if (rc != 0) {
 904                 spin_unlock(&imp->imp_lock);
 905                 req->rq_status = rc;
 906                 req->rq_phase = RQ_PHASE_INTERPRET;
 907                 RETURN(rc);
 908         }
 909
 910         /* XXX this is the same as ptlrpc_queue_wait */
 911         LASSERT(list_empty(&req->rq_list));
 912         list_add_tail(&req->rq_list, &imp->imp_sending_list);
 913         spin_unlock(&imp->imp_lock);
 914
 915         lustre_msg_set_status(req->rq_reqmsg, cfs_curproc_pid());
 916
 917         rc = sptlrpc_req_refresh_ctx(req, -1);
 918         if (rc) {
 919                 if (req->rq_err) {
 920                         req->rq_status = rc;
 921                         RETURN(1);
 922                 } else {
 923                         /* here begins timeout counting */
 924                         req->rq_sent = CURRENT_SECONDS;
 925                         req->rq_wait_ctx = 1;
 926                         RETURN(0);
 927                 }
 928         }
 929
 930         CDEBUG(D_RPCTRACE, "Sending RPC pname:cluuid:pid:xid:nid:opc"
 931                " %s:%s:%d:"LPU64":%s:%d\n", cfs_curproc_comm(),
 932                imp->imp_obd->obd_uuid.uuid,
 933                lustre_msg_get_status(req->rq_reqmsg), req->rq_xid,
 934                libcfs_nid2str(imp->imp_connection->c_peer.nid),
 935                lustre_msg_get_opc(req->rq_reqmsg));
 936
 937         rc = ptl_send_rpc(req, 0);
 938         if (rc) {
 939                 DEBUG_REQ(D_HA, req, "send failed (%d); expect timeout", rc);
 940                 req->rq_net_err = 1;
 941                 RETURN(rc);
 942         }
 943         RETURN(0);
 944 }
 945
 946 /* this sends any unsent RPCs in @set and returns TRUE if all are sent */
 947 int ptlrpc_check_set(struct ptlrpc_request_set *set)
 948 {
 949         struct list_head *tmp;
 950         int force_timer_recalc = 0;
 951         ENTRY;
 952
 953         if (set->set_remaining == 0)
 954                 RETURN(1);
 955
 956         list_for_each(tmp, &set->set_requests) {
 957                 struct ptlrpc_request *req =
 958                         list_entry(tmp, struct ptlrpc_request, rq_set_chain);
 959                 struct obd_import *imp = req->rq_import;
 960                 int rc = 0;
 961
 962                 if (req->rq_phase == RQ_PHASE_NEW &&
 963                     ptlrpc_send_new_req(req)) {
 964                         force_timer_recalc = 1;
 965                 }
 966                 /* delayed send - skip */
 967                 if (req->rq_phase == RQ_PHASE_NEW && req->rq_sent)
 968                         continue;
 969
 970                 if (!(req->rq_phase == RQ_PHASE_RPC ||
 971                       req->rq_phase == RQ_PHASE_BULK ||
 972                       req->rq_phase == RQ_PHASE_INTERPRET ||
 973                       req->rq_phase == RQ_PHASE_COMPLETE)) {
 974                         DEBUG_REQ(D_ERROR, req, "bad phase %x", req->rq_phase);
 975                         LBUG();
 976                 }
 977
 978                 if (req->rq_phase == RQ_PHASE_COMPLETE)
 979                         continue;
 980
 981                 if (req->rq_phase == RQ_PHASE_INTERPRET)
 982                         GOTO(interpret, req->rq_status);
 983
 984                 if (req->rq_net_err && !req->rq_timedout)
 985                         ptlrpc_expire_one_request(req);
 986
 987                 if (req->rq_err) {
 988                         ptlrpc_unregister_reply(req);
 989                         req->rq_replied = 0;
 990                         if (req->rq_status == 0)
 991                                 req->rq_status = -EIO;
 992                         req->rq_phase = RQ_PHASE_INTERPRET;
 993
 994                         spin_lock(&imp->imp_lock);
 995                         list_del_init(&req->rq_list);
 996                         spin_unlock(&imp->imp_lock);
 997
 998                         GOTO(interpret, req->rq_status);
 999                 }
1000
1001                 /* ptlrpc_queue_wait->l_wait_event guarantees that rq_intr
1002                  * will only be set after rq_timedout, but the oig waiting
1003                  * path sets rq_intr irrespective of whether ptlrpcd has
1004                  * seen a timeout.  our policy is to only interpret
1005                  * interrupted rpcs after they have timed out */
1006                 if (req->rq_intr && (req->rq_timedout || req->rq_waiting ||
1007                                      req->rq_wait_ctx)) {
1008                         /* NB could be on delayed list */
1009                         ptlrpc_unregister_reply(req);
1010                         req->rq_status = -EINTR;
1011                         req->rq_phase = RQ_PHASE_INTERPRET;
1012
1013                         spin_lock(&imp->imp_lock);
1014                         list_del_init(&req->rq_list);
1015                         spin_unlock(&imp->imp_lock);
1016
1017                         GOTO(interpret, req->rq_status);
1018                 }
1019
1020                 if (req->rq_phase == RQ_PHASE_RPC) {
1021                         if (req->rq_timedout || req->rq_resend ||
1022                             req->rq_waiting || req->rq_wait_ctx) {
1023                                 int status;
1024
1025                                 /* rq_wait_ctx is only touched in ptlrpcd,
1026                                  * no lock needed here.
1027                                  */
1028                                 if (req->rq_wait_ctx)
1029                                         goto check_ctx;
1030
1031                                 ptlrpc_unregister_reply(req);
1032
1033                                 spin_lock(&imp->imp_lock);
1034
1035                                 if (ptlrpc_import_delay_req(imp, req, &status)){
1036                                         spin_unlock(&imp->imp_lock);
1037                                         continue;
1038                                 }
1039
1040                                 list_del_init(&req->rq_list);
1041                                 if (status != 0)  {
1042                                         req->rq_status = status;
1043                                         req->rq_phase = RQ_PHASE_INTERPRET;
1044                                         spin_unlock(&imp->imp_lock);
1045                                         GOTO(interpret, req->rq_status);
1046                                 }
1047                                 if (req->rq_no_resend && !req->rq_wait_ctx) {
1048                                         req->rq_status = -ENOTCONN;
1049                                         req->rq_phase = RQ_PHASE_INTERPRET;
1050                                         spin_unlock(&imp->imp_lock);
1051                                         GOTO(interpret, req->rq_status);
1052                                 }
1053                                 list_add_tail(&req->rq_list,
1054                                               &imp->imp_sending_list);
1055
1056                                 spin_unlock(&imp->imp_lock);
1057
1058                                 req->rq_waiting = 0;
1059                                 if (req->rq_resend) {
1060                                         lustre_msg_add_flags(req->rq_reqmsg,
1061                                                              MSG_RESENT);
1062                                         if (req->rq_bulk) {
1063                                                 __u64 old_xid = req->rq_xid;
1064
1065                                                 ptlrpc_unregister_bulk (req);
1066
1067                                                 /* ensure previous bulk fails */
1068                                                 req->rq_xid = ptlrpc_next_xid();
1069                                                 CDEBUG(D_HA, "resend bulk "
1070                                                        "old x"LPU64
1071                                                        " new x"LPU64"\n",
1072                                                        old_xid, req->rq_xid);
1073                                         }
1074                                 }
1075 check_ctx:
1076                                 status = sptlrpc_req_refresh_ctx(req, -1);
1077                                 if (status) {
1078                                         if (req->rq_err) {
1079                                                 req->rq_status = status;
1080                                                 force_timer_recalc = 1;
1081                                         }
1082                                         if (!req->rq_wait_ctx) {
1083                                                 /* begins timeout counting */
1084                                                 req->rq_sent = CURRENT_SECONDS;
1085                                                 req->rq_wait_ctx = 1;
1086                                         }
1087                                         continue;
1088                                 } else {
1089                                         req->rq_sent = 0;
1090                                         req->rq_wait_ctx = 0;
1091                                 }
1092
1093                                 rc = ptl_send_rpc(req, 0);
1094                                 if (rc) {
1095                                         DEBUG_REQ(D_HA, req, "send failed (%d)",
1096                                                   rc);
1097                                         force_timer_recalc = 1;
1098                                         req->rq_net_err = 1;
1099                                 }
1100                                 /* need to reset the timeout */
1101                                 force_timer_recalc = 1;
1102                         }
1103
1104                         /* Still waiting for a reply? */
1105                         if (ptlrpc_client_receiving_reply(req))
1106                                 continue;
1107
1108                         /* Did we actually receive a reply? */
1109                         if (!ptlrpc_client_replied(req))
1110                                 continue;
1111
1112                         spin_lock(&imp->imp_lock);
1113                         list_del_init(&req->rq_list);
1114                         spin_unlock(&imp->imp_lock);
1115
1116                         req->rq_status = after_reply(req);
1117                         if (req->rq_resend) {
1118                                 /* Add this req to the delayed list so
1119                                    it can be errored if the import is
1120                                    evicted after recovery. */
1121                                 spin_lock(&imp->imp_lock);
1122                                 list_add_tail(&req->rq_list,
1123                                               &imp->imp_delayed_list);
1124                                 spin_unlock(&imp->imp_lock);
1125                                 continue;
1126                         }
1127
1128                         /* If there is no bulk associated with this request,
1129                          * then we're done and should let the interpreter
1130                          * process the reply.  Similarly if the RPC returned
1131                          * an error, and therefore the bulk will never arrive.
1132                          */
1133                         if (req->rq_bulk == NULL || req->rq_status != 0) {
1134                                 req->rq_phase = RQ_PHASE_INTERPRET;
1135                                 GOTO(interpret, req->rq_status);
1136                         }
1137
1138                         req->rq_phase = RQ_PHASE_BULK;
1139                 }
1140
1141                 LASSERT(req->rq_phase == RQ_PHASE_BULK);
1142                 if (ptlrpc_bulk_active(req->rq_bulk))
1143                         continue;
1144
1145                 if (!req->rq_bulk->bd_success) {
1146                         /* The RPC reply arrived OK, but the bulk screwed
1147                          * up!  Dead wierd since the server told us the RPC
1148                          * was good after getting the REPLY for her GET or
1149                          * the ACK for her PUT. */
1150                         DEBUG_REQ(D_ERROR, req, "bulk transfer failed");
1151                         LBUG();
1152                 }
1153
1154                 req->rq_phase = RQ_PHASE_INTERPRET;
1155
1156         interpret:
1157                 LASSERT(req->rq_phase == RQ_PHASE_INTERPRET);
1158                 LASSERT(!req->rq_receiving_reply);
1159
1160                 ptlrpc_unregister_reply(req);
1161                 if (req->rq_bulk != NULL)
1162                         ptlrpc_unregister_bulk (req);
1163
1164                 if (req->rq_interpret_reply != NULL) {
1165                         int (*interpreter)(struct ptlrpc_request *,void *,int) =
1166                                 req->rq_interpret_reply;
1167                         req->rq_status = interpreter(req, &req->rq_async_args,
1168                                                      req->rq_status);
1169                 }
1170                 req->rq_phase = RQ_PHASE_COMPLETE;
1171
1172                 CDEBUG(D_RPCTRACE, "Completed RPC pname:cluuid:pid:xid:nid:"
1173                        "opc %s:%s:%d:"LPU64":%s:%d\n", cfs_curproc_comm(),
1174                        imp->imp_obd->obd_uuid.uuid,
1175                        lustre_msg_get_status(req->rq_reqmsg), req->rq_xid,
1176                        libcfs_nid2str(imp->imp_connection->c_peer.nid),
1177                        lustre_msg_get_opc(req->rq_reqmsg));
1178
1179                 atomic_dec(&imp->imp_inflight);
1180                 set->set_remaining--;
1181                 cfs_waitq_signal(&imp->imp_recovery_waitq);
1182         }
1183
1184         /* If we hit an error, we want to recover promptly. */
1185         RETURN(set->set_remaining == 0 || force_timer_recalc);
1186 }
1187
1188 int ptlrpc_expire_one_request(struct ptlrpc_request *req)
1189 {
1190         struct obd_import *imp = req->rq_import;
1191         int rc = 0;
1192         ENTRY;
1193
1194         DEBUG_REQ(D_ERROR|D_NETERROR, req, "%s (sent at %lu, %lus ago)",
1195                   req->rq_net_err ? "network error" : "timeout",
1196                   (long)req->rq_sent, CURRENT_SECONDS - req->rq_sent);
1197
1198         if (imp != NULL && obd_debug_peer_on_timeout)
1199                 LNetCtl(IOC_LIBCFS_DEBUG_PEER, &imp->imp_connection->c_peer);
1200
1201         spin_lock(&req->rq_lock);
1202         req->rq_timedout = 1;
1203         req->rq_wait_ctx = 0;
1204         spin_unlock(&req->rq_lock);
1205
1206         ptlrpc_unregister_reply (req);
1207
1208         if (obd_dump_on_timeout)
1209                 libcfs_debug_dumplog();
1210
1211         if (req->rq_bulk != NULL)
1212                 ptlrpc_unregister_bulk (req);
1213
1214         if (imp == NULL) {
1215                 DEBUG_REQ(D_HA, req, "NULL import: already cleaned up?");
1216                 RETURN(1);
1217         }
1218
1219         /* The DLM server doesn't want recovery run on its imports. */
1220         if (imp->imp_dlm_fake)
1221                 RETURN(1);
1222
1223         /* If this request is for recovery or other primordial tasks,
1224          * then error it out here. */
1225         if (req->rq_ctx_init || req->rq_ctx_fini ||
1226             req->rq_send_state != LUSTRE_IMP_FULL ||
1227             imp->imp_obd->obd_no_recov) {
1228                 spin_lock(&req->rq_lock);
1229                 req->rq_status = -ETIMEDOUT;
1230                 req->rq_err = 1;
1231                 spin_unlock(&req->rq_lock);
1232                 RETURN(1);
1233         }
1234
1235         /* if request can't be resend we can't wait answer after timeout */
1236         if (req->rq_no_resend) {
1237                 DEBUG_REQ(D_RPCTRACE, req, "TIMEOUT-NORESEND:");
1238                 rc = 1;
1239         }
1240
1241         ptlrpc_fail_import(imp, lustre_msg_get_conn_cnt(req->rq_reqmsg));
1242
1243         RETURN(rc);
1244 }
1245
1246 int ptlrpc_expired_set(void *data)
1247 {
1248         struct ptlrpc_request_set *set = data;
1249         struct list_head          *tmp;
1250         time_t                     now = CURRENT_SECONDS;
1251         ENTRY;
1252
1253         LASSERT(set != NULL);
1254
1255         /* A timeout expired; see which reqs it applies to... */
1256         list_for_each (tmp, &set->set_requests) {
1257                 struct ptlrpc_request *req =
1258                         list_entry(tmp, struct ptlrpc_request, rq_set_chain);
1259
1260                 /* request in-flight? */
1261                 if (!((req->rq_phase == RQ_PHASE_RPC && !req->rq_waiting &&
1262                        !req->rq_resend) ||
1263                       (req->rq_phase == RQ_PHASE_BULK)))
1264                         continue;
1265
1266                 if (req->rq_timedout ||           /* already dealt with */
1267                     req->rq_sent + req->rq_timeout > now) /* not expired */
1268                         continue;
1269
1270                 /* deal with this guy */
1271                 ptlrpc_expire_one_request (req);
1272         }
1273
1274         /* When waiting for a whole set, we always to break out of the
1275          * sleep so we can recalculate the timeout, or enable interrupts
1276          * iff everyone's timed out.
1277          */
1278         RETURN(1);
1279 }
1280
1281 void ptlrpc_mark_interrupted(struct ptlrpc_request *req)
1282 {
1283         spin_lock(&req->rq_lock);
1284         req->rq_intr = 1;
1285         spin_unlock(&req->rq_lock);
1286 }
1287
1288 void ptlrpc_interrupted_set(void *data)
1289 {
1290         struct ptlrpc_request_set *set = data;
1291         struct list_head *tmp;
1292
1293         LASSERT(set != NULL);
1294         CERROR("INTERRUPTED SET %p\n", set);
1295
1296         list_for_each(tmp, &set->set_requests) {
1297                 struct ptlrpc_request *req =
1298                         list_entry(tmp, struct ptlrpc_request, rq_set_chain);
1299
1300                 if (req->rq_phase != RQ_PHASE_RPC)
1301                         continue;
1302
1303                 ptlrpc_mark_interrupted(req);
1304         }
1305 }
1306
1307 int ptlrpc_set_next_timeout(struct ptlrpc_request_set *set)
1308 {
1309         struct list_head      *tmp;
1310         time_t                 now = CURRENT_SECONDS;
1311         time_t                 deadline;
1312         int                    timeout = 0;
1313         struct ptlrpc_request *req;
1314         ENTRY;
1315
1316         SIGNAL_MASK_ASSERT(); /* XXX BUG 1511 */
1317
1318         list_for_each(tmp, &set->set_requests) {
1319                 req = list_entry(tmp, struct ptlrpc_request, rq_set_chain);
1320
1321                 /* request in-flight? */
1322                 if (!((req->rq_phase == RQ_PHASE_RPC && !req->rq_waiting) ||
1323                       (req->rq_phase == RQ_PHASE_BULK) ||
1324                       (req->rq_phase == RQ_PHASE_NEW)))
1325                         continue;
1326
1327                 if (req->rq_timedout)   /* already timed out */
1328                         continue;
1329
1330                 if (req->rq_phase == RQ_PHASE_NEW)
1331                         deadline = req->rq_sent;
1332                 else
1333                         deadline = req->rq_sent + req->rq_timeout;
1334
1335                 if (deadline <= now)    /* actually expired already */
1336                         timeout = 1;    /* ASAP */
1337                 else if (timeout == 0 || timeout > deadline - now)
1338                         timeout = deadline - now;
1339         }
1340         RETURN(timeout);
1341 }
1342
1343 int ptlrpc_set_wait(struct ptlrpc_request_set *set)
1344 {
1345         struct list_head      *tmp;
1346         struct ptlrpc_request *req;
1347         struct l_wait_info     lwi;
1348         int                    rc, timeout;
1349         ENTRY;
1350
1351         if (list_empty(&set->set_requests))
1352                 RETURN(0);
1353
1354         list_for_each(tmp, &set->set_requests) {
1355                 req = list_entry(tmp, struct ptlrpc_request, rq_set_chain);
1356                 if (req->rq_phase == RQ_PHASE_NEW)
1357                         (void)ptlrpc_send_new_req(req);
1358         }
1359
1360         do {
1361                 timeout = ptlrpc_set_next_timeout(set);
1362
1363                 /* wait until all complete, interrupted, or an in-flight
1364                  * req times out */
1365                 CDEBUG(D_RPCTRACE, "set %p going to sleep for %d seconds\n",
1366                        set, timeout);
1367                 lwi = LWI_TIMEOUT_INTR(cfs_time_seconds(timeout ? timeout : 1),
1368                                        ptlrpc_expired_set,
1369                                        ptlrpc_interrupted_set, set);
1370                 rc = l_wait_event(set->set_waitq, ptlrpc_check_set(set), &lwi);
1371
1372                 LASSERT(rc == 0 || rc == -EINTR || rc == -ETIMEDOUT);
1373
1374                 /* -EINTR => all requests have been flagged rq_intr so next
1375                  * check completes.
1376                  * -ETIMEOUTD => someone timed out.  When all reqs have
1377                  * timed out, signals are enabled allowing completion with
1378                  * EINTR.
1379                  * I don't really care if we go once more round the loop in
1380                  * the error cases -eeb. */
1381         } while (rc != 0 || set->set_remaining != 0);
1382
1383         LASSERT(set->set_remaining == 0);
1384
1385         rc = 0;
1386         list_for_each(tmp, &set->set_requests) {
1387                 req = list_entry(tmp, struct ptlrpc_request, rq_set_chain);
1388
1389                 LASSERT(req->rq_phase == RQ_PHASE_COMPLETE);
1390                 if (req->rq_status != 0)
1391                         rc = req->rq_status;
1392         }
1393
1394         if (set->set_interpret != NULL) {
1395                 int (*interpreter)(struct ptlrpc_request_set *set,void *,int) =
1396                         set->set_interpret;
1397                 rc = interpreter (set, set->set_arg, rc);
1398         } else {
1399                 struct ptlrpc_set_cbdata *cbdata, *n;
1400                 int err;
1401
1402                 list_for_each_entry_safe(cbdata, n,
1403                                          &set->set_cblist, psc_item) {
1404                         list_del_init(&cbdata->psc_item);
1405                         err = cbdata->psc_interpret(set, cbdata->psc_data, rc);
1406                         if (err && !rc)
1407                                 rc = err;
1408                         OBD_FREE_PTR(cbdata);
1409                 }
1410         }
1411
1412         RETURN(rc);
1413 }
1414
1415 static void __ptlrpc_free_req(struct ptlrpc_request *request, int locked)
1416 {
1417         ENTRY;
1418         if (request == NULL) {
1419                 EXIT;
1420                 return;
1421         }
1422
1423         LASSERTF(!request->rq_receiving_reply, "req %p\n", request);
1424         LASSERTF(request->rq_rqbd == NULL, "req %p\n",request);/* client-side */
1425         LASSERTF(list_empty(&request->rq_list), "req %p\n", request);
1426         LASSERTF(list_empty(&request->rq_set_chain), "req %p\n", request);
1427         LASSERT(request->rq_cli_ctx);
1428
1429         req_capsule_fini(&request->rq_pill);
1430
1431         /* We must take it off the imp_replay_list first.  Otherwise, we'll set
1432          * request->rq_reqmsg to NULL while osc_close is dereferencing it. */
1433         if (request->rq_import != NULL) {
1434                 if (!locked)
1435                         spin_lock(&request->rq_import->imp_lock);
1436                 list_del_init(&request->rq_mod_list);
1437                 list_del_init(&request->rq_replay_list);
1438                 if (!locked)
1439                         spin_unlock(&request->rq_import->imp_lock);
1440         }
1441         LASSERTF(list_empty(&request->rq_replay_list), "req %p\n", request);
1442
1443         if (atomic_read(&request->rq_refcount) != 0) {
1444                 DEBUG_REQ(D_ERROR, request,
1445                           "freeing request with nonzero refcount");
1446                 LBUG();
1447         }
1448
1449         if (request->rq_repbuf != NULL)
1450                 sptlrpc_cli_free_repbuf(request);
1451         if (request->rq_export != NULL) {
1452                 class_export_put(request->rq_export);
1453                 request->rq_export = NULL;
1454         }
1455         if (request->rq_import != NULL) {
1456                 class_import_put(request->rq_import);
1457                 request->rq_import = NULL;
1458         }
1459         if (request->rq_bulk != NULL)
1460                 ptlrpc_free_bulk(request->rq_bulk);
1461
1462         if (request->rq_reqbuf != NULL || request->rq_clrbuf != NULL)
1463                 sptlrpc_cli_free_reqbuf(request);
1464
1465         sptlrpc_req_put_ctx(request, !locked);
1466
1467         if (request->rq_pool)
1468                 __ptlrpc_free_req_to_pool(request);
1469         else
1470                 OBD_FREE(request, sizeof(*request));
1471         EXIT;
1472 }
1473
1474 void ptlrpc_free_req(struct ptlrpc_request *request)
1475 {
1476         __ptlrpc_free_req(request, 0);
1477 }
1478
1479 static int __ptlrpc_req_finished(struct ptlrpc_request *request, int locked);
1480 void ptlrpc_req_finished_with_imp_lock(struct ptlrpc_request *request)
1481 {
1482         LASSERT_SPIN_LOCKED(&request->rq_import->imp_lock);
1483         (void)__ptlrpc_req_finished(request, 1);
1484 }
1485
1486 static int __ptlrpc_req_finished(struct ptlrpc_request *request, int locked)
1487 {
1488         ENTRY;
1489         if (request == NULL)
1490                 RETURN(1);
1491
1492         if (request == LP_POISON ||
1493             request->rq_reqmsg == LP_POISON) {
1494                 CERROR("dereferencing freed request (bug 575)\n");
1495                 LBUG();
1496                 RETURN(1);
1497         }
1498
1499         DEBUG_REQ(D_INFO, request, "refcount now %u",
1500                   atomic_read(&request->rq_refcount) - 1);
1501
1502         if (atomic_dec_and_test(&request->rq_refcount)) {
1503                 __ptlrpc_free_req(request, locked);
1504                 RETURN(1);
1505         }
1506
1507         RETURN(0);
1508 }
1509
1510 void ptlrpc_req_finished(struct ptlrpc_request *request)
1511 {
1512         __ptlrpc_req_finished(request, 0);
1513 }
1514
1515 __u64 ptlrpc_req_xid(struct ptlrpc_request *request)
1516 {
1517         return request->rq_xid;
1518 }
1519 EXPORT_SYMBOL(ptlrpc_req_xid);
1520
1521 /* Disengage the client's reply buffer from the network
1522  * NB does _NOT_ unregister any client-side bulk.
1523  * IDEMPOTENT, but _not_ safe against concurrent callers.
1524  * The request owner (i.e. the thread doing the I/O) must call...
1525  */
1526 void ptlrpc_unregister_reply (struct ptlrpc_request *request)
1527 {
1528         int                rc;
1529         cfs_waitq_t       *wq;
1530         struct l_wait_info lwi;
1531
1532         LASSERT(!in_interrupt ());             /* might sleep */
1533
1534         if (!ptlrpc_client_receiving_reply(request))
1535                 return;
1536
1537         LNetMDUnlink (request->rq_reply_md_h);
1538
1539         /* We have to l_wait_event() whatever the result, to give liblustre
1540          * a chance to run reply_in_callback() */
1541
1542         if (request->rq_set != NULL)
1543                 wq = &request->rq_set->set_waitq;
1544         else
1545                 wq = &request->rq_reply_waitq;
1546
1547         for (;;) {
1548                 /* Network access will complete in finite time but the HUGE
1549                  * timeout lets us CWARN for visibility of sluggish NALs */
1550                 lwi = LWI_TIMEOUT(cfs_time_seconds(300), NULL, NULL);
1551                 rc = l_wait_event (*wq, !ptlrpc_client_receiving_reply(request), &lwi);
1552                 if (rc == 0)
1553                         return;
1554
1555                 LASSERT (rc == -ETIMEDOUT);
1556                 DEBUG_REQ(D_WARNING, request, "Unexpectedly long timeout");
1557         }
1558 }
1559
1560 /* caller must hold imp->imp_lock */
1561 void ptlrpc_free_committed(struct obd_import *imp)
1562 {
1563         struct list_head *tmp, *saved;
1564         struct ptlrpc_request *req;
1565         struct ptlrpc_request *last_req = NULL; /* temporary fire escape */
1566         ENTRY;
1567
1568         LASSERT(imp != NULL);
1569
1570         LASSERT_SPIN_LOCKED(&imp->imp_lock);
1571
1572
1573         if (imp->imp_peer_committed_transno == imp->imp_last_transno_checked &&
1574             imp->imp_generation == imp->imp_last_generation_checked) {
1575                 CDEBUG(D_RPCTRACE, "%s: skip recheck: last_committed "LPU64"\n",
1576                        imp->imp_obd->obd_name, imp->imp_peer_committed_transno);
1577                 return;
1578         }
1579
1580         CDEBUG(D_RPCTRACE, "%s: committing for last_committed "LPU64" gen %d\n",
1581                imp->imp_obd->obd_name, imp->imp_peer_committed_transno,
1582                imp->imp_generation);
1583         imp->imp_last_transno_checked = imp->imp_peer_committed_transno;
1584         imp->imp_last_generation_checked = imp->imp_generation;
1585
1586         list_for_each_safe(tmp, saved, &imp->imp_replay_list) {
1587                 req = list_entry(tmp, struct ptlrpc_request, rq_replay_list);
1588
1589                 /* XXX ok to remove when 1357 resolved - rread 05/29/03  */
1590                 LASSERT(req != last_req);
1591                 last_req = req;
1592
1593                 if (req->rq_import_generation < imp->imp_generation) {
1594                         DEBUG_REQ(D_RPCTRACE, req, "free request with old gen");
1595                         GOTO(free_req, 0);
1596                 }
1597
1598                 if (req->rq_replay) {
1599                         DEBUG_REQ(D_RPCTRACE, req, "keeping (FL_REPLAY)");
1600                         continue;
1601                 }
1602
1603                 /* not yet committed */
1604                 if (req->rq_transno > imp->imp_peer_committed_transno) {
1605                         DEBUG_REQ(D_RPCTRACE, req, "stopping search");
1606                         break;
1607                 }
1608
1609                 DEBUG_REQ(D_RPCTRACE, req, "commit (last_committed "LPU64")",
1610                           imp->imp_peer_committed_transno);
1611 free_req:
1612                 spin_lock(&req->rq_lock);
1613                 req->rq_replay = 0;
1614                 spin_unlock(&req->rq_lock);
1615                 if (req->rq_commit_cb != NULL)
1616                         req->rq_commit_cb(req);
1617                 list_del_init(&req->rq_replay_list);
1618                 __ptlrpc_req_finished(req, 1);
1619         }
1620
1621         EXIT;
1622         return;
1623 }
1624
1625 void ptlrpc_cleanup_client(struct obd_import *imp)
1626 {
1627         ENTRY;
1628         EXIT;
1629         return;
1630 }
1631
1632 void ptlrpc_resend_req(struct ptlrpc_request *req)
1633 {
1634         DEBUG_REQ(D_HA, req, "going to resend");
1635         lustre_msg_set_handle(req->rq_reqmsg, &(struct lustre_handle){ 0 });
1636         req->rq_status = -EAGAIN;
1637
1638         spin_lock(&req->rq_lock);
1639         req->rq_resend = 1;
1640         req->rq_net_err = 0;
1641         req->rq_timedout = 0;
1642         if (req->rq_bulk) {
1643                 __u64 old_xid = req->rq_xid;
1644
1645                 /* ensure previous bulk fails */
1646                 req->rq_xid = ptlrpc_next_xid();
1647                 CDEBUG(D_HA, "resend bulk old x"LPU64" new x"LPU64"\n",
1648                        old_xid, req->rq_xid);
1649         }
1650         ptlrpc_wake_client_req(req);
1651         spin_unlock(&req->rq_lock);
1652 }
1653
1654 /* XXX: this function and rq_status are currently unused */
1655 void ptlrpc_restart_req(struct ptlrpc_request *req)
1656 {
1657         DEBUG_REQ(D_HA, req, "restarting (possibly-)completed request");
1658         req->rq_status = -ERESTARTSYS;
1659
1660         spin_lock(&req->rq_lock);
1661         req->rq_restart = 1;
1662         req->rq_timedout = 0;
1663         ptlrpc_wake_client_req(req);
1664         spin_unlock(&req->rq_lock);
1665 }
1666
1667 static int expired_request(void *data)
1668 {
1669         struct ptlrpc_request *req = data;
1670         ENTRY;
1671
1672         /* some failure can suspend regular timeouts */
1673         if (ptlrpc_check_suspend())
1674                 RETURN(1);
1675
1676         RETURN(ptlrpc_expire_one_request(req));
1677 }
1678
1679 static void interrupted_request(void *data)
1680 {
1681         struct ptlrpc_request *req = data;
1682         DEBUG_REQ(D_HA, req, "request interrupted");
1683         spin_lock(&req->rq_lock);
1684         req->rq_intr = 1;
1685         spin_unlock(&req->rq_lock);
1686 }
1687
1688 struct ptlrpc_request *ptlrpc_request_addref(struct ptlrpc_request *req)
1689 {
1690         ENTRY;
1691         atomic_inc(&req->rq_refcount);
1692         RETURN(req);
1693 }
1694
1695 void ptlrpc_retain_replayable_request(struct ptlrpc_request *req,
1696                                       struct obd_import *imp)
1697 {
1698         struct list_head *tmp;
1699
1700         LASSERT_SPIN_LOCKED(&imp->imp_lock);
1701
1702         /* clear this for new requests that were resent as well
1703            as resent replayed requests. */
1704         lustre_msg_clear_flags(req->rq_reqmsg, MSG_RESENT);
1705
1706         /* don't re-add requests that have been replayed */
1707         if (!list_empty(&req->rq_replay_list))
1708                 return;
1709
1710         lustre_msg_add_flags(req->rq_reqmsg, MSG_REPLAY);
1711
1712         LASSERT(imp->imp_replayable);
1713         /* Balanced in ptlrpc_free_committed, usually. */
1714         ptlrpc_request_addref(req);
1715         list_for_each_prev(tmp, &imp->imp_replay_list) {
1716                 struct ptlrpc_request *iter =
1717                         list_entry(tmp, struct ptlrpc_request, rq_replay_list);
1718
1719                 /* We may have duplicate transnos if we create and then
1720                  * open a file, or for closes retained if to match creating
1721                  * opens, so use req->rq_xid as a secondary key.
1722                  * (See bugs 684, 685, and 428.)
1723                  * XXX no longer needed, but all opens need transnos!
1724                  */
1725                 if (iter->rq_transno > req->rq_transno)
1726                         continue;
1727
1728                 if (iter->rq_transno == req->rq_transno) {
1729                         LASSERT(iter->rq_xid != req->rq_xid);
1730                         if (iter->rq_xid > req->rq_xid)
1731                                 continue;
1732                 }
1733
1734                 list_add(&req->rq_replay_list, &iter->rq_replay_list);
1735                 return;
1736         }
1737
1738         list_add_tail(&req->rq_replay_list, &imp->imp_replay_list);
1739 }
1740
1741 int ptlrpc_queue_wait(struct ptlrpc_request *req)
1742 {
1743         int rc = 0;
1744         int brc;
1745         struct l_wait_info lwi;
1746         struct obd_import *imp = req->rq_import;
1747         cfs_duration_t timeout = 0;
1748         ENTRY;
1749
1750         LASSERT(req->rq_set == NULL);
1751         LASSERT(!req->rq_receiving_reply);
1752         atomic_inc(&imp->imp_inflight);
1753
1754         /* for distributed debugging */
1755         lustre_msg_set_status(req->rq_reqmsg, cfs_curproc_pid());
1756         LASSERT(imp->imp_obd != NULL);
1757         CDEBUG(D_RPCTRACE, "Sending RPC pname:cluuid:pid:xid:nid:opc "
1758                "%s:%s:%d:"LPU64":%s:%d\n", cfs_curproc_comm(),
1759                imp->imp_obd->obd_uuid.uuid,
1760                lustre_msg_get_status(req->rq_reqmsg), req->rq_xid,
1761                libcfs_nid2str(imp->imp_connection->c_peer.nid),
1762                lustre_msg_get_opc(req->rq_reqmsg));
1763
1764         /* Mark phase here for a little debug help */
1765         req->rq_phase = RQ_PHASE_RPC;
1766
1767         spin_lock(&imp->imp_lock);
1768         req->rq_import_generation = imp->imp_generation;
1769 restart:
1770         if (ptlrpc_import_delay_req(imp, req, &rc)) {
1771                 list_del(&req->rq_list);
1772
1773                 list_add_tail(&req->rq_list, &imp->imp_delayed_list);
1774                 spin_unlock(&imp->imp_lock);
1775
1776                 DEBUG_REQ(D_HA, req, "\"%s\" waiting for recovery: (%s != %s)",
1777                           cfs_curproc_comm(),
1778                           ptlrpc_import_state_name(req->rq_send_state),
1779                           ptlrpc_import_state_name(imp->imp_state));
1780                 lwi = LWI_INTR(interrupted_request, req);
1781                 rc = l_wait_event(req->rq_reply_waitq,
1782                                   (req->rq_send_state == imp->imp_state ||
1783                                    req->rq_err || req->rq_intr),
1784                                   &lwi);
1785                 DEBUG_REQ(D_HA, req, "\"%s\" awake: (%s == %s or %d/%d == 1)",
1786                           cfs_curproc_comm(),
1787                           ptlrpc_import_state_name(imp->imp_state),
1788                           ptlrpc_import_state_name(req->rq_send_state),
1789                           req->rq_err, req->rq_intr);
1790
1791                 spin_lock(&imp->imp_lock);
1792                 list_del_init(&req->rq_list);
1793
1794                 if (req->rq_err) {
1795                         rc = -EIO;
1796                 }
1797                 else if (req->rq_intr) {
1798                         rc = -EINTR;
1799                 }
1800                 else if (req->rq_no_resend) {
1801                         spin_unlock(&imp->imp_lock);
1802                         GOTO(out, rc = -ETIMEDOUT);
1803                 }
1804                 else {
1805                         GOTO(restart, rc);
1806                 }
1807         }
1808
1809         if (rc != 0) {
1810                 list_del_init(&req->rq_list);
1811                 spin_unlock(&imp->imp_lock);
1812                 req->rq_status = rc; // XXX this ok?
1813                 GOTO(out, rc);
1814         }
1815
1816         if (req->rq_resend) {
1817                 lustre_msg_add_flags(req->rq_reqmsg, MSG_RESENT);
1818
1819                 if (req->rq_bulk != NULL) {
1820                         ptlrpc_unregister_bulk (req);
1821
1822                         /* bulk requests are supposed to be
1823                          * idempotent, so we are free to bump the xid
1824                          * here, which we need to do before
1825                          * registering the bulk again (bug 6371).
1826                          * print the old xid first for sanity.
1827                          */
1828                         DEBUG_REQ(D_HA, req, "bumping xid for bulk: ");
1829                         req->rq_xid = ptlrpc_next_xid();
1830                 }
1831
1832                 DEBUG_REQ(D_HA, req, "resending: ");
1833         }
1834
1835         /* XXX this is the same as ptlrpc_set_wait */
1836         LASSERT(list_empty(&req->rq_list));
1837         list_add_tail(&req->rq_list, &imp->imp_sending_list);
1838         spin_unlock(&imp->imp_lock);
1839
1840         rc = sptlrpc_req_refresh_ctx(req, 0);
1841         if (rc) {
1842                 if (req->rq_err) {
1843                         /* we got fatal ctx refresh error, directly jump out
1844                          * thus we can pass back the actual error code.
1845                          */
1846                         spin_lock(&imp->imp_lock);
1847                         list_del_init(&req->rq_list);
1848                         spin_unlock(&imp->imp_lock);
1849
1850                         CERROR("Failed to refresh ctx of req %p: %d\n", req, rc);
1851                         GOTO(out, rc);
1852                 }
1853                 /* simulating we got error during send rpc */
1854                 goto after_send;
1855         }
1856
1857         rc = ptl_send_rpc(req, 0);
1858         if (rc) {
1859                 DEBUG_REQ(D_HA, req, "send failed (%d); recovering", rc);
1860                 timeout = CFS_TICK;
1861         } else {
1862                 timeout = cfs_timeout_cap(cfs_time_seconds(req->rq_timeout));
1863                 DEBUG_REQ(D_NET, req,
1864                           "-- sleeping for "CFS_DURATION_T" jiffies", timeout);
1865         }
1866 repeat:
1867         lwi = LWI_TIMEOUT_INTR(timeout, expired_request, interrupted_request,
1868                                req);
1869         rc = l_wait_event(req->rq_reply_waitq, ptlrpc_check_reply(req), &lwi);
1870         if (rc == -ETIMEDOUT && ptlrpc_check_and_wait_suspend(req))
1871                 goto repeat;
1872
1873 after_send:
1874         CDEBUG(D_RPCTRACE, "Completed RPC pname:cluuid:pid:xid:nid:opc "
1875                "%s:%s:%d:"LPU64":%s:%d\n", cfs_curproc_comm(),
1876                imp->imp_obd->obd_uuid.uuid,
1877                lustre_msg_get_status(req->rq_reqmsg), req->rq_xid,
1878                libcfs_nid2str(imp->imp_connection->c_peer.nid),
1879                lustre_msg_get_opc(req->rq_reqmsg));
1880
1881         spin_lock(&imp->imp_lock);
1882         list_del_init(&req->rq_list);
1883         spin_unlock(&imp->imp_lock);
1884
1885         /* If the reply was received normally, this just grabs the spinlock
1886          * (ensuring the reply callback has returned), sees that
1887          * req->rq_receiving_reply is clear and returns. */
1888         ptlrpc_unregister_reply (req);
1889
1890         if (req->rq_err)
1891                 GOTO(out, rc = -EIO);
1892
1893         /* Resend if we need to, unless we were interrupted. */
1894         if (req->rq_resend && !req->rq_intr) {
1895                 /* ...unless we were specifically told otherwise. */
1896                 if (req->rq_no_resend)
1897                         GOTO(out, rc = -ETIMEDOUT);
1898                 spin_lock(&imp->imp_lock);
1899                 goto restart;
1900         }
1901
1902         if (req->rq_intr) {
1903                 /* Should only be interrupted if we timed out. */
1904                 if (!req->rq_timedout)
1905                         DEBUG_REQ(D_ERROR, req,
1906                                   "rq_intr set but rq_timedout not");
1907                 GOTO(out, rc = -EINTR);
1908         }
1909
1910         if (req->rq_timedout) {                 /* non-recoverable timeout */
1911                 GOTO(out, rc = -ETIMEDOUT);
1912         }
1913
1914         if (!req->rq_replied) {
1915                 /* How can this be? -eeb */
1916                 DEBUG_REQ(D_ERROR, req, "!rq_replied: ");
1917                 LBUG();
1918                 GOTO(out, rc = req->rq_status);
1919         }
1920
1921         rc = after_reply(req);
1922         /* NB may return +ve success rc */
1923         if (req->rq_resend) {
1924                 spin_lock(&imp->imp_lock);
1925                 goto restart;
1926         }
1927
1928  out:
1929         if (req->rq_bulk != NULL) {
1930                 if (rc >= 0) {
1931                         /* success so far.  Note that anything going wrong
1932                          * with bulk now, is EXTREMELY strange, since the
1933                          * server must have believed that the bulk
1934                          * tranferred OK before she replied with success to
1935                          * me. */
1936                         lwi = LWI_TIMEOUT(timeout, NULL, NULL);
1937                         brc = l_wait_event(req->rq_reply_waitq,
1938                                            !ptlrpc_bulk_active(req->rq_bulk),
1939                                            &lwi);
1940                         LASSERT(brc == 0 || brc == -ETIMEDOUT);
1941                         if (brc != 0) {
1942                                 LASSERT(brc == -ETIMEDOUT);
1943                                 DEBUG_REQ(D_ERROR, req, "bulk timed out");
1944                                 rc = brc;
1945                         } else if (!req->rq_bulk->bd_success) {
1946                                 DEBUG_REQ(D_ERROR, req, "bulk transfer failed");
1947                                 rc = -EIO;
1948                         }
1949                 }
1950                 if (rc < 0)
1951                         ptlrpc_unregister_bulk (req);
1952         }
1953
1954         LASSERT(!req->rq_receiving_reply);
1955         req->rq_phase = RQ_PHASE_INTERPRET;
1956
1957         atomic_dec(&imp->imp_inflight);
1958         cfs_waitq_signal(&imp->imp_recovery_waitq);
1959         RETURN(rc);
1960 }
1961
1962 struct ptlrpc_replay_async_args {
1963         int praa_old_state;
1964         int praa_old_status;
1965 };
1966
1967 static int ptlrpc_replay_interpret(struct ptlrpc_request *req,
1968                                     void * data, int rc)
1969 {
1970         struct ptlrpc_replay_async_args *aa = data;
1971         struct obd_import *imp = req->rq_import;
1972
1973         ENTRY;
1974         atomic_dec(&imp->imp_replay_inflight);
1975
1976         if (!req->rq_replied) {
1977                 CERROR("request replay timed out, restarting recovery\n");
1978                 GOTO(out, rc = -ETIMEDOUT);
1979         }
1980
1981         if (lustre_msg_get_type(req->rq_repmsg) == PTL_RPC_MSG_ERR &&
1982             (lustre_msg_get_status(req->rq_repmsg) == -ENOTCONN ||
1983              lustre_msg_get_status(req->rq_repmsg) == -ENODEV))
1984                 GOTO(out, rc = lustre_msg_get_status(req->rq_repmsg));
1985
1986         /* The transno had better not change over replay. */
1987         LASSERT(lustre_msg_get_transno(req->rq_reqmsg) ==
1988                 lustre_msg_get_transno(req->rq_repmsg));
1989
1990         DEBUG_REQ(D_HA, req, "got rep");
1991
1992         /* let the callback do fixups, possibly including in the request */
1993         if (req->rq_replay_cb)
1994                 req->rq_replay_cb(req);
1995
1996         if (req->rq_replied &&
1997             lustre_msg_get_status(req->rq_repmsg) != aa->praa_old_status) {
1998                 DEBUG_REQ(D_ERROR, req, "status %d, old was %d",
1999                           lustre_msg_get_status(req->rq_repmsg),
2000                           aa->praa_old_status);
2001         } else {
2002                 /* Put it back for re-replay. */
2003                 lustre_msg_set_status(req->rq_repmsg, aa->praa_old_status);
2004         }
2005
2006         /*
2007          * Errors while replay can set transno to 0, but
2008          * imp_last_replay_transno shouldn't be set to 0 anyway
2009          */
2010         if (req->rq_transno > 0) {
2011                 spin_lock(&imp->imp_lock);
2012                 LASSERT(req->rq_transno <= imp->imp_last_replay_transno);
2013                 imp->imp_last_replay_transno = req->rq_transno;
2014                 spin_unlock(&imp->imp_lock);
2015         } else
2016                 CERROR("Transno is 0 during replay!\n");
2017         /* continue with recovery */
2018         rc = ptlrpc_import_recovery_state_machine(imp);
2019  out:
2020         req->rq_send_state = aa->praa_old_state;
2021
2022         if (rc != 0)
2023                 /* this replay failed, so restart recovery */
2024                 ptlrpc_connect_import(imp, NULL);
2025
2026         RETURN(rc);
2027 }
2028
2029 int ptlrpc_replay_req(struct ptlrpc_request *req)
2030 {
2031         struct ptlrpc_replay_async_args *aa;
2032         ENTRY;
2033
2034         LASSERT(req->rq_import->imp_state == LUSTRE_IMP_REPLAY);
2035
2036         /* Not handling automatic bulk replay yet (or ever?) */
2037         LASSERT(req->rq_bulk == NULL);
2038
2039         DEBUG_REQ(D_HA, req, "REPLAY");
2040
2041         LASSERT (sizeof (*aa) <= sizeof (req->rq_async_args));
2042         aa = (struct ptlrpc_replay_async_args *)&req->rq_async_args;
2043         memset(aa, 0, sizeof *aa);
2044
2045         /* Prepare request to be resent with ptlrpcd */
2046         aa->praa_old_state = req->rq_send_state;
2047         req->rq_send_state = LUSTRE_IMP_REPLAY;
2048         req->rq_phase = RQ_PHASE_NEW;
2049         aa->praa_old_status = lustre_msg_get_status(req->rq_repmsg);
2050         req->rq_status = 0;
2051
2052         req->rq_interpret_reply = ptlrpc_replay_interpret;
2053         atomic_inc(&req->rq_import->imp_replay_inflight);
2054         ptlrpc_request_addref(req); /* ptlrpcd needs a ref */
2055
2056         ptlrpcd_add_req(req);
2057         RETURN(0);
2058 }
2059
2060 void ptlrpc_abort_inflight(struct obd_import *imp)
2061 {
2062         struct list_head *tmp, *n;
2063         ENTRY;
2064
2065         /* Make sure that no new requests get processed for this import.
2066          * ptlrpc_{queue,set}_wait must (and does) hold imp_lock while testing
2067          * this flag and then putting requests on sending_list or delayed_list.
2068          */
2069         spin_lock(&imp->imp_lock);
2070
2071         /* XXX locking?  Maybe we should remove each request with the list
2072          * locked?  Also, how do we know if the requests on the list are
2073          * being freed at this time?
2074          */
2075         list_for_each_safe(tmp, n, &imp->imp_sending_list) {
2076                 struct ptlrpc_request *req =
2077                         list_entry(tmp, struct ptlrpc_request, rq_list);
2078
2079                 DEBUG_REQ(D_RPCTRACE, req, "inflight");
2080
2081                 spin_lock (&req->rq_lock);
2082                 if (req->rq_import_generation < imp->imp_generation) {
2083                         req->rq_err = 1;
2084                         ptlrpc_wake_client_req(req);
2085                 }
2086                 spin_unlock (&req->rq_lock);
2087         }
2088
2089         list_for_each_safe(tmp, n, &imp->imp_delayed_list) {
2090                 struct ptlrpc_request *req =
2091                         list_entry(tmp, struct ptlrpc_request, rq_list);
2092
2093                 DEBUG_REQ(D_RPCTRACE, req, "aborting waiting req");
2094
2095                 spin_lock (&req->rq_lock);
2096                 if (req->rq_import_generation < imp->imp_generation) {
2097                         req->rq_err = 1;
2098                         ptlrpc_wake_client_req(req);
2099                 }
2100                 spin_unlock (&req->rq_lock);
2101         }
2102
2103         /* Last chance to free reqs left on the replay list, but we
2104          * will still leak reqs that haven't committed.  */
2105         if (imp->imp_replayable)
2106                 ptlrpc_free_committed(imp);
2107
2108         spin_unlock(&imp->imp_lock);
2109
2110         EXIT;
2111 }
2112
2113 static __u64 ptlrpc_last_xid = 0;
2114 spinlock_t ptlrpc_last_xid_lock;
2115
2116 __u64 ptlrpc_next_xid(void)
2117 {
2118         __u64 tmp;
2119         spin_lock(&ptlrpc_last_xid_lock);
2120         tmp = ++ptlrpc_last_xid;
2121         spin_unlock(&ptlrpc_last_xid_lock);
2122         return tmp;
2123 }
2124
2125 __u64 ptlrpc_sample_next_xid(void)
2126 {
2127         __u64 tmp;
2128         spin_lock(&ptlrpc_last_xid_lock);
2129         tmp = ptlrpc_last_xid + 1;
2130         spin_unlock(&ptlrpc_last_xid_lock);
2131         return tmp;
2132 }
2133 EXPORT_SYMBOL(ptlrpc_sample_next_xid);