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         } else if (imp->imp_invalid && (!imp->imp_recon_bk ||
 667                                         imp->imp_obd->obd_no_recov)) {
 668                 /* If the import has been invalidated (such as by an OST
 669                  * failure), and if the import(MGC) tried all of its connection
 670                  * list (Bug 13464), the request must fail with -ESHUTDOWN.
 671                  * This indicates the requests should be discarded; an -EIO
 672                  * may result in a resend of the request. */
 673                 if (!imp->imp_deactive)
 674                           DEBUG_REQ(D_ERROR, req, "IMP_INVALID");
 675                 *status = -ESHUTDOWN; /* bz 12940 */
 676         } else if (req->rq_import_generation != imp->imp_generation) {
 677                 DEBUG_REQ(D_ERROR, req, "req wrong generation:");
 678                 *status = -EIO;
 679         } else if (req->rq_send_state != imp->imp_state) {
 680                 if (imp->imp_obd->obd_no_recov)
 681                         *status = -ESHUTDOWN;
 682                 else if (imp->imp_dlm_fake || req->rq_no_delay)
 683                         *status = -EWOULDBLOCK;
 684                 else
 685                         delay = 1;
 686         }
 687
 688         RETURN(delay);
 689 }
 690
 691 static int ptlrpc_check_reply(struct ptlrpc_request *req)
 692 {
 693         int rc = 0;
 694         ENTRY;
 695
 696         /* serialise with network callback */
 697         spin_lock(&req->rq_lock);
 698
 699         if (req->rq_replied)
 700                 GOTO(out, rc = 1);
 701
 702         if (req->rq_net_err && !req->rq_timedout) {
 703                 spin_unlock(&req->rq_lock);
 704                 rc = ptlrpc_expire_one_request(req);
 705                 spin_lock(&req->rq_lock);
 706                 GOTO(out, rc);
 707         }
 708
 709         if (req->rq_err)
 710                 GOTO(out, rc = 1);
 711
 712         if (req->rq_resend)
 713                 GOTO(out, rc = 1);
 714
 715         if (req->rq_restart)
 716                 GOTO(out, rc = 1);
 717         EXIT;
 718  out:
 719         spin_unlock(&req->rq_lock);
 720         DEBUG_REQ(D_NET, req, "rc = %d for", rc);
 721         return rc;
 722 }
 723
 724 static int ptlrpc_check_status(struct ptlrpc_request *req)
 725 {
 726         int err;
 727         ENTRY;
 728
 729         err = lustre_msg_get_status(req->rq_repmsg);
 730         if (lustre_msg_get_type(req->rq_repmsg) == PTL_RPC_MSG_ERR) {
 731                 struct obd_import *imp = req->rq_import;
 732                 __u32 opc = lustre_msg_get_opc(req->rq_reqmsg);
 733                 LCONSOLE_ERROR_MSG(0x011,"an error occurred while communicating"
 734                                 " with %s. The %s operation failed with %d\n",
 735                                 libcfs_nid2str(imp->imp_connection->c_peer.nid),
 736                                 ll_opcode2str(opc), err);
 737                 RETURN(err < 0 ? err : -EINVAL);
 738         }
 739
 740         if (err < 0) {
 741                 DEBUG_REQ(D_INFO, req, "status is %d", err);
 742         } else if (err > 0) {
 743                 /* XXX: translate this error from net to host */
 744                 DEBUG_REQ(D_INFO, req, "status is %d", err);
 745         }
 746
 747         RETURN(err);
 748 }
 749
 750 static int after_reply(struct ptlrpc_request *req)
 751 {
 752         struct obd_import *imp = req->rq_import;
 753         struct obd_device *obd = req->rq_import->imp_obd;
 754         int rc;
 755         struct timeval work_start;
 756         long timediff;
 757         ENTRY;
 758
 759         LASSERT(!req->rq_receiving_reply);
 760         LASSERT(obd);
 761         LASSERT(req->rq_nob_received <= req->rq_repbuf_len);
 762
 763         /* NB Until this point, the whole of the incoming message,
 764          * including buflens, status etc is in the sender's byte order. */
 765
 766         /* Clear reply swab mask; this is a new reply in sender's byte order */
 767         req->rq_rep_swab_mask = 0;
 768
 769         rc = sptlrpc_cli_unwrap_reply(req);
 770         if (rc) {
 771                 DEBUG_REQ(D_ERROR, req, "unwrap reply failed (%d):", rc);
 772                 RETURN(rc);
 773         }
 774
 775         /* security layer unwrap might ask resend this request */
 776         if (req->rq_resend)
 777                 RETURN(0);
 778
 779         rc = lustre_unpack_msg(req->rq_repmsg, req->rq_replen);
 780         if (rc) {
 781                 DEBUG_REQ(D_ERROR, req, "unpack_rep failed: %d", rc);
 782                 RETURN(-EPROTO);
 783         }
 784
 785         rc = lustre_unpack_rep_ptlrpc_body(req, MSG_PTLRPC_BODY_OFF);
 786         if (rc) {
 787                 DEBUG_REQ(D_ERROR, req, "unpack ptlrpc body failed: %d", rc);
 788                 RETURN(-EPROTO);
 789         }
 790
 791         do_gettimeofday(&work_start);
 792         timediff = cfs_timeval_sub(&work_start, &req->rq_arrival_time, NULL);
 793         if (obd->obd_svc_stats != NULL)
 794                 lprocfs_counter_add(obd->obd_svc_stats, PTLRPC_REQWAIT_CNTR,
 795                                     timediff);
 796
 797         if (lustre_msg_get_type(req->rq_repmsg) != PTL_RPC_MSG_REPLY &&
 798             lustre_msg_get_type(req->rq_repmsg) != PTL_RPC_MSG_ERR) {
 799                 DEBUG_REQ(D_ERROR, req, "invalid packet received (type=%u)",
 800                           lustre_msg_get_type(req->rq_repmsg));
 801                 RETURN(-EPROTO);
 802         }
 803
 804         rc = ptlrpc_check_status(req);
 805         imp->imp_connect_error = rc;
 806
 807         if (rc) {
 808                 /* Either we've been evicted, or the server has failed for
 809                  * some reason. Try to reconnect, and if that fails, punt to
 810                  * the upcall. */
 811                 if (ll_rpc_recoverable_error(rc)) {
 812                         if (req->rq_send_state != LUSTRE_IMP_FULL ||
 813                             imp->imp_obd->obd_no_recov || imp->imp_dlm_fake) {
 814                                 RETURN(rc);
 815                         }
 816                         ptlrpc_request_handle_notconn(req);
 817                         RETURN(rc);
 818                 }
 819         } else {
 820                 /* Let's look if server send slv. Do it only for RPC with
 821                  * rc == 0. */
 822                 if (imp->imp_obd->obd_namespace) {
 823                         /* Disconnect rpc is sent when namespace is already
 824                          * destroyed. Let's check this and will not try update
 825                          * pool. */
 826                         ldlm_cli_update_pool(req);
 827                 }
 828         }
 829
 830         /* Store transno in reqmsg for replay. */
 831         req->rq_transno = lustre_msg_get_transno(req->rq_repmsg);
 832         lustre_msg_set_transno(req->rq_reqmsg, req->rq_transno);
 833
 834         if (req->rq_import->imp_replayable) {
 835                 spin_lock(&imp->imp_lock);
 836                 /* no point in adding already-committed requests to the replay
 837                  * list, we will just remove them immediately. b=9829 */
 838                 if (req->rq_transno != 0 &&
 839                     (req->rq_transno >
 840                      lustre_msg_get_last_committed(req->rq_repmsg) ||
 841                      req->rq_replay))
 842                         ptlrpc_retain_replayable_request(req, imp);
 843                 else if (req->rq_commit_cb != NULL) {
 844                         spin_unlock(&imp->imp_lock);
 845                         req->rq_commit_cb(req);
 846                         spin_lock(&imp->imp_lock);
 847                 }
 848
 849                 /* Replay-enabled imports return commit-status information. */
 850                 if (lustre_msg_get_last_committed(req->rq_repmsg)) {
 851                         imp->imp_peer_committed_transno =
 852                                 lustre_msg_get_last_committed(req->rq_repmsg);
 853                 }
 854                 ptlrpc_free_committed(imp);
 855                 spin_unlock(&imp->imp_lock);
 856         }
 857
 858         RETURN(rc);
 859 }
 860
 861 static int ptlrpc_send_new_req(struct ptlrpc_request *req)
 862 {
 863         struct obd_import     *imp;
 864         int rc;
 865         ENTRY;
 866
 867         LASSERT(req->rq_phase == RQ_PHASE_NEW);
 868         if (req->rq_sent && (req->rq_sent > CURRENT_SECONDS))
 869                 RETURN (0);
 870
 871         req->rq_phase = RQ_PHASE_RPC;
 872
 873         imp = req->rq_import;
 874         spin_lock(&imp->imp_lock);
 875
 876         req->rq_import_generation = imp->imp_generation;
 877
 878         if (ptlrpc_import_delay_req(imp, req, &rc)) {
 879                 spin_lock (&req->rq_lock);
 880                 req->rq_waiting = 1;
 881                 spin_unlock (&req->rq_lock);
 882
 883                 DEBUG_REQ(D_HA, req, "req from PID %d waiting for recovery: "
 884                           "(%s != %s)",
 885                           lustre_msg_get_status(req->rq_reqmsg) ,
 886                           ptlrpc_import_state_name(req->rq_send_state),
 887                           ptlrpc_import_state_name(imp->imp_state));
 888                 LASSERT(list_empty (&req->rq_list));
 889
 890                 list_add_tail(&req->rq_list, &imp->imp_delayed_list);
 891                 spin_unlock(&imp->imp_lock);
 892                 RETURN(0);
 893         }
 894
 895         if (rc != 0) {
 896                 spin_unlock(&imp->imp_lock);
 897                 req->rq_status = rc;
 898                 req->rq_phase = RQ_PHASE_INTERPRET;
 899                 RETURN(rc);
 900         }
 901
 902         /* XXX this is the same as ptlrpc_queue_wait */
 903         LASSERT(list_empty(&req->rq_list));
 904         list_add_tail(&req->rq_list, &imp->imp_sending_list);
 905         spin_unlock(&imp->imp_lock);
 906
 907         lustre_msg_set_status(req->rq_reqmsg, cfs_curproc_pid());
 908
 909         rc = sptlrpc_req_refresh_ctx(req, -1);
 910         if (rc) {
 911                 if (req->rq_err) {
 912                         req->rq_status = rc;
 913                         RETURN(1);
 914                 } else {
 915                         /* here begins timeout counting */
 916                         req->rq_sent = CURRENT_SECONDS;
 917                         req->rq_wait_ctx = 1;
 918                         RETURN(0);
 919                 }
 920         }
 921
 922         CDEBUG(D_RPCTRACE, "Sending RPC pname:cluuid:pid:xid:nid:opc"
 923                " %s:%s:%d:"LPU64":%s:%d\n", cfs_curproc_comm(),
 924                imp->imp_obd->obd_uuid.uuid,
 925                lustre_msg_get_status(req->rq_reqmsg), req->rq_xid,
 926                libcfs_nid2str(imp->imp_connection->c_peer.nid),
 927                lustre_msg_get_opc(req->rq_reqmsg));
 928
 929         rc = ptl_send_rpc(req, 0);
 930         if (rc) {
 931                 DEBUG_REQ(D_HA, req, "send failed (%d); expect timeout", rc);
 932                 req->rq_net_err = 1;
 933                 RETURN(rc);
 934         }
 935         RETURN(0);
 936 }
 937
 938 /* this sends any unsent RPCs in @set and returns TRUE if all are sent */
 939 int ptlrpc_check_set(struct ptlrpc_request_set *set)
 940 {
 941         struct list_head *tmp;
 942         int force_timer_recalc = 0;
 943         ENTRY;
 944
 945         if (set->set_remaining == 0)
 946                 RETURN(1);
 947
 948         list_for_each(tmp, &set->set_requests) {
 949                 struct ptlrpc_request *req =
 950                         list_entry(tmp, struct ptlrpc_request, rq_set_chain);
 951                 struct obd_import *imp = req->rq_import;
 952                 int rc = 0;
 953
 954                 if (req->rq_phase == RQ_PHASE_NEW &&
 955                     ptlrpc_send_new_req(req)) {
 956                         force_timer_recalc = 1;
 957                 }
 958                 /* delayed send - skip */
 959                 if (req->rq_phase == RQ_PHASE_NEW && req->rq_sent)
 960                         continue;
 961
 962                 if (!(req->rq_phase == RQ_PHASE_RPC ||
 963                       req->rq_phase == RQ_PHASE_BULK ||
 964                       req->rq_phase == RQ_PHASE_INTERPRET ||
 965                       req->rq_phase == RQ_PHASE_COMPLETE)) {
 966                         DEBUG_REQ(D_ERROR, req, "bad phase %x", req->rq_phase);
 967                         LBUG();
 968                 }
 969
 970                 if (req->rq_phase == RQ_PHASE_COMPLETE)
 971                         continue;
 972
 973                 if (req->rq_phase == RQ_PHASE_INTERPRET)
 974                         GOTO(interpret, req->rq_status);
 975
 976                 if (req->rq_net_err && !req->rq_timedout)
 977                         ptlrpc_expire_one_request(req);
 978
 979                 if (req->rq_err) {
 980                         ptlrpc_unregister_reply(req);
 981                         req->rq_replied = 0;
 982                         if (req->rq_status == 0)
 983                                 req->rq_status = -EIO;
 984                         req->rq_phase = RQ_PHASE_INTERPRET;
 985
 986                         spin_lock(&imp->imp_lock);
 987                         list_del_init(&req->rq_list);
 988                         spin_unlock(&imp->imp_lock);
 989
 990                         GOTO(interpret, req->rq_status);
 991                 }
 992
 993                 /* ptlrpc_queue_wait->l_wait_event guarantees that rq_intr
 994                  * will only be set after rq_timedout, but the oig waiting
 995                  * path sets rq_intr irrespective of whether ptlrpcd has
 996                  * seen a timeout.  our policy is to only interpret
 997                  * interrupted rpcs after they have timed out */
 998                 if (req->rq_intr && (req->rq_timedout || req->rq_waiting ||
 999                                      req->rq_wait_ctx)) {
1000                         /* NB could be on delayed list */
1001                         ptlrpc_unregister_reply(req);
1002                         req->rq_status = -EINTR;
1003                         req->rq_phase = RQ_PHASE_INTERPRET;
1004
1005                         spin_lock(&imp->imp_lock);
1006                         list_del_init(&req->rq_list);
1007                         spin_unlock(&imp->imp_lock);
1008
1009                         GOTO(interpret, req->rq_status);
1010                 }
1011
1012                 if (req->rq_phase == RQ_PHASE_RPC) {
1013                         if (req->rq_timedout || req->rq_resend ||
1014                             req->rq_waiting || req->rq_wait_ctx) {
1015                                 int status;
1016
1017                                 /* rq_wait_ctx is only touched in ptlrpcd,
1018                                  * no lock needed here.
1019                                  */
1020                                 if (req->rq_wait_ctx)
1021                                         goto check_ctx;
1022
1023                                 ptlrpc_unregister_reply(req);
1024
1025                                 spin_lock(&imp->imp_lock);
1026
1027                                 if (ptlrpc_import_delay_req(imp, req, &status)){
1028                                         spin_unlock(&imp->imp_lock);
1029                                         continue;
1030                                 }
1031
1032                                 list_del_init(&req->rq_list);
1033                                 if (status != 0)  {
1034                                         req->rq_status = status;
1035                                         req->rq_phase = RQ_PHASE_INTERPRET;
1036                                         spin_unlock(&imp->imp_lock);
1037                                         GOTO(interpret, req->rq_status);
1038                                 }
1039                                 if (req->rq_no_resend && !req->rq_wait_ctx) {
1040                                         req->rq_status = -ENOTCONN;
1041                                         req->rq_phase = RQ_PHASE_INTERPRET;
1042                                         spin_unlock(&imp->imp_lock);
1043                                         GOTO(interpret, req->rq_status);
1044                                 }
1045                                 list_add_tail(&req->rq_list,
1046                                               &imp->imp_sending_list);
1047
1048                                 spin_unlock(&imp->imp_lock);
1049
1050                                 req->rq_waiting = 0;
1051                                 if (req->rq_resend) {
1052                                         lustre_msg_add_flags(req->rq_reqmsg,
1053                                                              MSG_RESENT);
1054                                         if (req->rq_bulk) {
1055                                                 __u64 old_xid = req->rq_xid;
1056
1057                                                 ptlrpc_unregister_bulk (req);
1058
1059                                                 /* ensure previous bulk fails */
1060                                                 req->rq_xid = ptlrpc_next_xid();
1061                                                 CDEBUG(D_HA, "resend bulk "
1062                                                        "old x"LPU64
1063                                                        " new x"LPU64"\n",
1064                                                        old_xid, req->rq_xid);
1065                                         }
1066                                 }
1067 check_ctx:
1068                                 status = sptlrpc_req_refresh_ctx(req, -1);
1069                                 if (status) {
1070                                         if (req->rq_err) {
1071                                                 req->rq_status = status;
1072                                                 force_timer_recalc = 1;
1073                                         }
1074                                         if (!req->rq_wait_ctx) {
1075                                                 /* begins timeout counting */
1076                                                 req->rq_sent = CURRENT_SECONDS;
1077                                                 req->rq_wait_ctx = 1;
1078                                         }
1079                                         continue;
1080                                 } else {
1081                                         req->rq_sent = 0;
1082                                         req->rq_wait_ctx = 0;
1083                                 }
1084
1085                                 rc = ptl_send_rpc(req, 0);
1086                                 if (rc) {
1087                                         DEBUG_REQ(D_HA, req, "send failed (%d)",
1088                                                   rc);
1089                                         force_timer_recalc = 1;
1090                                         req->rq_net_err = 1;
1091                                 }
1092                                 /* need to reset the timeout */
1093                                 force_timer_recalc = 1;
1094                         }
1095
1096                         /* Still waiting for a reply? */
1097                         if (ptlrpc_client_receiving_reply(req))
1098                                 continue;
1099
1100                         /* Did we actually receive a reply? */
1101                         if (!ptlrpc_client_replied(req))
1102                                 continue;
1103
1104                         spin_lock(&imp->imp_lock);
1105                         list_del_init(&req->rq_list);
1106                         spin_unlock(&imp->imp_lock);
1107
1108                         req->rq_status = after_reply(req);
1109                         if (req->rq_resend) {
1110                                 /* Add this req to the delayed list so
1111                                    it can be errored if the import is
1112                                    evicted after recovery. */
1113                                 spin_lock(&imp->imp_lock);
1114                                 list_add_tail(&req->rq_list,
1115                                               &imp->imp_delayed_list);
1116                                 spin_unlock(&imp->imp_lock);
1117                                 continue;
1118                         }
1119
1120                         /* If there is no bulk associated with this request,
1121                          * then we're done and should let the interpreter
1122                          * process the reply.  Similarly if the RPC returned
1123                          * an error, and therefore the bulk will never arrive.
1124                          */
1125                         if (req->rq_bulk == NULL || req->rq_status != 0) {
1126                                 req->rq_phase = RQ_PHASE_INTERPRET;
1127                                 GOTO(interpret, req->rq_status);
1128                         }
1129
1130                         req->rq_phase = RQ_PHASE_BULK;
1131                 }
1132
1133                 LASSERT(req->rq_phase == RQ_PHASE_BULK);
1134                 if (ptlrpc_bulk_active(req->rq_bulk))
1135                         continue;
1136
1137                 if (!req->rq_bulk->bd_success) {
1138                         /* The RPC reply arrived OK, but the bulk screwed
1139                          * up!  Dead wierd since the server told us the RPC
1140                          * was good after getting the REPLY for her GET or
1141                          * the ACK for her PUT. */
1142                         DEBUG_REQ(D_ERROR, req, "bulk transfer failed");
1143                         LBUG();
1144                 }
1145
1146                 req->rq_phase = RQ_PHASE_INTERPRET;
1147
1148         interpret:
1149                 LASSERT(req->rq_phase == RQ_PHASE_INTERPRET);
1150                 LASSERT(!req->rq_receiving_reply);
1151
1152                 ptlrpc_unregister_reply(req);
1153                 if (req->rq_bulk != NULL)
1154                         ptlrpc_unregister_bulk (req);
1155
1156                 if (req->rq_interpret_reply != NULL) {
1157                         int (*interpreter)(struct ptlrpc_request *,void *,int) =
1158                                 req->rq_interpret_reply;
1159                         req->rq_status = interpreter(req, &req->rq_async_args,
1160                                                      req->rq_status);
1161                 }
1162                 req->rq_phase = RQ_PHASE_COMPLETE;
1163
1164                 CDEBUG(D_RPCTRACE, "Completed RPC pname:cluuid:pid:xid:nid:"
1165                        "opc %s:%s:%d:"LPU64":%s:%d\n", cfs_curproc_comm(),
1166                        imp->imp_obd->obd_uuid.uuid,
1167                        lustre_msg_get_status(req->rq_reqmsg), req->rq_xid,
1168                        libcfs_nid2str(imp->imp_connection->c_peer.nid),
1169                        lustre_msg_get_opc(req->rq_reqmsg));
1170
1171                 set->set_remaining--;
1172
1173                 atomic_dec(&imp->imp_inflight);
1174                 cfs_waitq_signal(&imp->imp_recovery_waitq);
1175         }
1176
1177         /* If we hit an error, we want to recover promptly. */
1178         RETURN(set->set_remaining == 0 || force_timer_recalc);
1179 }
1180
1181 int ptlrpc_expire_one_request(struct ptlrpc_request *req)
1182 {
1183         struct obd_import *imp = req->rq_import;
1184         int rc = 0;
1185         ENTRY;
1186
1187         DEBUG_REQ(D_ERROR|D_NETERROR, req, "%s (sent at %lu, %lus ago)",
1188                   req->rq_net_err ? "network error" : "timeout",
1189                   (long)req->rq_sent, CURRENT_SECONDS - req->rq_sent);
1190
1191         if (imp != NULL && obd_debug_peer_on_timeout)
1192                 LNetCtl(IOC_LIBCFS_DEBUG_PEER, &imp->imp_connection->c_peer);
1193
1194         spin_lock(&req->rq_lock);
1195         req->rq_timedout = 1;
1196         req->rq_wait_ctx = 0;
1197         spin_unlock(&req->rq_lock);
1198
1199         ptlrpc_unregister_reply (req);
1200
1201         if (obd_dump_on_timeout)
1202                 libcfs_debug_dumplog();
1203
1204         if (req->rq_bulk != NULL)
1205                 ptlrpc_unregister_bulk (req);
1206
1207         if (imp == NULL) {
1208                 DEBUG_REQ(D_HA, req, "NULL import: already cleaned up?");
1209                 RETURN(1);
1210         }
1211
1212         /* The DLM server doesn't want recovery run on its imports. */
1213         if (imp->imp_dlm_fake)
1214                 RETURN(1);
1215
1216         /* If this request is for recovery or other primordial tasks,
1217          * then error it out here. */
1218         if (req->rq_ctx_init || req->rq_ctx_fini ||
1219             req->rq_send_state != LUSTRE_IMP_FULL ||
1220             imp->imp_obd->obd_no_recov) {
1221                 spin_lock(&req->rq_lock);
1222                 req->rq_status = -ETIMEDOUT;
1223                 req->rq_err = 1;
1224                 spin_unlock(&req->rq_lock);
1225                 RETURN(1);
1226         }
1227
1228         /* if request can't be resend we can't wait answer after timeout */
1229         if (req->rq_no_resend) {
1230                 DEBUG_REQ(D_RPCTRACE, req, "TIMEOUT-NORESEND:");
1231                 rc = 1;
1232         }
1233
1234         ptlrpc_fail_import(imp, lustre_msg_get_conn_cnt(req->rq_reqmsg));
1235
1236         RETURN(rc);
1237 }
1238
1239 int ptlrpc_expired_set(void *data)
1240 {
1241         struct ptlrpc_request_set *set = data;
1242         struct list_head          *tmp;
1243         time_t                     now = CURRENT_SECONDS;
1244         ENTRY;
1245
1246         LASSERT(set != NULL);
1247
1248         /* A timeout expired; see which reqs it applies to... */
1249         list_for_each (tmp, &set->set_requests) {
1250                 struct ptlrpc_request *req =
1251                         list_entry(tmp, struct ptlrpc_request, rq_set_chain);
1252
1253                 /* request in-flight? */
1254                 if (!((req->rq_phase == RQ_PHASE_RPC && !req->rq_waiting &&
1255                        !req->rq_resend) ||
1256                       (req->rq_phase == RQ_PHASE_BULK)))
1257                         continue;
1258
1259                 if (req->rq_timedout ||           /* already dealt with */
1260                     req->rq_sent + req->rq_timeout > now) /* not expired */
1261                         continue;
1262
1263                 /* deal with this guy */
1264                 ptlrpc_expire_one_request (req);
1265         }
1266
1267         /* When waiting for a whole set, we always to break out of the
1268          * sleep so we can recalculate the timeout, or enable interrupts
1269          * iff everyone's timed out.
1270          */
1271         RETURN(1);
1272 }
1273
1274 void ptlrpc_mark_interrupted(struct ptlrpc_request *req)
1275 {
1276         spin_lock(&req->rq_lock);
1277         req->rq_intr = 1;
1278         spin_unlock(&req->rq_lock);
1279 }
1280
1281 void ptlrpc_interrupted_set(void *data)
1282 {
1283         struct ptlrpc_request_set *set = data;
1284         struct list_head *tmp;
1285
1286         LASSERT(set != NULL);
1287         CERROR("INTERRUPTED SET %p\n", set);
1288
1289         list_for_each(tmp, &set->set_requests) {
1290                 struct ptlrpc_request *req =
1291                         list_entry(tmp, struct ptlrpc_request, rq_set_chain);
1292
1293                 if (req->rq_phase != RQ_PHASE_RPC)
1294                         continue;
1295
1296                 ptlrpc_mark_interrupted(req);
1297         }
1298 }
1299
1300 int ptlrpc_set_next_timeout(struct ptlrpc_request_set *set)
1301 {
1302         struct list_head      *tmp;
1303         time_t                 now = CURRENT_SECONDS;
1304         time_t                 deadline;
1305         int                    timeout = 0;
1306         struct ptlrpc_request *req;
1307         ENTRY;
1308
1309         SIGNAL_MASK_ASSERT(); /* XXX BUG 1511 */
1310
1311         list_for_each(tmp, &set->set_requests) {
1312                 req = list_entry(tmp, struct ptlrpc_request, rq_set_chain);
1313
1314                 /* request in-flight? */
1315                 if (!((req->rq_phase == RQ_PHASE_RPC && !req->rq_waiting) ||
1316                       (req->rq_phase == RQ_PHASE_BULK) ||
1317                       (req->rq_phase == RQ_PHASE_NEW)))
1318                         continue;
1319
1320                 if (req->rq_timedout)   /* already timed out */
1321                         continue;
1322
1323                 if (req->rq_phase == RQ_PHASE_NEW)
1324                         deadline = req->rq_sent;
1325                 else
1326                         deadline = req->rq_sent + req->rq_timeout;
1327
1328                 if (deadline <= now)    /* actually expired already */
1329                         timeout = 1;    /* ASAP */
1330                 else if (timeout == 0 || timeout > deadline - now)
1331                         timeout = deadline - now;
1332         }
1333         RETURN(timeout);
1334 }
1335
1336 int ptlrpc_set_wait(struct ptlrpc_request_set *set)
1337 {
1338         struct list_head      *tmp;
1339         struct ptlrpc_request *req;
1340         struct l_wait_info     lwi;
1341         int                    rc, timeout;
1342         ENTRY;
1343
1344         if (list_empty(&set->set_requests))
1345                 RETURN(0);
1346
1347         list_for_each(tmp, &set->set_requests) {
1348                 req = list_entry(tmp, struct ptlrpc_request, rq_set_chain);
1349                 if (req->rq_phase == RQ_PHASE_NEW)
1350                         (void)ptlrpc_send_new_req(req);
1351         }
1352
1353         do {
1354                 timeout = ptlrpc_set_next_timeout(set);
1355
1356                 /* wait until all complete, interrupted, or an in-flight
1357                  * req times out */
1358                 CDEBUG(D_RPCTRACE, "set %p going to sleep for %d seconds\n",
1359                        set, timeout);
1360                 lwi = LWI_TIMEOUT_INTR(cfs_time_seconds(timeout ? timeout : 1),
1361                                        ptlrpc_expired_set,
1362                                        ptlrpc_interrupted_set, set);
1363                 rc = l_wait_event(set->set_waitq, ptlrpc_check_set(set), &lwi);
1364
1365                 LASSERT(rc == 0 || rc == -EINTR || rc == -ETIMEDOUT);
1366
1367                 /* -EINTR => all requests have been flagged rq_intr so next
1368                  * check completes.
1369                  * -ETIMEOUTD => someone timed out.  When all reqs have
1370                  * timed out, signals are enabled allowing completion with
1371                  * EINTR.
1372                  * I don't really care if we go once more round the loop in
1373                  * the error cases -eeb. */
1374         } while (rc != 0 || set->set_remaining != 0);
1375
1376         LASSERT(set->set_remaining == 0);
1377
1378         rc = 0;
1379         list_for_each(tmp, &set->set_requests) {
1380                 req = list_entry(tmp, struct ptlrpc_request, rq_set_chain);
1381
1382                 LASSERT(req->rq_phase == RQ_PHASE_COMPLETE);
1383                 if (req->rq_status != 0)
1384                         rc = req->rq_status;
1385         }
1386
1387         if (set->set_interpret != NULL) {
1388                 int (*interpreter)(struct ptlrpc_request_set *set,void *,int) =
1389                         set->set_interpret;
1390                 rc = interpreter (set, set->set_arg, rc);
1391         } else {
1392                 struct ptlrpc_set_cbdata *cbdata, *n;
1393                 int err;
1394
1395                 list_for_each_entry_safe(cbdata, n,
1396                                          &set->set_cblist, psc_item) {
1397                         list_del_init(&cbdata->psc_item);
1398                         err = cbdata->psc_interpret(set, cbdata->psc_data, rc);
1399                         if (err && !rc)
1400                                 rc = err;
1401                         OBD_FREE_PTR(cbdata);
1402                 }
1403         }
1404
1405         RETURN(rc);
1406 }
1407
1408 static void __ptlrpc_free_req(struct ptlrpc_request *request, int locked)
1409 {
1410         ENTRY;
1411         if (request == NULL) {
1412                 EXIT;
1413                 return;
1414         }
1415
1416         LASSERTF(!request->rq_receiving_reply, "req %p\n", request);
1417         LASSERTF(request->rq_rqbd == NULL, "req %p\n",request);/* client-side */
1418         LASSERTF(list_empty(&request->rq_list), "req %p\n", request);
1419         LASSERTF(list_empty(&request->rq_set_chain), "req %p\n", request);
1420         LASSERT(request->rq_cli_ctx);
1421
1422         req_capsule_fini(&request->rq_pill);
1423
1424         /* We must take it off the imp_replay_list first.  Otherwise, we'll set
1425          * request->rq_reqmsg to NULL while osc_close is dereferencing it. */
1426         if (request->rq_import != NULL) {
1427                 if (!locked)
1428                         spin_lock(&request->rq_import->imp_lock);
1429                 list_del_init(&request->rq_mod_list);
1430                 list_del_init(&request->rq_replay_list);
1431                 if (!locked)
1432                         spin_unlock(&request->rq_import->imp_lock);
1433         }
1434         LASSERTF(list_empty(&request->rq_replay_list), "req %p\n", request);
1435
1436         if (atomic_read(&request->rq_refcount) != 0) {
1437                 DEBUG_REQ(D_ERROR, request,
1438                           "freeing request with nonzero refcount");
1439                 LBUG();
1440         }
1441
1442         if (request->rq_repbuf != NULL)
1443                 sptlrpc_cli_free_repbuf(request);
1444         if (request->rq_export != NULL) {
1445                 class_export_put(request->rq_export);
1446                 request->rq_export = NULL;
1447         }
1448         if (request->rq_import != NULL) {
1449                 class_import_put(request->rq_import);
1450                 request->rq_import = NULL;
1451         }
1452         if (request->rq_bulk != NULL)
1453                 ptlrpc_free_bulk(request->rq_bulk);
1454
1455         if (request->rq_reqbuf != NULL || request->rq_clrbuf != NULL)
1456                 sptlrpc_cli_free_reqbuf(request);
1457
1458         sptlrpc_req_put_ctx(request, !locked);
1459
1460         if (request->rq_pool)
1461                 __ptlrpc_free_req_to_pool(request);
1462         else
1463                 OBD_FREE(request, sizeof(*request));
1464         EXIT;
1465 }
1466
1467 void ptlrpc_free_req(struct ptlrpc_request *request)
1468 {
1469         __ptlrpc_free_req(request, 0);
1470 }
1471
1472 static int __ptlrpc_req_finished(struct ptlrpc_request *request, int locked);
1473 void ptlrpc_req_finished_with_imp_lock(struct ptlrpc_request *request)
1474 {
1475         LASSERT_SPIN_LOCKED(&request->rq_import->imp_lock);
1476         (void)__ptlrpc_req_finished(request, 1);
1477 }
1478
1479 static int __ptlrpc_req_finished(struct ptlrpc_request *request, int locked)
1480 {
1481         ENTRY;
1482         if (request == NULL)
1483                 RETURN(1);
1484
1485         if (request == LP_POISON ||
1486             request->rq_reqmsg == LP_POISON) {
1487                 CERROR("dereferencing freed request (bug 575)\n");
1488                 LBUG();
1489                 RETURN(1);
1490         }
1491
1492         DEBUG_REQ(D_INFO, request, "refcount now %u",
1493                   atomic_read(&request->rq_refcount) - 1);
1494
1495         if (atomic_dec_and_test(&request->rq_refcount)) {
1496                 __ptlrpc_free_req(request, locked);
1497                 RETURN(1);
1498         }
1499
1500         RETURN(0);
1501 }
1502
1503 void ptlrpc_req_finished(struct ptlrpc_request *request)
1504 {
1505         __ptlrpc_req_finished(request, 0);
1506 }
1507
1508 __u64 ptlrpc_req_xid(struct ptlrpc_request *request)
1509 {
1510         return request->rq_xid;
1511 }
1512 EXPORT_SYMBOL(ptlrpc_req_xid);
1513
1514 /* Disengage the client's reply buffer from the network
1515  * NB does _NOT_ unregister any client-side bulk.
1516  * IDEMPOTENT, but _not_ safe against concurrent callers.
1517  * The request owner (i.e. the thread doing the I/O) must call...
1518  */
1519 void ptlrpc_unregister_reply (struct ptlrpc_request *request)
1520 {
1521         int                rc;
1522         cfs_waitq_t       *wq;
1523         struct l_wait_info lwi;
1524
1525         LASSERT(!in_interrupt ());             /* might sleep */
1526
1527         if (!ptlrpc_client_receiving_reply(request))
1528                 return;
1529
1530         LNetMDUnlink (request->rq_reply_md_h);
1531
1532         /* We have to l_wait_event() whatever the result, to give liblustre
1533          * a chance to run reply_in_callback() */
1534
1535         if (request->rq_set != NULL)
1536                 wq = &request->rq_set->set_waitq;
1537         else
1538                 wq = &request->rq_reply_waitq;
1539
1540         for (;;) {
1541                 /* Network access will complete in finite time but the HUGE
1542                  * timeout lets us CWARN for visibility of sluggish NALs */
1543                 lwi = LWI_TIMEOUT(cfs_time_seconds(300), NULL, NULL);
1544                 rc = l_wait_event (*wq, !ptlrpc_client_receiving_reply(request), &lwi);
1545                 if (rc == 0)
1546                         return;
1547
1548                 LASSERT (rc == -ETIMEDOUT);
1549                 DEBUG_REQ(D_WARNING, request, "Unexpectedly long timeout");
1550         }
1551 }
1552
1553 /* caller must hold imp->imp_lock */
1554 void ptlrpc_free_committed(struct obd_import *imp)
1555 {
1556         struct list_head *tmp, *saved;
1557         struct ptlrpc_request *req;
1558         struct ptlrpc_request *last_req = NULL; /* temporary fire escape */
1559         ENTRY;
1560
1561         LASSERT(imp != NULL);
1562
1563         LASSERT_SPIN_LOCKED(&imp->imp_lock);
1564
1565
1566         if (imp->imp_peer_committed_transno == imp->imp_last_transno_checked &&
1567             imp->imp_generation == imp->imp_last_generation_checked) {
1568                 CDEBUG(D_RPCTRACE, "%s: skip recheck: last_committed "LPU64"\n",
1569                        imp->imp_obd->obd_name, imp->imp_peer_committed_transno);
1570                 return;
1571         }
1572
1573         CDEBUG(D_RPCTRACE, "%s: committing for last_committed "LPU64" gen %d\n",
1574                imp->imp_obd->obd_name, imp->imp_peer_committed_transno,
1575                imp->imp_generation);
1576         imp->imp_last_transno_checked = imp->imp_peer_committed_transno;
1577         imp->imp_last_generation_checked = imp->imp_generation;
1578
1579         list_for_each_safe(tmp, saved, &imp->imp_replay_list) {
1580                 req = list_entry(tmp, struct ptlrpc_request, rq_replay_list);
1581
1582                 /* XXX ok to remove when 1357 resolved - rread 05/29/03  */
1583                 LASSERT(req != last_req);
1584                 last_req = req;
1585
1586                 if (req->rq_import_generation < imp->imp_generation) {
1587                         DEBUG_REQ(D_RPCTRACE, req, "free request with old gen");
1588                         GOTO(free_req, 0);
1589                 }
1590
1591                 if (req->rq_replay) {
1592                         DEBUG_REQ(D_RPCTRACE, req, "keeping (FL_REPLAY)");
1593                         continue;
1594                 }
1595
1596                 /* not yet committed */
1597                 if (req->rq_transno > imp->imp_peer_committed_transno) {
1598                         DEBUG_REQ(D_RPCTRACE, req, "stopping search");
1599                         break;
1600                 }
1601
1602                 DEBUG_REQ(D_RPCTRACE, req, "commit (last_committed "LPU64")",
1603                           imp->imp_peer_committed_transno);
1604 free_req:
1605                 spin_lock(&req->rq_lock);
1606                 req->rq_replay = 0;
1607                 spin_unlock(&req->rq_lock);
1608                 if (req->rq_commit_cb != NULL)
1609                         req->rq_commit_cb(req);
1610                 list_del_init(&req->rq_replay_list);
1611                 __ptlrpc_req_finished(req, 1);
1612         }
1613
1614         EXIT;
1615         return;
1616 }
1617
1618 void ptlrpc_cleanup_client(struct obd_import *imp)
1619 {
1620         ENTRY;
1621         EXIT;
1622         return;
1623 }
1624
1625 void ptlrpc_resend_req(struct ptlrpc_request *req)
1626 {
1627         DEBUG_REQ(D_HA, req, "going to resend");
1628         lustre_msg_set_handle(req->rq_reqmsg, &(struct lustre_handle){ 0 });
1629         req->rq_status = -EAGAIN;
1630
1631         spin_lock(&req->rq_lock);
1632         req->rq_resend = 1;
1633         req->rq_net_err = 0;
1634         req->rq_timedout = 0;
1635         if (req->rq_bulk) {
1636                 __u64 old_xid = req->rq_xid;
1637
1638                 /* ensure previous bulk fails */
1639                 req->rq_xid = ptlrpc_next_xid();
1640                 CDEBUG(D_HA, "resend bulk old x"LPU64" new x"LPU64"\n",
1641                        old_xid, req->rq_xid);
1642         }
1643         ptlrpc_wake_client_req(req);
1644         spin_unlock(&req->rq_lock);
1645 }
1646
1647 /* XXX: this function and rq_status are currently unused */
1648 void ptlrpc_restart_req(struct ptlrpc_request *req)
1649 {
1650         DEBUG_REQ(D_HA, req, "restarting (possibly-)completed request");
1651         req->rq_status = -ERESTARTSYS;
1652
1653         spin_lock(&req->rq_lock);
1654         req->rq_restart = 1;
1655         req->rq_timedout = 0;
1656         ptlrpc_wake_client_req(req);
1657         spin_unlock(&req->rq_lock);
1658 }
1659
1660 static int expired_request(void *data)
1661 {
1662         struct ptlrpc_request *req = data;
1663         ENTRY;
1664
1665         /* some failure can suspend regular timeouts */
1666         if (ptlrpc_check_suspend())
1667                 RETURN(1);
1668
1669         RETURN(ptlrpc_expire_one_request(req));
1670 }
1671
1672 static void interrupted_request(void *data)
1673 {
1674         struct ptlrpc_request *req = data;
1675         DEBUG_REQ(D_HA, req, "request interrupted");
1676         spin_lock(&req->rq_lock);
1677         req->rq_intr = 1;
1678         spin_unlock(&req->rq_lock);
1679 }
1680
1681 struct ptlrpc_request *ptlrpc_request_addref(struct ptlrpc_request *req)
1682 {
1683         ENTRY;
1684         atomic_inc(&req->rq_refcount);
1685         RETURN(req);
1686 }
1687
1688 void ptlrpc_retain_replayable_request(struct ptlrpc_request *req,
1689                                       struct obd_import *imp)
1690 {
1691         struct list_head *tmp;
1692
1693         LASSERT_SPIN_LOCKED(&imp->imp_lock);
1694
1695         /* clear this for new requests that were resent as well
1696            as resent replayed requests. */
1697         lustre_msg_clear_flags(req->rq_reqmsg, MSG_RESENT);
1698
1699         /* don't re-add requests that have been replayed */
1700         if (!list_empty(&req->rq_replay_list))
1701                 return;
1702
1703         lustre_msg_add_flags(req->rq_reqmsg, MSG_REPLAY);
1704
1705         LASSERT(imp->imp_replayable);
1706         /* Balanced in ptlrpc_free_committed, usually. */
1707         ptlrpc_request_addref(req);
1708         list_for_each_prev(tmp, &imp->imp_replay_list) {
1709                 struct ptlrpc_request *iter =
1710                         list_entry(tmp, struct ptlrpc_request, rq_replay_list);
1711
1712                 /* We may have duplicate transnos if we create and then
1713                  * open a file, or for closes retained if to match creating
1714                  * opens, so use req->rq_xid as a secondary key.
1715                  * (See bugs 684, 685, and 428.)
1716                  * XXX no longer needed, but all opens need transnos!
1717                  */
1718                 if (iter->rq_transno > req->rq_transno)
1719                         continue;
1720
1721                 if (iter->rq_transno == req->rq_transno) {
1722                         LASSERT(iter->rq_xid != req->rq_xid);
1723                         if (iter->rq_xid > req->rq_xid)
1724                                 continue;
1725                 }
1726
1727                 list_add(&req->rq_replay_list, &iter->rq_replay_list);
1728                 return;
1729         }
1730
1731         list_add_tail(&req->rq_replay_list, &imp->imp_replay_list);
1732 }
1733
1734 int ptlrpc_queue_wait(struct ptlrpc_request *req)
1735 {
1736         int rc = 0;
1737         int brc;
1738         struct l_wait_info lwi;
1739         struct obd_import *imp = req->rq_import;
1740         cfs_duration_t timeout = 0;
1741         ENTRY;
1742
1743         LASSERT(req->rq_set == NULL);
1744         LASSERT(!req->rq_receiving_reply);
1745         atomic_inc(&imp->imp_inflight);
1746
1747         /* for distributed debugging */
1748         lustre_msg_set_status(req->rq_reqmsg, cfs_curproc_pid());
1749         LASSERT(imp->imp_obd != NULL);
1750         CDEBUG(D_RPCTRACE, "Sending RPC pname:cluuid:pid:xid:nid:opc "
1751                "%s:%s:%d:"LPU64":%s:%d\n", cfs_curproc_comm(),
1752                imp->imp_obd->obd_uuid.uuid,
1753                lustre_msg_get_status(req->rq_reqmsg), req->rq_xid,
1754                libcfs_nid2str(imp->imp_connection->c_peer.nid),
1755                lustre_msg_get_opc(req->rq_reqmsg));
1756
1757         /* Mark phase here for a little debug help */
1758         req->rq_phase = RQ_PHASE_RPC;
1759
1760         spin_lock(&imp->imp_lock);
1761 restart:
1762         req->rq_import_generation = imp->imp_generation;
1763         if (ptlrpc_import_delay_req(imp, req, &rc)) {
1764                 list_del(&req->rq_list);
1765
1766                 list_add_tail(&req->rq_list, &imp->imp_delayed_list);
1767                 spin_unlock(&imp->imp_lock);
1768
1769                 DEBUG_REQ(D_HA, req, "\"%s\" waiting for recovery: (%s != %s)",
1770                           cfs_curproc_comm(),
1771                           ptlrpc_import_state_name(req->rq_send_state),
1772                           ptlrpc_import_state_name(imp->imp_state));
1773                 lwi = LWI_INTR(interrupted_request, req);
1774                 rc = l_wait_event(req->rq_reply_waitq,
1775                                   (req->rq_send_state == imp->imp_state ||
1776                                    req->rq_err || req->rq_intr),
1777                                   &lwi);
1778                 DEBUG_REQ(D_HA, req, "\"%s\" awake: (%s == %s or %d/%d == 1)",
1779                           cfs_curproc_comm(),
1780                           ptlrpc_import_state_name(imp->imp_state),
1781                           ptlrpc_import_state_name(req->rq_send_state),
1782                           req->rq_err, req->rq_intr);
1783
1784                 spin_lock(&imp->imp_lock);
1785                 list_del_init(&req->rq_list);
1786
1787                 if (req->rq_err) {
1788                         rc = -EIO;
1789                 }
1790                 else if (req->rq_intr) {
1791                         rc = -EINTR;
1792                 }
1793                 else if (req->rq_no_resend) {
1794                         spin_unlock(&imp->imp_lock);
1795                         GOTO(out, rc = -ETIMEDOUT);
1796                 }
1797                 else {
1798                         GOTO(restart, rc);
1799                 }
1800         }
1801
1802         if (rc != 0) {
1803                 list_del_init(&req->rq_list);
1804                 spin_unlock(&imp->imp_lock);
1805                 req->rq_status = rc; // XXX this ok?
1806                 GOTO(out, rc);
1807         }
1808
1809         if (req->rq_resend) {
1810                 lustre_msg_add_flags(req->rq_reqmsg, MSG_RESENT);
1811
1812                 if (req->rq_bulk != NULL) {
1813                         ptlrpc_unregister_bulk (req);
1814
1815                         /* bulk requests are supposed to be
1816                          * idempotent, so we are free to bump the xid
1817                          * here, which we need to do before
1818                          * registering the bulk again (bug 6371).
1819                          * print the old xid first for sanity.
1820                          */
1821                         DEBUG_REQ(D_HA, req, "bumping xid for bulk: ");
1822                         req->rq_xid = ptlrpc_next_xid();
1823                 }
1824
1825                 DEBUG_REQ(D_HA, req, "resending: ");
1826         }
1827
1828         /* XXX this is the same as ptlrpc_set_wait */
1829         LASSERT(list_empty(&req->rq_list));
1830         list_add_tail(&req->rq_list, &imp->imp_sending_list);
1831         spin_unlock(&imp->imp_lock);
1832
1833         rc = sptlrpc_req_refresh_ctx(req, 0);
1834         if (rc) {
1835                 if (req->rq_err) {
1836                         /* we got fatal ctx refresh error, directly jump out
1837                          * thus we can pass back the actual error code.
1838                          */
1839                         spin_lock(&imp->imp_lock);
1840                         list_del_init(&req->rq_list);
1841                         spin_unlock(&imp->imp_lock);
1842
1843                         CERROR("Failed to refresh ctx of req %p: %d\n", req, rc);
1844                         GOTO(out, rc);
1845                 }
1846                 /* simulating we got error during send rpc */
1847                 goto after_send;
1848         }
1849
1850         rc = ptl_send_rpc(req, 0);
1851         if (rc) {
1852                 DEBUG_REQ(D_HA, req, "send failed (%d); recovering", rc);
1853                 timeout = CFS_TICK;
1854         } else {
1855                 timeout = cfs_timeout_cap(cfs_time_seconds(req->rq_timeout));
1856                 DEBUG_REQ(D_NET, req,
1857                           "-- sleeping for "CFS_DURATION_T" jiffies", timeout);
1858         }
1859 repeat:
1860         lwi = LWI_TIMEOUT_INTR(timeout, expired_request, interrupted_request,
1861                                req);
1862         rc = l_wait_event(req->rq_reply_waitq, ptlrpc_check_reply(req), &lwi);
1863         if (rc == -ETIMEDOUT && ptlrpc_check_and_wait_suspend(req))
1864                 goto repeat;
1865
1866 after_send:
1867         CDEBUG(D_RPCTRACE, "Completed RPC pname:cluuid:pid:xid:nid:opc "
1868                "%s:%s:%d:"LPU64":%s:%d\n", cfs_curproc_comm(),
1869                imp->imp_obd->obd_uuid.uuid,
1870                lustre_msg_get_status(req->rq_reqmsg), req->rq_xid,
1871                libcfs_nid2str(imp->imp_connection->c_peer.nid),
1872                lustre_msg_get_opc(req->rq_reqmsg));
1873
1874         spin_lock(&imp->imp_lock);
1875         list_del_init(&req->rq_list);
1876         spin_unlock(&imp->imp_lock);
1877
1878         /* If the reply was received normally, this just grabs the spinlock
1879          * (ensuring the reply callback has returned), sees that
1880          * req->rq_receiving_reply is clear and returns. */
1881         ptlrpc_unregister_reply (req);
1882
1883         if (req->rq_err)
1884                 GOTO(out, rc = -EIO);
1885
1886         /* Resend if we need to, unless we were interrupted. */
1887         if (req->rq_resend && !req->rq_intr) {
1888                 /* ...unless we were specifically told otherwise. */
1889                 if (req->rq_no_resend)
1890                         GOTO(out, rc = -ETIMEDOUT);
1891                 spin_lock(&imp->imp_lock);
1892                 goto restart;
1893         }
1894
1895         if (req->rq_intr) {
1896                 /* Should only be interrupted if we timed out. */
1897                 if (!req->rq_timedout)
1898                         DEBUG_REQ(D_ERROR, req,
1899                                   "rq_intr set but rq_timedout not");
1900                 GOTO(out, rc = -EINTR);
1901         }
1902
1903         if (req->rq_timedout) {                 /* non-recoverable timeout */
1904                 GOTO(out, rc = -ETIMEDOUT);
1905         }
1906
1907         if (!req->rq_replied) {
1908                 /* How can this be? -eeb */
1909                 DEBUG_REQ(D_ERROR, req, "!rq_replied: ");
1910                 LBUG();
1911                 GOTO(out, rc = req->rq_status);
1912         }
1913
1914         rc = after_reply(req);
1915         /* NB may return +ve success rc */
1916         if (req->rq_resend) {
1917                 spin_lock(&imp->imp_lock);
1918                 goto restart;
1919         }
1920
1921  out:
1922         if (req->rq_bulk != NULL) {
1923                 if (rc >= 0) {
1924                         /* success so far.  Note that anything going wrong
1925                          * with bulk now, is EXTREMELY strange, since the
1926                          * server must have believed that the bulk
1927                          * tranferred OK before she replied with success to
1928                          * me. */
1929                         lwi = LWI_TIMEOUT(timeout, NULL, NULL);
1930                         brc = l_wait_event(req->rq_reply_waitq,
1931                                            !ptlrpc_bulk_active(req->rq_bulk),
1932                                            &lwi);
1933                         LASSERT(brc == 0 || brc == -ETIMEDOUT);
1934                         if (brc != 0) {
1935                                 LASSERT(brc == -ETIMEDOUT);
1936                                 DEBUG_REQ(D_ERROR, req, "bulk timed out");
1937                                 rc = brc;
1938                         } else if (!req->rq_bulk->bd_success) {
1939                                 DEBUG_REQ(D_ERROR, req, "bulk transfer failed");
1940                                 rc = -EIO;
1941                         }
1942                 }
1943                 if (rc < 0)
1944                         ptlrpc_unregister_bulk (req);
1945         }
1946
1947         LASSERT(!req->rq_receiving_reply);
1948         req->rq_phase = RQ_PHASE_INTERPRET;
1949
1950         atomic_dec(&imp->imp_inflight);
1951         cfs_waitq_signal(&imp->imp_recovery_waitq);
1952         RETURN(rc);
1953 }
1954
1955 struct ptlrpc_replay_async_args {
1956         int praa_old_state;
1957         int praa_old_status;
1958 };
1959
1960 static int ptlrpc_replay_interpret(struct ptlrpc_request *req,
1961                                     void * data, int rc)
1962 {
1963         struct ptlrpc_replay_async_args *aa = data;
1964         struct obd_import *imp = req->rq_import;
1965
1966         ENTRY;
1967         atomic_dec(&imp->imp_replay_inflight);
1968
1969         if (!req->rq_replied) {
1970                 CERROR("request replay timed out, restarting recovery\n");
1971                 GOTO(out, rc = -ETIMEDOUT);
1972         }
1973
1974         if (lustre_msg_get_type(req->rq_repmsg) == PTL_RPC_MSG_ERR &&
1975             (lustre_msg_get_status(req->rq_repmsg) == -ENOTCONN ||
1976              lustre_msg_get_status(req->rq_repmsg) == -ENODEV))
1977                 GOTO(out, rc = lustre_msg_get_status(req->rq_repmsg));
1978
1979         /* The transno had better not change over replay. */
1980         LASSERT(lustre_msg_get_transno(req->rq_reqmsg) ==
1981                 lustre_msg_get_transno(req->rq_repmsg));
1982
1983         DEBUG_REQ(D_HA, req, "got rep");
1984
1985         /* let the callback do fixups, possibly including in the request */
1986         if (req->rq_replay_cb)
1987                 req->rq_replay_cb(req);
1988
1989         if (req->rq_replied &&
1990             lustre_msg_get_status(req->rq_repmsg) != aa->praa_old_status) {
1991                 DEBUG_REQ(D_ERROR, req, "status %d, old was %d",
1992                           lustre_msg_get_status(req->rq_repmsg),
1993                           aa->praa_old_status);
1994         } else {
1995                 /* Put it back for re-replay. */
1996                 lustre_msg_set_status(req->rq_repmsg, aa->praa_old_status);
1997         }
1998
1999         /*
2000          * Errors while replay can set transno to 0, but
2001          * imp_last_replay_transno shouldn't be set to 0 anyway
2002          */
2003         if (req->rq_transno > 0) {
2004                 spin_lock(&imp->imp_lock);
2005                 LASSERT(req->rq_transno <= imp->imp_last_replay_transno);
2006                 imp->imp_last_replay_transno = req->rq_transno;
2007                 spin_unlock(&imp->imp_lock);
2008         } else
2009                 CERROR("Transno is 0 during replay!\n");
2010         /* continue with recovery */
2011         rc = ptlrpc_import_recovery_state_machine(imp);
2012  out:
2013         req->rq_send_state = aa->praa_old_state;
2014
2015         if (rc != 0)
2016                 /* this replay failed, so restart recovery */
2017                 ptlrpc_connect_import(imp, NULL);
2018
2019         RETURN(rc);
2020 }
2021
2022 int ptlrpc_replay_req(struct ptlrpc_request *req)
2023 {
2024         struct ptlrpc_replay_async_args *aa;
2025         ENTRY;
2026
2027         LASSERT(req->rq_import->imp_state == LUSTRE_IMP_REPLAY);
2028
2029         /* Not handling automatic bulk replay yet (or ever?) */
2030         LASSERT(req->rq_bulk == NULL);
2031
2032         DEBUG_REQ(D_HA, req, "REPLAY");
2033
2034         LASSERT (sizeof (*aa) <= sizeof (req->rq_async_args));
2035         aa = (struct ptlrpc_replay_async_args *)&req->rq_async_args;
2036         memset(aa, 0, sizeof *aa);
2037
2038         /* Prepare request to be resent with ptlrpcd */
2039         aa->praa_old_state = req->rq_send_state;
2040         req->rq_send_state = LUSTRE_IMP_REPLAY;
2041         req->rq_phase = RQ_PHASE_NEW;
2042         aa->praa_old_status = lustre_msg_get_status(req->rq_repmsg);
2043         req->rq_status = 0;
2044
2045         req->rq_interpret_reply = ptlrpc_replay_interpret;
2046         atomic_inc(&req->rq_import->imp_replay_inflight);
2047         ptlrpc_request_addref(req); /* ptlrpcd needs a ref */
2048
2049         ptlrpcd_add_req(req);
2050         RETURN(0);
2051 }
2052
2053 void ptlrpc_abort_inflight(struct obd_import *imp)
2054 {
2055         struct list_head *tmp, *n;
2056         ENTRY;
2057
2058         /* Make sure that no new requests get processed for this import.
2059          * ptlrpc_{queue,set}_wait must (and does) hold imp_lock while testing
2060          * this flag and then putting requests on sending_list or delayed_list.
2061          */
2062         spin_lock(&imp->imp_lock);
2063
2064         /* XXX locking?  Maybe we should remove each request with the list
2065          * locked?  Also, how do we know if the requests on the list are
2066          * being freed at this time?
2067          */
2068         list_for_each_safe(tmp, n, &imp->imp_sending_list) {
2069                 struct ptlrpc_request *req =
2070                         list_entry(tmp, struct ptlrpc_request, rq_list);
2071
2072                 DEBUG_REQ(D_RPCTRACE, req, "inflight");
2073
2074                 spin_lock (&req->rq_lock);
2075                 if (req->rq_import_generation < imp->imp_generation) {
2076                         req->rq_err = 1;
2077                         ptlrpc_wake_client_req(req);
2078                 }
2079                 spin_unlock (&req->rq_lock);
2080         }
2081
2082         list_for_each_safe(tmp, n, &imp->imp_delayed_list) {
2083                 struct ptlrpc_request *req =
2084                         list_entry(tmp, struct ptlrpc_request, rq_list);
2085
2086                 DEBUG_REQ(D_RPCTRACE, req, "aborting waiting req");
2087
2088                 spin_lock (&req->rq_lock);
2089                 if (req->rq_import_generation < imp->imp_generation) {
2090                         req->rq_err = 1;
2091                         ptlrpc_wake_client_req(req);
2092                 }
2093                 spin_unlock (&req->rq_lock);
2094         }
2095
2096         /* Last chance to free reqs left on the replay list, but we
2097          * will still leak reqs that haven't committed.  */
2098         if (imp->imp_replayable)
2099                 ptlrpc_free_committed(imp);
2100
2101         spin_unlock(&imp->imp_lock);
2102
2103         EXIT;
2104 }
2105
2106 static __u64 ptlrpc_last_xid = 0;
2107 spinlock_t ptlrpc_last_xid_lock;
2108
2109 __u64 ptlrpc_next_xid(void)
2110 {
2111         __u64 tmp;
2112         spin_lock(&ptlrpc_last_xid_lock);
2113         tmp = ++ptlrpc_last_xid;
2114         spin_unlock(&ptlrpc_last_xid_lock);
2115         return tmp;
2116 }
2117
2118 __u64 ptlrpc_sample_next_xid(void)
2119 {
2120         __u64 tmp;
2121         spin_lock(&ptlrpc_last_xid_lock);
2122         tmp = ptlrpc_last_xid + 1;
2123         spin_unlock(&ptlrpc_last_xid_lock);
2124         return tmp;
2125 }
2126 EXPORT_SYMBOL(ptlrpc_sample_next_xid);