lustre/ptlrpc/client.c

   1 /* -*- mode: c; c-basic-offset: 8; indent-tabs-mode: nil; -*-
   2  * vim:expandtab:shiftwidth=8:tabstop=8:
   3  *
   4  * GPL HEADER START
   5  *
   6  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
   7  *
   8  * This program is free software; you can redistribute it and/or modify
   9  * it under the terms of the GNU General Public License version 2 only,
  10  * as published by the Free Software Foundation.
  11  *
  12  * This program is distributed in the hope that it will be useful, but
  13  * WITHOUT ANY WARRANTY; without even the implied warranty of
  14  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  15  * General Public License version 2 for more details (a copy is included
  16  * in the LICENSE file that accompanied this code).
  17  *
  18  * You should have received a copy of the GNU General Public License
  19  * version 2 along with this program; If not, see
  20  * http://www.sun.com/software/products/lustre/docs/GPLv2.pdf
  21  *
  22  * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
  23  * CA 95054 USA or visit www.sun.com if you need additional information or
  24  * have any questions.
  25  *
  26  * GPL HEADER END
  27  */
  28 /*
  29  * Copyright  2008 Sun Microsystems, Inc. All rights reserved
  30  * Use is subject to license terms.
  31  */
  32 /*
  33  * This file is part of Lustre, http://www.lustre.org/
  34  * Lustre is a trademark of Sun Microsystems, Inc.
  35  */
  36
  37 #define DEBUG_SUBSYSTEM S_RPC
  38 #ifndef __KERNEL__
  39 #include <errno.h>
  40 #include <signal.h>
  41 #include <liblustre.h>
  42 #endif
  43
  44 #include <obd_support.h>
  45 #include <obd_class.h>
  46 #include <lustre_lib.h>
  47 #include <lustre_ha.h>
  48 #include <lustre_import.h>
  49 #include <lustre_req_layout.h>
  50
  51 #include "ptlrpc_internal.h"
  52
  53 void ptlrpc_init_client(int req_portal, int rep_portal, char *name,
  54                         struct ptlrpc_client *cl)
  55 {
  56         cl->cli_request_portal = req_portal;
  57         cl->cli_reply_portal   = rep_portal;
  58         cl->cli_name           = name;
  59 }
  60
  61 struct ptlrpc_connection *ptlrpc_uuid_to_connection(struct obd_uuid *uuid)
  62 {
  63         struct ptlrpc_connection *c;
  64         lnet_nid_t                self;
  65         lnet_process_id_t         peer;
  66         int                       err;
  67
  68         err = ptlrpc_uuid_to_peer(uuid, &peer, &self);
  69         if (err != 0) {
  70                 CERROR("cannot find peer %s!\n", uuid->uuid);
  71                 return NULL;
  72         }
  73
  74         c = ptlrpc_connection_get(peer, self, uuid);
  75         if (c) {
  76                 memcpy(c->c_remote_uuid.uuid,
  77                        uuid->uuid, sizeof(c->c_remote_uuid.uuid));
  78         }
  79
  80         CDEBUG(D_INFO, "%s -> %p\n", uuid->uuid, c);
  81
  82         return c;
  83 }
  84
  85 static inline struct ptlrpc_bulk_desc *new_bulk(int npages, int type, int portal)
  86 {
  87         struct ptlrpc_bulk_desc *desc;
  88
  89         OBD_ALLOC(desc, offsetof (struct ptlrpc_bulk_desc, bd_iov[npages]));
  90         if (!desc)
  91                 return NULL;
  92
  93         spin_lock_init(&desc->bd_lock);
  94         cfs_waitq_init(&desc->bd_waitq);
  95         desc->bd_max_iov = npages;
  96         desc->bd_iov_count = 0;
  97         LNetInvalidateHandle(&desc->bd_md_h);
  98         desc->bd_portal = portal;
  99         desc->bd_type = type;
 100
 101         return desc;
 102 }
 103
 104 struct ptlrpc_bulk_desc *ptlrpc_prep_bulk_imp(struct ptlrpc_request *req,
 105                                               int npages, int type, int portal)
 106 {
 107         struct obd_import *imp = req->rq_import;
 108         struct ptlrpc_bulk_desc *desc;
 109
 110         ENTRY;
 111         LASSERT(type == BULK_PUT_SINK || type == BULK_GET_SOURCE);
 112         desc = new_bulk(npages, type, portal);
 113         if (desc == NULL)
 114                 RETURN(NULL);
 115
 116         desc->bd_import_generation = req->rq_import_generation;
 117         desc->bd_import = class_import_get(imp);
 118         desc->bd_req = req;
 119
 120         desc->bd_cbid.cbid_fn  = client_bulk_callback;
 121         desc->bd_cbid.cbid_arg = desc;
 122
 123         /* This makes req own desc, and free it when she frees herself */
 124         req->rq_bulk = desc;
 125
 126         return desc;
 127 }
 128
 129 struct ptlrpc_bulk_desc *ptlrpc_prep_bulk_exp(struct ptlrpc_request *req,
 130                                               int npages, int type, int portal)
 131 {
 132         struct obd_export *exp = req->rq_export;
 133         struct ptlrpc_bulk_desc *desc;
 134
 135         ENTRY;
 136         LASSERT(type == BULK_PUT_SOURCE || type == BULK_GET_SINK);
 137
 138         desc = new_bulk(npages, type, portal);
 139         if (desc == NULL)
 140                 RETURN(NULL);
 141
 142         desc->bd_export = class_export_get(exp);
 143         desc->bd_req = req;
 144
 145         desc->bd_cbid.cbid_fn  = server_bulk_callback;
 146         desc->bd_cbid.cbid_arg = desc;
 147
 148         /* NB we don't assign rq_bulk here; server-side requests are
 149          * re-used, and the handler frees the bulk desc explicitly. */
 150
 151         return desc;
 152 }
 153
 154 void ptlrpc_prep_bulk_page(struct ptlrpc_bulk_desc *desc,
 155                            cfs_page_t *page, int pageoffset, int len)
 156 {
 157         LASSERT(desc->bd_iov_count < desc->bd_max_iov);
 158         LASSERT(page != NULL);
 159         LASSERT(pageoffset >= 0);
 160         LASSERT(len > 0);
 161         LASSERT(pageoffset + len <= CFS_PAGE_SIZE);
 162
 163         desc->bd_nob += len;
 164
 165         ptlrpc_add_bulk_page(desc, page, pageoffset, len);
 166 }
 167
 168 void ptlrpc_free_bulk(struct ptlrpc_bulk_desc *desc)
 169 {
 170         ENTRY;
 171
 172         LASSERT(desc != NULL);
 173         LASSERT(desc->bd_iov_count != LI_POISON); /* not freed already */
 174         LASSERT(!desc->bd_network_rw);         /* network hands off or */
 175         LASSERT((desc->bd_export != NULL) ^ (desc->bd_import != NULL));
 176
 177         sptlrpc_enc_pool_put_pages(desc);
 178
 179         if (desc->bd_export)
 180                 class_export_put(desc->bd_export);
 181         else
 182                 class_import_put(desc->bd_import);
 183
 184         OBD_FREE(desc, offsetof(struct ptlrpc_bulk_desc,
 185                                 bd_iov[desc->bd_max_iov]));
 186         EXIT;
 187 }
 188
 189 /* Set server timelimit for this req */
 190 void ptlrpc_at_set_req_timeout(struct ptlrpc_request *req)
 191 {
 192         __u32 serv_est;
 193         int idx;
 194         struct imp_at *at;
 195
 196         LASSERT(req->rq_import);
 197
 198         if (AT_OFF) {
 199                 /* non-AT settings */
 200                 req->rq_timeout = req->rq_import->imp_server_timeout ?
 201                         obd_timeout / 2 : obd_timeout;
 202                 lustre_msg_set_timeout(req->rq_reqmsg, req->rq_timeout);
 203                 return;
 204         }
 205
 206         at = &req->rq_import->imp_at;
 207         idx = import_at_get_index(req->rq_import,
 208                                   req->rq_request_portal);
 209         serv_est = at_get(&at->iat_service_estimate[idx]);
 210         req->rq_timeout = at_est2timeout(serv_est);
 211         /* We could get even fancier here, using history to predict increased
 212            loading... */
 213
 214         /* Let the server know what this RPC timeout is by putting it in the
 215            reqmsg*/
 216         lustre_msg_set_timeout(req->rq_reqmsg, req->rq_timeout);
 217 }
 218
 219 /* Adjust max service estimate based on server value */
 220 static void ptlrpc_at_adj_service(struct ptlrpc_request *req,
 221                                   unsigned int serv_est)
 222 {
 223         int idx;
 224         unsigned int oldse;
 225         struct imp_at *at;
 226
 227         /* do estimate only if is not in recovery */
 228         if ((req->rq_send_state != LUSTRE_IMP_FULL) &&
 229              (req->rq_send_state != LUSTRE_IMP_CONNECTING))
 230                 return;
 231
 232         LASSERT(req->rq_import);
 233         at = &req->rq_import->imp_at;
 234
 235         idx = import_at_get_index(req->rq_import, req->rq_request_portal);
 236         /* max service estimates are tracked on the server side,
 237            so just keep minimal history here */
 238         oldse = at_add(&at->iat_service_estimate[idx], serv_est);
 239         if (oldse != 0)
 240                 CDEBUG(D_ADAPTTO, "The RPC service estimate for %s ptl %d "
 241                        "has changed from %d to %d\n",
 242                        req->rq_import->imp_obd->obd_name,req->rq_request_portal,
 243                        oldse, at_get(&at->iat_service_estimate[idx]));
 244 }
 245
 246 /* Expected network latency per remote node (secs) */
 247 int ptlrpc_at_get_net_latency(struct ptlrpc_request *req)
 248 {
 249         return AT_OFF ? 0 : at_get(&req->rq_import->imp_at.iat_net_latency);
 250 }
 251
 252 /* Adjust expected network latency */
 253 static void ptlrpc_at_adj_net_latency(struct ptlrpc_request *req,
 254                                       unsigned int service_time)
 255 {
 256         unsigned int nl, oldnl;
 257         struct imp_at *at;
 258         time_t now = cfs_time_current_sec();
 259
 260         LASSERT(req->rq_import);
 261         at = &req->rq_import->imp_at;
 262
 263         /* Network latency is total time less server processing time */
 264         nl = max_t(int, now - req->rq_sent - service_time, 0) +1/*st rounding*/;
 265         if (service_time > now - req->rq_sent + 3 /* bz16408 */)
 266                 CWARN("Reported service time %u > total measured time "
 267                       CFS_DURATION_T"\n", service_time,
 268                       cfs_time_sub(now, req->rq_sent));
 269
 270         oldnl = at_add(&at->iat_net_latency, nl);
 271         if (oldnl != 0)
 272                 CDEBUG(D_ADAPTTO, "The network latency for %s (nid %s) "
 273                        "has changed from %d to %d\n",
 274                        req->rq_import->imp_obd->obd_name,
 275                        obd_uuid2str(
 276                                &req->rq_import->imp_connection->c_remote_uuid),
 277                        oldnl, at_get(&at->iat_net_latency));
 278 }
 279
 280 static int unpack_reply(struct ptlrpc_request *req)
 281 {
 282         int rc;
 283
 284         /* Clear reply swab mask; we may have already swabbed an early reply */
 285         req->rq_rep_swab_mask = 0;
 286
 287         rc = lustre_unpack_msg(req->rq_repmsg, req->rq_replen);
 288         if (rc) {
 289                 DEBUG_REQ(D_ERROR, req, "unpack_rep failed: %d", rc);
 290                 return(-EPROTO);
 291         }
 292
 293         rc = lustre_unpack_rep_ptlrpc_body(req, MSG_PTLRPC_BODY_OFF);
 294         if (rc) {
 295                 DEBUG_REQ(D_ERROR, req, "unpack ptlrpc body failed: %d", rc);
 296                 return(-EPROTO);
 297         }
 298         return 0;
 299 }
 300
 301 /*
 302  * Handle an early reply message, called with the rq_lock held.
 303  * If anything goes wrong just ignore it - same as if it never happened
 304  */
 305 static int ptlrpc_at_recv_early_reply(struct ptlrpc_request *req)
 306 {
 307         struct ptlrpc_request *early_req;
 308         time_t                 olddl;
 309         int                    rc;
 310         ENTRY;
 311
 312         req->rq_early = 0;
 313         spin_unlock(&req->rq_lock);
 314
 315         rc = sptlrpc_cli_unwrap_early_reply(req, &early_req);
 316         if (rc) {
 317                 spin_lock(&req->rq_lock);
 318                 RETURN(rc);
 319         }
 320
 321         rc = unpack_reply(early_req);
 322         if (rc == 0) {
 323                 /* Expecting to increase the service time estimate here */
 324                 ptlrpc_at_adj_service(req,
 325                         lustre_msg_get_timeout(early_req->rq_repmsg));
 326                 ptlrpc_at_adj_net_latency(req,
 327                         lustre_msg_get_service_time(early_req->rq_repmsg));
 328         }
 329
 330         sptlrpc_cli_finish_early_reply(early_req);
 331
 332         spin_lock(&req->rq_lock);
 333
 334         if (rc == 0) {
 335                 /* Adjust the local timeout for this req */
 336                 ptlrpc_at_set_req_timeout(req);
 337
 338                 olddl = req->rq_deadline;
 339                 /* server assumes it now has rq_timeout from when it sent the
 340                    early reply, so client should give it at least that long. */
 341                 req->rq_deadline = cfs_time_current_sec() + req->rq_timeout +
 342                             ptlrpc_at_get_net_latency(req);
 343
 344                 DEBUG_REQ(D_ADAPTTO, req,
 345                           "Early reply #%d, new deadline in "CFS_DURATION_T"s "
 346                           "("CFS_DURATION_T"s)", req->rq_early_count,
 347                           cfs_time_sub(req->rq_deadline,
 348                                        cfs_time_current_sec()),
 349                           cfs_time_sub(req->rq_deadline, olddl));
 350         }
 351
 352         RETURN(rc);
 353 }
 354
 355 void ptlrpc_free_rq_pool(struct ptlrpc_request_pool *pool)
 356 {
 357         struct list_head *l, *tmp;
 358         struct ptlrpc_request *req;
 359
 360         LASSERT(pool != NULL);
 361
 362         spin_lock(&pool->prp_lock);
 363         list_for_each_safe(l, tmp, &pool->prp_req_list) {
 364                 req = list_entry(l, struct ptlrpc_request, rq_list);
 365                 list_del(&req->rq_list);
 366                 LASSERT(req->rq_reqbuf);
 367                 LASSERT(req->rq_reqbuf_len == pool->prp_rq_size);
 368                 OBD_FREE(req->rq_reqbuf, pool->prp_rq_size);
 369                 OBD_FREE(req, sizeof(*req));
 370         }
 371         spin_unlock(&pool->prp_lock);
 372         OBD_FREE(pool, sizeof(*pool));
 373 }
 374
 375 void ptlrpc_add_rqs_to_pool(struct ptlrpc_request_pool *pool, int num_rq)
 376 {
 377         int i;
 378         int size = 1;
 379
 380         while (size < pool->prp_rq_size + SPTLRPC_MAX_PAYLOAD)
 381                 size <<= 1;
 382
 383         LASSERTF(list_empty(&pool->prp_req_list) || size == pool->prp_rq_size,
 384                  "Trying to change pool size with nonempty pool "
 385                  "from %d to %d bytes\n", pool->prp_rq_size, size);
 386
 387         spin_lock(&pool->prp_lock);
 388         pool->prp_rq_size = size;
 389         for (i = 0; i < num_rq; i++) {
 390                 struct ptlrpc_request *req;
 391                 struct lustre_msg *msg;
 392
 393                 spin_unlock(&pool->prp_lock);
 394                 OBD_ALLOC(req, sizeof(struct ptlrpc_request));
 395                 if (!req)
 396                         return;
 397                 OBD_ALLOC_GFP(msg, size, CFS_ALLOC_STD);
 398                 if (!msg) {
 399                         OBD_FREE(req, sizeof(struct ptlrpc_request));
 400                         return;
 401                 }
 402                 req->rq_reqbuf = msg;
 403                 req->rq_reqbuf_len = size;
 404                 req->rq_pool = pool;
 405                 spin_lock(&pool->prp_lock);
 406                 list_add_tail(&req->rq_list, &pool->prp_req_list);
 407         }
 408         spin_unlock(&pool->prp_lock);
 409         return;
 410 }
 411
 412 struct ptlrpc_request_pool *
 413 ptlrpc_init_rq_pool(int num_rq, int msgsize,
 414                     void (*populate_pool)(struct ptlrpc_request_pool *, int))
 415 {
 416         struct ptlrpc_request_pool *pool;
 417
 418         OBD_ALLOC(pool, sizeof (struct ptlrpc_request_pool));
 419         if (!pool)
 420                 return NULL;
 421
 422         /* Request next power of two for the allocation, because internally
 423            kernel would do exactly this */
 424
 425         spin_lock_init(&pool->prp_lock);
 426         CFS_INIT_LIST_HEAD(&pool->prp_req_list);
 427         pool->prp_rq_size = msgsize;
 428         pool->prp_populate = populate_pool;
 429
 430         populate_pool(pool, num_rq);
 431
 432         if (list_empty(&pool->prp_req_list)) {
 433                 /* have not allocated a single request for the pool */
 434                 OBD_FREE(pool, sizeof (struct ptlrpc_request_pool));
 435                 pool = NULL;
 436         }
 437         return pool;
 438 }
 439
 440 static struct ptlrpc_request *
 441 ptlrpc_prep_req_from_pool(struct ptlrpc_request_pool *pool)
 442 {
 443         struct ptlrpc_request *request;
 444         struct lustre_msg *reqbuf;
 445
 446         if (!pool)
 447                 return NULL;
 448
 449         spin_lock(&pool->prp_lock);
 450
 451         /* See if we have anything in a pool, and bail out if nothing,
 452          * in writeout path, where this matters, this is safe to do, because
 453          * nothing is lost in this case, and when some in-flight requests
 454          * complete, this code will be called again. */
 455         if (unlikely(list_empty(&pool->prp_req_list))) {
 456                 spin_unlock(&pool->prp_lock);
 457                 return NULL;
 458         }
 459
 460         request = list_entry(pool->prp_req_list.next, struct ptlrpc_request,
 461                              rq_list);
 462         list_del_init(&request->rq_list);
 463         spin_unlock(&pool->prp_lock);
 464
 465         LASSERT(request->rq_reqbuf);
 466         LASSERT(request->rq_pool);
 467
 468         reqbuf = request->rq_reqbuf;
 469         memset(request, 0, sizeof(*request));
 470         request->rq_reqbuf = reqbuf;
 471         request->rq_reqbuf_len = pool->prp_rq_size;
 472         request->rq_pool = pool;
 473
 474         return request;
 475 }
 476
 477 static void __ptlrpc_free_req_to_pool(struct ptlrpc_request *request)
 478 {
 479         struct ptlrpc_request_pool *pool = request->rq_pool;
 480
 481         spin_lock(&pool->prp_lock);
 482         LASSERT(list_empty(&request->rq_list));
 483         LASSERT(!request->rq_receiving_reply);
 484         list_add_tail(&request->rq_list, &pool->prp_req_list);
 485         spin_unlock(&pool->prp_lock);
 486 }
 487
 488 static int __ptlrpc_request_bufs_pack(struct ptlrpc_request *request,
 489                                       __u32 version, int opcode,
 490                                       int count, __u32 *lengths, char **bufs,
 491                                       struct ptlrpc_cli_ctx *ctx)
 492 {
 493         struct obd_import  *imp = request->rq_import;
 494         int                 rc;
 495         ENTRY;
 496
 497         if (unlikely(ctx))
 498                 request->rq_cli_ctx = sptlrpc_cli_ctx_get(ctx);
 499         else {
 500                 rc = sptlrpc_req_get_ctx(request);
 501                 if (rc)
 502                         GOTO(out_free, rc);
 503         }
 504
 505         sptlrpc_req_set_flavor(request, opcode);
 506
 507         rc = lustre_pack_request(request, imp->imp_msg_magic, count,
 508                                  lengths, bufs);
 509         if (rc) {
 510                 LASSERT(!request->rq_pool);
 511                 GOTO(out_ctx, rc);
 512         }
 513
 514         lustre_msg_add_version(request->rq_reqmsg, version);
 515         request->rq_send_state = LUSTRE_IMP_FULL;
 516         request->rq_type = PTL_RPC_MSG_REQUEST;
 517         request->rq_export = NULL;
 518
 519         request->rq_req_cbid.cbid_fn  = request_out_callback;
 520         request->rq_req_cbid.cbid_arg = request;
 521
 522         request->rq_reply_cbid.cbid_fn  = reply_in_callback;
 523         request->rq_reply_cbid.cbid_arg = request;
 524
 525         request->rq_reply_deadline = 0;
 526         request->rq_phase = RQ_PHASE_NEW;
 527         request->rq_next_phase = RQ_PHASE_UNDEFINED;
 528
 529         request->rq_request_portal = imp->imp_client->cli_request_portal;
 530         request->rq_reply_portal = imp->imp_client->cli_reply_portal;
 531
 532         ptlrpc_at_set_req_timeout(request);
 533
 534         spin_lock_init(&request->rq_lock);
 535         CFS_INIT_LIST_HEAD(&request->rq_list);
 536         CFS_INIT_LIST_HEAD(&request->rq_timed_list);
 537         CFS_INIT_LIST_HEAD(&request->rq_replay_list);
 538         CFS_INIT_LIST_HEAD(&request->rq_ctx_chain);
 539         CFS_INIT_LIST_HEAD(&request->rq_set_chain);
 540         CFS_INIT_LIST_HEAD(&request->rq_history_list);
 541         CFS_INIT_LIST_HEAD(&request->rq_exp_list);
 542         cfs_waitq_init(&request->rq_reply_waitq);
 543         request->rq_xid = ptlrpc_next_xid();
 544         atomic_set(&request->rq_refcount, 1);
 545
 546         lustre_msg_set_opc(request->rq_reqmsg, opcode);
 547
 548         RETURN(0);
 549 out_ctx:
 550         sptlrpc_cli_ctx_put(request->rq_cli_ctx, 1);
 551 out_free:
 552         class_import_put(imp);
 553         return rc;
 554 }
 555
 556 int ptlrpc_request_bufs_pack(struct ptlrpc_request *request,
 557                              __u32 version, int opcode, char **bufs,
 558                              struct ptlrpc_cli_ctx *ctx)
 559 {
 560         int count;
 561
 562         count = req_capsule_filled_sizes(&request->rq_pill, RCL_CLIENT);
 563         return __ptlrpc_request_bufs_pack(request, version, opcode, count,
 564                                           request->rq_pill.rc_area[RCL_CLIENT],
 565                                           bufs, ctx);
 566 }
 567 EXPORT_SYMBOL(ptlrpc_request_bufs_pack);
 568
 569 int ptlrpc_request_pack(struct ptlrpc_request *request,
 570                         __u32 version, int opcode)
 571 {
 572         return ptlrpc_request_bufs_pack(request, version, opcode, NULL, NULL);
 573 }
 574
 575 static inline
 576 struct ptlrpc_request *__ptlrpc_request_alloc(struct obd_import *imp,
 577                                               struct ptlrpc_request_pool *pool)
 578 {
 579         struct ptlrpc_request *request = NULL;
 580
 581         if (pool)
 582                 request = ptlrpc_prep_req_from_pool(pool);
 583
 584         if (!request)
 585                 OBD_ALLOC_PTR(request);
 586
 587         if (request) {
 588                 LASSERTF((unsigned long)imp > 0x1000, "%p", imp);
 589                 LASSERT(imp != LP_POISON);
 590                 LASSERTF((unsigned long)imp->imp_client > 0x1000, "%p",
 591                         imp->imp_client);
 592                 LASSERT(imp->imp_client != LP_POISON);
 593
 594                 request->rq_import = class_import_get(imp);
 595         } else {
 596                 CERROR("request allocation out of memory\n");
 597         }
 598
 599         return request;
 600 }
 601
 602 static struct ptlrpc_request *
 603 ptlrpc_request_alloc_internal(struct obd_import *imp,
 604                               struct ptlrpc_request_pool * pool,
 605                               const struct req_format *format)
 606 {
 607         struct ptlrpc_request *request;
 608
 609         request = __ptlrpc_request_alloc(imp, pool);
 610         if (request == NULL)
 611                 return NULL;
 612
 613         req_capsule_init(&request->rq_pill, request, RCL_CLIENT);
 614         req_capsule_set(&request->rq_pill, format);
 615         return request;
 616 }
 617
 618 struct ptlrpc_request *ptlrpc_request_alloc(struct obd_import *imp,
 619                                             const struct req_format *format)
 620 {
 621         return ptlrpc_request_alloc_internal(imp, NULL, format);
 622 }
 623
 624 struct ptlrpc_request *ptlrpc_request_alloc_pool(struct obd_import *imp,
 625                                             struct ptlrpc_request_pool * pool,
 626                                             const struct req_format *format)
 627 {
 628         return ptlrpc_request_alloc_internal(imp, pool, format);
 629 }
 630
 631 void ptlrpc_request_free(struct ptlrpc_request *request)
 632 {
 633         if (request->rq_pool)
 634                 __ptlrpc_free_req_to_pool(request);
 635         else
 636                 OBD_FREE_PTR(request);
 637 }
 638
 639 struct ptlrpc_request *ptlrpc_request_alloc_pack(struct obd_import *imp,
 640                                                 const struct req_format *format,
 641                                                 __u32 version, int opcode)
 642 {
 643         struct ptlrpc_request *req = ptlrpc_request_alloc(imp, format);
 644         int                    rc;
 645
 646         if (req) {
 647                 rc = ptlrpc_request_pack(req, version, opcode);
 648                 if (rc) {
 649                         ptlrpc_request_free(req);
 650                         req = NULL;
 651                 }
 652         }
 653         return req;
 654 }
 655
 656 struct ptlrpc_request *
 657 ptlrpc_prep_req_pool(struct obd_import *imp,
 658                      __u32 version, int opcode,
 659                      int count, __u32 *lengths, char **bufs,
 660                      struct ptlrpc_request_pool *pool)
 661 {
 662         struct ptlrpc_request *request;
 663         int                    rc;
 664
 665         request = __ptlrpc_request_alloc(imp, pool);
 666         if (!request)
 667                 return NULL;
 668
 669         rc = __ptlrpc_request_bufs_pack(request, version, opcode, count,
 670                                         lengths, bufs, NULL);
 671         if (rc) {
 672                 ptlrpc_request_free(request);
 673                 request = NULL;
 674         }
 675         return request;
 676 }
 677
 678 struct ptlrpc_request *
 679 ptlrpc_prep_req(struct obd_import *imp, __u32 version, int opcode, int count,
 680                 __u32 *lengths, char **bufs)
 681 {
 682         return ptlrpc_prep_req_pool(imp, version, opcode, count, lengths, bufs,
 683                                     NULL);
 684 }
 685
 686 struct ptlrpc_request_set *ptlrpc_prep_set(void)
 687 {
 688         struct ptlrpc_request_set *set;
 689
 690         ENTRY;
 691         OBD_ALLOC(set, sizeof *set);
 692         if (!set)
 693                 RETURN(NULL);
 694         CFS_INIT_LIST_HEAD(&set->set_requests);
 695         cfs_waitq_init(&set->set_waitq);
 696         set->set_remaining = 0;
 697         spin_lock_init(&set->set_new_req_lock);
 698         CFS_INIT_LIST_HEAD(&set->set_new_requests);
 699         CFS_INIT_LIST_HEAD(&set->set_cblist);
 700
 701         RETURN(set);
 702 }
 703
 704 /* Finish with this set; opposite of prep_set. */
 705 void ptlrpc_set_destroy(struct ptlrpc_request_set *set)
 706 {
 707         struct list_head *tmp;
 708         struct list_head *next;
 709         int               expected_phase;
 710         int               n = 0;
 711         ENTRY;
 712
 713         /* Requests on the set should either all be completed, or all be new */
 714         expected_phase = (set->set_remaining == 0) ?
 715                          RQ_PHASE_COMPLETE : RQ_PHASE_NEW;
 716         list_for_each (tmp, &set->set_requests) {
 717                 struct ptlrpc_request *req =
 718                         list_entry(tmp, struct ptlrpc_request, rq_set_chain);
 719
 720                 LASSERT(req->rq_phase == expected_phase);
 721                 n++;
 722         }
 723
 724         LASSERT(set->set_remaining == 0 || set->set_remaining == n);
 725
 726         list_for_each_safe(tmp, next, &set->set_requests) {
 727                 struct ptlrpc_request *req =
 728                         list_entry(tmp, struct ptlrpc_request, rq_set_chain);
 729                 list_del_init(&req->rq_set_chain);
 730
 731                 LASSERT(req->rq_phase == expected_phase);
 732
 733                 if (req->rq_phase == RQ_PHASE_NEW) {
 734
 735                         if (req->rq_interpret_reply != NULL) {
 736                                 ptlrpc_interpterer_t interpreter =
 737                                         req->rq_interpret_reply;
 738
 739                                 /* higher level (i.e. LOV) failed;
 740                                  * let the sub reqs clean up */
 741                                 req->rq_status = -EBADR;
 742                                 interpreter(NULL, req, &req->rq_async_args,
 743                                             req->rq_status);
 744                         }
 745                         set->set_remaining--;
 746                 }
 747
 748                 req->rq_set = NULL;
 749                 ptlrpc_req_finished (req);
 750         }
 751
 752         LASSERT(set->set_remaining == 0);
 753
 754         OBD_FREE(set, sizeof(*set));
 755         EXIT;
 756 }
 757
 758 int ptlrpc_set_add_cb(struct ptlrpc_request_set *set,
 759                       set_interpreter_func fn, void *data)
 760 {
 761         struct ptlrpc_set_cbdata *cbdata;
 762
 763         OBD_ALLOC_PTR(cbdata);
 764         if (cbdata == NULL)
 765                 RETURN(-ENOMEM);
 766
 767         cbdata->psc_interpret = fn;
 768         cbdata->psc_data = data;
 769         list_add_tail(&cbdata->psc_item, &set->set_cblist);
 770
 771         RETURN(0);
 772 }
 773
 774 void ptlrpc_set_add_req(struct ptlrpc_request_set *set,
 775                         struct ptlrpc_request *req)
 776 {
 777         /* The set takes over the caller's request reference */
 778         list_add_tail(&req->rq_set_chain, &set->set_requests);
 779         req->rq_set = set;
 780         set->set_remaining++;
 781 }
 782
 783 /**
 784  * Lock so many callers can add things, the context that owns the set
 785  * is supposed to notice these and move them into the set proper.
 786  */
 787 int ptlrpc_set_add_new_req(struct ptlrpcd_ctl *pc,
 788                            struct ptlrpc_request *req)
 789 {
 790         struct ptlrpc_request_set *set = pc->pc_set;
 791
 792         /*
 793          * Let caller know that we stopped and will not handle this request.
 794          * It needs to take care itself of request.
 795          */
 796         if (test_bit(LIOD_STOP, &pc->pc_flags))
 797                 return -EALREADY;
 798
 799         spin_lock(&set->set_new_req_lock);
 800         /*
 801          * The set takes over the caller's request reference.
 802          */
 803         list_add_tail(&req->rq_set_chain, &set->set_new_requests);
 804         req->rq_set = set;
 805         spin_unlock(&set->set_new_req_lock);
 806
 807         /*
 808          * Let thead know that we added something and better it to wake up
 809          * and process.
 810          */
 811         cfs_waitq_signal(&set->set_waitq);
 812         return 0;
 813 }
 814
 815 /*
 816  * Based on the current state of the import, determine if the request
 817  * can be sent, is an error, or should be delayed.
 818  *
 819  * Returns true if this request should be delayed. If false, and
 820  * *status is set, then the request can not be sent and *status is the
 821  * error code.  If false and status is 0, then request can be sent.
 822  *
 823  * The imp->imp_lock must be held.
 824  */
 825 static int ptlrpc_import_delay_req(struct obd_import *imp,
 826                                    struct ptlrpc_request *req, int *status)
 827 {
 828         int delay = 0;
 829         ENTRY;
 830
 831         LASSERT (status != NULL);
 832         *status = 0;
 833
 834         if (req->rq_ctx_init || req->rq_ctx_fini) {
 835                 /* always allow ctx init/fini rpc go through */
 836         } else if (imp->imp_state == LUSTRE_IMP_NEW) {
 837                 DEBUG_REQ(D_ERROR, req, "Uninitialized import.");
 838                 *status = -EIO;
 839                 LBUG();
 840         } else if (imp->imp_state == LUSTRE_IMP_CLOSED) {
 841                 DEBUG_REQ(D_ERROR, req, "IMP_CLOSED ");
 842                 *status = -EIO;
 843         } else if (req->rq_send_state == LUSTRE_IMP_CONNECTING &&
 844                    imp->imp_state == LUSTRE_IMP_CONNECTING) {
 845                 /* allow CONNECT even if import is invalid */ ;
 846                 if (atomic_read(&imp->imp_inval_count) != 0) {
 847                         DEBUG_REQ(D_ERROR, req, "invalidate in flight");
 848                         *status = -EIO;
 849                 }
 850         } else if ((imp->imp_invalid && (!imp->imp_recon_bk)) ||
 851                                          imp->imp_obd->obd_no_recov) {
 852                 /* If the import has been invalidated (such as by an OST
 853                  * failure), and if the import(MGC) tried all of its connection
 854                  * list (Bug 13464), the request must fail with -ESHUTDOWN.
 855                  * This indicates the requests should be discarded; an -EIO
 856                  * may result in a resend of the request. */
 857                 if (!imp->imp_deactive)
 858                           DEBUG_REQ(D_ERROR, req, "IMP_INVALID");
 859                 *status = -ESHUTDOWN; /* bz 12940 */
 860         } else if (req->rq_import_generation != imp->imp_generation) {
 861                 DEBUG_REQ(D_ERROR, req, "req wrong generation:");
 862                 *status = -EIO;
 863         } else if (req->rq_send_state != imp->imp_state) {
 864                 /* invalidate in progress - any requests should be drop */
 865                 if (atomic_read(&imp->imp_inval_count) != 0) {
 866                         DEBUG_REQ(D_ERROR, req, "invalidate in flight");
 867                         *status = -EIO;
 868                 } else if (imp->imp_dlm_fake || req->rq_no_delay) {
 869                         *status = -EWOULDBLOCK;
 870                 } else {
 871                         delay = 1;
 872                 }
 873         }
 874
 875         RETURN(delay);
 876 }
 877
 878 static int ptlrpc_check_reply(struct ptlrpc_request *req)
 879 {
 880         int rc = 0;
 881         ENTRY;
 882
 883         /* serialise with network callback */
 884         spin_lock(&req->rq_lock);
 885
 886         if (ptlrpc_client_replied(req))
 887                 GOTO(out, rc = 1);
 888
 889         if (req->rq_net_err && !req->rq_timedout) {
 890                 spin_unlock(&req->rq_lock);
 891                 rc = ptlrpc_expire_one_request(req, 0);
 892                 spin_lock(&req->rq_lock);
 893                 GOTO(out, rc);
 894         }
 895
 896         if (req->rq_err)
 897                 GOTO(out, rc = 1);
 898
 899         if (req->rq_resend)
 900                 GOTO(out, rc = 1);
 901
 902         if (req->rq_restart)
 903                 GOTO(out, rc = 1);
 904
 905         if (ptlrpc_client_early(req)) {
 906                 ptlrpc_at_recv_early_reply(req);
 907                 GOTO(out, rc = 0); /* keep waiting */
 908         }
 909
 910         EXIT;
 911  out:
 912         spin_unlock(&req->rq_lock);
 913         DEBUG_REQ(D_NET, req, "rc = %d for", rc);
 914         return rc;
 915 }
 916
 917 static int ptlrpc_check_status(struct ptlrpc_request *req)
 918 {
 919         int err;
 920         ENTRY;
 921
 922         err = lustre_msg_get_status(req->rq_repmsg);
 923         if (lustre_msg_get_type(req->rq_repmsg) == PTL_RPC_MSG_ERR) {
 924                 struct obd_import *imp = req->rq_import;
 925                 __u32 opc = lustre_msg_get_opc(req->rq_reqmsg);
 926                 LCONSOLE_ERROR_MSG(0x011,"an error occurred while communicating"
 927                                 " with %s. The %s operation failed with %d\n",
 928                                 libcfs_nid2str(imp->imp_connection->c_peer.nid),
 929                                 ll_opcode2str(opc), err);
 930                 RETURN(err < 0 ? err : -EINVAL);
 931         }
 932
 933         if (err < 0) {
 934                 DEBUG_REQ(D_INFO, req, "status is %d", err);
 935         } else if (err > 0) {
 936                 /* XXX: translate this error from net to host */
 937                 DEBUG_REQ(D_INFO, req, "status is %d", err);
 938         }
 939
 940         RETURN(err);
 941 }
 942
 943 /**
 944  * save pre-versions for replay
 945  */
 946 static void ptlrpc_save_versions(struct ptlrpc_request *req)
 947 {
 948         struct lustre_msg *repmsg = req->rq_repmsg;
 949         struct lustre_msg *reqmsg = req->rq_reqmsg;
 950         __u64 *versions = lustre_msg_get_versions(repmsg);
 951         ENTRY;
 952
 953         if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY)
 954                 return;
 955
 956         LASSERT(versions);
 957         lustre_msg_set_versions(reqmsg, versions);
 958         CDEBUG(D_INFO, "Client save versions ["LPX64"/"LPX64"]\n",
 959                versions[0], versions[1]);
 960
 961         EXIT;
 962 }
 963
 964 /**
 965  * Callback function called when client receives RPC reply for \a req.
 966  */
 967 static int after_reply(struct ptlrpc_request *req)
 968 {
 969         struct obd_import *imp = req->rq_import;
 970         struct obd_device *obd = req->rq_import->imp_obd;
 971         int rc;
 972         struct timeval work_start;
 973         long timediff;
 974         ENTRY;
 975
 976         LASSERT(!req->rq_receiving_reply);
 977         LASSERT(obd);
 978         LASSERT(req->rq_nob_received <= req->rq_repbuf_len);
 979
 980         /*
 981          * NB Until this point, the whole of the incoming message,
 982          * including buflens, status etc is in the sender's byte order.
 983          */
 984
 985         rc = sptlrpc_cli_unwrap_reply(req);
 986         if (rc) {
 987                 DEBUG_REQ(D_ERROR, req, "unwrap reply failed (%d):", rc);
 988                 RETURN(rc);
 989         }
 990
 991         /*
 992          * Security layer unwrap might ask resend this request.
 993          */
 994         if (req->rq_resend)
 995                 RETURN(0);
 996
 997         rc = unpack_reply(req);
 998         if (rc)
 999                 RETURN(rc);
1000
1001         do_gettimeofday(&work_start);
1002         timediff = cfs_timeval_sub(&work_start, &req->rq_arrival_time, NULL);
1003         if (obd->obd_svc_stats != NULL) {
1004                 lprocfs_counter_add(obd->obd_svc_stats, PTLRPC_REQWAIT_CNTR,
1005                                     timediff);
1006                 ptlrpc_lprocfs_rpc_sent(req, timediff);
1007         }
1008
1009         if (lustre_msg_get_type(req->rq_repmsg) != PTL_RPC_MSG_REPLY &&
1010             lustre_msg_get_type(req->rq_repmsg) != PTL_RPC_MSG_ERR) {
1011                 DEBUG_REQ(D_ERROR, req, "invalid packet received (type=%u)",
1012                           lustre_msg_get_type(req->rq_repmsg));
1013                 RETURN(-EPROTO);
1014         }
1015
1016         OBD_FAIL_TIMEOUT(OBD_FAIL_PTLRPC_PAUSE_REP, obd_fail_val);
1017         ptlrpc_at_adj_service(req, lustre_msg_get_timeout(req->rq_repmsg));
1018         ptlrpc_at_adj_net_latency(req,
1019                                   lustre_msg_get_service_time(req->rq_repmsg));
1020
1021         rc = ptlrpc_check_status(req);
1022         imp->imp_connect_error = rc;
1023
1024         if (rc) {
1025                 /*
1026                  * Either we've been evicted, or the server has failed for
1027                  * some reason. Try to reconnect, and if that fails, punt to
1028                  * the upcall.
1029                  */
1030                 if (ll_rpc_recoverable_error(rc)) {
1031                         if (req->rq_send_state != LUSTRE_IMP_FULL ||
1032                             imp->imp_obd->obd_no_recov || imp->imp_dlm_fake) {
1033                                 RETURN(rc);
1034                         }
1035                         ptlrpc_request_handle_notconn(req);
1036                         RETURN(rc);
1037                 }
1038         } else {
1039                 /*
1040                  * Let's look if server sent slv. Do it only for RPC with
1041                  * rc == 0.
1042                  */
1043                 ldlm_cli_update_pool(req);
1044         }
1045
1046         /*
1047          * Store transno in reqmsg for replay.
1048          */
1049         if (!(lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY)) {
1050                 req->rq_transno = lustre_msg_get_transno(req->rq_repmsg);
1051                 lustre_msg_set_transno(req->rq_reqmsg, req->rq_transno);
1052         }
1053
1054         if (imp->imp_replayable) {
1055                 spin_lock(&imp->imp_lock);
1056                 /*
1057                  * No point in adding already-committed requests to the replay
1058                  * list, we will just remove them immediately. b=9829
1059                  */
1060                 if (req->rq_transno != 0 &&
1061                     (req->rq_transno >
1062                      lustre_msg_get_last_committed(req->rq_repmsg) ||
1063                      req->rq_replay)) {
1064                         /** version recovery */
1065                         ptlrpc_save_versions(req);
1066                         ptlrpc_retain_replayable_request(req, imp);
1067                 } else if (req->rq_commit_cb != NULL) {
1068                         spin_unlock(&imp->imp_lock);
1069                         req->rq_commit_cb(req);
1070                         spin_lock(&imp->imp_lock);
1071                 }
1072
1073                 /*
1074                  * Replay-enabled imports return commit-status information.
1075                  */
1076                 if (lustre_msg_get_last_committed(req->rq_repmsg)) {
1077                         imp->imp_peer_committed_transno =
1078                                 lustre_msg_get_last_committed(req->rq_repmsg);
1079                 }
1080                 ptlrpc_free_committed(imp);
1081                 spin_unlock(&imp->imp_lock);
1082         }
1083
1084         RETURN(rc);
1085 }
1086
1087 static int ptlrpc_send_new_req(struct ptlrpc_request *req)
1088 {
1089         struct obd_import     *imp;
1090         int rc;
1091         ENTRY;
1092
1093         LASSERT(req->rq_phase == RQ_PHASE_NEW);
1094         if (req->rq_sent && (req->rq_sent > cfs_time_current_sec()))
1095                 RETURN (0);
1096
1097         ptlrpc_rqphase_move(req, RQ_PHASE_RPC);
1098
1099         imp = req->rq_import;
1100         spin_lock(&imp->imp_lock);
1101
1102         req->rq_import_generation = imp->imp_generation;
1103
1104         if (ptlrpc_import_delay_req(imp, req, &rc)) {
1105                 spin_lock(&req->rq_lock);
1106                 req->rq_waiting = 1;
1107                 spin_unlock(&req->rq_lock);
1108
1109                 DEBUG_REQ(D_HA, req, "req from PID %d waiting for recovery: "
1110                           "(%s != %s)", lustre_msg_get_status(req->rq_reqmsg),
1111                           ptlrpc_import_state_name(req->rq_send_state),
1112                           ptlrpc_import_state_name(imp->imp_state));
1113                 LASSERT(list_empty(&req->rq_list));
1114                 list_add_tail(&req->rq_list, &imp->imp_delayed_list);
1115                 atomic_inc(&req->rq_import->imp_inflight);
1116                 spin_unlock(&imp->imp_lock);
1117                 RETURN(0);
1118         }
1119
1120         if (rc != 0) {
1121                 spin_unlock(&imp->imp_lock);
1122                 req->rq_status = rc;
1123                 ptlrpc_rqphase_move(req, RQ_PHASE_INTERPRET);
1124                 RETURN(rc);
1125         }
1126
1127         LASSERT(list_empty(&req->rq_list));
1128         list_add_tail(&req->rq_list, &imp->imp_sending_list);
1129         atomic_inc(&req->rq_import->imp_inflight);
1130         spin_unlock(&imp->imp_lock);
1131
1132         lustre_msg_set_status(req->rq_reqmsg, cfs_curproc_pid());
1133
1134         rc = sptlrpc_req_refresh_ctx(req, -1);
1135         if (rc) {
1136                 if (req->rq_err) {
1137                         req->rq_status = rc;
1138                         RETURN(1);
1139                 } else {
1140                         req->rq_wait_ctx = 1;
1141                         RETURN(0);
1142                 }
1143         }
1144
1145         CDEBUG(D_RPCTRACE, "Sending RPC pname:cluuid:pid:xid:nid:opc"
1146                " %s:%s:%d:"LPU64":%s:%d\n", cfs_curproc_comm(),
1147                imp->imp_obd->obd_uuid.uuid,
1148                lustre_msg_get_status(req->rq_reqmsg), req->rq_xid,
1149                libcfs_nid2str(imp->imp_connection->c_peer.nid),
1150                lustre_msg_get_opc(req->rq_reqmsg));
1151
1152         rc = ptl_send_rpc(req, 0);
1153         if (rc) {
1154                 DEBUG_REQ(D_HA, req, "send failed (%d); expect timeout", rc);
1155                 req->rq_net_err = 1;
1156                 RETURN(rc);
1157         }
1158         RETURN(0);
1159 }
1160
1161 /* this sends any unsent RPCs in @set and returns TRUE if all are sent */
1162 int ptlrpc_check_set(const struct lu_env *env, struct ptlrpc_request_set *set)
1163 {
1164         struct list_head *tmp;
1165         int force_timer_recalc = 0;
1166         ENTRY;
1167
1168         if (set->set_remaining == 0)
1169                 RETURN(1);
1170
1171         list_for_each(tmp, &set->set_requests) {
1172                 struct ptlrpc_request *req =
1173                         list_entry(tmp, struct ptlrpc_request, rq_set_chain);
1174                 struct obd_import *imp = req->rq_import;
1175                 int rc = 0;
1176
1177                 if (req->rq_phase == RQ_PHASE_NEW &&
1178                     ptlrpc_send_new_req(req)) {
1179                         force_timer_recalc = 1;
1180                 }
1181
1182                 /* delayed send - skip */
1183                 if (req->rq_phase == RQ_PHASE_NEW && req->rq_sent)
1184                         continue;
1185
1186                 if (!(req->rq_phase == RQ_PHASE_RPC ||
1187                       req->rq_phase == RQ_PHASE_BULK ||
1188                       req->rq_phase == RQ_PHASE_INTERPRET ||
1189                       req->rq_phase == RQ_PHASE_UNREGISTERING ||
1190                       req->rq_phase == RQ_PHASE_COMPLETE)) {
1191                         DEBUG_REQ(D_ERROR, req, "bad phase %x", req->rq_phase);
1192                         LBUG();
1193                 }
1194
1195                 if (req->rq_phase == RQ_PHASE_UNREGISTERING) {
1196                         LASSERT(req->rq_next_phase != req->rq_phase);
1197                         LASSERT(req->rq_next_phase != RQ_PHASE_UNDEFINED);
1198
1199                         /*
1200                          * Skip processing until reply is unlinked. We
1201                          * can't return to pool before that and we can't
1202                          * call interpret before that. We need to make
1203                          * sure that all rdma transfers finished and will
1204                          * not corrupt any data.
1205                          */
1206                         if (ptlrpc_client_recv_or_unlink(req) ||
1207                             ptlrpc_client_bulk_active(req))
1208                                 continue;
1209
1210                         /*
1211                          * Turn fail_loc off to prevent it from looping
1212                          * forever.
1213                          */
1214                         if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_REPL_UNLINK)) {
1215                                 OBD_FAIL_CHECK_ORSET(OBD_FAIL_PTLRPC_LONG_REPL_UNLINK,
1216                                                      OBD_FAIL_ONCE);
1217                         }
1218                         if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_BULK_UNLINK)) {
1219                                 OBD_FAIL_CHECK_ORSET(OBD_FAIL_PTLRPC_LONG_BULK_UNLINK,
1220                                                      OBD_FAIL_ONCE);
1221                         }
1222
1223                         /*
1224                          * Move to next phase if reply was successfully
1225                          * unlinked.
1226                          */
1227                         ptlrpc_rqphase_move(req, req->rq_next_phase);
1228                 }
1229
1230                 if (req->rq_phase == RQ_PHASE_COMPLETE)
1231                         continue;
1232
1233                 if (req->rq_phase == RQ_PHASE_INTERPRET)
1234                         GOTO(interpret, req->rq_status);
1235
1236                 /*
1237                  * Note that this also will start async reply unlink.
1238                  */
1239                 if (req->rq_net_err && !req->rq_timedout) {
1240                         ptlrpc_expire_one_request(req, 1);
1241
1242                         /*
1243                          * Check if we still need to wait for unlink.
1244                          */
1245                         if (ptlrpc_client_recv_or_unlink(req) ||
1246                             ptlrpc_client_bulk_active(req))
1247                                 continue;
1248                 }
1249
1250                 if (req->rq_err) {
1251                         req->rq_replied = 0;
1252                         if (req->rq_status == 0)
1253                                 req->rq_status = -EIO;
1254                         ptlrpc_rqphase_move(req, RQ_PHASE_INTERPRET);
1255                         GOTO(interpret, req->rq_status);
1256                 }
1257
1258                 /* ptlrpc_queue_wait->l_wait_event guarantees that rq_intr
1259                  * will only be set after rq_timedout, but the synchronous IO
1260                  * waiting path sets rq_intr irrespective of whether ptlrpcd
1261                  * has seen a timeout.  our policy is to only interpret
1262                  * interrupted rpcs after they have timed out */
1263                 if (req->rq_intr && (req->rq_timedout || req->rq_waiting ||
1264                                      req->rq_wait_ctx)) {
1265                         req->rq_status = -EINTR;
1266                         ptlrpc_rqphase_move(req, RQ_PHASE_INTERPRET);
1267                         GOTO(interpret, req->rq_status);
1268                 }
1269
1270                 if (req->rq_phase == RQ_PHASE_RPC) {
1271                         if (req->rq_timedout || req->rq_resend ||
1272                             req->rq_waiting || req->rq_wait_ctx) {
1273                                 int status;
1274
1275                                 if (!ptlrpc_unregister_reply(req, 1))
1276                                         continue;
1277
1278                                 spin_lock(&imp->imp_lock);
1279                                 if (ptlrpc_import_delay_req(imp, req, &status)){
1280                                         /* put on delay list - only if we wait
1281                                          * recovery finished - before send */
1282                                         list_del_init(&req->rq_list);
1283                                         list_add_tail(&req->rq_list, &imp->imp_delayed_list);
1284                                         spin_unlock(&imp->imp_lock);
1285                                         continue;
1286                                 }
1287
1288                                 if (status != 0)  {
1289                                         req->rq_status = status;
1290                                         ptlrpc_rqphase_move(req,
1291                                                 RQ_PHASE_INTERPRET);
1292                                         spin_unlock(&imp->imp_lock);
1293                                         GOTO(interpret, req->rq_status);
1294                                 }
1295                                 if (req->rq_no_resend && !req->rq_wait_ctx) {
1296                                         req->rq_status = -ENOTCONN;
1297                                         ptlrpc_rqphase_move(req,
1298                                                 RQ_PHASE_INTERPRET);
1299                                         spin_unlock(&imp->imp_lock);
1300                                         GOTO(interpret, req->rq_status);
1301                                 }
1302
1303                                 list_del_init(&req->rq_list);
1304                                 list_add_tail(&req->rq_list,
1305                                               &imp->imp_sending_list);
1306
1307                                 spin_unlock(&imp->imp_lock);
1308
1309                                 req->rq_waiting = 0;
1310
1311                                 if (req->rq_timedout||req->rq_resend) {
1312                                         /* This is re-sending anyways,
1313                                          * let's mark req as resend. */
1314                                         req->rq_resend = 1;
1315                                         if (req->rq_bulk) {
1316                                                 __u64 old_xid;
1317
1318                                                 if (!ptlrpc_unregister_bulk(req, 1))
1319                                                         continue;
1320
1321                                                 /* ensure previous bulk fails */
1322                                                 old_xid = req->rq_xid;
1323                                                 req->rq_xid = ptlrpc_next_xid();
1324                                                 CDEBUG(D_HA, "resend bulk "
1325                                                        "old x"LPU64
1326                                                        " new x"LPU64"\n",
1327                                                        old_xid, req->rq_xid);
1328                                         }
1329                                 }
1330                                 /*
1331                                  * rq_wait_ctx is only touched by ptlrpcd,
1332                                  * so no lock is needed here.
1333                                  */
1334                                 status = sptlrpc_req_refresh_ctx(req, -1);
1335                                 if (status) {
1336                                         if (req->rq_err) {
1337                                                 req->rq_status = status;
1338                                                 req->rq_wait_ctx = 0;
1339                                                 force_timer_recalc = 1;
1340                                         } else {
1341                                                 req->rq_wait_ctx = 1;
1342                                         }
1343
1344                                         continue;
1345                                 } else {
1346                                         req->rq_wait_ctx = 0;
1347                                 }
1348
1349                                 rc = ptl_send_rpc(req, 0);
1350                                 if (rc) {
1351                                         DEBUG_REQ(D_HA, req, "send failed (%d)",
1352                                                   rc);
1353                                         force_timer_recalc = 1;
1354                                         req->rq_net_err = 1;
1355                                 }
1356                                 /* need to reset the timeout */
1357                                 force_timer_recalc = 1;
1358                         }
1359
1360                         spin_lock(&req->rq_lock);
1361
1362                         if (ptlrpc_client_early(req)) {
1363                                 ptlrpc_at_recv_early_reply(req);
1364                                 spin_unlock(&req->rq_lock);
1365                                 continue;
1366                         }
1367
1368                         /* Still waiting for a reply? */
1369                         if (ptlrpc_client_recv(req)) {
1370                                 spin_unlock(&req->rq_lock);
1371                                 continue;
1372                         }
1373
1374                         /* Did we actually receive a reply? */
1375                         if (!ptlrpc_client_replied(req)) {
1376                                 spin_unlock(&req->rq_lock);
1377                                 continue;
1378                         }
1379
1380                         spin_unlock(&req->rq_lock);
1381
1382                         req->rq_status = after_reply(req);
1383                         if (req->rq_resend)
1384                                 continue;
1385
1386                         /* If there is no bulk associated with this request,
1387                          * then we're done and should let the interpreter
1388                          * process the reply. Similarly if the RPC returned
1389                          * an error, and therefore the bulk will never arrive.
1390                          */
1391                         if (req->rq_bulk == NULL || req->rq_status != 0) {
1392                                 ptlrpc_rqphase_move(req, RQ_PHASE_INTERPRET);
1393                                 GOTO(interpret, req->rq_status);
1394                         }
1395
1396                         ptlrpc_rqphase_move(req, RQ_PHASE_BULK);
1397                 }
1398
1399                 LASSERT(req->rq_phase == RQ_PHASE_BULK);
1400                 if (ptlrpc_client_bulk_active(req))
1401                         continue;
1402
1403                 if (!req->rq_bulk->bd_success) {
1404                         /* The RPC reply arrived OK, but the bulk screwed
1405                          * up!  Dead wierd since the server told us the RPC
1406                          * was good after getting the REPLY for her GET or
1407                          * the ACK for her PUT. */
1408                         DEBUG_REQ(D_ERROR, req, "bulk transfer failed");
1409                         LBUG();
1410                 }
1411
1412                 ptlrpc_rqphase_move(req, RQ_PHASE_INTERPRET);
1413
1414         interpret:
1415                 LASSERT(req->rq_phase == RQ_PHASE_INTERPRET);
1416
1417                 /* This moves to "unregistering" phase we need to wait for
1418                  * reply unlink. */
1419                 if (!ptlrpc_unregister_reply(req, 1))
1420                         continue;
1421
1422                 if (!ptlrpc_unregister_bulk(req, 1))
1423                         continue;
1424
1425                 /* When calling interpret receiving already should be
1426                  * finished. */
1427                 LASSERT(!req->rq_receiving_reply);
1428
1429                 if (req->rq_interpret_reply != NULL) {
1430                         ptlrpc_interpterer_t interpreter =
1431                                 req->rq_interpret_reply;
1432                         req->rq_status = interpreter(env, req,
1433                                                      &req->rq_async_args,
1434                                                      req->rq_status);
1435                 }
1436                 ptlrpc_rqphase_move(req, RQ_PHASE_COMPLETE);
1437
1438                 CDEBUG(D_RPCTRACE, "Completed RPC pname:cluuid:pid:xid:nid:"
1439                        "opc %s:%s:%d:"LPU64":%s:%d\n", cfs_curproc_comm(),
1440                        imp->imp_obd->obd_uuid.uuid,
1441                        lustre_msg_get_status(req->rq_reqmsg), req->rq_xid,
1442                        libcfs_nid2str(imp->imp_connection->c_peer.nid),
1443                        lustre_msg_get_opc(req->rq_reqmsg));
1444
1445                 spin_lock(&imp->imp_lock);
1446                 /* Request already may be not on sending or delaying list. This
1447                  * may happen in the case of marking it errorneous for the case
1448                  * ptlrpc_import_delay_req(req, status) find it impossible to
1449                  * allow sending this rpc and returns *status != 0. */
1450                 if (!list_empty(&req->rq_list)) {
1451                         list_del_init(&req->rq_list);
1452                         atomic_dec(&imp->imp_inflight);
1453                 }
1454                 spin_unlock(&imp->imp_lock);
1455
1456                 set->set_remaining--;
1457                 cfs_waitq_broadcast(&imp->imp_recovery_waitq);
1458         }
1459
1460         /* If we hit an error, we want to recover promptly. */
1461         RETURN(set->set_remaining == 0 || force_timer_recalc);
1462 }
1463
1464 /* Return 1 if we should give up, else 0 */
1465 int ptlrpc_expire_one_request(struct ptlrpc_request *req, int async_unlink)
1466 {
1467         struct obd_import *imp = req->rq_import;
1468         int rc = 0;
1469         ENTRY;
1470
1471         DEBUG_REQ(D_ERROR|D_NETERROR, req,
1472                   "%s (sent at "CFS_TIME_T", "CFS_DURATION_T"s ago)",
1473                   req->rq_net_err ? "network error" : "timeout",
1474                   req->rq_sent, cfs_time_sub(cfs_time_current_sec(),
1475                   req->rq_sent));
1476
1477         if (imp) {
1478                 LCONSOLE_WARN("Request x"LPU64" sent from %s to NID %s "
1479                               CFS_DURATION_T"s ago has timed out "
1480                               "(limit "CFS_DURATION_T"s).\n", req->rq_xid,
1481                               req->rq_import->imp_obd->obd_name,
1482                               libcfs_nid2str(imp->imp_connection->c_peer.nid),
1483                               cfs_time_sub(cfs_time_current_sec(), req->rq_sent),
1484                               cfs_time_sub(req->rq_deadline, req->rq_sent));
1485         }
1486
1487         if (imp != NULL && obd_debug_peer_on_timeout)
1488                 LNetCtl(IOC_LIBCFS_DEBUG_PEER, &imp->imp_connection->c_peer);
1489
1490         spin_lock(&req->rq_lock);
1491         req->rq_timedout = 1;
1492         spin_unlock(&req->rq_lock);
1493
1494         ptlrpc_unregister_reply(req, async_unlink);
1495         ptlrpc_unregister_bulk(req, async_unlink);
1496
1497         if (obd_dump_on_timeout)
1498                 libcfs_debug_dumplog();
1499
1500         if (imp == NULL) {
1501                 DEBUG_REQ(D_HA, req, "NULL import: already cleaned up?");
1502                 RETURN(1);
1503         }
1504
1505         atomic_inc(&imp->imp_timeouts);
1506
1507         /* The DLM server doesn't want recovery run on its imports. */
1508         if (imp->imp_dlm_fake)
1509                 RETURN(1);
1510
1511         /* If this request is for recovery or other primordial tasks,
1512          * then error it out here. */
1513         if (req->rq_ctx_init || req->rq_ctx_fini ||
1514             req->rq_send_state != LUSTRE_IMP_FULL ||
1515             imp->imp_obd->obd_no_recov) {
1516                 DEBUG_REQ(D_RPCTRACE, req, "err -110, sent_state=%s (now=%s)",
1517                           ptlrpc_import_state_name(req->rq_send_state),
1518                           ptlrpc_import_state_name(imp->imp_state));
1519                 spin_lock(&req->rq_lock);
1520                 req->rq_status = -ETIMEDOUT;
1521                 req->rq_err = 1;
1522                 spin_unlock(&req->rq_lock);
1523                 RETURN(1);
1524         }
1525
1526         /* if a request can't be resent we can't wait for an answer after
1527            the timeout */
1528         if (req->rq_no_resend) {
1529                 DEBUG_REQ(D_RPCTRACE, req, "TIMEOUT-NORESEND:");
1530                 rc = 1;
1531         }
1532
1533         ptlrpc_fail_import(imp, lustre_msg_get_conn_cnt(req->rq_reqmsg));
1534
1535         RETURN(rc);
1536 }
1537
1538 int ptlrpc_expired_set(void *data)
1539 {
1540         struct ptlrpc_request_set *set = data;
1541         struct list_head          *tmp;
1542         time_t                     now = cfs_time_current_sec();
1543         ENTRY;
1544
1545         LASSERT(set != NULL);
1546
1547         /*
1548          * A timeout expired. See which reqs it applies to...
1549          */
1550         list_for_each (tmp, &set->set_requests) {
1551                 struct ptlrpc_request *req =
1552                         list_entry(tmp, struct ptlrpc_request, rq_set_chain);
1553
1554                 /* don't expire request waiting for context */
1555                 if (req->rq_wait_ctx)
1556                         continue;
1557
1558                 /* Request in-flight? */
1559                 if (!((req->rq_phase == RQ_PHASE_RPC &&
1560                        !req->rq_waiting && !req->rq_resend) ||
1561                       (req->rq_phase == RQ_PHASE_BULK)))
1562                         continue;
1563
1564                 if (req->rq_timedout ||     /* already dealt with */
1565                     req->rq_deadline > now) /* not expired */
1566                         continue;
1567
1568                 /* Deal with this guy. Do it asynchronously to not block
1569                  * ptlrpcd thread. */
1570                 ptlrpc_expire_one_request(req, 1);
1571         }
1572
1573         /*
1574          * When waiting for a whole set, we always to break out of the
1575          * sleep so we can recalculate the timeout, or enable interrupts
1576          * if everyone's timed out.
1577          */
1578         RETURN(1);
1579 }
1580
1581 void ptlrpc_mark_interrupted(struct ptlrpc_request *req)
1582 {
1583         spin_lock(&req->rq_lock);
1584         req->rq_intr = 1;
1585         spin_unlock(&req->rq_lock);
1586 }
1587
1588 void ptlrpc_interrupted_set(void *data)
1589 {
1590         struct ptlrpc_request_set *set = data;
1591         struct list_head *tmp;
1592
1593         LASSERT(set != NULL);
1594         CERROR("INTERRUPTED SET %p\n", set);
1595
1596         list_for_each(tmp, &set->set_requests) {
1597                 struct ptlrpc_request *req =
1598                         list_entry(tmp, struct ptlrpc_request, rq_set_chain);
1599
1600                 if (req->rq_phase != RQ_PHASE_RPC &&
1601                     req->rq_phase != RQ_PHASE_UNREGISTERING)
1602                         continue;
1603
1604                 ptlrpc_mark_interrupted(req);
1605         }
1606 }
1607
1608 /**
1609  * Get the smallest timeout in the set; this does NOT set a timeout.
1610  */
1611 int ptlrpc_set_next_timeout(struct ptlrpc_request_set *set)
1612 {
1613         struct list_head      *tmp;
1614         time_t                 now = cfs_time_current_sec();
1615         int                    timeout = 0;
1616         struct ptlrpc_request *req;
1617         int                    deadline;
1618         ENTRY;
1619
1620         SIGNAL_MASK_ASSERT(); /* XXX BUG 1511 */
1621
1622         list_for_each(tmp, &set->set_requests) {
1623                 req = list_entry(tmp, struct ptlrpc_request, rq_set_chain);
1624
1625                 /*
1626                  * Request in-flight?
1627                  */
1628                 if (!(((req->rq_phase == RQ_PHASE_RPC) && !req->rq_waiting) ||
1629                       (req->rq_phase == RQ_PHASE_BULK) ||
1630                       (req->rq_phase == RQ_PHASE_NEW)))
1631                         continue;
1632
1633                 /*
1634                  * Already timed out.
1635                  */
1636                 if (req->rq_timedout)
1637                         continue;
1638
1639                 /*
1640                  * Waiting for ctx.
1641                  */
1642                 if (req->rq_wait_ctx)
1643                         continue;
1644
1645                 if (req->rq_phase == RQ_PHASE_NEW)
1646                         deadline = req->rq_sent;
1647                 else
1648                         deadline = req->rq_sent + req->rq_timeout;
1649
1650                 if (deadline <= now)    /* actually expired already */
1651                         timeout = 1;    /* ASAP */
1652                 else if (timeout == 0 || timeout > deadline - now)
1653                         timeout = deadline - now;
1654         }
1655         RETURN(timeout);
1656 }
1657
1658 int ptlrpc_set_wait(struct ptlrpc_request_set *set)
1659 {
1660         struct list_head      *tmp;
1661         struct ptlrpc_request *req;
1662         struct l_wait_info     lwi;
1663         int                    rc, timeout;
1664         ENTRY;
1665
1666         if (list_empty(&set->set_requests))
1667                 RETURN(0);
1668
1669         list_for_each(tmp, &set->set_requests) {
1670                 req = list_entry(tmp, struct ptlrpc_request, rq_set_chain);
1671                 if (req->rq_phase == RQ_PHASE_NEW)
1672                         (void)ptlrpc_send_new_req(req);
1673         }
1674
1675         do {
1676                 timeout = ptlrpc_set_next_timeout(set);
1677
1678                 /* wait until all complete, interrupted, or an in-flight
1679                  * req times out */
1680                 CDEBUG(D_RPCTRACE, "set %p going to sleep for %d seconds\n",
1681                        set, timeout);
1682                 lwi = LWI_TIMEOUT_INTR(cfs_time_seconds(timeout ? timeout : 1),
1683                                        ptlrpc_expired_set,
1684                                        ptlrpc_interrupted_set, set);
1685                 rc = l_wait_event(set->set_waitq,
1686                                   ptlrpc_check_set(NULL, set), &lwi);
1687
1688                 LASSERT(rc == 0 || rc == -EINTR || rc == -ETIMEDOUT);
1689
1690                 /* -EINTR => all requests have been flagged rq_intr so next
1691                  * check completes.
1692                  * -ETIMEOUTD => someone timed out.  When all reqs have
1693                  * timed out, signals are enabled allowing completion with
1694                  * EINTR.
1695                  * I don't really care if we go once more round the loop in
1696                  * the error cases -eeb. */
1697         } while (rc != 0 || set->set_remaining != 0);
1698
1699         LASSERT(set->set_remaining == 0);
1700
1701         rc = 0;
1702         list_for_each(tmp, &set->set_requests) {
1703                 req = list_entry(tmp, struct ptlrpc_request, rq_set_chain);
1704
1705                 LASSERT(req->rq_phase == RQ_PHASE_COMPLETE);
1706                 if (req->rq_status != 0)
1707                         rc = req->rq_status;
1708         }
1709
1710         if (set->set_interpret != NULL) {
1711                 int (*interpreter)(struct ptlrpc_request_set *set,void *,int) =
1712                         set->set_interpret;
1713                 rc = interpreter (set, set->set_arg, rc);
1714         } else {
1715                 struct ptlrpc_set_cbdata *cbdata, *n;
1716                 int err;
1717
1718                 list_for_each_entry_safe(cbdata, n,
1719                                          &set->set_cblist, psc_item) {
1720                         list_del_init(&cbdata->psc_item);
1721                         err = cbdata->psc_interpret(set, cbdata->psc_data, rc);
1722                         if (err && !rc)
1723                                 rc = err;
1724                         OBD_FREE_PTR(cbdata);
1725                 }
1726         }
1727
1728         RETURN(rc);
1729 }
1730
1731 static void __ptlrpc_free_req(struct ptlrpc_request *request, int locked)
1732 {
1733         ENTRY;
1734         if (request == NULL) {
1735                 EXIT;
1736                 return;
1737         }
1738
1739         LASSERTF(!request->rq_receiving_reply, "req %p\n", request);
1740         LASSERTF(request->rq_rqbd == NULL, "req %p\n",request);/* client-side */
1741         LASSERTF(list_empty(&request->rq_list), "req %p\n", request);
1742         LASSERTF(list_empty(&request->rq_set_chain), "req %p\n", request);
1743         LASSERTF(list_empty(&request->rq_exp_list), "req %p\n", request);
1744         LASSERTF(!request->rq_replay, "req %p\n", request);
1745         LASSERT(request->rq_cli_ctx);
1746
1747         req_capsule_fini(&request->rq_pill);
1748
1749         /* We must take it off the imp_replay_list first.  Otherwise, we'll set
1750          * request->rq_reqmsg to NULL while osc_close is dereferencing it. */
1751         if (request->rq_import != NULL) {
1752                 if (!locked)
1753                         spin_lock(&request->rq_import->imp_lock);
1754                 list_del_init(&request->rq_replay_list);
1755                 if (!locked)
1756                         spin_unlock(&request->rq_import->imp_lock);
1757         }
1758         LASSERTF(list_empty(&request->rq_replay_list), "req %p\n", request);
1759
1760         if (atomic_read(&request->rq_refcount) != 0) {
1761                 DEBUG_REQ(D_ERROR, request,
1762                           "freeing request with nonzero refcount");
1763                 LBUG();
1764         }
1765
1766         if (request->rq_repbuf != NULL)
1767                 sptlrpc_cli_free_repbuf(request);
1768         if (request->rq_export != NULL) {
1769                 class_export_put(request->rq_export);
1770                 request->rq_export = NULL;
1771         }
1772         if (request->rq_import != NULL) {
1773                 class_import_put(request->rq_import);
1774                 request->rq_import = NULL;
1775         }
1776         if (request->rq_bulk != NULL)
1777                 ptlrpc_free_bulk(request->rq_bulk);
1778
1779         if (request->rq_reqbuf != NULL || request->rq_clrbuf != NULL)
1780                 sptlrpc_cli_free_reqbuf(request);
1781
1782         sptlrpc_req_put_ctx(request, !locked);
1783
1784         if (request->rq_pool)
1785                 __ptlrpc_free_req_to_pool(request);
1786         else
1787                 OBD_FREE(request, sizeof(*request));
1788         EXIT;
1789 }
1790
1791 static int __ptlrpc_req_finished(struct ptlrpc_request *request, int locked);
1792 void ptlrpc_req_finished_with_imp_lock(struct ptlrpc_request *request)
1793 {
1794         LASSERT_SPIN_LOCKED(&request->rq_import->imp_lock);
1795         (void)__ptlrpc_req_finished(request, 1);
1796 }
1797
1798 static int __ptlrpc_req_finished(struct ptlrpc_request *request, int locked)
1799 {
1800         ENTRY;
1801         if (request == NULL)
1802                 RETURN(1);
1803
1804         if (request == LP_POISON ||
1805             request->rq_reqmsg == LP_POISON) {
1806                 CERROR("dereferencing freed request (bug 575)\n");
1807                 LBUG();
1808                 RETURN(1);
1809         }
1810
1811         DEBUG_REQ(D_INFO, request, "refcount now %u",
1812                   atomic_read(&request->rq_refcount) - 1);
1813
1814         if (atomic_dec_and_test(&request->rq_refcount)) {
1815                 __ptlrpc_free_req(request, locked);
1816                 RETURN(1);
1817         }
1818
1819         RETURN(0);
1820 }
1821
1822 void ptlrpc_req_finished(struct ptlrpc_request *request)
1823 {
1824         __ptlrpc_req_finished(request, 0);
1825 }
1826
1827 __u64 ptlrpc_req_xid(struct ptlrpc_request *request)
1828 {
1829         return request->rq_xid;
1830 }
1831 EXPORT_SYMBOL(ptlrpc_req_xid);
1832
1833 /* Disengage the client's reply buffer from the network
1834  * NB does _NOT_ unregister any client-side bulk.
1835  * IDEMPOTENT, but _not_ safe against concurrent callers.
1836  * The request owner (i.e. the thread doing the I/O) must call...
1837  */
1838 int ptlrpc_unregister_reply(struct ptlrpc_request *request, int async)
1839 {
1840         int                rc;
1841         cfs_waitq_t       *wq;
1842         struct l_wait_info lwi;
1843
1844         /*
1845          * Might sleep.
1846          */
1847         LASSERT(!in_interrupt());
1848
1849         /*
1850          * Let's setup deadline for reply unlink.
1851          */
1852         if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_REPL_UNLINK) &&
1853             async && request->rq_reply_deadline == 0)
1854                 request->rq_reply_deadline = cfs_time_current_sec()+LONG_UNLINK;
1855
1856         /*
1857          * Nothing left to do.
1858          */
1859         if (!ptlrpc_client_recv_or_unlink(request))
1860                 RETURN(1);
1861
1862         LNetMDUnlink(request->rq_reply_md_h);
1863
1864         /*
1865          * Let's check it once again.
1866          */
1867         if (!ptlrpc_client_recv_or_unlink(request))
1868                 RETURN(1);
1869
1870         /*
1871          * Move to "Unregistering" phase as reply was not unlinked yet.
1872          */
1873         ptlrpc_rqphase_move(request, RQ_PHASE_UNREGISTERING);
1874
1875         /*
1876          * Do not wait for unlink to finish.
1877          */
1878         if (async)
1879                 RETURN(0);
1880
1881         /*
1882          * We have to l_wait_event() whatever the result, to give liblustre
1883          * a chance to run reply_in_callback(), and to make sure we've
1884          * unlinked before returning a req to the pool.
1885          */
1886         if (request->rq_set != NULL)
1887                 wq = &request->rq_set->set_waitq;
1888         else
1889                 wq = &request->rq_reply_waitq;
1890
1891         for (;;) {
1892                 /* Network access will complete in finite time but the HUGE
1893                  * timeout lets us CWARN for visibility of sluggish NALs */
1894                 lwi = LWI_TIMEOUT_INTERVAL(cfs_time_seconds(LONG_UNLINK),
1895                                            cfs_time_seconds(1), NULL, NULL);
1896                 rc = l_wait_event(*wq, !ptlrpc_client_recv_or_unlink(request),
1897                                   &lwi);
1898                 if (rc == 0) {
1899                         ptlrpc_rqphase_move(request, request->rq_next_phase);
1900                         RETURN(1);
1901                 }
1902
1903                 LASSERT(rc == -ETIMEDOUT);
1904                 DEBUG_REQ(D_WARNING, request, "Unexpectedly long timeout "
1905                           "rvcng=%d unlnk=%d", request->rq_receiving_reply,
1906                           request->rq_must_unlink);
1907         }
1908         RETURN(0);
1909 }
1910
1911 /* caller must hold imp->imp_lock */
1912 void ptlrpc_free_committed(struct obd_import *imp)
1913 {
1914         struct list_head *tmp, *saved;
1915         struct ptlrpc_request *req;
1916         struct ptlrpc_request *last_req = NULL; /* temporary fire escape */
1917         ENTRY;
1918
1919         LASSERT(imp != NULL);
1920
1921         LASSERT_SPIN_LOCKED(&imp->imp_lock);
1922
1923
1924         if (imp->imp_peer_committed_transno == imp->imp_last_transno_checked &&
1925             imp->imp_generation == imp->imp_last_generation_checked) {
1926                 CDEBUG(D_RPCTRACE, "%s: skip recheck: last_committed "LPU64"\n",
1927                        imp->imp_obd->obd_name, imp->imp_peer_committed_transno);
1928                 EXIT;
1929                 return;
1930         }
1931         CDEBUG(D_RPCTRACE, "%s: committing for last_committed "LPU64" gen %d\n",
1932                imp->imp_obd->obd_name, imp->imp_peer_committed_transno,
1933                imp->imp_generation);
1934         imp->imp_last_transno_checked = imp->imp_peer_committed_transno;
1935         imp->imp_last_generation_checked = imp->imp_generation;
1936
1937         list_for_each_safe(tmp, saved, &imp->imp_replay_list) {
1938                 req = list_entry(tmp, struct ptlrpc_request, rq_replay_list);
1939
1940                 /* XXX ok to remove when 1357 resolved - rread 05/29/03  */
1941                 LASSERT(req != last_req);
1942                 last_req = req;
1943
1944                 if (req->rq_transno == 0) {
1945                         DEBUG_REQ(D_EMERG, req, "zero transno during replay");
1946                         LBUG();
1947                 }
1948                 if (req->rq_import_generation < imp->imp_generation) {
1949                         DEBUG_REQ(D_RPCTRACE, req, "free request with old gen");
1950                         GOTO(free_req, 0);
1951                 }
1952
1953                 if (req->rq_replay) {
1954                         DEBUG_REQ(D_RPCTRACE, req, "keeping (FL_REPLAY)");
1955                         continue;
1956                 }
1957
1958                 /* not yet committed */
1959                 if (req->rq_transno > imp->imp_peer_committed_transno) {
1960                         DEBUG_REQ(D_RPCTRACE, req, "stopping search");
1961                         break;
1962                 }
1963
1964                 DEBUG_REQ(D_RPCTRACE, req, "commit (last_committed "LPU64")",
1965                           imp->imp_peer_committed_transno);
1966 free_req:
1967                 spin_lock(&req->rq_lock);
1968                 req->rq_replay = 0;
1969                 spin_unlock(&req->rq_lock);
1970                 if (req->rq_commit_cb != NULL)
1971                         req->rq_commit_cb(req);
1972                 list_del_init(&req->rq_replay_list);
1973                 __ptlrpc_req_finished(req, 1);
1974         }
1975
1976         EXIT;
1977         return;
1978 }
1979
1980 void ptlrpc_cleanup_client(struct obd_import *imp)
1981 {
1982         ENTRY;
1983         EXIT;
1984         return;
1985 }
1986
1987 void ptlrpc_resend_req(struct ptlrpc_request *req)
1988 {
1989         DEBUG_REQ(D_HA, req, "going to resend");
1990         lustre_msg_set_handle(req->rq_reqmsg, &(struct lustre_handle){ 0 });
1991         req->rq_status = -EAGAIN;
1992
1993         spin_lock(&req->rq_lock);
1994         req->rq_resend = 1;
1995         req->rq_net_err = 0;
1996         req->rq_timedout = 0;
1997         if (req->rq_bulk) {
1998                 __u64 old_xid = req->rq_xid;
1999
2000                 /* ensure previous bulk fails */
2001                 req->rq_xid = ptlrpc_next_xid();
2002                 CDEBUG(D_HA, "resend bulk old x"LPU64" new x"LPU64"\n",
2003                        old_xid, req->rq_xid);
2004         }
2005         ptlrpc_client_wake_req(req);
2006         spin_unlock(&req->rq_lock);
2007 }
2008
2009 /* XXX: this function and rq_status are currently unused */
2010 void ptlrpc_restart_req(struct ptlrpc_request *req)
2011 {
2012         DEBUG_REQ(D_HA, req, "restarting (possibly-)completed request");
2013         req->rq_status = -ERESTARTSYS;
2014
2015         spin_lock(&req->rq_lock);
2016         req->rq_restart = 1;
2017         req->rq_timedout = 0;
2018         ptlrpc_client_wake_req(req);
2019         spin_unlock(&req->rq_lock);
2020 }
2021
2022 static int expired_request(void *data)
2023 {
2024         struct ptlrpc_request *req = data;
2025         ENTRY;
2026
2027         /*
2028          * Some failure can suspend regular timeouts.
2029          */
2030         if (ptlrpc_check_suspend())
2031                 RETURN(1);
2032
2033         /*
2034          * Deadline may have changed with an early reply.
2035          */
2036         if (req->rq_deadline > cfs_time_current_sec())
2037                 RETURN(1);
2038
2039         RETURN(ptlrpc_expire_one_request(req, 0));
2040 }
2041
2042 static void interrupted_request(void *data)
2043 {
2044         struct ptlrpc_request *req = data;
2045         DEBUG_REQ(D_HA, req, "request interrupted");
2046         spin_lock(&req->rq_lock);
2047         req->rq_intr = 1;
2048         spin_unlock(&req->rq_lock);
2049 }
2050
2051 struct ptlrpc_request *ptlrpc_request_addref(struct ptlrpc_request *req)
2052 {
2053         ENTRY;
2054         atomic_inc(&req->rq_refcount);
2055         RETURN(req);
2056 }
2057
2058 void ptlrpc_retain_replayable_request(struct ptlrpc_request *req,
2059                                       struct obd_import *imp)
2060 {
2061         struct list_head *tmp;
2062
2063         LASSERT_SPIN_LOCKED(&imp->imp_lock);
2064
2065         if (req->rq_transno == 0) {
2066                 DEBUG_REQ(D_EMERG, req, "saving request with zero transno");
2067                 LBUG();
2068         }
2069
2070         /* clear this for new requests that were resent as well
2071            as resent replayed requests. */
2072         lustre_msg_clear_flags(req->rq_reqmsg, MSG_RESENT);
2073
2074         /* don't re-add requests that have been replayed */
2075         if (!list_empty(&req->rq_replay_list))
2076                 return;
2077
2078         lustre_msg_add_flags(req->rq_reqmsg, MSG_REPLAY);
2079
2080         LASSERT(imp->imp_replayable);
2081         /* Balanced in ptlrpc_free_committed, usually. */
2082         ptlrpc_request_addref(req);
2083         list_for_each_prev(tmp, &imp->imp_replay_list) {
2084                 struct ptlrpc_request *iter =
2085                         list_entry(tmp, struct ptlrpc_request, rq_replay_list);
2086
2087                 /* We may have duplicate transnos if we create and then
2088                  * open a file, or for closes retained if to match creating
2089                  * opens, so use req->rq_xid as a secondary key.
2090                  * (See bugs 684, 685, and 428.)
2091                  * XXX no longer needed, but all opens need transnos!
2092                  */
2093                 if (iter->rq_transno > req->rq_transno)
2094                         continue;
2095
2096                 if (iter->rq_transno == req->rq_transno) {
2097                         LASSERT(iter->rq_xid != req->rq_xid);
2098                         if (iter->rq_xid > req->rq_xid)
2099                                 continue;
2100                 }
2101
2102                 list_add(&req->rq_replay_list, &iter->rq_replay_list);
2103                 return;
2104         }
2105
2106         list_add_tail(&req->rq_replay_list, &imp->imp_replay_list);
2107 }
2108
2109 int ptlrpc_queue_wait(struct ptlrpc_request *req)
2110 {
2111         int rc = 0;
2112         int brc;
2113         struct l_wait_info lwi;
2114         struct obd_import *imp = req->rq_import;
2115         cfs_duration_t timeout = CFS_TICK;
2116         long timeoutl;
2117         ENTRY;
2118
2119         LASSERT(req->rq_set == NULL);
2120         LASSERT(!req->rq_receiving_reply);
2121
2122         /* for distributed debugging */
2123         lustre_msg_set_status(req->rq_reqmsg, cfs_curproc_pid());
2124         LASSERT(imp->imp_obd != NULL);
2125         CDEBUG(D_RPCTRACE, "Sending RPC pname:cluuid:pid:xid:nid:opc "
2126                "%s:%s:%d:"LPU64":%s:%d\n", cfs_curproc_comm(),
2127                imp->imp_obd->obd_uuid.uuid,
2128                lustre_msg_get_status(req->rq_reqmsg), req->rq_xid,
2129                libcfs_nid2str(imp->imp_connection->c_peer.nid),
2130                lustre_msg_get_opc(req->rq_reqmsg));
2131
2132         /* Mark phase here for a little debug help */
2133         ptlrpc_rqphase_move(req, RQ_PHASE_RPC);
2134
2135         spin_lock(&imp->imp_lock);
2136         req->rq_import_generation = imp->imp_generation;
2137 restart:
2138         if (ptlrpc_import_delay_req(imp, req, &rc)) {
2139                 list_del_init(&req->rq_list);
2140                 list_add_tail(&req->rq_list, &imp->imp_delayed_list);
2141                 atomic_inc(&imp->imp_inflight);
2142                 spin_unlock(&imp->imp_lock);
2143
2144                 DEBUG_REQ(D_HA, req, "\"%s\" waiting for recovery: (%s != %s)",
2145                           cfs_curproc_comm(),
2146                           ptlrpc_import_state_name(req->rq_send_state),
2147                           ptlrpc_import_state_name(imp->imp_state));
2148                 lwi = LWI_INTR(interrupted_request, req);
2149                 rc = l_wait_event(req->rq_reply_waitq,
2150                                   (req->rq_send_state == imp->imp_state ||
2151                                    req->rq_err || req->rq_intr),
2152                                   &lwi);
2153                 DEBUG_REQ(D_HA, req, "\"%s\" awake: (%s == %s or %d/%d == 1)",
2154                           cfs_curproc_comm(),
2155                           ptlrpc_import_state_name(imp->imp_state),
2156                           ptlrpc_import_state_name(req->rq_send_state),
2157                           req->rq_err, req->rq_intr);
2158
2159                 spin_lock(&imp->imp_lock);
2160                 list_del_init(&req->rq_list);
2161                 atomic_dec(&imp->imp_inflight);
2162
2163                 if (req->rq_err) {
2164                         /* rq_status was set locally */
2165                         rc = -EIO;
2166                 }
2167                 else if (req->rq_intr) {
2168                         rc = -EINTR;
2169                 }
2170                 else if (req->rq_no_resend) {
2171                         spin_unlock(&imp->imp_lock);
2172                         GOTO(out, rc = -ETIMEDOUT);
2173                 }
2174                 else {
2175                         GOTO(restart, rc);
2176                 }
2177         }
2178
2179         if (rc != 0) {
2180                 spin_unlock(&imp->imp_lock);
2181                 req->rq_status = rc; // XXX this ok?
2182                 GOTO(out, rc);
2183         }
2184
2185         if (req->rq_resend) {
2186                 if (req->rq_bulk != NULL) {
2187                         ptlrpc_unregister_bulk(req, 0);
2188
2189                         /* bulk requests are supposed to be
2190                          * idempotent, so we are free to bump the xid
2191                          * here, which we need to do before
2192                          * registering the bulk again (bug 6371).
2193                          * print the old xid first for sanity.
2194                          */
2195                         DEBUG_REQ(D_HA, req, "bumping xid for bulk: ");
2196                         req->rq_xid = ptlrpc_next_xid();
2197                 }
2198
2199                 DEBUG_REQ(D_HA, req, "resending: ");
2200         }
2201
2202         /* XXX this is the same as ptlrpc_set_wait */
2203         LASSERT(list_empty(&req->rq_list));
2204         list_add_tail(&req->rq_list, &imp->imp_sending_list);
2205         atomic_inc(&imp->imp_inflight);
2206         spin_unlock(&imp->imp_lock);
2207
2208         rc = sptlrpc_req_refresh_ctx(req, 0);
2209         if (rc) {
2210                 if (req->rq_err) {
2211                         /* we got fatal ctx refresh error, directly jump out
2212                          * thus we can pass back the actual error code.
2213                          */
2214                         spin_lock(&imp->imp_lock);
2215                         list_del_init(&req->rq_list);
2216                         atomic_dec(&imp->imp_inflight);
2217                         spin_unlock(&imp->imp_lock);
2218
2219                         CERROR("Failed to refresh ctx of req %p: %d\n",
2220                                req, rc);
2221                         GOTO(out, rc);
2222                 }
2223                 /* simulating we got error during send rpc */
2224                 goto after_send;
2225         }
2226
2227         rc = ptl_send_rpc(req, 0);
2228         if (rc)
2229                 DEBUG_REQ(D_HA, req, "send failed (%d); recovering", rc);
2230 repeat:
2231         timeoutl = req->rq_deadline - cfs_time_current_sec();
2232         timeout = (timeoutl <= 0 || rc) ? CFS_TICK :
2233                 cfs_time_seconds(timeoutl);
2234         DEBUG_REQ(D_NET, req,
2235                   "-- sleeping for "CFS_DURATION_T" ticks", timeout);
2236         lwi = LWI_TIMEOUT_INTR(timeout, expired_request, interrupted_request,
2237                                req);
2238         brc = l_wait_event(req->rq_reply_waitq, ptlrpc_check_reply(req), &lwi);
2239         if (brc == -ETIMEDOUT && ((req->rq_deadline > cfs_time_current_sec()) ||
2240                                  ptlrpc_check_and_wait_suspend(req)))
2241                 goto repeat;
2242
2243 after_send:
2244         CDEBUG(D_RPCTRACE, "Completed RPC pname:cluuid:pid:xid:nid:opc "
2245                "%s:%s:%d:"LPU64":%s:%d\n", cfs_curproc_comm(),
2246                imp->imp_obd->obd_uuid.uuid,
2247                lustre_msg_get_status(req->rq_reqmsg), req->rq_xid,
2248                libcfs_nid2str(imp->imp_connection->c_peer.nid),
2249                lustre_msg_get_opc(req->rq_reqmsg));
2250
2251         /* If the reply was received normally, this just grabs the spinlock
2252          * (ensuring the reply callback has returned), sees that
2253          * req->rq_receiving_reply is clear and returns. */
2254         ptlrpc_unregister_reply(req, 0);
2255
2256         spin_lock(&imp->imp_lock);
2257         list_del_init(&req->rq_list);
2258         atomic_dec(&imp->imp_inflight);
2259         spin_unlock(&imp->imp_lock);
2260
2261         if (req->rq_err) {
2262                 DEBUG_REQ(D_RPCTRACE, req, "err rc=%d status=%d",
2263                           rc, req->rq_status);
2264                 GOTO(out, rc = rc ? rc : -EIO);
2265         }
2266
2267         if (req->rq_intr) {
2268                 /* Should only be interrupted if we timed out. */
2269                 if (!req->rq_timedout)
2270                         DEBUG_REQ(D_ERROR, req,
2271                                   "rq_intr set but rq_timedout not");
2272                 GOTO(out, rc = -EINTR);
2273         }
2274
2275         /* Resend if we need to */
2276         if (req->rq_resend||req->rq_timedout) {
2277                 /* ...unless we were specifically told otherwise. */
2278                 if (req->rq_no_resend)
2279                         GOTO(out, rc = -ETIMEDOUT);
2280                 spin_lock(&imp->imp_lock);
2281                 /* we can have rq_timeout on dlm fake import which not support
2282                  * recovery - but me need resend request on this import instead
2283                  * of return error */
2284                 req->rq_resend = 1;
2285                 goto restart;
2286         }
2287
2288         if (!ptlrpc_client_replied(req)) {
2289                 /* How can this be? -eeb */
2290                 DEBUG_REQ(D_ERROR, req, "!rq_replied: ");
2291                 LBUG();
2292                 GOTO(out, rc = req->rq_status);
2293         }
2294
2295         rc = after_reply(req);
2296         /* NB may return +ve success rc */
2297         if (req->rq_resend) {
2298                 spin_lock(&imp->imp_lock);
2299                 goto restart;
2300         }
2301
2302  out:
2303         if (req->rq_bulk != NULL) {
2304                 if (rc >= 0) {
2305                         /* success so far.  Note that anything going wrong
2306                          * with bulk now, is EXTREMELY strange, since the
2307                          * server must have believed that the bulk
2308                          * tranferred OK before she replied with success to
2309                          * me. */
2310                         lwi = LWI_TIMEOUT(timeout, NULL, NULL);
2311                         brc = l_wait_event(req->rq_reply_waitq,
2312                                            !ptlrpc_client_bulk_active(req),
2313                                            &lwi);
2314                         LASSERT(brc == 0 || brc == -ETIMEDOUT);
2315                         if (brc != 0) {
2316                                 LASSERT(brc == -ETIMEDOUT);
2317                                 DEBUG_REQ(D_ERROR, req, "bulk timed out");
2318                                 rc = brc;
2319                         } else if (!req->rq_bulk->bd_success) {
2320                                 DEBUG_REQ(D_ERROR, req, "bulk transfer failed");
2321                                 rc = -EIO;
2322                         }
2323                 }
2324                 if (rc < 0)
2325                         ptlrpc_unregister_bulk(req, 0);
2326         }
2327
2328         LASSERT(!req->rq_receiving_reply);
2329         ptlrpc_rqphase_move(req, RQ_PHASE_INTERPRET);
2330         cfs_waitq_broadcast(&imp->imp_recovery_waitq);
2331         RETURN(rc);
2332 }
2333
2334 struct ptlrpc_replay_async_args {
2335         int praa_old_state;
2336         int praa_old_status;
2337 };
2338
2339 static int ptlrpc_replay_interpret(const struct lu_env *env,
2340                                    struct ptlrpc_request *req,
2341                                    void * data, int rc)
2342 {
2343         struct ptlrpc_replay_async_args *aa = data;
2344         struct obd_import *imp = req->rq_import;
2345
2346         ENTRY;
2347         atomic_dec(&imp->imp_replay_inflight);
2348
2349         if (!ptlrpc_client_replied(req)) {
2350                 CERROR("request replay timed out, restarting recovery\n");
2351                 GOTO(out, rc = -ETIMEDOUT);
2352         }
2353
2354         if (lustre_msg_get_type(req->rq_repmsg) == PTL_RPC_MSG_ERR &&
2355             (lustre_msg_get_status(req->rq_repmsg) == -ENOTCONN ||
2356              lustre_msg_get_status(req->rq_repmsg) == -ENODEV))
2357                 GOTO(out, rc = lustre_msg_get_status(req->rq_repmsg));
2358
2359         /** VBR: check version failure */
2360         if (lustre_msg_get_status(req->rq_repmsg) == -EOVERFLOW) {
2361                 /** replay was failed due to version mismatch */
2362                 DEBUG_REQ(D_WARNING, req, "Version mismatch during replay\n");
2363                 spin_lock(&imp->imp_lock);
2364                 imp->imp_vbr_failed = 1;
2365                 imp->imp_no_lock_replay = 1;
2366                 spin_unlock(&imp->imp_lock);
2367         } else {
2368                 /** The transno had better not change over replay. */
2369                 LASSERTF(lustre_msg_get_transno(req->rq_reqmsg) ==
2370                          lustre_msg_get_transno(req->rq_repmsg) ||
2371                          lustre_msg_get_transno(req->rq_repmsg) == 0,
2372                          LPX64"/"LPX64"\n",
2373                          lustre_msg_get_transno(req->rq_reqmsg),
2374                          lustre_msg_get_transno(req->rq_repmsg));
2375         }
2376
2377         spin_lock(&imp->imp_lock);
2378         /** if replays by version then gap was occur on server, no trust to locks */
2379         if (lustre_msg_get_flags(req->rq_repmsg) & MSG_VERSION_REPLAY)
2380                 imp->imp_no_lock_replay = 1;
2381         imp->imp_last_replay_transno = lustre_msg_get_transno(req->rq_reqmsg);
2382         spin_unlock(&imp->imp_lock);
2383         LASSERT(imp->imp_last_replay_transno);
2384
2385         DEBUG_REQ(D_HA, req, "got rep");
2386
2387         /* let the callback do fixups, possibly including in the request */
2388         if (req->rq_replay_cb)
2389                 req->rq_replay_cb(req);
2390
2391         if (ptlrpc_client_replied(req) &&
2392             lustre_msg_get_status(req->rq_repmsg) != aa->praa_old_status) {
2393                 DEBUG_REQ(D_ERROR, req, "status %d, old was %d",
2394                           lustre_msg_get_status(req->rq_repmsg),
2395                           aa->praa_old_status);
2396         } else {
2397                 /* Put it back for re-replay. */
2398                 lustre_msg_set_status(req->rq_repmsg, aa->praa_old_status);
2399         }
2400
2401         /*
2402          * Errors while replay can set transno to 0, but
2403          * imp_last_replay_transno shouldn't be set to 0 anyway
2404          */
2405         if (req->rq_transno > 0) {
2406                 spin_lock(&imp->imp_lock);
2407                 LASSERT(req->rq_transno <= imp->imp_last_replay_transno);
2408                 imp->imp_last_replay_transno = req->rq_transno;
2409                 spin_unlock(&imp->imp_lock);
2410         } else
2411                 CERROR("Transno is 0 during replay!\n");
2412         /* continue with recovery */
2413         rc = ptlrpc_import_recovery_state_machine(imp);
2414  out:
2415         req->rq_send_state = aa->praa_old_state;
2416
2417         if (rc != 0)
2418                 /* this replay failed, so restart recovery */
2419                 ptlrpc_connect_import(imp, NULL);
2420
2421         RETURN(rc);
2422 }
2423
2424 int ptlrpc_replay_req(struct ptlrpc_request *req)
2425 {
2426         struct ptlrpc_replay_async_args *aa;
2427         ENTRY;
2428
2429         LASSERT(req->rq_import->imp_state == LUSTRE_IMP_REPLAY);
2430         /* Not handling automatic bulk replay yet (or ever?) */
2431         LASSERT(req->rq_bulk == NULL);
2432
2433         LASSERT (sizeof (*aa) <= sizeof (req->rq_async_args));
2434         aa = ptlrpc_req_async_args(req);
2435         memset(aa, 0, sizeof *aa);
2436
2437         /* Prepare request to be resent with ptlrpcd */
2438         aa->praa_old_state = req->rq_send_state;
2439         req->rq_send_state = LUSTRE_IMP_REPLAY;
2440         req->rq_phase = RQ_PHASE_NEW;
2441         req->rq_next_phase = RQ_PHASE_UNDEFINED;
2442         if (req->rq_repmsg)
2443                 aa->praa_old_status = lustre_msg_get_status(req->rq_repmsg);
2444         req->rq_status = 0;
2445         req->rq_interpret_reply = ptlrpc_replay_interpret;
2446         /* Readjust the timeout for current conditions */
2447         ptlrpc_at_set_req_timeout(req);
2448
2449         DEBUG_REQ(D_HA, req, "REPLAY");
2450
2451         atomic_inc(&req->rq_import->imp_replay_inflight);
2452         ptlrpc_request_addref(req); /* ptlrpcd needs a ref */
2453
2454         ptlrpcd_add_req(req, PSCOPE_OTHER);
2455         RETURN(0);
2456 }
2457
2458 void ptlrpc_abort_inflight(struct obd_import *imp)
2459 {
2460         struct list_head *tmp, *n;
2461         ENTRY;
2462
2463         /* Make sure that no new requests get processed for this import.
2464          * ptlrpc_{queue,set}_wait must (and does) hold imp_lock while testing
2465          * this flag and then putting requests on sending_list or delayed_list.
2466          */
2467         spin_lock(&imp->imp_lock);
2468
2469         /* XXX locking?  Maybe we should remove each request with the list
2470          * locked?  Also, how do we know if the requests on the list are
2471          * being freed at this time?
2472          */
2473         list_for_each_safe(tmp, n, &imp->imp_sending_list) {
2474                 struct ptlrpc_request *req =
2475                         list_entry(tmp, struct ptlrpc_request, rq_list);
2476
2477                 DEBUG_REQ(D_RPCTRACE, req, "inflight");
2478
2479                 spin_lock (&req->rq_lock);
2480                 if (req->rq_import_generation < imp->imp_generation) {
2481                         req->rq_err = 1;
2482                         req->rq_status = -EINTR;
2483                         ptlrpc_client_wake_req(req);
2484                 }
2485                 spin_unlock (&req->rq_lock);
2486         }
2487
2488         list_for_each_safe(tmp, n, &imp->imp_delayed_list) {
2489                 struct ptlrpc_request *req =
2490                         list_entry(tmp, struct ptlrpc_request, rq_list);
2491
2492                 DEBUG_REQ(D_RPCTRACE, req, "aborting waiting req");
2493
2494                 spin_lock (&req->rq_lock);
2495                 if (req->rq_import_generation < imp->imp_generation) {
2496                         req->rq_err = 1;
2497                         req->rq_status = -EINTR;
2498                         ptlrpc_client_wake_req(req);
2499                 }
2500                 spin_unlock (&req->rq_lock);
2501         }
2502
2503         /* Last chance to free reqs left on the replay list, but we
2504          * will still leak reqs that haven't committed.  */
2505         if (imp->imp_replayable)
2506                 ptlrpc_free_committed(imp);
2507
2508         spin_unlock(&imp->imp_lock);
2509
2510         EXIT;
2511 }
2512
2513 void ptlrpc_abort_set(struct ptlrpc_request_set *set)
2514 {
2515         struct list_head *tmp, *pos;
2516
2517         LASSERT(set != NULL);
2518
2519         list_for_each_safe(pos, tmp, &set->set_requests) {
2520                 struct ptlrpc_request *req =
2521                         list_entry(pos, struct ptlrpc_request, rq_set_chain);
2522
2523                 spin_lock(&req->rq_lock);
2524                 if (req->rq_phase != RQ_PHASE_RPC) {
2525                         spin_unlock(&req->rq_lock);
2526                         continue;
2527                 }
2528
2529                 req->rq_err = 1;
2530                 req->rq_status = -EINTR;
2531                 ptlrpc_client_wake_req(req);
2532                 spin_unlock(&req->rq_lock);
2533         }
2534 }
2535
2536 static __u64 ptlrpc_last_xid;
2537 static spinlock_t ptlrpc_last_xid_lock;
2538
2539 /* Initialize the XID for the node.  This is common among all requests on
2540  * this node, and only requires the property that it is monotonically
2541  * increasing.  It does not need to be sequential.  Since this is also used
2542  * as the RDMA match bits, it is important that a single client NOT have
2543  * the same match bits for two different in-flight requests, hence we do
2544  * NOT want to have an XID per target or similar.
2545  *
2546  * To avoid an unlikely collision between match bits after a client reboot
2547  * (which would deliver old data into the wrong RDMA buffer) initialize
2548  * the XID based on the current time, assuming a maximum RPC rate of 1M RPC/s.
2549  * If the time is clearly incorrect, we instead use a 62-bit random number.
2550  * In the worst case the random number will overflow 1M RPCs per second in
2551  * 9133 years, or permutations thereof.
2552  */
2553 #define YEAR_2004 (1ULL << 30)
2554 void ptlrpc_init_xid(void)
2555 {
2556         time_t now = cfs_time_current_sec();
2557
2558         spin_lock_init(&ptlrpc_last_xid_lock);
2559         if (now < YEAR_2004) {
2560                 ll_get_random_bytes(&ptlrpc_last_xid, sizeof(ptlrpc_last_xid));
2561                 ptlrpc_last_xid >>= 2;
2562                 ptlrpc_last_xid |= (1ULL << 61);
2563         } else {
2564                 ptlrpc_last_xid = (__u64)now << 20;
2565         }
2566 }
2567
2568 __u64 ptlrpc_next_xid(void)
2569 {
2570         __u64 tmp;
2571         spin_lock(&ptlrpc_last_xid_lock);
2572         tmp = ++ptlrpc_last_xid;
2573         spin_unlock(&ptlrpc_last_xid_lock);
2574         return tmp;
2575 }
2576
2577 __u64 ptlrpc_sample_next_xid(void)
2578 {
2579 #if BITS_PER_LONG == 32
2580         /* need to avoid possible word tearing on 32-bit systems */
2581         __u64 tmp;
2582         spin_lock(&ptlrpc_last_xid_lock);
2583         tmp = ptlrpc_last_xid + 1;
2584         spin_unlock(&ptlrpc_last_xid_lock);
2585         return tmp;
2586 #else
2587         /* No need to lock, since returned value is racy anyways */
2588         return ptlrpc_last_xid + 1;
2589 #endif
2590 }
2591 EXPORT_SYMBOL(ptlrpc_sample_next_xid);