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