lustre/ptlrpc/client.c

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