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         if (!pool)
 360                 return;
 361
 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         OBD_FREE(pool, sizeof(*pool));
 371 }
 372
 373 void ptlrpc_add_rqs_to_pool(struct ptlrpc_request_pool *pool, int num_rq)
 374 {
 375         int i;
 376         int size = 1;
 377
 378         while (size < pool->prp_rq_size + SPTLRPC_MAX_PAYLOAD)
 379                 size <<= 1;
 380
 381         LASSERTF(list_empty(&pool->prp_req_list) || size == pool->prp_rq_size,
 382                  "Trying to change pool size with nonempty pool "
 383                  "from %d to %d bytes\n", pool->prp_rq_size, size);
 384
 385         spin_lock(&pool->prp_lock);
 386         pool->prp_rq_size = size;
 387         for (i = 0; i < num_rq; i++) {
 388                 struct ptlrpc_request *req;
 389                 struct lustre_msg *msg;
 390
 391                 spin_unlock(&pool->prp_lock);
 392                 OBD_ALLOC(req, sizeof(struct ptlrpc_request));
 393                 if (!req)
 394                         return;
 395                 OBD_ALLOC_GFP(msg, size, CFS_ALLOC_STD);
 396                 if (!msg) {
 397                         OBD_FREE(req, sizeof(struct ptlrpc_request));
 398                         return;
 399                 }
 400                 req->rq_reqbuf = msg;
 401                 req->rq_reqbuf_len = size;
 402                 req->rq_pool = pool;
 403                 spin_lock(&pool->prp_lock);
 404                 list_add_tail(&req->rq_list, &pool->prp_req_list);
 405         }
 406         spin_unlock(&pool->prp_lock);
 407         return;
 408 }
 409
 410 struct ptlrpc_request_pool *ptlrpc_init_rq_pool(int num_rq, int msgsize,
 411                                                 void (*populate_pool)(struct ptlrpc_request_pool *, int))
 412 {
 413         struct ptlrpc_request_pool *pool;
 414
 415         OBD_ALLOC(pool, sizeof (struct ptlrpc_request_pool));
 416         if (!pool)
 417                 return NULL;
 418
 419         /* Request next power of two for the allocation, because internally
 420            kernel would do exactly this */
 421
 422         spin_lock_init(&pool->prp_lock);
 423         CFS_INIT_LIST_HEAD(&pool->prp_req_list);
 424         pool->prp_rq_size = msgsize;
 425         pool->prp_populate = populate_pool;
 426
 427         populate_pool(pool, num_rq);
 428
 429         if (list_empty(&pool->prp_req_list)) {
 430                 /* have not allocated a single request for the pool */
 431                 OBD_FREE(pool, sizeof (struct ptlrpc_request_pool));
 432                 pool = NULL;
 433         }
 434         return pool;
 435 }
 436
 437 static struct ptlrpc_request *ptlrpc_prep_req_from_pool(struct ptlrpc_request_pool *pool)
 438 {
 439         struct ptlrpc_request *request;
 440         struct lustre_msg *reqbuf;
 441
 442         if (!pool)
 443                 return NULL;
 444
 445         spin_lock(&pool->prp_lock);
 446
 447         /* See if we have anything in a pool, and bail out if nothing,
 448          * in writeout path, where this matters, this is safe to do, because
 449          * nothing is lost in this case, and when some in-flight requests
 450          * complete, this code will be called again. */
 451         if (unlikely(list_empty(&pool->prp_req_list))) {
 452                 spin_unlock(&pool->prp_lock);
 453                 return NULL;
 454         }
 455
 456         request = list_entry(pool->prp_req_list.next, struct ptlrpc_request,
 457                              rq_list);
 458         list_del(&request->rq_list);
 459         spin_unlock(&pool->prp_lock);
 460
 461         LASSERT(request->rq_reqbuf);
 462         LASSERT(request->rq_pool);
 463
 464         reqbuf = request->rq_reqbuf;
 465         memset(request, 0, sizeof(*request));
 466         request->rq_reqbuf = reqbuf;
 467         request->rq_reqbuf_len = pool->prp_rq_size;
 468         request->rq_pool = pool;
 469
 470         return request;
 471 }
 472
 473 static void __ptlrpc_free_req_to_pool(struct ptlrpc_request *request)
 474 {
 475         struct ptlrpc_request_pool *pool = request->rq_pool;
 476
 477         spin_lock(&pool->prp_lock);
 478         LASSERT(list_empty(&request->rq_list));
 479         list_add_tail(&request->rq_list, &pool->prp_req_list);
 480         spin_unlock(&pool->prp_lock);
 481 }
 482
 483 static int __ptlrpc_request_bufs_pack(struct ptlrpc_request *request,
 484                                       __u32 version, int opcode,
 485                                       int count, __u32 *lengths, char **bufs,
 486                                       struct ptlrpc_cli_ctx *ctx)
 487 {
 488         struct obd_import  *imp = request->rq_import;
 489         int                 rc;
 490         ENTRY;
 491
 492         if (unlikely(ctx))
 493                 request->rq_cli_ctx = sptlrpc_cli_ctx_get(ctx);
 494         else {
 495                 rc = sptlrpc_req_get_ctx(request);
 496                 if (rc)
 497                         GOTO(out_free, rc);
 498         }
 499
 500         sptlrpc_req_set_flavor(request, opcode);
 501
 502         rc = lustre_pack_request(request, imp->imp_msg_magic, count,
 503                                  lengths, bufs);
 504         if (rc) {
 505                 LASSERT(!request->rq_pool);
 506                 GOTO(out_ctx, rc);
 507         }
 508
 509         lustre_msg_add_version(request->rq_reqmsg, version);
 510         request->rq_send_state = LUSTRE_IMP_FULL;
 511         request->rq_type = PTL_RPC_MSG_REQUEST;
 512         request->rq_export = NULL;
 513
 514         request->rq_req_cbid.cbid_fn  = request_out_callback;
 515         request->rq_req_cbid.cbid_arg = request;
 516
 517         request->rq_reply_cbid.cbid_fn  = reply_in_callback;
 518         request->rq_reply_cbid.cbid_arg = request;
 519
 520         request->rq_phase = RQ_PHASE_NEW;
 521
 522         request->rq_request_portal = imp->imp_client->cli_request_portal;
 523         request->rq_reply_portal = imp->imp_client->cli_reply_portal;
 524
 525         ptlrpc_at_set_req_timeout(request);
 526
 527         spin_lock_init(&request->rq_lock);
 528         CFS_INIT_LIST_HEAD(&request->rq_list);
 529         CFS_INIT_LIST_HEAD(&request->rq_timed_list);
 530         CFS_INIT_LIST_HEAD(&request->rq_replay_list);
 531         CFS_INIT_LIST_HEAD(&request->rq_mod_list);
 532         CFS_INIT_LIST_HEAD(&request->rq_ctx_chain);
 533         CFS_INIT_LIST_HEAD(&request->rq_set_chain);
 534         CFS_INIT_LIST_HEAD(&request->rq_history_list);
 535         cfs_waitq_init(&request->rq_reply_waitq);
 536         request->rq_xid = ptlrpc_next_xid();
 537         atomic_set(&request->rq_refcount, 1);
 538
 539         lustre_msg_set_opc(request->rq_reqmsg, opcode);
 540
 541         RETURN(0);
 542 out_ctx:
 543         sptlrpc_cli_ctx_put(request->rq_cli_ctx, 1);
 544 out_free:
 545         class_import_put(imp);
 546         return rc;
 547 }
 548
 549 int ptlrpc_request_bufs_pack(struct ptlrpc_request *request,
 550                              __u32 version, int opcode, char **bufs,
 551                              struct ptlrpc_cli_ctx *ctx)
 552 {
 553         int count;
 554
 555         count = req_capsule_filled_sizes(&request->rq_pill, RCL_CLIENT);
 556         return __ptlrpc_request_bufs_pack(request, version, opcode, count,
 557                                           request->rq_pill.rc_area[RCL_CLIENT],
 558                                           bufs, ctx);
 559 }
 560 EXPORT_SYMBOL(ptlrpc_request_bufs_pack);
 561
 562 int ptlrpc_request_pack(struct ptlrpc_request *request,
 563                         __u32 version, int opcode)
 564 {
 565         return ptlrpc_request_bufs_pack(request, version, opcode, NULL, NULL);
 566 }
 567
 568 static inline
 569 struct ptlrpc_request *__ptlrpc_request_alloc(struct obd_import *imp,
 570                                               struct ptlrpc_request_pool *pool)
 571 {
 572         struct ptlrpc_request *request = NULL;
 573
 574         if (pool)
 575                 request = ptlrpc_prep_req_from_pool(pool);
 576
 577         if (!request)
 578                 OBD_ALLOC_PTR(request);
 579
 580         if (request) {
 581                 LASSERTF((unsigned long)imp > 0x1000, "%p", imp);
 582                 LASSERT(imp != LP_POISON);
 583                 LASSERTF((unsigned long)imp->imp_client > 0x1000, "%p",
 584                         imp->imp_client);
 585                 LASSERT(imp->imp_client != LP_POISON);
 586
 587                 request->rq_import = class_import_get(imp);
 588         } else {
 589                 CERROR("request allocation out of memory\n");
 590         }
 591
 592         return request;
 593 }
 594
 595 static struct ptlrpc_request *
 596 ptlrpc_request_alloc_internal(struct obd_import *imp,
 597                               struct ptlrpc_request_pool * pool,
 598                               const struct req_format *format)
 599 {
 600         struct ptlrpc_request *request;
 601
 602         request = __ptlrpc_request_alloc(imp, pool);
 603         if (request == NULL)
 604                 return NULL;
 605
 606         req_capsule_init(&request->rq_pill, request, RCL_CLIENT);
 607         req_capsule_set(&request->rq_pill, format);
 608         return request;
 609 }
 610
 611 struct ptlrpc_request *ptlrpc_request_alloc(struct obd_import *imp,
 612                                             const struct req_format *format)
 613 {
 614         return ptlrpc_request_alloc_internal(imp, NULL, format);
 615 }
 616
 617 struct ptlrpc_request *ptlrpc_request_alloc_pool(struct obd_import *imp,
 618                                             struct ptlrpc_request_pool * pool,
 619                                             const struct req_format *format)
 620 {
 621         return ptlrpc_request_alloc_internal(imp, pool, format);
 622 }
 623
 624 void ptlrpc_request_free(struct ptlrpc_request *request)
 625 {
 626         if (request->rq_pool)
 627                 __ptlrpc_free_req_to_pool(request);
 628         else
 629                 OBD_FREE_PTR(request);
 630 }
 631
 632 struct ptlrpc_request *ptlrpc_request_alloc_pack(struct obd_import *imp,
 633                                                 const struct req_format *format,
 634                                                 __u32 version, int opcode)
 635 {
 636         struct ptlrpc_request *req = ptlrpc_request_alloc(imp, format);
 637         int                    rc;
 638
 639         if (req) {
 640                 rc = ptlrpc_request_pack(req, version, opcode);
 641                 if (rc) {
 642                         ptlrpc_request_free(req);
 643                         req = NULL;
 644                 }
 645         }
 646         return req;
 647 }
 648
 649 struct ptlrpc_request *
 650 ptlrpc_prep_req_pool(struct obd_import *imp,
 651                      __u32 version, int opcode,
 652                      int count, __u32 *lengths, char **bufs,
 653                      struct ptlrpc_request_pool *pool)
 654 {
 655         struct ptlrpc_request *request;
 656         int                    rc;
 657
 658         request = __ptlrpc_request_alloc(imp, pool);
 659         if (!request)
 660                 return NULL;
 661
 662         rc = __ptlrpc_request_bufs_pack(request, version, opcode, count,
 663                                         lengths, bufs, NULL);
 664         if (rc) {
 665                 ptlrpc_request_free(request);
 666                 request = NULL;
 667         }
 668         return request;
 669 }
 670
 671 struct ptlrpc_request *
 672 ptlrpc_prep_req(struct obd_import *imp, __u32 version, int opcode, int count,
 673                 __u32 *lengths, char **bufs)
 674 {
 675         return ptlrpc_prep_req_pool(imp, version, opcode, count, lengths, bufs,
 676                                     NULL);
 677 }
 678
 679 struct ptlrpc_request_set *ptlrpc_prep_set(void)
 680 {
 681         struct ptlrpc_request_set *set;
 682
 683         ENTRY;
 684         OBD_ALLOC(set, sizeof *set);
 685         if (!set)
 686                 RETURN(NULL);
 687         CFS_INIT_LIST_HEAD(&set->set_requests);
 688         cfs_waitq_init(&set->set_waitq);
 689         set->set_remaining = 0;
 690         spin_lock_init(&set->set_new_req_lock);
 691         CFS_INIT_LIST_HEAD(&set->set_new_requests);
 692         CFS_INIT_LIST_HEAD(&set->set_cblist);
 693
 694         RETURN(set);
 695 }
 696
 697 /* Finish with this set; opposite of prep_set. */
 698 void ptlrpc_set_destroy(struct ptlrpc_request_set *set)
 699 {
 700         struct list_head *tmp;
 701         struct list_head *next;
 702         int               expected_phase;
 703         int               n = 0;
 704         ENTRY;
 705
 706         /* Requests on the set should either all be completed, or all be new */
 707         expected_phase = (set->set_remaining == 0) ?
 708                          RQ_PHASE_COMPLETE : RQ_PHASE_NEW;
 709         list_for_each (tmp, &set->set_requests) {
 710                 struct ptlrpc_request *req =
 711                         list_entry(tmp, struct ptlrpc_request, rq_set_chain);
 712
 713                 LASSERT(req->rq_phase == expected_phase);
 714                 n++;
 715         }
 716
 717         LASSERT(set->set_remaining == 0 || set->set_remaining == n);
 718
 719         list_for_each_safe(tmp, next, &set->set_requests) {
 720                 struct ptlrpc_request *req =
 721                         list_entry(tmp, struct ptlrpc_request, rq_set_chain);
 722                 list_del_init(&req->rq_set_chain);
 723
 724                 LASSERT(req->rq_phase == expected_phase);
 725
 726                 if (req->rq_phase == RQ_PHASE_NEW) {
 727
 728                         if (req->rq_interpret_reply != NULL) {
 729                                 ptlrpc_interpterer_t interpreter =
 730                                         req->rq_interpret_reply;
 731
 732                                 /* higher level (i.e. LOV) failed;
 733                                  * let the sub reqs clean up */
 734                                 req->rq_status = -EBADR;
 735                                 interpreter(NULL, req, &req->rq_async_args,
 736                                             req->rq_status);
 737                         }
 738                         set->set_remaining--;
 739                 }
 740
 741                 req->rq_set = NULL;
 742                 ptlrpc_req_finished (req);
 743         }
 744
 745         LASSERT(set->set_remaining == 0);
 746
 747         OBD_FREE(set, sizeof(*set));
 748         EXIT;
 749 }
 750
 751 int ptlrpc_set_add_cb(struct ptlrpc_request_set *set,
 752                       set_interpreter_func fn, void *data)
 753 {
 754         struct ptlrpc_set_cbdata *cbdata;
 755
 756         OBD_ALLOC_PTR(cbdata);
 757         if (cbdata == NULL)
 758                 RETURN(-ENOMEM);
 759
 760         cbdata->psc_interpret = fn;
 761         cbdata->psc_data = data;
 762         list_add_tail(&cbdata->psc_item, &set->set_cblist);
 763
 764         RETURN(0);
 765 }
 766
 767 void ptlrpc_set_add_req(struct ptlrpc_request_set *set,
 768                         struct ptlrpc_request *req)
 769 {
 770         /* The set takes over the caller's request reference */
 771         list_add_tail(&req->rq_set_chain, &set->set_requests);
 772         req->rq_set = set;
 773         set->set_remaining++;
 774
 775         atomic_inc(&req->rq_import->imp_inflight);
 776 }
 777
 778 /**
 779  * Lock so many callers can add things, the context that owns the set
 780  * is supposed to notice these and move them into the set proper.
 781  */
 782 int ptlrpc_set_add_new_req(struct ptlrpcd_ctl *pc,
 783                            struct ptlrpc_request *req)
 784 {
 785         struct ptlrpc_request_set *set = pc->pc_set;
 786
 787         /*
 788          * Let caller know that we stopped and will not handle this request.
 789          * It needs to take care itself of request.
 790          */
 791         if (test_bit(LIOD_STOP, &pc->pc_flags))
 792                 return -EALREADY;
 793
 794         spin_lock(&set->set_new_req_lock);
 795         /*
 796          * The set takes over the caller's request reference.
 797          */
 798         list_add_tail(&req->rq_set_chain, &set->set_new_requests);
 799         req->rq_set = set;
 800         spin_unlock(&set->set_new_req_lock);
 801
 802         /*
 803          * Let thead know that we added something and better it to wake up
 804          * and process.
 805          */
 806         cfs_waitq_signal(&set->set_waitq);
 807         return 0;
 808 }
 809
 810 /*
 811  * Based on the current state of the import, determine if the request
 812  * can be sent, is an error, or should be delayed.
 813  *
 814  * Returns true if this request should be delayed. If false, and
 815  * *status is set, then the request can not be sent and *status is the
 816  * error code.  If false and status is 0, then request can be sent.
 817  *
 818  * The imp->imp_lock must be held.
 819  */
 820 static int ptlrpc_import_delay_req(struct obd_import *imp,
 821                                    struct ptlrpc_request *req, int *status)
 822 {
 823         int delay = 0;
 824         ENTRY;
 825
 826         LASSERT (status != NULL);
 827         *status = 0;
 828
 829         if (req->rq_ctx_init || req->rq_ctx_fini) {
 830                 /* always allow ctx init/fini rpc go through */
 831         } else if (imp->imp_state == LUSTRE_IMP_NEW) {
 832                 DEBUG_REQ(D_ERROR, req, "Uninitialized import.");
 833                 *status = -EIO;
 834                 LBUG();
 835         } else if (imp->imp_state == LUSTRE_IMP_CLOSED) {
 836                 DEBUG_REQ(D_ERROR, req, "IMP_CLOSED ");
 837                 *status = -EIO;
 838         } else if (req->rq_send_state == LUSTRE_IMP_CONNECTING &&
 839                    imp->imp_state == LUSTRE_IMP_CONNECTING) {
 840                 /* allow CONNECT even if import is invalid */ ;
 841                 if (atomic_read(&imp->imp_inval_count) != 0) {
 842                         DEBUG_REQ(D_ERROR, req, "invalidate in flight");
 843                         *status = -EIO;
 844                 }
 845         } else if ((imp->imp_invalid && (!imp->imp_recon_bk)) ||
 846                                          imp->imp_obd->obd_no_recov) {
 847                 /* If the import has been invalidated (such as by an OST
 848                  * failure), and if the import(MGC) tried all of its connection
 849                  * list (Bug 13464), the request must fail with -ESHUTDOWN.
 850                  * This indicates the requests should be discarded; an -EIO
 851                  * may result in a resend of the request. */
 852                 if (!imp->imp_deactive)
 853                           DEBUG_REQ(D_ERROR, req, "IMP_INVALID");
 854                 *status = -ESHUTDOWN; /* bz 12940 */
 855         } else if (req->rq_import_generation != imp->imp_generation) {
 856                 DEBUG_REQ(D_ERROR, req, "req wrong generation:");
 857                 *status = -EIO;
 858         } else if (req->rq_send_state != imp->imp_state) {
 859                 /* invalidate in progress - any requests should be drop */
 860                 if (atomic_read(&imp->imp_inval_count) != 0) {
 861                         DEBUG_REQ(D_ERROR, req, "invalidate in flight");
 862                         *status = -EIO;
 863                 } else if (imp->imp_dlm_fake || req->rq_no_delay) {
 864                         *status = -EWOULDBLOCK;
 865                 } else {
 866                         delay = 1;
 867                 }
 868         }
 869
 870         RETURN(delay);
 871 }
 872
 873 static int ptlrpc_check_reply(struct ptlrpc_request *req)
 874 {
 875         int rc = 0;
 876         ENTRY;
 877
 878         /* serialise with network callback */
 879         spin_lock(&req->rq_lock);
 880
 881         if (req->rq_replied)
 882                 GOTO(out, rc = 1);
 883
 884         if (req->rq_net_err && !req->rq_timedout) {
 885                 spin_unlock(&req->rq_lock);
 886                 rc = ptlrpc_expire_one_request(req);
 887                 spin_lock(&req->rq_lock);
 888                 GOTO(out, rc);
 889         }
 890
 891         if (req->rq_err)
 892                 GOTO(out, rc = 1);
 893
 894         if (req->rq_resend)
 895                 GOTO(out, rc = 1);
 896
 897         if (req->rq_restart)
 898                 GOTO(out, rc = 1);
 899
 900         if (req->rq_early) {
 901                 ptlrpc_at_recv_early_reply(req);
 902                 GOTO(out, rc = 0); /* keep waiting */
 903         }
 904
 905         EXIT;
 906  out:
 907         spin_unlock(&req->rq_lock);
 908         DEBUG_REQ(D_NET, req, "rc = %d for", rc);
 909         return rc;
 910 }
 911
 912 static int ptlrpc_check_status(struct ptlrpc_request *req)
 913 {
 914         int err;
 915         ENTRY;
 916
 917         err = lustre_msg_get_status(req->rq_repmsg);
 918         if (lustre_msg_get_type(req->rq_repmsg) == PTL_RPC_MSG_ERR) {
 919                 struct obd_import *imp = req->rq_import;
 920                 __u32 opc = lustre_msg_get_opc(req->rq_reqmsg);
 921                 LCONSOLE_ERROR_MSG(0x011,"an error occurred while communicating"
 922                                 " with %s. The %s operation failed with %d\n",
 923                                 libcfs_nid2str(imp->imp_connection->c_peer.nid),
 924                                 ll_opcode2str(opc), err);
 925                 RETURN(err < 0 ? err : -EINVAL);
 926         }
 927
 928         if (err < 0) {
 929                 DEBUG_REQ(D_INFO, req, "status is %d", err);
 930         } else if (err > 0) {
 931                 /* XXX: translate this error from net to host */
 932                 DEBUG_REQ(D_INFO, req, "status is %d", err);
 933         }
 934
 935         RETURN(err);
 936 }
 937
 938 /**
 939  * Callback function called when client receives RPC reply for \a req.
 940  */
 941 static int after_reply(struct ptlrpc_request *req)
 942 {
 943         struct obd_import *imp = req->rq_import;
 944         struct obd_device *obd = req->rq_import->imp_obd;
 945         int rc;
 946         struct timeval work_start;
 947         long timediff;
 948         ENTRY;
 949
 950         LASSERT(!req->rq_receiving_reply);
 951         LASSERT(obd);
 952         LASSERT(req->rq_nob_received <= req->rq_repbuf_len);
 953
 954         /*
 955          * NB Until this point, the whole of the incoming message,
 956          * including buflens, status etc is in the sender's byte order.
 957          */
 958
 959         rc = sptlrpc_cli_unwrap_reply(req);
 960         if (rc) {
 961                 DEBUG_REQ(D_ERROR, req, "unwrap reply failed (%d):", rc);
 962                 RETURN(rc);
 963         }
 964
 965         /*
 966          * Security layer unwrap might ask resend this request.
 967          */
 968         if (req->rq_resend)
 969                 RETURN(0);
 970
 971         rc = unpack_reply(req);
 972         if (rc)
 973                 RETURN(rc);
 974
 975         do_gettimeofday(&work_start);
 976         timediff = cfs_timeval_sub(&work_start, &req->rq_arrival_time, NULL);
 977         if (obd->obd_svc_stats != NULL)
 978                 lprocfs_counter_add(obd->obd_svc_stats, PTLRPC_REQWAIT_CNTR,
 979                                     timediff);
 980
 981         if (lustre_msg_get_type(req->rq_repmsg) != PTL_RPC_MSG_REPLY &&
 982             lustre_msg_get_type(req->rq_repmsg) != PTL_RPC_MSG_ERR) {
 983                 DEBUG_REQ(D_ERROR, req, "invalid packet received (type=%u)",
 984                           lustre_msg_get_type(req->rq_repmsg));
 985                 RETURN(-EPROTO);
 986         }
 987
 988         OBD_FAIL_TIMEOUT(OBD_FAIL_PTLRPC_PAUSE_REP, obd_fail_val);
 989         ptlrpc_at_adj_service(req, lustre_msg_get_timeout(req->rq_repmsg));
 990         ptlrpc_at_adj_net_latency(req,
 991                                   lustre_msg_get_service_time(req->rq_repmsg));
 992
 993         rc = ptlrpc_check_status(req);
 994         imp->imp_connect_error = rc;
 995
 996         if (rc) {
 997                 /*
 998                  * Either we've been evicted, or the server has failed for
 999                  * some reason. Try to reconnect, and if that fails, punt to
1000                  * the upcall.
1001                  */
1002                 if (ll_rpc_recoverable_error(rc)) {
1003                         if (req->rq_send_state != LUSTRE_IMP_FULL ||
1004                             imp->imp_obd->obd_no_recov || imp->imp_dlm_fake) {
1005                                 RETURN(rc);
1006                         }
1007                         ptlrpc_request_handle_notconn(req);
1008                         RETURN(rc);
1009                 }
1010         } else {
1011                 /*
1012                  * Let's look if server sent slv. Do it only for RPC with
1013                  * rc == 0.
1014                  */
1015                 ldlm_cli_update_pool(req);
1016         }
1017
1018         /*
1019          * Store transno in reqmsg for replay.
1020          */
1021         req->rq_transno = lustre_msg_get_transno(req->rq_repmsg);
1022         lustre_msg_set_transno(req->rq_reqmsg, req->rq_transno);
1023
1024         if (req->rq_import->imp_replayable) {
1025                 spin_lock(&imp->imp_lock);
1026                 /*
1027                  * No point in adding already-committed requests to the replay
1028                  * list, we will just remove them immediately. b=9829
1029                  */
1030                 if (req->rq_transno != 0 &&
1031                     (req->rq_transno >
1032                      lustre_msg_get_last_committed(req->rq_repmsg) ||
1033                      req->rq_replay))
1034                         ptlrpc_retain_replayable_request(req, imp);
1035                 else if (req->rq_commit_cb != NULL) {
1036                         spin_unlock(&imp->imp_lock);
1037                         req->rq_commit_cb(req);
1038                         spin_lock(&imp->imp_lock);
1039                 }
1040
1041                 /*
1042                  * Replay-enabled imports return commit-status information.
1043                  */
1044                 if (lustre_msg_get_last_committed(req->rq_repmsg)) {
1045                         imp->imp_peer_committed_transno =
1046                                 lustre_msg_get_last_committed(req->rq_repmsg);
1047                 }
1048                 ptlrpc_free_committed(imp);
1049                 spin_unlock(&imp->imp_lock);
1050         }
1051
1052         RETURN(rc);
1053 }
1054
1055 static int ptlrpc_send_new_req(struct ptlrpc_request *req)
1056 {
1057         struct obd_import     *imp;
1058         int rc;
1059         ENTRY;
1060
1061         LASSERT(req->rq_phase == RQ_PHASE_NEW);
1062         if (req->rq_sent && (req->rq_sent > cfs_time_current_sec()))
1063                 RETURN (0);
1064
1065         req->rq_phase = RQ_PHASE_RPC;
1066
1067         imp = req->rq_import;
1068         spin_lock(&imp->imp_lock);
1069
1070         req->rq_import_generation = imp->imp_generation;
1071
1072         if (ptlrpc_import_delay_req(imp, req, &rc)) {
1073                 spin_lock (&req->rq_lock);
1074                 req->rq_waiting = 1;
1075                 spin_unlock (&req->rq_lock);
1076
1077                 DEBUG_REQ(D_HA, req, "req from PID %d waiting for recovery: "
1078                           "(%s != %s)",
1079                           lustre_msg_get_status(req->rq_reqmsg) ,
1080                           ptlrpc_import_state_name(req->rq_send_state),
1081                           ptlrpc_import_state_name(imp->imp_state));
1082                 LASSERT(list_empty (&req->rq_list));
1083
1084                 list_add_tail(&req->rq_list, &imp->imp_delayed_list);
1085                 spin_unlock(&imp->imp_lock);
1086                 RETURN(0);
1087         }
1088
1089         if (rc != 0) {
1090                 spin_unlock(&imp->imp_lock);
1091                 req->rq_status = rc;
1092                 req->rq_phase = RQ_PHASE_INTERPRET;
1093                 RETURN(rc);
1094         }
1095
1096         /* XXX this is the same as ptlrpc_queue_wait */
1097         LASSERT(list_empty(&req->rq_list));
1098         list_add_tail(&req->rq_list, &imp->imp_sending_list);
1099         spin_unlock(&imp->imp_lock);
1100
1101         lustre_msg_set_status(req->rq_reqmsg, cfs_curproc_pid());
1102
1103         rc = sptlrpc_req_refresh_ctx(req, -1);
1104         if (rc) {
1105                 if (req->rq_err) {
1106                         req->rq_status = rc;
1107                         RETURN(1);
1108                 } else {
1109                         req->rq_wait_ctx = 1;
1110                         RETURN(0);
1111                 }
1112         }
1113
1114         CDEBUG(D_RPCTRACE, "Sending RPC pname:cluuid:pid:xid:nid:opc"
1115                " %s:%s:%d:"LPU64":%s:%d\n", cfs_curproc_comm(),
1116                imp->imp_obd->obd_uuid.uuid,
1117                lustre_msg_get_status(req->rq_reqmsg), req->rq_xid,
1118                libcfs_nid2str(imp->imp_connection->c_peer.nid),
1119                lustre_msg_get_opc(req->rq_reqmsg));
1120
1121         rc = ptl_send_rpc(req, 0);
1122         if (rc) {
1123                 DEBUG_REQ(D_HA, req, "send failed (%d); expect timeout", rc);
1124                 req->rq_net_err = 1;
1125                 RETURN(rc);
1126         }
1127         RETURN(0);
1128 }
1129
1130 /* this sends any unsent RPCs in @set and returns TRUE if all are sent */
1131 int ptlrpc_check_set(const struct lu_env *env, struct ptlrpc_request_set *set)
1132 {
1133         struct list_head *tmp;
1134         int force_timer_recalc = 0;
1135         ENTRY;
1136
1137         if (set->set_remaining == 0)
1138                 RETURN(1);
1139
1140         list_for_each(tmp, &set->set_requests) {
1141                 struct ptlrpc_request *req =
1142                         list_entry(tmp, struct ptlrpc_request, rq_set_chain);
1143                 struct obd_import *imp = req->rq_import;
1144                 int rc = 0;
1145
1146                 if (req->rq_phase == RQ_PHASE_NEW &&
1147                     ptlrpc_send_new_req(req)) {
1148                         force_timer_recalc = 1;
1149                 }
1150                 /* delayed send - skip */
1151                 if (req->rq_phase == RQ_PHASE_NEW && req->rq_sent)
1152                         continue;
1153
1154                 if (!(req->rq_phase == RQ_PHASE_RPC ||
1155                       req->rq_phase == RQ_PHASE_BULK ||
1156                       req->rq_phase == RQ_PHASE_INTERPRET ||
1157                       req->rq_phase == RQ_PHASE_COMPLETE)) {
1158                         DEBUG_REQ(D_ERROR, req, "bad phase %x", req->rq_phase);
1159                         LBUG();
1160                 }
1161
1162                 if (req->rq_phase == RQ_PHASE_COMPLETE)
1163                         continue;
1164
1165                 if (req->rq_phase == RQ_PHASE_INTERPRET)
1166                         GOTO(interpret, req->rq_status);
1167
1168                 if (req->rq_net_err && !req->rq_timedout)
1169                         ptlrpc_expire_one_request(req);
1170
1171                 if (req->rq_err) {
1172                         ptlrpc_unregister_reply(req);
1173                         req->rq_replied = 0;
1174                         if (req->rq_status == 0)
1175                                 req->rq_status = -EIO;
1176                         req->rq_phase = RQ_PHASE_INTERPRET;
1177
1178                         spin_lock(&imp->imp_lock);
1179                         list_del_init(&req->rq_list);
1180                         spin_unlock(&imp->imp_lock);
1181
1182                         GOTO(interpret, req->rq_status);
1183                 }
1184
1185                 /* ptlrpc_queue_wait->l_wait_event guarantees that rq_intr
1186                  * will only be set after rq_timedout, but the oig waiting
1187                  * path sets rq_intr irrespective of whether ptlrpcd has
1188                  * seen a timeout.  our policy is to only interpret
1189                  * interrupted rpcs after they have timed out */
1190                 if (req->rq_intr && (req->rq_timedout || req->rq_waiting ||
1191                                      req->rq_wait_ctx)) {
1192                         /* NB could be on delayed list */
1193                         ptlrpc_unregister_reply(req);
1194                         req->rq_status = -EINTR;
1195                         req->rq_phase = RQ_PHASE_INTERPRET;
1196
1197                         spin_lock(&imp->imp_lock);
1198                         list_del_init(&req->rq_list);
1199                         spin_unlock(&imp->imp_lock);
1200
1201                         GOTO(interpret, req->rq_status);
1202                 }
1203
1204                 if (req->rq_phase == RQ_PHASE_RPC) {
1205                         if (req->rq_timedout || req->rq_resend ||
1206                             req->rq_waiting || req->rq_wait_ctx) {
1207                                 int status;
1208
1209                                 ptlrpc_unregister_reply(req);
1210
1211                                 spin_lock(&imp->imp_lock);
1212
1213                                 if (ptlrpc_import_delay_req(imp, req, &status)){
1214                                         spin_unlock(&imp->imp_lock);
1215                                         continue;
1216                                 }
1217
1218                                 list_del_init(&req->rq_list);
1219                                 if (status != 0)  {
1220                                         req->rq_status = status;
1221                                         req->rq_phase = RQ_PHASE_INTERPRET;
1222                                         spin_unlock(&imp->imp_lock);
1223                                         GOTO(interpret, req->rq_status);
1224                                 }
1225                                 if (req->rq_no_resend && !req->rq_wait_ctx) {
1226                                         req->rq_status = -ENOTCONN;
1227                                         req->rq_phase = RQ_PHASE_INTERPRET;
1228                                         spin_unlock(&imp->imp_lock);
1229                                         GOTO(interpret, req->rq_status);
1230                                 }
1231                                 list_add_tail(&req->rq_list,
1232                                               &imp->imp_sending_list);
1233
1234                                 spin_unlock(&imp->imp_lock);
1235
1236                                 req->rq_waiting = 0;
1237
1238                                 if (req->rq_timedout||req->rq_resend) {
1239                                         /* This is re-sending anyways,
1240                                          * let's mark req as resend. */
1241                                         req->rq_resend = 1;
1242                                         lustre_msg_add_flags(req->rq_reqmsg,
1243                                                              MSG_RESENT);
1244                                         if (req->rq_bulk) {
1245                                                 __u64 old_xid = req->rq_xid;
1246
1247                                                 ptlrpc_unregister_bulk (req);
1248
1249                                                 /* ensure previous bulk fails */
1250                                                 req->rq_xid = ptlrpc_next_xid();
1251                                                 CDEBUG(D_HA, "resend bulk "
1252                                                        "old x"LPU64
1253                                                        " new x"LPU64"\n",
1254                                                        old_xid, req->rq_xid);
1255                                         }
1256                                 }
1257                                 /*
1258                                  * rq_wait_ctx is only touched by ptlrpcd,
1259                                  * so no lock is needed here.
1260                                  */
1261                                 status = sptlrpc_req_refresh_ctx(req, -1);
1262                                 if (status) {
1263                                         if (req->rq_err) {
1264                                                 req->rq_status = status;
1265                                                 force_timer_recalc = 1;
1266                                         } else {
1267                                                 req->rq_wait_ctx = 1;
1268                                         }
1269
1270                                         continue;
1271                                 } else {
1272                                         req->rq_wait_ctx = 0;
1273                                 }
1274
1275                                 rc = ptl_send_rpc(req, 0);
1276                                 if (rc) {
1277                                         DEBUG_REQ(D_HA, req, "send failed (%d)",
1278                                                   rc);
1279                                         force_timer_recalc = 1;
1280                                         req->rq_net_err = 1;
1281                                 }
1282                                 /* need to reset the timeout */
1283                                 force_timer_recalc = 1;
1284                         }
1285
1286                         spin_lock(&req->rq_lock);
1287
1288                         if (req->rq_early) {
1289                                 ptlrpc_at_recv_early_reply(req);
1290                                 spin_unlock(&req->rq_lock);
1291                                 continue;
1292                         }
1293
1294                         /* Still waiting for a reply? */
1295                         if (req->rq_receiving_reply) {
1296                                 spin_unlock(&req->rq_lock);
1297                                 continue;
1298                         }
1299
1300                         /* Did we actually receive a reply? */
1301                         if (!req->rq_replied) {
1302                                 spin_unlock(&req->rq_lock);
1303                                 continue;
1304                         }
1305
1306                         spin_unlock(&req->rq_lock);
1307
1308                         spin_lock(&imp->imp_lock);
1309                         list_del_init(&req->rq_list);
1310                         spin_unlock(&imp->imp_lock);
1311
1312                         req->rq_status = after_reply(req);
1313                         if (req->rq_resend) {
1314                                 /* Add this req to the delayed list so
1315                                    it can be errored if the import is
1316                                    evicted after recovery. */
1317                                 spin_lock(&imp->imp_lock);
1318                                 list_add_tail(&req->rq_list,
1319                                               &imp->imp_delayed_list);
1320                                 spin_unlock(&imp->imp_lock);
1321                                 continue;
1322                         }
1323
1324                         /* If there is no bulk associated with this request,
1325                          * then we're done and should let the interpreter
1326                          * process the reply.  Similarly if the RPC returned
1327                          * an error, and therefore the bulk will never arrive.
1328                          */
1329                         if (req->rq_bulk == NULL || req->rq_status != 0) {
1330                                 req->rq_phase = RQ_PHASE_INTERPRET;
1331                                 GOTO(interpret, req->rq_status);
1332                         }
1333
1334                         req->rq_phase = RQ_PHASE_BULK;
1335                 }
1336
1337                 LASSERT(req->rq_phase == RQ_PHASE_BULK);
1338                 if (ptlrpc_bulk_active(req->rq_bulk))
1339                         continue;
1340
1341                 if (!req->rq_bulk->bd_success) {
1342                         /* The RPC reply arrived OK, but the bulk screwed
1343                          * up!  Dead wierd since the server told us the RPC
1344                          * was good after getting the REPLY for her GET or
1345                          * the ACK for her PUT. */
1346                         DEBUG_REQ(D_ERROR, req, "bulk transfer failed");
1347                         LBUG();
1348                 }
1349
1350                 req->rq_phase = RQ_PHASE_INTERPRET;
1351
1352         interpret:
1353                 LASSERT(req->rq_phase == RQ_PHASE_INTERPRET);
1354                 LASSERT(!req->rq_receiving_reply);
1355
1356                 ptlrpc_unregister_reply(req);
1357                 if (req->rq_bulk != NULL)
1358                         ptlrpc_unregister_bulk (req);
1359
1360                 if (req->rq_interpret_reply != NULL) {
1361                         ptlrpc_interpterer_t interpreter =
1362                                 req->rq_interpret_reply;
1363                         req->rq_status = interpreter(NULL, req,
1364                                                      &req->rq_async_args,
1365                                                      req->rq_status);
1366                 }
1367                 req->rq_phase = RQ_PHASE_COMPLETE;
1368
1369                 CDEBUG(D_RPCTRACE, "Completed RPC pname:cluuid:pid:xid:nid:"
1370                        "opc %s:%s:%d:"LPU64":%s:%d\n", cfs_curproc_comm(),
1371                        imp->imp_obd->obd_uuid.uuid,
1372                        lustre_msg_get_status(req->rq_reqmsg), req->rq_xid,
1373                        libcfs_nid2str(imp->imp_connection->c_peer.nid),
1374                        lustre_msg_get_opc(req->rq_reqmsg));
1375
1376                 atomic_dec(&imp->imp_inflight);
1377                 set->set_remaining--;
1378                 cfs_waitq_signal(&imp->imp_recovery_waitq);
1379         }
1380
1381         /* If we hit an error, we want to recover promptly. */
1382         RETURN(set->set_remaining == 0 || force_timer_recalc);
1383 }
1384
1385 /* Return 1 if we should give up, else 0 */
1386 int ptlrpc_expire_one_request(struct ptlrpc_request *req)
1387 {
1388         struct obd_import *imp = req->rq_import;
1389         int rc = 0;
1390         ENTRY;
1391
1392         DEBUG_REQ(D_ERROR|D_NETERROR, req,
1393                   "%s (sent at "CFS_TIME_T", "CFS_DURATION_T"s ago)",
1394                   req->rq_net_err ? "network error" : "timeout",
1395                   req->rq_sent, cfs_time_sub(cfs_time_current_sec(),
1396                   req->rq_sent));
1397
1398         if (imp) {
1399                 LCONSOLE_WARN("Request x"LPU64" sent from %s to NID %s "
1400                               CFS_DURATION_T"s ago has timed out "
1401                               "(limit "CFS_DURATION_T"s).\n", req->rq_xid,
1402                               req->rq_import->imp_obd->obd_name,
1403                               libcfs_nid2str(imp->imp_connection->c_peer.nid),
1404                               cfs_time_sub(cfs_time_current_sec(), req->rq_sent),
1405                               cfs_time_sub(req->rq_deadline, req->rq_sent));
1406         }
1407
1408         if (imp != NULL && obd_debug_peer_on_timeout)
1409                 LNetCtl(IOC_LIBCFS_DEBUG_PEER, &imp->imp_connection->c_peer);
1410
1411         spin_lock(&req->rq_lock);
1412         req->rq_timedout = 1;
1413         spin_unlock(&req->rq_lock);
1414
1415         ptlrpc_unregister_reply (req);
1416
1417         if (obd_dump_on_timeout)
1418                 libcfs_debug_dumplog();
1419
1420         if (req->rq_bulk != NULL)
1421                 ptlrpc_unregister_bulk (req);
1422
1423         if (imp == NULL) {
1424                 DEBUG_REQ(D_HA, req, "NULL import: already cleaned up?");
1425                 RETURN(1);
1426         }
1427
1428         /* The DLM server doesn't want recovery run on its imports. */
1429         if (imp->imp_dlm_fake)
1430                 RETURN(1);
1431
1432         /* If this request is for recovery or other primordial tasks,
1433          * then error it out here. */
1434         if (req->rq_ctx_init || req->rq_ctx_fini ||
1435             req->rq_send_state != LUSTRE_IMP_FULL ||
1436             imp->imp_obd->obd_no_recov) {
1437                 DEBUG_REQ(D_RPCTRACE, req, "err -110, sent_state=%s (now=%s)",
1438                           ptlrpc_import_state_name(req->rq_send_state),
1439                           ptlrpc_import_state_name(imp->imp_state));
1440                 spin_lock(&req->rq_lock);
1441                 req->rq_status = -ETIMEDOUT;
1442                 req->rq_err = 1;
1443                 spin_unlock(&req->rq_lock);
1444                 RETURN(1);
1445         }
1446
1447         /* if a request can't be resent we can't wait for an answer after
1448            the timeout */
1449         if (req->rq_no_resend) {
1450                 DEBUG_REQ(D_RPCTRACE, req, "TIMEOUT-NORESEND:");
1451                 rc = 1;
1452         }
1453
1454         ptlrpc_fail_import(imp, lustre_msg_get_conn_cnt(req->rq_reqmsg));
1455
1456         RETURN(rc);
1457 }
1458
1459 int ptlrpc_expired_set(void *data)
1460 {
1461         struct ptlrpc_request_set *set = data;
1462         struct list_head          *tmp;
1463         time_t                     now = cfs_time_current_sec();
1464         ENTRY;
1465
1466         LASSERT(set != NULL);
1467
1468         /* A timeout expired; see which reqs it applies to... */
1469         list_for_each (tmp, &set->set_requests) {
1470                 struct ptlrpc_request *req =
1471                         list_entry(tmp, struct ptlrpc_request, rq_set_chain);
1472
1473                 /* request in-flight? */
1474                 if (!(((req->rq_phase == RQ_PHASE_RPC) && !req->rq_waiting &&
1475                        !req->rq_resend) ||
1476                       (req->rq_phase == RQ_PHASE_BULK)))
1477                         continue;
1478
1479                 if (req->rq_timedout ||           /* already dealt with */
1480                     req->rq_deadline > now)       /* not expired */
1481                         continue;
1482
1483                 /* deal with this guy */
1484                 ptlrpc_expire_one_request (req);
1485         }
1486
1487         /* When waiting for a whole set, we always to break out of the
1488          * sleep so we can recalculate the timeout, or enable interrupts
1489          * iff everyone's timed out.
1490          */
1491         RETURN(1);
1492 }
1493
1494 void ptlrpc_mark_interrupted(struct ptlrpc_request *req)
1495 {
1496         spin_lock(&req->rq_lock);
1497         req->rq_intr = 1;
1498         spin_unlock(&req->rq_lock);
1499 }
1500
1501 void ptlrpc_interrupted_set(void *data)
1502 {
1503         struct ptlrpc_request_set *set = data;
1504         struct list_head *tmp;
1505
1506         LASSERT(set != NULL);
1507         CERROR("INTERRUPTED SET %p\n", set);
1508
1509         list_for_each(tmp, &set->set_requests) {
1510                 struct ptlrpc_request *req =
1511                         list_entry(tmp, struct ptlrpc_request, rq_set_chain);
1512
1513                 if (req->rq_phase != RQ_PHASE_RPC)
1514                         continue;
1515
1516                 ptlrpc_mark_interrupted(req);
1517         }
1518 }
1519
1520 /* get the smallest timeout in the set; this does NOT set a timeout. */
1521 int ptlrpc_set_next_timeout(struct ptlrpc_request_set *set)
1522 {
1523         struct list_head      *tmp;
1524         time_t                 now = cfs_time_current_sec();
1525         int                    timeout = 0;
1526         struct ptlrpc_request *req;
1527         int                    deadline;
1528         ENTRY;
1529
1530         SIGNAL_MASK_ASSERT(); /* XXX BUG 1511 */
1531
1532         list_for_each(tmp, &set->set_requests) {
1533                 req = list_entry(tmp, struct ptlrpc_request, rq_set_chain);
1534
1535                 /* request in-flight? */
1536                 if (!((req->rq_phase == RQ_PHASE_RPC && !req->rq_waiting) ||
1537                       (req->rq_phase == RQ_PHASE_BULK) ||
1538                       (req->rq_phase == RQ_PHASE_NEW)))
1539                         continue;
1540
1541                 if (req->rq_timedout)   /* already timed out */
1542                         continue;
1543
1544                 if (req->rq_wait_ctx)   /* waiting for ctx */
1545                         continue;
1546
1547                 if (req->rq_phase == RQ_PHASE_NEW)
1548                         deadline = req->rq_sent;
1549                 else
1550                         deadline = req->rq_sent + req->rq_timeout;
1551
1552                 if (deadline <= now)    /* actually expired already */
1553                         timeout = 1;    /* ASAP */
1554                 else if (timeout == 0 || timeout > deadline - now)
1555                         timeout = deadline - now;
1556         }
1557         RETURN(timeout);
1558 }
1559
1560 int ptlrpc_set_wait(struct ptlrpc_request_set *set)
1561 {
1562         struct list_head      *tmp;
1563         struct ptlrpc_request *req;
1564         struct l_wait_info     lwi;
1565         int                    rc, timeout;
1566         ENTRY;
1567
1568         if (list_empty(&set->set_requests))
1569                 RETURN(0);
1570
1571         list_for_each(tmp, &set->set_requests) {
1572                 req = list_entry(tmp, struct ptlrpc_request, rq_set_chain);
1573                 if (req->rq_phase == RQ_PHASE_NEW)
1574                         (void)ptlrpc_send_new_req(req);
1575         }
1576
1577         do {
1578                 timeout = ptlrpc_set_next_timeout(set);
1579
1580                 /* wait until all complete, interrupted, or an in-flight
1581                  * req times out */
1582                 CDEBUG(D_RPCTRACE, "set %p going to sleep for %d seconds\n",
1583                        set, timeout);
1584                 lwi = LWI_TIMEOUT_INTR(cfs_time_seconds(timeout ? timeout : 1),
1585                                        ptlrpc_expired_set,
1586                                        ptlrpc_interrupted_set, set);
1587                 rc = l_wait_event(set->set_waitq,
1588                                   ptlrpc_check_set(NULL, set), &lwi);
1589
1590                 LASSERT(rc == 0 || rc == -EINTR || rc == -ETIMEDOUT);
1591
1592                 /* -EINTR => all requests have been flagged rq_intr so next
1593                  * check completes.
1594                  * -ETIMEOUTD => someone timed out.  When all reqs have
1595                  * timed out, signals are enabled allowing completion with
1596                  * EINTR.
1597                  * I don't really care if we go once more round the loop in
1598                  * the error cases -eeb. */
1599         } while (rc != 0 || set->set_remaining != 0);
1600
1601         LASSERT(set->set_remaining == 0);
1602
1603         rc = 0;
1604         list_for_each(tmp, &set->set_requests) {
1605                 req = list_entry(tmp, struct ptlrpc_request, rq_set_chain);
1606
1607                 LASSERT(req->rq_phase == RQ_PHASE_COMPLETE);
1608                 if (req->rq_status != 0)
1609                         rc = req->rq_status;
1610         }
1611
1612         if (set->set_interpret != NULL) {
1613                 int (*interpreter)(struct ptlrpc_request_set *set,void *,int) =
1614                         set->set_interpret;
1615                 rc = interpreter (set, set->set_arg, rc);
1616         } else {
1617                 struct ptlrpc_set_cbdata *cbdata, *n;
1618                 int err;
1619
1620                 list_for_each_entry_safe(cbdata, n,
1621                                          &set->set_cblist, psc_item) {
1622                         list_del_init(&cbdata->psc_item);
1623                         err = cbdata->psc_interpret(set, cbdata->psc_data, rc);
1624                         if (err && !rc)
1625                                 rc = err;
1626                         OBD_FREE_PTR(cbdata);
1627                 }
1628         }
1629
1630         RETURN(rc);
1631 }
1632
1633 static void __ptlrpc_free_req(struct ptlrpc_request *request, int locked)
1634 {
1635         ENTRY;
1636         if (request == NULL) {
1637                 EXIT;
1638                 return;
1639         }
1640
1641         LASSERTF(!request->rq_receiving_reply, "req %p\n", request);
1642         LASSERTF(request->rq_rqbd == NULL, "req %p\n",request);/* client-side */
1643         LASSERTF(list_empty(&request->rq_list), "req %p\n", request);
1644         LASSERTF(list_empty(&request->rq_set_chain), "req %p\n", request);
1645         LASSERTF(!request->rq_replay, "req %p\n", request);
1646         LASSERT(request->rq_cli_ctx);
1647
1648         req_capsule_fini(&request->rq_pill);
1649
1650         /* We must take it off the imp_replay_list first.  Otherwise, we'll set
1651          * request->rq_reqmsg to NULL while osc_close is dereferencing it. */
1652         if (request->rq_import != NULL) {
1653                 if (!locked)
1654                         spin_lock(&request->rq_import->imp_lock);
1655                 list_del_init(&request->rq_mod_list);
1656                 list_del_init(&request->rq_replay_list);
1657                 if (!locked)
1658                         spin_unlock(&request->rq_import->imp_lock);
1659         }
1660         LASSERTF(list_empty(&request->rq_replay_list), "req %p\n", request);
1661
1662         if (atomic_read(&request->rq_refcount) != 0) {
1663                 DEBUG_REQ(D_ERROR, request,
1664                           "freeing request with nonzero refcount");
1665                 LBUG();
1666         }
1667
1668         if (request->rq_repbuf != NULL)
1669                 sptlrpc_cli_free_repbuf(request);
1670         if (request->rq_export != NULL) {
1671                 class_export_put(request->rq_export);
1672                 request->rq_export = NULL;
1673         }
1674         if (request->rq_import != NULL) {
1675                 class_import_put(request->rq_import);
1676                 request->rq_import = NULL;
1677         }
1678         if (request->rq_bulk != NULL)
1679                 ptlrpc_free_bulk(request->rq_bulk);
1680
1681         if (request->rq_reqbuf != NULL || request->rq_clrbuf != NULL)
1682                 sptlrpc_cli_free_reqbuf(request);
1683
1684         sptlrpc_req_put_ctx(request, !locked);
1685
1686         if (request->rq_pool)
1687                 __ptlrpc_free_req_to_pool(request);
1688         else
1689                 OBD_FREE(request, sizeof(*request));
1690         EXIT;
1691 }
1692
1693 static int __ptlrpc_req_finished(struct ptlrpc_request *request, int locked);
1694 void ptlrpc_req_finished_with_imp_lock(struct ptlrpc_request *request)
1695 {
1696         LASSERT_SPIN_LOCKED(&request->rq_import->imp_lock);
1697         (void)__ptlrpc_req_finished(request, 1);
1698 }
1699
1700 static int __ptlrpc_req_finished(struct ptlrpc_request *request, int locked)
1701 {
1702         ENTRY;
1703         if (request == NULL)
1704                 RETURN(1);
1705
1706         if (request == LP_POISON ||
1707             request->rq_reqmsg == LP_POISON) {
1708                 CERROR("dereferencing freed request (bug 575)\n");
1709                 LBUG();
1710                 RETURN(1);
1711         }
1712
1713         DEBUG_REQ(D_INFO, request, "refcount now %u",
1714                   atomic_read(&request->rq_refcount) - 1);
1715
1716         if (atomic_dec_and_test(&request->rq_refcount)) {
1717                 __ptlrpc_free_req(request, locked);
1718                 RETURN(1);
1719         }
1720
1721         RETURN(0);
1722 }
1723
1724 void ptlrpc_req_finished(struct ptlrpc_request *request)
1725 {
1726         __ptlrpc_req_finished(request, 0);
1727 }
1728
1729 __u64 ptlrpc_req_xid(struct ptlrpc_request *request)
1730 {
1731         return request->rq_xid;
1732 }
1733 EXPORT_SYMBOL(ptlrpc_req_xid);
1734
1735 /* Disengage the client's reply buffer from the network
1736  * NB does _NOT_ unregister any client-side bulk.
1737  * IDEMPOTENT, but _not_ safe against concurrent callers.
1738  * The request owner (i.e. the thread doing the I/O) must call...
1739  */
1740 void ptlrpc_unregister_reply (struct ptlrpc_request *request)
1741 {
1742         int                rc;
1743         cfs_waitq_t       *wq;
1744         struct l_wait_info lwi;
1745
1746         LASSERT(!in_interrupt ());             /* might sleep */
1747         if (!ptlrpc_client_recv_or_unlink(request))
1748                 /* Nothing left to do */
1749                 return;
1750
1751         LNetMDUnlink (request->rq_reply_md_h);
1752
1753         /* We have to l_wait_event() whatever the result, to give liblustre
1754          * a chance to run reply_in_callback(), and to make sure we've
1755          * unlinked before returning a req to the pool */
1756
1757         if (request->rq_set != NULL)
1758                 wq = &request->rq_set->set_waitq;
1759         else
1760                 wq = &request->rq_reply_waitq;
1761
1762         for (;;) {
1763                 /* Network access will complete in finite time but the HUGE
1764                  * timeout lets us CWARN for visibility of sluggish NALs */
1765                 lwi = LWI_TIMEOUT(cfs_time_seconds(LONG_UNLINK), NULL, NULL);
1766                 rc = l_wait_event (*wq, !ptlrpc_client_recv_or_unlink(request),
1767                                    &lwi);
1768                 if (rc == 0)
1769                         return;
1770
1771                 LASSERT (rc == -ETIMEDOUT);
1772                 DEBUG_REQ(D_WARNING, request, "Unexpectedly long timeout "
1773                           "rvcng=%d unlnk=%d", request->rq_receiving_reply,
1774                           request->rq_must_unlink);
1775         }
1776 }
1777
1778 /* caller must hold imp->imp_lock */
1779 void ptlrpc_free_committed(struct obd_import *imp)
1780 {
1781         struct list_head *tmp, *saved;
1782         struct ptlrpc_request *req;
1783         struct ptlrpc_request *last_req = NULL; /* temporary fire escape */
1784         ENTRY;
1785
1786         LASSERT(imp != NULL);
1787
1788         LASSERT_SPIN_LOCKED(&imp->imp_lock);
1789
1790
1791         if (imp->imp_peer_committed_transno == imp->imp_last_transno_checked &&
1792             imp->imp_generation == imp->imp_last_generation_checked) {
1793                 CDEBUG(D_RPCTRACE, "%s: skip recheck: last_committed "LPU64"\n",
1794                        imp->imp_obd->obd_name, imp->imp_peer_committed_transno);
1795                 EXIT;
1796                 return;
1797         }
1798
1799         CDEBUG(D_RPCTRACE, "%s: committing for last_committed "LPU64" gen %d\n",
1800                imp->imp_obd->obd_name, imp->imp_peer_committed_transno,
1801                imp->imp_generation);
1802         imp->imp_last_transno_checked = imp->imp_peer_committed_transno;
1803         imp->imp_last_generation_checked = imp->imp_generation;
1804
1805         list_for_each_safe(tmp, saved, &imp->imp_replay_list) {
1806                 req = list_entry(tmp, struct ptlrpc_request, rq_replay_list);
1807
1808                 /* XXX ok to remove when 1357 resolved - rread 05/29/03  */
1809                 LASSERT(req != last_req);
1810                 last_req = req;
1811
1812                 if (req->rq_import_generation < imp->imp_generation) {
1813                         DEBUG_REQ(D_RPCTRACE, req, "free request with old gen");
1814                         GOTO(free_req, 0);
1815                 }
1816
1817                 if (req->rq_replay) {
1818                         DEBUG_REQ(D_RPCTRACE, req, "keeping (FL_REPLAY)");
1819                         continue;
1820                 }
1821
1822                 /* not yet committed */
1823                 if (req->rq_transno > imp->imp_peer_committed_transno) {
1824                         DEBUG_REQ(D_RPCTRACE, req, "stopping search");
1825                         break;
1826                 }
1827
1828                 DEBUG_REQ(D_RPCTRACE, req, "commit (last_committed "LPU64")",
1829                           imp->imp_peer_committed_transno);
1830 free_req:
1831                 spin_lock(&req->rq_lock);
1832                 req->rq_replay = 0;
1833                 spin_unlock(&req->rq_lock);
1834                 if (req->rq_commit_cb != NULL)
1835                         req->rq_commit_cb(req);
1836                 list_del_init(&req->rq_replay_list);
1837                 __ptlrpc_req_finished(req, 1);
1838         }
1839
1840         EXIT;
1841         return;
1842 }
1843
1844 void ptlrpc_cleanup_client(struct obd_import *imp)
1845 {
1846         ENTRY;
1847         EXIT;
1848         return;
1849 }
1850
1851 void ptlrpc_resend_req(struct ptlrpc_request *req)
1852 {
1853         DEBUG_REQ(D_HA, req, "going to resend");
1854         lustre_msg_set_handle(req->rq_reqmsg, &(struct lustre_handle){ 0 });
1855         req->rq_status = -EAGAIN;
1856
1857         spin_lock(&req->rq_lock);
1858         req->rq_resend = 1;
1859         req->rq_net_err = 0;
1860         req->rq_timedout = 0;
1861         if (req->rq_bulk) {
1862                 __u64 old_xid = req->rq_xid;
1863
1864                 /* ensure previous bulk fails */
1865                 req->rq_xid = ptlrpc_next_xid();
1866                 CDEBUG(D_HA, "resend bulk old x"LPU64" new x"LPU64"\n",
1867                        old_xid, req->rq_xid);
1868         }
1869         ptlrpc_wake_client_req(req);
1870         spin_unlock(&req->rq_lock);
1871 }
1872
1873 /* XXX: this function and rq_status are currently unused */
1874 void ptlrpc_restart_req(struct ptlrpc_request *req)
1875 {
1876         DEBUG_REQ(D_HA, req, "restarting (possibly-)completed request");
1877         req->rq_status = -ERESTARTSYS;
1878
1879         spin_lock(&req->rq_lock);
1880         req->rq_restart = 1;
1881         req->rq_timedout = 0;
1882         ptlrpc_wake_client_req(req);
1883         spin_unlock(&req->rq_lock);
1884 }
1885
1886 static int expired_request(void *data)
1887 {
1888         struct ptlrpc_request *req = data;
1889         ENTRY;
1890
1891         /* some failure can suspend regular timeouts */
1892         if (ptlrpc_check_suspend())
1893                 RETURN(1);
1894
1895         /* deadline may have changed with an early reply */
1896         if (req->rq_deadline > cfs_time_current_sec())
1897                 RETURN(1);
1898
1899         RETURN(ptlrpc_expire_one_request(req));
1900 }
1901
1902 static void interrupted_request(void *data)
1903 {
1904         struct ptlrpc_request *req = data;
1905         DEBUG_REQ(D_HA, req, "request interrupted");
1906         spin_lock(&req->rq_lock);
1907         req->rq_intr = 1;
1908         spin_unlock(&req->rq_lock);
1909 }
1910
1911 struct ptlrpc_request *ptlrpc_request_addref(struct ptlrpc_request *req)
1912 {
1913         ENTRY;
1914         atomic_inc(&req->rq_refcount);
1915         RETURN(req);
1916 }
1917
1918 void ptlrpc_retain_replayable_request(struct ptlrpc_request *req,
1919                                       struct obd_import *imp)
1920 {
1921         struct list_head *tmp;
1922
1923         LASSERT_SPIN_LOCKED(&imp->imp_lock);
1924
1925         /* clear this for new requests that were resent as well
1926            as resent replayed requests. */
1927         lustre_msg_clear_flags(req->rq_reqmsg, MSG_RESENT);
1928
1929         /* don't re-add requests that have been replayed */
1930         if (!list_empty(&req->rq_replay_list))
1931                 return;
1932
1933         lustre_msg_add_flags(req->rq_reqmsg, MSG_REPLAY);
1934
1935         LASSERT(imp->imp_replayable);
1936         /* Balanced in ptlrpc_free_committed, usually. */
1937         ptlrpc_request_addref(req);
1938         list_for_each_prev(tmp, &imp->imp_replay_list) {
1939                 struct ptlrpc_request *iter =
1940                         list_entry(tmp, struct ptlrpc_request, rq_replay_list);
1941
1942                 /* We may have duplicate transnos if we create and then
1943                  * open a file, or for closes retained if to match creating
1944                  * opens, so use req->rq_xid as a secondary key.
1945                  * (See bugs 684, 685, and 428.)
1946                  * XXX no longer needed, but all opens need transnos!
1947                  */
1948                 if (iter->rq_transno > req->rq_transno)
1949                         continue;
1950
1951                 if (iter->rq_transno == req->rq_transno) {
1952                         LASSERT(iter->rq_xid != req->rq_xid);
1953                         if (iter->rq_xid > req->rq_xid)
1954                                 continue;
1955                 }
1956
1957                 list_add(&req->rq_replay_list, &iter->rq_replay_list);
1958                 return;
1959         }
1960
1961         list_add_tail(&req->rq_replay_list, &imp->imp_replay_list);
1962 }
1963
1964 int ptlrpc_queue_wait(struct ptlrpc_request *req)
1965 {
1966         int rc = 0;
1967         int brc;
1968         struct l_wait_info lwi;
1969         struct obd_import *imp = req->rq_import;
1970         cfs_duration_t timeout = CFS_TICK;
1971         long timeoutl;
1972         ENTRY;
1973
1974         LASSERT(req->rq_set == NULL);
1975         LASSERT(!req->rq_receiving_reply);
1976         atomic_inc(&imp->imp_inflight);
1977
1978         /* for distributed debugging */
1979         lustre_msg_set_status(req->rq_reqmsg, cfs_curproc_pid());
1980         LASSERT(imp->imp_obd != NULL);
1981         CDEBUG(D_RPCTRACE, "Sending RPC pname:cluuid:pid:xid:nid:opc "
1982                "%s:%s:%d:"LPU64":%s:%d\n", cfs_curproc_comm(),
1983                imp->imp_obd->obd_uuid.uuid,
1984                lustre_msg_get_status(req->rq_reqmsg), req->rq_xid,
1985                libcfs_nid2str(imp->imp_connection->c_peer.nid),
1986                lustre_msg_get_opc(req->rq_reqmsg));
1987
1988         /* Mark phase here for a little debug help */
1989         req->rq_phase = RQ_PHASE_RPC;
1990
1991         spin_lock(&imp->imp_lock);
1992         req->rq_import_generation = imp->imp_generation;
1993 restart:
1994         if (ptlrpc_import_delay_req(imp, req, &rc)) {
1995                 list_del(&req->rq_list);
1996
1997                 list_add_tail(&req->rq_list, &imp->imp_delayed_list);
1998                 spin_unlock(&imp->imp_lock);
1999
2000                 DEBUG_REQ(D_HA, req, "\"%s\" waiting for recovery: (%s != %s)",
2001                           cfs_curproc_comm(),
2002                           ptlrpc_import_state_name(req->rq_send_state),
2003                           ptlrpc_import_state_name(imp->imp_state));
2004                 lwi = LWI_INTR(interrupted_request, req);
2005                 rc = l_wait_event(req->rq_reply_waitq,
2006                                   (req->rq_send_state == imp->imp_state ||
2007                                    req->rq_err || req->rq_intr),
2008                                   &lwi);
2009                 DEBUG_REQ(D_HA, req, "\"%s\" awake: (%s == %s or %d/%d == 1)",
2010                           cfs_curproc_comm(),
2011                           ptlrpc_import_state_name(imp->imp_state),
2012                           ptlrpc_import_state_name(req->rq_send_state),
2013                           req->rq_err, req->rq_intr);
2014
2015                 spin_lock(&imp->imp_lock);
2016                 list_del_init(&req->rq_list);
2017
2018                 if (req->rq_err) {
2019                         /* rq_status was set locally */
2020                         rc = -EIO;
2021                 }
2022                 else if (req->rq_intr) {
2023                         rc = -EINTR;
2024                 }
2025                 else if (req->rq_no_resend) {
2026                         spin_unlock(&imp->imp_lock);
2027                         GOTO(out, rc = -ETIMEDOUT);
2028                 }
2029                 else {
2030                         GOTO(restart, rc);
2031                 }
2032         }
2033
2034         if (rc != 0) {
2035                 list_del_init(&req->rq_list);
2036                 spin_unlock(&imp->imp_lock);
2037                 req->rq_status = rc; // XXX this ok?
2038                 GOTO(out, rc);
2039         }
2040
2041         if (req->rq_resend) {
2042                 lustre_msg_add_flags(req->rq_reqmsg, MSG_RESENT);
2043
2044                 if (req->rq_bulk != NULL) {
2045                         ptlrpc_unregister_bulk (req);
2046
2047                         /* bulk requests are supposed to be
2048                          * idempotent, so we are free to bump the xid
2049                          * here, which we need to do before
2050                          * registering the bulk again (bug 6371).
2051                          * print the old xid first for sanity.
2052                          */
2053                         DEBUG_REQ(D_HA, req, "bumping xid for bulk: ");
2054                         req->rq_xid = ptlrpc_next_xid();
2055                 }
2056
2057                 DEBUG_REQ(D_HA, req, "resending: ");
2058         }
2059
2060         /* XXX this is the same as ptlrpc_set_wait */
2061         LASSERT(list_empty(&req->rq_list));
2062         list_add_tail(&req->rq_list, &imp->imp_sending_list);
2063         spin_unlock(&imp->imp_lock);
2064
2065         rc = sptlrpc_req_refresh_ctx(req, 0);
2066         if (rc) {
2067                 if (req->rq_err) {
2068                         /* we got fatal ctx refresh error, directly jump out
2069                          * thus we can pass back the actual error code.
2070                          */
2071                         spin_lock(&imp->imp_lock);
2072                         list_del_init(&req->rq_list);
2073                         spin_unlock(&imp->imp_lock);
2074
2075                         CERROR("Failed to refresh ctx of req %p: %d\n", req, rc);
2076                         GOTO(out, rc);
2077                 }
2078                 /* simulating we got error during send rpc */
2079                 goto after_send;
2080         }
2081
2082         rc = ptl_send_rpc(req, 0);
2083         if (rc)
2084                 DEBUG_REQ(D_HA, req, "send failed (%d); recovering", rc);
2085
2086 repeat:
2087         timeoutl = req->rq_deadline - cfs_time_current_sec();
2088         timeout = (timeoutl <= 0 || rc) ? CFS_TICK :
2089                 cfs_time_seconds(timeoutl);
2090         DEBUG_REQ(D_NET, req,
2091                   "-- sleeping for "CFS_DURATION_T" ticks", timeout);
2092         lwi = LWI_TIMEOUT_INTR(timeout, expired_request, interrupted_request,
2093                                req);
2094         brc = l_wait_event(req->rq_reply_waitq, ptlrpc_check_reply(req), &lwi);
2095         if (brc == -ETIMEDOUT && ((req->rq_deadline > cfs_time_current_sec()) ||
2096                                  ptlrpc_check_and_wait_suspend(req)))
2097                 goto repeat;
2098
2099 after_send:
2100         CDEBUG(D_RPCTRACE, "Completed RPC pname:cluuid:pid:xid:nid:opc "
2101                "%s:%s:%d:"LPU64":%s:%d\n", cfs_curproc_comm(),
2102                imp->imp_obd->obd_uuid.uuid,
2103                lustre_msg_get_status(req->rq_reqmsg), req->rq_xid,
2104                libcfs_nid2str(imp->imp_connection->c_peer.nid),
2105                lustre_msg_get_opc(req->rq_reqmsg));
2106
2107         spin_lock(&imp->imp_lock);
2108         list_del_init(&req->rq_list);
2109         spin_unlock(&imp->imp_lock);
2110
2111         /* If the reply was received normally, this just grabs the spinlock
2112          * (ensuring the reply callback has returned), sees that
2113          * req->rq_receiving_reply is clear and returns. */
2114         ptlrpc_unregister_reply (req);
2115
2116
2117         if (req->rq_err) {
2118                 DEBUG_REQ(D_RPCTRACE, req, "err rc=%d status=%d",
2119                           rc, req->rq_status);
2120                 GOTO(out, rc = rc ? rc : -EIO);
2121         }
2122
2123         if (req->rq_intr) {
2124                 /* Should only be interrupted if we timed out. */
2125                 if (!req->rq_timedout)
2126                         DEBUG_REQ(D_ERROR, req,
2127                                   "rq_intr set but rq_timedout not");
2128                 GOTO(out, rc = -EINTR);
2129         }
2130
2131         /* Resend if we need to */
2132         if (req->rq_resend) {
2133                 /* ...unless we were specifically told otherwise. */
2134                 if (req->rq_no_resend)
2135                         GOTO(out, rc = -ETIMEDOUT);
2136                 spin_lock(&imp->imp_lock);
2137                 goto restart;
2138         }
2139
2140         if (req->rq_timedout) {                 /* non-recoverable timeout */
2141                 GOTO(out, rc = -ETIMEDOUT);
2142         }
2143
2144         if (!req->rq_replied) {
2145                 /* How can this be? -eeb */
2146                 DEBUG_REQ(D_ERROR, req, "!rq_replied: ");
2147                 LBUG();
2148                 GOTO(out, rc = req->rq_status);
2149         }
2150
2151         rc = after_reply(req);
2152         /* NB may return +ve success rc */
2153         if (req->rq_resend) {
2154                 spin_lock(&imp->imp_lock);
2155                 goto restart;
2156         }
2157
2158  out:
2159         if (req->rq_bulk != NULL) {
2160                 if (rc >= 0) {
2161                         /* success so far.  Note that anything going wrong
2162                          * with bulk now, is EXTREMELY strange, since the
2163                          * server must have believed that the bulk
2164                          * tranferred OK before she replied with success to
2165                          * me. */
2166                         lwi = LWI_TIMEOUT(timeout, NULL, NULL);
2167                         brc = l_wait_event(req->rq_reply_waitq,
2168                                            !ptlrpc_bulk_active(req->rq_bulk),
2169                                            &lwi);
2170                         LASSERT(brc == 0 || brc == -ETIMEDOUT);
2171                         if (brc != 0) {
2172                                 LASSERT(brc == -ETIMEDOUT);
2173                                 DEBUG_REQ(D_ERROR, req, "bulk timed out");
2174                                 rc = brc;
2175                         } else if (!req->rq_bulk->bd_success) {
2176                                 DEBUG_REQ(D_ERROR, req, "bulk transfer failed");
2177                                 rc = -EIO;
2178                         }
2179                 }
2180                 if (rc < 0)
2181                         ptlrpc_unregister_bulk (req);
2182         }
2183
2184         LASSERT(!req->rq_receiving_reply);
2185         req->rq_phase = RQ_PHASE_INTERPRET;
2186
2187         atomic_dec(&imp->imp_inflight);
2188         cfs_waitq_signal(&imp->imp_recovery_waitq);
2189         RETURN(rc);
2190 }
2191
2192 struct ptlrpc_replay_async_args {
2193         int praa_old_state;
2194         int praa_old_status;
2195 };
2196
2197 static int ptlrpc_replay_interpret(const struct lu_env *env,
2198                                    struct ptlrpc_request *req,
2199                                     void * data, int rc)
2200 {
2201         struct ptlrpc_replay_async_args *aa = data;
2202         struct obd_import *imp = req->rq_import;
2203
2204         ENTRY;
2205         atomic_dec(&imp->imp_replay_inflight);
2206
2207         if (!req->rq_replied) {
2208                 CERROR("request replay timed out, restarting recovery\n");
2209                 GOTO(out, rc = -ETIMEDOUT);
2210         }
2211
2212         if (lustre_msg_get_type(req->rq_repmsg) == PTL_RPC_MSG_ERR &&
2213             (lustre_msg_get_status(req->rq_repmsg) == -ENOTCONN ||
2214              lustre_msg_get_status(req->rq_repmsg) == -ENODEV))
2215                 GOTO(out, rc = lustre_msg_get_status(req->rq_repmsg));
2216
2217         /* The transno had better not change over replay. */
2218         LASSERT(lustre_msg_get_transno(req->rq_reqmsg) ==
2219                 lustre_msg_get_transno(req->rq_repmsg));
2220
2221         DEBUG_REQ(D_HA, req, "got rep");
2222
2223         /* let the callback do fixups, possibly including in the request */
2224         if (req->rq_replay_cb)
2225                 req->rq_replay_cb(req);
2226
2227         if (req->rq_replied &&
2228             lustre_msg_get_status(req->rq_repmsg) != aa->praa_old_status) {
2229                 DEBUG_REQ(D_ERROR, req, "status %d, old was %d",
2230                           lustre_msg_get_status(req->rq_repmsg),
2231                           aa->praa_old_status);
2232         } else {
2233                 /* Put it back for re-replay. */
2234                 lustre_msg_set_status(req->rq_repmsg, aa->praa_old_status);
2235         }
2236
2237         /*
2238          * Errors while replay can set transno to 0, but
2239          * imp_last_replay_transno shouldn't be set to 0 anyway
2240          */
2241         if (req->rq_transno > 0) {
2242                 spin_lock(&imp->imp_lock);
2243                 LASSERT(req->rq_transno <= imp->imp_last_replay_transno);
2244                 imp->imp_last_replay_transno = req->rq_transno;
2245                 spin_unlock(&imp->imp_lock);
2246         } else
2247                 CERROR("Transno is 0 during replay!\n");
2248         /* continue with recovery */
2249         rc = ptlrpc_import_recovery_state_machine(imp);
2250  out:
2251         req->rq_send_state = aa->praa_old_state;
2252
2253         if (rc != 0)
2254                 /* this replay failed, so restart recovery */
2255                 ptlrpc_connect_import(imp, NULL);
2256
2257         RETURN(rc);
2258 }
2259
2260 int ptlrpc_replay_req(struct ptlrpc_request *req)
2261 {
2262         struct ptlrpc_replay_async_args *aa;
2263         ENTRY;
2264
2265         LASSERT(req->rq_import->imp_state == LUSTRE_IMP_REPLAY);
2266         /* Not handling automatic bulk replay yet (or ever?) */
2267         LASSERT(req->rq_bulk == NULL);
2268
2269         LASSERT (sizeof (*aa) <= sizeof (req->rq_async_args));
2270         aa = ptlrpc_req_async_args(req);
2271         memset(aa, 0, sizeof *aa);
2272
2273         /* Prepare request to be resent with ptlrpcd */
2274         aa->praa_old_state = req->rq_send_state;
2275         req->rq_send_state = LUSTRE_IMP_REPLAY;
2276         req->rq_phase = RQ_PHASE_NEW;
2277         if (req->rq_repmsg)
2278                 aa->praa_old_status = lustre_msg_get_status(req->rq_repmsg);
2279         req->rq_status = 0;
2280         req->rq_interpret_reply = ptlrpc_replay_interpret;
2281         /* Readjust the timeout for current conditions */
2282         ptlrpc_at_set_req_timeout(req);
2283
2284         DEBUG_REQ(D_HA, req, "REPLAY");
2285
2286         atomic_inc(&req->rq_import->imp_replay_inflight);
2287         ptlrpc_request_addref(req); /* ptlrpcd needs a ref */
2288
2289         ptlrpcd_add_req(req);
2290         RETURN(0);
2291 }
2292
2293 void ptlrpc_abort_inflight(struct obd_import *imp)
2294 {
2295         struct list_head *tmp, *n;
2296         ENTRY;
2297
2298         /* Make sure that no new requests get processed for this import.
2299          * ptlrpc_{queue,set}_wait must (and does) hold imp_lock while testing
2300          * this flag and then putting requests on sending_list or delayed_list.
2301          */
2302         spin_lock(&imp->imp_lock);
2303
2304         /* XXX locking?  Maybe we should remove each request with the list
2305          * locked?  Also, how do we know if the requests on the list are
2306          * being freed at this time?
2307          */
2308         list_for_each_safe(tmp, n, &imp->imp_sending_list) {
2309                 struct ptlrpc_request *req =
2310                         list_entry(tmp, struct ptlrpc_request, rq_list);
2311
2312                 DEBUG_REQ(D_RPCTRACE, req, "inflight");
2313
2314                 spin_lock (&req->rq_lock);
2315                 if (req->rq_import_generation < imp->imp_generation) {
2316                         req->rq_err = 1;
2317                         req->rq_status = -EINTR;
2318                         ptlrpc_wake_client_req(req);
2319                 }
2320                 spin_unlock (&req->rq_lock);
2321         }
2322
2323         list_for_each_safe(tmp, n, &imp->imp_delayed_list) {
2324                 struct ptlrpc_request *req =
2325                         list_entry(tmp, struct ptlrpc_request, rq_list);
2326
2327                 DEBUG_REQ(D_RPCTRACE, req, "aborting waiting req");
2328
2329                 spin_lock (&req->rq_lock);
2330                 if (req->rq_import_generation < imp->imp_generation) {
2331                         req->rq_err = 1;
2332                         req->rq_status = -EINTR;
2333                         ptlrpc_wake_client_req(req);
2334                 }
2335                 spin_unlock (&req->rq_lock);
2336         }
2337
2338         /* Last chance to free reqs left on the replay list, but we
2339          * will still leak reqs that haven't committed.  */
2340         if (imp->imp_replayable)
2341                 ptlrpc_free_committed(imp);
2342
2343         spin_unlock(&imp->imp_lock);
2344
2345         EXIT;
2346 }
2347
2348 void ptlrpc_abort_set(struct ptlrpc_request_set *set)
2349 {
2350         struct list_head *tmp, *n;
2351
2352         LASSERT(set != NULL);
2353
2354         list_for_each_safe(tmp, n, &set->set_requests) {
2355                 struct ptlrpc_request *req =
2356                         list_entry(tmp, struct ptlrpc_request, rq_set_chain);
2357
2358                 spin_lock (&req->rq_lock);
2359                 if (req->rq_phase != RQ_PHASE_RPC) {
2360                         spin_unlock (&req->rq_lock);
2361                         continue;
2362                 }
2363
2364                 req->rq_err = 1;
2365                 req->rq_status = -EINTR;
2366                 ptlrpc_wake_client_req(req);
2367                 spin_unlock (&req->rq_lock);
2368         }
2369 }
2370
2371 static __u64 ptlrpc_last_xid;
2372 static spinlock_t ptlrpc_last_xid_lock;
2373
2374 /* Initialize the XID for the node.  This is common among all requests on
2375  * this node, and only requires the property that it is monotonically
2376  * increasing.  It does not need to be sequential.  Since this is also used
2377  * as the RDMA match bits, it is important that a single client NOT have
2378  * the same match bits for two different in-flight requests, hence we do
2379  * NOT want to have an XID per target or similar.
2380  *
2381  * To avoid an unlikely collision between match bits after a client reboot
2382  * (which would cause old to be delivered into the wrong buffer) we initialize
2383  * the XID based on the current time, assuming a maximum RPC rate of 1M RPC/s.
2384  * If the time is clearly incorrect, we instead use a 62-bit random number.
2385  * In the worst case the random number will overflow 1M RPCs per second in
2386  * 9133 years, or permutations thereof.
2387  */
2388 #define YEAR_2004 (1ULL << 30)
2389 void ptlrpc_init_xid(void)
2390 {
2391         time_t now = cfs_time_current_sec();
2392
2393         spin_lock_init(&ptlrpc_last_xid_lock);
2394         if (now < YEAR_2004) {
2395                 ll_get_random_bytes(&ptlrpc_last_xid, sizeof(ptlrpc_last_xid));
2396                 ptlrpc_last_xid >>= 2;
2397                 ptlrpc_last_xid |= (1ULL << 61);
2398         } else {
2399                 ptlrpc_last_xid = (now << 20);
2400         }
2401 }
2402
2403 __u64 ptlrpc_next_xid(void)
2404 {
2405         __u64 tmp;
2406         spin_lock(&ptlrpc_last_xid_lock);
2407         tmp = ++ptlrpc_last_xid;
2408         spin_unlock(&ptlrpc_last_xid_lock);
2409         return tmp;
2410 }
2411
2412 __u64 ptlrpc_sample_next_xid(void)
2413 {
2414 #if BITS_PER_LONG == 32
2415         /* need to avoid possible word tearing on 32-bit systems */
2416         __u64 tmp;
2417         spin_lock(&ptlrpc_last_xid_lock);
2418         tmp = ptlrpc_last_xid + 1;
2419         spin_unlock(&ptlrpc_last_xid_lock);
2420         return tmp;
2421 #else
2422         /* No need to lock, since returned value is racy anyways */
2423         return ptlrpc_last_xid + 1;
2424 #endif
2425 }
2426 EXPORT_SYMBOL(ptlrpc_sample_next_xid);