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