lustre/ptlrpc/client.c

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