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
1235                                 if (req->rq_timedout||req->rq_resend) {
1236                                         /* This is re-sending anyways,
1237                                          * let's mark req as resend. */
1238                                         req->rq_resend = 1;
1239                                         lustre_msg_add_flags(req->rq_reqmsg,
1240                                                              MSG_RESENT);
1241                                         if (req->rq_bulk) {
1242                                                 __u64 old_xid = req->rq_xid;
1243
1244                                                 ptlrpc_unregister_bulk (req);
1245
1246                                                 /* ensure previous bulk fails */
1247                                                 req->rq_xid = ptlrpc_next_xid();
1248                                                 CDEBUG(D_HA, "resend bulk "
1249                                                        "old x"LPU64
1250                                                        " new x"LPU64"\n",
1251                                                        old_xid, req->rq_xid);
1252                                         }
1253                                 }
1254                                 /*
1255                                  * rq_wait_ctx is only touched by ptlrpcd,
1256                                  * so no lock is needed here.
1257                                  */
1258                                 status = sptlrpc_req_refresh_ctx(req, -1);
1259                                 if (status) {
1260                                         if (req->rq_err) {
1261                                                 req->rq_status = status;
1262                                                 force_timer_recalc = 1;
1263                                         } else {
1264                                                 req->rq_wait_ctx = 1;
1265                                         }
1266
1267                                         continue;
1268                                 } else {
1269                                         req->rq_wait_ctx = 0;
1270                                 }
1271
1272                                 rc = ptl_send_rpc(req, 0);
1273                                 if (rc) {
1274                                         DEBUG_REQ(D_HA, req, "send failed (%d)",
1275                                                   rc);
1276                                         force_timer_recalc = 1;
1277                                         req->rq_net_err = 1;
1278                                 }
1279                                 /* need to reset the timeout */
1280                                 force_timer_recalc = 1;
1281                         }
1282
1283                         spin_lock(&req->rq_lock);
1284
1285                         if (req->rq_early) {
1286                                 ptlrpc_at_recv_early_reply(req);
1287                                 spin_unlock(&req->rq_lock);
1288                                 continue;
1289                         }
1290
1291                         /* Still waiting for a reply? */
1292                         if (req->rq_receiving_reply) {
1293                                 spin_unlock(&req->rq_lock);
1294                                 continue;
1295                         }
1296
1297                         /* Did we actually receive a reply? */
1298                         if (!req->rq_replied) {
1299                                 spin_unlock(&req->rq_lock);
1300                                 continue;
1301                         }
1302
1303                         spin_unlock(&req->rq_lock);
1304
1305                         spin_lock(&imp->imp_lock);
1306                         list_del_init(&req->rq_list);
1307                         spin_unlock(&imp->imp_lock);
1308
1309                         req->rq_status = after_reply(req);
1310                         if (req->rq_resend) {
1311                                 /* Add this req to the delayed list so
1312                                    it can be errored if the import is
1313                                    evicted after recovery. */
1314                                 spin_lock(&imp->imp_lock);
1315                                 list_add_tail(&req->rq_list,
1316                                               &imp->imp_delayed_list);
1317                                 spin_unlock(&imp->imp_lock);
1318                                 continue;
1319                         }
1320
1321                         /* If there is no bulk associated with this request,
1322                          * then we're done and should let the interpreter
1323                          * process the reply.  Similarly if the RPC returned
1324                          * an error, and therefore the bulk will never arrive.
1325                          */
1326                         if (req->rq_bulk == NULL || req->rq_status != 0) {
1327                                 req->rq_phase = RQ_PHASE_INTERPRET;
1328                                 GOTO(interpret, req->rq_status);
1329                         }
1330
1331                         req->rq_phase = RQ_PHASE_BULK;
1332                 }
1333
1334                 LASSERT(req->rq_phase == RQ_PHASE_BULK);
1335                 if (ptlrpc_bulk_active(req->rq_bulk))
1336                         continue;
1337
1338                 if (!req->rq_bulk->bd_success) {
1339                         /* The RPC reply arrived OK, but the bulk screwed
1340                          * up!  Dead wierd since the server told us the RPC
1341                          * was good after getting the REPLY for her GET or
1342                          * the ACK for her PUT. */
1343                         DEBUG_REQ(D_ERROR, req, "bulk transfer failed");
1344                         LBUG();
1345                 }
1346
1347                 req->rq_phase = RQ_PHASE_INTERPRET;
1348
1349         interpret:
1350                 LASSERT(req->rq_phase == RQ_PHASE_INTERPRET);
1351                 LASSERT(!req->rq_receiving_reply);
1352
1353                 ptlrpc_unregister_reply(req);
1354                 if (req->rq_bulk != NULL)
1355                         ptlrpc_unregister_bulk (req);
1356
1357                 if (req->rq_interpret_reply != NULL) {
1358                         ptlrpc_interpterer_t interpreter =
1359                                 req->rq_interpret_reply;
1360                         req->rq_status = interpreter(NULL, req,
1361                                                      &req->rq_async_args,
1362                                                      req->rq_status);
1363                 }
1364                 req->rq_phase = RQ_PHASE_COMPLETE;
1365
1366                 CDEBUG(D_RPCTRACE, "Completed RPC pname:cluuid:pid:xid:nid:"
1367                        "opc %s:%s:%d:"LPU64":%s:%d\n", cfs_curproc_comm(),
1368                        imp->imp_obd->obd_uuid.uuid,
1369                        lustre_msg_get_status(req->rq_reqmsg), req->rq_xid,
1370                        libcfs_nid2str(imp->imp_connection->c_peer.nid),
1371                        lustre_msg_get_opc(req->rq_reqmsg));
1372
1373                 atomic_dec(&imp->imp_inflight);
1374                 set->set_remaining--;
1375                 cfs_waitq_signal(&imp->imp_recovery_waitq);
1376         }
1377
1378         /* If we hit an error, we want to recover promptly. */
1379         RETURN(set->set_remaining == 0 || force_timer_recalc);
1380 }
1381
1382 /* Return 1 if we should give up, else 0 */
1383 int ptlrpc_expire_one_request(struct ptlrpc_request *req)
1384 {
1385         struct obd_import *imp = req->rq_import;
1386         int rc = 0;
1387         ENTRY;
1388
1389         DEBUG_REQ(D_ERROR|D_NETERROR, req,
1390                   "%s (sent at "CFS_TIME_T", "CFS_DURATION_T"s ago)",
1391                   req->rq_net_err ? "network error" : "timeout",
1392                   req->rq_sent, cfs_time_sub(cfs_time_current_sec(),
1393                   req->rq_sent));
1394
1395         if (imp) {
1396                 LCONSOLE_WARN("Request x"LPU64" sent from %s to NID %s "
1397                               CFS_DURATION_T"s ago has timed out "
1398                               "(limit "CFS_DURATION_T"s).\n", req->rq_xid,
1399                               req->rq_import->imp_obd->obd_name,
1400                               libcfs_nid2str(imp->imp_connection->c_peer.nid),
1401                               cfs_time_sub(cfs_time_current_sec(), req->rq_sent),
1402                               cfs_time_sub(req->rq_deadline, req->rq_sent));
1403         }
1404
1405         if (imp != NULL && obd_debug_peer_on_timeout)
1406                 LNetCtl(IOC_LIBCFS_DEBUG_PEER, &imp->imp_connection->c_peer);
1407
1408         spin_lock(&req->rq_lock);
1409         req->rq_timedout = 1;
1410         spin_unlock(&req->rq_lock);
1411
1412         ptlrpc_unregister_reply (req);
1413
1414         if (obd_dump_on_timeout)
1415                 libcfs_debug_dumplog();
1416
1417         if (req->rq_bulk != NULL)
1418                 ptlrpc_unregister_bulk (req);
1419
1420         if (imp == NULL) {
1421                 DEBUG_REQ(D_HA, req, "NULL import: already cleaned up?");
1422                 RETURN(1);
1423         }
1424
1425         /* The DLM server doesn't want recovery run on its imports. */
1426         if (imp->imp_dlm_fake)
1427                 RETURN(1);
1428
1429         /* If this request is for recovery or other primordial tasks,
1430          * then error it out here. */
1431         if (req->rq_ctx_init || req->rq_ctx_fini ||
1432             req->rq_send_state != LUSTRE_IMP_FULL ||
1433             imp->imp_obd->obd_no_recov) {
1434                 DEBUG_REQ(D_RPCTRACE, req, "err -110, sent_state=%s (now=%s)",
1435                           ptlrpc_import_state_name(req->rq_send_state),
1436                           ptlrpc_import_state_name(imp->imp_state));
1437                 spin_lock(&req->rq_lock);
1438                 req->rq_status = -ETIMEDOUT;
1439                 req->rq_err = 1;
1440                 spin_unlock(&req->rq_lock);
1441                 RETURN(1);
1442         }
1443
1444         /* if a request can't be resent we can't wait for an answer after
1445            the timeout */
1446         if (req->rq_no_resend) {
1447                 DEBUG_REQ(D_RPCTRACE, req, "TIMEOUT-NORESEND:");
1448                 rc = 1;
1449         }
1450
1451         ptlrpc_fail_import(imp, lustre_msg_get_conn_cnt(req->rq_reqmsg));
1452
1453         RETURN(rc);
1454 }
1455
1456 int ptlrpc_expired_set(void *data)
1457 {
1458         struct ptlrpc_request_set *set = data;
1459         struct list_head          *tmp;
1460         time_t                     now = cfs_time_current_sec();
1461         ENTRY;
1462
1463         LASSERT(set != NULL);
1464
1465         /* A timeout expired; see which reqs it applies to... */
1466         list_for_each (tmp, &set->set_requests) {
1467                 struct ptlrpc_request *req =
1468                         list_entry(tmp, struct ptlrpc_request, rq_set_chain);
1469
1470                 /* request in-flight? */
1471                 if (!(((req->rq_phase == RQ_PHASE_RPC) && !req->rq_waiting &&
1472                        !req->rq_resend) ||
1473                       (req->rq_phase == RQ_PHASE_BULK)))
1474                         continue;
1475
1476                 if (req->rq_timedout ||           /* already dealt with */
1477                     req->rq_deadline > now)       /* not expired */
1478                         continue;
1479
1480                 /* deal with this guy */
1481                 ptlrpc_expire_one_request (req);
1482         }
1483
1484         /* When waiting for a whole set, we always to break out of the
1485          * sleep so we can recalculate the timeout, or enable interrupts
1486          * iff everyone's timed out.
1487          */
1488         RETURN(1);
1489 }
1490
1491 void ptlrpc_mark_interrupted(struct ptlrpc_request *req)
1492 {
1493         spin_lock(&req->rq_lock);
1494         req->rq_intr = 1;
1495         spin_unlock(&req->rq_lock);
1496 }
1497
1498 void ptlrpc_interrupted_set(void *data)
1499 {
1500         struct ptlrpc_request_set *set = data;
1501         struct list_head *tmp;
1502
1503         LASSERT(set != NULL);
1504         CERROR("INTERRUPTED SET %p\n", set);
1505
1506         list_for_each(tmp, &set->set_requests) {
1507                 struct ptlrpc_request *req =
1508                         list_entry(tmp, struct ptlrpc_request, rq_set_chain);
1509
1510                 if (req->rq_phase != RQ_PHASE_RPC)
1511                         continue;
1512
1513                 ptlrpc_mark_interrupted(req);
1514         }
1515 }
1516
1517 /* get the smallest timeout in the set; this does NOT set a timeout. */
1518 int ptlrpc_set_next_timeout(struct ptlrpc_request_set *set)
1519 {
1520         struct list_head      *tmp;
1521         time_t                 now = cfs_time_current_sec();
1522         int                    timeout = 0;
1523         struct ptlrpc_request *req;
1524         int                    deadline;
1525         ENTRY;
1526
1527         SIGNAL_MASK_ASSERT(); /* XXX BUG 1511 */
1528
1529         list_for_each(tmp, &set->set_requests) {
1530                 req = list_entry(tmp, struct ptlrpc_request, rq_set_chain);
1531
1532                 /* request in-flight? */
1533                 if (!((req->rq_phase == RQ_PHASE_RPC && !req->rq_waiting) ||
1534                       (req->rq_phase == RQ_PHASE_BULK) ||
1535                       (req->rq_phase == RQ_PHASE_NEW)))
1536                         continue;
1537
1538                 if (req->rq_timedout)   /* already timed out */
1539                         continue;
1540
1541                 if (req->rq_wait_ctx)   /* waiting for ctx */
1542                         continue;
1543
1544                 if (req->rq_phase == RQ_PHASE_NEW)
1545                         deadline = req->rq_sent;
1546                 else
1547                         deadline = req->rq_sent + req->rq_timeout;
1548
1549                 if (deadline <= now)    /* actually expired already */
1550                         timeout = 1;    /* ASAP */
1551                 else if (timeout == 0 || timeout > deadline - now)
1552                         timeout = deadline - now;
1553         }
1554         RETURN(timeout);
1555 }
1556
1557 int ptlrpc_set_wait(struct ptlrpc_request_set *set)
1558 {
1559         struct list_head      *tmp;
1560         struct ptlrpc_request *req;
1561         struct l_wait_info     lwi;
1562         int                    rc, timeout;
1563         ENTRY;
1564
1565         if (list_empty(&set->set_requests))
1566                 RETURN(0);
1567
1568         list_for_each(tmp, &set->set_requests) {
1569                 req = list_entry(tmp, struct ptlrpc_request, rq_set_chain);
1570                 if (req->rq_phase == RQ_PHASE_NEW)
1571                         (void)ptlrpc_send_new_req(req);
1572         }
1573
1574         do {
1575                 timeout = ptlrpc_set_next_timeout(set);
1576
1577                 /* wait until all complete, interrupted, or an in-flight
1578                  * req times out */
1579                 CDEBUG(D_RPCTRACE, "set %p going to sleep for %d seconds\n",
1580                        set, timeout);
1581                 lwi = LWI_TIMEOUT_INTR(cfs_time_seconds(timeout ? timeout : 1),
1582                                        ptlrpc_expired_set,
1583                                        ptlrpc_interrupted_set, set);
1584                 rc = l_wait_event(set->set_waitq,
1585                                   ptlrpc_check_set(NULL, set), &lwi);
1586
1587                 LASSERT(rc == 0 || rc == -EINTR || rc == -ETIMEDOUT);
1588
1589                 /* -EINTR => all requests have been flagged rq_intr so next
1590                  * check completes.
1591                  * -ETIMEOUTD => someone timed out.  When all reqs have
1592                  * timed out, signals are enabled allowing completion with
1593                  * EINTR.
1594                  * I don't really care if we go once more round the loop in
1595                  * the error cases -eeb. */
1596         } while (rc != 0 || set->set_remaining != 0);
1597
1598         LASSERT(set->set_remaining == 0);
1599
1600         rc = 0;
1601         list_for_each(tmp, &set->set_requests) {
1602                 req = list_entry(tmp, struct ptlrpc_request, rq_set_chain);
1603
1604                 LASSERT(req->rq_phase == RQ_PHASE_COMPLETE);
1605                 if (req->rq_status != 0)
1606                         rc = req->rq_status;
1607         }
1608
1609         if (set->set_interpret != NULL) {
1610                 int (*interpreter)(struct ptlrpc_request_set *set,void *,int) =
1611                         set->set_interpret;
1612                 rc = interpreter (set, set->set_arg, rc);
1613         } else {
1614                 struct ptlrpc_set_cbdata *cbdata, *n;
1615                 int err;
1616
1617                 list_for_each_entry_safe(cbdata, n,
1618                                          &set->set_cblist, psc_item) {
1619                         list_del_init(&cbdata->psc_item);
1620                         err = cbdata->psc_interpret(set, cbdata->psc_data, rc);
1621                         if (err && !rc)
1622                                 rc = err;
1623                         OBD_FREE_PTR(cbdata);
1624                 }
1625         }
1626
1627         RETURN(rc);
1628 }
1629
1630 static void __ptlrpc_free_req(struct ptlrpc_request *request, int locked)
1631 {
1632         ENTRY;
1633         if (request == NULL) {
1634                 EXIT;
1635                 return;
1636         }
1637
1638         LASSERTF(!request->rq_receiving_reply, "req %p\n", request);
1639         LASSERTF(request->rq_rqbd == NULL, "req %p\n",request);/* client-side */
1640         LASSERTF(list_empty(&request->rq_list), "req %p\n", request);
1641         LASSERTF(list_empty(&request->rq_set_chain), "req %p\n", request);
1642         LASSERTF(!request->rq_replay, "req %p\n", request);
1643         LASSERT(request->rq_cli_ctx);
1644
1645         req_capsule_fini(&request->rq_pill);
1646
1647         /* We must take it off the imp_replay_list first.  Otherwise, we'll set
1648          * request->rq_reqmsg to NULL while osc_close is dereferencing it. */
1649         if (request->rq_import != NULL) {
1650                 if (!locked)
1651                         spin_lock(&request->rq_import->imp_lock);
1652                 list_del_init(&request->rq_mod_list);
1653                 list_del_init(&request->rq_replay_list);
1654                 if (!locked)
1655                         spin_unlock(&request->rq_import->imp_lock);
1656         }
1657         LASSERTF(list_empty(&request->rq_replay_list), "req %p\n", request);
1658
1659         if (atomic_read(&request->rq_refcount) != 0) {
1660                 DEBUG_REQ(D_ERROR, request,
1661                           "freeing request with nonzero refcount");
1662                 LBUG();
1663         }
1664
1665         if (request->rq_repbuf != NULL)
1666                 sptlrpc_cli_free_repbuf(request);
1667         if (request->rq_export != NULL) {
1668                 class_export_put(request->rq_export);
1669                 request->rq_export = NULL;
1670         }
1671         if (request->rq_import != NULL) {
1672                 class_import_put(request->rq_import);
1673                 request->rq_import = NULL;
1674         }
1675         if (request->rq_bulk != NULL)
1676                 ptlrpc_free_bulk(request->rq_bulk);
1677
1678         if (request->rq_reqbuf != NULL || request->rq_clrbuf != NULL)
1679                 sptlrpc_cli_free_reqbuf(request);
1680
1681         sptlrpc_req_put_ctx(request, !locked);
1682
1683         if (request->rq_pool)
1684                 __ptlrpc_free_req_to_pool(request);
1685         else
1686                 OBD_FREE(request, sizeof(*request));
1687         EXIT;
1688 }
1689
1690 static int __ptlrpc_req_finished(struct ptlrpc_request *request, int locked);
1691 void ptlrpc_req_finished_with_imp_lock(struct ptlrpc_request *request)
1692 {
1693         LASSERT_SPIN_LOCKED(&request->rq_import->imp_lock);
1694         (void)__ptlrpc_req_finished(request, 1);
1695 }
1696
1697 static int __ptlrpc_req_finished(struct ptlrpc_request *request, int locked)
1698 {
1699         ENTRY;
1700         if (request == NULL)
1701                 RETURN(1);
1702
1703         if (request == LP_POISON ||
1704             request->rq_reqmsg == LP_POISON) {
1705                 CERROR("dereferencing freed request (bug 575)\n");
1706                 LBUG();
1707                 RETURN(1);
1708         }
1709
1710         DEBUG_REQ(D_INFO, request, "refcount now %u",
1711                   atomic_read(&request->rq_refcount) - 1);
1712
1713         if (atomic_dec_and_test(&request->rq_refcount)) {
1714                 __ptlrpc_free_req(request, locked);
1715                 RETURN(1);
1716         }
1717
1718         RETURN(0);
1719 }
1720
1721 void ptlrpc_req_finished(struct ptlrpc_request *request)
1722 {
1723         __ptlrpc_req_finished(request, 0);
1724 }
1725
1726 __u64 ptlrpc_req_xid(struct ptlrpc_request *request)
1727 {
1728         return request->rq_xid;
1729 }
1730 EXPORT_SYMBOL(ptlrpc_req_xid);
1731
1732 /* Disengage the client's reply buffer from the network
1733  * NB does _NOT_ unregister any client-side bulk.
1734  * IDEMPOTENT, but _not_ safe against concurrent callers.
1735  * The request owner (i.e. the thread doing the I/O) must call...
1736  */
1737 void ptlrpc_unregister_reply (struct ptlrpc_request *request)
1738 {
1739         int                rc;
1740         cfs_waitq_t       *wq;
1741         struct l_wait_info lwi;
1742
1743         LASSERT(!in_interrupt ());             /* might sleep */
1744         if (!ptlrpc_client_recv_or_unlink(request))
1745                 /* Nothing left to do */
1746                 return;
1747
1748         LNetMDUnlink (request->rq_reply_md_h);
1749
1750         /* We have to l_wait_event() whatever the result, to give liblustre
1751          * a chance to run reply_in_callback(), and to make sure we've
1752          * unlinked before returning a req to the pool */
1753
1754         if (request->rq_set != NULL)
1755                 wq = &request->rq_set->set_waitq;
1756         else
1757                 wq = &request->rq_reply_waitq;
1758
1759         for (;;) {
1760                 /* Network access will complete in finite time but the HUGE
1761                  * timeout lets us CWARN for visibility of sluggish NALs */
1762                 lwi = LWI_TIMEOUT(cfs_time_seconds(LONG_UNLINK), NULL, NULL);
1763                 rc = l_wait_event (*wq, !ptlrpc_client_recv_or_unlink(request),
1764                                    &lwi);
1765                 if (rc == 0)
1766                         return;
1767
1768                 LASSERT (rc == -ETIMEDOUT);
1769                 DEBUG_REQ(D_WARNING, request, "Unexpectedly long timeout "
1770                           "rvcng=%d unlnk=%d", request->rq_receiving_reply,
1771                           request->rq_must_unlink);
1772         }
1773 }
1774
1775 /* caller must hold imp->imp_lock */
1776 void ptlrpc_free_committed(struct obd_import *imp)
1777 {
1778         struct list_head *tmp, *saved;
1779         struct ptlrpc_request *req;
1780         struct ptlrpc_request *last_req = NULL; /* temporary fire escape */
1781         ENTRY;
1782
1783         LASSERT(imp != NULL);
1784
1785         LASSERT_SPIN_LOCKED(&imp->imp_lock);
1786
1787
1788         if (imp->imp_peer_committed_transno == imp->imp_last_transno_checked &&
1789             imp->imp_generation == imp->imp_last_generation_checked) {
1790                 CDEBUG(D_RPCTRACE, "%s: skip recheck: last_committed "LPU64"\n",
1791                        imp->imp_obd->obd_name, imp->imp_peer_committed_transno);
1792                 EXIT;
1793                 return;
1794         }
1795
1796         CDEBUG(D_RPCTRACE, "%s: committing for last_committed "LPU64" gen %d\n",
1797                imp->imp_obd->obd_name, imp->imp_peer_committed_transno,
1798                imp->imp_generation);
1799         imp->imp_last_transno_checked = imp->imp_peer_committed_transno;
1800         imp->imp_last_generation_checked = imp->imp_generation;
1801
1802         list_for_each_safe(tmp, saved, &imp->imp_replay_list) {
1803                 req = list_entry(tmp, struct ptlrpc_request, rq_replay_list);
1804
1805                 /* XXX ok to remove when 1357 resolved - rread 05/29/03  */
1806                 LASSERT(req != last_req);
1807                 last_req = req;
1808
1809                 if (req->rq_import_generation < imp->imp_generation) {
1810                         DEBUG_REQ(D_RPCTRACE, req, "free request with old gen");
1811                         GOTO(free_req, 0);
1812                 }
1813
1814                 if (req->rq_replay) {
1815                         DEBUG_REQ(D_RPCTRACE, req, "keeping (FL_REPLAY)");
1816                         continue;
1817                 }
1818
1819                 /* not yet committed */
1820                 if (req->rq_transno > imp->imp_peer_committed_transno) {
1821                         DEBUG_REQ(D_RPCTRACE, req, "stopping search");
1822                         break;
1823                 }
1824
1825                 DEBUG_REQ(D_RPCTRACE, req, "commit (last_committed "LPU64")",
1826                           imp->imp_peer_committed_transno);
1827 free_req:
1828                 spin_lock(&req->rq_lock);
1829                 req->rq_replay = 0;
1830                 spin_unlock(&req->rq_lock);
1831                 if (req->rq_commit_cb != NULL)
1832                         req->rq_commit_cb(req);
1833                 list_del_init(&req->rq_replay_list);
1834                 __ptlrpc_req_finished(req, 1);
1835         }
1836
1837         EXIT;
1838         return;
1839 }
1840
1841 void ptlrpc_cleanup_client(struct obd_import *imp)
1842 {
1843         ENTRY;
1844         EXIT;
1845         return;
1846 }
1847
1848 void ptlrpc_resend_req(struct ptlrpc_request *req)
1849 {
1850         DEBUG_REQ(D_HA, req, "going to resend");
1851         lustre_msg_set_handle(req->rq_reqmsg, &(struct lustre_handle){ 0 });
1852         req->rq_status = -EAGAIN;
1853
1854         spin_lock(&req->rq_lock);
1855         req->rq_resend = 1;
1856         req->rq_net_err = 0;
1857         req->rq_timedout = 0;
1858         if (req->rq_bulk) {
1859                 __u64 old_xid = req->rq_xid;
1860
1861                 /* ensure previous bulk fails */
1862                 req->rq_xid = ptlrpc_next_xid();
1863                 CDEBUG(D_HA, "resend bulk old x"LPU64" new x"LPU64"\n",
1864                        old_xid, req->rq_xid);
1865         }
1866         ptlrpc_wake_client_req(req);
1867         spin_unlock(&req->rq_lock);
1868 }
1869
1870 /* XXX: this function and rq_status are currently unused */
1871 void ptlrpc_restart_req(struct ptlrpc_request *req)
1872 {
1873         DEBUG_REQ(D_HA, req, "restarting (possibly-)completed request");
1874         req->rq_status = -ERESTARTSYS;
1875
1876         spin_lock(&req->rq_lock);
1877         req->rq_restart = 1;
1878         req->rq_timedout = 0;
1879         ptlrpc_wake_client_req(req);
1880         spin_unlock(&req->rq_lock);
1881 }
1882
1883 static int expired_request(void *data)
1884 {
1885         struct ptlrpc_request *req = data;
1886         ENTRY;
1887
1888         /* some failure can suspend regular timeouts */
1889         if (ptlrpc_check_suspend())
1890                 RETURN(1);
1891
1892         /* deadline may have changed with an early reply */
1893         if (req->rq_deadline > cfs_time_current_sec())
1894                 RETURN(1);
1895
1896         RETURN(ptlrpc_expire_one_request(req));
1897 }
1898
1899 static void interrupted_request(void *data)
1900 {
1901         struct ptlrpc_request *req = data;
1902         DEBUG_REQ(D_HA, req, "request interrupted");
1903         spin_lock(&req->rq_lock);
1904         req->rq_intr = 1;
1905         spin_unlock(&req->rq_lock);
1906 }
1907
1908 struct ptlrpc_request *ptlrpc_request_addref(struct ptlrpc_request *req)
1909 {
1910         ENTRY;
1911         atomic_inc(&req->rq_refcount);
1912         RETURN(req);
1913 }
1914
1915 void ptlrpc_retain_replayable_request(struct ptlrpc_request *req,
1916                                       struct obd_import *imp)
1917 {
1918         struct list_head *tmp;
1919
1920         LASSERT_SPIN_LOCKED(&imp->imp_lock);
1921
1922         /* clear this for new requests that were resent as well
1923            as resent replayed requests. */
1924         lustre_msg_clear_flags(req->rq_reqmsg, MSG_RESENT);
1925
1926         /* don't re-add requests that have been replayed */
1927         if (!list_empty(&req->rq_replay_list))
1928                 return;
1929
1930         lustre_msg_add_flags(req->rq_reqmsg, MSG_REPLAY);
1931
1932         LASSERT(imp->imp_replayable);
1933         /* Balanced in ptlrpc_free_committed, usually. */
1934         ptlrpc_request_addref(req);
1935         list_for_each_prev(tmp, &imp->imp_replay_list) {
1936                 struct ptlrpc_request *iter =
1937                         list_entry(tmp, struct ptlrpc_request, rq_replay_list);
1938
1939                 /* We may have duplicate transnos if we create and then
1940                  * open a file, or for closes retained if to match creating
1941                  * opens, so use req->rq_xid as a secondary key.
1942                  * (See bugs 684, 685, and 428.)
1943                  * XXX no longer needed, but all opens need transnos!
1944                  */
1945                 if (iter->rq_transno > req->rq_transno)
1946                         continue;
1947
1948                 if (iter->rq_transno == req->rq_transno) {
1949                         LASSERT(iter->rq_xid != req->rq_xid);
1950                         if (iter->rq_xid > req->rq_xid)
1951                                 continue;
1952                 }
1953
1954                 list_add(&req->rq_replay_list, &iter->rq_replay_list);
1955                 return;
1956         }
1957
1958         list_add_tail(&req->rq_replay_list, &imp->imp_replay_list);
1959 }
1960
1961 int ptlrpc_queue_wait(struct ptlrpc_request *req)
1962 {
1963         int rc = 0;
1964         int brc;
1965         struct l_wait_info lwi;
1966         struct obd_import *imp = req->rq_import;
1967         cfs_duration_t timeout = CFS_TICK;
1968         long timeoutl;
1969         ENTRY;
1970
1971         LASSERT(req->rq_set == NULL);
1972         LASSERT(!req->rq_receiving_reply);
1973         atomic_inc(&imp->imp_inflight);
1974
1975         /* for distributed debugging */
1976         lustre_msg_set_status(req->rq_reqmsg, cfs_curproc_pid());
1977         LASSERT(imp->imp_obd != NULL);
1978         CDEBUG(D_RPCTRACE, "Sending RPC pname:cluuid:pid:xid:nid:opc "
1979                "%s:%s:%d:"LPU64":%s:%d\n", cfs_curproc_comm(),
1980                imp->imp_obd->obd_uuid.uuid,
1981                lustre_msg_get_status(req->rq_reqmsg), req->rq_xid,
1982                libcfs_nid2str(imp->imp_connection->c_peer.nid),
1983                lustre_msg_get_opc(req->rq_reqmsg));
1984
1985         /* Mark phase here for a little debug help */
1986         req->rq_phase = RQ_PHASE_RPC;
1987
1988         spin_lock(&imp->imp_lock);
1989         req->rq_import_generation = imp->imp_generation;
1990 restart:
1991         if (ptlrpc_import_delay_req(imp, req, &rc)) {
1992                 list_del(&req->rq_list);
1993
1994                 list_add_tail(&req->rq_list, &imp->imp_delayed_list);
1995                 spin_unlock(&imp->imp_lock);
1996
1997                 DEBUG_REQ(D_HA, req, "\"%s\" waiting for recovery: (%s != %s)",
1998                           cfs_curproc_comm(),
1999                           ptlrpc_import_state_name(req->rq_send_state),
2000                           ptlrpc_import_state_name(imp->imp_state));
2001                 lwi = LWI_INTR(interrupted_request, req);
2002                 rc = l_wait_event(req->rq_reply_waitq,
2003                                   (req->rq_send_state == imp->imp_state ||
2004                                    req->rq_err || req->rq_intr),
2005                                   &lwi);
2006                 DEBUG_REQ(D_HA, req, "\"%s\" awake: (%s == %s or %d/%d == 1)",
2007                           cfs_curproc_comm(),
2008                           ptlrpc_import_state_name(imp->imp_state),
2009                           ptlrpc_import_state_name(req->rq_send_state),
2010                           req->rq_err, req->rq_intr);
2011
2012                 spin_lock(&imp->imp_lock);
2013                 list_del_init(&req->rq_list);
2014
2015                 if (req->rq_err) {
2016                         /* rq_status was set locally */
2017                         rc = -EIO;
2018                 }
2019                 else if (req->rq_intr) {
2020                         rc = -EINTR;
2021                 }
2022                 else if (req->rq_no_resend) {
2023                         spin_unlock(&imp->imp_lock);
2024                         GOTO(out, rc = -ETIMEDOUT);
2025                 }
2026                 else {
2027                         GOTO(restart, rc);
2028                 }
2029         }
2030
2031         if (rc != 0) {
2032                 list_del_init(&req->rq_list);
2033                 spin_unlock(&imp->imp_lock);
2034                 req->rq_status = rc; // XXX this ok?
2035                 GOTO(out, rc);
2036         }
2037
2038         if (req->rq_resend) {
2039                 lustre_msg_add_flags(req->rq_reqmsg, MSG_RESENT);
2040
2041                 if (req->rq_bulk != NULL) {
2042                         ptlrpc_unregister_bulk (req);
2043
2044                         /* bulk requests are supposed to be
2045                          * idempotent, so we are free to bump the xid
2046                          * here, which we need to do before
2047                          * registering the bulk again (bug 6371).
2048                          * print the old xid first for sanity.
2049                          */
2050                         DEBUG_REQ(D_HA, req, "bumping xid for bulk: ");
2051                         req->rq_xid = ptlrpc_next_xid();
2052                 }
2053
2054                 DEBUG_REQ(D_HA, req, "resending: ");
2055         }
2056
2057         /* XXX this is the same as ptlrpc_set_wait */
2058         LASSERT(list_empty(&req->rq_list));
2059         list_add_tail(&req->rq_list, &imp->imp_sending_list);
2060         spin_unlock(&imp->imp_lock);
2061
2062         rc = sptlrpc_req_refresh_ctx(req, 0);
2063         if (rc) {
2064                 if (req->rq_err) {
2065                         /* we got fatal ctx refresh error, directly jump out
2066                          * thus we can pass back the actual error code.
2067                          */
2068                         spin_lock(&imp->imp_lock);
2069                         list_del_init(&req->rq_list);
2070                         spin_unlock(&imp->imp_lock);
2071
2072                         CERROR("Failed to refresh ctx of req %p: %d\n", req, rc);
2073                         GOTO(out, rc);
2074                 }
2075                 /* simulating we got error during send rpc */
2076                 goto after_send;
2077         }
2078
2079         rc = ptl_send_rpc(req, 0);
2080         if (rc)
2081                 DEBUG_REQ(D_HA, req, "send failed (%d); recovering", rc);
2082
2083 repeat:
2084         timeoutl = req->rq_deadline - cfs_time_current_sec();
2085         timeout = (timeoutl <= 0 || rc) ? CFS_TICK :
2086                 cfs_time_seconds(timeoutl);
2087         DEBUG_REQ(D_NET, req,
2088                   "-- sleeping for "CFS_DURATION_T" ticks", timeout);
2089         lwi = LWI_TIMEOUT_INTR(timeout, expired_request, interrupted_request,
2090                                req);
2091         brc = l_wait_event(req->rq_reply_waitq, ptlrpc_check_reply(req), &lwi);
2092         if (brc == -ETIMEDOUT && ((req->rq_deadline > cfs_time_current_sec()) ||
2093                                  ptlrpc_check_and_wait_suspend(req)))
2094                 goto repeat;
2095
2096 after_send:
2097         CDEBUG(D_RPCTRACE, "Completed RPC pname:cluuid:pid:xid:nid:opc "
2098                "%s:%s:%d:"LPU64":%s:%d\n", cfs_curproc_comm(),
2099                imp->imp_obd->obd_uuid.uuid,
2100                lustre_msg_get_status(req->rq_reqmsg), req->rq_xid,
2101                libcfs_nid2str(imp->imp_connection->c_peer.nid),
2102                lustre_msg_get_opc(req->rq_reqmsg));
2103
2104         spin_lock(&imp->imp_lock);
2105         list_del_init(&req->rq_list);
2106         spin_unlock(&imp->imp_lock);
2107
2108         /* If the reply was received normally, this just grabs the spinlock
2109          * (ensuring the reply callback has returned), sees that
2110          * req->rq_receiving_reply is clear and returns. */
2111         ptlrpc_unregister_reply (req);
2112
2113
2114         if (req->rq_err) {
2115                 DEBUG_REQ(D_RPCTRACE, req, "err rc=%d status=%d",
2116                           rc, req->rq_status);
2117                 GOTO(out, rc = rc ? rc : -EIO);
2118         }
2119
2120         if (req->rq_intr) {
2121                 /* Should only be interrupted if we timed out. */
2122                 if (!req->rq_timedout)
2123                         DEBUG_REQ(D_ERROR, req,
2124                                   "rq_intr set but rq_timedout not");
2125                 GOTO(out, rc = -EINTR);
2126         }
2127
2128         /* Resend if we need to */
2129         if (req->rq_resend) {
2130                 /* ...unless we were specifically told otherwise. */
2131                 if (req->rq_no_resend)
2132                         GOTO(out, rc = -ETIMEDOUT);
2133                 spin_lock(&imp->imp_lock);
2134                 goto restart;
2135         }
2136
2137         if (req->rq_timedout) {                 /* non-recoverable timeout */
2138                 GOTO(out, rc = -ETIMEDOUT);
2139         }
2140
2141         if (!req->rq_replied) {
2142                 /* How can this be? -eeb */
2143                 DEBUG_REQ(D_ERROR, req, "!rq_replied: ");
2144                 LBUG();
2145                 GOTO(out, rc = req->rq_status);
2146         }
2147
2148         rc = after_reply(req);
2149         /* NB may return +ve success rc */
2150         if (req->rq_resend) {
2151                 spin_lock(&imp->imp_lock);
2152                 goto restart;
2153         }
2154
2155  out:
2156         if (req->rq_bulk != NULL) {
2157                 if (rc >= 0) {
2158                         /* success so far.  Note that anything going wrong
2159                          * with bulk now, is EXTREMELY strange, since the
2160                          * server must have believed that the bulk
2161                          * tranferred OK before she replied with success to
2162                          * me. */
2163                         lwi = LWI_TIMEOUT(timeout, NULL, NULL);
2164                         brc = l_wait_event(req->rq_reply_waitq,
2165                                            !ptlrpc_bulk_active(req->rq_bulk),
2166                                            &lwi);
2167                         LASSERT(brc == 0 || brc == -ETIMEDOUT);
2168                         if (brc != 0) {
2169                                 LASSERT(brc == -ETIMEDOUT);
2170                                 DEBUG_REQ(D_ERROR, req, "bulk timed out");
2171                                 rc = brc;
2172                         } else if (!req->rq_bulk->bd_success) {
2173                                 DEBUG_REQ(D_ERROR, req, "bulk transfer failed");
2174                                 rc = -EIO;
2175                         }
2176                 }
2177                 if (rc < 0)
2178                         ptlrpc_unregister_bulk (req);
2179         }
2180
2181         LASSERT(!req->rq_receiving_reply);
2182         req->rq_phase = RQ_PHASE_INTERPRET;
2183
2184         atomic_dec(&imp->imp_inflight);
2185         cfs_waitq_signal(&imp->imp_recovery_waitq);
2186         RETURN(rc);
2187 }
2188
2189 struct ptlrpc_replay_async_args {
2190         int praa_old_state;
2191         int praa_old_status;
2192 };
2193
2194 static int ptlrpc_replay_interpret(const struct lu_env *env,
2195                                    struct ptlrpc_request *req,
2196                                     void * data, int rc)
2197 {
2198         struct ptlrpc_replay_async_args *aa = data;
2199         struct obd_import *imp = req->rq_import;
2200
2201         ENTRY;
2202         atomic_dec(&imp->imp_replay_inflight);
2203
2204         if (!req->rq_replied) {
2205                 CERROR("request replay timed out, restarting recovery\n");
2206                 GOTO(out, rc = -ETIMEDOUT);
2207         }
2208
2209         if (lustre_msg_get_type(req->rq_repmsg) == PTL_RPC_MSG_ERR &&
2210             (lustre_msg_get_status(req->rq_repmsg) == -ENOTCONN ||
2211              lustre_msg_get_status(req->rq_repmsg) == -ENODEV))
2212                 GOTO(out, rc = lustre_msg_get_status(req->rq_repmsg));
2213
2214         /* The transno had better not change over replay. */
2215         LASSERT(lustre_msg_get_transno(req->rq_reqmsg) ==
2216                 lustre_msg_get_transno(req->rq_repmsg));
2217
2218         DEBUG_REQ(D_HA, req, "got rep");
2219
2220         /* let the callback do fixups, possibly including in the request */
2221         if (req->rq_replay_cb)
2222                 req->rq_replay_cb(req);
2223
2224         if (req->rq_replied &&
2225             lustre_msg_get_status(req->rq_repmsg) != aa->praa_old_status) {
2226                 DEBUG_REQ(D_ERROR, req, "status %d, old was %d",
2227                           lustre_msg_get_status(req->rq_repmsg),
2228                           aa->praa_old_status);
2229         } else {
2230                 /* Put it back for re-replay. */
2231                 lustre_msg_set_status(req->rq_repmsg, aa->praa_old_status);
2232         }
2233
2234         /*
2235          * Errors while replay can set transno to 0, but
2236          * imp_last_replay_transno shouldn't be set to 0 anyway
2237          */
2238         if (req->rq_transno > 0) {
2239                 spin_lock(&imp->imp_lock);
2240                 LASSERT(req->rq_transno <= imp->imp_last_replay_transno);
2241                 imp->imp_last_replay_transno = req->rq_transno;
2242                 spin_unlock(&imp->imp_lock);
2243         } else
2244                 CERROR("Transno is 0 during replay!\n");
2245         /* continue with recovery */
2246         rc = ptlrpc_import_recovery_state_machine(imp);
2247  out:
2248         req->rq_send_state = aa->praa_old_state;
2249
2250         if (rc != 0)
2251                 /* this replay failed, so restart recovery */
2252                 ptlrpc_connect_import(imp, NULL);
2253
2254         RETURN(rc);
2255 }
2256
2257 int ptlrpc_replay_req(struct ptlrpc_request *req)
2258 {
2259         struct ptlrpc_replay_async_args *aa;
2260         ENTRY;
2261
2262         LASSERT(req->rq_import->imp_state == LUSTRE_IMP_REPLAY);
2263         /* Not handling automatic bulk replay yet (or ever?) */
2264         LASSERT(req->rq_bulk == NULL);
2265
2266         LASSERT (sizeof (*aa) <= sizeof (req->rq_async_args));
2267         aa = ptlrpc_req_async_args(req);
2268         memset(aa, 0, sizeof *aa);
2269
2270         /* Prepare request to be resent with ptlrpcd */
2271         aa->praa_old_state = req->rq_send_state;
2272         req->rq_send_state = LUSTRE_IMP_REPLAY;
2273         req->rq_phase = RQ_PHASE_NEW;
2274         if (req->rq_repmsg)
2275                 aa->praa_old_status = lustre_msg_get_status(req->rq_repmsg);
2276         req->rq_status = 0;
2277         req->rq_interpret_reply = ptlrpc_replay_interpret;
2278         /* Readjust the timeout for current conditions */
2279         ptlrpc_at_set_req_timeout(req);
2280
2281         DEBUG_REQ(D_HA, req, "REPLAY");
2282
2283         atomic_inc(&req->rq_import->imp_replay_inflight);
2284         ptlrpc_request_addref(req); /* ptlrpcd needs a ref */
2285
2286         ptlrpcd_add_req(req);
2287         RETURN(0);
2288 }
2289
2290 void ptlrpc_abort_inflight(struct obd_import *imp)
2291 {
2292         struct list_head *tmp, *n;
2293         ENTRY;
2294
2295         /* Make sure that no new requests get processed for this import.
2296          * ptlrpc_{queue,set}_wait must (and does) hold imp_lock while testing
2297          * this flag and then putting requests on sending_list or delayed_list.
2298          */
2299         spin_lock(&imp->imp_lock);
2300
2301         /* XXX locking?  Maybe we should remove each request with the list
2302          * locked?  Also, how do we know if the requests on the list are
2303          * being freed at this time?
2304          */
2305         list_for_each_safe(tmp, n, &imp->imp_sending_list) {
2306                 struct ptlrpc_request *req =
2307                         list_entry(tmp, struct ptlrpc_request, rq_list);
2308
2309                 DEBUG_REQ(D_RPCTRACE, req, "inflight");
2310
2311                 spin_lock (&req->rq_lock);
2312                 if (req->rq_import_generation < imp->imp_generation) {
2313                         req->rq_err = 1;
2314                         req->rq_status = -EINTR;
2315                         ptlrpc_wake_client_req(req);
2316                 }
2317                 spin_unlock (&req->rq_lock);
2318         }
2319
2320         list_for_each_safe(tmp, n, &imp->imp_delayed_list) {
2321                 struct ptlrpc_request *req =
2322                         list_entry(tmp, struct ptlrpc_request, rq_list);
2323
2324                 DEBUG_REQ(D_RPCTRACE, req, "aborting waiting req");
2325
2326                 spin_lock (&req->rq_lock);
2327                 if (req->rq_import_generation < imp->imp_generation) {
2328                         req->rq_err = 1;
2329                         req->rq_status = -EINTR;
2330                         ptlrpc_wake_client_req(req);
2331                 }
2332                 spin_unlock (&req->rq_lock);
2333         }
2334
2335         /* Last chance to free reqs left on the replay list, but we
2336          * will still leak reqs that haven't committed.  */
2337         if (imp->imp_replayable)
2338                 ptlrpc_free_committed(imp);
2339
2340         spin_unlock(&imp->imp_lock);
2341
2342         EXIT;
2343 }
2344
2345 void ptlrpc_abort_set(struct ptlrpc_request_set *set)
2346 {
2347         struct list_head *tmp, *n;
2348
2349         LASSERT(set != NULL);
2350
2351         list_for_each_safe(tmp, n, &set->set_requests) {
2352                 struct ptlrpc_request *req =
2353                         list_entry(tmp, struct ptlrpc_request, rq_set_chain);
2354
2355                 spin_lock (&req->rq_lock);
2356                 if (req->rq_phase != RQ_PHASE_RPC) {
2357                         spin_unlock (&req->rq_lock);
2358                         continue;
2359                 }
2360
2361                 req->rq_err = 1;
2362                 req->rq_status = -EINTR;
2363                 ptlrpc_wake_client_req(req);
2364                 spin_unlock (&req->rq_lock);
2365         }
2366 }
2367
2368 static __u64 ptlrpc_last_xid;
2369 static spinlock_t ptlrpc_last_xid_lock;
2370
2371 /* Initialize the XID for the node.  This is common among all requests on
2372  * this node, and only requires the property that it is monotonically
2373  * increasing.  It does not need to be sequential.  Since this is also used
2374  * as the RDMA match bits, it is important that a single client NOT have
2375  * the same match bits for two different in-flight requests, hence we do
2376  * NOT want to have an XID per target or similar.
2377  *
2378  * To avoid an unlikely collision between match bits after a client reboot
2379  * (which would cause old to be delivered into the wrong buffer) we initialize
2380  * the XID based on the current time, assuming a maximum RPC rate of 1M RPC/s.
2381  * If the time is clearly incorrect, we instead use a 62-bit random number.
2382  * In the worst case the random number will overflow 1M RPCs per second in
2383  * 9133 years, or permutations thereof.
2384  */
2385 #define YEAR_2004 (1ULL << 30)
2386 void ptlrpc_init_xid(void)
2387 {
2388         time_t now = cfs_time_current_sec();
2389
2390         spin_lock_init(&ptlrpc_last_xid_lock);
2391         if (now < YEAR_2004) {
2392                 ll_get_random_bytes(&ptlrpc_last_xid, sizeof(ptlrpc_last_xid));
2393                 ptlrpc_last_xid >>= 2;
2394                 ptlrpc_last_xid |= (1ULL << 61);
2395         } else {
2396                 ptlrpc_last_xid = (now << 20);
2397         }
2398 }
2399
2400 __u64 ptlrpc_next_xid(void)
2401 {
2402         __u64 tmp;
2403         spin_lock(&ptlrpc_last_xid_lock);
2404         tmp = ++ptlrpc_last_xid;
2405         spin_unlock(&ptlrpc_last_xid_lock);
2406         return tmp;
2407 }
2408
2409 __u64 ptlrpc_sample_next_xid(void)
2410 {
2411 #if BITS_PER_LONG == 32
2412         /* need to avoid possible word tearing on 32-bit systems */
2413         __u64 tmp;
2414         spin_lock(&ptlrpc_last_xid_lock);
2415         tmp = ptlrpc_last_xid + 1;
2416         spin_unlock(&ptlrpc_last_xid_lock);
2417         return tmp;
2418 #else
2419         /* No need to lock, since returned value is racy anyways */
2420         return ptlrpc_last_xid + 1;
2421 #endif
2422 }
2423 EXPORT_SYMBOL(ptlrpc_sample_next_xid);