lustre/ptlrpc/client.c

   1 /* -*- mode: c; c-basic-offset: 8; indent-tabs-mode: nil; -*-
   2  * vim:expandtab:shiftwidth=8:tabstop=8:
   3  *
   4  *  Copyright (c) 2002, 2003 Cluster File Systems, Inc.
   5  *
   6  *   This file is part of Lustre, http://www.lustre.org.
   7  *
   8  *   Lustre is free software; you can redistribute it and/or
   9  *   modify it under the terms of version 2 of the GNU General Public
  10  *   License as published by the Free Software Foundation.
  11  *
  12  *   Lustre is distributed in the hope that it will be useful,
  13  *   but WITHOUT ANY WARRANTY; without even the implied warranty of
  14  *   MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
  15  *   GNU General Public License for more details.
  16  *
  17  *   You should have received a copy of the GNU General Public License
  18  *   along with Lustre; if not, write to the Free Software
  19  *   Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA.
  20  *
  21  */
  22
  23 #define DEBUG_SUBSYSTEM S_RPC
  24 #ifndef __KERNEL__
  25 #include <errno.h>
  26 #include <signal.h>
  27 #include <liblustre.h>
  28 #endif
  29
  30 #include <linux/obd_support.h>
  31 #include <linux/obd_class.h>
  32 #include <linux/lustre_lib.h>
  33 #include <linux/lustre_ha.h>
  34 #include <linux/lustre_import.h>
  35 #include <linux/lustre_sec.h>
  36
  37 #include "ptlrpc_internal.h"
  38
  39 void ptlrpc_init_client(int req_portal, int rep_portal, char *name,
  40                         struct ptlrpc_client *cl)
  41 {
  42         cl->cli_request_portal = req_portal;
  43         cl->cli_reply_portal   = rep_portal;
  44         cl->cli_name           = name;
  45 }
  46
  47 struct ptlrpc_connection *ptlrpc_uuid_to_connection(struct obd_uuid *uuid)
  48 {
  49         struct ptlrpc_connection *c;
  50         struct ptlrpc_peer peer;
  51         int err;
  52
  53         err = ptlrpc_uuid_to_peer(uuid, &peer);
  54         if (err != 0) {
  55                 CERROR("cannot find peer %s!\n", uuid->uuid);
  56                 return NULL;
  57         }
  58
  59         c = ptlrpc_get_connection(&peer, uuid);
  60         if (c) {
  61                 memcpy(c->c_remote_uuid.uuid,
  62                        uuid->uuid, sizeof(c->c_remote_uuid.uuid));
  63         }
  64
  65         CDEBUG(D_INFO, "%s -> %p\n", uuid->uuid, c);
  66
  67         return c;
  68 }
  69
  70 void ptlrpc_readdress_connection(struct ptlrpc_connection *conn,
  71                                  struct obd_uuid *uuid)
  72 {
  73         struct ptlrpc_peer peer;
  74         int err;
  75
  76         err = ptlrpc_uuid_to_peer(uuid, &peer);
  77         if (err != 0) {
  78                 CERROR("cannot find peer %s!\n", uuid->uuid);
  79                 return;
  80         }
  81
  82         memcpy(&conn->c_peer, &peer, sizeof (peer));
  83         return;
  84 }
  85
  86 static inline struct ptlrpc_bulk_desc *new_bulk(int npages, int type, int portal)
  87 {
  88         struct ptlrpc_bulk_desc *desc;
  89
  90         OBD_ALLOC(desc, offsetof (struct ptlrpc_bulk_desc, bd_iov[npages]));
  91         if (!desc)
  92                 return NULL;
  93
  94         spin_lock_init(&desc->bd_lock);
  95         init_waitqueue_head(&desc->bd_waitq);
  96         desc->bd_max_iov = npages;
  97         desc->bd_iov_count = 0;
  98         desc->bd_md_h = PTL_INVALID_HANDLE;
  99         desc->bd_portal = portal;
 100         desc->bd_type = type;
 101
 102         return desc;
 103 }
 104
 105 struct ptlrpc_bulk_desc *ptlrpc_prep_bulk_imp (struct ptlrpc_request *req,
 106                                                int npages, int type, int portal)
 107 {
 108         struct obd_import *imp = req->rq_import;
 109         struct ptlrpc_bulk_desc *desc;
 110
 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         LASSERT(type == BULK_PUT_SOURCE || type == BULK_GET_SINK);
 136
 137         desc = new_bulk(npages, type, portal);
 138         if (desc == NULL)
 139                 RETURN(NULL);
 140
 141         desc->bd_export = class_export_get(exp);
 142         desc->bd_req = req;
 143
 144         desc->bd_cbid.cbid_fn  = server_bulk_callback;
 145         desc->bd_cbid.cbid_arg = desc;
 146
 147         /* NB we don't assign rq_bulk here; server-side requests are
 148          * re-used, and the handler frees the bulk desc explicitly. */
 149
 150         return desc;
 151 }
 152
 153 void ptlrpc_prep_bulk_page(struct ptlrpc_bulk_desc *desc,
 154                            struct page *page, int pageoffset, int len)
 155 {
 156         LASSERT(desc->bd_iov_count < desc->bd_max_iov);
 157         LASSERT(page != NULL);
 158         LASSERT(pageoffset >= 0);
 159         LASSERT(len > 0);
 160         LASSERT(pageoffset + len <= PAGE_SIZE);
 161
 162         desc->bd_nob += len;
 163
 164         ptlrpc_add_bulk_page(desc, page, pageoffset, len);
 165 }
 166
 167 void ptlrpc_free_bulk(struct ptlrpc_bulk_desc *desc)
 168 {
 169         ENTRY;
 170
 171         LASSERT(desc != NULL);
 172         LASSERT(desc->bd_iov_count != LI_POISON); /* not freed already */
 173         LASSERT(!desc->bd_network_rw);         /* network hands off or */
 174         LASSERT((desc->bd_export != NULL) ^ (desc->bd_import != NULL));
 175         if (desc->bd_export)
 176                 class_export_put(desc->bd_export);
 177         else
 178                 class_import_put(desc->bd_import);
 179
 180         OBD_FREE(desc, offsetof(struct ptlrpc_bulk_desc,
 181                                 bd_iov[desc->bd_max_iov]));
 182         EXIT;
 183 }
 184
 185 /* FIXME prep_req now should return error code other than NULL. but
 186  * this is called everywhere :(
 187  */
 188 struct ptlrpc_request *ptlrpc_prep_req(struct obd_import *imp, __u32 version,
 189                                        int opcode, int count, int *lengths,
 190                                        char **bufs)
 191 {
 192         struct ptlrpc_request *request;
 193         int rc;
 194         ENTRY;
 195
 196         LASSERT((unsigned long)imp > 0x1000);
 197
 198         OBD_ALLOC(request, sizeof(*request));
 199         if (!request) {
 200                 CERROR("request allocation out of memory\n");
 201                 RETURN(NULL);
 202         }
 203
 204         request->rq_import = class_import_get(imp);
 205
 206         rc = ptlrpcs_req_get_cred(request);
 207         if (rc) {
 208                 CDEBUG(D_SEC, "failed to get credential\n");
 209                 GOTO(out_free, rc);
 210         }
 211
 212         /* try to refresh the cred. we do this here in order to let fewer
 213          * refresh be performed in ptlrpcd context (which might block ptlrpcd).
 214          * fail out only if a fatal ptlrpcs error occured.
 215          */
 216         ptlrpcs_req_refresh_cred(request);
 217         if (request->rq_ptlrpcs_err)
 218                 GOTO(out_cred, rc = -EPERM);
 219
 220         /* set default sec flavor for this req. in the future we might need
 221          * increase security strengh, e.g. AUTH -> PRIV
 222          */
 223         request->rq_req_secflvr = imp->imp_sec->ps_flavor;
 224
 225         rc = lustre_pack_request(request, count, lengths, bufs);
 226         if (rc) {
 227                 CERROR("cannot pack request %d\n", rc);
 228                 GOTO(out_cred, rc);
 229         }
 230         request->rq_reqmsg->version |= version;
 231
 232         if (imp->imp_server_timeout)
 233                 request->rq_timeout = obd_timeout / 2;
 234         else
 235                 request->rq_timeout = obd_timeout;
 236
 237         request->rq_send_state = LUSTRE_IMP_FULL;
 238         request->rq_type = PTL_RPC_MSG_REQUEST;
 239
 240         request->rq_req_cbid.cbid_fn  = request_out_callback;
 241         request->rq_req_cbid.cbid_arg = request;
 242
 243         request->rq_reply_cbid.cbid_fn  = reply_in_callback;
 244         request->rq_reply_cbid.cbid_arg = request;
 245
 246         request->rq_phase = RQ_PHASE_NEW;
 247
 248         /* XXX FIXME bug 249 */
 249         request->rq_request_portal = imp->imp_client->cli_request_portal;
 250         request->rq_reply_portal = imp->imp_client->cli_reply_portal;
 251
 252         spin_lock_init(&request->rq_lock);
 253         INIT_LIST_HEAD(&request->rq_list);
 254         INIT_LIST_HEAD(&request->rq_replay_list);
 255         INIT_LIST_HEAD(&request->rq_set_chain);
 256         init_waitqueue_head(&request->rq_reply_waitq);
 257         request->rq_xid = ptlrpc_next_xid();
 258         atomic_set(&request->rq_refcount, 1);
 259
 260         request->rq_reqmsg->opc = opcode;
 261         request->rq_reqmsg->flags = 0;
 262         RETURN(request);
 263 out_cred:
 264         ptlrpcs_req_drop_cred(request);
 265 out_free:
 266         class_import_put(imp);
 267         OBD_FREE(request, sizeof(*request));
 268         RETURN(NULL);
 269 }
 270
 271 struct ptlrpc_request_set *ptlrpc_prep_set(void)
 272 {
 273         struct ptlrpc_request_set *set;
 274
 275         OBD_ALLOC(set, sizeof *set);
 276         if (!set)
 277                 RETURN(NULL);
 278         INIT_LIST_HEAD(&set->set_requests);
 279         init_waitqueue_head(&set->set_waitq);
 280         set->set_remaining = 0;
 281         spin_lock_init(&set->set_new_req_lock);
 282         INIT_LIST_HEAD(&set->set_new_requests);
 283
 284         RETURN(set);
 285 }
 286
 287 /* Finish with this set; opposite of prep_set. */
 288 void ptlrpc_set_destroy(struct ptlrpc_request_set *set)
 289 {
 290         struct list_head *tmp;
 291         struct list_head *next;
 292         int               expected_phase;
 293         int               n = 0;
 294         ENTRY;
 295
 296         /* Requests on the set should either all be completed, or all be new */
 297         expected_phase = (set->set_remaining == 0) ?
 298                          RQ_PHASE_COMPLETE : RQ_PHASE_NEW;
 299         list_for_each (tmp, &set->set_requests) {
 300                 struct ptlrpc_request *req =
 301                         list_entry(tmp, struct ptlrpc_request, rq_set_chain);
 302
 303                 LASSERT(req->rq_phase == expected_phase);
 304                 n++;
 305         }
 306
 307         LASSERT(set->set_remaining == 0 || set->set_remaining == n);
 308
 309         list_for_each_safe(tmp, next, &set->set_requests) {
 310                 struct ptlrpc_request *req =
 311                         list_entry(tmp, struct ptlrpc_request, rq_set_chain);
 312                 list_del_init(&req->rq_set_chain);
 313
 314                 LASSERT(req->rq_phase == expected_phase);
 315
 316                 if (req->rq_phase == RQ_PHASE_NEW) {
 317
 318                         if (req->rq_interpret_reply != NULL) {
 319                                 int (*interpreter)(struct ptlrpc_request *,
 320                                                    void *, int) =
 321                                         req->rq_interpret_reply;
 322
 323                                 /* higher level (i.e. LOV) failed;
 324                                  * let the sub reqs clean up */
 325                                 req->rq_status = -EBADR;
 326                                 interpreter(req, &req->rq_async_args,
 327                                             req->rq_status);
 328                         }
 329                         set->set_remaining--;
 330                 }
 331
 332                 req->rq_set = NULL;
 333                 ptlrpc_req_finished (req);
 334         }
 335
 336         LASSERT(set->set_remaining == 0);
 337
 338         OBD_FREE(set, sizeof(*set));
 339         EXIT;
 340 }
 341
 342 void ptlrpc_set_add_req(struct ptlrpc_request_set *set,
 343                         struct ptlrpc_request *req)
 344 {
 345         /* The set takes over the caller's request reference */
 346         list_add_tail(&req->rq_set_chain, &set->set_requests);
 347         req->rq_set = set;
 348         set->set_remaining++;
 349         atomic_inc(&req->rq_import->imp_inflight);
 350 }
 351
 352 /* lock so many callers can add things, the context that owns the set
 353  * is supposed to notice these and move them into the set proper. */
 354 void ptlrpc_set_add_new_req(struct ptlrpc_request_set *set,
 355                             struct ptlrpc_request *req)
 356 {
 357         unsigned long flags;
 358         spin_lock_irqsave(&set->set_new_req_lock, flags);
 359         /* The set takes over the caller's request reference */
 360         list_add_tail(&req->rq_set_chain, &set->set_new_requests);
 361         req->rq_set = set;
 362         spin_unlock_irqrestore(&set->set_new_req_lock, flags);
 363 }
 364
 365 /*
 366  * Based on the current state of the import, determine if the request
 367  * can be sent, is an error, or should be delayed.
 368  *
 369  * Returns true if this request should be delayed. If false, and
 370  * *status is set, then the request can not be sent and *status is the
 371  * error code.  If false and status is 0, then request can be sent.
 372  *
 373  * The imp->imp_lock must be held.
 374  */
 375 static int ptlrpc_import_delay_req(struct obd_import *imp,
 376                                    struct ptlrpc_request *req, int *status)
 377 {
 378         int delay = 0;
 379         ENTRY;
 380
 381         LASSERT (status != NULL);
 382         *status = 0;
 383
 384         if (imp->imp_state == LUSTRE_IMP_NEW) {
 385                 DEBUG_REQ(D_ERROR, req, "Uninitialized import.");
 386                 *status = -EIO;
 387                 LBUG();
 388         }
 389         else if (imp->imp_state == LUSTRE_IMP_CLOSED) {
 390                 DEBUG_REQ(D_ERROR, req, "IMP_CLOSED ");
 391                 *status = -EIO;
 392         }
 393         /* allow CONNECT even if import is invalid */
 394         else if (req->rq_send_state == LUSTRE_IMP_CONNECTING &&
 395                  imp->imp_state == LUSTRE_IMP_CONNECTING) {
 396                 ;
 397         }
 398         /*
 399          * If the import has been invalidated (such as by an OST failure), the
 400          * request must fail with -EIO.
 401          */
 402         else if (imp->imp_invalid) {
 403                 DEBUG_REQ(D_ERROR, req, "IMP_INVALID");
 404                 *status = -EIO;
 405         }
 406         else if (req->rq_import_generation != imp->imp_generation) {
 407                 DEBUG_REQ(D_ERROR, req, "req wrong generation:");
 408                 *status = -EIO;
 409         }
 410         else if (req->rq_send_state != imp->imp_state) {
 411                 if (imp->imp_obd->obd_no_recov || imp->imp_dlm_fake
 412                     || req->rq_no_delay)
 413                         *status = -EWOULDBLOCK;
 414                 else
 415                         delay = 1;
 416         }
 417
 418         RETURN(delay);
 419 }
 420
 421 static int ptlrpc_check_reply(struct ptlrpc_request *req)
 422 {
 423         unsigned long flags;
 424         int rc = 0;
 425         ENTRY;
 426
 427         /* serialise with network callback */
 428         spin_lock_irqsave (&req->rq_lock, flags);
 429
 430         if (req->rq_replied) {
 431                 DEBUG_REQ(D_NET, req, "REPLIED:");
 432                 GOTO(out, rc = 1);
 433         }
 434
 435         if (req->rq_net_err && !req->rq_timedout) {
 436                 spin_unlock_irqrestore (&req->rq_lock, flags);
 437                 rc = ptlrpc_expire_one_request(req);
 438                 spin_lock_irqsave (&req->rq_lock, flags);
 439                 GOTO(out, rc);
 440         }
 441
 442         if (req->rq_err) {
 443                 DEBUG_REQ(D_ERROR, req, "ABORTED:");
 444                 GOTO(out, rc = 1);
 445         }
 446
 447         if (req->rq_resend) {
 448                 DEBUG_REQ(D_ERROR, req, "RESEND:");
 449                 GOTO(out, rc = 1);
 450         }
 451
 452         if (req->rq_restart) {
 453                 DEBUG_REQ(D_ERROR, req, "RESTART:");
 454                 GOTO(out, rc = 1);
 455         }
 456         EXIT;
 457  out:
 458         spin_unlock_irqrestore (&req->rq_lock, flags);
 459         DEBUG_REQ(D_NET, req, "rc = %d for", rc);
 460         return rc;
 461 }
 462
 463 static int ptlrpc_check_status(struct ptlrpc_request *req)
 464 {
 465         int err;
 466         ENTRY;
 467
 468         err = req->rq_repmsg->status;
 469         if (req->rq_repmsg->type == PTL_RPC_MSG_ERR) {
 470                 DEBUG_REQ(D_ERROR, req, "type == PTL_RPC_MSG_ERR, err == %d",
 471                           err);
 472                 RETURN(err < 0 ? err : -EINVAL);
 473         }
 474
 475         if (err < 0) {
 476                 DEBUG_REQ(D_INFO, req, "status is %d", err);
 477         } else if (err > 0) {
 478                 /* XXX: translate this error from net to host */
 479                 DEBUG_REQ(D_INFO, req, "status is %d", err);
 480         }
 481
 482         RETURN(err);
 483 }
 484
 485 static int after_reply(struct ptlrpc_request *req)
 486 {
 487         unsigned long flags;
 488         struct obd_import *imp = req->rq_import;
 489         int rc;
 490         ENTRY;
 491
 492         LASSERT(!req->rq_receiving_reply);
 493
 494         /* NB Until this point, the whole of the incoming message,
 495          * including buflens, status etc is in the sender's byte order. */
 496
 497 #if SWAB_PARANOIA
 498         /* Clear reply swab mask; this is a new reply in sender's byte order */
 499         req->rq_rep_swab_mask = 0;
 500 #endif
 501         LASSERT (req->rq_nob_received <= req->rq_repbuf_len);
 502         rc = ptlrpcs_cli_unwrap_reply(req);
 503         if (rc) {
 504                 CERROR("verify reply error: %d\n", rc);
 505                 RETURN(rc);
 506         }
 507         /* unwrap_reply may request rpc be resend */
 508         if (req->rq_ptlrpcs_restart) {
 509                 req->rq_resend = 1;
 510                 RETURN(0);
 511         }
 512
 513         /* unwrap_reply will set rq_replen as the actual received
 514          * lustre_msg length
 515          */
 516         rc = lustre_unpack_msg(req->rq_repmsg, req->rq_replen);
 517         if (rc) {
 518                 CERROR("unpack_rep failed: %d\n", rc);
 519                 RETURN(-EPROTO);
 520         }
 521
 522         if (req->rq_repmsg->type != PTL_RPC_MSG_REPLY &&
 523             req->rq_repmsg->type != PTL_RPC_MSG_ERR) {
 524                 CERROR("invalid packet type received (type=%u)\n",
 525                        req->rq_repmsg->type);
 526                 RETURN(-EPROTO);
 527         }
 528
 529         rc = ptlrpc_check_status(req);
 530
 531         /* Either we've been evicted, or the server has failed for
 532          * some reason. Try to reconnect, and if that fails, punt to the
 533          * upcall. */
 534         if (rc == -ENOTCONN) {
 535                 if (req->rq_send_state != LUSTRE_IMP_FULL ||
 536                     imp->imp_obd->obd_no_recov || imp->imp_dlm_fake) {
 537                         RETURN(-ENOTCONN);
 538                 }
 539
 540                 ptlrpc_request_handle_notconn(req);
 541
 542                 RETURN(rc);
 543         }
 544
 545         /* Store transno in reqmsg for replay. */
 546         req->rq_reqmsg->transno = req->rq_transno = req->rq_repmsg->transno;
 547
 548         if (req->rq_import->imp_replayable) {
 549                 spin_lock_irqsave(&imp->imp_lock, flags);
 550                 if (req->rq_transno != 0)
 551                         ptlrpc_retain_replayable_request(req, imp);
 552                 else if (req->rq_commit_cb != NULL) {
 553                         spin_unlock_irqrestore(&imp->imp_lock, flags);
 554                         req->rq_commit_cb(req);
 555                         spin_lock_irqsave(&imp->imp_lock, flags);
 556                 }
 557
 558                 if (req->rq_transno > imp->imp_max_transno)
 559                         imp->imp_max_transno = req->rq_transno;
 560
 561                 /* Replay-enabled imports return commit-status information. */
 562                 if (req->rq_repmsg->last_committed)
 563                         imp->imp_peer_committed_transno =
 564                                 req->rq_repmsg->last_committed;
 565                 ptlrpc_free_committed(imp);
 566                 spin_unlock_irqrestore(&imp->imp_lock, flags);
 567         }
 568
 569         RETURN(rc);
 570 }
 571
 572 static int ptlrpc_send_new_req(struct ptlrpc_request *req)
 573 {
 574         char                   str[PTL_NALFMT_SIZE];
 575         struct obd_import     *imp;
 576         unsigned long          flags;
 577         int rc;
 578         ENTRY;
 579
 580         LASSERT(req->rq_phase == RQ_PHASE_NEW);
 581         req->rq_phase = RQ_PHASE_RPC;
 582
 583         imp = req->rq_import;
 584         spin_lock_irqsave(&imp->imp_lock, flags);
 585
 586         req->rq_import_generation = imp->imp_generation;
 587
 588         if (ptlrpc_import_delay_req(imp, req, &rc)) {
 589                 spin_lock (&req->rq_lock);
 590                 req->rq_waiting = 1;
 591                 spin_unlock (&req->rq_lock);
 592
 593                 DEBUG_REQ(D_HA, req, "req from PID %d waiting for recovery: "
 594                           "(%s != %s)",
 595                           req->rq_reqmsg->status,
 596                           ptlrpc_import_state_name(req->rq_send_state),
 597                           ptlrpc_import_state_name(imp->imp_state));
 598                 LASSERT(list_empty (&req->rq_list));
 599
 600                 list_add_tail(&req->rq_list, &imp->imp_delayed_list);
 601                 spin_unlock_irqrestore(&imp->imp_lock, flags);
 602                 RETURN(0);
 603         }
 604
 605         if (rc != 0) {
 606                 spin_unlock_irqrestore(&imp->imp_lock, flags);
 607                 req->rq_status = rc;
 608                 req->rq_phase = RQ_PHASE_INTERPRET;
 609                 RETURN(rc);
 610         }
 611
 612         /* XXX this is the same as ptlrpc_queue_wait */
 613         LASSERT(list_empty(&req->rq_list));
 614         list_add_tail(&req->rq_list, &imp->imp_sending_list);
 615         spin_unlock_irqrestore(&imp->imp_lock, flags);
 616
 617         req->rq_reqmsg->status = current->pid;
 618         CDEBUG(D_RPCTRACE, "Sending RPC pname:cluuid:pid:xid:ni:nid:opc"
 619                " %s:%s:%d:"LPU64":%s:%s:%d\n", current->comm,
 620                imp->imp_obd->obd_uuid.uuid, req->rq_reqmsg->status,
 621                req->rq_xid,
 622                imp->imp_connection->c_peer.peer_ni->pni_name,
 623                ptlrpc_peernid2str(&imp->imp_connection->c_peer, str),
 624                req->rq_reqmsg->opc);
 625
 626         rc = ptl_send_rpc(req);
 627         if (rc) {
 628                 DEBUG_REQ(D_HA, req, "send failed (%d); expect timeout", rc);
 629                 req->rq_net_err = 1;
 630                 RETURN(rc);
 631         }
 632         RETURN(0);
 633 }
 634
 635 int ptlrpc_check_set(struct ptlrpc_request_set *set)
 636 {
 637         char str[PTL_NALFMT_SIZE];
 638         unsigned long flags;
 639         struct list_head *tmp;
 640         int force_timer_recalc = 0;
 641         ENTRY;
 642
 643         if (set->set_remaining == 0)
 644                 RETURN(1);
 645
 646         list_for_each(tmp, &set->set_requests) {
 647                 struct ptlrpc_request *req =
 648                         list_entry(tmp, struct ptlrpc_request, rq_set_chain);
 649                 struct obd_import *imp = req->rq_import;
 650                 int rc = 0;
 651
 652                 if (req->rq_phase == RQ_PHASE_NEW &&
 653                     ptlrpc_send_new_req(req)) {
 654                         force_timer_recalc = 1;
 655                 }
 656
 657                 if (!(req->rq_phase == RQ_PHASE_RPC ||
 658                       req->rq_phase == RQ_PHASE_BULK ||
 659                       req->rq_phase == RQ_PHASE_INTERPRET ||
 660                       req->rq_phase == RQ_PHASE_COMPLETE)) {
 661                         DEBUG_REQ(D_ERROR, req, "bad phase %x", req->rq_phase);
 662                         LBUG();
 663                 }
 664
 665                 if (req->rq_phase == RQ_PHASE_COMPLETE)
 666                         continue;
 667
 668                 if (req->rq_phase == RQ_PHASE_INTERPRET)
 669                         GOTO(interpret, req->rq_status);
 670
 671                 if (req->rq_net_err && !req->rq_timedout)
 672                         ptlrpc_expire_one_request(req);
 673
 674                 if (req->rq_err || req->rq_ptlrpcs_err) {
 675                         ptlrpc_unregister_reply(req);
 676                         if (req->rq_status == 0)
 677                                 req->rq_status = req->rq_err ? -EIO : -EPERM;
 678                         req->rq_phase = RQ_PHASE_INTERPRET;
 679
 680                         spin_lock_irqsave(&imp->imp_lock, flags);
 681                         list_del_init(&req->rq_list);
 682                         spin_unlock_irqrestore(&imp->imp_lock, flags);
 683
 684                         GOTO(interpret, req->rq_status);
 685                 }
 686
 687                 /* ptlrpc_queue_wait->l_wait_event guarantees that rq_intr
 688                  * will only be set after rq_timedout, but the oig waiting
 689                  * path sets rq_intr irrespective of whether ptlrpcd has
 690                  * seen a timeout.  our policy is to only interpret
 691                  * interrupted rpcs after they have timed out */
 692                 if (req->rq_intr && (req->rq_timedout || req->rq_waiting)) {
 693                         /* NB could be on delayed list */
 694                         ptlrpc_unregister_reply(req);
 695                         req->rq_status = -EINTR;
 696                         req->rq_phase = RQ_PHASE_INTERPRET;
 697
 698                         spin_lock_irqsave(&imp->imp_lock, flags);
 699                         list_del_init(&req->rq_list);
 700                         spin_unlock_irqrestore(&imp->imp_lock, flags);
 701
 702                         GOTO(interpret, req->rq_status);
 703                 }
 704
 705                 if (req->rq_phase == RQ_PHASE_RPC) {
 706                         if (req->rq_timedout||req->rq_waiting||req->rq_resend) {
 707                                 int status;
 708
 709                                 ptlrpc_unregister_reply(req);
 710
 711                                 spin_lock_irqsave(&imp->imp_lock, flags);
 712
 713                                 if (ptlrpc_import_delay_req(imp, req, &status)){
 714                                         spin_unlock_irqrestore(&imp->imp_lock,
 715                                                                flags);
 716                                         continue;
 717                                 }
 718
 719                                 list_del_init(&req->rq_list);
 720                                 if (status != 0)  {
 721                                         req->rq_status = status;
 722                                         req->rq_phase = RQ_PHASE_INTERPRET;
 723                                         spin_unlock_irqrestore(&imp->imp_lock,
 724                                                                flags);
 725                                         GOTO(interpret, req->rq_status);
 726                                 }
 727                                 if (req->rq_no_resend) {
 728                                         req->rq_status = -ENOTCONN;
 729                                         req->rq_phase = RQ_PHASE_INTERPRET;
 730                                         spin_unlock_irqrestore(&imp->imp_lock,
 731                                                                flags);
 732                                         GOTO(interpret, req->rq_status);
 733                                 }
 734                                 list_add_tail(&req->rq_list,
 735                                               &imp->imp_sending_list);
 736
 737                                 spin_unlock_irqrestore(&imp->imp_lock, flags);
 738
 739                                 req->rq_waiting = 0;
 740                                 if (req->rq_resend) {
 741                                         if (!req->rq_ptlrpcs_restart)
 742                                                 lustre_msg_add_flags(
 743                                                         req->rq_reqmsg,
 744                                                         MSG_RESENT);
 745                                         if (req->rq_bulk) {
 746                                                 __u64 old_xid = req->rq_xid;
 747
 748                                                 ptlrpc_unregister_bulk (req);
 749
 750                                                 /* ensure previous bulk fails */
 751                                                 req->rq_xid = ptlrpc_next_xid();
 752                                                 CDEBUG(D_HA, "resend bulk "
 753                                                        "old x"LPU64
 754                                                        " new x"LPU64"\n",
 755                                                        old_xid, req->rq_xid);
 756                                         }
 757                                 }
 758
 759                                 rc = ptl_send_rpc(req);
 760                                 if (rc) {
 761                                         DEBUG_REQ(D_HA, req, "send failed (%d)",
 762                                                   rc);
 763                                         force_timer_recalc = 1;
 764                                         req->rq_net_err = 1;
 765                                 }
 766                                 /* need to reset the timeout */
 767                                 force_timer_recalc = 1;
 768                         }
 769
 770                         /* Still waiting for a reply? */
 771                         if (ptlrpc_client_receiving_reply(req))
 772                                 continue;
 773
 774                         /* Did we actually receive a reply? */
 775                         if (!ptlrpc_client_replied(req))
 776                                 continue;
 777
 778                         spin_lock_irqsave(&imp->imp_lock, flags);
 779                         list_del_init(&req->rq_list);
 780                         spin_unlock_irqrestore(&imp->imp_lock, flags);
 781
 782                         req->rq_status = after_reply(req);
 783                         if (req->rq_resend) {
 784                                 /* Add this req to the delayed list so
 785                                    it can be errored if the import is
 786                                    evicted after recovery. */
 787                                 spin_lock_irqsave (&req->rq_lock, flags);
 788                                 list_add_tail(&req->rq_list,
 789                                               &imp->imp_delayed_list);
 790                                 spin_unlock_irqrestore(&req->rq_lock, flags);
 791                                 continue;
 792                         }
 793
 794                         /* If there is no bulk associated with this request,
 795                          * then we're done and should let the interpreter
 796                          * process the reply.  Similarly if the RPC returned
 797                          * an error, and therefore the bulk will never arrive.
 798                          */
 799                         if (req->rq_bulk == NULL || req->rq_status != 0) {
 800                                 req->rq_phase = RQ_PHASE_INTERPRET;
 801                                 GOTO(interpret, req->rq_status);
 802                         }
 803
 804                         req->rq_phase = RQ_PHASE_BULK;
 805                 }
 806
 807                 LASSERT(req->rq_phase == RQ_PHASE_BULK);
 808                 if (ptlrpc_bulk_active(req->rq_bulk))
 809                         continue;
 810
 811                 if (!req->rq_bulk->bd_success) {
 812                         /* The RPC reply arrived OK, but the bulk screwed
 813                          * up!  Dead wierd since the server told us the RPC
 814                          * was good after getting the REPLY for her GET or
 815                          * the ACK for her PUT. */
 816                         DEBUG_REQ(D_ERROR, req, "bulk transfer failed");
 817                         LBUG();
 818                 }
 819
 820                 req->rq_phase = RQ_PHASE_INTERPRET;
 821
 822         interpret:
 823                 LASSERT(req->rq_phase == RQ_PHASE_INTERPRET);
 824                 LASSERT(!req->rq_receiving_reply);
 825
 826                 ptlrpc_unregister_reply(req);
 827                 if (req->rq_bulk != NULL)
 828                         ptlrpc_unregister_bulk (req);
 829
 830                 req->rq_phase = RQ_PHASE_COMPLETE;
 831
 832                 if (req->rq_interpret_reply != NULL) {
 833                         int (*interpreter)(struct ptlrpc_request *,void *,int) =
 834                                 req->rq_interpret_reply;
 835                         req->rq_status = interpreter(req, &req->rq_async_args,
 836                                                      req->rq_status);
 837                 }
 838
 839                 CDEBUG(D_RPCTRACE, "Completed RPC pname:cluuid:pid:xid:ni:nid:"
 840                        "opc %s:%s:%d:"LPU64":%s:%s:%d\n", current->comm,
 841                        imp->imp_obd->obd_uuid.uuid, req->rq_reqmsg->status,
 842                        req->rq_xid,
 843                        imp->imp_connection->c_peer.peer_ni->pni_name,
 844                        ptlrpc_peernid2str(&imp->imp_connection->c_peer, str),
 845                        req->rq_reqmsg->opc);
 846
 847                 set->set_remaining--;
 848
 849                 atomic_dec(&imp->imp_inflight);
 850                 wake_up(&imp->imp_recovery_waitq);
 851         }
 852
 853         /* If we hit an error, we want to recover promptly. */
 854         RETURN(set->set_remaining == 0 || force_timer_recalc);
 855 }
 856
 857 int ptlrpc_expire_one_request(struct ptlrpc_request *req)
 858 {
 859         unsigned long      flags;
 860         struct obd_import *imp = req->rq_import;
 861         int replied = 0;
 862         ENTRY;
 863
 864         DEBUG_REQ(D_ERROR, req, "timeout (sent at %lu, %lus ago)",
 865                   (long)req->rq_sent, LTIME_S(CURRENT_TIME) - req->rq_sent);
 866
 867         spin_lock_irqsave (&req->rq_lock, flags);
 868         replied = req->rq_replied;
 869         if (!replied)
 870                 req->rq_timedout = 1;
 871         spin_unlock_irqrestore (&req->rq_lock, flags);
 872
 873         if (replied)
 874                 RETURN(0);
 875
 876         ptlrpc_unregister_reply (req);
 877
 878         if (obd_dump_on_timeout)
 879                 portals_debug_dumplog();
 880
 881         if (req->rq_bulk != NULL)
 882                 ptlrpc_unregister_bulk (req);
 883
 884         if (imp == NULL) {
 885                 DEBUG_REQ(D_HA, req, "NULL import: already cleaned up?");
 886                 RETURN(1);
 887         }
 888
 889         /* The DLM server doesn't want recovery run on its imports. */
 890         if (imp->imp_dlm_fake)
 891                 RETURN(1);
 892
 893         /* If this request is for recovery or other primordial tasks,
 894          * then error it out here. */
 895         if (req->rq_send_state != LUSTRE_IMP_FULL ||
 896             imp->imp_obd->obd_no_recov) {
 897                 spin_lock_irqsave (&req->rq_lock, flags);
 898                 req->rq_status = -ETIMEDOUT;
 899                 req->rq_err = 1;
 900                 spin_unlock_irqrestore (&req->rq_lock, flags);
 901                 RETURN(1);
 902         }
 903
 904         ptlrpc_fail_import(imp, req->rq_import_generation);
 905
 906         RETURN(0);
 907 }
 908
 909 int ptlrpc_expired_set(void *data)
 910 {
 911         struct ptlrpc_request_set *set = data;
 912         struct list_head          *tmp;
 913         time_t                     now = LTIME_S(CURRENT_TIME);
 914         ENTRY;
 915
 916         LASSERT(set != NULL);
 917
 918         /* A timeout expired; see which reqs it applies to... */
 919         list_for_each (tmp, &set->set_requests) {
 920                 struct ptlrpc_request *req =
 921                         list_entry(tmp, struct ptlrpc_request, rq_set_chain);
 922
 923                 /* request in-flight? */
 924                 if (!((req->rq_phase == RQ_PHASE_RPC && !req->rq_waiting
 925                        && !req->rq_resend) ||
 926                       (req->rq_phase == RQ_PHASE_BULK)))
 927                         continue;
 928
 929                 if (req->rq_timedout ||           /* already dealt with */
 930                     req->rq_sent + req->rq_timeout > now) /* not expired */
 931                         continue;
 932
 933                 /* deal with this guy */
 934                 ptlrpc_expire_one_request (req);
 935         }
 936
 937         /* When waiting for a whole set, we always to break out of the
 938          * sleep so we can recalculate the timeout, or enable interrupts
 939          * iff everyone's timed out.
 940          */
 941         RETURN(1);
 942 }
 943
 944 void ptlrpc_mark_interrupted(struct ptlrpc_request *req)
 945 {
 946         unsigned long flags;
 947         spin_lock_irqsave(&req->rq_lock, flags);
 948         req->rq_intr = 1;
 949         spin_unlock_irqrestore(&req->rq_lock, flags);
 950 }
 951
 952 void ptlrpc_interrupted_set(void *data)
 953 {
 954         struct ptlrpc_request_set *set = data;
 955         struct list_head *tmp;
 956
 957         LASSERT(set != NULL);
 958         CERROR("INTERRUPTED SET %p\n", set);
 959
 960         list_for_each(tmp, &set->set_requests) {
 961                 struct ptlrpc_request *req =
 962                         list_entry(tmp, struct ptlrpc_request, rq_set_chain);
 963
 964                 if (req->rq_phase != RQ_PHASE_RPC)
 965                         continue;
 966
 967                 ptlrpc_mark_interrupted(req);
 968         }
 969 }
 970
 971 int ptlrpc_set_next_timeout(struct ptlrpc_request_set *set)
 972 {
 973         struct list_head      *tmp;
 974         time_t                 now = LTIME_S(CURRENT_TIME);
 975         time_t                 deadline;
 976         int                    timeout = 0;
 977         struct ptlrpc_request *req;
 978         ENTRY;
 979
 980         SIGNAL_MASK_ASSERT(); /* XXX BUG 1511 */
 981
 982         list_for_each(tmp, &set->set_requests) {
 983                 req = list_entry(tmp, struct ptlrpc_request, rq_set_chain);
 984
 985                 /* request in-flight? */
 986                 if (!((req->rq_phase == RQ_PHASE_RPC && !req->rq_waiting) ||
 987                       (req->rq_phase == RQ_PHASE_BULK)))
 988                         continue;
 989
 990                 if (req->rq_timedout)   /* already timed out */
 991                         continue;
 992
 993                 deadline = req->rq_sent + req->rq_timeout;
 994                 if (deadline <= now)    /* actually expired already */
 995                         timeout = 1;    /* ASAP */
 996                 else if (timeout == 0 || timeout > deadline - now)
 997                         timeout = deadline - now;
 998         }
 999         RETURN(timeout);
1000 }
1001
1002
1003 int ptlrpc_set_wait(struct ptlrpc_request_set *set)
1004 {
1005         struct list_head      *tmp;
1006         struct ptlrpc_request *req;
1007         struct l_wait_info     lwi;
1008         int                    rc, timeout;
1009         ENTRY;
1010
1011         LASSERT(!list_empty(&set->set_requests));
1012         list_for_each(tmp, &set->set_requests) {
1013                 req = list_entry(tmp, struct ptlrpc_request, rq_set_chain);
1014                 if (req->rq_phase == RQ_PHASE_NEW)
1015                         (void)ptlrpc_send_new_req(req);
1016         }
1017
1018         do {
1019                 timeout = ptlrpc_set_next_timeout(set);
1020
1021                 /* wait until all complete, interrupted, or an in-flight
1022                  * req times out */
1023                 CDEBUG(D_HA, "set %p going to sleep for %d seconds\n",
1024                        set, timeout);
1025                 lwi = LWI_TIMEOUT_INTR((timeout ? timeout : 1) * HZ,
1026                                        ptlrpc_expired_set,
1027                                        ptlrpc_interrupted_set, set);
1028                 rc = l_wait_event(set->set_waitq, ptlrpc_check_set(set), &lwi);
1029
1030                 LASSERT(rc == 0 || rc == -EINTR || rc == -ETIMEDOUT);
1031
1032                 /* -EINTR => all requests have been flagged rq_intr so next
1033                  * check completes.
1034                  * -ETIMEOUTD => someone timed out.  When all reqs have
1035                  * timed out, signals are enabled allowing completion with
1036                  * EINTR.
1037                  * I don't really care if we go once more round the loop in
1038                  * the error cases -eeb. */
1039         } while (rc != 0 || set->set_remaining != 0);
1040
1041         LASSERT(set->set_remaining == 0);
1042
1043         rc = 0;
1044         list_for_each(tmp, &set->set_requests) {
1045                 req = list_entry(tmp, struct ptlrpc_request, rq_set_chain);
1046
1047                 LASSERT(req->rq_phase == RQ_PHASE_COMPLETE);
1048                 if (req->rq_status != 0)
1049                         rc = req->rq_status;
1050         }
1051
1052         if (set->set_interpret != NULL) {
1053                 int (*interpreter)(struct ptlrpc_request_set *set,void *,int) =
1054                         set->set_interpret;
1055                 rc = interpreter (set, set->set_arg, rc);
1056         }
1057
1058         RETURN(rc);
1059 }
1060
1061 static void __ptlrpc_free_req(struct ptlrpc_request *request, int locked)
1062 {
1063         ENTRY;
1064         if (request == NULL) {
1065                 EXIT;
1066                 return;
1067         }
1068
1069         LASSERTF(!request->rq_receiving_reply, "req %p\n", request);
1070         LASSERTF(request->rq_rqbd == NULL, "req %p\n",request);/* client-side */
1071         LASSERTF(list_empty(&request->rq_list), "req %p\n", request);
1072         LASSERTF(list_empty(&request->rq_set_chain), "req %p\n", request);
1073         LASSERT(request->rq_cred);
1074
1075         /* We must take it off the imp_replay_list first.  Otherwise, we'll set
1076          * request->rq_reqmsg to NULL while osc_close is dereferencing it. */
1077         if (request->rq_import != NULL) {
1078                 unsigned long flags = 0;
1079                 if (!locked)
1080                         spin_lock_irqsave(&request->rq_import->imp_lock, flags);
1081                 list_del_init(&request->rq_replay_list);
1082                 if (!locked)
1083                         spin_unlock_irqrestore(&request->rq_import->imp_lock,
1084                                                flags);
1085         }
1086         LASSERTF(list_empty(&request->rq_replay_list), "req %p\n", request);
1087
1088         if (atomic_read(&request->rq_refcount) != 0) {
1089                 DEBUG_REQ(D_ERROR, request,
1090                           "freeing request with nonzero refcount");
1091                 LBUG();
1092         }
1093
1094         if (request->rq_repbuf != NULL)
1095                 ptlrpcs_cli_free_repbuf(request);
1096         if (request->rq_reqbuf != NULL)
1097                 ptlrpcs_cli_free_reqbuf(request);
1098
1099         if (request->rq_export != NULL) {
1100                 class_export_put(request->rq_export);
1101                 request->rq_export = NULL;
1102         }
1103         if (request->rq_import != NULL) {
1104                 class_import_put(request->rq_import);
1105                 request->rq_import = NULL;
1106         }
1107         if (request->rq_bulk != NULL)
1108                 ptlrpc_free_bulk(request->rq_bulk);
1109
1110         ptlrpcs_req_drop_cred(request);
1111         OBD_FREE(request, sizeof(*request));
1112         EXIT;
1113 }
1114
1115 void ptlrpc_free_req(struct ptlrpc_request *request)
1116 {
1117         __ptlrpc_free_req(request, 0);
1118 }
1119
1120 static int __ptlrpc_req_finished(struct ptlrpc_request *request, int locked);
1121 void ptlrpc_req_finished_with_imp_lock(struct ptlrpc_request *request)
1122 {
1123         LASSERT_SPIN_LOCKED(&request->rq_import->imp_lock);
1124         (void)__ptlrpc_req_finished(request, 1);
1125 }
1126
1127 static int __ptlrpc_req_finished(struct ptlrpc_request *request, int locked)
1128 {
1129         ENTRY;
1130         if (request == NULL)
1131                 RETURN(1);
1132
1133         if (request == LP_POISON ||
1134             request->rq_reqmsg == LP_POISON) {
1135                 CERROR("dereferencing freed request (bug 575)\n");
1136                 LBUG();
1137                 RETURN(1);
1138         }
1139
1140         DEBUG_REQ(D_INFO, request, "refcount now %u",
1141                   atomic_read(&request->rq_refcount) - 1);
1142
1143         if (atomic_dec_and_test(&request->rq_refcount)) {
1144                 __ptlrpc_free_req(request, locked);
1145                 RETURN(1);
1146         }
1147
1148         RETURN(0);
1149 }
1150
1151 void ptlrpc_req_finished(struct ptlrpc_request *request)
1152 {
1153         __ptlrpc_req_finished(request, 0);
1154 }
1155
1156 /* Disengage the client's reply buffer from the network
1157  * NB does _NOT_ unregister any client-side bulk.
1158  * IDEMPOTENT, but _not_ safe against concurrent callers.
1159  * The request owner (i.e. the thread doing the I/O) must call...
1160  */
1161 void ptlrpc_unregister_reply (struct ptlrpc_request *request)
1162 {
1163         int                rc;
1164         wait_queue_head_t *wq;
1165         struct l_wait_info lwi;
1166
1167         LASSERT(!in_interrupt ());             /* might sleep */
1168
1169         if (!ptlrpc_client_receiving_reply(request))
1170                 return;
1171
1172         PtlMDUnlink (request->rq_reply_md_h);
1173
1174         /* We have to l_wait_event() whatever the result, to give liblustre
1175          * a chance to run reply_in_callback() */
1176
1177         if (request->rq_set != NULL)
1178                 wq = &request->rq_set->set_waitq;
1179         else
1180                 wq = &request->rq_reply_waitq;
1181
1182         for (;;) {
1183                 /* Network access will complete in finite time but the HUGE
1184                  * timeout lets us CWARN for visibility of sluggish NALs */
1185                 lwi = LWI_TIMEOUT(300 * HZ, NULL, NULL);
1186                 rc = l_wait_event (*wq, !ptlrpc_client_receiving_reply(request), &lwi);
1187                 if (rc == 0)
1188                         return;
1189
1190                 LASSERT (rc == -ETIMEDOUT);
1191                 DEBUG_REQ(D_WARNING, request, "Unexpectedly long timeout");
1192         }
1193 }
1194
1195 /* caller must hold imp->imp_lock */
1196 void ptlrpc_free_committed(struct obd_import *imp)
1197 {
1198         struct list_head *tmp, *saved;
1199         struct ptlrpc_request *req;
1200         struct ptlrpc_request *last_req = NULL; /* temporary fire escape */
1201         ENTRY;
1202
1203         LASSERT(imp != NULL);
1204
1205         LASSERT_SPIN_LOCKED(&imp->imp_lock);
1206
1207         CDEBUG(D_HA, "%s: committing for last_committed "LPU64"\n",
1208                imp->imp_obd->obd_name, imp->imp_peer_committed_transno);
1209
1210         list_for_each_safe(tmp, saved, &imp->imp_replay_list) {
1211                 req = list_entry(tmp, struct ptlrpc_request, rq_replay_list);
1212
1213                 /* XXX ok to remove when 1357 resolved - rread 05/29/03  */
1214                 LASSERT(req != last_req);
1215                 last_req = req;
1216
1217                 if (req->rq_import_generation < imp->imp_generation) {
1218                         DEBUG_REQ(D_HA, req, "freeing request with old gen");
1219                         GOTO(free_req, 0);
1220                 }
1221
1222                 if (req->rq_replay) {
1223                         /* this is debug flag to avoid flooding of logs -bzzz */
1224                         if (imp->imp_debug_open_replays)
1225                                 DEBUG_REQ(D_HA, req, "keeping (FL_REPLAY)");
1226                         continue;
1227                 }
1228
1229                 /* not yet committed */
1230                 if (req->rq_transno > imp->imp_peer_committed_transno) {
1231                         DEBUG_REQ(D_HA, req, "stopping search");
1232                         break;
1233                 }
1234
1235                 DEBUG_REQ(D_HA, req, "committing (last_committed "LPU64")",
1236                           imp->imp_peer_committed_transno);
1237 free_req:
1238                 if (req->rq_commit_cb != NULL)
1239                         req->rq_commit_cb(req);
1240                 list_del_init(&req->rq_replay_list);
1241                 __ptlrpc_req_finished(req, 1);
1242         }
1243         imp->imp_debug_open_replays = 0;
1244
1245         EXIT;
1246         return;
1247 }
1248
1249 void ptlrpc_cleanup_client(struct obd_import *imp)
1250 {
1251         ENTRY;
1252         EXIT;
1253         return;
1254 }
1255
1256 void ptlrpc_resend_req(struct ptlrpc_request *req)
1257 {
1258         unsigned long flags;
1259
1260         DEBUG_REQ(D_HA, req, "going to resend");
1261         req->rq_reqmsg->handle.cookie = 0;
1262         req->rq_status = -EAGAIN;
1263
1264         spin_lock_irqsave (&req->rq_lock, flags);
1265         req->rq_resend = 1;
1266         req->rq_net_err = 0;
1267         req->rq_timedout = 0;
1268         if (req->rq_bulk) {
1269                 __u64 old_xid = req->rq_xid;
1270
1271                 /* ensure previous bulk fails */
1272                 req->rq_xid = ptlrpc_next_xid();
1273                 CDEBUG(D_HA, "resend bulk old x"LPU64" new x"LPU64"\n",
1274                        old_xid, req->rq_xid);
1275         }
1276         ptlrpc_wake_client_req(req);
1277         spin_unlock_irqrestore (&req->rq_lock, flags);
1278 }
1279
1280 /* XXX: this function and rq_status are currently unused */
1281 void ptlrpc_restart_req(struct ptlrpc_request *req)
1282 {
1283         unsigned long flags;
1284
1285         DEBUG_REQ(D_HA, req, "restarting (possibly-)completed request");
1286         req->rq_status = -ERESTARTSYS;
1287
1288         spin_lock_irqsave (&req->rq_lock, flags);
1289         req->rq_restart = 1;
1290         req->rq_timedout = 0;
1291         ptlrpc_wake_client_req(req);
1292         spin_unlock_irqrestore (&req->rq_lock, flags);
1293 }
1294
1295 static int expired_request(void *data)
1296 {
1297         struct ptlrpc_request *req = data;
1298         ENTRY;
1299
1300         /* some failure can suspend regular timeouts */
1301         if (ptlrpc_check_suspend())
1302                 RETURN(1);
1303
1304         RETURN(ptlrpc_expire_one_request(req));
1305 }
1306
1307 static void interrupted_request(void *data)
1308 {
1309         unsigned long flags;
1310
1311         struct ptlrpc_request *req = data;
1312         DEBUG_REQ(D_HA, req, "request interrupted");
1313         spin_lock_irqsave (&req->rq_lock, flags);
1314         req->rq_intr = 1;
1315         spin_unlock_irqrestore (&req->rq_lock, flags);
1316 }
1317
1318 struct ptlrpc_request *ptlrpc_request_addref(struct ptlrpc_request *req)
1319 {
1320         ENTRY;
1321         atomic_inc(&req->rq_refcount);
1322         RETURN(req);
1323 }
1324
1325 void ptlrpc_retain_replayable_request(struct ptlrpc_request *req,
1326                                       struct obd_import *imp)
1327 {
1328         struct list_head *tmp;
1329
1330         LASSERT_SPIN_LOCKED(&imp->imp_lock);
1331
1332         /* clear this  for new requests that were resent as well
1333            as resent replayed requests. */
1334         lustre_msg_clear_flags(req->rq_reqmsg, MSG_RESENT);
1335
1336         /* don't re-add requests that have been replayed */
1337         if (!list_empty(&req->rq_replay_list))
1338                 return;
1339
1340         lustre_msg_add_flags(req->rq_reqmsg, MSG_REPLAY);
1341
1342         LASSERT(imp->imp_replayable);
1343         /* Balanced in ptlrpc_free_committed, usually. */
1344         ptlrpc_request_addref(req);
1345         list_for_each_prev(tmp, &imp->imp_replay_list) {
1346                 struct ptlrpc_request *iter =
1347                         list_entry(tmp, struct ptlrpc_request, rq_replay_list);
1348
1349                 /* We may have duplicate transnos if we create and then
1350                  * open a file, or for closes retained if to match creating
1351                  * opens, so use req->rq_xid as a secondary key.
1352                  * (See bugs 684, 685, and 428.)
1353                  * XXX no longer needed, but all opens need transnos!
1354                  */
1355                 if (iter->rq_transno > req->rq_transno)
1356                         continue;
1357
1358                 if (iter->rq_transno == req->rq_transno) {
1359                         LASSERT(iter->rq_xid != req->rq_xid);
1360                         if (iter->rq_xid > req->rq_xid)
1361                                 continue;
1362                 }
1363
1364                 list_add(&req->rq_replay_list, &iter->rq_replay_list);
1365                 return;
1366         }
1367
1368         list_add_tail(&req->rq_replay_list, &imp->imp_replay_list);
1369 }
1370
1371 int ptlrpc_queue_wait(struct ptlrpc_request *req)
1372 {
1373         char str[PTL_NALFMT_SIZE];
1374         int rc = 0;
1375         int brc;
1376         struct l_wait_info lwi;
1377         struct obd_import *imp = req->rq_import;
1378         unsigned long flags;
1379         int timeout = 0;
1380         ENTRY;
1381
1382         LASSERT(req->rq_set == NULL);
1383         LASSERT(!req->rq_receiving_reply);
1384         atomic_inc(&imp->imp_inflight);
1385
1386         if (imp->imp_connection == NULL) {
1387                 CERROR("request on not connected import %s\n",
1388                         imp->imp_obd->obd_name);
1389                 RETURN(-EINVAL);
1390         }
1391
1392         /* for distributed debugging */
1393         req->rq_reqmsg->status = current->pid;
1394         LASSERT(imp->imp_obd != NULL);
1395         CDEBUG(D_RPCTRACE, "Sending RPC pname:cluuid:pid:xid:ni:nid:opc "
1396                "%s:%s:%d:"LPU64":%s:%s:%d\n", current->comm,
1397                imp->imp_obd->obd_uuid.uuid,
1398                req->rq_reqmsg->status, req->rq_xid,
1399                imp->imp_connection->c_peer.peer_ni->pni_name,
1400                ptlrpc_peernid2str(&imp->imp_connection->c_peer, str),
1401                req->rq_reqmsg->opc);
1402
1403         /* Mark phase here for a little debug help */
1404         req->rq_phase = RQ_PHASE_RPC;
1405
1406         spin_lock_irqsave(&imp->imp_lock, flags);
1407         req->rq_import_generation = imp->imp_generation;
1408 restart:
1409         if (ptlrpc_import_delay_req(imp, req, &rc)) {
1410                 list_del(&req->rq_list);
1411
1412                 list_add_tail(&req->rq_list, &imp->imp_delayed_list);
1413                 spin_unlock_irqrestore(&imp->imp_lock, flags);
1414
1415                 DEBUG_REQ(D_HA, req, "\"%s\" waiting for recovery: (%s != %s)",
1416                           current->comm,
1417                           ptlrpc_import_state_name(req->rq_send_state),
1418                           ptlrpc_import_state_name(imp->imp_state));
1419                 lwi = LWI_INTR(interrupted_request, req);
1420                 rc = l_wait_event(req->rq_reply_waitq,
1421                                   (req->rq_send_state == imp->imp_state ||
1422                                    req->rq_err),
1423                                   &lwi);
1424                 DEBUG_REQ(D_HA, req, "\"%s\" awake: (%s == %s or %d == 1)",
1425                           current->comm,
1426                           ptlrpc_import_state_name(imp->imp_state),
1427                           ptlrpc_import_state_name(req->rq_send_state),
1428                           req->rq_err);
1429
1430                 spin_lock_irqsave(&imp->imp_lock, flags);
1431                 list_del_init(&req->rq_list);
1432
1433                 if (req->rq_err) {
1434                         rc = -EIO;
1435                 }
1436                 else if (req->rq_intr) {
1437                         rc = -EINTR;
1438                 }
1439                 else if (req->rq_no_resend) {
1440                         spin_unlock_irqrestore(&imp->imp_lock, flags);
1441                         GOTO(out, rc = -ETIMEDOUT);
1442                 }
1443                 else {
1444                         GOTO(restart, rc);
1445                 }
1446         }
1447
1448         if (rc != 0) {
1449                 list_del_init(&req->rq_list);
1450                 spin_unlock_irqrestore(&imp->imp_lock, flags);
1451                 req->rq_status = rc; // XXX this ok?
1452                 GOTO(out, rc);
1453         }
1454
1455         if (req->rq_resend) {
1456                 if (!req->rq_ptlrpcs_restart)
1457                         lustre_msg_add_flags(req->rq_reqmsg, MSG_RESENT);
1458
1459                 if (req->rq_bulk != NULL)
1460                         ptlrpc_unregister_bulk (req);
1461
1462                 DEBUG_REQ(D_HA, req, "resending: ");
1463         }
1464
1465         /* XXX this is the same as ptlrpc_set_wait */
1466         LASSERT(list_empty(&req->rq_list));
1467         list_add_tail(&req->rq_list, &imp->imp_sending_list);
1468         spin_unlock_irqrestore(&imp->imp_lock, flags);
1469
1470         rc = ptl_send_rpc(req);
1471         if (rc) {
1472                 DEBUG_REQ(D_HA, req, "send failed (%d); recovering", rc);
1473                 timeout = 1;
1474         } else {
1475                 timeout = MAX(req->rq_timeout * HZ, 1);
1476                 DEBUG_REQ(D_NET, req, "-- sleeping for %d jiffies", timeout);
1477         }
1478 repeat:
1479         lwi = LWI_TIMEOUT_INTR(timeout, expired_request, interrupted_request,
1480                                req);
1481         rc = l_wait_event(req->rq_reply_waitq, ptlrpc_check_reply(req), &lwi);
1482         if (rc == -ETIMEDOUT && ptlrpc_check_and_wait_suspend(req))
1483                 goto repeat;
1484         DEBUG_REQ(D_NET, req, "-- done sleeping");
1485
1486         CDEBUG(D_RPCTRACE, "Completed RPC pname:cluuid:pid:xid:ni:nid:opc "
1487                "%s:%s:%d:"LPU64":%s:%s:%d\n", current->comm,
1488                imp->imp_obd->obd_uuid.uuid,
1489                req->rq_reqmsg->status, req->rq_xid,
1490                imp->imp_connection->c_peer.peer_ni->pni_name,
1491                ptlrpc_peernid2str(&imp->imp_connection->c_peer, str),
1492                req->rq_reqmsg->opc);
1493
1494         spin_lock_irqsave(&imp->imp_lock, flags);
1495         list_del_init(&req->rq_list);
1496         spin_unlock_irqrestore(&imp->imp_lock, flags);
1497
1498         /* If the reply was received normally, this just grabs the spinlock
1499          * (ensuring the reply callback has returned), sees that
1500          * req->rq_receiving_reply is clear and returns. */
1501         ptlrpc_unregister_reply (req);
1502
1503         if (req->rq_err)
1504                 GOTO(out, rc = -EIO);
1505
1506         if (req->rq_ptlrpcs_err)
1507                 GOTO(out, rc = -EPERM);
1508
1509         /* Resend if we need to, unless we were interrupted. */
1510         if (req->rq_resend && !req->rq_intr) {
1511                 /* ...unless we were specifically told otherwise. */
1512                 if (req->rq_no_resend)
1513                         GOTO(out, rc = -ETIMEDOUT);
1514                 spin_lock_irqsave(&imp->imp_lock, flags);
1515                 goto restart;
1516         }
1517
1518         if (req->rq_intr) {
1519                 /* Should only be interrupted if we timed out. */
1520                 if (!req->rq_timedout)
1521                         DEBUG_REQ(D_ERROR, req,
1522                                   "rq_intr set but rq_timedout not");
1523                 GOTO(out, rc = -EINTR);
1524         }
1525
1526         if (req->rq_timedout) {                 /* non-recoverable timeout */
1527                 GOTO(out, rc = -ETIMEDOUT);
1528         }
1529
1530         if (!req->rq_replied) {
1531                 /* How can this be? -eeb */
1532                 DEBUG_REQ(D_ERROR, req, "!rq_replied: ");
1533                 LBUG();
1534                 GOTO(out, rc = req->rq_status);
1535         }
1536
1537         rc = after_reply (req);
1538         /* NB may return +ve success rc */
1539         if (req->rq_resend) {
1540                 spin_lock_irqsave(&imp->imp_lock, flags);
1541                 goto restart;
1542         }
1543
1544  out:
1545         if (req->rq_bulk != NULL) {
1546                 if (rc >= 0) {
1547                         /* success so far.  Note that anything going wrong
1548                          * with bulk now, is EXTREMELY strange, since the
1549                          * server must have believed that the bulk
1550                          * tranferred OK before she replied with success to
1551                          * me. */
1552                         lwi = LWI_TIMEOUT(timeout, NULL, NULL);
1553                         brc = l_wait_event(req->rq_reply_waitq,
1554                                            !ptlrpc_bulk_active(req->rq_bulk),
1555                                            &lwi);
1556                         LASSERT(brc == 0 || brc == -ETIMEDOUT);
1557                         if (brc != 0) {
1558                                 LASSERT(brc == -ETIMEDOUT);
1559                                 DEBUG_REQ(D_ERROR, req, "bulk timed out");
1560                                 rc = brc;
1561                         } else if (!req->rq_bulk->bd_success) {
1562                                 DEBUG_REQ(D_ERROR, req, "bulk transfer failed");
1563                                 rc = -EIO;
1564                         }
1565                 }
1566                 if (rc < 0)
1567                         ptlrpc_unregister_bulk (req);
1568         }
1569
1570         LASSERT(!req->rq_receiving_reply);
1571         req->rq_phase = RQ_PHASE_INTERPRET;
1572
1573         atomic_dec(&imp->imp_inflight);
1574         wake_up(&imp->imp_recovery_waitq);
1575         RETURN(rc);
1576 }
1577
1578 struct ptlrpc_replay_async_args {
1579         int praa_old_state;
1580         int praa_old_status;
1581 };
1582
1583 static int ptlrpc_replay_interpret(struct ptlrpc_request *req,
1584                                     void * data, int rc)
1585 {
1586         struct ptlrpc_replay_async_args *aa = data;
1587         struct obd_import *imp = req->rq_import;
1588         unsigned long flags;
1589
1590         atomic_dec(&imp->imp_replay_inflight);
1591
1592         if (!req->rq_replied) {
1593                 CERROR("request replay timed out, restarting recovery\n");
1594                 GOTO(out, rc = -ETIMEDOUT);
1595         }
1596
1597 #if SWAB_PARANOIA
1598         /* Clear reply swab mask; this is a new reply in sender's byte order */
1599         req->rq_rep_swab_mask = 0;
1600 #endif
1601         LASSERT (req->rq_nob_received <= req->rq_repbuf_len);
1602         rc = lustre_unpack_msg(req->rq_repmsg, req->rq_replen);
1603         if (rc) {
1604                 CERROR("unpack_rep failed: %d\n", rc);
1605                 GOTO(out, rc = -EPROTO);
1606         }
1607
1608         if (req->rq_repmsg->type == PTL_RPC_MSG_ERR &&
1609             req->rq_repmsg->status == -ENOTCONN)
1610                 GOTO(out, rc = req->rq_repmsg->status);
1611
1612         /* The transno had better not change over replay. */
1613         LASSERT(req->rq_reqmsg->transno == req->rq_repmsg->transno);
1614
1615         DEBUG_REQ(D_HA, req, "got rep");
1616
1617         /* let the callback do fixups, possibly including in the request */
1618         if (req->rq_replay_cb)
1619                 req->rq_replay_cb(req);
1620
1621         if (req->rq_replied && req->rq_repmsg->status != aa->praa_old_status) {
1622                 DEBUG_REQ(D_HA, req, "status %d, old was %d",
1623                           req->rq_repmsg->status, aa->praa_old_status);
1624         } else {
1625                 /* Put it back for re-replay. */
1626                 req->rq_repmsg->status = aa->praa_old_status;
1627         }
1628
1629         spin_lock_irqsave(&imp->imp_lock, flags);
1630         imp->imp_last_replay_transno = req->rq_transno;
1631         spin_unlock_irqrestore(&imp->imp_lock, flags);
1632
1633         /* continue with recovery */
1634         rc = ptlrpc_import_recovery_state_machine(imp);
1635  out:
1636         req->rq_send_state = aa->praa_old_state;
1637
1638         if (rc != 0)
1639                 /* this replay failed, so restart recovery */
1640                 ptlrpc_connect_import(imp, NULL);
1641
1642         RETURN(rc);
1643 }
1644
1645
1646 int ptlrpc_replay_req(struct ptlrpc_request *req)
1647 {
1648         struct ptlrpc_replay_async_args *aa;
1649         ENTRY;
1650
1651         LASSERT(req->rq_import->imp_state == LUSTRE_IMP_REPLAY);
1652
1653         /* Not handling automatic bulk replay yet (or ever?) */
1654         LASSERT(req->rq_bulk == NULL);
1655
1656         DEBUG_REQ(D_HA, req, "REPLAY");
1657
1658         LASSERT (sizeof (*aa) <= sizeof (req->rq_async_args));
1659         aa = (struct ptlrpc_replay_async_args *)&req->rq_async_args;
1660         memset(aa, 0, sizeof *aa);
1661
1662         /* Prepare request to be resent with ptlrpcd */
1663         aa->praa_old_state = req->rq_send_state;
1664         req->rq_send_state = LUSTRE_IMP_REPLAY;
1665         req->rq_phase = RQ_PHASE_NEW;
1666         aa->praa_old_status = req->rq_repmsg->status;
1667         req->rq_status = 0;
1668
1669         req->rq_interpret_reply = ptlrpc_replay_interpret;
1670         atomic_inc(&req->rq_import->imp_replay_inflight);
1671         ptlrpc_request_addref(req); /* ptlrpcd needs a ref */
1672
1673         ptlrpcd_add_req(req);
1674         RETURN(0);
1675 }
1676
1677 void ptlrpc_abort_inflight(struct obd_import *imp)
1678 {
1679         unsigned long flags;
1680         struct list_head *tmp, *n;
1681         ENTRY;
1682
1683         /* Make sure that no new requests get processed for this import.
1684          * ptlrpc_{queue,set}_wait must (and does) hold imp_lock while testing
1685          * this flag and then putting requests on sending_list or delayed_list.
1686          */
1687         spin_lock_irqsave(&imp->imp_lock, flags);
1688
1689         /* XXX locking?  Maybe we should remove each request with the list
1690          * locked?  Also, how do we know if the requests on the list are
1691          * being freed at this time?
1692          */
1693         list_for_each_safe(tmp, n, &imp->imp_sending_list) {
1694                 struct ptlrpc_request *req =
1695                         list_entry(tmp, struct ptlrpc_request, rq_list);
1696
1697                 DEBUG_REQ(D_HA, req, "inflight");
1698
1699                 spin_lock (&req->rq_lock);
1700                 if (req->rq_import_generation < imp->imp_generation) {
1701                         req->rq_err = 1;
1702                         ptlrpc_wake_client_req(req);
1703                 }
1704                 spin_unlock (&req->rq_lock);
1705         }
1706
1707         list_for_each_safe(tmp, n, &imp->imp_delayed_list) {
1708                 struct ptlrpc_request *req =
1709                         list_entry(tmp, struct ptlrpc_request, rq_list);
1710
1711                 DEBUG_REQ(D_HA, req, "aborting waiting req");
1712
1713                 spin_lock (&req->rq_lock);
1714                 if (req->rq_import_generation < imp->imp_generation) {
1715                         req->rq_err = 1;
1716                         ptlrpc_wake_client_req(req);
1717                 }
1718                 spin_unlock (&req->rq_lock);
1719         }
1720
1721         list_for_each_safe(tmp, n, &imp->imp_rawrpc_list) {
1722                 struct ptlrpc_request *req =
1723                         list_entry(tmp, struct ptlrpc_request, rq_list);
1724
1725                 DEBUG_REQ(D_HA, req, "aborting raw rpc");
1726
1727                 spin_lock(&req->rq_lock);
1728                 req->rq_err = 1;
1729                 ptlrpc_wake_client_req(req);
1730                 spin_unlock(&req->rq_lock);
1731         }
1732
1733         /* Last chance to free reqs left on the replay list, but we
1734          * will still leak reqs that haven't comitted.  */
1735         if (imp->imp_replayable)
1736                 ptlrpc_free_committed(imp);
1737
1738         spin_unlock_irqrestore(&imp->imp_lock, flags);
1739
1740         EXIT;
1741 }
1742
1743 static __u64 ptlrpc_last_xid = 0;
1744 static spinlock_t ptlrpc_last_xid_lock = SPIN_LOCK_UNLOCKED;
1745
1746 __u64 ptlrpc_next_xid(void)
1747 {
1748         __u64 tmp;
1749         spin_lock(&ptlrpc_last_xid_lock);
1750         tmp = ++ptlrpc_last_xid;
1751         spin_unlock(&ptlrpc_last_xid_lock);
1752         return tmp;
1753 }
1754
1755