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 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
  25 #include <linux/obd_support.h>
  26 #include <linux/obd_class.h>
  27 #include <linux/lustre_lib.h>
  28 #include <linux/lustre_ha.h>
  29 #include <linux/lustre_import.h>
  30
  31 void ptlrpc_init_client(int req_portal, int rep_portal, char *name,
  32                         struct ptlrpc_client *cl)
  33 {
  34         cl->cli_request_portal = req_portal;
  35         cl->cli_reply_portal   = rep_portal;
  36         cl->cli_name           = name;
  37 }
  38
  39 __u8 *ptlrpc_req_to_uuid(struct ptlrpc_request *req)
  40 {
  41         return req->rq_connection->c_remote_uuid;
  42 }
  43
  44 struct ptlrpc_connection *ptlrpc_uuid_to_connection(obd_uuid_t uuid)
  45 {
  46         struct ptlrpc_connection *c;
  47         struct lustre_peer peer;
  48         int err;
  49
  50         err = kportal_uuid_to_peer(uuid, &peer);
  51         if (err != 0) {
  52                 CERROR("cannot find peer %s!\n", uuid);
  53                 return NULL;
  54         }
  55
  56         c = ptlrpc_get_connection(&peer, uuid);
  57         if (c) {
  58                 memcpy(c->c_remote_uuid, uuid, sizeof(c->c_remote_uuid));
  59                 c->c_epoch++;
  60         }
  61
  62         CDEBUG(D_INFO, "%s -> %p\n", uuid, c);
  63
  64         return c;
  65 }
  66
  67 void ptlrpc_readdress_connection(struct ptlrpc_connection *conn,obd_uuid_t uuid)
  68 {
  69         struct lustre_peer peer;
  70         int err;
  71
  72         err = kportal_uuid_to_peer(uuid, &peer);
  73         if (err != 0) {
  74                 CERROR("cannot find peer %s!\n", uuid);
  75                 return;
  76         }
  77
  78         memcpy(&conn->c_peer, &peer, sizeof(peer));
  79         return;
  80 }
  81
  82 struct ptlrpc_bulk_desc *ptlrpc_prep_bulk(struct ptlrpc_connection *conn)
  83 {
  84         struct ptlrpc_bulk_desc *desc;
  85
  86         OBD_ALLOC(desc, sizeof(*desc));
  87         if (desc != NULL) {
  88                 desc->bd_connection = ptlrpc_connection_addref(conn);
  89                 atomic_set(&desc->bd_refcount, 1);
  90                 init_waitqueue_head(&desc->bd_waitq);
  91                 INIT_LIST_HEAD(&desc->bd_page_list);
  92                 INIT_LIST_HEAD(&desc->bd_set_chain);
  93                 ptl_set_inv_handle(&desc->bd_md_h);
  94                 ptl_set_inv_handle(&desc->bd_me_h);
  95         }
  96
  97         return desc;
  98 }
  99
 100 int ptlrpc_bulk_error(struct ptlrpc_bulk_desc *desc)
 101 {
 102         int rc = 0;
 103         if (desc->bd_flags & PTL_RPC_FL_TIMEOUT) {
 104                 rc = (desc->bd_flags & PTL_RPC_FL_INTR ? -ERESTARTSYS :
 105                       -ETIMEDOUT);
 106         }
 107         return rc;
 108 }
 109
 110 struct ptlrpc_bulk_page *ptlrpc_prep_bulk_page(struct ptlrpc_bulk_desc *desc)
 111 {
 112         struct ptlrpc_bulk_page *bulk;
 113
 114         OBD_ALLOC(bulk, sizeof(*bulk));
 115         if (bulk != NULL) {
 116                 bulk->bp_desc = desc;
 117                 list_add_tail(&bulk->bp_link, &desc->bd_page_list);
 118                 desc->bd_page_count++;
 119         }
 120         return bulk;
 121 }
 122
 123 void ptlrpc_free_bulk(struct ptlrpc_bulk_desc *desc)
 124 {
 125         struct list_head *tmp, *next;
 126         ENTRY;
 127         if (desc == NULL) {
 128                 EXIT;
 129                 return;
 130         }
 131
 132         LASSERT(list_empty(&desc->bd_set_chain));
 133
 134         if (atomic_read(&desc->bd_refcount) != 0)
 135                 CERROR("freeing desc %p with refcount %d!\n", desc,
 136                        atomic_read(&desc->bd_refcount));
 137
 138         list_for_each_safe(tmp, next, &desc->bd_page_list) {
 139                 struct ptlrpc_bulk_page *bulk;
 140                 bulk = list_entry(tmp, struct ptlrpc_bulk_page, bp_link);
 141                 ptlrpc_free_bulk_page(bulk);
 142         }
 143
 144         ptlrpc_put_connection(desc->bd_connection);
 145
 146         OBD_FREE(desc, sizeof(*desc));
 147         EXIT;
 148 }
 149
 150 void ptlrpc_free_bulk_page(struct ptlrpc_bulk_page *bulk)
 151 {
 152         ENTRY;
 153         if (bulk == NULL) {
 154                 EXIT;
 155                 return;
 156         }
 157
 158         list_del(&bulk->bp_link);
 159         bulk->bp_desc->bd_page_count--;
 160         OBD_FREE(bulk, sizeof(*bulk));
 161         EXIT;
 162 }
 163
 164 static int ll_sync_brw_timeout(void *data)
 165 {
 166         struct obd_brw_set *set = data;
 167         struct list_head *tmp;
 168         int failed = 0;
 169         ENTRY;
 170
 171         LASSERT(set);
 172
 173         set->brw_flags |= PTL_RPC_FL_TIMEOUT;
 174
 175         list_for_each(tmp, &set->brw_desc_head) {
 176                 struct ptlrpc_bulk_desc *desc =
 177                         list_entry(tmp, struct ptlrpc_bulk_desc, bd_set_chain);
 178
 179                 /* Skip descriptors that were completed successfully. */
 180                 if (desc->bd_flags & (PTL_BULK_FL_RCVD | PTL_BULK_FL_SENT))
 181                         continue;
 182
 183                 LASSERT(desc->bd_connection);
 184
 185                 /* If PtlMDUnlink succeeds, then it hasn't completed yet.  If it
 186                  * fails, the bulk finished _just_ in time (after the timeout
 187                  * fired but before we got this far) and we'll let it live.
 188                  */
 189                 if (PtlMDUnlink(desc->bd_md_h) != 0) {
 190                         CERROR("Near-miss on OST %s -- need to adjust "
 191                                "obd_timeout?\n",
 192                                desc->bd_connection->c_remote_uuid);
 193                         continue;
 194                 }
 195
 196                 CERROR("IO of %d pages to/from %s:%d (conn %p) timed out\n",
 197                        desc->bd_page_count, desc->bd_connection->c_remote_uuid,
 198                        desc->bd_portal, desc->bd_connection);
 199
 200                 /* This one will "never" arrive, don't wait for it. */
 201                 if (atomic_dec_and_test(&set->brw_refcount))
 202                         wake_up(&set->brw_waitq);
 203
 204                 if (class_signal_connection_failure)
 205                         class_signal_connection_failure(desc->bd_connection);
 206                 else
 207                         failed = 1;
 208         }
 209
 210         /* 0 = We go back to sleep, until we're resumed or interrupted */
 211         /* 1 = We can't be recovered, just abort the syscall with -ETIMEDOUT */
 212         RETURN(failed);
 213 }
 214
 215 static int ll_sync_brw_intr(void *data)
 216 {
 217         struct obd_brw_set *set = data;
 218
 219         ENTRY;
 220         set->brw_flags |= PTL_RPC_FL_INTR;
 221         RETURN(1); /* ignored, as of this writing */
 222 }
 223
 224 int ll_brw_sync_wait(struct obd_brw_set *set, int phase)
 225 {
 226         struct l_wait_info lwi;
 227         struct list_head *tmp, *next;
 228         int rc = 0;
 229         ENTRY;
 230
 231         switch(phase) {
 232         case CB_PHASE_START:
 233                 lwi = LWI_TIMEOUT_INTR(obd_timeout * HZ, ll_sync_brw_timeout,
 234                                        ll_sync_brw_intr, set);
 235                 rc = l_wait_event(set->brw_waitq,
 236                                   atomic_read(&set->brw_refcount) == 0, &lwi);
 237
 238                 list_for_each_safe(tmp, next, &set->brw_desc_head) {
 239                         struct ptlrpc_bulk_desc *desc =
 240                                 list_entry(tmp, struct ptlrpc_bulk_desc,
 241                                            bd_set_chain);
 242                         list_del_init(&desc->bd_set_chain);
 243                         ptlrpc_bulk_decref(desc);
 244                 }
 245                 break;
 246         case CB_PHASE_FINISH:
 247                 if (atomic_dec_and_test(&set->brw_refcount))
 248                         wake_up(&set->brw_waitq);
 249                 break;
 250         default:
 251                 LBUG();
 252         }
 253
 254         RETURN(rc);
 255 }
 256
 257 struct ptlrpc_request *ptlrpc_prep_req(struct obd_import *imp, int opcode,
 258                                        int count, int *lengths, char **bufs)
 259 {
 260         struct ptlrpc_connection *conn;
 261         struct ptlrpc_request *request;
 262         int rc;
 263         ENTRY;
 264
 265         LASSERT((unsigned long)imp > 0x1000);
 266         conn = imp->imp_connection;
 267
 268         OBD_ALLOC(request, sizeof(*request));
 269         if (!request) {
 270                 CERROR("request allocation out of memory\n");
 271                 RETURN(NULL);
 272         }
 273
 274         rc = lustre_pack_msg(count, lengths, bufs,
 275                              &request->rq_reqlen, &request->rq_reqmsg);
 276         if (rc) {
 277                 CERROR("cannot pack request %d\n", rc);
 278                 OBD_FREE(request, sizeof(*request));
 279                 RETURN(NULL);
 280         }
 281
 282         request->rq_level = LUSTRE_CONN_FULL;
 283         request->rq_type = PTL_RPC_MSG_REQUEST;
 284         request->rq_import = imp;
 285
 286         /* XXX FIXME bug 625069 */
 287         request->rq_request_portal = imp->imp_client->cli_request_portal;
 288         request->rq_reply_portal = imp->imp_client->cli_reply_portal;
 289
 290         request->rq_connection = ptlrpc_connection_addref(conn);
 291
 292         INIT_LIST_HEAD(&request->rq_list);
 293         atomic_set(&request->rq_refcount, 1);
 294
 295         spin_lock(&imp->imp_lock);
 296         request->rq_xid = HTON__u32(++imp->imp_last_xid);
 297         spin_unlock(&imp->imp_lock);
 298
 299         request->rq_reqmsg->magic = PTLRPC_MSG_MAGIC;
 300         request->rq_reqmsg->version = PTLRPC_MSG_VERSION;
 301         request->rq_reqmsg->opc = HTON__u32(opcode);
 302         request->rq_reqmsg->flags = 0;
 303
 304         ptlrpc_hdl2req(request, &imp->imp_handle);
 305         RETURN(request);
 306 }
 307
 308 static void __ptlrpc_free_req(struct ptlrpc_request *request, int locked)
 309 {
 310         ENTRY;
 311         if (request == NULL) {
 312                 EXIT;
 313                 return;
 314         }
 315
 316         if (atomic_read(&request->rq_refcount) != 0) {
 317                 CERROR("freeing request %p (%d->%s:%d) with refcount %d\n",
 318                        request, request->rq_reqmsg->opc,
 319                        request->rq_connection->c_remote_uuid,
 320                        request->rq_import->imp_client->cli_request_portal,
 321                        request->rq_refcount);
 322                 /* LBUG(); */
 323         }
 324
 325         if (request->rq_repmsg != NULL) {
 326                 OBD_FREE(request->rq_repmsg, request->rq_replen);
 327                 request->rq_repmsg = NULL;
 328                 request->rq_reply_md.start = NULL;
 329         }
 330         if (request->rq_reqmsg != NULL) {
 331                 OBD_FREE(request->rq_reqmsg, request->rq_reqlen);
 332                 request->rq_reqmsg = NULL;
 333         }
 334
 335         if (request->rq_import) {
 336                 if (!locked)
 337                         spin_lock(&request->rq_import->imp_lock);
 338                 list_del_init(&request->rq_list);
 339                 if (!locked)
 340                         spin_unlock(&request->rq_import->imp_lock);
 341         }
 342
 343         ptlrpc_put_connection(request->rq_connection);
 344         OBD_FREE(request, sizeof(*request));
 345         EXIT;
 346 }
 347
 348 void ptlrpc_free_req(struct ptlrpc_request *request)
 349 {
 350         __ptlrpc_free_req(request, 0);
 351 }
 352
 353 static int __ptlrpc_req_finished(struct ptlrpc_request *request, int locked)
 354 {
 355         ENTRY;
 356         if (request == NULL)
 357                 RETURN(1);
 358
 359         DEBUG_REQ(D_INFO, request, "refcount now %u",
 360                   atomic_read(&request->rq_refcount) - 1);
 361
 362         if (atomic_dec_and_test(&request->rq_refcount)) {
 363                 __ptlrpc_free_req(request, locked);
 364                 RETURN(1);
 365         }
 366
 367         RETURN(0);
 368 }
 369
 370 void ptlrpc_req_finished(struct ptlrpc_request *request)
 371 {
 372         __ptlrpc_req_finished(request, 0);
 373 }
 374
 375 static int ptlrpc_check_reply(struct ptlrpc_request *req)
 376 {
 377         int rc = 0;
 378
 379         ENTRY;
 380         if (req->rq_repmsg != NULL) {
 381                 req->rq_transno = NTOH__u64(req->rq_repmsg->transno);
 382                 req->rq_flags |= PTL_RPC_FL_REPLIED;
 383                 GOTO(out, rc = 1);
 384         }
 385
 386         if (req->rq_flags & PTL_RPC_FL_RESEND) {
 387                 ENTRY;
 388                 DEBUG_REQ(D_ERROR, req, "RESEND:");
 389                 GOTO(out, rc = 1);
 390         }
 391
 392         if (req->rq_flags & PTL_RPC_FL_ERR) {
 393                 ENTRY;
 394                 DEBUG_REQ(D_ERROR, req, "ABORTED:");
 395                 GOTO(out, rc = 1);
 396         }
 397
 398         if (req->rq_flags & PTL_RPC_FL_RESTART) {
 399                 DEBUG_REQ(D_ERROR, req, "RESTART:");
 400                 GOTO(out, rc = 1);
 401         }
 402         EXIT;
 403  out:
 404         DEBUG_REQ(D_NET, req, "rc = %d for", rc);
 405         return rc;
 406 }
 407
 408 static int ptlrpc_check_status(struct ptlrpc_request *req)
 409 {
 410         int err;
 411         ENTRY;
 412
 413         err = req->rq_repmsg->status;
 414         if (req->rq_repmsg->type == NTOH__u32(PTL_RPC_MSG_ERR)) {
 415                 DEBUG_REQ(D_ERROR, req, "type == PTL_RPC_MSG_ERR (%d)\n", err);
 416                 RETURN(err ? err : -EINVAL);
 417         }
 418
 419         if (err < 0) {
 420                 DEBUG_REQ(D_ERROR, req, "status is %d", err);
 421         } else if (err > 0) {
 422                 /* XXX: translate this error from net to host */
 423                 DEBUG_REQ(D_INFO, req, "status is %d", err);
 424         }
 425
 426         RETURN(err);
 427 }
 428
 429 static void ptlrpc_cleanup_request_buf(struct ptlrpc_request *request)
 430 {
 431         OBD_FREE(request->rq_reqmsg, request->rq_reqlen);
 432         request->rq_reqmsg = NULL;
 433         request->rq_reqlen = 0;
 434 }
 435
 436 /* Abort this request and cleanup any resources associated with it. */
 437 static int ptlrpc_abort(struct ptlrpc_request *request)
 438 {
 439         /* First remove the ME for the reply; in theory, this means
 440          * that we can tear down the buffer safely. */
 441         PtlMEUnlink(request->rq_reply_me_h);
 442         OBD_FREE(request->rq_reply_md.start, request->rq_replen);
 443         request->rq_repmsg = NULL;
 444         request->rq_replen = 0;
 445         return 0;
 446 }
 447
 448 /* caller must hold imp->imp_lock */
 449 void ptlrpc_free_committed(struct obd_import *imp)
 450 {
 451         struct list_head *tmp, *saved;
 452         struct ptlrpc_request *req;
 453         ENTRY;
 454
 455 #ifdef CONFIG_SMP
 456         LASSERT(spin_is_locked(&imp->imp_lock));
 457 #endif
 458
 459         CDEBUG(D_HA, "committing for xid "LPU64", last_committed "LPU64"\n",
 460                imp->imp_peer_last_xid, imp->imp_peer_committed_transno);
 461
 462         list_for_each_safe(tmp, saved, &imp->imp_replay_list) {
 463                 req = list_entry(tmp, struct ptlrpc_request, rq_list);
 464
 465                 if (req->rq_flags & PTL_RPC_FL_REPLAY) {
 466                         DEBUG_REQ(D_HA, req, "keeping (FL_REPLAY)");
 467                         continue;
 468                 }
 469
 470                 /* not yet committed */
 471                 if (req->rq_transno > imp->imp_peer_committed_transno) {
 472                         DEBUG_REQ(D_HA, req, "stopping search");
 473                         break;
 474                 }
 475
 476                 DEBUG_REQ(D_HA, req, "committing (last_committed "LPU64")",
 477                           imp->imp_peer_committed_transno);
 478                 __ptlrpc_req_finished(req, 1);
 479         }
 480
 481         EXIT;
 482         return;
 483 }
 484
 485 void ptlrpc_cleanup_client(struct obd_import *imp)
 486 {
 487         struct list_head *tmp, *saved;
 488         struct ptlrpc_request *req;
 489         struct ptlrpc_connection *conn = imp->imp_connection;
 490         ENTRY;
 491
 492         LASSERT(conn);
 493
 494         spin_lock(&imp->imp_lock);
 495         list_for_each_safe(tmp, saved, &imp->imp_replay_list) {
 496                 req = list_entry(tmp, struct ptlrpc_request, rq_list);
 497
 498                 /* XXX we should make sure that nobody's sleeping on these! */
 499                 DEBUG_REQ(D_HA, req, "cleaning up from sending list");
 500                 list_del_init(&req->rq_list);
 501                 req->rq_import = NULL;
 502                 __ptlrpc_req_finished(req, 0);
 503         }
 504         spin_unlock(&imp->imp_lock);
 505
 506         EXIT;
 507         return;
 508 }
 509
 510 void ptlrpc_continue_req(struct ptlrpc_request *req)
 511 {
 512         ENTRY;
 513         DEBUG_REQ(D_HA, req, "continuing delayed request");
 514         req->rq_reqmsg->addr = req->rq_import->imp_handle.addr;
 515         req->rq_reqmsg->cookie = req->rq_import->imp_handle.cookie;
 516         wake_up(&req->rq_wait_for_rep);
 517         EXIT;
 518 }
 519
 520 void ptlrpc_resend_req(struct ptlrpc_request *req)
 521 {
 522         ENTRY;
 523         DEBUG_REQ(D_HA, req, "resending");
 524         req->rq_reqmsg->addr = req->rq_import->imp_handle.addr;
 525         req->rq_reqmsg->cookie = req->rq_import->imp_handle.cookie;
 526         req->rq_status = -EAGAIN;
 527         req->rq_level = LUSTRE_CONN_RECOVD;
 528         req->rq_flags |= PTL_RPC_FL_RESEND;
 529         req->rq_flags &= ~PTL_RPC_FL_TIMEOUT;
 530         wake_up(&req->rq_wait_for_rep);
 531         EXIT;
 532 }
 533
 534 void ptlrpc_restart_req(struct ptlrpc_request *req)
 535 {
 536         ENTRY;
 537         DEBUG_REQ(D_HA, req, "restarting (possibly-)completed request");
 538         req->rq_status = -ERESTARTSYS;
 539         req->rq_flags |= PTL_RPC_FL_RESTART;
 540         req->rq_flags &= ~PTL_RPC_FL_TIMEOUT;
 541         wake_up(&req->rq_wait_for_rep);
 542         EXIT;
 543 }
 544
 545 static int expired_request(void *data)
 546 {
 547         struct ptlrpc_request *req = data;
 548
 549         ENTRY;
 550         if (!req) {
 551                 CERROR("NULL req!");
 552                 LBUG();
 553                 RETURN(0);
 554         }
 555
 556         DEBUG_REQ(D_ERROR, req, "timeout");
 557         req->rq_flags |= PTL_RPC_FL_TIMEOUT;
 558
 559         if (!req->rq_import) {
 560                 DEBUG_REQ(D_ERROR, req, "NULL import");
 561                 LBUG();
 562                 RETURN(0);
 563         }
 564
 565         if (!req->rq_import->imp_connection) {
 566                 DEBUG_REQ(D_ERROR, req, "NULL connection");
 567                 LBUG();
 568                 RETURN(0);
 569         }
 570
 571         if (!req->rq_import->imp_connection->c_recovd_data.rd_recovd)
 572                 RETURN(1);
 573
 574         req->rq_timeout = 0;
 575         recovd_conn_fail(req->rq_import->imp_connection);
 576
 577 #if 0
 578         /* If this request is for recovery or other primordial tasks,
 579          * don't go back to sleep.
 580          */
 581         if (req->rq_level < LUSTRE_CONN_FULL)
 582                 RETURN(1);
 583 #endif
 584         RETURN(0);
 585 }
 586
 587 static int interrupted_request(void *data)
 588 {
 589         struct ptlrpc_request *req = data;
 590         ENTRY;
 591         req->rq_flags |= PTL_RPC_FL_INTR;
 592         RETURN(1); /* ignored, as of this writing */
 593 }
 594
 595 /* If the import has been invalidated (such as by an OST failure), the
 596  * request must fail with -EIO.
 597  *
 598  * Must be called with imp_lock held, will drop it if it returns -EIO.
 599  */
 600 #define EIO_IF_INVALID(req)                                                   \
 601 if (req->rq_import->imp_flags & IMP_INVALID) {                                \
 602         DEBUG_REQ(D_ERROR, req, "IMP_INVALID:");                              \
 603         spin_unlock(&imp->imp_lock);                                          \
 604         RETURN(-EIO);                                                         \
 605 }
 606
 607 int ptlrpc_queue_wait(struct ptlrpc_request *req)
 608 {
 609         int rc = 0;
 610         struct l_wait_info lwi;
 611         struct obd_import *imp = req->rq_import;
 612         struct ptlrpc_connection *conn = imp->imp_connection;
 613         ENTRY;
 614
 615         init_waitqueue_head(&req->rq_wait_for_rep);
 616
 617         /* for distributed debugging */
 618         req->rq_reqmsg->status = HTON__u32(current->pid);
 619         CDEBUG(D_RPCTRACE, "Sending RPC pid:xid:nid:opc %d:"LPU64":%x:%d\n",
 620                NTOH__u32(req->rq_reqmsg->status), req->rq_xid,
 621                conn->c_peer.peer_nid, NTOH__u32(req->rq_reqmsg->opc));
 622
 623         spin_lock(&imp->imp_lock);
 624         EIO_IF_INVALID(req);
 625         if (req->rq_level > imp->imp_level) {
 626                 list_del(&req->rq_list);
 627                 list_add_tail(&req->rq_list, &imp->imp_delayed_list);
 628                 spin_unlock(&imp->imp_lock);
 629
 630                 DEBUG_REQ(D_HA, req, "\"%s\" waiting for recovery: (%d < %d)",
 631                           current->comm, req->rq_level, imp->imp_level);
 632                 lwi = LWI_INTR(NULL, NULL);
 633                 rc = l_wait_event(req->rq_wait_for_rep,
 634                                   (req->rq_level <= imp->imp_level) ||
 635                                   (req->rq_flags & PTL_RPC_FL_ERR), &lwi);
 636
 637                 spin_lock(&imp->imp_lock);
 638                 list_del_init(&req->rq_list);
 639                 spin_unlock(&imp->imp_lock);
 640
 641                 if (req->rq_flags & PTL_RPC_FL_ERR)
 642                         RETURN(-EIO);
 643
 644                 if (rc)
 645                         RETURN(rc);
 646
 647                 CERROR("process %d resumed\n", current->pid);
 648         }
 649  resend:
 650
 651         LASSERT(list_empty(&req->rq_list));
 652         list_add_tail(&req->rq_list, &imp->imp_sending_list);
 653         spin_unlock(&imp->imp_lock);
 654         rc = ptl_send_rpc(req);
 655         if (rc) {
 656                 CDEBUG(D_HA, "error %d, opcode %d, need recovery\n", rc,
 657                        req->rq_reqmsg->opc);
 658                 /* sleep for a jiffy, then trigger recovery */
 659                 lwi = LWI_TIMEOUT_INTR(1, expired_request,
 660                                        interrupted_request, req);
 661         } else {
 662                 DEBUG_REQ(D_NET, req, "-- sleeping");
 663                 lwi = LWI_TIMEOUT_INTR(req->rq_timeout * HZ, expired_request,
 664                                        interrupted_request, req);
 665         }
 666         l_wait_event(req->rq_wait_for_rep, ptlrpc_check_reply(req), &lwi);
 667         DEBUG_REQ(D_NET, req, "-- done sleeping");
 668
 669         spin_lock(&imp->imp_lock);
 670         list_del_init(&req->rq_list);
 671         spin_unlock(&imp->imp_lock);
 672
 673         if (req->rq_flags & PTL_RPC_FL_ERR) {
 674                 ptlrpc_abort(req);
 675                 GOTO(out, rc = -EIO);
 676         }
 677
 678         /* Don't resend if we were interrupted. */
 679         if ((req->rq_flags & (PTL_RPC_FL_RESEND | PTL_RPC_FL_INTR)) ==
 680             PTL_RPC_FL_RESEND) {
 681                 req->rq_flags &= ~PTL_RPC_FL_RESEND;
 682                 lustre_msg_add_flags(req->rq_reqmsg, MSG_RESENT);
 683                 DEBUG_REQ(D_HA, req, "resending: ");
 684                 goto resend;
 685         }
 686
 687         if (req->rq_flags & PTL_RPC_FL_INTR) {
 688                 if (!(req->rq_flags & PTL_RPC_FL_TIMEOUT))
 689                         LBUG(); /* should only be interrupted if we timed out */
 690                 /* Clean up the dangling reply buffers */
 691                 ptlrpc_abort(req);
 692                 GOTO(out, rc = -EINTR);
 693         }
 694
 695         if (req->rq_flags & PTL_RPC_FL_TIMEOUT)
 696                 GOTO(out, rc = -ETIMEDOUT);
 697
 698         if (!(req->rq_flags & PTL_RPC_FL_REPLIED))
 699                 GOTO(out, rc = req->rq_status);
 700
 701         rc = lustre_unpack_msg(req->rq_repmsg, req->rq_replen);
 702         if (rc) {
 703                 CERROR("unpack_rep failed: %d\n", rc);
 704                 GOTO(out, rc);
 705         }
 706 #if 0
 707         /* FIXME: Enable when BlueArc makes new release */
 708         if (req->rq_repmsg->type != PTL_RPC_MSG_REPLY &&
 709             req->rq_repmsg->type != PTL_RPC_MSG_ERR) {
 710                 CERROR("invalid packet type received (type=%u)\n",
 711                        req->rq_repmsg->type);
 712                 LBUG();
 713                 GOTO(out, rc = -EINVAL);
 714         }
 715 #endif
 716         DEBUG_REQ(D_NET, req, "status %d\n", req->rq_repmsg->status);
 717
 718         /* We're a rejected connection, need to invalidate and rebuild. */
 719         if (req->rq_repmsg->status == -ENOTCONN) {
 720                 spin_lock(&imp->imp_lock);
 721                 /* If someone else is reconnecting us (CONN_RECOVD) or has
 722                  * already completed it (handle mismatch), then we just need
 723                  * to get out.
 724                  */
 725                 if (imp->imp_level == LUSTRE_CONN_RECOVD ||
 726                     imp->imp_handle.addr != req->rq_reqmsg->addr ||
 727                     imp->imp_handle.cookie != req->rq_reqmsg->cookie) {
 728                         spin_unlock(&imp->imp_lock);
 729                         GOTO(out, rc = -EIO);
 730                 }
 731                 imp->imp_level = LUSTRE_CONN_RECOVD;
 732                 spin_unlock(&imp->imp_lock);
 733                 rc = imp->imp_recover(imp, PTLRPC_RECOVD_PHASE_NOTCONN);
 734                 if (rc)
 735                         LBUG();
 736                 GOTO(out, rc = -EIO);
 737         }
 738
 739         if (req->rq_import->imp_flags & IMP_REPLAYABLE) {
 740                 spin_lock(&imp->imp_lock);
 741                 if (req->rq_flags & PTL_RPC_FL_REPLAY || req->rq_transno != 0) {
 742                         /* Balanced in ptlrpc_free_committed, usually. */
 743                         atomic_inc(&req->rq_refcount);
 744                         list_add_tail(&req->rq_list, &imp->imp_replay_list);
 745                 }
 746
 747                 if (req->rq_transno > imp->imp_max_transno) {
 748                         imp->imp_max_transno = req->rq_transno;
 749                 } else if (req->rq_transno != 0 &&
 750                            imp->imp_level == LUSTRE_CONN_FULL) {
 751                         CDEBUG(D_HA, "got transno "LPD64" after "LPD64
 752                                ": recovery may not work\n", req->rq_transno,
 753                                imp->imp_max_transno);
 754                 }
 755
 756                 /* Replay-enabled imports return commit-status information. */
 757                 imp->imp_peer_last_xid = req->rq_repmsg->last_xid;
 758                 imp->imp_peer_committed_transno =
 759                         req->rq_repmsg->last_committed;
 760                 ptlrpc_free_committed(imp);
 761                 spin_unlock(&imp->imp_lock);
 762         }
 763
 764         rc = ptlrpc_check_status(req);
 765
 766         EXIT;
 767  out:
 768         return rc;
 769 }
 770
 771 #undef EIO_IF_INVALID
 772
 773 int ptlrpc_replay_req(struct ptlrpc_request *req)
 774 {
 775         int rc = 0, old_level, old_status = 0;
 776         // struct ptlrpc_client *cli = req->rq_import->imp_client;
 777         struct l_wait_info lwi;
 778         ENTRY;
 779
 780         init_waitqueue_head(&req->rq_wait_for_rep);
 781         DEBUG_REQ(D_NET, req, "");
 782
 783         req->rq_timeout = obd_timeout;
 784         req->rq_reqmsg->addr = req->rq_import->imp_handle.addr;
 785         req->rq_reqmsg->cookie = req->rq_import->imp_handle.cookie;
 786
 787         /* temporarily set request to RECOVD level (reset at out:) */
 788         old_level = req->rq_level;
 789         if (req->rq_flags & PTL_RPC_FL_REPLIED)
 790                 old_status = req->rq_repmsg->status;
 791         req->rq_level = LUSTRE_CONN_RECOVD;
 792         rc = ptl_send_rpc(req);
 793         if (rc) {
 794                 CERROR("error %d, opcode %d\n", rc, req->rq_reqmsg->opc);
 795                 ptlrpc_cleanup_request_buf(req);
 796                 // up(&cli->cli_rpc_sem);
 797                 GOTO(out, rc = -rc);
 798         }
 799
 800         CDEBUG(D_OTHER, "-- sleeping\n");
 801         lwi = LWI_INTR(NULL, NULL); /* XXX needs timeout, nested recovery */
 802         l_wait_event(req->rq_wait_for_rep, ptlrpc_check_reply(req), &lwi);
 803         CDEBUG(D_OTHER, "-- done\n");
 804
 805         // up(&cli->cli_rpc_sem);
 806
 807         if (!(req->rq_flags & PTL_RPC_FL_REPLIED)) {
 808                 CERROR("Unknown reason for wakeup\n");
 809                 /* XXX Phil - I end up here when I kill obdctl */
 810                 ptlrpc_abort(req);
 811                 GOTO(out, rc = -EINTR);
 812         }
 813
 814         rc = lustre_unpack_msg(req->rq_repmsg, req->rq_replen);
 815         if (rc) {
 816                 CERROR("unpack_rep failed: %d\n", rc);
 817                 GOTO(out, rc);
 818         }
 819
 820         CDEBUG(D_NET, "got rep "LPD64"\n", req->rq_xid);
 821
 822         /* let the callback do fixups, possibly including in the request */
 823         if (req->rq_replay_cb)
 824                 req->rq_replay_cb(req);
 825
 826         if ((req->rq_flags & PTL_RPC_FL_REPLIED) &&
 827             req->rq_repmsg->status != old_status) {
 828                 DEBUG_REQ(D_HA, req, "status %d, old was %d",
 829                           req->rq_repmsg->status, old_status);
 830         }
 831
 832  out:
 833         req->rq_level = old_level;
 834         RETURN(rc);
 835 }
 836
 837 /* XXX looks a lot like super.c:invalidate_request_list, don't it? */
 838 void ptlrpc_abort_inflight(struct obd_import *imp)
 839 {
 840         struct list_head *tmp, *n;
 841
 842         /* Make sure that no new requests get processed for this import.
 843          * ptlrpc_queue_wait must (and does) hold imp_lock while testing this
 844          * flag and then putting requests on sending_list or delayed_list.
 845          */
 846         spin_lock(&imp->imp_lock);
 847         imp->imp_flags |= IMP_INVALID;
 848         spin_unlock(&imp->imp_lock);
 849
 850         list_for_each_safe(tmp, n, &imp->imp_sending_list) {
 851                 struct ptlrpc_request *req =
 852                         list_entry(tmp, struct ptlrpc_request, rq_list);
 853
 854                 DEBUG_REQ(D_HA, req, "inflight");
 855                 req->rq_flags |= PTL_RPC_FL_ERR;
 856                 wake_up(&req->rq_wait_for_rep);
 857         }
 858
 859         list_for_each_safe(tmp, n, &imp->imp_delayed_list) {
 860                 struct ptlrpc_request *req =
 861                         list_entry(tmp, struct ptlrpc_request, rq_list);
 862
 863                 DEBUG_REQ(D_HA, req, "aborting waiting req");
 864                 req->rq_flags |= PTL_RPC_FL_ERR;
 865                 wake_up(&req->rq_wait_for_rep);
 866         }
 867 }