lustre/ldlm/ldlm_lib.c

   1 /* -*- mode: c; c-basic-offset: 8; indent-tabs-mode: nil; -*-
   2  * vim:expandtab:shiftwidth=8:tabstop=8:
   3  *
   4  *  Copyright (c) 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 #ifndef EXPORT_SYMTAB
  23 # define EXPORT_SYMTAB
  24 #endif
  25 #define DEBUG_SUBSYSTEM S_LDLM
  26
  27 #ifdef __KERNEL__
  28 # include <linux/module.h>
  29 #else
  30 # include <liblustre.h>
  31 #endif
  32 #include <linux/obd.h>
  33 #include <linux/obd_ost.h> /* for LUSTRE_OSC_NAME */
  34 #include <linux/lustre_mds.h> /* for LUSTRE_MDC_NAME */
  35 #include <linux/lustre_mgmt.h>
  36 #include <linux/lustre_dlm.h>
  37 #include <linux/lustre_net.h>
  38
  39 int client_obd_setup(struct obd_device *obddev, obd_count len, void *buf)
  40 {
  41         struct ptlrpc_connection *conn;
  42         struct lustre_cfg* lcfg = buf;
  43         struct client_obd *cli = &obddev->u.cli;
  44         struct obd_import *imp;
  45         struct obd_uuid server_uuid;
  46         int rq_portal, rp_portal, connect_op;
  47         char *name = obddev->obd_type->typ_name;
  48         char *mgmt_name = NULL;
  49         int rc = 0;
  50         struct obd_device *mgmt_obd;
  51         mgmtcli_register_for_events_t register_f;
  52         ENTRY;
  53
  54         /* In a more perfect world, we would hang a ptlrpc_client off of
  55          * obd_type and just use the values from there. */
  56         if (!strcmp(name, LUSTRE_OSC_NAME)) {
  57                 rq_portal = OST_REQUEST_PORTAL;
  58                 rp_portal = OSC_REPLY_PORTAL;
  59                 connect_op = OST_CONNECT;
  60         } else if (!strcmp(name, LUSTRE_MDC_NAME)) {
  61                 rq_portal = MDS_REQUEST_PORTAL;
  62                 rp_portal = MDC_REPLY_PORTAL;
  63                 connect_op = MDS_CONNECT;
  64         } else if (!strcmp(name, LUSTRE_MGMTCLI_NAME)) {
  65                 rq_portal = MGMT_REQUEST_PORTAL;
  66                 rp_portal = MGMT_REPLY_PORTAL;
  67                 connect_op = MGMT_CONNECT;
  68         } else {
  69                 CERROR("unknown client OBD type \"%s\", can't setup\n",
  70                        name);
  71                 RETURN(-EINVAL);
  72         }
  73
  74         if (lcfg->lcfg_inllen1 < 1) {
  75                 CERROR("requires a TARGET UUID\n");
  76                 RETURN(-EINVAL);
  77         }
  78
  79         if (lcfg->lcfg_inllen1 > 37) {
  80                 CERROR("client UUID must be less than 38 characters\n");
  81                 RETURN(-EINVAL);
  82         }
  83
  84         if (lcfg->lcfg_inllen2 < 1) {
  85                 CERROR("setup requires a SERVER UUID\n");
  86                 RETURN(-EINVAL);
  87         }
  88
  89         if (lcfg->lcfg_inllen2 > 37) {
  90                 CERROR("target UUID must be less than 38 characters\n");
  91                 RETURN(-EINVAL);
  92         }
  93
  94         sema_init(&cli->cl_sem, 1);
  95         cli->cl_conn_count = 0;
  96         memcpy(server_uuid.uuid, lcfg->lcfg_inlbuf2,
  97                min_t(unsigned int, lcfg->lcfg_inllen2, sizeof(server_uuid)));
  98
  99         cli->cl_dirty = 0;
 100         cli->cl_avail_grant = 0;
 101         cli->cl_dirty_max = OSC_MAX_DIRTY_DEFAULT * 1024 * 1024;
 102         INIT_LIST_HEAD(&cli->cl_cache_waiters);
 103         INIT_LIST_HEAD(&cli->cl_loi_ready_list);
 104         INIT_LIST_HEAD(&cli->cl_loi_write_list);
 105         spin_lock_init(&cli->cl_loi_list_lock);
 106         cli->cl_brw_in_flight = 0;
 107         spin_lock_init(&cli->cl_read_rpc_hist.oh_lock);
 108         spin_lock_init(&cli->cl_write_rpc_hist.oh_lock);
 109         spin_lock_init(&cli->cl_read_page_hist.oh_lock);
 110         spin_lock_init(&cli->cl_write_page_hist.oh_lock);
 111         cli->cl_max_pages_per_rpc = PTL_MD_MAX_PAGES;
 112         cli->cl_max_rpcs_in_flight = OSC_MAX_RIF_DEFAULT;
 113
 114         ldlm_get_ref();
 115         if (rc) {
 116                 CERROR("ldlm_get_ref failed: %d\n", rc);
 117                 GOTO(err, rc);
 118         }
 119
 120         conn = ptlrpc_uuid_to_connection(&server_uuid);
 121         if (conn == NULL)
 122                 GOTO(err_ldlm, rc = -ENOENT);
 123
 124         ptlrpc_init_client(rq_portal, rp_portal, name,
 125                            &obddev->obd_ldlm_client);
 126
 127         imp = class_new_import();
 128         if (imp == NULL) {
 129                 ptlrpc_put_connection(conn);
 130                 GOTO(err_ldlm, rc = -ENOENT);
 131         }
 132         imp->imp_connection = conn;
 133         imp->imp_client = &obddev->obd_ldlm_client;
 134         imp->imp_obd = obddev;
 135         imp->imp_connect_op = connect_op;
 136         imp->imp_generation = 0;
 137         imp->imp_initial_recov = 1;
 138         INIT_LIST_HEAD(&imp->imp_pinger_chain);
 139         memcpy(imp->imp_target_uuid.uuid, lcfg->lcfg_inlbuf1,
 140               lcfg->lcfg_inllen1);
 141         class_import_put(imp);
 142
 143         cli->cl_import = imp;
 144         cli->cl_max_mds_easize = sizeof(struct lov_mds_md);
 145         cli->cl_max_mds_cookiesize = sizeof(struct llog_cookie);
 146         cli->cl_sandev = to_kdev_t(0);
 147
 148         if (lcfg->lcfg_inllen3 != 0) {
 149                 if (!strcmp(lcfg->lcfg_inlbuf3, "inactive")) {
 150                         CDEBUG(D_HA, "marking %s %s->%s as inactive\n",
 151                                name, obddev->obd_name,
 152                                imp->imp_target_uuid.uuid);
 153                         imp->imp_invalid = 1;
 154
 155                         if (lcfg->lcfg_inllen4 != 0)
 156                                 mgmt_name = lcfg->lcfg_inlbuf4;
 157                 } else {
 158                         mgmt_name = lcfg->lcfg_inlbuf3;
 159                 }
 160         }
 161
 162         if (mgmt_name != NULL) {
 163                 /* Register with management client if we need to. */
 164                 CDEBUG(D_HA, "%s registering with %s for events about %s\n",
 165                        obddev->obd_name, mgmt_name, server_uuid.uuid);
 166
 167                 mgmt_obd = class_name2obd(mgmt_name);
 168                 if (!mgmt_obd) {
 169                         CERROR("can't find mgmtcli %s to register\n",
 170                                mgmt_name);
 171                         GOTO(err_import, rc = -ENOSYS);
 172                 }
 173
 174                 register_f = inter_module_get("mgmtcli_register_for_events");
 175                 if (!register_f) {
 176                         CERROR("can't i_m_g mgmtcli_register_for_events\n");
 177                         GOTO(err_import, rc = -ENOSYS);
 178                 }
 179
 180                 rc = register_f(mgmt_obd, obddev, &imp->imp_target_uuid);
 181                 inter_module_put("mgmtcli_register_for_events");
 182
 183                 if (!rc)
 184                         cli->cl_mgmtcli_obd = mgmt_obd;
 185         }
 186
 187         RETURN(rc);
 188
 189 err_import:
 190         class_destroy_import(imp);
 191 err_ldlm:
 192         ldlm_put_ref(0);
 193 err:
 194         RETURN(rc);
 195
 196 }
 197
 198 int client_obd_cleanup(struct obd_device *obddev, int flags)
 199 {
 200         struct client_obd *cli = &obddev->u.cli;
 201
 202         if (!cli->cl_import)
 203                 RETURN(-EINVAL);
 204         if (cli->cl_mgmtcli_obd) {
 205                 mgmtcli_deregister_for_events_t dereg_f;
 206
 207                 dereg_f = inter_module_get("mgmtcli_deregister_for_events");
 208                 dereg_f(cli->cl_mgmtcli_obd, obddev);
 209                 inter_module_put("mgmtcli_deregister_for_events");
 210         }
 211         class_destroy_import(cli->cl_import);
 212         cli->cl_import = NULL;
 213
 214         ldlm_put_ref(flags & OBD_OPT_FORCE);
 215
 216         RETURN(0);
 217 }
 218
 219 int client_connect_import(struct lustre_handle *dlm_handle,
 220                           struct obd_device *obd,
 221                           struct obd_uuid *cluuid)
 222 {
 223         struct client_obd *cli = &obd->u.cli;
 224         struct obd_import *imp = cli->cl_import;
 225         struct obd_export *exp;
 226         int rc;
 227         ENTRY;
 228
 229         down(&cli->cl_sem);
 230         rc = class_connect(dlm_handle, obd, cluuid);
 231         if (rc)
 232                 GOTO(out_sem, rc);
 233
 234         cli->cl_conn_count++;
 235         if (cli->cl_conn_count > 1)
 236                 GOTO(out_sem, rc);
 237         exp = class_conn2export(dlm_handle);
 238
 239         if (obd->obd_namespace != NULL)
 240                 CERROR("already have namespace!\n");
 241         obd->obd_namespace = ldlm_namespace_new(obd->obd_name,
 242                                                 LDLM_NAMESPACE_CLIENT);
 243         if (obd->obd_namespace == NULL)
 244                 GOTO(out_disco, rc = -ENOMEM);
 245
 246         imp->imp_dlm_handle = *dlm_handle;
 247         rc = ptlrpc_init_import(imp);
 248         if (rc != 0)
 249                 GOTO(out_ldlm, rc);
 250
 251         exp->exp_connection = ptlrpc_connection_addref(imp->imp_connection);
 252         rc = ptlrpc_connect_import(imp, NULL);
 253         if (rc != 0) {
 254                 LASSERT (imp->imp_state == LUSTRE_IMP_DISCON);
 255                 GOTO(out_ldlm, rc);
 256         }
 257
 258         ptlrpc_pinger_add_import(imp);
 259         EXIT;
 260
 261         if (rc) {
 262 out_ldlm:
 263                 ldlm_namespace_free(obd->obd_namespace, 0);
 264                 obd->obd_namespace = NULL;
 265 out_disco:
 266                 cli->cl_conn_count--;
 267                 class_disconnect(exp, 0);
 268         } else {
 269                 class_export_put(exp);
 270         }
 271 out_sem:
 272         up(&cli->cl_sem);
 273         return rc;
 274 }
 275
 276 int client_disconnect_export(struct obd_export *exp, int failover)
 277 {
 278         struct obd_device *obd = class_exp2obd(exp);
 279         struct client_obd *cli = &obd->u.cli;
 280         struct obd_import *imp = cli->cl_import;
 281         int rc = 0, err;
 282         ENTRY;
 283
 284         if (!obd) {
 285                 CERROR("invalid export for disconnect: exp %p cookie "LPX64"\n",
 286                        exp, exp ? exp->exp_handle.h_cookie : -1);
 287                 RETURN(-EINVAL);
 288         }
 289
 290         down(&cli->cl_sem);
 291         if (!cli->cl_conn_count) {
 292                 CERROR("disconnecting disconnected device (%s)\n",
 293                        obd->obd_name);
 294                 GOTO(out_sem, rc = -EINVAL);
 295         }
 296
 297         cli->cl_conn_count--;
 298         if (cli->cl_conn_count)
 299                 GOTO(out_no_disconnect, rc = 0);
 300
 301         /* Some non-replayable imports (MDS's OSCs) are pinged, so just
 302          * delete it regardless.  (It's safe to delete an import that was
 303          * never added.) */
 304         (void)ptlrpc_pinger_del_import(imp);
 305
 306         if (obd->obd_namespace != NULL) {
 307                 /* obd_no_recov == local only */
 308                 ldlm_cli_cancel_unused(obd->obd_namespace, NULL,
 309                                        obd->obd_no_recov, NULL);
 310                 ldlm_namespace_free(obd->obd_namespace, obd->obd_no_recov);
 311                 obd->obd_namespace = NULL;
 312         }
 313
 314         /* Yeah, obd_no_recov also (mainly) means "forced shutdown". */
 315         if (obd->obd_no_recov)
 316                 ptlrpc_invalidate_import(imp);
 317         else
 318                 rc = ptlrpc_disconnect_import(imp);
 319
 320         EXIT;
 321  out_no_disconnect:
 322         err = class_disconnect(exp, 0);
 323         if (!rc && err)
 324                 rc = err;
 325  out_sem:
 326         up(&cli->cl_sem);
 327         RETURN(rc);
 328 }
 329
 330 /* --------------------------------------------------------------------------
 331  * from old lib/target.c
 332  * -------------------------------------------------------------------------- */
 333
 334 int target_handle_reconnect(struct lustre_handle *conn, struct obd_export *exp,
 335                             struct obd_uuid *cluuid)
 336 {
 337         if (exp->exp_connection) {
 338                 struct lustre_handle *hdl;
 339                 hdl = &exp->exp_imp_reverse->imp_remote_handle;
 340                 /* Might be a re-connect after a partition. */
 341                 if (!memcmp(&conn->cookie, &hdl->cookie, sizeof conn->cookie)) {
 342                         CERROR("%s reconnecting\n", cluuid->uuid);
 343                         conn->cookie = exp->exp_handle.h_cookie;
 344                         RETURN(EALREADY);
 345                 } else {
 346                         CERROR("%s reconnecting from %s, "
 347                                "handle mismatch (ours "LPX64", theirs "
 348                                LPX64")\n", cluuid->uuid,
 349                                exp->exp_connection->c_remote_uuid.uuid,
 350                                hdl->cookie, conn->cookie);
 351                         memset(conn, 0, sizeof *conn);
 352                         RETURN(-EALREADY);
 353                 }
 354         }
 355
 356         conn->cookie = exp->exp_handle.h_cookie;
 357         CDEBUG(D_INFO, "existing export for UUID '%s' at %p\n",
 358                cluuid->uuid, exp);
 359         CDEBUG(D_IOCTL,"connect: cookie "LPX64"\n", conn->cookie);
 360         RETURN(0);
 361 }
 362
 363 int target_handle_connect(struct ptlrpc_request *req, svc_handler_t handler)
 364 {
 365         struct obd_device *target;
 366         struct obd_export *export = NULL;
 367         struct obd_import *revimp;
 368         struct lustre_handle conn;
 369         struct obd_uuid tgtuuid;
 370         struct obd_uuid cluuid;
 371         struct obd_uuid remote_uuid;
 372         struct list_head *p;
 373         char *str, *tmp;
 374         int rc = 0, abort_recovery;
 375         unsigned long flags;
 376         ENTRY;
 377
 378         OBD_RACE(OBD_FAIL_TGT_CONN_RACE);
 379
 380         LASSERT_REQSWAB (req, 0);
 381         str = lustre_msg_string(req->rq_reqmsg, 0, sizeof(tgtuuid) - 1);
 382         if (str == NULL) {
 383                 CERROR("bad target UUID for connect\n");
 384                 GOTO(out, rc = -EINVAL);
 385         }
 386
 387         obd_str2uuid (&tgtuuid, str);
 388         target = class_uuid2obd(&tgtuuid);
 389         if (!target) {
 390                 target = class_name2obd(str);
 391         }
 392
 393         if (!target || target->obd_stopping || !target->obd_set_up) {
 394                 CERROR("UUID '%s' is not available for connect\n", str);
 395                 GOTO(out, rc = -ENODEV);
 396         }
 397
 398         LASSERT_REQSWAB (req, 1);
 399         str = lustre_msg_string(req->rq_reqmsg, 1, sizeof(cluuid) - 1);
 400         if (str == NULL) {
 401                 CERROR("bad client UUID for connect\n");
 402                 GOTO(out, rc = -EINVAL);
 403         }
 404
 405         obd_str2uuid (&cluuid, str);
 406
 407         /* XXX extract a nettype and format accordingly */
 408         snprintf(remote_uuid.uuid, sizeof remote_uuid,
 409                  "NET_"LPX64"_UUID", req->rq_peer.peer_nid);
 410
 411         spin_lock_bh(&target->obd_processing_task_lock);
 412         abort_recovery = target->obd_abort_recovery;
 413         spin_unlock_bh(&target->obd_processing_task_lock);
 414         if (abort_recovery)
 415                 target_abort_recovery(target);
 416
 417         tmp = lustre_msg_buf(req->rq_reqmsg, 2, sizeof conn);
 418         if (tmp == NULL)
 419                 GOTO(out, rc = -EPROTO);
 420
 421         memcpy(&conn, tmp, sizeof conn);
 422
 423         rc = lustre_pack_reply(req, 0, NULL, NULL);
 424         if (rc)
 425                 GOTO(out, rc);
 426
 427         /* lctl gets a backstage, all-access pass. */
 428         if (obd_uuid_equals(&cluuid, &target->obd_uuid))
 429                 goto dont_check_exports;
 430
 431         spin_lock(&target->obd_dev_lock);
 432         list_for_each(p, &target->obd_exports) {
 433                 export = list_entry(p, struct obd_export, exp_obd_chain);
 434                 if (obd_uuid_equals(&cluuid, &export->exp_client_uuid)) {
 435                         spin_unlock(&target->obd_dev_lock);
 436                         LASSERT(export->exp_obd == target);
 437
 438                         rc = target_handle_reconnect(&conn, export, &cluuid);
 439                         break;
 440                 }
 441                 export = NULL;
 442         }
 443         /* If we found an export, we already unlocked. */
 444         if (!export) {
 445                 spin_unlock(&target->obd_dev_lock);
 446         } else if (req->rq_reqmsg->conn_cnt == 1) {
 447                 CERROR("%s reconnected with 1 conn_cnt; cookies not random?\n",
 448                        cluuid.uuid);
 449                 GOTO(out, rc = -EALREADY);
 450         }
 451
 452         /* Tell the client if we're in recovery. */
 453         /* If this is the first client, start the recovery timer */
 454         if (target->obd_recovering) {
 455                 lustre_msg_add_op_flags(req->rq_repmsg, MSG_CONNECT_RECOVERING);
 456                 target_start_recovery_timer(target, handler);
 457         }
 458
 459         /* Tell the client if we support replayable requests */
 460         if (target->obd_replayable)
 461                 lustre_msg_add_op_flags(req->rq_repmsg, MSG_CONNECT_REPLAYABLE);
 462
 463         if (export == NULL) {
 464                 if (target->obd_recovering) {
 465                         CERROR("denying connection for new client %s: "
 466                                "%d clients in recovery for %lds\n", cluuid.uuid,
 467                                target->obd_recoverable_clients,
 468                                (target->obd_recovery_timer.expires-jiffies)/HZ);
 469                         rc = -EBUSY;
 470                 } else {
 471  dont_check_exports:
 472                         rc = obd_connect(&conn, target, &cluuid);
 473                 }
 474         }
 475
 476
 477         /* If all else goes well, this is our RPC return code. */
 478         req->rq_status = 0;
 479
 480         if (rc && rc != EALREADY)
 481                 GOTO(out, rc);
 482
 483         /* XXX track this all the time? */
 484         if (target->obd_recovering) {
 485                 target->obd_connected_clients++;
 486         }
 487
 488         req->rq_repmsg->handle = conn;
 489
 490         /* If the client and the server are the same node, we will already
 491          * have an export that really points to the client's DLM export,
 492          * because we have a shared handles table.
 493          *
 494          * XXX this will go away when shaver stops sending the "connect" handle
 495          * in the real "remote handle" field of the request --phik 24 Apr 2003
 496          */
 497         if (req->rq_export != NULL)
 498                 class_export_put(req->rq_export);
 499
 500         /* ownership of this export ref transfers to the request */
 501         export = req->rq_export = class_conn2export(&conn);
 502         LASSERT(export != NULL);
 503
 504         spin_lock_irqsave(&export->exp_lock, flags);
 505         if (export->exp_conn_cnt >= req->rq_reqmsg->conn_cnt) {
 506                 CERROR("%s: already connected at a higher conn_cnt: %d > %d\n",
 507                        cluuid.uuid, export->exp_conn_cnt,
 508                        req->rq_reqmsg->conn_cnt);
 509                 spin_unlock_irqrestore(&export->exp_lock, flags);
 510                 GOTO(out, rc = -EALREADY);
 511         }
 512         export->exp_conn_cnt = req->rq_reqmsg->conn_cnt;
 513         spin_unlock_irqrestore(&export->exp_lock, flags);
 514
 515         /* request from liblustre? */
 516         if (lustre_msg_get_op_flags(req->rq_reqmsg) & MSG_CONNECT_LIBCLIENT)
 517                 export->exp_libclient = 1;
 518
 519         if (export->exp_connection != NULL)
 520                 ptlrpc_put_connection(export->exp_connection);
 521         export->exp_connection = ptlrpc_get_connection(&req->rq_peer,
 522                                                        &remote_uuid);
 523
 524         if (rc == EALREADY) {
 525                 /* We indicate the reconnection in a flag, not an error code. */
 526                 lustre_msg_add_op_flags(req->rq_repmsg, MSG_CONNECT_RECONNECT);
 527                 GOTO(out, rc = 0);
 528         }
 529
 530         memcpy(&conn, lustre_msg_buf(req->rq_reqmsg, 2, sizeof conn),
 531                sizeof conn);
 532
 533         if (export->exp_imp_reverse != NULL)
 534                 class_destroy_import(export->exp_imp_reverse);
 535         revimp = export->exp_imp_reverse = class_new_import();
 536         revimp->imp_connection = ptlrpc_connection_addref(export->exp_connection);
 537         revimp->imp_client = &export->exp_obd->obd_ldlm_client;
 538         revimp->imp_remote_handle = conn;
 539         revimp->imp_obd = target;
 540         revimp->imp_dlm_fake = 1;
 541         revimp->imp_state = LUSTRE_IMP_FULL;
 542         class_import_put(revimp);
 543 out:
 544         if (rc)
 545                 req->rq_status = rc;
 546         RETURN(rc);
 547 }
 548
 549 int target_handle_disconnect(struct ptlrpc_request *req)
 550 {
 551         struct obd_export *exp;
 552         int rc;
 553         ENTRY;
 554
 555         rc = lustre_pack_reply(req, 0, NULL, NULL);
 556         if (rc)
 557                 RETURN(rc);
 558
 559         /* keep the rq_export around so we can send the reply */
 560         exp = class_export_get(req->rq_export);
 561         req->rq_status = obd_disconnect(exp, 0);
 562         RETURN(0);
 563 }
 564
 565 void target_destroy_export(struct obd_export *exp)
 566 {
 567         /* exports created from last_rcvd data, and "fake"
 568            exports created by lctl don't have an import */
 569         if (exp->exp_imp_reverse != NULL)
 570                 class_destroy_import(exp->exp_imp_reverse);
 571
 572         /* We cancel locks at disconnect time, but this will catch any locks
 573          * granted in a race with recovery-induced disconnect. */
 574         if (exp->exp_obd->obd_namespace != NULL)
 575                 ldlm_cancel_locks_for_export(exp);
 576 }
 577
 578 /*
 579  * Recovery functions
 580  */
 581
 582 static void abort_delayed_replies(struct obd_device *obd)
 583 {
 584         struct ptlrpc_request *req;
 585         struct list_head *tmp, *n;
 586         list_for_each_safe(tmp, n, &obd->obd_delayed_reply_queue) {
 587                 req = list_entry(tmp, struct ptlrpc_request, rq_list);
 588                 DEBUG_REQ(D_ERROR, req, "aborted:");
 589                 req->rq_status = -ENOTCONN;
 590                 req->rq_type = PTL_RPC_MSG_ERR;
 591                 ptlrpc_reply(req);
 592                 class_export_put(req->rq_export);
 593                 list_del(&req->rq_list);
 594                 OBD_FREE(req->rq_reqmsg, req->rq_reqlen);
 595                 OBD_FREE(req, sizeof *req);
 596         }
 597 }
 598
 599 static void abort_recovery_queue(struct obd_device *obd)
 600 {
 601         struct ptlrpc_request *req;
 602         struct list_head *tmp, *n;
 603         int rc;
 604
 605         list_for_each_safe(tmp, n, &obd->obd_recovery_queue) {
 606                 req = list_entry(tmp, struct ptlrpc_request, rq_list);
 607                 DEBUG_REQ(D_ERROR, req, "aborted:");
 608                 req->rq_status = -ENOTCONN;
 609                 req->rq_type = PTL_RPC_MSG_ERR;
 610                 rc = lustre_pack_reply(req, 0, NULL, NULL);
 611                 if (rc == 0) {
 612                         ptlrpc_reply(req);
 613                 } else {
 614                         DEBUG_REQ(D_ERROR, req,
 615                                   "packing failed for abort-reply; skipping");
 616                 }
 617                 list_del(&req->rq_list);
 618                 class_export_put(req->rq_export);
 619                 OBD_FREE(req->rq_reqmsg, req->rq_reqlen);
 620                 OBD_FREE(req, sizeof *req);
 621         }
 622 }
 623
 624 void target_abort_recovery(void *data)
 625 {
 626         struct obd_device *obd = data;
 627         int rc;
 628
 629         CERROR("disconnecting clients and aborting recovery\n");
 630         spin_lock_bh(&obd->obd_processing_task_lock);
 631         if (!obd->obd_recovering) {
 632                 spin_unlock_bh(&obd->obd_processing_task_lock);
 633                 EXIT;
 634                 return;
 635         }
 636
 637         obd->obd_recovering = obd->obd_abort_recovery = 0;
 638
 639         wake_up(&obd->obd_next_transno_waitq);
 640         target_cancel_recovery_timer(obd);
 641         spin_unlock_bh(&obd->obd_processing_task_lock);
 642
 643         class_disconnect_exports(obd, 0);
 644
 645         /* when recovery was aborted, cleanup orphans on mds and ost */
 646         if (OBT(obd) && OBP(obd, postrecov)) {
 647                 rc = OBP(obd, postrecov)(obd);
 648                 if (rc >= 0)
 649                         CWARN("Cleanup %d orphans after recovery was aborted\n", rc);
 650                 else
 651                         CERROR("postrecov failed %d\n", rc);
 652         }
 653
 654         abort_delayed_replies(obd);
 655         abort_recovery_queue(obd);
 656         ptlrpc_run_recovery_over_upcall(obd);
 657 }
 658
 659 static void target_recovery_expired(unsigned long castmeharder)
 660 {
 661         struct obd_device *obd = (struct obd_device *)castmeharder;
 662         CERROR("recovery timed out, aborting\n");
 663         spin_lock_bh(&obd->obd_processing_task_lock);
 664         obd->obd_abort_recovery = 1;
 665         wake_up(&obd->obd_next_transno_waitq);
 666         spin_unlock_bh(&obd->obd_processing_task_lock);
 667 }
 668
 669
 670 /* obd_processing_task_lock should be held */
 671 void target_cancel_recovery_timer(struct obd_device *obd)
 672 {
 673         CDEBUG(D_HA, "%s: cancel recovery timer\n", obd->obd_name);
 674         del_timer(&obd->obd_recovery_timer);
 675 }
 676
 677 static void reset_recovery_timer(struct obd_device *obd)
 678 {
 679         spin_lock_bh(&obd->obd_processing_task_lock);
 680         if (!obd->obd_recovering) {
 681                 spin_unlock_bh(&obd->obd_processing_task_lock);
 682                 return;
 683         }
 684         CDEBUG(D_HA, "timer will expire in %u seconds\n",
 685                OBD_RECOVERY_TIMEOUT / HZ);
 686         mod_timer(&obd->obd_recovery_timer, jiffies + OBD_RECOVERY_TIMEOUT);
 687         spin_unlock_bh(&obd->obd_processing_task_lock);
 688 }
 689
 690
 691 /* Only start it the first time called */
 692 void target_start_recovery_timer(struct obd_device *obd, svc_handler_t handler)
 693 {
 694         spin_lock_bh(&obd->obd_processing_task_lock);
 695         if (obd->obd_recovery_handler) {
 696                 spin_unlock_bh(&obd->obd_processing_task_lock);
 697                 return;
 698         }
 699         CWARN("%s: starting recovery timer (%us)\n", obd->obd_name,
 700                OBD_RECOVERY_TIMEOUT / HZ);
 701         obd->obd_recovery_handler = handler;
 702         obd->obd_recovery_timer.function = target_recovery_expired;
 703         obd->obd_recovery_timer.data = (unsigned long)obd;
 704         init_timer(&obd->obd_recovery_timer);
 705         spin_unlock_bh(&obd->obd_processing_task_lock);
 706
 707         reset_recovery_timer(obd);
 708 }
 709
 710 static int check_for_next_transno(struct obd_device *obd)
 711 {
 712         struct ptlrpc_request *req;
 713         int wake_up = 0, connected, completed, queue_len, max;
 714         __u64 next_transno, req_transno;
 715
 716         spin_lock_bh(&obd->obd_processing_task_lock);
 717         req = list_entry(obd->obd_recovery_queue.next,
 718                          struct ptlrpc_request, rq_list);
 719         max = obd->obd_max_recoverable_clients;
 720         req_transno = req->rq_reqmsg->transno;
 721         connected = obd->obd_connected_clients;
 722         completed = max - obd->obd_recoverable_clients;
 723         queue_len = obd->obd_requests_queued_for_recovery;
 724         next_transno = obd->obd_next_recovery_transno;
 725
 726         if (obd->obd_abort_recovery) {
 727                 CDEBUG(D_HA, "waking for aborted recovery\n");
 728                 wake_up = 1;
 729         } else if (!obd->obd_recovering) {
 730                 CDEBUG(D_HA, "waking for completed recovery (?)\n");
 731                 wake_up = 1;
 732         } else if (req_transno == next_transno) {
 733                 CDEBUG(D_HA, "waking for next ("LPD64")\n", next_transno);
 734                 wake_up = 1;
 735         } else if (queue_len + completed == max) {
 736                 CDEBUG(D_ERROR,
 737                        "waking for skipped transno (skip: "LPD64
 738                        ", ql: %d, comp: %d, conn: %d, next: "LPD64")\n",
 739                        next_transno, queue_len, completed, max, req_transno);
 740                 obd->obd_next_recovery_transno = req_transno;
 741                 wake_up = 1;
 742         }
 743         spin_unlock_bh(&obd->obd_processing_task_lock);
 744         LASSERT(req->rq_reqmsg->transno >= next_transno);
 745         return wake_up;
 746 }
 747
 748 static void process_recovery_queue(struct obd_device *obd)
 749 {
 750         struct ptlrpc_request *req;
 751         int abort_recovery = 0;
 752         struct l_wait_info lwi = { 0 };
 753         ENTRY;
 754
 755         for (;;) {
 756                 spin_lock_bh(&obd->obd_processing_task_lock);
 757                 LASSERT(obd->obd_processing_task == current->pid);
 758                 req = list_entry(obd->obd_recovery_queue.next,
 759                                  struct ptlrpc_request, rq_list);
 760
 761                 if (req->rq_reqmsg->transno != obd->obd_next_recovery_transno) {
 762                         spin_unlock_bh(&obd->obd_processing_task_lock);
 763                         CDEBUG(D_HA, "Waiting for transno "LPD64" (1st is "
 764                                LPD64")\n",
 765                                obd->obd_next_recovery_transno,
 766                                req->rq_reqmsg->transno);
 767                         l_wait_event(obd->obd_next_transno_waitq,
 768                                      check_for_next_transno(obd), &lwi);
 769                         spin_lock_bh(&obd->obd_processing_task_lock);
 770                         abort_recovery = obd->obd_abort_recovery;
 771                         spin_unlock_bh(&obd->obd_processing_task_lock);
 772                         if (abort_recovery) {
 773                                 target_abort_recovery(obd);
 774                                 return;
 775                         }
 776                         continue;
 777                 }
 778                 list_del_init(&req->rq_list);
 779                 obd->obd_requests_queued_for_recovery--;
 780                 spin_unlock_bh(&obd->obd_processing_task_lock);
 781
 782                 DEBUG_REQ(D_HA, req, "processing: ");
 783                 (void)obd->obd_recovery_handler(req);
 784                 obd->obd_replayed_requests++;
 785                 reset_recovery_timer(obd);
 786                 /* bug 1580: decide how to properly sync() in recovery */
 787                 //mds_fsync_super(mds->mds_sb);
 788                 class_export_put(req->rq_export);
 789                 OBD_FREE(req->rq_reqmsg, req->rq_reqlen);
 790                 OBD_FREE(req, sizeof *req);
 791                 spin_lock_bh(&obd->obd_processing_task_lock);
 792                 obd->obd_next_recovery_transno++;
 793                 if (list_empty(&obd->obd_recovery_queue)) {
 794                         obd->obd_processing_task = 0;
 795                         spin_unlock_bh(&obd->obd_processing_task_lock);
 796                         break;
 797                 }
 798                 spin_unlock_bh(&obd->obd_processing_task_lock);
 799         }
 800         EXIT;
 801 }
 802
 803 int target_queue_recovery_request(struct ptlrpc_request *req,
 804                                   struct obd_device *obd)
 805 {
 806         struct list_head *tmp;
 807         int inserted = 0;
 808         __u64 transno = req->rq_reqmsg->transno;
 809         struct ptlrpc_request *saved_req;
 810         struct lustre_msg *reqmsg;
 811
 812         /* CAVEAT EMPTOR: The incoming request message has been swabbed
 813          * (i.e. buflens etc are in my own byte order), but type-dependent
 814          * buffers (eg mds_body, ost_body etc) have NOT been swabbed. */
 815
 816         if (!transno) {
 817                 INIT_LIST_HEAD(&req->rq_list);
 818                 DEBUG_REQ(D_HA, req, "not queueing");
 819                 return 1;
 820         }
 821
 822         /* XXX If I were a real man, these LBUGs would be sane cleanups. */
 823         /* XXX just like the request-dup code in queue_final_reply */
 824         OBD_ALLOC(saved_req, sizeof *saved_req);
 825         if (!saved_req)
 826                 LBUG();
 827         OBD_ALLOC(reqmsg, req->rq_reqlen);
 828         if (!reqmsg)
 829                 LBUG();
 830
 831         spin_lock_bh(&obd->obd_processing_task_lock);
 832
 833         /* If we're processing the queue, we want don't want to queue this
 834          * message.
 835          *
 836          * Also, if this request has a transno less than the one we're waiting
 837          * for, we should process it now.  It could (and currently always will)
 838          * be an open request for a descriptor that was opened some time ago.
 839          */
 840         if (obd->obd_processing_task == current->pid ||
 841             transno < obd->obd_next_recovery_transno) {
 842                 /* Processing the queue right now, don't re-add. */
 843                 LASSERT(list_empty(&req->rq_list));
 844                 spin_unlock_bh(&obd->obd_processing_task_lock);
 845                 OBD_FREE(reqmsg, req->rq_reqlen);
 846                 OBD_FREE(saved_req, sizeof *saved_req);
 847                 return 1;
 848         }
 849
 850         memcpy(saved_req, req, sizeof *req);
 851         memcpy(reqmsg, req->rq_reqmsg, req->rq_reqlen);
 852         req = saved_req;
 853         req->rq_reqmsg = reqmsg;
 854         class_export_get(req->rq_export);
 855         INIT_LIST_HEAD(&req->rq_list);
 856
 857         /* XXX O(n^2) */
 858         list_for_each(tmp, &obd->obd_recovery_queue) {
 859                 struct ptlrpc_request *reqiter =
 860                         list_entry(tmp, struct ptlrpc_request, rq_list);
 861
 862                 if (reqiter->rq_reqmsg->transno > transno) {
 863                         list_add_tail(&req->rq_list, &reqiter->rq_list);
 864                         inserted = 1;
 865                         break;
 866                 }
 867         }
 868
 869         if (!inserted) {
 870                 list_add_tail(&req->rq_list, &obd->obd_recovery_queue);
 871         }
 872
 873         obd->obd_requests_queued_for_recovery++;
 874
 875         if (obd->obd_processing_task != 0) {
 876                 /* Someone else is processing this queue, we'll leave it to
 877                  * them.
 878                  */
 879                 wake_up(&obd->obd_next_transno_waitq);
 880                 spin_unlock_bh(&obd->obd_processing_task_lock);
 881                 return 0;
 882         }
 883
 884         /* Nobody is processing, and we know there's (at least) one to process
 885          * now, so we'll do the honours.
 886          */
 887         obd->obd_processing_task = current->pid;
 888         spin_unlock_bh(&obd->obd_processing_task_lock);
 889
 890         process_recovery_queue(obd);
 891         return 0;
 892 }
 893
 894 struct obd_device * target_req2obd(struct ptlrpc_request *req)
 895 {
 896         return req->rq_export->exp_obd;
 897 }
 898
 899 int target_queue_final_reply(struct ptlrpc_request *req, int rc)
 900 {
 901         struct obd_device *obd = target_req2obd(req);
 902         struct ptlrpc_request *saved_req;
 903         struct lustre_msg *reqmsg;
 904         int recovery_done = 0;
 905         int rc2;
 906
 907         LASSERT ((rc == 0) == (req->rq_reply_state != NULL));
 908
 909         if (rc) {
 910                 /* Just like ptlrpc_error, but without the sending. */
 911                 rc = lustre_pack_reply(req, 0, NULL, NULL);
 912                 LASSERT(rc == 0); /* XXX handle this */
 913                 req->rq_type = PTL_RPC_MSG_ERR;
 914         }
 915
 916         LASSERT (!req->rq_reply_state->rs_difficult);
 917         LASSERT(list_empty(&req->rq_list));
 918         /* XXX a bit like the request-dup code in queue_recovery_request */
 919         OBD_ALLOC(saved_req, sizeof *saved_req);
 920         if (!saved_req)
 921                 LBUG();
 922         OBD_ALLOC(reqmsg, req->rq_reqlen);
 923         if (!reqmsg)
 924                 LBUG();
 925         memcpy(saved_req, req, sizeof *saved_req);
 926         memcpy(reqmsg, req->rq_reqmsg, req->rq_reqlen);
 927         /* the copied req takes over the reply state */
 928         req->rq_reply_state = NULL;
 929         req = saved_req;
 930         req->rq_reqmsg = reqmsg;
 931         class_export_get(req->rq_export);
 932         list_add(&req->rq_list, &obd->obd_delayed_reply_queue);
 933
 934         spin_lock_bh(&obd->obd_processing_task_lock);
 935         --obd->obd_recoverable_clients;
 936         recovery_done = (obd->obd_recoverable_clients == 0);
 937         spin_unlock_bh(&obd->obd_processing_task_lock);
 938
 939         if (recovery_done) {
 940                 struct list_head *tmp, *n;
 941                 ldlm_reprocess_all_ns(req->rq_export->exp_obd->obd_namespace);
 942                 CWARN("%s: all clients recovered, sending delayed replies\n",
 943                        obd->obd_name);
 944                 spin_lock_bh(&obd->obd_processing_task_lock);
 945                 obd->obd_recovering = 0;
 946                 target_cancel_recovery_timer(obd);
 947                 spin_unlock_bh(&obd->obd_processing_task_lock);
 948
 949                 /* when recovery finished, cleanup orphans on mds and ost */
 950                 if (OBT(obd) && OBP(obd, postrecov)) {
 951                         rc2 = OBP(obd, postrecov)(obd);
 952                         if (rc2 >= 0)
 953                                 CWARN("%s: all clients recovered, %d MDS "
 954                                       "orphans deleted\n", obd->obd_name, rc2);
 955                         else
 956                                 CERROR("postrecov failed %d\n", rc2);
 957                 }
 958
 959                 list_for_each_safe(tmp, n, &obd->obd_delayed_reply_queue) {
 960                         req = list_entry(tmp, struct ptlrpc_request, rq_list);
 961                         DEBUG_REQ(D_ERROR, req, "delayed:");
 962                         ptlrpc_reply(req);
 963                         class_export_put(req->rq_export);
 964                         list_del(&req->rq_list);
 965                         OBD_FREE(req->rq_reqmsg, req->rq_reqlen);
 966                         OBD_FREE(req, sizeof *req);
 967                 }
 968                 ptlrpc_run_recovery_over_upcall(obd);
 969         } else {
 970                 CWARN("%s: %d recoverable clients remain\n",
 971                        obd->obd_name, obd->obd_recoverable_clients);
 972                 wake_up(&obd->obd_next_transno_waitq);
 973         }
 974
 975         return 1;
 976 }
 977
 978 int
 979 target_send_reply_msg (struct ptlrpc_request *req, int rc, int fail_id)
 980 {
 981         if (OBD_FAIL_CHECK(fail_id | OBD_FAIL_ONCE)) {
 982                 obd_fail_loc |= OBD_FAIL_ONCE | OBD_FAILED;
 983                 DEBUG_REQ(D_ERROR, req, "dropping reply");
 984                 /* NB this does _not_ send with ACK disabled, to simulate
 985                  * sending OK, but timing out for the ACK */
 986                 if (req->rq_reply_state != NULL) {
 987                         if (!req->rq_reply_state->rs_difficult) {
 988                                 lustre_free_reply_state (req->rq_reply_state);
 989                                 req->rq_reply_state = NULL;
 990                         } else {
 991                                 struct ptlrpc_service *svc =
 992                                         req->rq_rqbd->rqbd_srv_ni->sni_service;
 993                                 atomic_inc(&svc->srv_outstanding_replies);
 994                         }
 995                 }
 996                 return (-ECOMM);
 997         }
 998
 999         if (rc) {
1000                 DEBUG_REQ(D_ERROR, req, "processing error (%d)", rc);
1001                 if (req->rq_reply_state == NULL) {
1002                         rc = lustre_pack_reply (req, 0, NULL, NULL);
1003                         if (rc != 0) {
1004                                 CERROR ("can't allocate reply\n");
1005                                 return (rc);
1006                         }
1007                 }
1008                 req->rq_type = PTL_RPC_MSG_ERR;
1009         } else {
1010                 DEBUG_REQ(D_NET, req, "sending reply");
1011         }
1012
1013         return (ptlrpc_send_reply(req, 1));
1014 }
1015
1016 void
1017 target_send_reply(struct ptlrpc_request *req, int rc, int fail_id)
1018 {
1019         int                        netrc;
1020         unsigned long              flags;
1021         struct ptlrpc_reply_state *rs;
1022         struct obd_device         *obd;
1023         struct obd_export         *exp;
1024         struct ptlrpc_srv_ni      *sni;
1025         struct ptlrpc_service     *svc;
1026
1027         sni = req->rq_rqbd->rqbd_srv_ni;
1028         svc = sni->sni_service;
1029
1030         rs = req->rq_reply_state;
1031         if (rs == NULL || !rs->rs_difficult) {
1032                 /* The easy case; no notifiers and reply_out_callback()
1033                  * cleans up (i.e. we can't look inside rs after a
1034                  * successful send) */
1035                 netrc = target_send_reply_msg (req, rc, fail_id);
1036
1037                 LASSERT (netrc == 0 || req->rq_reply_state == NULL);
1038                 return;
1039         }
1040
1041         /* must be an export if locks saved */
1042         LASSERT (req->rq_export != NULL);
1043         /* req/reply consistent */
1044         LASSERT (rs->rs_srv_ni == sni);
1045
1046         /* "fresh" reply */
1047         LASSERT (!rs->rs_scheduled);
1048         LASSERT (!rs->rs_scheduled_ever);
1049         LASSERT (!rs->rs_handled);
1050         LASSERT (!rs->rs_on_net);
1051         LASSERT (rs->rs_export == NULL);
1052         LASSERT (list_empty(&rs->rs_obd_list));
1053         LASSERT (list_empty(&rs->rs_exp_list));
1054
1055         exp = class_export_get (req->rq_export);
1056         obd = exp->exp_obd;
1057
1058         /* disable reply scheduling onto srv_reply_queue while I'm setting up */
1059         rs->rs_scheduled = 1;
1060         rs->rs_on_net    = 1;
1061         rs->rs_xid       = req->rq_xid;
1062         rs->rs_transno   = req->rq_transno;
1063         rs->rs_export    = exp;
1064
1065         spin_lock_irqsave (&obd->obd_uncommitted_replies_lock, flags);
1066
1067         if (rs->rs_transno > obd->obd_last_committed) {
1068                 /* not committed already */
1069                 list_add_tail (&rs->rs_obd_list,
1070                                &obd->obd_uncommitted_replies);
1071         }
1072
1073         spin_unlock (&obd->obd_uncommitted_replies_lock);
1074         spin_lock (&exp->exp_lock);
1075
1076         list_add_tail (&rs->rs_exp_list, &exp->exp_outstanding_replies);
1077
1078         spin_unlock_irqrestore (&exp->exp_lock, flags);
1079
1080         netrc = target_send_reply_msg (req, rc, fail_id);
1081
1082         spin_lock_irqsave (&svc->srv_lock, flags);
1083
1084         svc->srv_n_difficult_replies++;
1085
1086         if (netrc != 0) /* error sending: reply is off the net */
1087                 rs->rs_on_net = 0;
1088
1089         if (!rs->rs_on_net ||                   /* some notifier */
1090             list_empty(&rs->rs_exp_list) ||     /* completed already */
1091             list_empty(&rs->rs_obd_list)) {
1092                 list_add_tail (&rs->rs_list, &svc->srv_reply_queue);
1093                 wake_up (&svc->srv_waitq);
1094         } else {
1095                 list_add (&rs->rs_list, &sni->sni_active_replies);
1096                 rs->rs_scheduled = 0;           /* allow notifier to schedule */
1097         }
1098
1099         spin_unlock_irqrestore (&svc->srv_lock, flags);
1100 }
1101
1102 int target_handle_ping(struct ptlrpc_request *req)
1103 {
1104         return lustre_pack_reply(req, 0, NULL, NULL);
1105 }