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         INIT_LIST_HEAD(&cli->cl_loi_read_list);
 106         spin_lock_init(&cli->cl_loi_list_lock);
 107         cli->cl_brw_in_flight = 0;
 108         spin_lock_init(&cli->cl_read_rpc_hist.oh_lock);
 109         spin_lock_init(&cli->cl_write_rpc_hist.oh_lock);
 110         spin_lock_init(&cli->cl_read_page_hist.oh_lock);
 111         spin_lock_init(&cli->cl_write_page_hist.oh_lock);
 112         cli->cl_max_pages_per_rpc = PTL_MD_MAX_PAGES;
 113         cli->cl_max_rpcs_in_flight = OSC_MAX_RIF_DEFAULT;
 114
 115         ldlm_get_ref();
 116         if (rc) {
 117                 CERROR("ldlm_get_ref failed: %d\n", rc);
 118                 GOTO(err, rc);
 119         }
 120
 121         conn = ptlrpc_uuid_to_connection(&server_uuid);
 122         if (conn == NULL)
 123                 GOTO(err_ldlm, rc = -ENOENT);
 124
 125         ptlrpc_init_client(rq_portal, rp_portal, name,
 126                            &obddev->obd_ldlm_client);
 127
 128         imp = class_new_import();
 129         if (imp == NULL) {
 130                 ptlrpc_put_connection(conn);
 131                 GOTO(err_ldlm, rc = -ENOENT);
 132         }
 133         imp->imp_connection = conn;
 134         imp->imp_client = &obddev->obd_ldlm_client;
 135         imp->imp_obd = obddev;
 136         imp->imp_connect_op = connect_op;
 137         imp->imp_generation = 0;
 138         imp->imp_initial_recov = 1;
 139         INIT_LIST_HEAD(&imp->imp_pinger_chain);
 140         memcpy(imp->imp_target_uuid.uuid, lcfg->lcfg_inlbuf1,
 141               lcfg->lcfg_inllen1);
 142         class_import_put(imp);
 143
 144         cli->cl_import = imp;
 145         cli->cl_max_mds_easize = sizeof(struct lov_mds_md);
 146         cli->cl_max_mds_cookiesize = sizeof(struct llog_cookie);
 147         cli->cl_sandev = to_kdev_t(0);
 148
 149         if (lcfg->lcfg_inllen3 != 0) {
 150                 if (!strcmp(lcfg->lcfg_inlbuf3, "inactive")) {
 151                         CDEBUG(D_HA, "marking %s %s->%s as inactive\n",
 152                                name, obddev->obd_name,
 153                                imp->imp_target_uuid.uuid);
 154                         imp->imp_invalid = 1;
 155
 156                         if (lcfg->lcfg_inllen4 != 0)
 157                                 mgmt_name = lcfg->lcfg_inlbuf4;
 158                 } else {
 159                         mgmt_name = lcfg->lcfg_inlbuf3;
 160                 }
 161         }
 162
 163         if (mgmt_name != NULL) {
 164                 /* Register with management client if we need to. */
 165                 CDEBUG(D_HA, "%s registering with %s for events about %s\n",
 166                        obddev->obd_name, mgmt_name, server_uuid.uuid);
 167
 168                 mgmt_obd = class_name2obd(mgmt_name);
 169                 if (!mgmt_obd) {
 170                         CERROR("can't find mgmtcli %s to register\n",
 171                                mgmt_name);
 172                         GOTO(err_import, rc = -ENOSYS);
 173                 }
 174
 175                 register_f = inter_module_get("mgmtcli_register_for_events");
 176                 if (!register_f) {
 177                         CERROR("can't i_m_g mgmtcli_register_for_events\n");
 178                         GOTO(err_import, rc = -ENOSYS);
 179                 }
 180
 181                 rc = register_f(mgmt_obd, obddev, &imp->imp_target_uuid);
 182                 inter_module_put("mgmtcli_register_for_events");
 183
 184                 if (!rc)
 185                         cli->cl_mgmtcli_obd = mgmt_obd;
 186         }
 187
 188         RETURN(rc);
 189
 190 err_import:
 191         class_destroy_import(imp);
 192 err_ldlm:
 193         ldlm_put_ref(0);
 194 err:
 195         RETURN(rc);
 196
 197 }
 198
 199 int client_obd_cleanup(struct obd_device *obddev, int flags)
 200 {
 201         struct client_obd *cli = &obddev->u.cli;
 202
 203         if (!cli->cl_import)
 204                 RETURN(-EINVAL);
 205         if (cli->cl_mgmtcli_obd) {
 206                 mgmtcli_deregister_for_events_t dereg_f;
 207
 208                 dereg_f = inter_module_get("mgmtcli_deregister_for_events");
 209                 dereg_f(cli->cl_mgmtcli_obd, obddev);
 210                 inter_module_put("mgmtcli_deregister_for_events");
 211         }
 212         class_destroy_import(cli->cl_import);
 213         cli->cl_import = NULL;
 214
 215         ldlm_put_ref(flags & OBD_OPT_FORCE);
 216
 217         RETURN(0);
 218 }
 219
 220 int client_connect_import(struct lustre_handle *dlm_handle,
 221                           struct obd_device *obd,
 222                           struct obd_uuid *cluuid)
 223 {
 224         struct client_obd *cli = &obd->u.cli;
 225         struct obd_import *imp = cli->cl_import;
 226         struct obd_export *exp;
 227         int rc;
 228         ENTRY;
 229
 230         down(&cli->cl_sem);
 231         rc = class_connect(dlm_handle, obd, cluuid);
 232         if (rc)
 233                 GOTO(out_sem, rc);
 234
 235         cli->cl_conn_count++;
 236         if (cli->cl_conn_count > 1)
 237                 GOTO(out_sem, rc);
 238         exp = class_conn2export(dlm_handle);
 239
 240         if (obd->obd_namespace != NULL)
 241                 CERROR("already have namespace!\n");
 242         obd->obd_namespace = ldlm_namespace_new(obd->obd_name,
 243                                                 LDLM_NAMESPACE_CLIENT);
 244         if (obd->obd_namespace == NULL)
 245                 GOTO(out_disco, rc = -ENOMEM);
 246
 247         imp->imp_dlm_handle = *dlm_handle;
 248         rc = ptlrpc_init_import(imp);
 249         if (rc != 0)
 250                 GOTO(out_ldlm, rc);
 251
 252         exp->exp_connection = ptlrpc_connection_addref(imp->imp_connection);
 253         rc = ptlrpc_connect_import(imp, NULL);
 254         if (rc != 0) {
 255                 LASSERT (imp->imp_state == LUSTRE_IMP_DISCON);
 256                 GOTO(out_ldlm, rc);
 257         }
 258
 259         ptlrpc_pinger_add_import(imp);
 260         EXIT;
 261
 262         if (rc) {
 263 out_ldlm:
 264                 ldlm_namespace_free(obd->obd_namespace, 0);
 265                 obd->obd_namespace = NULL;
 266 out_disco:
 267                 cli->cl_conn_count--;
 268                 class_disconnect(exp, 0);
 269         } else {
 270                 class_export_put(exp);
 271         }
 272 out_sem:
 273         up(&cli->cl_sem);
 274         return rc;
 275 }
 276
 277 int client_disconnect_export(struct obd_export *exp, int failover)
 278 {
 279         struct obd_device *obd = class_exp2obd(exp);
 280         struct client_obd *cli = &obd->u.cli;
 281         struct obd_import *imp = cli->cl_import;
 282         int rc = 0, err;
 283         ENTRY;
 284
 285         if (!obd) {
 286                 CERROR("invalid export for disconnect: exp %p cookie "LPX64"\n",
 287                        exp, exp ? exp->exp_handle.h_cookie : -1);
 288                 RETURN(-EINVAL);
 289         }
 290
 291         down(&cli->cl_sem);
 292         if (!cli->cl_conn_count) {
 293                 CERROR("disconnecting disconnected device (%s)\n",
 294                        obd->obd_name);
 295                 GOTO(out_sem, rc = -EINVAL);
 296         }
 297
 298         cli->cl_conn_count--;
 299         if (cli->cl_conn_count)
 300                 GOTO(out_no_disconnect, rc = 0);
 301
 302         /* Some non-replayable imports (MDS's OSCs) are pinged, so just
 303          * delete it regardless.  (It's safe to delete an import that was
 304          * never added.) */
 305         (void)ptlrpc_pinger_del_import(imp);
 306
 307         if (obd->obd_namespace != NULL) {
 308                 /* obd_no_recov == local only */
 309                 ldlm_cli_cancel_unused(obd->obd_namespace, NULL,
 310                                        obd->obd_no_recov, NULL);
 311                 ldlm_namespace_free(obd->obd_namespace, obd->obd_no_recov);
 312                 obd->obd_namespace = NULL;
 313         }
 314
 315         /* Yeah, obd_no_recov also (mainly) means "forced shutdown". */
 316         if (obd->obd_no_recov)
 317                 ptlrpc_invalidate_import(imp, 0);
 318         else
 319                 rc = ptlrpc_disconnect_import(imp);
 320
 321         EXIT;
 322  out_no_disconnect:
 323         err = class_disconnect(exp, 0);
 324         if (!rc && err)
 325                 rc = err;
 326  out_sem:
 327         up(&cli->cl_sem);
 328         RETURN(rc);
 329 }
 330
 331 /* --------------------------------------------------------------------------
 332  * from old lib/target.c
 333  * -------------------------------------------------------------------------- */
 334
 335 int target_handle_reconnect(struct lustre_handle *conn, struct obd_export *exp,
 336                             struct obd_uuid *cluuid)
 337 {
 338         if (exp->exp_connection) {
 339                 struct lustre_handle *hdl;
 340                 hdl = &exp->exp_imp_reverse->imp_remote_handle;
 341                 /* Might be a re-connect after a partition. */
 342                 if (!memcmp(&conn->cookie, &hdl->cookie, sizeof conn->cookie)) {
 343                         CERROR("%s reconnecting\n", cluuid->uuid);
 344                         conn->cookie = exp->exp_handle.h_cookie;
 345                         RETURN(EALREADY);
 346                 } else {
 347                         CERROR("%s reconnecting from %s, "
 348                                "handle mismatch (ours "LPX64", theirs "
 349                                LPX64")\n", cluuid->uuid,
 350                                exp->exp_connection->c_remote_uuid.uuid,
 351                                hdl->cookie, conn->cookie);
 352                         memset(conn, 0, sizeof *conn);
 353                         RETURN(-EALREADY);
 354                 }
 355         }
 356
 357         conn->cookie = exp->exp_handle.h_cookie;
 358         CDEBUG(D_INFO, "existing export for UUID '%s' at %p\n",
 359                cluuid->uuid, exp);
 360         CDEBUG(D_IOCTL,"connect: cookie "LPX64"\n", conn->cookie);
 361         RETURN(0);
 362 }
 363
 364 int target_handle_connect(struct ptlrpc_request *req, svc_handler_t handler)
 365 {
 366         struct obd_device *target;
 367         struct obd_export *export = NULL;
 368         struct obd_import *revimp;
 369         struct lustre_handle conn;
 370         struct obd_uuid tgtuuid;
 371         struct obd_uuid cluuid;
 372         struct obd_uuid remote_uuid;
 373         struct list_head *p;
 374         char *str, *tmp;
 375         int rc = 0, abort_recovery;
 376         unsigned long flags;
 377         ENTRY;
 378
 379         OBD_RACE(OBD_FAIL_TGT_CONN_RACE);
 380
 381         LASSERT_REQSWAB (req, 0);
 382         str = lustre_msg_string(req->rq_reqmsg, 0, sizeof(tgtuuid) - 1);
 383         if (str == NULL) {
 384                 CERROR("bad target UUID for connect\n");
 385                 GOTO(out, rc = -EINVAL);
 386         }
 387
 388         obd_str2uuid (&tgtuuid, str);
 389         target = class_uuid2obd(&tgtuuid);
 390         if (!target) {
 391                 target = class_name2obd(str);
 392         }
 393
 394         if (!target || target->obd_stopping || !target->obd_set_up) {
 395                 CERROR("UUID '%s' is not available for connect\n", str);
 396                 GOTO(out, rc = -ENODEV);
 397         }
 398
 399         LASSERT_REQSWAB (req, 1);
 400         str = lustre_msg_string(req->rq_reqmsg, 1, sizeof(cluuid) - 1);
 401         if (str == NULL) {
 402                 CERROR("bad client UUID for connect\n");
 403                 GOTO(out, rc = -EINVAL);
 404         }
 405
 406         obd_str2uuid (&cluuid, str);
 407
 408         /* XXX extract a nettype and format accordingly */
 409         snprintf(remote_uuid.uuid, sizeof remote_uuid,
 410                  "NET_"LPX64"_UUID", req->rq_peer.peer_nid);
 411
 412         spin_lock_bh(&target->obd_processing_task_lock);
 413         abort_recovery = target->obd_abort_recovery;
 414         spin_unlock_bh(&target->obd_processing_task_lock);
 415         if (abort_recovery)
 416                 target_abort_recovery(target);
 417
 418         tmp = lustre_msg_buf(req->rq_reqmsg, 2, sizeof conn);
 419         if (tmp == NULL)
 420                 GOTO(out, rc = -EPROTO);
 421
 422         memcpy(&conn, tmp, sizeof conn);
 423
 424         rc = lustre_pack_reply(req, 0, NULL, NULL);
 425         if (rc)
 426                 GOTO(out, rc);
 427
 428         /* lctl gets a backstage, all-access pass. */
 429         if (obd_uuid_equals(&cluuid, &target->obd_uuid))
 430                 goto dont_check_exports;
 431
 432         spin_lock(&target->obd_dev_lock);
 433         list_for_each(p, &target->obd_exports) {
 434                 export = list_entry(p, struct obd_export, exp_obd_chain);
 435                 if (obd_uuid_equals(&cluuid, &export->exp_client_uuid)) {
 436                         spin_unlock(&target->obd_dev_lock);
 437                         LASSERT(export->exp_obd == target);
 438
 439                         rc = target_handle_reconnect(&conn, export, &cluuid);
 440                         break;
 441                 }
 442                 export = NULL;
 443         }
 444         /* If we found an export, we already unlocked. */
 445         if (!export) {
 446                 spin_unlock(&target->obd_dev_lock);
 447         } else if (req->rq_reqmsg->conn_cnt == 1) {
 448                 CERROR("%s reconnected with 1 conn_cnt; cookies not random?\n",
 449                        cluuid.uuid);
 450                 GOTO(out, rc = -EALREADY);
 451         }
 452
 453         /* Tell the client if we're in recovery. */
 454         /* If this is the first client, start the recovery timer */
 455         if (target->obd_recovering) {
 456                 lustre_msg_add_op_flags(req->rq_repmsg, MSG_CONNECT_RECOVERING);
 457                 target_start_recovery_timer(target, handler);
 458         }
 459
 460         /* Tell the client if we support replayable requests */
 461         if (target->obd_replayable)
 462                 lustre_msg_add_op_flags(req->rq_repmsg, MSG_CONNECT_REPLAYABLE);
 463
 464         if (export == NULL) {
 465                 if (target->obd_recovering) {
 466                         CERROR("denying connection for new client %s: "
 467                                "%d clients in recovery for %lds\n", cluuid.uuid,
 468                                target->obd_recoverable_clients,
 469                                (target->obd_recovery_timer.expires-jiffies)/HZ);
 470                         rc = -EBUSY;
 471                 } else {
 472  dont_check_exports:
 473                         rc = obd_connect(&conn, target, &cluuid);
 474                 }
 475         }
 476
 477
 478         /* If all else goes well, this is our RPC return code. */
 479         req->rq_status = 0;
 480
 481         if (rc && rc != EALREADY)
 482                 GOTO(out, rc);
 483
 484         /* XXX track this all the time? */
 485         if (target->obd_recovering) {
 486                 target->obd_connected_clients++;
 487         }
 488
 489         req->rq_repmsg->handle = conn;
 490
 491         /* If the client and the server are the same node, we will already
 492          * have an export that really points to the client's DLM export,
 493          * because we have a shared handles table.
 494          *
 495          * XXX this will go away when shaver stops sending the "connect" handle
 496          * in the real "remote handle" field of the request --phik 24 Apr 2003
 497          */
 498         if (req->rq_export != NULL)
 499                 class_export_put(req->rq_export);
 500
 501         /* ownership of this export ref transfers to the request */
 502         export = req->rq_export = class_conn2export(&conn);
 503         LASSERT(export != NULL);
 504
 505         spin_lock_irqsave(&export->exp_lock, flags);
 506         if (export->exp_conn_cnt >= req->rq_reqmsg->conn_cnt) {
 507                 CERROR("%s: already connected at a higher conn_cnt: %d > %d\n",
 508                        cluuid.uuid, export->exp_conn_cnt,
 509                        req->rq_reqmsg->conn_cnt);
 510                 spin_unlock_irqrestore(&export->exp_lock, flags);
 511                 GOTO(out, rc = -EALREADY);
 512         }
 513         export->exp_conn_cnt = req->rq_reqmsg->conn_cnt;
 514         spin_unlock_irqrestore(&export->exp_lock, flags);
 515
 516         /* request from liblustre? */
 517         if (lustre_msg_get_op_flags(req->rq_reqmsg) & MSG_CONNECT_LIBCLIENT)
 518                 export->exp_libclient = 1;
 519
 520         if (export->exp_connection != NULL)
 521                 ptlrpc_put_connection(export->exp_connection);
 522         export->exp_connection = ptlrpc_get_connection(&req->rq_peer,
 523                                                        &remote_uuid);
 524
 525         if (rc == EALREADY) {
 526                 /* We indicate the reconnection in a flag, not an error code. */
 527                 lustre_msg_add_op_flags(req->rq_repmsg, MSG_CONNECT_RECONNECT);
 528                 GOTO(out, rc = 0);
 529         }
 530
 531         memcpy(&conn, lustre_msg_buf(req->rq_reqmsg, 2, sizeof conn),
 532                sizeof conn);
 533
 534         if (export->exp_imp_reverse != NULL)
 535                 class_destroy_import(export->exp_imp_reverse);
 536         revimp = export->exp_imp_reverse = class_new_import();
 537         revimp->imp_connection = ptlrpc_connection_addref(export->exp_connection);
 538         revimp->imp_client = &export->exp_obd->obd_ldlm_client;
 539         revimp->imp_remote_handle = conn;
 540         revimp->imp_obd = target;
 541         revimp->imp_dlm_fake = 1;
 542         revimp->imp_state = LUSTRE_IMP_FULL;
 543         class_import_put(revimp);
 544 out:
 545         if (rc)
 546                 req->rq_status = rc;
 547         RETURN(rc);
 548 }
 549
 550 int target_handle_disconnect(struct ptlrpc_request *req)
 551 {
 552         struct obd_export *exp;
 553         int rc;
 554         ENTRY;
 555
 556         rc = lustre_pack_reply(req, 0, NULL, NULL);
 557         if (rc)
 558                 RETURN(rc);
 559
 560         /* keep the rq_export around so we can send the reply */
 561         exp = class_export_get(req->rq_export);
 562         req->rq_status = obd_disconnect(exp, 0);
 563         RETURN(0);
 564 }
 565
 566 void target_destroy_export(struct obd_export *exp)
 567 {
 568         /* exports created from last_rcvd data, and "fake"
 569            exports created by lctl don't have an import */
 570         if (exp->exp_imp_reverse != NULL)
 571                 class_destroy_import(exp->exp_imp_reverse);
 572
 573         /* We cancel locks at disconnect time, but this will catch any locks
 574          * granted in a race with recovery-induced disconnect. */
 575         if (exp->exp_obd->obd_namespace != NULL)
 576                 ldlm_cancel_locks_for_export(exp);
 577 }
 578
 579 /*
 580  * Recovery functions
 581  */
 582
 583 static void abort_delayed_replies(struct obd_device *obd)
 584 {
 585         struct ptlrpc_request *req;
 586         struct list_head *tmp, *n;
 587         list_for_each_safe(tmp, n, &obd->obd_delayed_reply_queue) {
 588                 req = list_entry(tmp, struct ptlrpc_request, rq_list);
 589                 DEBUG_REQ(D_ERROR, req, "aborted:");
 590                 req->rq_status = -ENOTCONN;
 591                 req->rq_type = PTL_RPC_MSG_ERR;
 592                 ptlrpc_reply(req);
 593                 class_export_put(req->rq_export);
 594                 list_del(&req->rq_list);
 595                 OBD_FREE(req->rq_reqmsg, req->rq_reqlen);
 596                 OBD_FREE(req, sizeof *req);
 597         }
 598 }
 599
 600 static void abort_recovery_queue(struct obd_device *obd)
 601 {
 602         struct ptlrpc_request *req;
 603         struct list_head *tmp, *n;
 604         int rc;
 605
 606         list_for_each_safe(tmp, n, &obd->obd_recovery_queue) {
 607                 req = list_entry(tmp, struct ptlrpc_request, rq_list);
 608                 DEBUG_REQ(D_ERROR, req, "aborted:");
 609                 req->rq_status = -ENOTCONN;
 610                 req->rq_type = PTL_RPC_MSG_ERR;
 611                 rc = lustre_pack_reply(req, 0, NULL, NULL);
 612                 if (rc == 0) {
 613                         ptlrpc_reply(req);
 614                 } else {
 615                         DEBUG_REQ(D_ERROR, req,
 616                                   "packing failed for abort-reply; skipping");
 617                 }
 618                 list_del(&req->rq_list);
 619                 class_export_put(req->rq_export);
 620                 OBD_FREE(req->rq_reqmsg, req->rq_reqlen);
 621                 OBD_FREE(req, sizeof *req);
 622         }
 623 }
 624
 625 void target_abort_recovery(void *data)
 626 {
 627         struct obd_device *obd = data;
 628         int rc;
 629
 630         CERROR("disconnecting clients and aborting recovery\n");
 631         spin_lock_bh(&obd->obd_processing_task_lock);
 632         if (!obd->obd_recovering) {
 633                 spin_unlock_bh(&obd->obd_processing_task_lock);
 634                 EXIT;
 635                 return;
 636         }
 637
 638         obd->obd_recovering = obd->obd_abort_recovery = 0;
 639
 640         wake_up(&obd->obd_next_transno_waitq);
 641         target_cancel_recovery_timer(obd);
 642         spin_unlock_bh(&obd->obd_processing_task_lock);
 643
 644         class_disconnect_exports(obd, 0);
 645
 646         /* when recovery was aborted, cleanup orphans on mds and ost */
 647         if (OBT(obd) && OBP(obd, postrecov)) {
 648                 rc = OBP(obd, postrecov)(obd);
 649                 if (rc >= 0)
 650                         CWARN("Cleanup %d orphans after recovery was aborted\n", rc);
 651                 else
 652                         CERROR("postrecov failed %d\n", rc);
 653         }
 654
 655         abort_delayed_replies(obd);
 656         abort_recovery_queue(obd);
 657         ptlrpc_run_recovery_over_upcall(obd);
 658 }
 659
 660 static void target_recovery_expired(unsigned long castmeharder)
 661 {
 662         struct obd_device *obd = (struct obd_device *)castmeharder;
 663         CERROR("recovery timed out, aborting\n");
 664         spin_lock_bh(&obd->obd_processing_task_lock);
 665         obd->obd_abort_recovery = 1;
 666         wake_up(&obd->obd_next_transno_waitq);
 667         spin_unlock_bh(&obd->obd_processing_task_lock);
 668 }
 669
 670
 671 /* obd_processing_task_lock should be held */
 672 void target_cancel_recovery_timer(struct obd_device *obd)
 673 {
 674         CDEBUG(D_HA, "%s: cancel recovery timer\n", obd->obd_name);
 675         del_timer(&obd->obd_recovery_timer);
 676 }
 677
 678 static void reset_recovery_timer(struct obd_device *obd)
 679 {
 680         spin_lock_bh(&obd->obd_processing_task_lock);
 681         if (!obd->obd_recovering) {
 682                 spin_unlock_bh(&obd->obd_processing_task_lock);
 683                 return;
 684         }
 685         CDEBUG(D_HA, "timer will expire in %u seconds\n",
 686                OBD_RECOVERY_TIMEOUT / HZ);
 687         mod_timer(&obd->obd_recovery_timer, jiffies + OBD_RECOVERY_TIMEOUT);
 688         spin_unlock_bh(&obd->obd_processing_task_lock);
 689 }
 690
 691
 692 /* Only start it the first time called */
 693 void target_start_recovery_timer(struct obd_device *obd, svc_handler_t handler)
 694 {
 695         spin_lock_bh(&obd->obd_processing_task_lock);
 696         if (obd->obd_recovery_handler) {
 697                 spin_unlock_bh(&obd->obd_processing_task_lock);
 698                 return;
 699         }
 700         CWARN("%s: starting recovery timer (%us)\n", obd->obd_name,
 701                OBD_RECOVERY_TIMEOUT / HZ);
 702         obd->obd_recovery_handler = handler;
 703         obd->obd_recovery_timer.function = target_recovery_expired;
 704         obd->obd_recovery_timer.data = (unsigned long)obd;
 705         init_timer(&obd->obd_recovery_timer);
 706         spin_unlock_bh(&obd->obd_processing_task_lock);
 707
 708         reset_recovery_timer(obd);
 709 }
 710
 711 static int check_for_next_transno(struct obd_device *obd)
 712 {
 713         struct ptlrpc_request *req;
 714         int wake_up = 0, connected, completed, queue_len, max;
 715         __u64 next_transno, req_transno;
 716
 717         spin_lock_bh(&obd->obd_processing_task_lock);
 718         req = list_entry(obd->obd_recovery_queue.next,
 719                          struct ptlrpc_request, rq_list);
 720         max = obd->obd_max_recoverable_clients;
 721         req_transno = req->rq_reqmsg->transno;
 722         connected = obd->obd_connected_clients;
 723         completed = max - obd->obd_recoverable_clients;
 724         queue_len = obd->obd_requests_queued_for_recovery;
 725         next_transno = obd->obd_next_recovery_transno;
 726
 727         if (obd->obd_abort_recovery) {
 728                 CDEBUG(D_HA, "waking for aborted recovery\n");
 729                 wake_up = 1;
 730         } else if (!obd->obd_recovering) {
 731                 CDEBUG(D_HA, "waking for completed recovery (?)\n");
 732                 wake_up = 1;
 733         } else if (req_transno == next_transno) {
 734                 CDEBUG(D_HA, "waking for next ("LPD64")\n", next_transno);
 735                 wake_up = 1;
 736         } else if (queue_len + completed == max) {
 737                 CDEBUG(D_ERROR,
 738                        "waking for skipped transno (skip: "LPD64
 739                        ", ql: %d, comp: %d, conn: %d, next: "LPD64")\n",
 740                        next_transno, queue_len, completed, max, req_transno);
 741                 obd->obd_next_recovery_transno = req_transno;
 742                 wake_up = 1;
 743         }
 744         spin_unlock_bh(&obd->obd_processing_task_lock);
 745         LASSERT(req->rq_reqmsg->transno >= next_transno);
 746         return wake_up;
 747 }
 748
 749 static void process_recovery_queue(struct obd_device *obd)
 750 {
 751         struct ptlrpc_request *req;
 752         int abort_recovery = 0;
 753         struct l_wait_info lwi = { 0 };
 754         ENTRY;
 755
 756         for (;;) {
 757                 spin_lock_bh(&obd->obd_processing_task_lock);
 758                 LASSERT(obd->obd_processing_task == current->pid);
 759                 req = list_entry(obd->obd_recovery_queue.next,
 760                                  struct ptlrpc_request, rq_list);
 761
 762                 if (req->rq_reqmsg->transno != obd->obd_next_recovery_transno) {
 763                         spin_unlock_bh(&obd->obd_processing_task_lock);
 764                         CDEBUG(D_HA, "Waiting for transno "LPD64" (1st is "
 765                                LPD64")\n",
 766                                obd->obd_next_recovery_transno,
 767                                req->rq_reqmsg->transno);
 768                         l_wait_event(obd->obd_next_transno_waitq,
 769                                      check_for_next_transno(obd), &lwi);
 770                         spin_lock_bh(&obd->obd_processing_task_lock);
 771                         abort_recovery = obd->obd_abort_recovery;
 772                         spin_unlock_bh(&obd->obd_processing_task_lock);
 773                         if (abort_recovery) {
 774                                 target_abort_recovery(obd);
 775                                 return;
 776                         }
 777                         continue;
 778                 }
 779                 list_del_init(&req->rq_list);
 780                 obd->obd_requests_queued_for_recovery--;
 781                 spin_unlock_bh(&obd->obd_processing_task_lock);
 782
 783                 DEBUG_REQ(D_HA, req, "processing: ");
 784                 (void)obd->obd_recovery_handler(req);
 785                 obd->obd_replayed_requests++;
 786                 reset_recovery_timer(obd);
 787                 /* bug 1580: decide how to properly sync() in recovery */
 788                 //mds_fsync_super(mds->mds_sb);
 789                 class_export_put(req->rq_export);
 790                 OBD_FREE(req->rq_reqmsg, req->rq_reqlen);
 791                 OBD_FREE(req, sizeof *req);
 792                 spin_lock_bh(&obd->obd_processing_task_lock);
 793                 obd->obd_next_recovery_transno++;
 794                 if (list_empty(&obd->obd_recovery_queue)) {
 795                         obd->obd_processing_task = 0;
 796                         spin_unlock_bh(&obd->obd_processing_task_lock);
 797                         break;
 798                 }
 799                 spin_unlock_bh(&obd->obd_processing_task_lock);
 800         }
 801         EXIT;
 802 }
 803
 804 int target_queue_recovery_request(struct ptlrpc_request *req,
 805                                   struct obd_device *obd)
 806 {
 807         struct list_head *tmp;
 808         int inserted = 0;
 809         __u64 transno = req->rq_reqmsg->transno;
 810         struct ptlrpc_request *saved_req;
 811         struct lustre_msg *reqmsg;
 812
 813         /* CAVEAT EMPTOR: The incoming request message has been swabbed
 814          * (i.e. buflens etc are in my own byte order), but type-dependent
 815          * buffers (eg mds_body, ost_body etc) have NOT been swabbed. */
 816
 817         if (!transno) {
 818                 INIT_LIST_HEAD(&req->rq_list);
 819                 DEBUG_REQ(D_HA, req, "not queueing");
 820                 return 1;
 821         }
 822
 823         /* XXX If I were a real man, these LBUGs would be sane cleanups. */
 824         /* XXX just like the request-dup code in queue_final_reply */
 825         OBD_ALLOC(saved_req, sizeof *saved_req);
 826         if (!saved_req)
 827                 LBUG();
 828         OBD_ALLOC(reqmsg, req->rq_reqlen);
 829         if (!reqmsg)
 830                 LBUG();
 831
 832         spin_lock_bh(&obd->obd_processing_task_lock);
 833
 834         /* If we're processing the queue, we want don't want to queue this
 835          * message.
 836          *
 837          * Also, if this request has a transno less than the one we're waiting
 838          * for, we should process it now.  It could (and currently always will)
 839          * be an open request for a descriptor that was opened some time ago.
 840          */
 841         if (obd->obd_processing_task == current->pid ||
 842             transno < obd->obd_next_recovery_transno) {
 843                 /* Processing the queue right now, don't re-add. */
 844                 LASSERT(list_empty(&req->rq_list));
 845                 spin_unlock_bh(&obd->obd_processing_task_lock);
 846                 OBD_FREE(reqmsg, req->rq_reqlen);
 847                 OBD_FREE(saved_req, sizeof *saved_req);
 848                 return 1;
 849         }
 850
 851         memcpy(saved_req, req, sizeof *req);
 852         memcpy(reqmsg, req->rq_reqmsg, req->rq_reqlen);
 853         req = saved_req;
 854         req->rq_reqmsg = reqmsg;
 855         class_export_get(req->rq_export);
 856         INIT_LIST_HEAD(&req->rq_list);
 857
 858         /* XXX O(n^2) */
 859         list_for_each(tmp, &obd->obd_recovery_queue) {
 860                 struct ptlrpc_request *reqiter =
 861                         list_entry(tmp, struct ptlrpc_request, rq_list);
 862
 863                 if (reqiter->rq_reqmsg->transno > transno) {
 864                         list_add_tail(&req->rq_list, &reqiter->rq_list);
 865                         inserted = 1;
 866                         break;
 867                 }
 868         }
 869
 870         if (!inserted) {
 871                 list_add_tail(&req->rq_list, &obd->obd_recovery_queue);
 872         }
 873
 874         obd->obd_requests_queued_for_recovery++;
 875
 876         if (obd->obd_processing_task != 0) {
 877                 /* Someone else is processing this queue, we'll leave it to
 878                  * them.
 879                  */
 880                 wake_up(&obd->obd_next_transno_waitq);
 881                 spin_unlock_bh(&obd->obd_processing_task_lock);
 882                 return 0;
 883         }
 884
 885         /* Nobody is processing, and we know there's (at least) one to process
 886          * now, so we'll do the honours.
 887          */
 888         obd->obd_processing_task = current->pid;
 889         spin_unlock_bh(&obd->obd_processing_task_lock);
 890
 891         process_recovery_queue(obd);
 892         return 0;
 893 }
 894
 895 struct obd_device * target_req2obd(struct ptlrpc_request *req)
 896 {
 897         return req->rq_export->exp_obd;
 898 }
 899
 900 int target_queue_final_reply(struct ptlrpc_request *req, int rc)
 901 {
 902         struct obd_device *obd = target_req2obd(req);
 903         struct ptlrpc_request *saved_req;
 904         struct lustre_msg *reqmsg;
 905         int recovery_done = 0;
 906         int rc2;
 907
 908         LASSERT ((rc == 0) == (req->rq_reply_state != NULL));
 909
 910         if (rc) {
 911                 /* Just like ptlrpc_error, but without the sending. */
 912                 rc = lustre_pack_reply(req, 0, NULL, NULL);
 913                 LASSERT(rc == 0); /* XXX handle this */
 914                 req->rq_type = PTL_RPC_MSG_ERR;
 915         }
 916
 917         LASSERT (!req->rq_reply_state->rs_difficult);
 918         LASSERT(list_empty(&req->rq_list));
 919         /* XXX a bit like the request-dup code in queue_recovery_request */
 920         OBD_ALLOC(saved_req, sizeof *saved_req);
 921         if (!saved_req)
 922                 LBUG();
 923         OBD_ALLOC(reqmsg, req->rq_reqlen);
 924         if (!reqmsg)
 925                 LBUG();
 926         memcpy(saved_req, req, sizeof *saved_req);
 927         memcpy(reqmsg, req->rq_reqmsg, req->rq_reqlen);
 928         /* the copied req takes over the reply state */
 929         req->rq_reply_state = NULL;
 930         req = saved_req;
 931         req->rq_reqmsg = reqmsg;
 932         class_export_get(req->rq_export);
 933         list_add(&req->rq_list, &obd->obd_delayed_reply_queue);
 934
 935         spin_lock_bh(&obd->obd_processing_task_lock);
 936         --obd->obd_recoverable_clients;
 937         recovery_done = (obd->obd_recoverable_clients == 0);
 938         spin_unlock_bh(&obd->obd_processing_task_lock);
 939
 940         if (recovery_done) {
 941                 struct list_head *tmp, *n;
 942                 ldlm_reprocess_all_ns(req->rq_export->exp_obd->obd_namespace);
 943                 CWARN("%s: all clients recovered, sending delayed replies\n",
 944                        obd->obd_name);
 945                 spin_lock_bh(&obd->obd_processing_task_lock);
 946                 obd->obd_recovering = 0;
 947                 target_cancel_recovery_timer(obd);
 948                 spin_unlock_bh(&obd->obd_processing_task_lock);
 949
 950                 /* when recovery finished, cleanup orphans on mds and ost */
 951                 if (OBT(obd) && OBP(obd, postrecov)) {
 952                         rc2 = OBP(obd, postrecov)(obd);
 953                         if (rc2 >= 0)
 954                                 CWARN("%s: all clients recovered, %d MDS "
 955                                       "orphans deleted\n", obd->obd_name, rc2);
 956                         else
 957                                 CERROR("postrecov failed %d\n", rc2);
 958                 }
 959
 960                 list_for_each_safe(tmp, n, &obd->obd_delayed_reply_queue) {
 961                         req = list_entry(tmp, struct ptlrpc_request, rq_list);
 962                         DEBUG_REQ(D_ERROR, req, "delayed:");
 963                         ptlrpc_reply(req);
 964                         class_export_put(req->rq_export);
 965                         list_del(&req->rq_list);
 966                         OBD_FREE(req->rq_reqmsg, req->rq_reqlen);
 967                         OBD_FREE(req, sizeof *req);
 968                 }
 969                 ptlrpc_run_recovery_over_upcall(obd);
 970         } else {
 971                 CWARN("%s: %d recoverable clients remain\n",
 972                        obd->obd_name, obd->obd_recoverable_clients);
 973                 wake_up(&obd->obd_next_transno_waitq);
 974         }
 975
 976         return 1;
 977 }
 978
 979 int
 980 target_send_reply_msg (struct ptlrpc_request *req, int rc, int fail_id)
 981 {
 982         if (OBD_FAIL_CHECK(fail_id | OBD_FAIL_ONCE)) {
 983                 obd_fail_loc |= OBD_FAIL_ONCE | OBD_FAILED;
 984                 DEBUG_REQ(D_ERROR, req, "dropping reply");
 985                 /* NB this does _not_ send with ACK disabled, to simulate
 986                  * sending OK, but timing out for the ACK */
 987                 if (req->rq_reply_state != NULL) {
 988                         if (!req->rq_reply_state->rs_difficult) {
 989                                 lustre_free_reply_state (req->rq_reply_state);
 990                                 req->rq_reply_state = NULL;
 991                         } else {
 992                                 struct ptlrpc_service *svc =
 993                                         req->rq_rqbd->rqbd_srv_ni->sni_service;
 994                                 atomic_inc(&svc->srv_outstanding_replies);
 995                         }
 996                 }
 997                 return (-ECOMM);
 998         }
 999
1000         if (rc) {
1001                 DEBUG_REQ(D_ERROR, req, "processing error (%d)", rc);
1002                 if (req->rq_reply_state == NULL) {
1003                         rc = lustre_pack_reply (req, 0, NULL, NULL);
1004                         if (rc != 0) {
1005                                 CERROR ("can't allocate reply\n");
1006                                 return (rc);
1007                         }
1008                 }
1009                 req->rq_type = PTL_RPC_MSG_ERR;
1010         } else {
1011                 DEBUG_REQ(D_NET, req, "sending reply");
1012         }
1013
1014         return (ptlrpc_send_reply(req, 1));
1015 }
1016
1017 void
1018 target_send_reply(struct ptlrpc_request *req, int rc, int fail_id)
1019 {
1020         int                        netrc;
1021         unsigned long              flags;
1022         struct ptlrpc_reply_state *rs;
1023         struct obd_device         *obd;
1024         struct obd_export         *exp;
1025         struct ptlrpc_srv_ni      *sni;
1026         struct ptlrpc_service     *svc;
1027
1028         sni = req->rq_rqbd->rqbd_srv_ni;
1029         svc = sni->sni_service;
1030
1031         rs = req->rq_reply_state;
1032         if (rs == NULL || !rs->rs_difficult) {
1033                 /* The easy case; no notifiers and reply_out_callback()
1034                  * cleans up (i.e. we can't look inside rs after a
1035                  * successful send) */
1036                 netrc = target_send_reply_msg (req, rc, fail_id);
1037
1038                 LASSERT (netrc == 0 || req->rq_reply_state == NULL);
1039                 return;
1040         }
1041
1042         /* must be an export if locks saved */
1043         LASSERT (req->rq_export != NULL);
1044         /* req/reply consistent */
1045         LASSERT (rs->rs_srv_ni == sni);
1046
1047         /* "fresh" reply */
1048         LASSERT (!rs->rs_scheduled);
1049         LASSERT (!rs->rs_scheduled_ever);
1050         LASSERT (!rs->rs_handled);
1051         LASSERT (!rs->rs_on_net);
1052         LASSERT (rs->rs_export == NULL);
1053         LASSERT (list_empty(&rs->rs_obd_list));
1054         LASSERT (list_empty(&rs->rs_exp_list));
1055
1056         exp = class_export_get (req->rq_export);
1057         obd = exp->exp_obd;
1058
1059         /* disable reply scheduling onto srv_reply_queue while I'm setting up */
1060         rs->rs_scheduled = 1;
1061         rs->rs_on_net    = 1;
1062         rs->rs_xid       = req->rq_xid;
1063         rs->rs_transno   = req->rq_transno;
1064         rs->rs_export    = exp;
1065
1066         spin_lock_irqsave (&obd->obd_uncommitted_replies_lock, flags);
1067
1068         if (rs->rs_transno > obd->obd_last_committed) {
1069                 /* not committed already */
1070                 list_add_tail (&rs->rs_obd_list,
1071                                &obd->obd_uncommitted_replies);
1072         }
1073
1074         spin_unlock (&obd->obd_uncommitted_replies_lock);
1075         spin_lock (&exp->exp_lock);
1076
1077         list_add_tail (&rs->rs_exp_list, &exp->exp_outstanding_replies);
1078
1079         spin_unlock_irqrestore (&exp->exp_lock, flags);
1080
1081         netrc = target_send_reply_msg (req, rc, fail_id);
1082
1083         spin_lock_irqsave (&svc->srv_lock, flags);
1084
1085         svc->srv_n_difficult_replies++;
1086
1087         if (netrc != 0) /* error sending: reply is off the net */
1088                 rs->rs_on_net = 0;
1089
1090         if (!rs->rs_on_net ||                   /* some notifier */
1091             list_empty(&rs->rs_exp_list) ||     /* completed already */
1092             list_empty(&rs->rs_obd_list)) {
1093                 list_add_tail (&rs->rs_list, &svc->srv_reply_queue);
1094                 wake_up (&svc->srv_waitq);
1095         } else {
1096                 list_add (&rs->rs_list, &sni->sni_active_replies);
1097                 rs->rs_scheduled = 0;           /* allow notifier to schedule */
1098         }
1099
1100         spin_unlock_irqrestore (&svc->srv_lock, flags);
1101 }
1102
1103 int target_handle_ping(struct ptlrpc_request *req)
1104 {
1105         return lustre_pack_reply(req, 0, NULL, NULL);
1106 }