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, min(lcfg->lcfg_inllen2,
  97                                                         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         EXIT;
 259
 260         if (rc) {
 261 out_ldlm:
 262                 ldlm_namespace_free(obd->obd_namespace, 0);
 263                 obd->obd_namespace = NULL;
 264 out_disco:
 265                 cli->cl_conn_count--;
 266                 class_disconnect(exp, 0);
 267         } else {
 268                 class_export_put(exp);
 269         }
 270 out_sem:
 271         up(&cli->cl_sem);
 272         return rc;
 273 }
 274
 275 int client_disconnect_export(struct obd_export *exp, int failover)
 276 {
 277         struct obd_device *obd = class_exp2obd(exp);
 278         struct client_obd *cli = &obd->u.cli;
 279         struct obd_import *imp = cli->cl_import;
 280         int rc = 0, err;
 281         ENTRY;
 282
 283         if (!obd) {
 284                 CERROR("invalid export for disconnect: exp %p cookie "LPX64"\n",
 285                        exp, exp ? exp->exp_handle.h_cookie : -1);
 286                 RETURN(-EINVAL);
 287         }
 288
 289         down(&cli->cl_sem);
 290         if (!cli->cl_conn_count) {
 291                 CERROR("disconnecting disconnected device (%s)\n",
 292                        obd->obd_name);
 293                 GOTO(out_sem, rc = -EINVAL);
 294         }
 295
 296         cli->cl_conn_count--;
 297         if (cli->cl_conn_count)
 298                 GOTO(out_no_disconnect, rc = 0);
 299
 300         /* Some non-replayable imports (MDS's OSCs) are pinged, so just
 301          * delete it regardless.  (It's safe to delete an import that was
 302          * never added.) */
 303         (void)ptlrpc_pinger_del_import(imp);
 304
 305         if (obd->obd_namespace != NULL) {
 306                 /* obd_no_recov == local only */
 307                 ldlm_cli_cancel_unused(obd->obd_namespace, NULL,
 308                                        obd->obd_no_recov, NULL);
 309                 ldlm_namespace_free(obd->obd_namespace, obd->obd_no_recov);
 310                 obd->obd_namespace = NULL;
 311         }
 312
 313         /* Yeah, obd_no_recov also (mainly) means "forced shutdown". */
 314         if (obd->obd_no_recov)
 315                 ptlrpc_set_import_active(imp, 0);
 316         else
 317                 rc = ptlrpc_disconnect_import(imp);
 318
 319         EXIT;
 320  out_no_disconnect:
 321         err = class_disconnect(exp, 0);
 322         if (!rc && err)
 323                 rc = err;
 324  out_sem:
 325         up(&cli->cl_sem);
 326         RETURN(rc);
 327 }
 328
 329 /* --------------------------------------------------------------------------
 330  * from old lib/target.c
 331  * -------------------------------------------------------------------------- */
 332
 333 int target_handle_reconnect(struct lustre_handle *conn, struct obd_export *exp,
 334                             struct obd_uuid *cluuid)
 335 {
 336         if (exp->exp_connection) {
 337                 struct lustre_handle *hdl;
 338                 hdl = &exp->exp_imp_reverse->imp_remote_handle;
 339                 /* Might be a re-connect after a partition. */
 340                 if (!memcmp(&conn->cookie, &hdl->cookie, sizeof conn->cookie)) {
 341                         CERROR("%s reconnecting\n", cluuid->uuid);
 342                         conn->cookie = exp->exp_handle.h_cookie;
 343                         RETURN(EALREADY);
 344                 } else {
 345                         CERROR("%s reconnecting from %s, "
 346                                "handle mismatch (ours "LPX64", theirs "
 347                                LPX64")\n", cluuid->uuid,
 348                                exp->exp_connection->c_remote_uuid.uuid,
 349                                hdl->cookie, conn->cookie);
 350                         memset(conn, 0, sizeof *conn);
 351                         RETURN(-EALREADY);
 352                 }
 353         }
 354
 355         conn->cookie = exp->exp_handle.h_cookie;
 356         CDEBUG(D_INFO, "existing export for UUID '%s' at %p\n",
 357                cluuid->uuid, exp);
 358         CDEBUG(D_IOCTL,"connect: cookie "LPX64"\n", conn->cookie);
 359         RETURN(0);
 360 }
 361
 362 int target_handle_connect(struct ptlrpc_request *req, svc_handler_t handler)
 363 {
 364         struct obd_device *target;
 365         struct obd_export *export = NULL;
 366         struct obd_import *revimp;
 367         struct lustre_handle conn;
 368         struct obd_uuid tgtuuid;
 369         struct obd_uuid cluuid;
 370         struct obd_uuid remote_uuid;
 371         struct list_head *p;
 372         char *str, *tmp;
 373         int rc = 0, abort_recovery;
 374         ENTRY;
 375
 376         LASSERT_REQSWAB (req, 0);
 377         str = lustre_msg_string(req->rq_reqmsg, 0, sizeof(tgtuuid) - 1);
 378         if (str == NULL) {
 379                 CERROR("bad target UUID for connect\n");
 380                 GOTO(out, rc = -EINVAL);
 381         }
 382
 383         obd_str2uuid (&tgtuuid, str);
 384         target = class_uuid2obd(&tgtuuid);
 385         if (!target) {
 386                 target = class_name2obd(str);
 387         }
 388
 389         if (!target || target->obd_stopping || !target->obd_set_up) {
 390                 CERROR("UUID '%s' is not available for connect\n", str);
 391                 GOTO(out, rc = -ENODEV);
 392         }
 393
 394         LASSERT_REQSWAB (req, 1);
 395         str = lustre_msg_string(req->rq_reqmsg, 1, sizeof(cluuid) - 1);
 396         if (str == NULL) {
 397                 CERROR("bad client UUID for connect\n");
 398                 GOTO(out, rc = -EINVAL);
 399         }
 400
 401         obd_str2uuid (&cluuid, str);
 402
 403         /* XXX extract a nettype and format accordingly */
 404         snprintf(remote_uuid.uuid, sizeof remote_uuid,
 405                  "NET_"LPX64"_UUID", req->rq_peer.peer_nid);
 406
 407         spin_lock_bh(&target->obd_processing_task_lock);
 408         abort_recovery = target->obd_abort_recovery;
 409         spin_unlock_bh(&target->obd_processing_task_lock);
 410         if (abort_recovery)
 411                 target_abort_recovery(target);
 412
 413         tmp = lustre_msg_buf(req->rq_reqmsg, 2, sizeof conn);
 414         if (tmp == NULL)
 415                 GOTO(out, rc = -EPROTO);
 416
 417         memcpy(&conn, tmp, sizeof conn);
 418
 419         rc = lustre_pack_reply(req, 0, NULL, NULL);
 420         if (rc)
 421                 GOTO(out, rc);
 422
 423         /* lctl gets a backstage, all-access pass. */
 424         if (obd_uuid_equals(&cluuid, &target->obd_uuid))
 425                 goto dont_check_exports;
 426
 427         spin_lock(&target->obd_dev_lock);
 428         list_for_each(p, &target->obd_exports) {
 429                 export = list_entry(p, struct obd_export, exp_obd_chain);
 430                 if (obd_uuid_equals(&cluuid, &export->exp_client_uuid)) {
 431                         spin_unlock(&target->obd_dev_lock);
 432                         LASSERT(export->exp_obd == target);
 433
 434                         rc = target_handle_reconnect(&conn, export, &cluuid);
 435                         break;
 436                 }
 437                 export = NULL;
 438         }
 439         /* If we found an export, we already unlocked. */
 440         if (!export) {
 441                 spin_unlock(&target->obd_dev_lock);
 442         } else if (req->rq_reqmsg->conn_cnt == 1) {
 443                 CERROR("%s reconnected with 1 conn_cnt; cookies not random?\n",
 444                        cluuid.uuid);
 445                 GOTO(out, rc = -EALREADY);
 446         }
 447
 448         /* Tell the client if we're in recovery. */
 449         /* If this is the first client, start the recovery timer */
 450         if (target->obd_recovering) {
 451                 lustre_msg_add_op_flags(req->rq_repmsg, MSG_CONNECT_RECOVERING);
 452                 target_start_recovery_timer(target, handler);
 453         }
 454
 455         /* Tell the client if we support replayable requests */
 456         if (target->obd_replayable)
 457                 lustre_msg_add_op_flags(req->rq_repmsg, MSG_CONNECT_REPLAYABLE);
 458
 459         if (export == NULL) {
 460                 if (target->obd_recovering) {
 461                         CERROR("denying connection for new client %s: "
 462                                "%d clients in recovery for %lds\n", cluuid.uuid,
 463                                target->obd_recoverable_clients,
 464                                (target->obd_recovery_timer.expires-jiffies)/HZ);
 465                         rc = -EBUSY;
 466                 } else {
 467  dont_check_exports:
 468                         rc = obd_connect(&conn, target, &cluuid);
 469                 }
 470         }
 471
 472
 473         /* If all else goes well, this is our RPC return code. */
 474         req->rq_status = 0;
 475
 476         if (rc && rc != EALREADY)
 477                 GOTO(out, rc);
 478
 479         /* XXX track this all the time? */
 480         if (target->obd_recovering) {
 481                 target->obd_connected_clients++;
 482         }
 483
 484         req->rq_repmsg->handle = conn;
 485
 486         /* If the client and the server are the same node, we will already
 487          * have an export that really points to the client's DLM export,
 488          * because we have a shared handles table.
 489          *
 490          * XXX this will go away when shaver stops sending the "connect" handle
 491          * in the real "remote handle" field of the request --phik 24 Apr 2003
 492          */
 493         if (req->rq_export != NULL)
 494                 class_export_put(req->rq_export);
 495
 496         /* ownership of this export ref transfers to the request */
 497         export = req->rq_export = class_conn2export(&conn);
 498         LASSERT(export != NULL);
 499
 500         /* request from liblustre? */
 501         if (lustre_msg_get_op_flags(req->rq_reqmsg) & MSG_CONNECT_LIBCLIENT)
 502                 export->exp_libclient = 1;
 503
 504         if (export->exp_connection != NULL)
 505                 ptlrpc_put_connection(export->exp_connection);
 506         export->exp_connection = ptlrpc_get_connection(&req->rq_peer,
 507                                                        &remote_uuid);
 508
 509         LASSERT(export->exp_conn_cnt < req->rq_reqmsg->conn_cnt);
 510         export->exp_conn_cnt = req->rq_reqmsg->conn_cnt;
 511
 512         if (rc == EALREADY) {
 513                 /* We indicate the reconnection in a flag, not an error code. */
 514                 lustre_msg_add_op_flags(req->rq_repmsg, MSG_CONNECT_RECONNECT);
 515                 GOTO(out, rc = 0);
 516         }
 517
 518         memcpy(&conn, lustre_msg_buf(req->rq_reqmsg, 2, sizeof conn),
 519                sizeof conn);
 520
 521         if (export->exp_imp_reverse != NULL)
 522                 class_destroy_import(export->exp_imp_reverse);
 523         revimp = export->exp_imp_reverse = class_new_import();
 524         revimp->imp_connection = ptlrpc_connection_addref(export->exp_connection);
 525         revimp->imp_client = &export->exp_obd->obd_ldlm_client;
 526         revimp->imp_remote_handle = conn;
 527         revimp->imp_obd = target;
 528         revimp->imp_dlm_fake = 1;
 529         revimp->imp_state = LUSTRE_IMP_FULL;
 530         class_import_put(revimp);
 531 out:
 532         if (rc)
 533                 req->rq_status = rc;
 534         RETURN(rc);
 535 }
 536
 537 int target_handle_disconnect(struct ptlrpc_request *req)
 538 {
 539         struct obd_export *exp;
 540         int rc;
 541         ENTRY;
 542
 543         rc = lustre_pack_reply(req, 0, NULL, NULL);
 544         if (rc)
 545                 RETURN(rc);
 546
 547         /* keep the rq_export around so we can send the reply */
 548         exp = class_export_get(req->rq_export);
 549         req->rq_status = obd_disconnect(exp, 0);
 550         RETURN(0);
 551 }
 552
 553 void target_destroy_export(struct obd_export *exp)
 554 {
 555         /* exports created from last_rcvd data, and "fake"
 556            exports created by lctl don't have an import */
 557         if (exp->exp_imp_reverse != NULL)
 558                 class_destroy_import(exp->exp_imp_reverse);
 559
 560         /* We cancel locks at disconnect time, but this will catch any locks
 561          * granted in a race with recovery-induced disconnect. */
 562         if (exp->exp_obd->obd_namespace != NULL)
 563                 ldlm_cancel_locks_for_export(exp);
 564 }
 565
 566 /*
 567  * Recovery functions
 568  */
 569
 570 static void abort_delayed_replies(struct obd_device *obd)
 571 {
 572         struct ptlrpc_request *req;
 573         struct list_head *tmp, *n;
 574         list_for_each_safe(tmp, n, &obd->obd_delayed_reply_queue) {
 575                 req = list_entry(tmp, struct ptlrpc_request, rq_list);
 576                 DEBUG_REQ(D_ERROR, req, "aborted:");
 577                 req->rq_status = -ENOTCONN;
 578                 req->rq_type = PTL_RPC_MSG_ERR;
 579                 ptlrpc_reply(req);
 580                 class_export_put(req->rq_export);
 581                 list_del(&req->rq_list);
 582                 OBD_FREE(req->rq_reqmsg, req->rq_reqlen);
 583                 OBD_FREE(req, sizeof *req);
 584         }
 585 }
 586
 587 static void abort_recovery_queue(struct obd_device *obd)
 588 {
 589         struct ptlrpc_request *req;
 590         struct list_head *tmp, *n;
 591         int rc;
 592
 593         list_for_each_safe(tmp, n, &obd->obd_recovery_queue) {
 594                 req = list_entry(tmp, struct ptlrpc_request, rq_list);
 595                 DEBUG_REQ(D_ERROR, req, "aborted:");
 596                 req->rq_status = -ENOTCONN;
 597                 req->rq_type = PTL_RPC_MSG_ERR;
 598                 rc = lustre_pack_reply(req, 0, NULL, NULL);
 599                 if (rc == 0) {
 600                         ptlrpc_reply(req);
 601                 } else {
 602                         DEBUG_REQ(D_ERROR, req,
 603                                   "packing failed for abort-reply; skipping");
 604                 }
 605                 list_del(&req->rq_list);
 606                 class_export_put(req->rq_export);
 607                 OBD_FREE(req->rq_reqmsg, req->rq_reqlen);
 608                 OBD_FREE(req, sizeof *req);
 609         }
 610 }
 611
 612 void target_abort_recovery(void *data)
 613 {
 614         struct obd_device *obd = data;
 615         int rc;
 616
 617         CERROR("disconnecting clients and aborting recovery\n");
 618         spin_lock_bh(&obd->obd_processing_task_lock);
 619         if (!obd->obd_recovering) {
 620                 spin_unlock_bh(&obd->obd_processing_task_lock);
 621                 EXIT;
 622                 return;
 623         }
 624
 625         obd->obd_recovering = obd->obd_abort_recovery = 0;
 626
 627         wake_up(&obd->obd_next_transno_waitq);
 628         target_cancel_recovery_timer(obd);
 629         spin_unlock_bh(&obd->obd_processing_task_lock);
 630
 631         class_disconnect_exports(obd, 0);
 632
 633         /* when recovery was aborted, cleanup orphans on mds and ost */
 634         if (OBT(obd) && OBP(obd, postrecov)) {
 635                 rc = OBP(obd, postrecov)(obd);
 636                 if (rc >= 0)
 637                         CWARN("Cleanup %d orphans after recovery was aborted\n", rc);
 638                 else
 639                         CERROR("postrecov failed %d\n", rc);
 640         }
 641
 642         abort_delayed_replies(obd);
 643         abort_recovery_queue(obd);
 644         ptlrpc_run_recovery_over_upcall(obd);
 645 }
 646
 647 static void target_recovery_expired(unsigned long castmeharder)
 648 {
 649         struct obd_device *obd = (struct obd_device *)castmeharder;
 650         CERROR("recovery timed out, aborting\n");
 651         spin_lock_bh(&obd->obd_processing_task_lock);
 652         obd->obd_abort_recovery = 1;
 653         wake_up(&obd->obd_next_transno_waitq);
 654         spin_unlock_bh(&obd->obd_processing_task_lock);
 655 }
 656
 657
 658 /* obd_processing_task_lock should be held */
 659 void target_cancel_recovery_timer(struct obd_device *obd)
 660 {
 661         CDEBUG(D_HA, "%s: cancel recovery timer\n", obd->obd_name);
 662         del_timer(&obd->obd_recovery_timer);
 663 }
 664
 665 static void reset_recovery_timer(struct obd_device *obd)
 666 {
 667         spin_lock_bh(&obd->obd_processing_task_lock);
 668         if (!obd->obd_recovering) {
 669                 spin_unlock_bh(&obd->obd_processing_task_lock);
 670                 return;
 671         }
 672         CDEBUG(D_HA, "timer will expire in %u seconds\n",
 673                OBD_RECOVERY_TIMEOUT / HZ);
 674         mod_timer(&obd->obd_recovery_timer, jiffies + OBD_RECOVERY_TIMEOUT);
 675         spin_unlock_bh(&obd->obd_processing_task_lock);
 676 }
 677
 678
 679 /* Only start it the first time called */
 680 void target_start_recovery_timer(struct obd_device *obd, svc_handler_t handler)
 681 {
 682         spin_lock_bh(&obd->obd_processing_task_lock);
 683         if (obd->obd_recovery_handler) {
 684                 spin_unlock_bh(&obd->obd_processing_task_lock);
 685                 return;
 686         }
 687         CWARN("%s: starting recovery timer (%us)\n", obd->obd_name,
 688                OBD_RECOVERY_TIMEOUT / HZ);
 689         obd->obd_recovery_handler = handler;
 690         obd->obd_recovery_timer.function = target_recovery_expired;
 691         obd->obd_recovery_timer.data = (unsigned long)obd;
 692         init_timer(&obd->obd_recovery_timer);
 693         spin_unlock_bh(&obd->obd_processing_task_lock);
 694
 695         reset_recovery_timer(obd);
 696 }
 697
 698 static int check_for_next_transno(struct obd_device *obd)
 699 {
 700         struct ptlrpc_request *req;
 701         int wake_up = 0, connected, completed, queue_len, max;
 702         __u64 next_transno, req_transno;
 703
 704         spin_lock_bh(&obd->obd_processing_task_lock);
 705         req = list_entry(obd->obd_recovery_queue.next,
 706                          struct ptlrpc_request, rq_list);
 707         max = obd->obd_max_recoverable_clients;
 708         req_transno = req->rq_reqmsg->transno;
 709         connected = obd->obd_connected_clients;
 710         completed = max - obd->obd_recoverable_clients;
 711         queue_len = obd->obd_requests_queued_for_recovery;
 712         next_transno = obd->obd_next_recovery_transno;
 713
 714         if (obd->obd_abort_recovery) {
 715                 CDEBUG(D_HA, "waking for aborted recovery\n");
 716                 wake_up = 1;
 717         } else if (!obd->obd_recovering) {
 718                 CDEBUG(D_HA, "waking for completed recovery (?)\n");
 719                 wake_up = 1;
 720         } else if (req_transno == next_transno) {
 721                 CDEBUG(D_HA, "waking for next ("LPD64")\n", next_transno);
 722                 wake_up = 1;
 723         } else if (queue_len + completed == max) {
 724                 CDEBUG(D_ERROR,
 725                        "waking for skipped transno (skip: "LPD64
 726                        ", ql: %d, comp: %d, conn: %d, next: "LPD64")\n",
 727                        next_transno, queue_len, completed, max, req_transno);
 728                 obd->obd_next_recovery_transno = req_transno;
 729                 wake_up = 1;
 730         }
 731         spin_unlock_bh(&obd->obd_processing_task_lock);
 732         LASSERT(req->rq_reqmsg->transno >= next_transno);
 733         return wake_up;
 734 }
 735
 736 static void process_recovery_queue(struct obd_device *obd)
 737 {
 738         struct ptlrpc_request *req;
 739         int abort_recovery = 0;
 740         struct l_wait_info lwi = { 0 };
 741         ENTRY;
 742
 743         for (;;) {
 744                 spin_lock_bh(&obd->obd_processing_task_lock);
 745                 LASSERT(obd->obd_processing_task == current->pid);
 746                 req = list_entry(obd->obd_recovery_queue.next,
 747                                  struct ptlrpc_request, rq_list);
 748
 749                 if (req->rq_reqmsg->transno != obd->obd_next_recovery_transno) {
 750                         spin_unlock_bh(&obd->obd_processing_task_lock);
 751                         CDEBUG(D_HA, "Waiting for transno "LPD64" (1st is "
 752                                LPD64")\n",
 753                                obd->obd_next_recovery_transno,
 754                                req->rq_reqmsg->transno);
 755                         l_wait_event(obd->obd_next_transno_waitq,
 756                                      check_for_next_transno(obd), &lwi);
 757                         spin_lock_bh(&obd->obd_processing_task_lock);
 758                         abort_recovery = obd->obd_abort_recovery;
 759                         spin_unlock_bh(&obd->obd_processing_task_lock);
 760                         if (abort_recovery) {
 761                                 target_abort_recovery(obd);
 762                                 return;
 763                         }
 764                         continue;
 765                 }
 766                 list_del_init(&req->rq_list);
 767                 obd->obd_requests_queued_for_recovery--;
 768                 spin_unlock_bh(&obd->obd_processing_task_lock);
 769
 770                 DEBUG_REQ(D_HA, req, "processing: ");
 771                 (void)obd->obd_recovery_handler(req);
 772                 obd->obd_replayed_requests++;
 773                 reset_recovery_timer(obd);
 774                 /* bug 1580: decide how to properly sync() in recovery */
 775                 //mds_fsync_super(mds->mds_sb);
 776                 class_export_put(req->rq_export);
 777                 OBD_FREE(req->rq_reqmsg, req->rq_reqlen);
 778                 OBD_FREE(req, sizeof *req);
 779                 spin_lock_bh(&obd->obd_processing_task_lock);
 780                 obd->obd_next_recovery_transno++;
 781                 if (list_empty(&obd->obd_recovery_queue)) {
 782                         obd->obd_processing_task = 0;
 783                         spin_unlock_bh(&obd->obd_processing_task_lock);
 784                         break;
 785                 }
 786                 spin_unlock_bh(&obd->obd_processing_task_lock);
 787         }
 788         EXIT;
 789 }
 790
 791 int target_queue_recovery_request(struct ptlrpc_request *req,
 792                                   struct obd_device *obd)
 793 {
 794         struct list_head *tmp;
 795         int inserted = 0;
 796         __u64 transno = req->rq_reqmsg->transno;
 797         struct ptlrpc_request *saved_req;
 798         struct lustre_msg *reqmsg;
 799
 800         /* CAVEAT EMPTOR: The incoming request message has been swabbed
 801          * (i.e. buflens etc are in my own byte order), but type-dependent
 802          * buffers (eg mds_body, ost_body etc) have NOT been swabbed. */
 803
 804         if (!transno) {
 805                 INIT_LIST_HEAD(&req->rq_list);
 806                 DEBUG_REQ(D_HA, req, "not queueing");
 807                 return 1;
 808         }
 809
 810         /* XXX If I were a real man, these LBUGs would be sane cleanups. */
 811         /* XXX just like the request-dup code in queue_final_reply */
 812         OBD_ALLOC(saved_req, sizeof *saved_req);
 813         if (!saved_req)
 814                 LBUG();
 815         OBD_ALLOC(reqmsg, req->rq_reqlen);
 816         if (!reqmsg)
 817                 LBUG();
 818
 819         spin_lock_bh(&obd->obd_processing_task_lock);
 820
 821         /* If we're processing the queue, we want don't want to queue this
 822          * message.
 823          *
 824          * Also, if this request has a transno less than the one we're waiting
 825          * for, we should process it now.  It could (and currently always will)
 826          * be an open request for a descriptor that was opened some time ago.
 827          */
 828         if (obd->obd_processing_task == current->pid ||
 829             transno < obd->obd_next_recovery_transno) {
 830                 /* Processing the queue right now, don't re-add. */
 831                 LASSERT(list_empty(&req->rq_list));
 832                 spin_unlock_bh(&obd->obd_processing_task_lock);
 833                 OBD_FREE(reqmsg, req->rq_reqlen);
 834                 OBD_FREE(saved_req, sizeof *saved_req);
 835                 return 1;
 836         }
 837
 838         memcpy(saved_req, req, sizeof *req);
 839         memcpy(reqmsg, req->rq_reqmsg, req->rq_reqlen);
 840         req = saved_req;
 841         req->rq_reqmsg = reqmsg;
 842         class_export_get(req->rq_export);
 843         INIT_LIST_HEAD(&req->rq_list);
 844
 845         /* XXX O(n^2) */
 846         list_for_each(tmp, &obd->obd_recovery_queue) {
 847                 struct ptlrpc_request *reqiter =
 848                         list_entry(tmp, struct ptlrpc_request, rq_list);
 849
 850                 if (reqiter->rq_reqmsg->transno > transno) {
 851                         list_add_tail(&req->rq_list, &reqiter->rq_list);
 852                         inserted = 1;
 853                         break;
 854                 }
 855         }
 856
 857         if (!inserted) {
 858                 list_add_tail(&req->rq_list, &obd->obd_recovery_queue);
 859         }
 860
 861         obd->obd_requests_queued_for_recovery++;
 862
 863         if (obd->obd_processing_task != 0) {
 864                 /* Someone else is processing this queue, we'll leave it to
 865                  * them.
 866                  */
 867                 wake_up(&obd->obd_next_transno_waitq);
 868                 spin_unlock_bh(&obd->obd_processing_task_lock);
 869                 return 0;
 870         }
 871
 872         /* Nobody is processing, and we know there's (at least) one to process
 873          * now, so we'll do the honours.
 874          */
 875         obd->obd_processing_task = current->pid;
 876         spin_unlock_bh(&obd->obd_processing_task_lock);
 877
 878         process_recovery_queue(obd);
 879         return 0;
 880 }
 881
 882 struct obd_device * target_req2obd(struct ptlrpc_request *req)
 883 {
 884         return req->rq_export->exp_obd;
 885 }
 886
 887 int target_queue_final_reply(struct ptlrpc_request *req, int rc)
 888 {
 889         struct obd_device *obd = target_req2obd(req);
 890         struct ptlrpc_request *saved_req;
 891         struct lustre_msg *reqmsg;
 892         int recovery_done = 0;
 893         int rc2;
 894
 895         LASSERT ((rc == 0) == (req->rq_reply_state != NULL));
 896
 897         if (rc) {
 898                 /* Just like ptlrpc_error, but without the sending. */
 899                 rc = lustre_pack_reply(req, 0, NULL, NULL);
 900                 LASSERT(rc == 0); /* XXX handle this */
 901                 req->rq_type = PTL_RPC_MSG_ERR;
 902         }
 903
 904         LASSERT (!req->rq_reply_state->rs_difficult);
 905         LASSERT(list_empty(&req->rq_list));
 906         /* XXX a bit like the request-dup code in queue_recovery_request */
 907         OBD_ALLOC(saved_req, sizeof *saved_req);
 908         if (!saved_req)
 909                 LBUG();
 910         OBD_ALLOC(reqmsg, req->rq_reqlen);
 911         if (!reqmsg)
 912                 LBUG();
 913         memcpy(saved_req, req, sizeof *saved_req);
 914         memcpy(reqmsg, req->rq_reqmsg, req->rq_reqlen);
 915         /* the copied req takes over the reply state */
 916         req->rq_reply_state = NULL;
 917         req = saved_req;
 918         req->rq_reqmsg = reqmsg;
 919         class_export_get(req->rq_export);
 920         list_add(&req->rq_list, &obd->obd_delayed_reply_queue);
 921
 922         spin_lock_bh(&obd->obd_processing_task_lock);
 923         --obd->obd_recoverable_clients;
 924         recovery_done = (obd->obd_recoverable_clients == 0);
 925         spin_unlock_bh(&obd->obd_processing_task_lock);
 926
 927         if (recovery_done) {
 928                 struct list_head *tmp, *n;
 929                 ldlm_reprocess_all_ns(req->rq_export->exp_obd->obd_namespace);
 930                 CWARN("%s: all clients recovered, sending delayed replies\n",
 931                        obd->obd_name);
 932                 spin_lock_bh(&obd->obd_processing_task_lock);
 933                 obd->obd_recovering = 0;
 934                 target_cancel_recovery_timer(obd);
 935                 spin_unlock_bh(&obd->obd_processing_task_lock);
 936
 937                 /* when recovery finished, cleanup orphans on mds and ost */
 938                 if (OBT(obd) && OBP(obd, postrecov)) {
 939                         rc2 = OBP(obd, postrecov)(obd);
 940                         if (rc2 >= 0)
 941                                 CWARN("%s: all clients recovered, %d MDS "
 942                                       "orphans deleted\n", obd->obd_name, rc2);
 943                         else
 944                                 CERROR("postrecov failed %d\n", rc2);
 945                 }
 946
 947                 list_for_each_safe(tmp, n, &obd->obd_delayed_reply_queue) {
 948                         req = list_entry(tmp, struct ptlrpc_request, rq_list);
 949                         DEBUG_REQ(D_ERROR, req, "delayed:");
 950                         ptlrpc_reply(req);
 951                         class_export_put(req->rq_export);
 952                         list_del(&req->rq_list);
 953                         OBD_FREE(req->rq_reqmsg, req->rq_reqlen);
 954                         OBD_FREE(req, sizeof *req);
 955                 }
 956                 ptlrpc_run_recovery_over_upcall(obd);
 957         } else {
 958                 CWARN("%s: %d recoverable clients remain\n",
 959                        obd->obd_name, obd->obd_recoverable_clients);
 960                 wake_up(&obd->obd_next_transno_waitq);
 961         }
 962
 963         return 1;
 964 }
 965
 966 int
 967 target_send_reply_msg (struct ptlrpc_request *req, int rc, int fail_id)
 968 {
 969         if (OBD_FAIL_CHECK(fail_id | OBD_FAIL_ONCE)) {
 970                 obd_fail_loc |= OBD_FAIL_ONCE | OBD_FAILED;
 971                 DEBUG_REQ(D_ERROR, req, "dropping reply");
 972                 /* NB this does _not_ send with ACK disabled, to simulate
 973                  * sending OK, but timing out for the ACK */
 974                 if (req->rq_reply_state != NULL) {
 975                         if (!req->rq_reply_state->rs_difficult) {
 976                                 lustre_free_reply_state (req->rq_reply_state);
 977                                 req->rq_reply_state = NULL;
 978                         } else {
 979                                 struct ptlrpc_service *svc =
 980                                         req->rq_rqbd->rqbd_srv_ni->sni_service;
 981                                 atomic_inc(&svc->srv_outstanding_replies);
 982                         }
 983                 }
 984                 return (-ECOMM);
 985         }
 986
 987         if (rc) {
 988                 DEBUG_REQ(D_ERROR, req, "processing error (%d)", rc);
 989                 if (req->rq_reply_state == NULL) {
 990                         rc = lustre_pack_reply (req, 0, NULL, NULL);
 991                         if (rc != 0) {
 992                                 CERROR ("can't allocate reply\n");
 993                                 return (rc);
 994                         }
 995                 }
 996                 req->rq_type = PTL_RPC_MSG_ERR;
 997         } else {
 998                 DEBUG_REQ(D_NET, req, "sending reply");
 999         }
1000
1001         return (ptlrpc_send_reply(req, 1));
1002 }
1003
1004 void
1005 target_send_reply(struct ptlrpc_request *req, int rc, int fail_id)
1006 {
1007         int                        netrc;
1008         unsigned long              flags;
1009         struct ptlrpc_reply_state *rs;
1010         struct obd_device         *obd;
1011         struct obd_export         *exp;
1012         struct ptlrpc_srv_ni      *sni;
1013         struct ptlrpc_service     *svc;
1014
1015         sni = req->rq_rqbd->rqbd_srv_ni;
1016         svc = sni->sni_service;
1017
1018         rs = req->rq_reply_state;
1019         if (rs == NULL || !rs->rs_difficult) {
1020                 /* The easy case; no notifiers and reply_out_callback()
1021                  * cleans up (i.e. we can't look inside rs after a
1022                  * successful send) */
1023                 netrc = target_send_reply_msg (req, rc, fail_id);
1024
1025                 LASSERT (netrc == 0 || req->rq_reply_state == NULL);
1026                 return;
1027         }
1028
1029         /* must be an export if locks saved */
1030         LASSERT (req->rq_export != NULL);
1031         /* req/reply consistent */
1032         LASSERT (rs->rs_srv_ni == sni);
1033
1034         /* "fresh" reply */
1035         LASSERT (!rs->rs_scheduled);
1036         LASSERT (!rs->rs_scheduled_ever);
1037         LASSERT (!rs->rs_handled);
1038         LASSERT (!rs->rs_on_net);
1039         LASSERT (rs->rs_export == NULL);
1040         LASSERT (list_empty(&rs->rs_obd_list));
1041         LASSERT (list_empty(&rs->rs_exp_list));
1042
1043         exp = class_export_get (req->rq_export);
1044         obd = exp->exp_obd;
1045
1046         /* disable reply scheduling onto srv_reply_queue while I'm setting up */
1047         rs->rs_scheduled = 1;
1048         rs->rs_on_net    = 1;
1049         rs->rs_xid       = req->rq_xid;
1050         rs->rs_transno   = req->rq_transno;
1051         rs->rs_export    = exp;
1052
1053         spin_lock_irqsave (&obd->obd_uncommitted_replies_lock, flags);
1054
1055         if (rs->rs_transno > obd->obd_last_committed) {
1056                 /* not committed already */
1057                 list_add_tail (&rs->rs_obd_list,
1058                                &obd->obd_uncommitted_replies);
1059         }
1060
1061         spin_unlock (&obd->obd_uncommitted_replies_lock);
1062         spin_lock (&exp->exp_lock);
1063
1064         list_add_tail (&rs->rs_exp_list, &exp->exp_outstanding_replies);
1065
1066         spin_unlock_irqrestore (&exp->exp_lock, flags);
1067
1068         netrc = target_send_reply_msg (req, rc, fail_id);
1069
1070         spin_lock_irqsave (&svc->srv_lock, flags);
1071
1072         svc->srv_n_difficult_replies++;
1073
1074         if (netrc != 0) /* error sending: reply is off the net */
1075                 rs->rs_on_net = 0;
1076
1077         if (!rs->rs_on_net ||                   /* some notifier */
1078             list_empty(&rs->rs_exp_list) ||     /* completed already */
1079             list_empty(&rs->rs_obd_list)) {
1080                 list_add_tail (&rs->rs_list, &svc->srv_reply_queue);
1081                 wake_up (&svc->srv_waitq);
1082         } else {
1083                 list_add (&rs->rs_list, &sni->sni_active_replies);
1084                 rs->rs_scheduled = 0;           /* allow notifier to schedule */
1085         }
1086
1087         spin_unlock_irqrestore (&svc->srv_lock, flags);
1088 }
1089
1090 int target_handle_ping(struct ptlrpc_request *req)
1091 {
1092         return lustre_pack_reply(req, 0, NULL, NULL);
1093 }