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
  95         sema_init(&cli->cl_sem, 1);
  96         cli->cl_conn_count = 0;
  97         memcpy(server_uuid.uuid, lcfg->lcfg_inlbuf2, MIN(lcfg->lcfg_inllen2,
  98                                                         sizeof(server_uuid)));
  99
 100         init_MUTEX(&cli->cl_dirty_sem);
 101         cli->cl_dirty = 0;
 102         cli->cl_dirty_granted = 0;
 103         cli->cl_dirty_max = OSC_MAX_DIRTY_DEFAULT * 1024 * 1024;
 104         cli->cl_ost_can_grant = 1;
 105         INIT_LIST_HEAD(&cli->cl_cache_waiters);
 106         INIT_LIST_HEAD(&cli->cl_loi_ready_list);
 107         INIT_LIST_HEAD(&cli->cl_loi_write_list);
 108         spin_lock_init(&cli->cl_loi_list_lock);
 109         cli->cl_brw_in_flight = 0;
 110         spin_lock_init(&cli->cl_read_rpc_hist.oh_lock);
 111         spin_lock_init(&cli->cl_write_rpc_hist.oh_lock);
 112         spin_lock_init(&cli->cl_read_page_hist.oh_lock);
 113         spin_lock_init(&cli->cl_write_page_hist.oh_lock);
 114         cli->cl_max_pages_per_rpc = PTL_MD_MAX_PAGES;
 115         cli->cl_max_rpcs_in_flight = OSC_MAX_RIF_DEFAULT;
 116
 117         ldlm_get_ref();
 118         if (rc) {
 119                 CERROR("ldlm_get_ref failed: %d\n", rc);
 120                 GOTO(err, rc);
 121         }
 122
 123         conn = ptlrpc_uuid_to_connection(&server_uuid);
 124         if (conn == NULL)
 125                 GOTO(err_ldlm, rc = -ENOENT);
 126
 127         ptlrpc_init_client(rq_portal, rp_portal, name,
 128                            &obddev->obd_ldlm_client);
 129
 130         imp = class_new_import();
 131         if (imp == NULL) {
 132                 ptlrpc_put_connection(conn);
 133                 GOTO(err_ldlm, rc = -ENOENT);
 134         }
 135         imp->imp_connection = conn;
 136         imp->imp_client = &obddev->obd_ldlm_client;
 137         imp->imp_obd = obddev;
 138         imp->imp_connect_op = connect_op;
 139         imp->imp_generation = 0;
 140         imp->imp_initial_recov = 1;
 141         INIT_LIST_HEAD(&imp->imp_pinger_chain);
 142         memcpy(imp->imp_target_uuid.uuid, lcfg->lcfg_inlbuf1,
 143               lcfg->lcfg_inllen1);
 144         class_import_put(imp);
 145
 146         cli->cl_import = imp;
 147         cli->cl_max_mds_easize = sizeof(struct lov_mds_md);
 148         cli->cl_max_mds_cookiesize = sizeof(struct llog_cookie);
 149         cli->cl_sandev = to_kdev_t(0);
 150
 151         if (lcfg->lcfg_inllen3 != 0) {
 152                 if (!strcmp(lcfg->lcfg_inlbuf3, "inactive")) {
 153                         CDEBUG(D_HA, "marking %s %s->%s as inactive\n",
 154                                name, obddev->obd_name,
 155                                imp->imp_target_uuid.uuid);
 156                         imp->imp_invalid = 1;
 157
 158                         if (lcfg->lcfg_inllen4 != 0)
 159                                 mgmt_name = lcfg->lcfg_inlbuf4;
 160                 } else {
 161                         mgmt_name = lcfg->lcfg_inlbuf3;
 162                 }
 163         }
 164
 165         if (mgmt_name != NULL) {
 166                 /* Register with management client if we need to. */
 167                 CDEBUG(D_HA, "%s registering with %s for events about %s\n",
 168                        obddev->obd_name, mgmt_name, server_uuid.uuid);
 169
 170                 mgmt_obd = class_name2obd(mgmt_name);
 171                 if (!mgmt_obd) {
 172                         CERROR("can't find mgmtcli %s to register\n",
 173                                mgmt_name);
 174                         GOTO(err_import, rc = -ENOSYS);
 175                 }
 176
 177                 register_f = inter_module_get("mgmtcli_register_for_events");
 178                 if (!register_f) {
 179                         CERROR("can't i_m_g mgmtcli_register_for_events\n");
 180                         GOTO(err_import, rc = -ENOSYS);
 181                 }
 182
 183                 rc = register_f(mgmt_obd, obddev, &imp->imp_target_uuid);
 184                 inter_module_put("mgmtcli_register_for_events");
 185
 186                 if (!rc)
 187                         cli->cl_mgmtcli_obd = mgmt_obd;
 188         }
 189
 190         RETURN(rc);
 191
 192 err_import:
 193         class_destroy_import(imp);
 194 err_ldlm:
 195         ldlm_put_ref(0);
 196 err:
 197         RETURN(rc);
 198
 199 }
 200
 201 int client_obd_cleanup(struct obd_device *obddev, int flags)
 202 {
 203         struct client_obd *cli = &obddev->u.cli;
 204
 205         if (!cli->cl_import)
 206                 RETURN(-EINVAL);
 207         if (cli->cl_mgmtcli_obd) {
 208                 mgmtcli_deregister_for_events_t dereg_f;
 209
 210                 dereg_f = inter_module_get("mgmtcli_deregister_for_events");
 211                 dereg_f(cli->cl_mgmtcli_obd, obddev);
 212                 inter_module_put("mgmtcli_deregister_for_events");
 213         }
 214         class_destroy_import(cli->cl_import);
 215         cli->cl_import = NULL;
 216
 217         ldlm_put_ref(flags & OBD_OPT_FORCE);
 218
 219         RETURN(0);
 220 }
 221
 222 int client_connect_import(struct lustre_handle *dlm_handle,
 223                           struct obd_device *obd,
 224                           struct obd_uuid *cluuid)
 225 {
 226         struct client_obd *cli = &obd->u.cli;
 227         struct obd_import *imp = cli->cl_import;
 228         struct obd_export *exp;
 229         int rc;
 230         ENTRY;
 231
 232         down(&cli->cl_sem);
 233         rc = class_connect(dlm_handle, obd, cluuid);
 234         if (rc)
 235                 GOTO(out_sem, rc);
 236
 237         cli->cl_conn_count++;
 238         if (cli->cl_conn_count > 1)
 239                 GOTO(out_sem, rc);
 240         exp = class_conn2export(dlm_handle);
 241
 242         if (obd->obd_namespace != NULL)
 243                 CERROR("already have namespace!\n");
 244         obd->obd_namespace = ldlm_namespace_new(obd->obd_name,
 245                                                 LDLM_NAMESPACE_CLIENT);
 246         if (obd->obd_namespace == NULL)
 247                 GOTO(out_disco, rc = -ENOMEM);
 248
 249         imp->imp_dlm_handle = *dlm_handle;
 250         rc = ptlrpc_init_import(imp);
 251         if (rc != 0)
 252                 GOTO(out_ldlm, rc);
 253
 254         exp->exp_connection = ptlrpc_connection_addref(imp->imp_connection);
 255         rc = ptlrpc_connect_import(imp, NULL);
 256         if (rc != 0) {
 257                 LASSERT (imp->imp_state == LUSTRE_IMP_DISCON);
 258                 GOTO(out_ldlm, rc);
 259         }
 260
 261         EXIT;
 262
 263         if (rc) {
 264 out_ldlm:
 265                 ldlm_namespace_free(obd->obd_namespace, 0);
 266                 obd->obd_namespace = NULL;
 267 out_disco:
 268                 cli->cl_conn_count--;
 269                 class_disconnect(exp, 0);
 270         } else {
 271                 class_export_put(exp);
 272         }
 273 out_sem:
 274         up(&cli->cl_sem);
 275         return rc;
 276 }
 277
 278 int client_disconnect_export(struct obd_export *exp, int failover)
 279 {
 280         struct obd_device *obd = class_exp2obd(exp);
 281         struct client_obd *cli = &obd->u.cli;
 282         struct obd_import *imp = cli->cl_import;
 283         int rc = 0, err;
 284         ENTRY;
 285
 286         if (!obd) {
 287                 CERROR("invalid export for disconnect: exp %p cookie "LPX64"\n",
 288                        exp, exp ? exp->exp_handle.h_cookie : -1);
 289                 RETURN(-EINVAL);
 290         }
 291
 292         down(&cli->cl_sem);
 293         if (!cli->cl_conn_count) {
 294                 CERROR("disconnecting disconnected device (%s)\n",
 295                        obd->obd_name);
 296                 GOTO(out_sem, rc = -EINVAL);
 297         }
 298
 299         cli->cl_conn_count--;
 300         if (cli->cl_conn_count)
 301                 GOTO(out_no_disconnect, rc = 0);
 302
 303         /* Some non-replayable imports (MDS's OSCs) are pinged, so just
 304          * delete it regardless.  (It's safe to delete an import that was
 305          * never added.) */
 306         (void)ptlrpc_pinger_del_import(imp);
 307
 308         if (obd->obd_namespace != NULL) {
 309                 /* obd_no_recov == local only */
 310                 ldlm_cli_cancel_unused(obd->obd_namespace, NULL,
 311                                        obd->obd_no_recov, NULL);
 312                 ldlm_namespace_free(obd->obd_namespace, obd->obd_no_recov);
 313                 obd->obd_namespace = NULL;
 314         }
 315
 316         /* Yeah, obd_no_recov also (mainly) means "forced shutdown". */
 317         if (obd->obd_no_recov)
 318                 ptlrpc_set_import_active(imp, 0);
 319         else
 320                 rc = ptlrpc_disconnect_import(imp);
 321
 322         EXIT;
 323  out_no_disconnect:
 324         err = class_disconnect(exp, 0);
 325         if (!rc && err)
 326                 rc = err;
 327  out_sem:
 328         up(&cli->cl_sem);
 329         RETURN(rc);
 330 }
 331
 332 /* --------------------------------------------------------------------------
 333  * from old lib/target.c
 334  * -------------------------------------------------------------------------- */
 335
 336 int target_handle_reconnect(struct lustre_handle *conn, struct obd_export *exp,
 337                             struct obd_uuid *cluuid)
 338 {
 339         if (exp->exp_connection) {
 340                 struct lustre_handle *hdl;
 341                 hdl = &exp->exp_imp_reverse->imp_remote_handle;
 342                 /* Might be a re-connect after a partition. */
 343                 if (!memcmp(&conn->cookie, &hdl->cookie, sizeof conn->cookie)) {
 344                         CERROR("%s reconnecting\n", cluuid->uuid);
 345                         conn->cookie = exp->exp_handle.h_cookie;
 346                         RETURN(EALREADY);
 347                 } else {
 348                         CERROR("%s reconnecting from %s, "
 349                                "handle mismatch (ours "LPX64", theirs "
 350                                LPX64")\n", cluuid->uuid,
 351                                exp->exp_connection->c_remote_uuid.uuid,
 352                                hdl->cookie, conn->cookie);
 353                         memset(conn, 0, sizeof *conn);
 354                         RETURN(-EALREADY);
 355                 }
 356         }
 357
 358         conn->cookie = exp->exp_handle.h_cookie;
 359         CDEBUG(D_INFO, "existing export for UUID '%s' at %p\n",
 360                cluuid->uuid, exp);
 361         CDEBUG(D_IOCTL,"connect: cookie "LPX64"\n", conn->cookie);
 362         RETURN(0);
 363 }
 364
 365 int target_handle_connect(struct ptlrpc_request *req, svc_handler_t handler)
 366 {
 367         struct obd_device *target;
 368         struct obd_export *export = NULL;
 369         struct obd_import *revimp;
 370         struct lustre_handle conn;
 371         struct obd_uuid tgtuuid;
 372         struct obd_uuid cluuid;
 373         struct obd_uuid remote_uuid;
 374         struct list_head *p;
 375         char *str, *tmp;
 376         int rc = 0, abort_recovery;
 377         ENTRY;
 378
 379         LASSERT_REQSWAB (req, 0);
 380         str = lustre_msg_string(req->rq_reqmsg, 0, sizeof(tgtuuid) - 1);
 381         if (str == NULL) {
 382                 CERROR("bad target UUID for connect\n");
 383                 GOTO(out, rc = -EINVAL);
 384         }
 385
 386         obd_str2uuid (&tgtuuid, str);
 387         target = class_uuid2obd(&tgtuuid);
 388         if (!target) {
 389                 target = class_name2obd(str);
 390         }
 391
 392         if (!target || target->obd_stopping || !target->obd_set_up) {
 393                 CERROR("UUID '%s' is not available for connect\n", str);
 394                 GOTO(out, rc = -ENODEV);
 395         }
 396
 397         LASSERT_REQSWAB (req, 1);
 398         str = lustre_msg_string(req->rq_reqmsg, 1, sizeof(cluuid) - 1);
 399         if (str == NULL) {
 400                 CERROR("bad client UUID for connect\n");
 401                 GOTO(out, rc = -EINVAL);
 402         }
 403
 404         obd_str2uuid (&cluuid, str);
 405
 406         /* XXX extract a nettype and format accordingly */
 407         snprintf(remote_uuid.uuid, sizeof remote_uuid,
 408                  "NET_"LPX64"_UUID", req->rq_peer.peer_nid);
 409
 410         spin_lock_bh(&target->obd_processing_task_lock);
 411         abort_recovery = target->obd_abort_recovery;
 412         spin_unlock_bh(&target->obd_processing_task_lock);
 413         if (abort_recovery)
 414                 target_abort_recovery(target);
 415
 416         tmp = lustre_msg_buf(req->rq_reqmsg, 2, sizeof conn);
 417         if (tmp == NULL)
 418                 GOTO(out, rc = -EPROTO);
 419
 420         memcpy(&conn, tmp, sizeof conn);
 421
 422         rc = lustre_pack_reply(req, 0, NULL, NULL);
 423         if (rc)
 424                 GOTO(out, rc);
 425
 426         /* lctl gets a backstage, all-access pass. */
 427         if (obd_uuid_equals(&cluuid, &target->obd_uuid))
 428                 goto dont_check_exports;
 429
 430         spin_lock(&target->obd_dev_lock);
 431         list_for_each(p, &target->obd_exports) {
 432                 export = list_entry(p, struct obd_export, exp_obd_chain);
 433                 if (obd_uuid_equals(&cluuid, &export->exp_client_uuid)) {
 434                         spin_unlock(&target->obd_dev_lock);
 435                         LASSERT(export->exp_obd == target);
 436
 437                         rc = target_handle_reconnect(&conn, export, &cluuid);
 438                         break;
 439                 }
 440                 export = NULL;
 441         }
 442         /* If we found an export, we already unlocked. */
 443         if (!export) {
 444                 spin_unlock(&target->obd_dev_lock);
 445         } else if (req->rq_reqmsg->conn_cnt == 1) {
 446                 CERROR("%s reconnected with 1 conn_cnt; cookies not random?\n",
 447                        cluuid.uuid);
 448                 GOTO(out, rc = -EALREADY);
 449         }
 450
 451         /* Tell the client if we're in recovery. */
 452         /* If this is the first client, start the recovery timer */
 453         if (target->obd_recovering) {
 454                 lustre_msg_add_op_flags(req->rq_repmsg, MSG_CONNECT_RECOVERING);
 455                 target_start_recovery_timer(target, handler);
 456         }
 457
 458         /* Tell the client if we support replayable requests */
 459         if (target->obd_replayable)
 460                 lustre_msg_add_op_flags(req->rq_repmsg, MSG_CONNECT_REPLAYABLE);
 461
 462         if (export == NULL) {
 463                 if (target->obd_recovering) {
 464                         CERROR("denying connection for new client %s: "
 465                                "%d clients in recovery for %lds\n", cluuid.uuid,
 466                                target->obd_recoverable_clients,
 467                                (target->obd_recovery_timer.expires-jiffies)/HZ);
 468                         rc = -EBUSY;
 469                 } else {
 470  dont_check_exports:
 471                         rc = obd_connect(&conn, target, &cluuid);
 472                 }
 473         }
 474
 475         /* If all else goes well, this is our RPC return code. */
 476         req->rq_status = 0;
 477
 478         if (rc && rc != EALREADY)
 479                 GOTO(out, rc);
 480
 481         /* XXX track this all the time? */
 482         if (target->obd_recovering) {
 483                 target->obd_connected_clients++;
 484         }
 485
 486         req->rq_repmsg->handle = conn;
 487
 488         /* If the client and the server are the same node, we will already
 489          * have an export that really points to the client's DLM export,
 490          * because we have a shared handles table.
 491          *
 492          * XXX this will go away when shaver stops sending the "connect" handle
 493          * in the real "remote handle" field of the request --phik 24 Apr 2003
 494          */
 495         if (req->rq_export != NULL)
 496                 class_export_put(req->rq_export);
 497
 498         /* ownership of this export ref transfers to the request */
 499         export = req->rq_export = class_conn2export(&conn);
 500         LASSERT(export != NULL);
 501
 502         if (export->exp_connection != NULL)
 503                 ptlrpc_put_connection(export->exp_connection);
 504         export->exp_connection = ptlrpc_get_connection(&req->rq_peer,
 505                                                        &remote_uuid);
 506
 507         LASSERT(export->exp_conn_cnt < req->rq_reqmsg->conn_cnt);
 508         export->exp_conn_cnt = req->rq_reqmsg->conn_cnt;
 509
 510         if (rc == EALREADY) {
 511                 /* We indicate the reconnection in a flag, not an error code. */
 512                 lustre_msg_add_op_flags(req->rq_repmsg, MSG_CONNECT_RECONNECT);
 513                 GOTO(out, rc = 0);
 514         }
 515
 516         memcpy(&conn, lustre_msg_buf(req->rq_reqmsg, 2, sizeof conn),
 517                sizeof conn);
 518
 519         if (export->exp_imp_reverse != NULL)
 520                 class_destroy_import(export->exp_imp_reverse);
 521         revimp = export->exp_imp_reverse = class_new_import();
 522         revimp->imp_connection = ptlrpc_connection_addref(export->exp_connection);
 523         revimp->imp_client = &export->exp_obd->obd_ldlm_client;
 524         revimp->imp_remote_handle = conn;
 525         revimp->imp_obd = target;
 526         revimp->imp_dlm_fake = 1;
 527         revimp->imp_state = LUSTRE_IMP_FULL;
 528         class_import_put(revimp);
 529 out:
 530         if (rc)
 531                 req->rq_status = rc;
 532         RETURN(rc);
 533 }
 534
 535 int target_handle_disconnect(struct ptlrpc_request *req)
 536 {
 537         struct obd_export *exp;
 538         int rc;
 539         ENTRY;
 540
 541         rc = lustre_pack_reply(req, 0, NULL, NULL);
 542         if (rc)
 543                 RETURN(rc);
 544
 545         /* keep the rq_export around so we can send the reply */
 546         exp = class_export_get(req->rq_export);
 547         req->rq_status = obd_disconnect(exp, 0);
 548         RETURN(0);
 549 }
 550
 551 void target_destroy_export(struct obd_export *exp)
 552 {
 553         /* exports created from last_rcvd data, and "fake"
 554            exports created by lctl don't have an import */
 555         if (exp->exp_imp_reverse != NULL)
 556                 class_destroy_import(exp->exp_imp_reverse);
 557
 558         /* We cancel locks at disconnect time, but this will catch any locks
 559          * granted in a race with recovery-induced disconnect. */
 560         if (exp->exp_obd->obd_namespace != NULL)
 561                 ldlm_cancel_locks_for_export(exp);
 562 }
 563
 564 /*
 565  * Recovery functions
 566  */
 567
 568 static void abort_delayed_replies(struct obd_device *obd)
 569 {
 570         struct ptlrpc_request *req;
 571         struct list_head *tmp, *n;
 572         list_for_each_safe(tmp, n, &obd->obd_delayed_reply_queue) {
 573                 req = list_entry(tmp, struct ptlrpc_request, rq_list);
 574                 DEBUG_REQ(D_ERROR, req, "aborted:");
 575                 req->rq_status = -ENOTCONN;
 576                 req->rq_type = PTL_RPC_MSG_ERR;
 577                 ptlrpc_reply(req);
 578                 class_export_put(req->rq_export);
 579                 list_del(&req->rq_list);
 580                 OBD_FREE(req->rq_reqmsg, req->rq_reqlen);
 581                 OBD_FREE(req, sizeof *req);
 582         }
 583 }
 584
 585 static void abort_recovery_queue(struct obd_device *obd)
 586 {
 587         struct ptlrpc_request *req;
 588         struct list_head *tmp, *n;
 589         int rc;
 590
 591         list_for_each_safe(tmp, n, &obd->obd_recovery_queue) {
 592                 req = list_entry(tmp, struct ptlrpc_request, rq_list);
 593                 DEBUG_REQ(D_ERROR, req, "aborted:");
 594                 req->rq_status = -ENOTCONN;
 595                 req->rq_type = PTL_RPC_MSG_ERR;
 596                 rc = lustre_pack_reply(req, 0, NULL, NULL);
 597                 if (rc == 0) {
 598                         ptlrpc_reply(req);
 599                 } else {
 600                         DEBUG_REQ(D_ERROR, req,
 601                                   "packing failed for abort-reply; skipping");
 602                 }
 603                 list_del(&req->rq_list);
 604                 class_export_put(req->rq_export);
 605                 OBD_FREE(req->rq_reqmsg, req->rq_reqlen);
 606                 OBD_FREE(req, sizeof *req);
 607         }
 608 }
 609
 610 void target_abort_recovery(void *data)
 611 {
 612         struct obd_device *obd = data;
 613         int rc;
 614
 615         CERROR("disconnecting clients and aborting recovery\n");
 616         spin_lock_bh(&obd->obd_processing_task_lock);
 617         if (!obd->obd_recovering) {
 618                 spin_unlock_bh(&obd->obd_processing_task_lock);
 619                 EXIT;
 620                 return;
 621         }
 622
 623         obd->obd_recovering = obd->obd_abort_recovery = 0;
 624
 625         wake_up(&obd->obd_next_transno_waitq);
 626         target_cancel_recovery_timer(obd);
 627         spin_unlock_bh(&obd->obd_processing_task_lock);
 628
 629         class_disconnect_exports(obd, 0);
 630
 631         /* when recovery was aborted, cleanup orphans on mds and ost */
 632         if (OBT(obd) && OBP(obd, postrecov)) {
 633                 rc = OBP(obd, postrecov)(obd);
 634                 if (rc >= 0)
 635                         CWARN("Cleanup %d orphans after recovery was aborted\n", rc);
 636                 else
 637                         CERROR("postrecov failed %d\n", rc);
 638         }
 639
 640         abort_delayed_replies(obd);
 641         abort_recovery_queue(obd);
 642         ptlrpc_run_recovery_over_upcall(obd);
 643 }
 644
 645 static void target_recovery_expired(unsigned long castmeharder)
 646 {
 647         struct obd_device *obd = (struct obd_device *)castmeharder;
 648         CERROR("recovery timed out, aborting\n");
 649         spin_lock_bh(&obd->obd_processing_task_lock);
 650         obd->obd_abort_recovery = 1;
 651         wake_up(&obd->obd_next_transno_waitq);
 652         spin_unlock_bh(&obd->obd_processing_task_lock);
 653 }
 654
 655
 656 /* obd_processing_task_lock should be held */
 657 void target_cancel_recovery_timer(struct obd_device *obd)
 658 {
 659         CDEBUG(D_HA, "%s: cancel recovery timer\n", obd->obd_name);
 660         del_timer(&obd->obd_recovery_timer);
 661 }
 662
 663 static void reset_recovery_timer(struct obd_device *obd)
 664 {
 665         spin_lock_bh(&obd->obd_processing_task_lock);
 666         if (!obd->obd_recovering) {
 667                 spin_unlock_bh(&obd->obd_processing_task_lock);
 668                 return;
 669         }
 670         CDEBUG(D_HA, "timer will expire in %u seconds\n",
 671                OBD_RECOVERY_TIMEOUT / HZ);
 672         mod_timer(&obd->obd_recovery_timer, jiffies + OBD_RECOVERY_TIMEOUT);
 673         spin_unlock_bh(&obd->obd_processing_task_lock);
 674 }
 675
 676
 677 /* Only start it the first time called */
 678 void target_start_recovery_timer(struct obd_device *obd, svc_handler_t handler)
 679 {
 680         spin_lock_bh(&obd->obd_processing_task_lock);
 681         if (obd->obd_recovery_handler) {
 682                 spin_unlock_bh(&obd->obd_processing_task_lock);
 683                 return;
 684         }
 685         CWARN("%s: starting recovery timer (%us)\n", obd->obd_name,
 686                OBD_RECOVERY_TIMEOUT / HZ);
 687         obd->obd_recovery_handler = handler;
 688         obd->obd_recovery_timer.function = target_recovery_expired;
 689         obd->obd_recovery_timer.data = (unsigned long)obd;
 690         init_timer(&obd->obd_recovery_timer);
 691         spin_unlock_bh(&obd->obd_processing_task_lock);
 692
 693         reset_recovery_timer(obd);
 694 }
 695
 696 static int check_for_next_transno(struct obd_device *obd)
 697 {
 698         struct ptlrpc_request *req;
 699         int wake_up = 0, connected, completed, queue_len, max;
 700         __u64 next_transno, req_transno;
 701
 702         spin_lock_bh(&obd->obd_processing_task_lock);
 703         req = list_entry(obd->obd_recovery_queue.next,
 704                          struct ptlrpc_request, rq_list);
 705         max = obd->obd_max_recoverable_clients;
 706         req_transno = req->rq_reqmsg->transno;
 707         connected = obd->obd_connected_clients;
 708         completed = max - obd->obd_recoverable_clients;
 709         queue_len = obd->obd_requests_queued_for_recovery;
 710         next_transno = obd->obd_next_recovery_transno;
 711
 712         if (obd->obd_abort_recovery) {
 713                 CDEBUG(D_HA, "waking for aborted recovery\n");
 714                 wake_up = 1;
 715         } else if (!obd->obd_recovering) {
 716                 CDEBUG(D_HA, "waking for completed recovery (?)\n");
 717                 wake_up = 1;
 718         } else if (req_transno == next_transno) {
 719                 CDEBUG(D_HA, "waking for next ("LPD64")\n", next_transno);
 720                 wake_up = 1;
 721         } else if (queue_len + completed == max) {
 722                 CDEBUG(D_ERROR,
 723                        "waking for skipped transno (skip: "LPD64
 724                        ", ql: %d, comp: %d, conn: %d, next: "LPD64")\n",
 725                        next_transno, queue_len, completed, max, req_transno);
 726                 obd->obd_next_recovery_transno = req_transno;
 727                 wake_up = 1;
 728         }
 729         spin_unlock_bh(&obd->obd_processing_task_lock);
 730         LASSERT(req->rq_reqmsg->transno >= next_transno);
 731         return wake_up;
 732 }
 733
 734 static void process_recovery_queue(struct obd_device *obd)
 735 {
 736         struct ptlrpc_request *req;
 737         int abort_recovery = 0;
 738         struct l_wait_info lwi = { 0 };
 739         ENTRY;
 740
 741         for (;;) {
 742                 spin_lock_bh(&obd->obd_processing_task_lock);
 743                 LASSERT(obd->obd_processing_task == current->pid);
 744                 req = list_entry(obd->obd_recovery_queue.next,
 745                                  struct ptlrpc_request, rq_list);
 746
 747                 if (req->rq_reqmsg->transno != obd->obd_next_recovery_transno) {
 748                         spin_unlock_bh(&obd->obd_processing_task_lock);
 749                         CDEBUG(D_HA, "Waiting for transno "LPD64" (1st is "
 750                                LPD64")\n",
 751                                obd->obd_next_recovery_transno,
 752                                req->rq_reqmsg->transno);
 753                         l_wait_event(obd->obd_next_transno_waitq,
 754                                      check_for_next_transno(obd), &lwi);
 755                         spin_lock_bh(&obd->obd_processing_task_lock);
 756                         abort_recovery = obd->obd_abort_recovery;
 757                         spin_unlock_bh(&obd->obd_processing_task_lock);
 758                         if (abort_recovery) {
 759                                 target_abort_recovery(obd);
 760                                 return;
 761                         }
 762                         continue;
 763                 }
 764                 list_del_init(&req->rq_list);
 765                 obd->obd_requests_queued_for_recovery--;
 766                 spin_unlock_bh(&obd->obd_processing_task_lock);
 767
 768                 DEBUG_REQ(D_HA, req, "processing: ");
 769                 (void)obd->obd_recovery_handler(req);
 770                 obd->obd_replayed_requests++;
 771                 reset_recovery_timer(obd);
 772                 /* bug 1580: decide how to properly sync() in recovery */
 773                 //mds_fsync_super(mds->mds_sb);
 774                 class_export_put(req->rq_export);
 775                 OBD_FREE(req->rq_reqmsg, req->rq_reqlen);
 776                 OBD_FREE(req, sizeof *req);
 777                 spin_lock_bh(&obd->obd_processing_task_lock);
 778                 obd->obd_next_recovery_transno++;
 779                 if (list_empty(&obd->obd_recovery_queue)) {
 780                         obd->obd_processing_task = 0;
 781                         spin_unlock_bh(&obd->obd_processing_task_lock);
 782                         break;
 783                 }
 784                 spin_unlock_bh(&obd->obd_processing_task_lock);
 785         }
 786         EXIT;
 787 }
 788
 789 int target_queue_recovery_request(struct ptlrpc_request *req,
 790                                   struct obd_device *obd)
 791 {
 792         struct list_head *tmp;
 793         int inserted = 0;
 794         __u64 transno = req->rq_reqmsg->transno;
 795         struct ptlrpc_request *saved_req;
 796         struct lustre_msg *reqmsg;
 797
 798         /* CAVEAT EMPTOR: The incoming request message has been swabbed
 799          * (i.e. buflens etc are in my own byte order), but type-dependent
 800          * buffers (eg mds_body, ost_body etc) have NOT been swabbed. */
 801
 802         if (!transno) {
 803                 INIT_LIST_HEAD(&req->rq_list);
 804                 DEBUG_REQ(D_HA, req, "not queueing");
 805                 return 1;
 806         }
 807
 808         /* XXX If I were a real man, these LBUGs would be sane cleanups. */
 809         /* XXX just like the request-dup code in queue_final_reply */
 810         OBD_ALLOC(saved_req, sizeof *saved_req);
 811         if (!saved_req)
 812                 LBUG();
 813         OBD_ALLOC(reqmsg, req->rq_reqlen);
 814         if (!reqmsg)
 815                 LBUG();
 816
 817         spin_lock_bh(&obd->obd_processing_task_lock);
 818
 819         /* If we're processing the queue, we want don't want to queue this
 820          * message.
 821          *
 822          * Also, if this request has a transno less than the one we're waiting
 823          * for, we should process it now.  It could (and currently always will)
 824          * be an open request for a descriptor that was opened some time ago.
 825          */
 826         if (obd->obd_processing_task == current->pid ||
 827             transno < obd->obd_next_recovery_transno) {
 828                 /* Processing the queue right now, don't re-add. */
 829                 LASSERT(list_empty(&req->rq_list));
 830                 spin_unlock_bh(&obd->obd_processing_task_lock);
 831                 OBD_FREE(reqmsg, req->rq_reqlen);
 832                 OBD_FREE(saved_req, sizeof *saved_req);
 833                 return 1;
 834         }
 835
 836         memcpy(saved_req, req, sizeof *req);
 837         memcpy(reqmsg, req->rq_reqmsg, req->rq_reqlen);
 838         req = saved_req;
 839         req->rq_reqmsg = reqmsg;
 840         class_export_get(req->rq_export);
 841         INIT_LIST_HEAD(&req->rq_list);
 842
 843         /* XXX O(n^2) */
 844         list_for_each(tmp, &obd->obd_recovery_queue) {
 845                 struct ptlrpc_request *reqiter =
 846                         list_entry(tmp, struct ptlrpc_request, rq_list);
 847
 848                 if (reqiter->rq_reqmsg->transno > transno) {
 849                         list_add_tail(&req->rq_list, &reqiter->rq_list);
 850                         inserted = 1;
 851                         break;
 852                 }
 853         }
 854
 855         if (!inserted) {
 856                 list_add_tail(&req->rq_list, &obd->obd_recovery_queue);
 857         }
 858
 859         obd->obd_requests_queued_for_recovery++;
 860
 861         if (obd->obd_processing_task != 0) {
 862                 /* Someone else is processing this queue, we'll leave it to
 863                  * them.
 864                  */
 865                 wake_up(&obd->obd_next_transno_waitq);
 866                 spin_unlock_bh(&obd->obd_processing_task_lock);
 867                 return 0;
 868         }
 869
 870         /* Nobody is processing, and we know there's (at least) one to process
 871          * now, so we'll do the honours.
 872          */
 873         obd->obd_processing_task = current->pid;
 874         spin_unlock_bh(&obd->obd_processing_task_lock);
 875
 876         process_recovery_queue(obd);
 877         return 0;
 878 }
 879
 880 struct obd_device * target_req2obd(struct ptlrpc_request *req)
 881 {
 882         return req->rq_export->exp_obd;
 883 }
 884
 885 int target_queue_final_reply(struct ptlrpc_request *req, int rc)
 886 {
 887         struct obd_device *obd = target_req2obd(req);
 888         struct ptlrpc_request *saved_req;
 889         struct lustre_msg *reqmsg;
 890         int recovery_done = 0;
 891         int rc2;
 892
 893         if (rc) {
 894                 /* Just like ptlrpc_error, but without the sending. */
 895                 rc = lustre_pack_reply(req, 0, NULL, NULL);
 896                 LASSERT(rc == 0); /* XXX handle this */
 897                 req->rq_type = PTL_RPC_MSG_ERR;
 898         }
 899
 900         LASSERT(list_empty(&req->rq_list));
 901         /* XXX a bit like the request-dup code in queue_recovery_request */
 902         OBD_ALLOC(saved_req, sizeof *saved_req);
 903         if (!saved_req)
 904                 LBUG();
 905         OBD_ALLOC(reqmsg, req->rq_reqlen);
 906         if (!reqmsg)
 907                 LBUG();
 908         memcpy(saved_req, req, sizeof *saved_req);
 909         memcpy(reqmsg, req->rq_reqmsg, req->rq_reqlen);
 910         req = saved_req;
 911         req->rq_reqmsg = reqmsg;
 912         class_export_get(req->rq_export);
 913         list_add(&req->rq_list, &obd->obd_delayed_reply_queue);
 914
 915         spin_lock_bh(&obd->obd_processing_task_lock);
 916         --obd->obd_recoverable_clients;
 917         recovery_done = (obd->obd_recoverable_clients == 0);
 918         spin_unlock_bh(&obd->obd_processing_task_lock);
 919
 920         if (recovery_done) {
 921                 struct list_head *tmp, *n;
 922                 ldlm_reprocess_all_ns(req->rq_export->exp_obd->obd_namespace);
 923                 CWARN("%s: all clients recovered, sending delayed replies\n",
 924                        obd->obd_name);
 925                 spin_lock_bh(&obd->obd_processing_task_lock);
 926                 obd->obd_recovering = 0;
 927                 target_cancel_recovery_timer(obd);
 928                 spin_unlock_bh(&obd->obd_processing_task_lock);
 929
 930                 /* when recovery finished, cleanup orphans on mds and ost */
 931                 if (OBT(obd) && OBP(obd, postrecov)) {
 932                         rc2 = OBP(obd, postrecov)(obd);
 933                         if (rc2 >= 0)
 934                                 CWARN("%s: all clients recovered, %d MDS "
 935                                       "orphans deleted\n", obd->obd_name, rc2);
 936                         else
 937                                 CERROR("postrecov failed %d\n", rc2);
 938                 }
 939
 940                 list_for_each_safe(tmp, n, &obd->obd_delayed_reply_queue) {
 941                         req = list_entry(tmp, struct ptlrpc_request, rq_list);
 942                         DEBUG_REQ(D_ERROR, req, "delayed:");
 943                         ptlrpc_reply(req);
 944                         class_export_put(req->rq_export);
 945                         list_del(&req->rq_list);
 946                         OBD_FREE(req->rq_reqmsg, req->rq_reqlen);
 947                         OBD_FREE(req, sizeof *req);
 948                 }
 949                 ptlrpc_run_recovery_over_upcall(obd);
 950         } else {
 951                 CWARN("%s: %d recoverable clients remain\n",
 952                        obd->obd_name, obd->obd_recoverable_clients);
 953                 wake_up(&obd->obd_next_transno_waitq);
 954         }
 955
 956         return 1;
 957 }
 958
 959 static void ptlrpc_abort_reply (struct ptlrpc_request *req)
 960 {
 961         /* On return, we must be sure that the ACK callback has either
 962          * happened or will not happen.  Note that the SENT callback will
 963          * happen come what may since we successfully posted the PUT. */
 964         int rc;
 965         struct l_wait_info lwi;
 966         unsigned long flags;
 967
 968  again:
 969         /* serialise with ACK callback */
 970         spin_lock_irqsave (&req->rq_lock, flags);
 971         if (!req->rq_want_ack) {
 972                 spin_unlock_irqrestore (&req->rq_lock, flags);
 973                 /* The ACK callback has happened already.  Although the
 974                  * SENT callback might still be outstanding (yes really) we
 975                  * don't care; this is just like normal completion. */
 976                 return;
 977         }
 978         spin_unlock_irqrestore (&req->rq_lock, flags);
 979
 980         /* Have a bash at unlinking the MD.  This will fail until the SENT
 981          * callback has happened since the MD is busy from the PUT.  If the
 982          * ACK still hasn't arrived after then, a successful unlink will
 983          * ensure the ACK callback never happens. */
 984         rc = PtlMDUnlink (req->rq_reply_md_h);
 985         switch (rc) {
 986         default:
 987                 LBUG ();
 988         case PTL_OK:
 989                 /* SENT callback happened; ACK callback preempted */
 990                 LASSERT (req->rq_want_ack);
 991                 spin_lock_irqsave (&req->rq_lock, flags);
 992                 req->rq_want_ack = 0;
 993                 spin_unlock_irqrestore (&req->rq_lock, flags);
 994                 return;
 995         case PTL_INV_MD:
 996                 return;
 997         case PTL_MD_INUSE:
 998                 /* Still sending or ACK callback in progress: wait until
 999                  * either callback has completed and try again.
1000                  * Actually we can't wait for the SENT callback because
1001                  * there's no state the SENT callback can touch that will
1002                  * allow it to communicate with us!  So we just wait here
1003                  * for a short time, effectively polling for the SENT
1004                  * callback by calling PtlMDUnlink() again, to see if it
1005                  * has finished.  Note that if the ACK does arrive, its
1006                  * callback wakes us in short order. --eeb */
1007                 lwi = LWI_TIMEOUT (HZ/4, NULL, NULL);
1008                 rc = l_wait_event(req->rq_reply_waitq, !req->rq_want_ack,
1009                                   &lwi);
1010                 CDEBUG (D_HA, "Retrying req %p: %d\n", req, rc);
1011                 /* NB go back and test rq_want_ack with locking, to ensure
1012                  * if ACK callback happened, it has completed stopped
1013                  * referencing this req. */
1014                 goto again;
1015         }
1016 }
1017
1018 void target_send_reply(struct ptlrpc_request *req, int rc, int fail_id)
1019 {
1020         int i;
1021         int netrc;
1022         unsigned long flags;
1023         struct ptlrpc_req_ack_lock *ack_lock;
1024         struct l_wait_info lwi = { 0 };
1025         wait_queue_t commit_wait;
1026         struct obd_device *obd =
1027                 req->rq_export ? req->rq_export->exp_obd : NULL;
1028         struct obd_export *exp = NULL;
1029
1030         if (req->rq_export) {
1031                 for (i = 0; i < REQ_MAX_ACK_LOCKS; i++) {
1032                         if (req->rq_ack_locks[i].mode) {
1033                                 exp = req->rq_export;
1034                                 break;
1035                         }
1036                 }
1037         }
1038
1039         if (exp) {
1040                 exp->exp_outstanding_reply = req;
1041                 spin_lock_irqsave (&req->rq_lock, flags);
1042                 req->rq_want_ack = 1;
1043                 spin_unlock_irqrestore (&req->rq_lock, flags);
1044         }
1045
1046         if (!OBD_FAIL_CHECK(fail_id | OBD_FAIL_ONCE)) {
1047                 if (rc == 0) {
1048                         DEBUG_REQ(D_NET, req, "sending reply");
1049                         netrc = ptlrpc_reply(req);
1050                 } else if (rc == -ENOTCONN) {
1051                         DEBUG_REQ(D_HA, req, "processing error (%d)", rc);
1052                         netrc = ptlrpc_error(req);
1053                 } else {
1054                         DEBUG_REQ(D_ERROR, req, "processing error (%d)", rc);
1055                         netrc = ptlrpc_error(req);
1056                 }
1057         } else {
1058                 obd_fail_loc |= OBD_FAIL_ONCE | OBD_FAILED;
1059                 DEBUG_REQ(D_ERROR, req, "dropping reply");
1060                 if (req->rq_repmsg) {
1061                         OBD_FREE(req->rq_repmsg, req->rq_replen);
1062                         req->rq_repmsg = NULL;
1063                 }
1064                 init_waitqueue_head(&req->rq_reply_waitq);
1065                 netrc = 0;
1066         }
1067
1068         /* a failed send simulates the callbacks */
1069         LASSERT(netrc == 0 || req->rq_want_ack == 0);
1070         if (exp == NULL) {
1071                 LASSERT(req->rq_want_ack == 0);
1072                 return;
1073         }
1074         LASSERT(obd != NULL);
1075
1076         init_waitqueue_entry(&commit_wait, current);
1077         add_wait_queue(&obd->obd_commit_waitq, &commit_wait);
1078         rc = l_wait_event(req->rq_reply_waitq,
1079                           !req->rq_want_ack || req->rq_resent ||
1080                           req->rq_transno <= obd->obd_last_committed, &lwi);
1081         remove_wait_queue(&obd->obd_commit_waitq, &commit_wait);
1082
1083         spin_lock_irqsave (&req->rq_lock, flags);
1084         /* If we got here because the ACK callback ran, this acts as a
1085          * barrier to ensure the callback completed the wakeup. */
1086         spin_unlock_irqrestore (&req->rq_lock, flags);
1087
1088         /* If we committed the transno already, then we might wake up before
1089          * the ack arrives.  We need to stop waiting for the ack before we can
1090          * reuse this request structure.  We are guaranteed by this point that
1091          * this cannot abort the sending of the actual reply.*/
1092         ptlrpc_abort_reply(req);
1093
1094         if (req->rq_resent) {
1095                 DEBUG_REQ(D_HA, req, "resent: not cancelling locks");
1096                 return;
1097         }
1098
1099         LASSERT(rc == 0);
1100         DEBUG_REQ(D_HA, req, "cancelling locks for %s",
1101                   req->rq_want_ack ? "commit" : "ack");
1102
1103         exp->exp_outstanding_reply = NULL;
1104
1105         for (ack_lock = req->rq_ack_locks, i = 0;
1106              i < REQ_MAX_ACK_LOCKS; i++, ack_lock++) {
1107                 if (!ack_lock->mode)
1108                         continue;
1109                 ldlm_lock_decref(&ack_lock->lock, ack_lock->mode);
1110         }
1111 }
1112
1113 int target_handle_ping(struct ptlrpc_request *req)
1114 {
1115         return lustre_pack_reply(req, 0, NULL, NULL);
1116 }
1117
1118 void *ldlm_put_lock_into_req(struct ptlrpc_request *req,
1119                                 struct lustre_handle *lock, int mode)
1120 {
1121         int i;
1122
1123         for (i = 0; i < REQ_MAX_ACK_LOCKS; i++) {
1124                 if (req->rq_ack_locks[i].mode)
1125                         continue;
1126                 CDEBUG(D_HA, "saving lock "LPX64" in req %p ack_lock[%d]\n",
1127                        lock->cookie, req, i);
1128                 memcpy(&req->rq_ack_locks[i].lock, lock, sizeof(*lock));
1129                 req->rq_ack_locks[i].mode = mode;
1130                 return &req->rq_ack_locks[i];
1131         }
1132         CERROR("no space for lock in struct ptlrpc_request\n");
1133         LBUG();
1134         return NULL;
1135 }