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         init_MUTEX(&cli->cl_dirty_sem);
 100         cli->cl_dirty = 0;
 101         cli->cl_dirty_granted = 0;
 102         cli->cl_dirty_max = OSC_MAX_DIRTY_DEFAULT * 1024 * 1024;
 103         cli->cl_ost_can_grant = 1;
 104         INIT_LIST_HEAD(&cli->cl_cache_waiters);
 105         INIT_LIST_HEAD(&cli->cl_loi_ready_list);
 106         INIT_LIST_HEAD(&cli->cl_loi_write_list);
 107         spin_lock_init(&cli->cl_loi_list_lock);
 108         cli->cl_brw_in_flight = 0;
 109         spin_lock_init(&cli->cl_read_rpc_hist.oh_lock);
 110         spin_lock_init(&cli->cl_write_rpc_hist.oh_lock);
 111         spin_lock_init(&cli->cl_read_page_hist.oh_lock);
 112         spin_lock_init(&cli->cl_write_page_hist.oh_lock);
 113         cli->cl_max_pages_per_rpc = PTL_MD_MAX_PAGES;
 114         cli->cl_max_rpcs_in_flight = OSC_MAX_RIF_DEFAULT;
 115
 116         ldlm_get_ref();
 117         if (rc) {
 118                 CERROR("ldlm_get_ref failed: %d\n", rc);
 119                 GOTO(err, rc);
 120         }
 121
 122         conn = ptlrpc_uuid_to_connection(&server_uuid);
 123         if (conn == NULL)
 124                 GOTO(err_ldlm, rc = -ENOENT);
 125
 126         ptlrpc_init_client(rq_portal, rp_portal, name,
 127                            &obddev->obd_ldlm_client);
 128
 129         imp = class_new_import();
 130         if (imp == NULL) {
 131                 ptlrpc_put_connection(conn);
 132                 GOTO(err_ldlm, rc = -ENOENT);
 133         }
 134         imp->imp_connection = conn;
 135         imp->imp_client = &obddev->obd_ldlm_client;
 136         imp->imp_obd = obddev;
 137         imp->imp_connect_op = connect_op;
 138         imp->imp_generation = 0;
 139         INIT_LIST_HEAD(&imp->imp_pinger_chain);
 140         memcpy(imp->imp_target_uuid.uuid, lcfg->lcfg_inlbuf1,
 141               lcfg->lcfg_inllen1);
 142         class_import_put(imp);
 143
 144         cli->cl_import = imp;
 145         cli->cl_max_mds_easize = sizeof(struct lov_mds_md);
 146         cli->cl_max_mds_cookiesize = sizeof(struct llog_cookie);
 147         cli->cl_sandev = to_kdev_t(0);
 148
 149         if (lcfg->lcfg_inllen3 != 0) {
 150                 if (!strcmp(lcfg->lcfg_inlbuf3, "inactive")) {
 151                         CDEBUG(D_HA, "marking %s %s->%s as inactive\n",
 152                                name, obddev->obd_name,
 153                                imp->imp_target_uuid.uuid);
 154                         imp->imp_invalid = 1;
 155
 156                         if (lcfg->lcfg_inllen4 != 0)
 157                                 mgmt_name = lcfg->lcfg_inlbuf4;
 158                 } else {
 159                         mgmt_name = lcfg->lcfg_inlbuf3;
 160                 }
 161         }
 162
 163         if (mgmt_name != NULL) {
 164                 /* Register with management client if we need to. */
 165                 CDEBUG(D_HA, "%s registering with %s for events about %s\n",
 166                        obddev->obd_name, mgmt_name, server_uuid.uuid);
 167
 168                 mgmt_obd = class_name2obd(mgmt_name);
 169                 if (!mgmt_obd) {
 170                         CERROR("can't find mgmtcli %s to register\n",
 171                                mgmt_name);
 172                         GOTO(err_import, rc = -ENOSYS);
 173                 }
 174
 175                 register_f = inter_module_get("mgmtcli_register_for_events");
 176                 if (!register_f) {
 177                         CERROR("can't i_m_g mgmtcli_register_for_events\n");
 178                         GOTO(err_import, rc = -ENOSYS);
 179                 }
 180
 181                 rc = register_f(mgmt_obd, obddev, &imp->imp_target_uuid);
 182                 inter_module_put("mgmtcli_register_for_events");
 183
 184                 if (!rc)
 185                         cli->cl_mgmtcli_obd = mgmt_obd;
 186         }
 187
 188         RETURN(rc);
 189
 190 err_import:
 191         class_destroy_import(imp);
 192 err_ldlm:
 193         ldlm_put_ref(0);
 194 err:
 195         RETURN(rc);
 196
 197 }
 198
 199 int client_obd_cleanup(struct obd_device *obddev, int flags)
 200 {
 201         struct client_obd *cli = &obddev->u.cli;
 202
 203         if (!cli->cl_import)
 204                 RETURN(-EINVAL);
 205         if (cli->cl_mgmtcli_obd) {
 206                 mgmtcli_deregister_for_events_t dereg_f;
 207
 208                 dereg_f = inter_module_get("mgmtcli_deregister_for_events");
 209                 dereg_f(cli->cl_mgmtcli_obd, obddev);
 210                 inter_module_put("mgmtcli_deregister_for_events");
 211         }
 212         class_destroy_import(cli->cl_import);
 213         cli->cl_import = NULL;
 214
 215         ldlm_put_ref(flags & OBD_OPT_FORCE);
 216
 217         RETURN(0);
 218 }
 219
 220 int client_connect_import(struct lustre_handle *dlm_handle,
 221                           struct obd_device *obd,
 222                           struct obd_uuid *cluuid)
 223 {
 224         struct client_obd *cli = &obd->u.cli;
 225         struct obd_import *imp = cli->cl_import;
 226         struct obd_export *exp;
 227         int rc;
 228         ENTRY;
 229
 230         down(&cli->cl_sem);
 231         rc = class_connect(dlm_handle, obd, cluuid);
 232         if (rc)
 233                 GOTO(out_sem, rc);
 234
 235         cli->cl_conn_count++;
 236         if (cli->cl_conn_count > 1)
 237                 GOTO(out_sem, rc);
 238         exp = class_conn2export(dlm_handle);
 239
 240         if (obd->obd_namespace != NULL)
 241                 CERROR("already have namespace!\n");
 242         obd->obd_namespace = ldlm_namespace_new(obd->obd_name,
 243                                                 LDLM_NAMESPACE_CLIENT);
 244         if (obd->obd_namespace == NULL)
 245                 GOTO(out_disco, rc = -ENOMEM);
 246
 247         imp->imp_dlm_handle = *dlm_handle;
 248         imp->imp_state = LUSTRE_IMP_DISCON;
 249
 250         rc = ptlrpc_connect_import(imp);
 251         if (rc != 0) {
 252                 LASSERT (imp->imp_state == LUSTRE_IMP_DISCON);
 253                 GOTO(out_ldlm, rc);
 254         }
 255
 256         LASSERT (imp->imp_state == LUSTRE_IMP_FULL);
 257
 258         exp->exp_connection = ptlrpc_connection_addref(imp->imp_connection);
 259
 260         if (imp->imp_replayable) {
 261                 CDEBUG(D_HA, "connected to replayable target: %s\n",
 262                        imp->imp_target_uuid.uuid);
 263                 ptlrpc_pinger_add_import(imp);
 264         }
 265
 266         CDEBUG(D_HA, "local import: %p, remote handle: "LPX64"\n", imp,
 267                imp->imp_remote_handle.cookie);
 268
 269         EXIT;
 270
 271         if (rc) {
 272 out_ldlm:
 273                 ldlm_namespace_free(obd->obd_namespace, 0);
 274                 obd->obd_namespace = NULL;
 275 out_disco:
 276                 cli->cl_conn_count--;
 277                 class_disconnect(exp, 0);
 278         } else {
 279                 class_export_put(exp);
 280         }
 281 out_sem:
 282         up(&cli->cl_sem);
 283         return rc;
 284 }
 285
 286 int client_disconnect_export(struct obd_export *exp, int failover)
 287 {
 288         struct obd_device *obd = class_exp2obd(exp);
 289         struct client_obd *cli = &obd->u.cli;
 290         struct obd_import *imp = cli->cl_import;
 291         int rc = 0, err;
 292         ENTRY;
 293
 294         if (!obd) {
 295                 CERROR("invalid export for disconnect: exp %p cookie "LPX64"\n",
 296                        exp, exp ? exp->exp_handle.h_cookie : -1);
 297                 RETURN(-EINVAL);
 298         }
 299
 300         down(&cli->cl_sem);
 301         if (!cli->cl_conn_count) {
 302                 CERROR("disconnecting disconnected device (%s)\n",
 303                        obd->obd_name);
 304                 GOTO(out_sem, rc = -EINVAL);
 305         }
 306
 307         cli->cl_conn_count--;
 308         if (cli->cl_conn_count)
 309                 GOTO(out_no_disconnect, rc = 0);
 310
 311         /* Some non-replayable imports (MDS's OSCs) are pinged, so just
 312          * delete it regardless.  (It's safe to delete an import that was
 313          * never added.) */
 314         (void)ptlrpc_pinger_del_import(imp);
 315
 316         if (obd->obd_namespace != NULL) {
 317                 /* obd_no_recov == local only */
 318                 ldlm_cli_cancel_unused(obd->obd_namespace, NULL,
 319                                        obd->obd_no_recov, NULL);
 320                 ldlm_namespace_free(obd->obd_namespace, obd->obd_no_recov);
 321                 obd->obd_namespace = NULL;
 322         }
 323
 324         /* Yeah, obd_no_recov also (mainly) means "forced shutdown". */
 325         if (obd->obd_no_recov)
 326                 ptlrpc_set_import_active(imp, 0);
 327         else
 328                 rc = ptlrpc_disconnect_import(imp);
 329
 330         imp->imp_state = LUSTRE_IMP_NEW;
 331
 332         EXIT;
 333  out_no_disconnect:
 334         err = class_disconnect(exp, 0);
 335         if (!rc && err)
 336                 rc = err;
 337  out_sem:
 338         up(&cli->cl_sem);
 339         RETURN(rc);
 340 }
 341
 342 /* --------------------------------------------------------------------------
 343  * from old lib/target.c
 344  * -------------------------------------------------------------------------- */
 345
 346 int target_handle_reconnect(struct lustre_handle *conn, struct obd_export *exp,
 347                             struct obd_uuid *cluuid)
 348 {
 349         if (exp->exp_connection) {
 350                 struct lustre_handle *hdl;
 351                 hdl = &exp->exp_imp_reverse->imp_remote_handle;
 352                 /* Might be a re-connect after a partition. */
 353                 if (!memcmp(&conn->cookie, &hdl->cookie, sizeof conn->cookie)) {
 354                         CERROR("%s reconnecting\n", cluuid->uuid);
 355                         conn->cookie = exp->exp_handle.h_cookie;
 356                         RETURN(EALREADY);
 357                 } else {
 358                         CERROR("%s reconnecting from %s, "
 359                                "handle mismatch (ours "LPX64", theirs "
 360                                LPX64")\n", cluuid->uuid,
 361                                exp->exp_connection->c_remote_uuid.uuid,
 362                                hdl->cookie, conn->cookie);
 363                         memset(conn, 0, sizeof *conn);
 364                         RETURN(-EALREADY);
 365                 }
 366         }
 367
 368         conn->cookie = exp->exp_handle.h_cookie;
 369         CDEBUG(D_INFO, "existing export for UUID '%s' at %p\n",
 370                cluuid->uuid, exp);
 371         CDEBUG(D_IOCTL,"connect: cookie "LPX64"\n", conn->cookie);
 372         RETURN(0);
 373 }
 374
 375 int target_handle_connect(struct ptlrpc_request *req, svc_handler_t handler)
 376 {
 377         struct obd_device *target;
 378         struct obd_export *export = NULL;
 379         struct obd_import *revimp;
 380         struct lustre_handle conn;
 381         struct obd_uuid tgtuuid;
 382         struct obd_uuid cluuid;
 383         struct obd_uuid remote_uuid;
 384         struct list_head *p;
 385         char *str, *tmp;
 386         int rc = 0, abort_recovery;
 387         ENTRY;
 388
 389         LASSERT_REQSWAB (req, 0);
 390         str = lustre_msg_string(req->rq_reqmsg, 0, sizeof(tgtuuid) - 1);
 391         if (str == NULL) {
 392                 CERROR("bad target UUID for connect\n");
 393                 GOTO(out, rc = -EINVAL);
 394         }
 395
 396         obd_str2uuid (&tgtuuid, str);
 397         target = class_uuid2obd(&tgtuuid);
 398         if (!target) {
 399                 target = class_name2obd(str);
 400         }
 401
 402         if (!target || target->obd_stopping || !target->obd_set_up) {
 403                 CERROR("UUID '%s' is not available for connect\n", str);
 404                 GOTO(out, rc = -ENODEV);
 405         }
 406
 407         LASSERT_REQSWAB (req, 1);
 408         str = lustre_msg_string(req->rq_reqmsg, 1, sizeof(cluuid) - 1);
 409         if (str == NULL) {
 410                 CERROR("bad client UUID for connect\n");
 411                 GOTO(out, rc = -EINVAL);
 412         }
 413
 414         obd_str2uuid (&cluuid, str);
 415
 416         /* XXX extract a nettype and format accordingly */
 417         snprintf(remote_uuid.uuid, sizeof remote_uuid,
 418                  "NET_"LPX64"_UUID", req->rq_peer.peer_nid);
 419
 420         spin_lock_bh(&target->obd_processing_task_lock);
 421         abort_recovery = target->obd_abort_recovery;
 422         spin_unlock_bh(&target->obd_processing_task_lock);
 423         if (abort_recovery)
 424                 target_abort_recovery(target);
 425
 426         tmp = lustre_msg_buf(req->rq_reqmsg, 2, sizeof conn);
 427         if (tmp == NULL)
 428                 GOTO(out, rc = -EPROTO);
 429
 430         memcpy(&conn, tmp, sizeof conn);
 431
 432         rc = lustre_pack_reply(req, 0, NULL, NULL);
 433         if (rc)
 434                 GOTO(out, rc);
 435
 436         /* lctl gets a backstage, all-access pass. */
 437         if (obd_uuid_equals(&cluuid, &target->obd_uuid))
 438                 goto dont_check_exports;
 439
 440         spin_lock(&target->obd_dev_lock);
 441         list_for_each(p, &target->obd_exports) {
 442                 export = list_entry(p, struct obd_export, exp_obd_chain);
 443                 if (obd_uuid_equals(&cluuid, &export->exp_client_uuid)) {
 444                         spin_unlock(&target->obd_dev_lock);
 445                         LASSERT(export->exp_obd == target);
 446
 447                         rc = target_handle_reconnect(&conn, export, &cluuid);
 448                         break;
 449                 }
 450                 export = NULL;
 451         }
 452         /* If we found an export, we already unlocked. */
 453         if (!export) {
 454                 spin_unlock(&target->obd_dev_lock);
 455         } else if (req->rq_reqmsg->conn_cnt == 1) {
 456                 CERROR("%s reconnected with 1 conn_cnt; cookies not random?\n",
 457                        cluuid.uuid);
 458                 GOTO(out, rc = -EALREADY);
 459         }
 460
 461         /* Tell the client if we're in recovery. */
 462         /* If this is the first client, start the recovery timer */
 463         if (target->obd_recovering) {
 464                 lustre_msg_add_op_flags(req->rq_repmsg, MSG_CONNECT_RECOVERING);
 465                 target_start_recovery_timer(target, handler);
 466         }
 467
 468         /* Tell the client if we support replayable requests */
 469         if (target->obd_replayable)
 470                 lustre_msg_add_op_flags(req->rq_repmsg, MSG_CONNECT_REPLAYABLE);
 471
 472         if (export == NULL) {
 473                 if (target->obd_recovering) {
 474                         CERROR("denying connection for new client %s: "
 475                                "%d clients in recovery for %lds\n", cluuid.uuid,
 476                                target->obd_recoverable_clients,
 477                                (target->obd_recovery_timer.expires-jiffies)/HZ);
 478                         rc = -EBUSY;
 479                 } else {
 480  dont_check_exports:
 481                         rc = obd_connect(&conn, target, &cluuid);
 482                 }
 483         }
 484
 485         /* If all else goes well, this is our RPC return code. */
 486         req->rq_status = 0;
 487
 488         if (rc && rc != EALREADY)
 489                 GOTO(out, rc);
 490
 491         /* XXX track this all the time? */
 492         if (target->obd_recovering) {
 493                 target->obd_connected_clients++;
 494         }
 495
 496         req->rq_repmsg->handle = conn;
 497
 498         /* If the client and the server are the same node, we will already
 499          * have an export that really points to the client's DLM export,
 500          * because we have a shared handles table.
 501          *
 502          * XXX this will go away when shaver stops sending the "connect" handle
 503          * in the real "remote handle" field of the request --phik 24 Apr 2003
 504          */
 505         if (req->rq_export != NULL)
 506                 class_export_put(req->rq_export);
 507
 508         /* ownership of this export ref transfers to the request */
 509         export = req->rq_export = class_conn2export(&conn);
 510         LASSERT(export != NULL);
 511
 512         if (req->rq_connection != NULL)
 513                 ptlrpc_put_connection(req->rq_connection);
 514         if (export->exp_connection != NULL)
 515                 ptlrpc_put_connection(export->exp_connection);
 516         export->exp_connection = ptlrpc_get_connection(&req->rq_peer,
 517                                                        &remote_uuid);
 518         req->rq_connection = ptlrpc_connection_addref(export->exp_connection);
 519
 520         LASSERT(export->exp_conn_cnt < req->rq_reqmsg->conn_cnt);
 521         export->exp_conn_cnt = req->rq_reqmsg->conn_cnt;
 522
 523         if (rc == EALREADY) {
 524                 /* We indicate the reconnection in a flag, not an error code. */
 525                 lustre_msg_add_op_flags(req->rq_repmsg, MSG_CONNECT_RECONNECT);
 526                 GOTO(out, rc = 0);
 527         }
 528
 529         memcpy(&conn, lustre_msg_buf(req->rq_reqmsg, 2, sizeof conn),
 530                sizeof conn);
 531
 532         if (export->exp_imp_reverse != NULL)
 533                 class_destroy_import(export->exp_imp_reverse);
 534         revimp = export->exp_imp_reverse = class_new_import();
 535         revimp->imp_connection = ptlrpc_connection_addref(req->rq_connection);
 536         revimp->imp_client = &export->exp_obd->obd_ldlm_client;
 537         revimp->imp_remote_handle = conn;
 538         revimp->imp_obd = target;
 539         revimp->imp_dlm_fake = 1;
 540         revimp->imp_state = LUSTRE_IMP_FULL;
 541         class_import_put(revimp);
 542 out:
 543         if (rc)
 544                 req->rq_status = rc;
 545         RETURN(rc);
 546 }
 547
 548 int target_handle_disconnect(struct ptlrpc_request *req)
 549 {
 550         int rc;
 551         ENTRY;
 552
 553         rc = lustre_pack_reply(req, 0, NULL, NULL);
 554         if (rc)
 555                 RETURN(rc);
 556
 557         req->rq_status = obd_disconnect(req->rq_export, 0);
 558         req->rq_export = NULL;
 559         RETURN(0);
 560 }
 561
 562 void target_destroy_export(struct obd_export *exp)
 563 {
 564         /* exports created from last_rcvd data, and "fake"
 565            exports created by lctl don't have an import */
 566         if (exp->exp_imp_reverse != NULL)
 567                 class_destroy_import(exp->exp_imp_reverse);
 568
 569         /* We cancel locks at disconnect time, but this will catch any locks
 570          * granted in a race with recovery-induced disconnect. */
 571         ldlm_cancel_locks_for_export(exp);
 572 }
 573
 574 /*
 575  * Recovery functions
 576  */
 577
 578 void target_cancel_recovery_timer(struct obd_device *obd)
 579 {
 580         del_timer(&obd->obd_recovery_timer);
 581 }
 582
 583 static void abort_delayed_replies(struct obd_device *obd)
 584 {
 585         struct ptlrpc_request *req;
 586         struct list_head *tmp, *n;
 587         list_for_each_safe(tmp, n, &obd->obd_delayed_reply_queue) {
 588                 req = list_entry(tmp, struct ptlrpc_request, rq_list);
 589                 DEBUG_REQ(D_ERROR, req, "aborted:");
 590                 req->rq_status = -ENOTCONN;
 591                 req->rq_type = PTL_RPC_MSG_ERR;
 592                 ptlrpc_reply(req);
 593                 list_del(&req->rq_list);
 594                 OBD_FREE(req->rq_reqmsg, req->rq_reqlen);
 595                 OBD_FREE(req, sizeof *req);
 596         }
 597 }
 598
 599 static void abort_recovery_queue(struct obd_device *obd)
 600 {
 601         struct ptlrpc_request *req;
 602         struct list_head *tmp, *n;
 603         int rc;
 604
 605         list_for_each_safe(tmp, n, &obd->obd_recovery_queue) {
 606                 req = list_entry(tmp, struct ptlrpc_request, rq_list);
 607                 DEBUG_REQ(D_ERROR, req, "aborted:");
 608                 req->rq_status = -ENOTCONN;
 609                 req->rq_type = PTL_RPC_MSG_ERR;
 610                 rc = lustre_pack_reply(req, 0, NULL, NULL);
 611                 if (rc == 0) {
 612                         ptlrpc_reply(req);
 613                 } else {
 614                         DEBUG_REQ(D_ERROR, req,
 615                                   "packing failed for abort-reply; skipping");
 616                 }
 617                 list_del(&req->rq_list);
 618                 class_export_put(req->rq_export);
 619                 OBD_FREE(req->rq_reqmsg, req->rq_reqlen);
 620                 OBD_FREE(req, sizeof *req);
 621         }
 622 }
 623
 624 void target_abort_recovery(void *data)
 625 {
 626         struct obd_device *obd = data;
 627         int rc;
 628
 629         CERROR("disconnecting clients and aborting recovery\n");
 630         spin_lock_bh(&obd->obd_processing_task_lock);
 631         if (!obd->obd_recovering) {
 632                 spin_unlock_bh(&obd->obd_processing_task_lock);
 633                 EXIT;
 634                 return;
 635         }
 636
 637         obd->obd_recovering = obd->obd_abort_recovery = 0;
 638
 639         wake_up(&obd->obd_next_transno_waitq);
 640         target_cancel_recovery_timer(obd);
 641         spin_unlock_bh(&obd->obd_processing_task_lock);
 642
 643         class_disconnect_exports(obd, 0);
 644
 645         /* when recovery was abort, cleanup orphans for mds */
 646         if (OBT(obd) && OBP(obd, postrecov)) {
 647                 rc = OBP(obd, postrecov)(obd);
 648                 if (rc >= 0)
 649                         CWARN("Cleanup %d orphans after recovery was aborted\n", rc);
 650                 else
 651                         CERROR("postrecov failed %d\n", rc);
 652         }
 653
 654         abort_delayed_replies(obd);
 655         abort_recovery_queue(obd);
 656         ptlrpc_run_recovery_over_upcall(obd);
 657 }
 658
 659 static void target_recovery_expired(unsigned long castmeharder)
 660 {
 661         struct obd_device *obd = (struct obd_device *)castmeharder;
 662         CERROR("recovery timed out, aborting\n");
 663         spin_lock_bh(&obd->obd_processing_task_lock);
 664         obd->obd_abort_recovery = 1;
 665         wake_up(&obd->obd_next_transno_waitq);
 666         spin_unlock_bh(&obd->obd_processing_task_lock);
 667 }
 668
 669 static void reset_recovery_timer(struct obd_device *obd)
 670 {
 671         int recovering;
 672         spin_lock(&obd->obd_dev_lock);
 673         recovering = obd->obd_recovering;
 674         spin_unlock(&obd->obd_dev_lock);
 675
 676         if (!recovering)
 677                 return;
 678         CDEBUG(D_HA, "timer will expire in %u seconds\n",
 679                OBD_RECOVERY_TIMEOUT / HZ);
 680         mod_timer(&obd->obd_recovery_timer, jiffies + OBD_RECOVERY_TIMEOUT);
 681 }
 682
 683
 684 /* Only start it the first time called */
 685 void target_start_recovery_timer(struct obd_device *obd, svc_handler_t handler)
 686 {
 687         spin_lock_bh(&obd->obd_processing_task_lock);
 688         if (obd->obd_recovery_handler) {
 689                 spin_unlock_bh(&obd->obd_processing_task_lock);
 690                 return;
 691         }
 692         CWARN("%s: starting recovery timer (%us)\n", obd->obd_name,
 693                OBD_RECOVERY_TIMEOUT / HZ);
 694         obd->obd_recovery_handler = handler;
 695         obd->obd_recovery_timer.function = target_recovery_expired;
 696         obd->obd_recovery_timer.data = (unsigned long)obd;
 697         init_timer(&obd->obd_recovery_timer);
 698         spin_unlock_bh(&obd->obd_processing_task_lock);
 699
 700         reset_recovery_timer(obd);
 701 }
 702
 703 static int check_for_next_transno(struct obd_device *obd)
 704 {
 705         struct ptlrpc_request *req;
 706         int wake_up = 0, connected, completed, queue_len, max;
 707         __u64 next_transno, req_transno;
 708
 709         spin_lock_bh(&obd->obd_processing_task_lock);
 710         req = list_entry(obd->obd_recovery_queue.next,
 711                          struct ptlrpc_request, rq_list);
 712         max = obd->obd_max_recoverable_clients;
 713         req_transno = req->rq_reqmsg->transno;
 714         connected = obd->obd_connected_clients;
 715         completed = max - obd->obd_recoverable_clients;
 716         queue_len = obd->obd_requests_queued_for_recovery;
 717         next_transno = obd->obd_next_recovery_transno;
 718
 719         if (obd->obd_abort_recovery) {
 720                 CDEBUG(D_HA, "waking for aborted recovery\n");
 721                 wake_up = 1;
 722         } else if (!obd->obd_recovering) {
 723                 CDEBUG(D_HA, "waking for completed recovery (?)\n");
 724                 wake_up = 1;
 725         } else if (req_transno == next_transno) {
 726                 CDEBUG(D_HA, "waking for next ("LPD64")\n", next_transno);
 727                 wake_up = 1;
 728         } else if (queue_len + completed == max) {
 729                 CDEBUG(D_ERROR,
 730                        "waking for skipped transno (skip: "LPD64
 731                        ", ql: %d, comp: %d, conn: %d, next: "LPD64")\n",
 732                        next_transno, queue_len, completed, max, req_transno);
 733                 obd->obd_next_recovery_transno = req_transno;
 734                 wake_up = 1;
 735         }
 736         spin_unlock_bh(&obd->obd_processing_task_lock);
 737         LASSERT(req->rq_reqmsg->transno >= next_transno);
 738         return wake_up;
 739 }
 740
 741 static void process_recovery_queue(struct obd_device *obd)
 742 {
 743         struct ptlrpc_request *req;
 744         int abort_recovery = 0;
 745         struct l_wait_info lwi = { 0 };
 746         ENTRY;
 747
 748         for (;;) {
 749                 spin_lock_bh(&obd->obd_processing_task_lock);
 750                 LASSERT(obd->obd_processing_task == current->pid);
 751                 req = list_entry(obd->obd_recovery_queue.next,
 752                                  struct ptlrpc_request, rq_list);
 753
 754                 if (req->rq_reqmsg->transno != obd->obd_next_recovery_transno) {
 755                         spin_unlock_bh(&obd->obd_processing_task_lock);
 756                         CDEBUG(D_HA, "Waiting for transno "LPD64" (1st is "
 757                                LPD64")\n",
 758                                obd->obd_next_recovery_transno,
 759                                req->rq_reqmsg->transno);
 760                         l_wait_event(obd->obd_next_transno_waitq,
 761                                      check_for_next_transno(obd), &lwi);
 762                         spin_lock_bh(&obd->obd_processing_task_lock);
 763                         abort_recovery = obd->obd_abort_recovery;
 764                         spin_unlock_bh(&obd->obd_processing_task_lock);
 765                         if (abort_recovery) {
 766                                 target_abort_recovery(obd);
 767                                 return;
 768                         }
 769                         continue;
 770                 }
 771                 list_del_init(&req->rq_list);
 772                 obd->obd_requests_queued_for_recovery--;
 773                 spin_unlock_bh(&obd->obd_processing_task_lock);
 774
 775                 DEBUG_REQ(D_HA, req, "processing: ");
 776                 (void)obd->obd_recovery_handler(req);
 777                 obd->obd_replayed_requests++;
 778                 reset_recovery_timer(obd);
 779                 /* bug 1580: decide how to properly sync() in recovery */
 780                 //mds_fsync_super(mds->mds_sb);
 781                 class_export_put(req->rq_export);
 782                 OBD_FREE(req->rq_reqmsg, req->rq_reqlen);
 783                 OBD_FREE(req, sizeof *req);
 784                 spin_lock_bh(&obd->obd_processing_task_lock);
 785                 obd->obd_next_recovery_transno++;
 786                 if (list_empty(&obd->obd_recovery_queue)) {
 787                         obd->obd_processing_task = 0;
 788                         spin_unlock_bh(&obd->obd_processing_task_lock);
 789                         break;
 790                 }
 791                 spin_unlock_bh(&obd->obd_processing_task_lock);
 792         }
 793         EXIT;
 794 }
 795
 796 int target_queue_recovery_request(struct ptlrpc_request *req,
 797                                   struct obd_device *obd)
 798 {
 799         struct list_head *tmp;
 800         int inserted = 0;
 801         __u64 transno = req->rq_reqmsg->transno;
 802         struct ptlrpc_request *saved_req;
 803         struct lustre_msg *reqmsg;
 804
 805         /* CAVEAT EMPTOR: The incoming request message has been swabbed
 806          * (i.e. buflens etc are in my own byte order), but type-dependent
 807          * buffers (eg mds_body, ost_body etc) have NOT been swabbed. */
 808
 809         if (!transno) {
 810                 INIT_LIST_HEAD(&req->rq_list);
 811                 DEBUG_REQ(D_HA, req, "not queueing");
 812                 return 1;
 813         }
 814
 815         /* XXX If I were a real man, these LBUGs would be sane cleanups. */
 816         /* XXX just like the request-dup code in queue_final_reply */
 817         OBD_ALLOC(saved_req, sizeof *saved_req);
 818         if (!saved_req)
 819                 LBUG();
 820         OBD_ALLOC(reqmsg, req->rq_reqlen);
 821         if (!reqmsg)
 822                 LBUG();
 823
 824         spin_lock_bh(&obd->obd_processing_task_lock);
 825
 826         /* If we're processing the queue, we want don't want to queue this
 827          * message.
 828          *
 829          * Also, if this request has a transno less than the one we're waiting
 830          * for, we should process it now.  It could (and currently always will)
 831          * be an open request for a descriptor that was opened some time ago.
 832          */
 833         if (obd->obd_processing_task == current->pid ||
 834             transno < obd->obd_next_recovery_transno) {
 835                 /* Processing the queue right now, don't re-add. */
 836                 LASSERT(list_empty(&req->rq_list));
 837                 spin_unlock_bh(&obd->obd_processing_task_lock);
 838                 OBD_FREE(reqmsg, req->rq_reqlen);
 839                 OBD_FREE(saved_req, sizeof *saved_req);
 840                 return 1;
 841         }
 842
 843         memcpy(saved_req, req, sizeof *req);
 844         memcpy(reqmsg, req->rq_reqmsg, req->rq_reqlen);
 845         req = saved_req;
 846         req->rq_reqmsg = reqmsg;
 847         class_export_get(req->rq_export);
 848         INIT_LIST_HEAD(&req->rq_list);
 849
 850         /* XXX O(n^2) */
 851         list_for_each(tmp, &obd->obd_recovery_queue) {
 852                 struct ptlrpc_request *reqiter =
 853                         list_entry(tmp, struct ptlrpc_request, rq_list);
 854
 855                 if (reqiter->rq_reqmsg->transno > transno) {
 856                         list_add_tail(&req->rq_list, &reqiter->rq_list);
 857                         inserted = 1;
 858                         break;
 859                 }
 860         }
 861
 862         if (!inserted) {
 863                 list_add_tail(&req->rq_list, &obd->obd_recovery_queue);
 864         }
 865
 866         obd->obd_requests_queued_for_recovery++;
 867
 868         if (obd->obd_processing_task != 0) {
 869                 /* Someone else is processing this queue, we'll leave it to
 870                  * them.
 871                  */
 872                 wake_up(&obd->obd_next_transno_waitq);
 873                 spin_unlock_bh(&obd->obd_processing_task_lock);
 874                 return 0;
 875         }
 876
 877         /* Nobody is processing, and we know there's (at least) one to process
 878          * now, so we'll do the honours.
 879          */
 880         obd->obd_processing_task = current->pid;
 881         spin_unlock_bh(&obd->obd_processing_task_lock);
 882
 883         process_recovery_queue(obd);
 884         return 0;
 885 }
 886
 887 struct obd_device * target_req2obd(struct ptlrpc_request *req)
 888 {
 889         return req->rq_export->exp_obd;
 890 }
 891
 892 int target_queue_final_reply(struct ptlrpc_request *req, int rc)
 893 {
 894         struct obd_device *obd = target_req2obd(req);
 895         struct ptlrpc_request *saved_req;
 896         struct lustre_msg *reqmsg;
 897         int recovery_done = 0;
 898         int rc2;
 899
 900         if (rc) {
 901                 /* Just like ptlrpc_error, but without the sending. */
 902                 rc = lustre_pack_reply(req, 0, NULL, NULL);
 903                 LASSERT(rc == 0); /* XXX handle this */
 904                 req->rq_type = PTL_RPC_MSG_ERR;
 905         }
 906
 907         LASSERT(list_empty(&req->rq_list));
 908         /* XXX a bit like the request-dup code in queue_recovery_request */
 909         OBD_ALLOC(saved_req, sizeof *saved_req);
 910         if (!saved_req)
 911                 LBUG();
 912         OBD_ALLOC(reqmsg, req->rq_reqlen);
 913         if (!reqmsg)
 914                 LBUG();
 915         memcpy(saved_req, req, sizeof *saved_req);
 916         memcpy(reqmsg, req->rq_reqmsg, req->rq_reqlen);
 917         req = saved_req;
 918         req->rq_reqmsg = reqmsg;
 919         list_add(&req->rq_list, &obd->obd_delayed_reply_queue);
 920
 921         spin_lock_bh(&obd->obd_processing_task_lock);
 922         --obd->obd_recoverable_clients;
 923         recovery_done = (obd->obd_recoverable_clients == 0);
 924         spin_unlock_bh(&obd->obd_processing_task_lock);
 925
 926         if (recovery_done) {
 927                 struct list_head *tmp, *n;
 928                 ldlm_reprocess_all_ns(req->rq_export->exp_obd->obd_namespace);
 929                 CWARN("%s: all clients recovered, sending delayed replies\n",
 930                        obd->obd_name);
 931                 obd->obd_recovering = 0;
 932
 933                 /* when recovering finished, cleanup orphans for mds */
 934                 if (OBT(obd) && OBP(obd, postrecov)) {
 935                         rc2 = OBP(obd, postrecov)(obd);
 936                         if (rc2 >= 0)
 937                                 CWARN("%s: all clients recovered, %d MDS orphans "
 938                                        "deleted\n", obd->obd_name, rc2);
 939                         else
 940                                 CERROR("postrecov failed %d\n", rc2);
 941                 }
 942
 943                 list_for_each_safe(tmp, n, &obd->obd_delayed_reply_queue) {
 944                         req = list_entry(tmp, struct ptlrpc_request, rq_list);
 945                         DEBUG_REQ(D_ERROR, req, "delayed:");
 946                         ptlrpc_reply(req);
 947                         list_del(&req->rq_list);
 948                         OBD_FREE(req->rq_reqmsg, req->rq_reqlen);
 949                         OBD_FREE(req, sizeof *req);
 950                 }
 951                 target_cancel_recovery_timer(obd);
 952         } else {
 953                 CWARN("%s: %d recoverable clients remain\n",
 954                        obd->obd_name, obd->obd_recoverable_clients);
 955                 wake_up(&obd->obd_next_transno_waitq);
 956         }
 957
 958         return 1;
 959 }
 960
 961 static void ptlrpc_abort_reply (struct ptlrpc_request *req)
 962 {
 963         /* On return, we must be sure that the ACK callback has either
 964          * happened or will not happen.  Note that the SENT callback will
 965          * happen come what may since we successfully posted the PUT. */
 966         int rc;
 967         struct l_wait_info lwi;
 968         unsigned long flags;
 969
 970  again:
 971         /* serialise with ACK callback */
 972         spin_lock_irqsave (&req->rq_lock, flags);
 973         if (!req->rq_want_ack) {
 974                 spin_unlock_irqrestore (&req->rq_lock, flags);
 975                 /* The ACK callback has happened already.  Although the
 976                  * SENT callback might still be outstanding (yes really) we
 977                  * don't care; this is just like normal completion. */
 978                 return;
 979         }
 980         spin_unlock_irqrestore (&req->rq_lock, flags);
 981
 982         /* Have a bash at unlinking the MD.  This will fail until the SENT
 983          * callback has happened since the MD is busy from the PUT.  If the
 984          * ACK still hasn't arrived after then, a successful unlink will
 985          * ensure the ACK callback never happens. */
 986         rc = PtlMDUnlink (req->rq_reply_md_h);
 987         switch (rc) {
 988         default:
 989                 LBUG ();
 990         case PTL_OK:
 991                 /* SENT callback happened; ACK callback preempted */
 992                 LASSERT (req->rq_want_ack);
 993                 spin_lock_irqsave (&req->rq_lock, flags);
 994                 req->rq_want_ack = 0;
 995                 spin_unlock_irqrestore (&req->rq_lock, flags);
 996                 return;
 997         case PTL_INV_MD:
 998                 return;
 999         case PTL_MD_INUSE:
1000                 /* Still sending or ACK callback in progress: wait until
1001                  * either callback has completed and try again.
1002                  * Actually we can't wait for the SENT callback because
1003                  * there's no state the SENT callback can touch that will
1004                  * allow it to communicate with us!  So we just wait here
1005                  * for a short time, effectively polling for the SENT
1006                  * callback by calling PtlMDUnlink() again, to see if it
1007                  * has finished.  Note that if the ACK does arrive, its
1008                  * callback wakes us in short order. --eeb */
1009                 lwi = LWI_TIMEOUT (HZ/4, NULL, NULL);
1010                 rc = l_wait_event(req->rq_reply_waitq, !req->rq_want_ack,
1011                                   &lwi);
1012                 CDEBUG (D_HA, "Retrying req %p: %d\n", req, rc);
1013                 /* NB go back and test rq_want_ack with locking, to ensure
1014                  * if ACK callback happened, it has completed stopped
1015                  * referencing this req. */
1016                 goto again;
1017         }
1018 }
1019
1020 void target_send_reply(struct ptlrpc_request *req, int rc, int fail_id)
1021 {
1022         int i;
1023         int netrc;
1024         unsigned long flags;
1025         struct ptlrpc_req_ack_lock *ack_lock;
1026         struct l_wait_info lwi = { 0 };
1027         wait_queue_t commit_wait;
1028         struct obd_device *obd =
1029                 req->rq_export ? req->rq_export->exp_obd : NULL;
1030         struct obd_export *exp = NULL;
1031
1032         if (req->rq_export) {
1033                 for (i = 0; i < REQ_MAX_ACK_LOCKS; i++) {
1034                         if (req->rq_ack_locks[i].mode) {
1035                                 exp = req->rq_export;
1036                                 break;
1037                         }
1038                 }
1039         }
1040
1041         if (exp) {
1042                 exp->exp_outstanding_reply = req;
1043                 spin_lock_irqsave (&req->rq_lock, flags);
1044                 req->rq_want_ack = 1;
1045                 spin_unlock_irqrestore (&req->rq_lock, flags);
1046         }
1047
1048         if (!OBD_FAIL_CHECK(fail_id | OBD_FAIL_ONCE)) {
1049                 if (rc == 0) {
1050                         DEBUG_REQ(D_NET, req, "sending reply");
1051                         netrc = ptlrpc_reply(req);
1052                 } else if (rc == -ENOTCONN) {
1053                         DEBUG_REQ(D_HA, req, "processing error (%d)", rc);
1054                         netrc = ptlrpc_error(req);
1055                 } else {
1056                         DEBUG_REQ(D_ERROR, req, "processing error (%d)", rc);
1057                         netrc = ptlrpc_error(req);
1058                 }
1059         } else {
1060                 obd_fail_loc |= OBD_FAIL_ONCE | OBD_FAILED;
1061                 DEBUG_REQ(D_ERROR, req, "dropping reply");
1062                 if (req->rq_repmsg) {
1063                         OBD_FREE(req->rq_repmsg, req->rq_replen);
1064                         req->rq_repmsg = NULL;
1065                 }
1066                 init_waitqueue_head(&req->rq_reply_waitq);
1067                 netrc = 0;
1068         }
1069
1070         /* a failed send simulates the callbacks */
1071         LASSERT(netrc == 0 || req->rq_want_ack == 0);
1072         if (exp == NULL) {
1073                 LASSERT(req->rq_want_ack == 0);
1074                 return;
1075         }
1076         LASSERT(obd != NULL);
1077
1078         init_waitqueue_entry(&commit_wait, current);
1079         add_wait_queue(&obd->obd_commit_waitq, &commit_wait);
1080         rc = l_wait_event(req->rq_reply_waitq,
1081                           !req->rq_want_ack || req->rq_resent ||
1082                           req->rq_transno <= obd->obd_last_committed, &lwi);
1083         remove_wait_queue(&obd->obd_commit_waitq, &commit_wait);
1084
1085         spin_lock_irqsave (&req->rq_lock, flags);
1086         /* If we got here because the ACK callback ran, this acts as a
1087          * barrier to ensure the callback completed the wakeup. */
1088         spin_unlock_irqrestore (&req->rq_lock, flags);
1089
1090         /* If we committed the transno already, then we might wake up before
1091          * the ack arrives.  We need to stop waiting for the ack before we can
1092          * reuse this request structure.  We are guaranteed by this point that
1093          * this cannot abort the sending of the actual reply.*/
1094         ptlrpc_abort_reply(req);
1095
1096         if (req->rq_resent) {
1097                 DEBUG_REQ(D_HA, req, "resent: not cancelling locks");
1098                 return;
1099         }
1100
1101         LASSERT(rc == 0);
1102         DEBUG_REQ(D_HA, req, "cancelling locks for %s",
1103                   req->rq_want_ack ? "commit" : "ack");
1104
1105         exp->exp_outstanding_reply = NULL;
1106
1107         for (ack_lock = req->rq_ack_locks, i = 0;
1108              i < REQ_MAX_ACK_LOCKS; i++, ack_lock++) {
1109                 if (!ack_lock->mode)
1110                         continue;
1111                 ldlm_lock_decref(&ack_lock->lock, ack_lock->mode);
1112         }
1113 }
1114
1115 int target_handle_ping(struct ptlrpc_request *req)
1116 {
1117         return lustre_pack_reply(req, 0, NULL, NULL);
1118 }
1119
1120 void *ldlm_put_lock_into_req(struct ptlrpc_request *req,
1121                                 struct lustre_handle *lock, int mode)
1122 {
1123         int i;
1124
1125         for (i = 0; i < REQ_MAX_ACK_LOCKS; i++) {
1126                 if (req->rq_ack_locks[i].mode)
1127                         continue;
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 }
1136