lustre/ldlm/ldlm_lib.c

   1 /* -*- mode: c; c-basic-offset: 8; indent-tabs-mode: nil; -*-
   2  * vim:expandtab:shiftwidth=8:tabstop=8:
   3  *
   4  *  Copyright (c) 2003 Cluster File Systems, Inc.
   5  *
   6  *   This file is part of Lustre, http://www.lustre.org.
   7  *
   8  *   Lustre is free software; you can redistribute it and/or
   9  *   modify it under the terms of version 2 of the GNU General Public
  10  *   License as published by the Free Software Foundation.
  11  *
  12  *   Lustre is distributed in the hope that it will be useful,
  13  *   but WITHOUT ANY WARRANTY; without even the implied warranty of
  14  *   MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
  15  *   GNU General Public License for more details.
  16  *
  17  *   You should have received a copy of the GNU General Public License
  18  *   along with Lustre; if not, write to the Free Software
  19  *   Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA.
  20  */
  21
  22 #ifndef EXPORT_SYMTAB
  23 # define EXPORT_SYMTAB
  24 #endif
  25 #define DEBUG_SUBSYSTEM S_LDLM
  26
  27 #ifdef __KERNEL__
  28 # include <linux/module.h>
  29 #else
  30 # include <liblustre.h>
  31 #endif
  32 #include <linux/obd.h>
  33 #include <linux/obd_ost.h> /* for LUSTRE_OSC_NAME */
  34 #include <linux/lustre_mds.h> /* for LUSTRE_MDC_NAME */
  35 #include <linux/lustre_mgmt.h>
  36 #include <linux/lustre_dlm.h>
  37 #include <linux/lustre_net.h>
  38
  39 int client_obd_setup(struct obd_device *obddev, obd_count len, void *buf)
  40 {
  41         struct ptlrpc_connection *conn;
  42         struct lustre_cfg* lcfg = buf;
  43         struct client_obd *cli = &obddev->u.cli;
  44         struct obd_import *imp;
  45         struct obd_uuid server_uuid;
  46         int rq_portal, rp_portal, connect_op;
  47         char *name = obddev->obd_type->typ_name;
  48         char *mgmt_name = NULL;
  49         int rc = 0;
  50         struct obd_device *mgmt_obd;
  51         mgmtcli_register_for_events_t register_f;
  52         ENTRY;
  53
  54         /* In a more perfect world, we would hang a ptlrpc_client off of
  55          * obd_type and just use the values from there. */
  56         if (!strcmp(name, LUSTRE_OSC_NAME)) {
  57                 rq_portal = OST_REQUEST_PORTAL;
  58                 rp_portal = OSC_REPLY_PORTAL;
  59                 connect_op = OST_CONNECT;
  60         } else if (!strcmp(name, LUSTRE_MDC_NAME)) {
  61                 rq_portal = MDS_REQUEST_PORTAL;
  62                 rp_portal = MDC_REPLY_PORTAL;
  63                 connect_op = MDS_CONNECT;
  64         } else if (!strcmp(name, LUSTRE_MGMTCLI_NAME)) {
  65                 rq_portal = MGMT_REQUEST_PORTAL;
  66                 rp_portal = MGMT_REPLY_PORTAL;
  67                 connect_op = MGMT_CONNECT;
  68         } else {
  69                 CERROR("unknown client OBD type \"%s\", can't setup\n",
  70                        name);
  71                 RETURN(-EINVAL);
  72         }
  73
  74         if (lcfg->lcfg_inllen1 < 1) {
  75                 CERROR("requires a TARGET UUID\n");
  76                 RETURN(-EINVAL);
  77         }
  78
  79         if (lcfg->lcfg_inllen1 > 37) {
  80                 CERROR("client UUID must be less than 38 characters\n");
  81                 RETURN(-EINVAL);
  82         }
  83
  84         if (lcfg->lcfg_inllen2 < 1) {
  85                 CERROR("setup requires a SERVER UUID\n");
  86                 RETURN(-EINVAL);
  87         }
  88
  89         if (lcfg->lcfg_inllen2 > 37) {
  90                 CERROR("target UUID must be less than 38 characters\n");
  91                 RETURN(-EINVAL);
  92         }
  93
  94         sema_init(&cli->cl_sem, 1);
  95         cli->cl_conn_count = 0;
  96         memcpy(server_uuid.uuid, lcfg->lcfg_inlbuf2, MIN(lcfg->lcfg_inllen2,
  97                                                         sizeof(server_uuid)));
  98
  99         cli->cl_dirty = 0;
 100         cli->cl_dirty_granted = 0;
 101         cli->cl_dirty_max = 64*1024*1024; /* some default */
 102         cli->cl_ost_can_grant = 1;
 103         INIT_LIST_HEAD(&cli->cl_cache_waiters);
 104         INIT_LIST_HEAD(&cli->cl_loi_ready_list);
 105         spin_lock_init(&cli->cl_loi_list_lock);
 106         cli->cl_brw_in_flight = 0;
 107         spin_lock_init(&cli->cl_read_rpc_hist.oh_lock);
 108         spin_lock_init(&cli->cl_write_rpc_hist.oh_lock);
 109         spin_lock_init(&cli->cl_read_page_hist.oh_lock);
 110         spin_lock_init(&cli->cl_write_page_hist.oh_lock);
 111         cli->cl_max_pages_per_rpc = PTL_MD_MAX_PAGES;
 112         cli->cl_max_rpcs_in_flight = 8;
 113
 114         ldlm_get_ref();
 115         if (rc) {
 116                 CERROR("ldlm_get_ref failed: %d\n", rc);
 117                 GOTO(err, rc);
 118         }
 119
 120         conn = ptlrpc_uuid_to_connection(&server_uuid);
 121         if (conn == NULL)
 122                 GOTO(err_ldlm, rc = -ENOENT);
 123
 124         ptlrpc_init_client(rq_portal, rp_portal, name,
 125                            &obddev->obd_ldlm_client);
 126
 127         imp = class_new_import();
 128         if (imp == NULL) {
 129                 ptlrpc_put_connection(conn);
 130                 GOTO(err_ldlm, rc = -ENOENT);
 131         }
 132         imp->imp_connection = conn;
 133         imp->imp_client = &obddev->obd_ldlm_client;
 134         imp->imp_obd = obddev;
 135         imp->imp_connect_op = connect_op;
 136         imp->imp_generation = 0;
 137         INIT_LIST_HEAD(&imp->imp_pinger_chain);
 138         memcpy(imp->imp_target_uuid.uuid, lcfg->lcfg_inlbuf1,
 139               lcfg->lcfg_inllen1);
 140         class_import_put(imp);
 141
 142         cli->cl_import = imp;
 143         cli->cl_max_mds_easize = sizeof(struct lov_mds_md);
 144         cli->cl_max_mds_cookiesize = sizeof(struct llog_cookie);
 145         cli->cl_sandev = to_kdev_t(0);
 146
 147         if (lcfg->lcfg_inllen3 != 0) {
 148                 if (!strcmp(lcfg->lcfg_inlbuf3, "inactive")) {
 149                         CDEBUG(D_HA, "marking %s %s->%s as inactive\n",
 150                                name, obddev->obd_name,
 151                                imp->imp_target_uuid.uuid);
 152                         imp->imp_invalid = 1;
 153
 154                         if (lcfg->lcfg_inllen4 != 0)
 155                                 mgmt_name = lcfg->lcfg_inlbuf4;
 156                 } else {
 157                         mgmt_name = lcfg->lcfg_inlbuf3;
 158                 }
 159         }
 160
 161         if (mgmt_name != NULL) {
 162                 /* Register with management client if we need to. */
 163                 CDEBUG(D_HA, "%s registering with %s for events about %s\n",
 164                        obddev->obd_name, mgmt_name, server_uuid.uuid);
 165
 166                 mgmt_obd = class_name2obd(mgmt_name);
 167                 if (!mgmt_obd) {
 168                         CERROR("can't find mgmtcli %s to register\n",
 169                                mgmt_name);
 170                         GOTO(err_import, rc = -ENOSYS);
 171                 }
 172
 173                 register_f = inter_module_get("mgmtcli_register_for_events");
 174                 if (!register_f) {
 175                         CERROR("can't i_m_g mgmtcli_register_for_events\n");
 176                         GOTO(err_import, rc = -ENOSYS);
 177                 }
 178
 179                 rc = register_f(mgmt_obd, obddev, &imp->imp_target_uuid);
 180                 inter_module_put("mgmtcli_register_for_events");
 181
 182                 if (!rc)
 183                         cli->cl_mgmtcli_obd = mgmt_obd;
 184         }
 185
 186         RETURN(rc);
 187
 188 err_import:
 189         class_destroy_import(imp);
 190 err_ldlm:
 191         ldlm_put_ref(0);
 192 err:
 193         RETURN(rc);
 194
 195 }
 196
 197 int client_obd_cleanup(struct obd_device *obddev, int flags)
 198 {
 199         struct client_obd *cli = &obddev->u.cli;
 200
 201         if (!cli->cl_import)
 202                 RETURN(-EINVAL);
 203         if (cli->cl_mgmtcli_obd) {
 204                 mgmtcli_deregister_for_events_t dereg_f;
 205
 206                 dereg_f = inter_module_get("mgmtcli_deregister_for_events");
 207                 dereg_f(cli->cl_mgmtcli_obd, obddev);
 208                 inter_module_put("mgmtcli_deregister_for_events");
 209         }
 210         class_destroy_import(cli->cl_import);
 211         cli->cl_import = NULL;
 212
 213         ldlm_put_ref(flags & OBD_OPT_FORCE);
 214
 215         RETURN(0);
 216 }
 217
 218 int client_connect_import(struct lustre_handle *dlm_handle,
 219                           struct obd_device *obd,
 220                           struct obd_uuid *cluuid)
 221 {
 222         struct client_obd *cli = &obd->u.cli;
 223         struct obd_import *imp = cli->cl_import;
 224         struct obd_export *exp;
 225         int rc;
 226         ENTRY;
 227
 228         down(&cli->cl_sem);
 229         rc = class_connect(dlm_handle, obd, cluuid);
 230         if (rc)
 231                 GOTO(out_sem, rc);
 232
 233         cli->cl_conn_count++;
 234         if (cli->cl_conn_count > 1)
 235                 GOTO(out_sem, rc);
 236         exp = class_conn2export(dlm_handle);
 237
 238         if (obd->obd_namespace != NULL)
 239                 CERROR("already have namespace!\n");
 240         obd->obd_namespace = ldlm_namespace_new(obd->obd_name,
 241                                                 LDLM_NAMESPACE_CLIENT);
 242         if (obd->obd_namespace == NULL)
 243                 GOTO(out_disco, rc = -ENOMEM);
 244
 245         imp->imp_dlm_handle = *dlm_handle;
 246         imp->imp_state = LUSTRE_IMP_DISCON;
 247
 248         rc = ptlrpc_connect_import(imp);
 249         if (rc != 0) {
 250                 LASSERT (imp->imp_state == LUSTRE_IMP_DISCON);
 251                 GOTO(out_ldlm, rc);
 252         }
 253
 254         LASSERT (imp->imp_state == LUSTRE_IMP_FULL);
 255
 256         exp->exp_connection = ptlrpc_connection_addref(imp->imp_connection);
 257
 258         if (imp->imp_replayable) {
 259                 CDEBUG(D_HA, "connected to replayable target: %s\n",
 260                        imp->imp_target_uuid.uuid);
 261                 ptlrpc_pinger_add_import(imp);
 262         }
 263
 264         CDEBUG(D_HA, "local import: %p, remote handle: "LPX64"\n", imp,
 265                imp->imp_remote_handle.cookie);
 266
 267         EXIT;
 268
 269         if (rc) {
 270 out_ldlm:
 271                 ldlm_namespace_free(obd->obd_namespace, 0);
 272                 obd->obd_namespace = NULL;
 273 out_disco:
 274                 cli->cl_conn_count--;
 275                 class_disconnect(exp, 0);
 276         } else {
 277                 class_export_put(exp);
 278         }
 279 out_sem:
 280         up(&cli->cl_sem);
 281         return rc;
 282 }
 283
 284 int client_disconnect_export(struct obd_export *exp, int failover)
 285 {
 286         struct obd_device *obd = class_exp2obd(exp);
 287         struct client_obd *cli = &obd->u.cli;
 288         struct obd_import *imp = cli->cl_import;
 289         int rc = 0, err;
 290         ENTRY;
 291
 292         if (!obd) {
 293                 CERROR("invalid export for disconnect: exp %p cookie "LPX64"\n",
 294                        exp, exp ? exp->exp_handle.h_cookie : -1);
 295                 RETURN(-EINVAL);
 296         }
 297
 298         down(&cli->cl_sem);
 299         if (!cli->cl_conn_count) {
 300                 CERROR("disconnecting disconnected device (%s)\n",
 301                        obd->obd_name);
 302                 GOTO(out_sem, rc = -EINVAL);
 303         }
 304
 305         cli->cl_conn_count--;
 306         if (cli->cl_conn_count)
 307                 GOTO(out_no_disconnect, rc = 0);
 308
 309         /* Some non-replayable imports (MDS's OSCs) are pinged, so just
 310          * delete it regardless.  (It's safe to delete an import that was
 311          * never added.) */
 312         (void)ptlrpc_pinger_del_import(imp);
 313
 314         if (obd->obd_namespace != NULL) {
 315                 /* obd_no_recov == local only */
 316                 ldlm_cli_cancel_unused(obd->obd_namespace, NULL,
 317                                        obd->obd_no_recov, NULL);
 318                 ldlm_namespace_free(obd->obd_namespace, obd->obd_no_recov);
 319                 obd->obd_namespace = NULL;
 320         }
 321
 322         /* Yeah, obd_no_recov also (mainly) means "forced shutdown". */
 323         if (obd->obd_no_recov)
 324                 ptlrpc_set_import_active(imp, 0);
 325         else
 326                 rc = ptlrpc_disconnect_import(imp);
 327
 328         imp->imp_state = LUSTRE_IMP_NEW;
 329
 330         EXIT;
 331  out_no_disconnect:
 332         err = class_disconnect(exp, 0);
 333         if (!rc && err)
 334                 rc = err;
 335  out_sem:
 336         up(&cli->cl_sem);
 337         RETURN(rc);
 338 }
 339
 340 /* --------------------------------------------------------------------------
 341  * from old lib/target.c
 342  * -------------------------------------------------------------------------- */
 343
 344 int target_handle_reconnect(struct lustre_handle *conn, struct obd_export *exp,
 345                             struct obd_uuid *cluuid)
 346 {
 347         if (exp->exp_connection) {
 348                 struct lustre_handle *hdl;
 349                 hdl = &exp->exp_imp_reverse->imp_remote_handle;
 350                 /* Might be a re-connect after a partition. */
 351                 if (!memcmp(&conn->cookie, &hdl->cookie, sizeof conn->cookie)) {
 352                         CERROR("%s reconnecting\n", cluuid->uuid);
 353                         conn->cookie = exp->exp_handle.h_cookie;
 354                         RETURN(EALREADY);
 355                 } else {
 356                         CERROR("%s reconnecting from %s, "
 357                                "handle mismatch (ours "LPX64", theirs "
 358                                LPX64")\n", cluuid->uuid,
 359                                exp->exp_connection->c_remote_uuid.uuid,
 360                                hdl->cookie, conn->cookie);
 361                         memset(conn, 0, sizeof *conn);
 362                         RETURN(-EALREADY);
 363                 }
 364         }
 365
 366         conn->cookie = exp->exp_handle.h_cookie;
 367         CDEBUG(D_INFO, "existing export for UUID '%s' at %p\n",
 368                cluuid->uuid, exp);
 369         CDEBUG(D_IOCTL,"connect: cookie "LPX64"\n", conn->cookie);
 370         RETURN(0);
 371 }
 372
 373 int target_handle_connect(struct ptlrpc_request *req, svc_handler_t handler)
 374 {
 375         struct obd_device *target;
 376         struct obd_export *export = NULL;
 377         struct obd_import *revimp;
 378         struct lustre_handle conn;
 379         struct obd_uuid tgtuuid;
 380         struct obd_uuid cluuid;
 381         struct obd_uuid remote_uuid;
 382         struct list_head *p;
 383         char *str, *tmp;
 384         int rc = 0, abort_recovery;
 385         ENTRY;
 386
 387         LASSERT_REQSWAB (req, 0);
 388         str = lustre_msg_string(req->rq_reqmsg, 0, sizeof(tgtuuid) - 1);
 389         if (str == NULL) {
 390                 CERROR("bad target UUID for connect\n");
 391                 GOTO(out, rc = -EINVAL);
 392         }
 393
 394         obd_str2uuid (&tgtuuid, str);
 395         target = class_uuid2obd(&tgtuuid);
 396         if (!target) {
 397                 target = class_name2obd(str);
 398         }
 399
 400         if (!target || target->obd_stopping || !target->obd_set_up) {
 401                 CERROR("UUID '%s' is not available for connect\n", str);
 402                 GOTO(out, rc = -ENODEV);
 403         }
 404
 405         LASSERT_REQSWAB (req, 1);
 406         str = lustre_msg_string(req->rq_reqmsg, 1, sizeof(cluuid) - 1);
 407         if (str == NULL) {
 408                 CERROR("bad client UUID for connect\n");
 409                 GOTO(out, rc = -EINVAL);
 410         }
 411
 412         obd_str2uuid (&cluuid, str);
 413
 414         /* XXX extract a nettype and format accordingly */
 415         snprintf(remote_uuid.uuid, sizeof remote_uuid,
 416                  "NET_"LPX64"_UUID", req->rq_peer.peer_nid);
 417
 418         spin_lock_bh(&target->obd_processing_task_lock);
 419         abort_recovery = target->obd_abort_recovery;
 420         spin_unlock_bh(&target->obd_processing_task_lock);
 421         if (abort_recovery)
 422                 target_abort_recovery(target);
 423
 424         tmp = lustre_msg_buf(req->rq_reqmsg, 2, sizeof conn);
 425         if (tmp == NULL)
 426                 GOTO(out, rc = -EPROTO);
 427
 428         memcpy(&conn, tmp, sizeof conn);
 429
 430         rc = lustre_pack_reply(req, 0, NULL, NULL);
 431         if (rc)
 432                 GOTO(out, rc);
 433
 434         /* lctl gets a backstage, all-access pass. */
 435         if (obd_uuid_equals(&cluuid, &target->obd_uuid))
 436                 goto dont_check_exports;
 437
 438         spin_lock(&target->obd_dev_lock);
 439         list_for_each(p, &target->obd_exports) {
 440                 export = list_entry(p, struct obd_export, exp_obd_chain);
 441                 if (obd_uuid_equals(&cluuid, &export->exp_client_uuid)) {
 442                         spin_unlock(&target->obd_dev_lock);
 443                         LASSERT(export->exp_obd == target);
 444
 445                         rc = target_handle_reconnect(&conn, export, &cluuid);
 446                         break;
 447                 }
 448                 export = NULL;
 449         }
 450         /* If we found an export, we already unlocked. */
 451         if (!export) {
 452                 spin_unlock(&target->obd_dev_lock);
 453         } else if (req->rq_reqmsg->conn_cnt == 1) {
 454                 CERROR("%s reconnected with 1 conn_cnt; cookies not random?\n",
 455                        cluuid.uuid);
 456                 GOTO(out, rc = -EALREADY);
 457         }
 458
 459         /* Tell the client if we're in recovery. */
 460         /* If this is the first client, start the recovery timer */
 461         if (target->obd_recovering) {
 462                 lustre_msg_add_op_flags(req->rq_repmsg, MSG_CONNECT_RECOVERING);
 463                 target_start_recovery_timer(target, handler);
 464         }
 465
 466         /* Tell the client if we support replayable requests */
 467         if (target->obd_replayable)
 468                 lustre_msg_add_op_flags(req->rq_repmsg, MSG_CONNECT_REPLAYABLE);
 469
 470         if (export == NULL) {
 471                 if (target->obd_recovering) {
 472                         CERROR("denying connection for new client %s: "
 473                                "%d clients in recovery for %lds\n", cluuid.uuid,
 474                                target->obd_recoverable_clients,
 475                                (target->obd_recovery_timer.expires-jiffies)/HZ);
 476                         rc = -EBUSY;
 477                 } else {
 478  dont_check_exports:
 479                         rc = obd_connect(&conn, target, &cluuid);
 480                 }
 481         }
 482
 483         /* If all else goes well, this is our RPC return code. */
 484         req->rq_status = 0;
 485
 486         if (rc && rc != EALREADY)
 487                 GOTO(out, rc);
 488
 489         /* XXX track this all the time? */
 490         if (target->obd_recovering) {
 491                 target->obd_connected_clients++;
 492         }
 493
 494         req->rq_repmsg->handle = conn;
 495
 496         /* If the client and the server are the same node, we will already
 497          * have an export that really points to the client's DLM export,
 498          * because we have a shared handles table.
 499          *
 500          * XXX this will go away when shaver stops sending the "connect" handle
 501          * in the real "remote handle" field of the request --phik 24 Apr 2003
 502          */
 503         if (req->rq_export != NULL)
 504                 class_export_put(req->rq_export);
 505
 506         /* ownership of this export ref transfers to the request */
 507         export = req->rq_export = class_conn2export(&conn);
 508         LASSERT(export != NULL);
 509
 510         if (req->rq_connection != NULL)
 511                 ptlrpc_put_connection(req->rq_connection);
 512         if (export->exp_connection != NULL)
 513                 ptlrpc_put_connection(export->exp_connection);
 514         export->exp_connection = ptlrpc_get_connection(&req->rq_peer,
 515                                                        &remote_uuid);
 516         req->rq_connection = ptlrpc_connection_addref(export->exp_connection);
 517
 518         LASSERT(export->exp_conn_cnt < req->rq_reqmsg->conn_cnt);
 519         export->exp_conn_cnt = req->rq_reqmsg->conn_cnt;
 520
 521         if (rc == EALREADY) {
 522                 /* We indicate the reconnection in a flag, not an error code. */
 523                 lustre_msg_add_op_flags(req->rq_repmsg, MSG_CONNECT_RECONNECT);
 524                 GOTO(out, rc = 0);
 525         }
 526
 527         memcpy(&conn, lustre_msg_buf(req->rq_reqmsg, 2, sizeof conn),
 528                sizeof conn);
 529
 530         if (export->exp_imp_reverse != NULL)
 531                 class_destroy_import(export->exp_imp_reverse);
 532         revimp = export->exp_imp_reverse = class_new_import();
 533         revimp->imp_connection = ptlrpc_connection_addref(req->rq_connection);
 534         revimp->imp_client = &export->exp_obd->obd_ldlm_client;
 535         revimp->imp_remote_handle = conn;
 536         revimp->imp_obd = target;
 537         revimp->imp_dlm_fake = 1;
 538         revimp->imp_state = LUSTRE_IMP_FULL;
 539         class_import_put(revimp);
 540 out:
 541         if (rc)
 542                 req->rq_status = rc;
 543         RETURN(rc);
 544 }
 545
 546 int target_handle_disconnect(struct ptlrpc_request *req)
 547 {
 548         int rc;
 549         ENTRY;
 550
 551         rc = lustre_pack_reply(req, 0, NULL, NULL);
 552         if (rc)
 553                 RETURN(rc);
 554
 555         req->rq_status = obd_disconnect(req->rq_export, 0);
 556         req->rq_export = NULL;
 557         RETURN(0);
 558 }
 559
 560 void target_destroy_export(struct obd_export *exp)
 561 {
 562         /* exports created from last_rcvd data, and "fake"
 563            exports created by lctl don't have an import */
 564         if (exp->exp_imp_reverse != NULL)
 565                 class_destroy_import(exp->exp_imp_reverse);
 566
 567         /* We cancel locks at disconnect time, but this will catch any locks
 568          * granted in a race with recovery-induced disconnect. */
 569         ldlm_cancel_locks_for_export(exp);
 570 }
 571
 572 /*
 573  * Recovery functions
 574  */
 575
 576 void target_cancel_recovery_timer(struct obd_device *obd)
 577 {
 578         del_timer(&obd->obd_recovery_timer);
 579 }
 580
 581 static void abort_delayed_replies(struct obd_device *obd)
 582 {
 583         struct ptlrpc_request *req;
 584         struct list_head *tmp, *n;
 585         list_for_each_safe(tmp, n, &obd->obd_delayed_reply_queue) {
 586                 req = list_entry(tmp, struct ptlrpc_request, rq_list);
 587                 DEBUG_REQ(D_ERROR, req, "aborted:");
 588                 req->rq_status = -ENOTCONN;
 589                 req->rq_type = PTL_RPC_MSG_ERR;
 590                 ptlrpc_reply(req);
 591                 list_del(&req->rq_list);
 592                 OBD_FREE(req->rq_reqmsg, req->rq_reqlen);
 593                 OBD_FREE(req, sizeof *req);
 594         }
 595 }
 596
 597 static void abort_recovery_queue(struct obd_device *obd)
 598 {
 599         struct ptlrpc_request *req;
 600         struct list_head *tmp, *n;
 601         int rc;
 602
 603         list_for_each_safe(tmp, n, &obd->obd_recovery_queue) {
 604                 req = list_entry(tmp, struct ptlrpc_request, rq_list);
 605                 DEBUG_REQ(D_ERROR, req, "aborted:");
 606                 req->rq_status = -ENOTCONN;
 607                 req->rq_type = PTL_RPC_MSG_ERR;
 608                 rc = lustre_pack_reply(req, 0, NULL, NULL);
 609                 if (rc == 0) {
 610                         ptlrpc_reply(req);
 611                 } else {
 612                         DEBUG_REQ(D_ERROR, req,
 613                                   "packing failed for abort-reply; skipping");
 614                 }
 615                 list_del(&req->rq_list);
 616                 class_export_put(req->rq_export);
 617                 OBD_FREE(req->rq_reqmsg, req->rq_reqlen);
 618                 OBD_FREE(req, sizeof *req);
 619         }
 620 }
 621
 622 void target_abort_recovery(void *data)
 623 {
 624         struct obd_device *obd = data;
 625         int rc;
 626
 627         CERROR("disconnecting clients and aborting recovery\n");
 628         spin_lock_bh(&obd->obd_processing_task_lock);
 629         if (!obd->obd_recovering) {
 630                 spin_unlock_bh(&obd->obd_processing_task_lock);
 631                 EXIT;
 632                 return;
 633         }
 634
 635         obd->obd_recovering = obd->obd_abort_recovery = 0;
 636
 637         wake_up(&obd->obd_next_transno_waitq);
 638         target_cancel_recovery_timer(obd);
 639         spin_unlock_bh(&obd->obd_processing_task_lock);
 640
 641         class_disconnect_exports(obd, 0);
 642
 643         /* when recovery was abort, cleanup orphans for mds */
 644         if (OBT(obd) && OBP(obd, postrecov)) {
 645                 rc = OBP(obd, postrecov)(obd);
 646                 if (rc >= 0)
 647                         CWARN("Cleanup %d orphans after recovery was aborted\n", rc);
 648                 else
 649                         CERROR("postrecov failed %d\n", rc);
 650         }
 651
 652         abort_delayed_replies(obd);
 653         abort_recovery_queue(obd);
 654         ptlrpc_run_recovery_over_upcall(obd);
 655 }
 656
 657 static void target_recovery_expired(unsigned long castmeharder)
 658 {
 659         struct obd_device *obd = (struct obd_device *)castmeharder;
 660         CERROR("recovery timed out, aborting\n");
 661         spin_lock_bh(&obd->obd_processing_task_lock);
 662         obd->obd_abort_recovery = 1;
 663         wake_up(&obd->obd_next_transno_waitq);
 664         spin_unlock_bh(&obd->obd_processing_task_lock);
 665 }
 666
 667 static void reset_recovery_timer(struct obd_device *obd)
 668 {
 669         int recovering;
 670         spin_lock(&obd->obd_dev_lock);
 671         recovering = obd->obd_recovering;
 672         spin_unlock(&obd->obd_dev_lock);
 673
 674         if (!recovering)
 675                 return;
 676         CDEBUG(D_HA, "timer will expire in %u seconds\n",
 677                OBD_RECOVERY_TIMEOUT / HZ);
 678         mod_timer(&obd->obd_recovery_timer, jiffies + OBD_RECOVERY_TIMEOUT);
 679 }
 680
 681
 682 /* Only start it the first time called */
 683 void target_start_recovery_timer(struct obd_device *obd, svc_handler_t handler)
 684 {
 685         spin_lock_bh(&obd->obd_processing_task_lock);
 686         if (obd->obd_recovery_handler) {
 687                 spin_unlock_bh(&obd->obd_processing_task_lock);
 688                 return;
 689         }
 690         CWARN("%s: starting recovery timer (%us)\n", obd->obd_name,
 691                OBD_RECOVERY_TIMEOUT / HZ);
 692         obd->obd_recovery_handler = handler;
 693         obd->obd_recovery_timer.function = target_recovery_expired;
 694         obd->obd_recovery_timer.data = (unsigned long)obd;
 695         init_timer(&obd->obd_recovery_timer);
 696         spin_unlock_bh(&obd->obd_processing_task_lock);
 697
 698         reset_recovery_timer(obd);
 699 }
 700
 701 static int check_for_next_transno(struct obd_device *obd)
 702 {
 703         struct ptlrpc_request *req;
 704         int wake_up = 0, connected, completed, queue_len, max;
 705         __u64 next_transno, req_transno;
 706
 707         spin_lock_bh(&obd->obd_processing_task_lock);
 708         req = list_entry(obd->obd_recovery_queue.next,
 709                          struct ptlrpc_request, rq_list);
 710         max = obd->obd_max_recoverable_clients;
 711         req_transno = req->rq_reqmsg->transno;
 712         connected = obd->obd_connected_clients;
 713         completed = max - obd->obd_recoverable_clients;
 714         queue_len = obd->obd_requests_queued_for_recovery;
 715         next_transno = obd->obd_next_recovery_transno;
 716
 717         if (obd->obd_abort_recovery) {
 718                 CDEBUG(D_HA, "waking for aborted recovery\n");
 719                 wake_up = 1;
 720         } else if (!obd->obd_recovering) {
 721                 CDEBUG(D_HA, "waking for completed recovery (?)\n");
 722                 wake_up = 1;
 723         } else if (req_transno == next_transno) {
 724                 CDEBUG(D_HA, "waking for next ("LPD64")\n", next_transno);
 725                 wake_up = 1;
 726         } else if (queue_len + completed == max) {
 727                 CDEBUG(D_ERROR,
 728                        "waking for skipped transno (skip: "LPD64
 729                        ", ql: %d, comp: %d, conn: %d, next: "LPD64")\n",
 730                        next_transno, queue_len, completed, max, req_transno);
 731                 obd->obd_next_recovery_transno = req_transno;
 732                 wake_up = 1;
 733         }
 734         spin_unlock_bh(&obd->obd_processing_task_lock);
 735         LASSERT(req->rq_reqmsg->transno >= next_transno);
 736         return wake_up;
 737 }
 738
 739 static void process_recovery_queue(struct obd_device *obd)
 740 {
 741         struct ptlrpc_request *req;
 742         int abort_recovery = 0;
 743         struct l_wait_info lwi = { 0 };
 744         ENTRY;
 745
 746         for (;;) {
 747                 spin_lock_bh(&obd->obd_processing_task_lock);
 748                 LASSERT(obd->obd_processing_task == current->pid);
 749                 req = list_entry(obd->obd_recovery_queue.next,
 750                                  struct ptlrpc_request, rq_list);
 751
 752                 if (req->rq_reqmsg->transno != obd->obd_next_recovery_transno) {
 753                         spin_unlock_bh(&obd->obd_processing_task_lock);
 754                         CDEBUG(D_HA, "Waiting for transno "LPD64" (1st is "
 755                                LPD64")\n",
 756                                obd->obd_next_recovery_transno,
 757                                req->rq_reqmsg->transno);
 758                         l_wait_event(obd->obd_next_transno_waitq,
 759                                      check_for_next_transno(obd), &lwi);
 760                         spin_lock_bh(&obd->obd_processing_task_lock);
 761                         abort_recovery = obd->obd_abort_recovery;
 762                         spin_unlock_bh(&obd->obd_processing_task_lock);
 763                         if (abort_recovery) {
 764                                 target_abort_recovery(obd);
 765                                 return;
 766                         }
 767                         continue;
 768                 }
 769                 list_del_init(&req->rq_list);
 770                 obd->obd_requests_queued_for_recovery--;
 771                 spin_unlock_bh(&obd->obd_processing_task_lock);
 772
 773                 DEBUG_REQ(D_HA, req, "processing: ");
 774                 (void)obd->obd_recovery_handler(req);
 775                 obd->obd_replayed_requests++;
 776                 reset_recovery_timer(obd);
 777                 /* bug 1580: decide how to properly sync() in recovery */
 778                 //mds_fsync_super(mds->mds_sb);
 779                 class_export_put(req->rq_export);
 780                 OBD_FREE(req->rq_reqmsg, req->rq_reqlen);
 781                 OBD_FREE(req, sizeof *req);
 782                 spin_lock_bh(&obd->obd_processing_task_lock);
 783                 obd->obd_next_recovery_transno++;
 784                 if (list_empty(&obd->obd_recovery_queue)) {
 785                         obd->obd_processing_task = 0;
 786                         spin_unlock_bh(&obd->obd_processing_task_lock);
 787                         break;
 788                 }
 789                 spin_unlock_bh(&obd->obd_processing_task_lock);
 790         }
 791         EXIT;
 792 }
 793
 794 int target_queue_recovery_request(struct ptlrpc_request *req,
 795                                   struct obd_device *obd)
 796 {
 797         struct list_head *tmp;
 798         int inserted = 0;
 799         __u64 transno = req->rq_reqmsg->transno;
 800         struct ptlrpc_request *saved_req;
 801         struct lustre_msg *reqmsg;
 802
 803         /* CAVEAT EMPTOR: The incoming request message has been swabbed
 804          * (i.e. buflens etc are in my own byte order), but type-dependent
 805          * buffers (eg mds_body, ost_body etc) have NOT been swabbed. */
 806
 807         if (!transno) {
 808                 INIT_LIST_HEAD(&req->rq_list);
 809                 DEBUG_REQ(D_HA, req, "not queueing");
 810                 return 1;
 811         }
 812
 813         /* XXX If I were a real man, these LBUGs would be sane cleanups. */
 814         /* XXX just like the request-dup code in queue_final_reply */
 815         OBD_ALLOC(saved_req, sizeof *saved_req);
 816         if (!saved_req)
 817                 LBUG();
 818         OBD_ALLOC(reqmsg, req->rq_reqlen);
 819         if (!reqmsg)
 820                 LBUG();
 821
 822         spin_lock_bh(&obd->obd_processing_task_lock);
 823
 824         /* If we're processing the queue, we want don't want to queue this
 825          * message.
 826          *
 827          * Also, if this request has a transno less than the one we're waiting
 828          * for, we should process it now.  It could (and currently always will)
 829          * be an open request for a descriptor that was opened some time ago.
 830          */
 831         if (obd->obd_processing_task == current->pid ||
 832             transno < obd->obd_next_recovery_transno) {
 833                 /* Processing the queue right now, don't re-add. */
 834                 LASSERT(list_empty(&req->rq_list));
 835                 spin_unlock_bh(&obd->obd_processing_task_lock);
 836                 OBD_FREE(reqmsg, req->rq_reqlen);
 837                 OBD_FREE(saved_req, sizeof *saved_req);
 838                 return 1;
 839         }
 840
 841         memcpy(saved_req, req, sizeof *req);
 842         memcpy(reqmsg, req->rq_reqmsg, req->rq_reqlen);
 843         req = saved_req;
 844         req->rq_reqmsg = reqmsg;
 845         class_export_get(req->rq_export);
 846         INIT_LIST_HEAD(&req->rq_list);
 847
 848         /* XXX O(n^2) */
 849         list_for_each(tmp, &obd->obd_recovery_queue) {
 850                 struct ptlrpc_request *reqiter =
 851                         list_entry(tmp, struct ptlrpc_request, rq_list);
 852
 853                 if (reqiter->rq_reqmsg->transno > transno) {
 854                         list_add_tail(&req->rq_list, &reqiter->rq_list);
 855                         inserted = 1;
 856                         break;
 857                 }
 858         }
 859
 860         if (!inserted) {
 861                 list_add_tail(&req->rq_list, &obd->obd_recovery_queue);
 862         }
 863
 864         obd->obd_requests_queued_for_recovery++;
 865
 866         if (obd->obd_processing_task != 0) {
 867                 /* Someone else is processing this queue, we'll leave it to
 868                  * them.
 869                  */
 870                 wake_up(&obd->obd_next_transno_waitq);
 871                 spin_unlock_bh(&obd->obd_processing_task_lock);
 872                 return 0;
 873         }
 874
 875         /* Nobody is processing, and we know there's (at least) one to process
 876          * now, so we'll do the honours.
 877          */
 878         obd->obd_processing_task = current->pid;
 879         spin_unlock_bh(&obd->obd_processing_task_lock);
 880
 881         process_recovery_queue(obd);
 882         return 0;
 883 }
 884
 885 struct obd_device * target_req2obd(struct ptlrpc_request *req)
 886 {
 887         return req->rq_export->exp_obd;
 888 }
 889
 890 int target_queue_final_reply(struct ptlrpc_request *req, int rc)
 891 {
 892         struct obd_device *obd = target_req2obd(req);
 893         struct ptlrpc_request *saved_req;
 894         struct lustre_msg *reqmsg;
 895         int recovery_done = 0;
 896         int rc2;
 897
 898         if (rc) {
 899                 /* Just like ptlrpc_error, but without the sending. */
 900                 rc = lustre_pack_reply(req, 0, NULL, NULL);
 901                 LASSERT(rc == 0); /* XXX handle this */
 902                 req->rq_type = PTL_RPC_MSG_ERR;
 903         }
 904
 905         LASSERT(list_empty(&req->rq_list));
 906         /* XXX a bit like the request-dup code in queue_recovery_request */
 907         OBD_ALLOC(saved_req, sizeof *saved_req);
 908         if (!saved_req)
 909                 LBUG();
 910         OBD_ALLOC(reqmsg, req->rq_reqlen);
 911         if (!reqmsg)
 912                 LBUG();
 913         memcpy(saved_req, req, sizeof *saved_req);
 914         memcpy(reqmsg, req->rq_reqmsg, req->rq_reqlen);
 915         req = saved_req;
 916         req->rq_reqmsg = reqmsg;
 917         list_add(&req->rq_list, &obd->obd_delayed_reply_queue);
 918
 919         spin_lock_bh(&obd->obd_processing_task_lock);
 920         --obd->obd_recoverable_clients;
 921         recovery_done = (obd->obd_recoverable_clients == 0);
 922         spin_unlock_bh(&obd->obd_processing_task_lock);
 923
 924         if (recovery_done) {
 925                 struct list_head *tmp, *n;
 926                 ldlm_reprocess_all_ns(req->rq_export->exp_obd->obd_namespace);
 927                 CWARN("%s: all clients recovered, sending delayed replies\n",
 928                        obd->obd_name);
 929                 obd->obd_recovering = 0;
 930
 931                 /* when recovering finished, cleanup orphans for mds */
 932                 if (OBT(obd) && OBP(obd, postrecov)) {
 933                         rc2 = OBP(obd, postrecov)(obd);
 934                         if (rc2 >= 0)
 935                                 CWARN("%s: all clients recovered, %d MDS orphans "
 936                                        "deleted\n", obd->obd_name, rc2);
 937                         else
 938                                 CERROR("postrecov failed %d\n", rc2);
 939                 }
 940
 941                 list_for_each_safe(tmp, n, &obd->obd_delayed_reply_queue) {
 942                         req = list_entry(tmp, struct ptlrpc_request, rq_list);
 943                         DEBUG_REQ(D_ERROR, req, "delayed:");
 944                         ptlrpc_reply(req);
 945                         list_del(&req->rq_list);
 946                         OBD_FREE(req->rq_reqmsg, req->rq_reqlen);
 947                         OBD_FREE(req, sizeof *req);
 948                 }
 949                 target_cancel_recovery_timer(obd);
 950         } else {
 951                 CERROR("%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                 memcpy(&req->rq_ack_locks[i].lock, lock, sizeof(*lock));
1127                 req->rq_ack_locks[i].mode = mode;
1128                 return &req->rq_ack_locks[i];
1129         }
1130         CERROR("no space for lock in struct ptlrpc_request\n");
1131         LBUG();
1132         return NULL;
1133 }
1134