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;
  50         struct obd_device *mgmt_obd;
  51         mgmtcli_register_for_events_t register_f;
  52         ENTRY;
  53
  54         /* In a more perfect world, we would hang a ptlrpc_client off of
  55          * obd_type and just use the values from there. */
  56         if (!strcmp(name, LUSTRE_OSC_NAME)) {
  57                 rq_portal = OST_REQUEST_PORTAL;
  58                 rp_portal = OSC_REPLY_PORTAL;
  59                 connect_op = OST_CONNECT;
  60         } else if (!strcmp(name, LUSTRE_MDC_NAME)) {
  61                 rq_portal = MDS_REQUEST_PORTAL;
  62                 rp_portal = MDC_REPLY_PORTAL;
  63                 connect_op = MDS_CONNECT;
  64         } else if (!strcmp(name, LUSTRE_MGMTCLI_NAME)) {
  65                 rq_portal = MGMT_REQUEST_PORTAL;
  66                 rp_portal = MGMT_REPLY_PORTAL;
  67                 connect_op = MGMT_CONNECT;
  68         } else {
  69                 CERROR("unknown client OBD type \"%s\", can't setup\n",
  70                        name);
  71                 RETURN(-EINVAL);
  72         }
  73
  74         if (lcfg->lcfg_inllen1 < 1) {
  75                 CERROR("requires a TARGET UUID\n");
  76                 RETURN(-EINVAL);
  77         }
  78
  79         if (lcfg->lcfg_inllen1 > 37) {
  80                 CERROR("client UUID must be less than 38 characters\n");
  81                 RETURN(-EINVAL);
  82         }
  83
  84         if (lcfg->lcfg_inllen2 < 1) {
  85                 CERROR("setup requires a SERVER UUID\n");
  86                 RETURN(-EINVAL);
  87         }
  88
  89         if (lcfg->lcfg_inllen2 > 37) {
  90                 CERROR("target UUID must be less than 38 characters\n");
  91                 RETURN(-EINVAL);
  92         }
  93
  94         sema_init(&cli->cl_sem, 1);
  95         cli->cl_conn_count = 0;
  96         memcpy(server_uuid.uuid, lcfg->lcfg_inlbuf2,
  97                min_t(unsigned int, lcfg->lcfg_inllen2, sizeof(server_uuid)));
  98
  99         cli->cl_dirty = 0;
 100         cli->cl_avail_grant = 0;
 101         /* FIXME: should limit this for the sum of all cl_dirty_max */
 102         cli->cl_dirty_max = OSC_MAX_DIRTY_DEFAULT * 1024 * 1024;
 103         if (cli->cl_dirty_max >> PAGE_SHIFT > num_physpages / 8)
 104                 cli->cl_dirty_max = num_physpages << (PAGE_SHIFT - 3);
 105         INIT_LIST_HEAD(&cli->cl_cache_waiters);
 106         INIT_LIST_HEAD(&cli->cl_loi_ready_list);
 107         INIT_LIST_HEAD(&cli->cl_loi_write_list);
 108         INIT_LIST_HEAD(&cli->cl_loi_read_list);
 109         spin_lock_init(&cli->cl_loi_list_lock);
 110         cli->cl_r_in_flight = 0;
 111         cli->cl_w_in_flight = 0;
 112         spin_lock_init(&cli->cl_read_rpc_hist.oh_lock);
 113         spin_lock_init(&cli->cl_write_rpc_hist.oh_lock);
 114         spin_lock_init(&cli->cl_read_page_hist.oh_lock);
 115         spin_lock_init(&cli->cl_write_page_hist.oh_lock);
 116         cli->cl_max_pages_per_rpc = PTLRPC_MAX_BRW_PAGES;
 117         cli->cl_max_rpcs_in_flight = OSC_MAX_RIF_DEFAULT;
 118
 119         rc = ldlm_get_ref();
 120         if (rc) {
 121                 CERROR("ldlm_get_ref failed: %d\n", rc);
 122                 GOTO(err, rc);
 123         }
 124
 125         conn = ptlrpc_uuid_to_connection(&server_uuid);
 126         if (conn == NULL)
 127                 GOTO(err_ldlm, rc = -ENOENT);
 128
 129         ptlrpc_init_client(rq_portal, rp_portal, name,
 130                            &obddev->obd_ldlm_client);
 131
 132         imp = class_new_import();
 133         if (imp == NULL) {
 134                 ptlrpc_put_connection(conn);
 135                 GOTO(err_ldlm, rc = -ENOENT);
 136         }
 137         imp->imp_connection = conn;
 138         imp->imp_client = &obddev->obd_ldlm_client;
 139         imp->imp_obd = obddev;
 140         imp->imp_connect_op = connect_op;
 141         imp->imp_generation = 0;
 142         imp->imp_initial_recov = 1;
 143         INIT_LIST_HEAD(&imp->imp_pinger_chain);
 144         memcpy(imp->imp_target_uuid.uuid, lcfg->lcfg_inlbuf1,
 145               lcfg->lcfg_inllen1);
 146         class_import_put(imp);
 147
 148         cli->cl_import = imp;
 149         cli->cl_max_mds_easize = sizeof(struct lov_mds_md);
 150         cli->cl_max_mds_cookiesize = sizeof(struct llog_cookie);
 151         cli->cl_sandev = to_kdev_t(0);
 152
 153         if (lcfg->lcfg_inllen3 != 0) {
 154                 if (!strcmp(lcfg->lcfg_inlbuf3, "inactive")) {
 155                         CDEBUG(D_HA, "marking %s %s->%s as inactive\n",
 156                                name, obddev->obd_name,
 157                                imp->imp_target_uuid.uuid);
 158                         imp->imp_invalid = 1;
 159
 160                         if (lcfg->lcfg_inllen4 != 0)
 161                                 mgmt_name = lcfg->lcfg_inlbuf4;
 162                 } else {
 163                         mgmt_name = lcfg->lcfg_inlbuf3;
 164                 }
 165         }
 166
 167         if (mgmt_name != NULL) {
 168                 /* Register with management client if we need to. */
 169                 CDEBUG(D_HA, "%s registering with %s for events about %s\n",
 170                        obddev->obd_name, mgmt_name, server_uuid.uuid);
 171
 172                 mgmt_obd = class_name2obd(mgmt_name);
 173                 if (!mgmt_obd) {
 174                         CERROR("can't find mgmtcli %s to register\n",
 175                                mgmt_name);
 176                         GOTO(err_import, rc = -ENOSYS);
 177                 }
 178
 179                 register_f = inter_module_get("mgmtcli_register_for_events");
 180                 if (!register_f) {
 181                         CERROR("can't i_m_g mgmtcli_register_for_events\n");
 182                         GOTO(err_import, rc = -ENOSYS);
 183                 }
 184
 185                 rc = register_f(mgmt_obd, obddev, &imp->imp_target_uuid);
 186                 inter_module_put("mgmtcli_register_for_events");
 187
 188                 if (!rc)
 189                         cli->cl_mgmtcli_obd = mgmt_obd;
 190         }
 191
 192         RETURN(rc);
 193
 194 err_import:
 195         class_destroy_import(imp);
 196 err_ldlm:
 197         ldlm_put_ref(0);
 198 err:
 199         RETURN(rc);
 200
 201 }
 202
 203 int client_obd_cleanup(struct obd_device *obddev, int flags)
 204 {
 205         struct client_obd *cli = &obddev->u.cli;
 206
 207         if (!cli->cl_import)
 208                 RETURN(-EINVAL);
 209         if (cli->cl_mgmtcli_obd) {
 210                 mgmtcli_deregister_for_events_t dereg_f;
 211
 212                 dereg_f = inter_module_get("mgmtcli_deregister_for_events");
 213                 dereg_f(cli->cl_mgmtcli_obd, obddev);
 214                 inter_module_put("mgmtcli_deregister_for_events");
 215         }
 216         class_destroy_import(cli->cl_import);
 217         cli->cl_import = NULL;
 218
 219         ldlm_put_ref(flags & OBD_OPT_FORCE);
 220
 221         RETURN(0);
 222 }
 223
 224 int client_connect_import(struct lustre_handle *dlm_handle,
 225                           struct obd_device *obd,
 226                           struct obd_uuid *cluuid)
 227 {
 228         struct client_obd *cli = &obd->u.cli;
 229         struct obd_import *imp = cli->cl_import;
 230         struct obd_export *exp;
 231         int rc;
 232         ENTRY;
 233
 234         down(&cli->cl_sem);
 235         rc = class_connect(dlm_handle, obd, cluuid);
 236         if (rc)
 237                 GOTO(out_sem, rc);
 238
 239         cli->cl_conn_count++;
 240         if (cli->cl_conn_count > 1)
 241                 GOTO(out_sem, rc);
 242         exp = class_conn2export(dlm_handle);
 243
 244         if (obd->obd_namespace != NULL)
 245                 CERROR("already have namespace!\n");
 246         obd->obd_namespace = ldlm_namespace_new(obd->obd_name,
 247                                                 LDLM_NAMESPACE_CLIENT);
 248         if (obd->obd_namespace == NULL)
 249                 GOTO(out_disco, rc = -ENOMEM);
 250
 251         imp->imp_dlm_handle = *dlm_handle;
 252         rc = ptlrpc_init_import(imp);
 253         if (rc != 0)
 254                 GOTO(out_ldlm, rc);
 255
 256         exp->exp_connection = ptlrpc_connection_addref(imp->imp_connection);
 257         rc = ptlrpc_connect_import(imp, NULL);
 258         if (rc != 0) {
 259                 LASSERT (imp->imp_state == LUSTRE_IMP_DISCON);
 260                 GOTO(out_ldlm, rc);
 261         }
 262
 263         ptlrpc_pinger_add_import(imp);
 264         EXIT;
 265
 266         if (rc) {
 267 out_ldlm:
 268                 ldlm_namespace_free(obd->obd_namespace, 0);
 269                 obd->obd_namespace = NULL;
 270 out_disco:
 271                 cli->cl_conn_count--;
 272                 class_disconnect(exp, 0);
 273         } else {
 274                 class_export_put(exp);
 275         }
 276 out_sem:
 277         up(&cli->cl_sem);
 278         return rc;
 279 }
 280
 281 int client_disconnect_export(struct obd_export *exp, int failover)
 282 {
 283         struct obd_device *obd = class_exp2obd(exp);
 284         struct client_obd *cli = &obd->u.cli;
 285         struct obd_import *imp = cli->cl_import;
 286         int rc = 0, err;
 287         ENTRY;
 288
 289         if (!obd) {
 290                 CERROR("invalid export for disconnect: exp %p cookie "LPX64"\n",
 291                        exp, exp ? exp->exp_handle.h_cookie : -1);
 292                 RETURN(-EINVAL);
 293         }
 294
 295         down(&cli->cl_sem);
 296         if (!cli->cl_conn_count) {
 297                 CERROR("disconnecting disconnected device (%s)\n",
 298                        obd->obd_name);
 299                 GOTO(out_sem, rc = -EINVAL);
 300         }
 301
 302         cli->cl_conn_count--;
 303         if (cli->cl_conn_count)
 304                 GOTO(out_no_disconnect, rc = 0);
 305
 306         /* Some non-replayable imports (MDS's OSCs) are pinged, so just
 307          * delete it regardless.  (It's safe to delete an import that was
 308          * never added.) */
 309         (void)ptlrpc_pinger_del_import(imp);
 310
 311         if (obd->obd_namespace != NULL) {
 312                 /* obd_no_recov == local only */
 313                 ldlm_cli_cancel_unused(obd->obd_namespace, NULL,
 314                                        obd->obd_no_recov, NULL);
 315                 ldlm_namespace_free(obd->obd_namespace, obd->obd_no_recov);
 316                 obd->obd_namespace = NULL;
 317         }
 318
 319         /* Yeah, obd_no_recov also (mainly) means "forced shutdown". */
 320         if (obd->obd_no_recov)
 321                 ptlrpc_invalidate_import(imp, 0);
 322         else
 323                 rc = ptlrpc_disconnect_import(imp);
 324
 325         EXIT;
 326  out_no_disconnect:
 327         err = class_disconnect(exp, 0);
 328         if (!rc && err)
 329                 rc = err;
 330  out_sem:
 331         up(&cli->cl_sem);
 332         RETURN(rc);
 333 }
 334
 335 /* --------------------------------------------------------------------------
 336  * from old lib/target.c
 337  * -------------------------------------------------------------------------- */
 338
 339 int target_handle_reconnect(struct lustre_handle *conn, struct obd_export *exp,
 340                             struct obd_uuid *cluuid)
 341 {
 342         if (exp->exp_connection) {
 343                 struct lustre_handle *hdl;
 344                 hdl = &exp->exp_imp_reverse->imp_remote_handle;
 345                 /* Might be a re-connect after a partition. */
 346                 if (!memcmp(&conn->cookie, &hdl->cookie, sizeof conn->cookie)) {
 347                         CERROR("%s reconnecting\n", cluuid->uuid);
 348                         conn->cookie = exp->exp_handle.h_cookie;
 349                         RETURN(EALREADY);
 350                 } else {
 351                         CERROR("%s reconnecting from %s, "
 352                                "handle mismatch (ours "LPX64", theirs "
 353                                LPX64")\n", cluuid->uuid,
 354                                exp->exp_connection->c_remote_uuid.uuid,
 355                                hdl->cookie, conn->cookie);
 356                         memset(conn, 0, sizeof *conn);
 357                         RETURN(-EALREADY);
 358                 }
 359         }
 360
 361         conn->cookie = exp->exp_handle.h_cookie;
 362         CDEBUG(D_INFO, "existing export for UUID '%s' at %p\n",
 363                cluuid->uuid, exp);
 364         CDEBUG(D_IOCTL,"connect: cookie "LPX64"\n", conn->cookie);
 365         RETURN(0);
 366 }
 367
 368 int target_handle_connect(struct ptlrpc_request *req, svc_handler_t handler)
 369 {
 370         struct obd_device *target;
 371         struct obd_export *export = NULL;
 372         struct obd_import *revimp;
 373         struct lustre_handle conn;
 374         struct obd_uuid tgtuuid;
 375         struct obd_uuid cluuid;
 376         struct obd_uuid remote_uuid;
 377         struct list_head *p;
 378         char *str, *tmp;
 379         int rc = 0, abort_recovery;
 380         unsigned long flags;
 381         ENTRY;
 382
 383         OBD_RACE(OBD_FAIL_TGT_CONN_RACE);
 384
 385         LASSERT_REQSWAB (req, 0);
 386         str = lustre_msg_string(req->rq_reqmsg, 0, sizeof(tgtuuid) - 1);
 387         if (str == NULL) {
 388                 CERROR("bad target UUID for connect\n");
 389                 GOTO(out, rc = -EINVAL);
 390         }
 391
 392         obd_str2uuid (&tgtuuid, str);
 393         target = class_uuid2obd(&tgtuuid);
 394         if (!target) {
 395                 target = class_name2obd(str);
 396         }
 397
 398         if (!target || target->obd_stopping || !target->obd_set_up) {
 399                 CERROR("UUID '%s' is not available for connect\n", str);
 400                 GOTO(out, rc = -ENODEV);
 401         }
 402
 403         LASSERT_REQSWAB (req, 1);
 404         str = lustre_msg_string(req->rq_reqmsg, 1, sizeof(cluuid) - 1);
 405         if (str == NULL) {
 406                 CERROR("bad client UUID for connect\n");
 407                 GOTO(out, rc = -EINVAL);
 408         }
 409
 410         obd_str2uuid (&cluuid, str);
 411
 412         /* XXX extract a nettype and format accordingly */
 413         snprintf(remote_uuid.uuid, sizeof remote_uuid,
 414                  "NET_"LPX64"_UUID", req->rq_peer.peer_nid);
 415
 416         spin_lock_bh(&target->obd_processing_task_lock);
 417         abort_recovery = target->obd_abort_recovery;
 418         spin_unlock_bh(&target->obd_processing_task_lock);
 419         if (abort_recovery)
 420                 target_abort_recovery(target);
 421
 422         tmp = lustre_msg_buf(req->rq_reqmsg, 2, sizeof conn);
 423         if (tmp == NULL)
 424                 GOTO(out, rc = -EPROTO);
 425
 426         memcpy(&conn, tmp, sizeof conn);
 427
 428         rc = lustre_pack_reply(req, 0, NULL, NULL);
 429         if (rc)
 430                 GOTO(out, rc);
 431
 432         /* lctl gets a backstage, all-access pass. */
 433         if (obd_uuid_equals(&cluuid, &target->obd_uuid))
 434                 goto dont_check_exports;
 435
 436         spin_lock(&target->obd_dev_lock);
 437         list_for_each(p, &target->obd_exports) {
 438                 export = list_entry(p, struct obd_export, exp_obd_chain);
 439                 if (obd_uuid_equals(&cluuid, &export->exp_client_uuid)) {
 440                         spin_unlock(&target->obd_dev_lock);
 441                         LASSERT(export->exp_obd == target);
 442
 443                         rc = target_handle_reconnect(&conn, export, &cluuid);
 444                         break;
 445                 }
 446                 export = NULL;
 447         }
 448         /* If we found an export, we already unlocked. */
 449         if (!export) {
 450                 spin_unlock(&target->obd_dev_lock);
 451         } else if (req->rq_reqmsg->conn_cnt == 1) {
 452                 CERROR("%s reconnected with 1 conn_cnt; cookies not random?\n",
 453                        cluuid.uuid);
 454                 GOTO(out, rc = -EALREADY);
 455         }
 456
 457         /* Tell the client if we're in recovery. */
 458         /* If this is the first client, start the recovery timer */
 459         if (target->obd_recovering) {
 460                 lustre_msg_add_op_flags(req->rq_repmsg, MSG_CONNECT_RECOVERING);
 461                 target_start_recovery_timer(target, handler);
 462         }
 463
 464         /* Tell the client if we support replayable requests */
 465         if (target->obd_replayable)
 466                 lustre_msg_add_op_flags(req->rq_repmsg, MSG_CONNECT_REPLAYABLE);
 467
 468         if (export == NULL) {
 469                 if (target->obd_recovering) {
 470                         CERROR("denying connection for new client %s: "
 471                                "%d clients in recovery for %lds\n", cluuid.uuid,
 472                                target->obd_recoverable_clients,
 473                                (target->obd_recovery_timer.expires-jiffies)/HZ);
 474                         rc = -EBUSY;
 475                 } else {
 476  dont_check_exports:
 477                         rc = obd_connect(&conn, target, &cluuid);
 478                 }
 479         }
 480
 481
 482         /* If all else goes well, this is our RPC return code. */
 483         req->rq_status = 0;
 484
 485         if (rc && rc != EALREADY)
 486                 GOTO(out, rc);
 487
 488         req->rq_repmsg->handle = conn;
 489
 490         /* If the client and the server are the same node, we will already
 491          * have an export that really points to the client's DLM export,
 492          * because we have a shared handles table.
 493          *
 494          * XXX this will go away when shaver stops sending the "connect" handle
 495          * in the real "remote handle" field of the request --phik 24 Apr 2003
 496          */
 497         if (req->rq_export != NULL)
 498                 class_export_put(req->rq_export);
 499
 500         /* ownership of this export ref transfers to the request */
 501         export = req->rq_export = class_conn2export(&conn);
 502         LASSERT(export != NULL);
 503
 504         spin_lock_irqsave(&export->exp_lock, flags);
 505         if (export->exp_conn_cnt >= req->rq_reqmsg->conn_cnt) {
 506                 CERROR("%s: already connected at a higher conn_cnt: %d > %d\n",
 507                        cluuid.uuid, export->exp_conn_cnt,
 508                        req->rq_reqmsg->conn_cnt);
 509                 spin_unlock_irqrestore(&export->exp_lock, flags);
 510                 GOTO(out, rc = -EALREADY);
 511         }
 512         export->exp_conn_cnt = req->rq_reqmsg->conn_cnt;
 513         spin_unlock_irqrestore(&export->exp_lock, flags);
 514
 515         /* request from liblustre? */
 516         if (lustre_msg_get_op_flags(req->rq_reqmsg) & MSG_CONNECT_LIBCLIENT)
 517                 export->exp_libclient = 1;
 518
 519         if (export->exp_connection != NULL)
 520                 ptlrpc_put_connection(export->exp_connection);
 521         export->exp_connection = ptlrpc_get_connection(&req->rq_peer,
 522                                                        &remote_uuid);
 523
 524         if (rc == EALREADY) {
 525                 /* We indicate the reconnection in a flag, not an error code. */
 526                 lustre_msg_add_op_flags(req->rq_repmsg, MSG_CONNECT_RECONNECT);
 527                 GOTO(out, rc = 0);
 528         }
 529
 530         if (target->obd_recovering)
 531                 target->obd_connected_clients++;
 532
 533         memcpy(&conn, lustre_msg_buf(req->rq_reqmsg, 2, sizeof conn),
 534                sizeof conn);
 535
 536         if (export->exp_imp_reverse != NULL)
 537                 class_destroy_import(export->exp_imp_reverse);
 538         revimp = export->exp_imp_reverse = class_new_import();
 539         revimp->imp_connection = ptlrpc_connection_addref(export->exp_connection);
 540         revimp->imp_client = &export->exp_obd->obd_ldlm_client;
 541         revimp->imp_remote_handle = conn;
 542         revimp->imp_obd = target;
 543         revimp->imp_dlm_fake = 1;
 544         revimp->imp_state = LUSTRE_IMP_FULL;
 545         class_import_put(revimp);
 546 out:
 547         if (rc)
 548                 req->rq_status = rc;
 549         RETURN(rc);
 550 }
 551
 552 int target_handle_disconnect(struct ptlrpc_request *req)
 553 {
 554         struct obd_export *exp;
 555         int rc;
 556         ENTRY;
 557
 558         rc = lustre_pack_reply(req, 0, NULL, NULL);
 559         if (rc)
 560                 RETURN(rc);
 561
 562         /* keep the rq_export around so we can send the reply */
 563         exp = class_export_get(req->rq_export);
 564         req->rq_status = obd_disconnect(exp, 0);
 565         RETURN(0);
 566 }
 567
 568 void target_destroy_export(struct obd_export *exp)
 569 {
 570         /* exports created from last_rcvd data, and "fake"
 571            exports created by lctl don't have an import */
 572         if (exp->exp_imp_reverse != NULL)
 573                 class_destroy_import(exp->exp_imp_reverse);
 574
 575         /* We cancel locks at disconnect time, but this will catch any locks
 576          * granted in a race with recovery-induced disconnect. */
 577         if (exp->exp_obd->obd_namespace != NULL)
 578                 ldlm_cancel_locks_for_export(exp);
 579 }
 580
 581 /*
 582  * Recovery functions
 583  */
 584
 585
 586 static void target_release_saved_req(struct ptlrpc_request *req)
 587 {
 588         class_export_put(req->rq_export);
 589         OBD_FREE(req->rq_reqmsg, req->rq_reqlen);
 590         OBD_FREE(req, sizeof *req);
 591 }
 592
 593 static void target_finish_recovery(struct obd_device *obd)
 594 {
 595         struct list_head *tmp, *n;
 596         int rc;
 597
 598         CWARN("%s: sending delayed replies to recovered clients\n",
 599               obd->obd_name);
 600
 601         ldlm_reprocess_all_ns(obd->obd_namespace);
 602
 603         /* when recovery finished, cleanup orphans on mds and ost */
 604         if (OBT(obd) && OBP(obd, postrecov)) {
 605                 rc = OBP(obd, postrecov)(obd);
 606                 if (rc >= 0)
 607                         CWARN("%s: all clients recovered, %d MDS "
 608                               "orphans deleted\n", obd->obd_name, rc);
 609                 else
 610                         CERROR("postrecov failed %d\n", rc);
 611         }
 612
 613         list_for_each_safe(tmp, n, &obd->obd_delayed_reply_queue) {
 614                 struct ptlrpc_request *req;
 615                 req = list_entry(tmp, struct ptlrpc_request, rq_list);
 616                 list_del(&req->rq_list);
 617                 DEBUG_REQ(D_ERROR, req, "delayed:");
 618                 ptlrpc_reply(req);
 619                 target_release_saved_req(req);
 620         }
 621         obd->obd_recovery_end = LTIME_S(CURRENT_TIME);
 622         return;
 623 }
 624
 625 static void abort_recovery_queue(struct obd_device *obd)
 626 {
 627         struct ptlrpc_request *req;
 628         struct list_head *tmp, *n;
 629         int rc;
 630
 631         list_for_each_safe(tmp, n, &obd->obd_recovery_queue) {
 632                 req = list_entry(tmp, struct ptlrpc_request, rq_list);
 633                 list_del(&req->rq_list);
 634                 DEBUG_REQ(D_ERROR, req, "aborted:");
 635                 req->rq_status = -ENOTCONN;
 636                 req->rq_type = PTL_RPC_MSG_ERR;
 637                 rc = lustre_pack_reply(req, 0, NULL, NULL);
 638                 if (rc == 0) {
 639                         ptlrpc_reply(req);
 640                 } else {
 641                         DEBUG_REQ(D_ERROR, req,
 642                                   "packing failed for abort-reply; skipping");
 643                 }
 644                 target_release_saved_req(req);
 645         }
 646 }
 647
 648 /* Called from a cleanup function if the device is being cleaned up
 649    forcefully.  The exports should all have been disconnected already,
 650    the only thing left to do is
 651      - clear the recovery flags
 652      - cancel the timer
 653      - free queued requests and replies, but don't send replies
 654    Because the obd_stopping flag is set, no new requests should be received.
 655
 656 */
 657 void target_cleanup_recovery(struct obd_device *obd)
 658 {
 659         struct list_head *tmp, *n;
 660         struct ptlrpc_request *req;
 661
 662         spin_lock_bh(&obd->obd_processing_task_lock);
 663         if (!obd->obd_recovering) {
 664                 spin_unlock_bh(&obd->obd_processing_task_lock);
 665                 EXIT;
 666                 return;
 667         }
 668         obd->obd_recovering = obd->obd_abort_recovery = 0;
 669         target_cancel_recovery_timer(obd);
 670         spin_unlock_bh(&obd->obd_processing_task_lock);
 671
 672
 673         list_for_each_safe(tmp, n, &obd->obd_delayed_reply_queue) {
 674                 req = list_entry(tmp, struct ptlrpc_request, rq_list);
 675                 list_del(&req->rq_list);
 676                 LASSERT (req->rq_reply_state);
 677                 lustre_free_reply_state(req->rq_reply_state);
 678                 target_release_saved_req(req);
 679         }
 680
 681         list_for_each_safe(tmp, n, &obd->obd_recovery_queue) {
 682                 req = list_entry(tmp, struct ptlrpc_request, rq_list);
 683                 list_del(&req->rq_list);
 684                 LASSERT (req->rq_reply_state == 0);
 685                 target_release_saved_req(req);
 686         }
 687 }
 688
 689 void target_abort_recovery(void *data)
 690 {
 691         struct obd_device *obd = data;
 692
 693         spin_lock_bh(&obd->obd_processing_task_lock);
 694         if (!obd->obd_recovering) {
 695                 spin_unlock_bh(&obd->obd_processing_task_lock);
 696                 EXIT;
 697                 return;
 698         }
 699         obd->obd_recovering = obd->obd_abort_recovery = 0;
 700         target_cancel_recovery_timer(obd);
 701         spin_unlock_bh(&obd->obd_processing_task_lock);
 702
 703         CERROR("%s: recovery period over; disconnecting unfinished clients.\n",
 704                obd->obd_name);
 705         class_disconnect_stale_exports(obd, 0);
 706         abort_recovery_queue(obd);
 707
 708         target_finish_recovery(obd);
 709
 710         ptlrpc_run_recovery_over_upcall(obd);
 711 }
 712
 713 static void target_recovery_expired(unsigned long castmeharder)
 714 {
 715         struct obd_device *obd = (struct obd_device *)castmeharder;
 716         CERROR("recovery timed out, aborting\n");
 717         spin_lock_bh(&obd->obd_processing_task_lock);
 718         if (obd->obd_recovering)
 719                 obd->obd_abort_recovery = 1;
 720         wake_up(&obd->obd_next_transno_waitq);
 721         spin_unlock_bh(&obd->obd_processing_task_lock);
 722 }
 723
 724
 725 /* obd_processing_task_lock should be held */
 726 void target_cancel_recovery_timer(struct obd_device *obd)
 727 {
 728         CDEBUG(D_HA, "%s: cancel recovery timer\n", obd->obd_name);
 729         del_timer(&obd->obd_recovery_timer);
 730 }
 731
 732 static void reset_recovery_timer(struct obd_device *obd)
 733 {
 734         spin_lock_bh(&obd->obd_processing_task_lock);
 735         if (!obd->obd_recovering) {
 736                 spin_unlock_bh(&obd->obd_processing_task_lock);
 737                 return;
 738         }
 739         CDEBUG(D_HA, "timer will expire in %u seconds\n",
 740                OBD_RECOVERY_TIMEOUT / HZ);
 741         mod_timer(&obd->obd_recovery_timer, jiffies + OBD_RECOVERY_TIMEOUT);
 742         spin_unlock_bh(&obd->obd_processing_task_lock);
 743 }
 744
 745
 746 /* Only start it the first time called */
 747 void target_start_recovery_timer(struct obd_device *obd, svc_handler_t handler)
 748 {
 749         spin_lock_bh(&obd->obd_processing_task_lock);
 750         if (obd->obd_recovery_handler) {
 751                 spin_unlock_bh(&obd->obd_processing_task_lock);
 752                 return;
 753         }
 754         CWARN("%s: starting recovery timer (%us)\n", obd->obd_name,
 755                OBD_RECOVERY_TIMEOUT / HZ);
 756         obd->obd_recovery_handler = handler;
 757         obd->obd_recovery_timer.function = target_recovery_expired;
 758         obd->obd_recovery_timer.data = (unsigned long)obd;
 759         spin_unlock_bh(&obd->obd_processing_task_lock);
 760
 761         reset_recovery_timer(obd);
 762 }
 763
 764 static int check_for_next_transno(struct obd_device *obd)
 765 {
 766         struct ptlrpc_request *req;
 767         int wake_up = 0, connected, completed, queue_len, max;
 768         __u64 next_transno, req_transno;
 769
 770         spin_lock_bh(&obd->obd_processing_task_lock);
 771         req = list_entry(obd->obd_recovery_queue.next,
 772                          struct ptlrpc_request, rq_list);
 773         max = obd->obd_max_recoverable_clients;
 774         req_transno = req->rq_reqmsg->transno;
 775         connected = obd->obd_connected_clients;
 776         completed = max - obd->obd_recoverable_clients;
 777         queue_len = obd->obd_requests_queued_for_recovery;
 778         next_transno = obd->obd_next_recovery_transno;
 779
 780         CDEBUG(D_HA,"max: %d, connected: %d, completed: %d, queue_len: %d, "
 781                "req_transno: "LPU64", next_transno: "LPU64"\n",
 782                max, connected, completed, queue_len, req_transno, next_transno);
 783         if (obd->obd_abort_recovery) {
 784                 CDEBUG(D_HA, "waking for aborted recovery\n");
 785                 wake_up = 1;
 786         } else if (!obd->obd_recovering) {
 787                 CDEBUG(D_HA, "waking for completed recovery (?)\n");
 788                 wake_up = 1;
 789         } else if (req_transno == next_transno) {
 790                 CDEBUG(D_HA, "waking for next ("LPD64")\n", next_transno);
 791                 wake_up = 1;
 792         } else if (queue_len + completed == max) {
 793                 CDEBUG(D_ERROR,
 794                        "waking for skipped transno (skip: "LPD64
 795                        ", ql: %d, comp: %d, conn: %d, next: "LPD64")\n",
 796                        next_transno, queue_len, completed, max, req_transno);
 797                 obd->obd_next_recovery_transno = req_transno;
 798                 wake_up = 1;
 799         }
 800         spin_unlock_bh(&obd->obd_processing_task_lock);
 801         LASSERT(req->rq_reqmsg->transno >= next_transno);
 802         return wake_up;
 803 }
 804
 805 static void process_recovery_queue(struct obd_device *obd)
 806 {
 807         struct ptlrpc_request *req;
 808         int abort_recovery = 0;
 809         struct l_wait_info lwi = { 0 };
 810         ENTRY;
 811
 812         for (;;) {
 813                 spin_lock_bh(&obd->obd_processing_task_lock);
 814                 LASSERT(obd->obd_processing_task == current->pid);
 815                 req = list_entry(obd->obd_recovery_queue.next,
 816                                  struct ptlrpc_request, rq_list);
 817
 818                 if (req->rq_reqmsg->transno != obd->obd_next_recovery_transno) {
 819                         spin_unlock_bh(&obd->obd_processing_task_lock);
 820                         CDEBUG(D_HA, "Waiting for transno "LPD64" (1st is "
 821                                LPD64")\n",
 822                                obd->obd_next_recovery_transno,
 823                                req->rq_reqmsg->transno);
 824                         l_wait_event(obd->obd_next_transno_waitq,
 825                                      check_for_next_transno(obd), &lwi);
 826                         spin_lock_bh(&obd->obd_processing_task_lock);
 827                         abort_recovery = obd->obd_abort_recovery;
 828                         spin_unlock_bh(&obd->obd_processing_task_lock);
 829                         if (abort_recovery) {
 830                                 target_abort_recovery(obd);
 831                                 return;
 832                         }
 833                         continue;
 834                 }
 835                 list_del_init(&req->rq_list);
 836                 obd->obd_requests_queued_for_recovery--;
 837                 spin_unlock_bh(&obd->obd_processing_task_lock);
 838
 839                 DEBUG_REQ(D_HA, req, "processing: ");
 840                 (void)obd->obd_recovery_handler(req);
 841                 obd->obd_replayed_requests++;
 842                 reset_recovery_timer(obd);
 843                 /* bug 1580: decide how to properly sync() in recovery */
 844                 //mds_fsync_super(mds->mds_sb);
 845                 class_export_put(req->rq_export);
 846                 OBD_FREE(req->rq_reqmsg, req->rq_reqlen);
 847                 OBD_FREE(req, sizeof *req);
 848                 spin_lock_bh(&obd->obd_processing_task_lock);
 849                 obd->obd_next_recovery_transno++;
 850                 if (list_empty(&obd->obd_recovery_queue)) {
 851                         obd->obd_processing_task = 0;
 852                         spin_unlock_bh(&obd->obd_processing_task_lock);
 853                         break;
 854                 }
 855                 spin_unlock_bh(&obd->obd_processing_task_lock);
 856         }
 857         EXIT;
 858 }
 859
 860 int target_queue_recovery_request(struct ptlrpc_request *req,
 861                                   struct obd_device *obd)
 862 {
 863         struct list_head *tmp;
 864         int inserted = 0;
 865         __u64 transno = req->rq_reqmsg->transno;
 866         struct ptlrpc_request *saved_req;
 867         struct lustre_msg *reqmsg;
 868
 869         /* CAVEAT EMPTOR: The incoming request message has been swabbed
 870          * (i.e. buflens etc are in my own byte order), but type-dependent
 871          * buffers (eg mds_body, ost_body etc) have NOT been swabbed. */
 872
 873         if (!transno) {
 874                 INIT_LIST_HEAD(&req->rq_list);
 875                 DEBUG_REQ(D_HA, req, "not queueing");
 876                 return 1;
 877         }
 878
 879         /* XXX If I were a real man, these LBUGs would be sane cleanups. */
 880         /* XXX just like the request-dup code in queue_final_reply */
 881         OBD_ALLOC(saved_req, sizeof *saved_req);
 882         if (!saved_req)
 883                 LBUG();
 884         OBD_ALLOC(reqmsg, req->rq_reqlen);
 885         if (!reqmsg)
 886                 LBUG();
 887
 888         spin_lock_bh(&obd->obd_processing_task_lock);
 889
 890         /* If we're processing the queue, we want don't want to queue this
 891          * message.
 892          *
 893          * Also, if this request has a transno less than the one we're waiting
 894          * for, we should process it now.  It could (and currently always will)
 895          * be an open request for a descriptor that was opened some time ago.
 896          *
 897          * Also, a resent, replayed request that has already been
 898          * handled will pass through here and be processed immediately.
 899          */
 900         if (obd->obd_processing_task == current->pid ||
 901             transno < obd->obd_next_recovery_transno) {
 902                 /* Processing the queue right now, don't re-add. */
 903                 LASSERT(list_empty(&req->rq_list));
 904                 spin_unlock_bh(&obd->obd_processing_task_lock);
 905                 OBD_FREE(reqmsg, req->rq_reqlen);
 906                 OBD_FREE(saved_req, sizeof *saved_req);
 907                 return 1;
 908         }
 909
 910         /* A resent, replayed request that is still on the queue; just drop it.
 911            The queued request will handle this. */
 912         if ((lustre_msg_get_flags(req->rq_reqmsg) & (MSG_RESENT | MSG_REPLAY)) ==
 913             (MSG_RESENT | MSG_REPLAY)) {
 914                 DEBUG_REQ(D_ERROR, req, "dropping resent queued req");
 915                 spin_unlock_bh(&obd->obd_processing_task_lock);
 916                 OBD_FREE(reqmsg, req->rq_reqlen);
 917                 OBD_FREE(saved_req, sizeof *saved_req);
 918                 return 0;
 919         }
 920
 921         memcpy(saved_req, req, sizeof *req);
 922         memcpy(reqmsg, req->rq_reqmsg, req->rq_reqlen);
 923         req = saved_req;
 924         req->rq_reqmsg = reqmsg;
 925         class_export_get(req->rq_export);
 926         INIT_LIST_HEAD(&req->rq_list);
 927
 928         /* XXX O(n^2) */
 929         list_for_each(tmp, &obd->obd_recovery_queue) {
 930                 struct ptlrpc_request *reqiter =
 931                         list_entry(tmp, struct ptlrpc_request, rq_list);
 932
 933                 if (reqiter->rq_reqmsg->transno > transno) {
 934                         list_add_tail(&req->rq_list, &reqiter->rq_list);
 935                         inserted = 1;
 936                         break;
 937                 }
 938         }
 939
 940         if (!inserted) {
 941                 list_add_tail(&req->rq_list, &obd->obd_recovery_queue);
 942         }
 943
 944         obd->obd_requests_queued_for_recovery++;
 945
 946         if (obd->obd_processing_task != 0) {
 947                 /* Someone else is processing this queue, we'll leave it to
 948                  * them.
 949                  */
 950                 wake_up(&obd->obd_next_transno_waitq);
 951                 spin_unlock_bh(&obd->obd_processing_task_lock);
 952                 return 0;
 953         }
 954
 955         /* Nobody is processing, and we know there's (at least) one to process
 956          * now, so we'll do the honours.
 957          */
 958         obd->obd_processing_task = current->pid;
 959         spin_unlock_bh(&obd->obd_processing_task_lock);
 960
 961         process_recovery_queue(obd);
 962         return 0;
 963 }
 964
 965 struct obd_device * target_req2obd(struct ptlrpc_request *req)
 966 {
 967         return req->rq_export->exp_obd;
 968 }
 969
 970 int target_queue_final_reply(struct ptlrpc_request *req, int rc)
 971 {
 972         struct obd_device *obd = target_req2obd(req);
 973         struct ptlrpc_request *saved_req;
 974         struct lustre_msg *reqmsg;
 975         int recovery_done = 0;
 976
 977         LASSERT ((rc == 0) == (req->rq_reply_state != NULL));
 978
 979         if (rc) {
 980                 /* Just like ptlrpc_error, but without the sending. */
 981                 rc = lustre_pack_reply(req, 0, NULL, NULL);
 982                 LASSERT(rc == 0); /* XXX handle this */
 983                 req->rq_type = PTL_RPC_MSG_ERR;
 984         }
 985
 986         LASSERT (!req->rq_reply_state->rs_difficult);
 987         LASSERT(list_empty(&req->rq_list));
 988         /* XXX a bit like the request-dup code in queue_recovery_request */
 989         OBD_ALLOC(saved_req, sizeof *saved_req);
 990         if (!saved_req)
 991                 LBUG();
 992         OBD_ALLOC(reqmsg, req->rq_reqlen);
 993         if (!reqmsg)
 994                 LBUG();
 995         memcpy(saved_req, req, sizeof *saved_req);
 996         memcpy(reqmsg, req->rq_reqmsg, req->rq_reqlen);
 997         /* the copied req takes over the reply state */
 998         req->rq_reply_state = NULL;
 999         req = saved_req;
1000         req->rq_reqmsg = reqmsg;
1001         class_export_get(req->rq_export);
1002         list_add(&req->rq_list, &obd->obd_delayed_reply_queue);
1003
1004         spin_lock_bh(&obd->obd_processing_task_lock);
1005         /* only count the first "replay over" request from each
1006            export */
1007         if (req->rq_export->exp_replay_needed) {
1008                 --obd->obd_recoverable_clients;
1009                 req->rq_export->exp_replay_needed = 0;
1010         }
1011         recovery_done = (obd->obd_recoverable_clients == 0);
1012         spin_unlock_bh(&obd->obd_processing_task_lock);
1013
1014         if (recovery_done) {
1015                 spin_lock_bh(&obd->obd_processing_task_lock);
1016                 obd->obd_recovering = obd->obd_abort_recovery = 0;
1017                 target_cancel_recovery_timer(obd);
1018                 spin_unlock_bh(&obd->obd_processing_task_lock);
1019
1020                 target_finish_recovery(obd);
1021                 ptlrpc_run_recovery_over_upcall(obd);
1022         } else {
1023                 CWARN("%s: %d recoverable clients remain\n",
1024                        obd->obd_name, obd->obd_recoverable_clients);
1025                 wake_up(&obd->obd_next_transno_waitq);
1026         }
1027
1028         return 1;
1029 }
1030
1031 int
1032 target_send_reply_msg (struct ptlrpc_request *req, int rc, int fail_id)
1033 {
1034         if (OBD_FAIL_CHECK(fail_id | OBD_FAIL_ONCE)) {
1035                 obd_fail_loc |= OBD_FAIL_ONCE | OBD_FAILED;
1036                 DEBUG_REQ(D_ERROR, req, "dropping reply");
1037                 /* NB this does _not_ send with ACK disabled, to simulate
1038                  * sending OK, but timing out for the ACK */
1039                 if (req->rq_reply_state != NULL) {
1040                         if (!req->rq_reply_state->rs_difficult) {
1041                                 lustre_free_reply_state (req->rq_reply_state);
1042                                 req->rq_reply_state = NULL;
1043                         } else {
1044                                 struct ptlrpc_service *svc =
1045                                         req->rq_rqbd->rqbd_srv_ni->sni_service;
1046                                 atomic_inc(&svc->srv_outstanding_replies);
1047                         }
1048                 }
1049                 return (-ECOMM);
1050         }
1051
1052         if (rc) {
1053                 DEBUG_REQ(D_ERROR, req, "processing error (%d)", rc);
1054                 req->rq_status = rc;
1055                 return (ptlrpc_error(req));
1056         } else {
1057                 DEBUG_REQ(D_NET, req, "sending reply");
1058         }
1059
1060         return (ptlrpc_send_reply(req, 1));
1061 }
1062
1063 void
1064 target_send_reply(struct ptlrpc_request *req, int rc, int fail_id)
1065 {
1066         int                        netrc;
1067         unsigned long              flags;
1068         struct ptlrpc_reply_state *rs;
1069         struct obd_device         *obd;
1070         struct obd_export         *exp;
1071         struct ptlrpc_srv_ni      *sni;
1072         struct ptlrpc_service     *svc;
1073
1074         sni = req->rq_rqbd->rqbd_srv_ni;
1075         svc = sni->sni_service;
1076
1077         rs = req->rq_reply_state;
1078         if (rs == NULL || !rs->rs_difficult) {
1079                 /* The easy case; no notifiers and reply_out_callback()
1080                  * cleans up (i.e. we can't look inside rs after a
1081                  * successful send) */
1082                 netrc = target_send_reply_msg (req, rc, fail_id);
1083
1084                 LASSERT (netrc == 0 || req->rq_reply_state == NULL);
1085                 return;
1086         }
1087
1088         /* must be an export if locks saved */
1089         LASSERT (req->rq_export != NULL);
1090         /* req/reply consistent */
1091         LASSERT (rs->rs_srv_ni == sni);
1092
1093         /* "fresh" reply */
1094         LASSERT (!rs->rs_scheduled);
1095         LASSERT (!rs->rs_scheduled_ever);
1096         LASSERT (!rs->rs_handled);
1097         LASSERT (!rs->rs_on_net);
1098         LASSERT (rs->rs_export == NULL);
1099         LASSERT (list_empty(&rs->rs_obd_list));
1100         LASSERT (list_empty(&rs->rs_exp_list));
1101
1102         exp = class_export_get (req->rq_export);
1103         obd = exp->exp_obd;
1104
1105         /* disable reply scheduling onto srv_reply_queue while I'm setting up */
1106         rs->rs_scheduled = 1;
1107         rs->rs_on_net    = 1;
1108         rs->rs_xid       = req->rq_xid;
1109         rs->rs_transno   = req->rq_transno;
1110         rs->rs_export    = exp;
1111
1112         spin_lock_irqsave (&obd->obd_uncommitted_replies_lock, flags);
1113
1114         if (rs->rs_transno > obd->obd_last_committed) {
1115                 /* not committed already */
1116                 list_add_tail (&rs->rs_obd_list,
1117                                &obd->obd_uncommitted_replies);
1118         }
1119
1120         spin_unlock (&obd->obd_uncommitted_replies_lock);
1121         spin_lock (&exp->exp_lock);
1122
1123         list_add_tail (&rs->rs_exp_list, &exp->exp_outstanding_replies);
1124
1125         spin_unlock_irqrestore (&exp->exp_lock, flags);
1126
1127         netrc = target_send_reply_msg (req, rc, fail_id);
1128
1129         spin_lock_irqsave (&svc->srv_lock, flags);
1130
1131         svc->srv_n_difficult_replies++;
1132
1133         if (netrc != 0) /* error sending: reply is off the net */
1134                 rs->rs_on_net = 0;
1135
1136         if (!rs->rs_on_net ||                   /* some notifier */
1137             list_empty(&rs->rs_exp_list) ||     /* completed already */
1138             list_empty(&rs->rs_obd_list)) {
1139                 list_add_tail (&rs->rs_list, &svc->srv_reply_queue);
1140                 wake_up (&svc->srv_waitq);
1141         } else {
1142                 list_add (&rs->rs_list, &sni->sni_active_replies);
1143                 rs->rs_scheduled = 0;           /* allow notifier to schedule */
1144         }
1145
1146         spin_unlock_irqrestore (&svc->srv_lock, flags);
1147 }
1148
1149 int target_handle_ping(struct ptlrpc_request *req)
1150 {
1151         return lustre_pack_reply(req, 0, NULL, NULL);
1152 }