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  * GPL HEADER START
   5  *
   6  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
   7  *
   8  * This program is free software; you can redistribute it and/or modify
   9  * it under the terms of the GNU General Public License version 2 only,
  10  * as published by the Free Software Foundation.
  11  *
  12  * This program is distributed in the hope that it will be useful, but
  13  * WITHOUT ANY WARRANTY; without even the implied warranty of
  14  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  15  * General Public License version 2 for more details (a copy is included
  16  * in the LICENSE file that accompanied this code).
  17  *
  18  * You should have received a copy of the GNU General Public License
  19  * version 2 along with this program; If not, see
  20  * http://www.sun.com/software/products/lustre/docs/GPLv2.pdf
  21  *
  22  * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
  23  * CA 95054 USA or visit www.sun.com if you need additional information or
  24  * have any questions.
  25  *
  26  * GPL HEADER END
  27  */
  28 /*
  29  * Copyright  2008 Sun Microsystems, Inc. All rights reserved
  30  * Use is subject to license terms.
  31  */
  32 /*
  33  * This file is part of Lustre, http://www.lustre.org/
  34  * Lustre is a trademark of Sun Microsystems, Inc.
  35  */
  36
  37 #ifndef EXPORT_SYMTAB
  38 # define EXPORT_SYMTAB
  39 #endif
  40 #define DEBUG_SUBSYSTEM S_LDLM
  41
  42 #ifdef __KERNEL__
  43 # include <libcfs/libcfs.h>
  44 #else
  45 # include <liblustre.h>
  46 #endif
  47 #include <obd.h>
  48 #include <lustre_mds.h>
  49 #include <lustre_dlm.h>
  50 #include <lustre_net.h>
  51 #include <lustre_sec.h>
  52 #include "ldlm_internal.h"
  53
  54
  55 /* @priority: if non-zero, move the selected to the list head
  56  * @create: if zero, only search in existed connections
  57  */
  58 static int import_set_conn(struct obd_import *imp, struct obd_uuid *uuid,
  59                            int priority, int create)
  60 {
  61         struct ptlrpc_connection *ptlrpc_conn;
  62         struct obd_import_conn *imp_conn = NULL, *item;
  63         int rc = 0;
  64         ENTRY;
  65
  66         if (!create && !priority) {
  67                 CDEBUG(D_HA, "Nothing to do\n");
  68                 RETURN(-EINVAL);
  69         }
  70
  71         ptlrpc_conn = ptlrpc_uuid_to_connection(uuid);
  72         if (!ptlrpc_conn) {
  73                 CDEBUG(D_HA, "can't find connection %s\n", uuid->uuid);
  74                 RETURN (-ENOENT);
  75         }
  76
  77         if (create) {
  78                 OBD_ALLOC(imp_conn, sizeof(*imp_conn));
  79                 if (!imp_conn) {
  80                         GOTO(out_put, rc = -ENOMEM);
  81                 }
  82         }
  83
  84         spin_lock(&imp->imp_lock);
  85         list_for_each_entry(item, &imp->imp_conn_list, oic_item) {
  86                 if (obd_uuid_equals(uuid, &item->oic_uuid)) {
  87                         if (priority) {
  88                                 list_del(&item->oic_item);
  89                                 list_add(&item->oic_item, &imp->imp_conn_list);
  90                                 item->oic_last_attempt = 0;
  91                         }
  92                         CDEBUG(D_HA, "imp %p@%s: found existing conn %s%s\n",
  93                                imp, imp->imp_obd->obd_name, uuid->uuid,
  94                                (priority ? ", moved to head" : ""));
  95                         spin_unlock(&imp->imp_lock);
  96                         GOTO(out_free, rc = 0);
  97                 }
  98         }
  99         /* not found */
 100         if (create) {
 101                 imp_conn->oic_conn = ptlrpc_conn;
 102                 imp_conn->oic_uuid = *uuid;
 103                 item->oic_last_attempt = 0;
 104                 if (priority)
 105                         list_add(&imp_conn->oic_item, &imp->imp_conn_list);
 106                 else
 107                         list_add_tail(&imp_conn->oic_item, &imp->imp_conn_list);
 108                 CDEBUG(D_HA, "imp %p@%s: add connection %s at %s\n",
 109                        imp, imp->imp_obd->obd_name, uuid->uuid,
 110                        (priority ? "head" : "tail"));
 111         } else {
 112                 spin_unlock(&imp->imp_lock);
 113                 GOTO(out_free, rc = -ENOENT);
 114         }
 115
 116         spin_unlock(&imp->imp_lock);
 117         RETURN(0);
 118 out_free:
 119         if (imp_conn)
 120                 OBD_FREE(imp_conn, sizeof(*imp_conn));
 121 out_put:
 122         ptlrpc_connection_put(ptlrpc_conn);
 123         RETURN(rc);
 124 }
 125
 126 int import_set_conn_priority(struct obd_import *imp, struct obd_uuid *uuid)
 127 {
 128         return import_set_conn(imp, uuid, 1, 0);
 129 }
 130
 131 int client_import_add_conn(struct obd_import *imp, struct obd_uuid *uuid,
 132                            int priority)
 133 {
 134         return import_set_conn(imp, uuid, priority, 1);
 135 }
 136
 137 int client_import_del_conn(struct obd_import *imp, struct obd_uuid *uuid)
 138 {
 139         struct obd_import_conn *imp_conn;
 140         struct obd_export *dlmexp;
 141         int rc = -ENOENT;
 142         ENTRY;
 143
 144         spin_lock(&imp->imp_lock);
 145         if (list_empty(&imp->imp_conn_list)) {
 146                 LASSERT(!imp->imp_connection);
 147                 GOTO(out, rc);
 148         }
 149
 150         list_for_each_entry(imp_conn, &imp->imp_conn_list, oic_item) {
 151                 if (!obd_uuid_equals(uuid, &imp_conn->oic_uuid))
 152                         continue;
 153                 LASSERT(imp_conn->oic_conn);
 154
 155                 /* is current conn? */
 156                 if (imp_conn == imp->imp_conn_current) {
 157                         LASSERT(imp_conn->oic_conn == imp->imp_connection);
 158
 159                         if (imp->imp_state != LUSTRE_IMP_CLOSED &&
 160                             imp->imp_state != LUSTRE_IMP_DISCON) {
 161                                 CERROR("can't remove current connection\n");
 162                                 GOTO(out, rc = -EBUSY);
 163                         }
 164
 165                         ptlrpc_connection_put(imp->imp_connection);
 166                         imp->imp_connection = NULL;
 167
 168                         dlmexp = class_conn2export(&imp->imp_dlm_handle);
 169                         if (dlmexp && dlmexp->exp_connection) {
 170                                 LASSERT(dlmexp->exp_connection ==
 171                                         imp_conn->oic_conn);
 172                                 ptlrpc_connection_put(dlmexp->exp_connection);
 173                                 dlmexp->exp_connection = NULL;
 174                         }
 175                 }
 176
 177                 list_del(&imp_conn->oic_item);
 178                 ptlrpc_connection_put(imp_conn->oic_conn);
 179                 OBD_FREE(imp_conn, sizeof(*imp_conn));
 180                 CDEBUG(D_HA, "imp %p@%s: remove connection %s\n",
 181                        imp, imp->imp_obd->obd_name, uuid->uuid);
 182                 rc = 0;
 183                 break;
 184         }
 185 out:
 186         spin_unlock(&imp->imp_lock);
 187         if (rc == -ENOENT)
 188                 CERROR("connection %s not found\n", uuid->uuid);
 189         RETURN(rc);
 190 }
 191
 192 static void destroy_import(struct obd_import *imp)
 193 {
 194         /* drop security policy instance after all rpc finished/aborted
 195          * to let all busy contexts be released. */
 196         class_import_get(imp);
 197         class_destroy_import(imp);
 198         sptlrpc_import_sec_put(imp);
 199         class_import_put(imp);
 200 }
 201
 202 /* configure an RPC client OBD device
 203  *
 204  * lcfg parameters:
 205  * 1 - client UUID
 206  * 2 - server UUID
 207  * 3 - inactive-on-startup
 208  */
 209 int client_obd_setup(struct obd_device *obddev, struct lustre_cfg *lcfg)
 210 {
 211         struct client_obd *cli = &obddev->u.cli;
 212         struct obd_import *imp;
 213         struct obd_uuid server_uuid;
 214         int rq_portal, rp_portal, connect_op;
 215         char *name = obddev->obd_type->typ_name;
 216         int rc;
 217         ENTRY;
 218
 219         /* In a more perfect world, we would hang a ptlrpc_client off of
 220          * obd_type and just use the values from there. */
 221         if (!strcmp(name, LUSTRE_OSC_NAME)) {
 222                 rq_portal = OST_REQUEST_PORTAL;
 223                 rp_portal = OSC_REPLY_PORTAL;
 224                 connect_op = OST_CONNECT;
 225         } else if (!strcmp(name, LUSTRE_MDC_NAME)) {
 226                 rq_portal = MDS_REQUEST_PORTAL;
 227                 rp_portal = MDC_REPLY_PORTAL;
 228                 connect_op = MDS_CONNECT;
 229         } else if (!strcmp(name, LUSTRE_MGC_NAME)) {
 230                 rq_portal = MGS_REQUEST_PORTAL;
 231                 rp_portal = MGC_REPLY_PORTAL;
 232                 connect_op = MGS_CONNECT;
 233         } else {
 234                 CERROR("unknown client OBD type \"%s\", can't setup\n",
 235                        name);
 236                 RETURN(-EINVAL);
 237         }
 238
 239         if (LUSTRE_CFG_BUFLEN(lcfg, 1) < 1) {
 240                 CERROR("requires a TARGET UUID\n");
 241                 RETURN(-EINVAL);
 242         }
 243
 244         if (LUSTRE_CFG_BUFLEN(lcfg, 1) > 37) {
 245                 CERROR("client UUID must be less than 38 characters\n");
 246                 RETURN(-EINVAL);
 247         }
 248
 249         if (LUSTRE_CFG_BUFLEN(lcfg, 2) < 1) {
 250                 CERROR("setup requires a SERVER UUID\n");
 251                 RETURN(-EINVAL);
 252         }
 253
 254         if (LUSTRE_CFG_BUFLEN(lcfg, 2) > 37) {
 255                 CERROR("target UUID must be less than 38 characters\n");
 256                 RETURN(-EINVAL);
 257         }
 258
 259         init_rwsem(&cli->cl_sem);
 260         sema_init(&cli->cl_mgc_sem, 1);
 261         sptlrpc_rule_set_init(&cli->cl_sptlrpc_rset);
 262         cli->cl_sec_part = LUSTRE_SP_ANY;
 263         cli->cl_conn_count = 0;
 264         memcpy(server_uuid.uuid, lustre_cfg_buf(lcfg, 2),
 265                min_t(unsigned int, LUSTRE_CFG_BUFLEN(lcfg, 2),
 266                      sizeof(server_uuid)));
 267
 268         cli->cl_dirty = 0;
 269         cli->cl_avail_grant = 0;
 270         /* FIXME: should limit this for the sum of all cl_dirty_max */
 271         cli->cl_dirty_max = OSC_MAX_DIRTY_DEFAULT * 1024 * 1024;
 272         if (cli->cl_dirty_max >> CFS_PAGE_SHIFT > num_physpages / 8)
 273                 cli->cl_dirty_max = num_physpages << (CFS_PAGE_SHIFT - 3);
 274         CFS_INIT_LIST_HEAD(&cli->cl_cache_waiters);
 275         CFS_INIT_LIST_HEAD(&cli->cl_loi_ready_list);
 276         CFS_INIT_LIST_HEAD(&cli->cl_loi_write_list);
 277         CFS_INIT_LIST_HEAD(&cli->cl_loi_read_list);
 278         client_obd_list_lock_init(&cli->cl_loi_list_lock);
 279         cli->cl_r_in_flight = 0;
 280         cli->cl_w_in_flight = 0;
 281
 282         spin_lock_init(&cli->cl_read_rpc_hist.oh_lock);
 283         spin_lock_init(&cli->cl_write_rpc_hist.oh_lock);
 284         spin_lock_init(&cli->cl_read_page_hist.oh_lock);
 285         spin_lock_init(&cli->cl_write_page_hist.oh_lock);
 286         spin_lock_init(&cli->cl_read_offset_hist.oh_lock);
 287         spin_lock_init(&cli->cl_write_offset_hist.oh_lock);
 288         cfs_waitq_init(&cli->cl_destroy_waitq);
 289         atomic_set(&cli->cl_destroy_in_flight, 0);
 290 #ifdef ENABLE_CHECKSUM
 291         /* Turn on checksumming by default. */
 292         cli->cl_checksum = 1;
 293         /*
 294          * The supported checksum types will be worked out at connect time
 295          * Set cl_chksum* to CRC32 for now to avoid returning screwed info
 296          * through procfs.
 297          */
 298         cli->cl_cksum_type = cli->cl_supp_cksum_types = OBD_CKSUM_CRC32;
 299 #endif
 300         atomic_set(&cli->cl_resends, OSC_DEFAULT_RESENDS);
 301
 302         /* This value may be changed at connect time in
 303            ptlrpc_connect_interpret. */
 304         cli->cl_max_pages_per_rpc = min((int)PTLRPC_MAX_BRW_PAGES,
 305                                         (int)(1024 * 1024 >> CFS_PAGE_SHIFT));
 306
 307         if (!strcmp(name, LUSTRE_MDC_NAME)) {
 308                 cli->cl_max_rpcs_in_flight = MDC_MAX_RIF_DEFAULT;
 309         } else if (num_physpages >> (20 - CFS_PAGE_SHIFT) <= 128 /* MB */) {
 310                 cli->cl_max_rpcs_in_flight = 2;
 311         } else if (num_physpages >> (20 - CFS_PAGE_SHIFT) <= 256 /* MB */) {
 312                 cli->cl_max_rpcs_in_flight = 3;
 313         } else if (num_physpages >> (20 - CFS_PAGE_SHIFT) <= 512 /* MB */) {
 314                 cli->cl_max_rpcs_in_flight = 4;
 315         } else {
 316                 cli->cl_max_rpcs_in_flight = OSC_MAX_RIF_DEFAULT;
 317         }
 318
 319         rc = ldlm_get_ref();
 320         if (rc) {
 321                 CERROR("ldlm_get_ref failed: %d\n", rc);
 322                 GOTO(err, rc);
 323         }
 324
 325         ptlrpc_init_client(rq_portal, rp_portal, name,
 326                            &obddev->obd_ldlm_client);
 327
 328         imp = class_new_import(obddev);
 329         if (imp == NULL)
 330                 GOTO(err_ldlm, rc = -ENOENT);
 331         imp->imp_client = &obddev->obd_ldlm_client;
 332         imp->imp_connect_op = connect_op;
 333         imp->imp_initial_recov = 1;
 334         imp->imp_initial_recov_bk = 0;
 335         CFS_INIT_LIST_HEAD(&imp->imp_pinger_chain);
 336         memcpy(cli->cl_target_uuid.uuid, lustre_cfg_buf(lcfg, 1),
 337                LUSTRE_CFG_BUFLEN(lcfg, 1));
 338         class_import_put(imp);
 339
 340         rc = client_import_add_conn(imp, &server_uuid, 1);
 341         if (rc) {
 342                 CERROR("can't add initial connection\n");
 343                 GOTO(err_import, rc);
 344         }
 345
 346         cli->cl_import = imp;
 347         /* cli->cl_max_mds_{easize,cookiesize} updated by mdc_init_ea_size() */
 348         cli->cl_max_mds_easize = sizeof(struct lov_mds_md_v3);
 349         cli->cl_max_mds_cookiesize = sizeof(struct llog_cookie);
 350
 351         if (LUSTRE_CFG_BUFLEN(lcfg, 3) > 0) {
 352                 if (!strcmp(lustre_cfg_string(lcfg, 3), "inactive")) {
 353                         CDEBUG(D_HA, "marking %s %s->%s as inactive\n",
 354                                name, obddev->obd_name,
 355                                cli->cl_target_uuid.uuid);
 356                         spin_lock(&imp->imp_lock);
 357                         imp->imp_invalid = 1;
 358                         spin_unlock(&imp->imp_lock);
 359                 }
 360         }
 361
 362         cli->cl_qchk_stat = CL_NOT_QUOTACHECKED;
 363
 364         RETURN(rc);
 365
 366 err_import:
 367         class_destroy_import(imp);
 368 err_ldlm:
 369         ldlm_put_ref();
 370 err:
 371         RETURN(rc);
 372
 373 }
 374
 375 int client_obd_cleanup(struct obd_device *obddev)
 376 {
 377         ENTRY;
 378         sptlrpc_rule_set_free(&obddev->u.cli.cl_sptlrpc_rset);
 379         ldlm_put_ref();
 380         RETURN(0);
 381 }
 382
 383 /* ->o_connect() method for client side (OSC and MDC and MGC) */
 384 int client_connect_import(const struct lu_env *env,
 385                           struct lustre_handle *dlm_handle,
 386                           struct obd_device *obd, struct obd_uuid *cluuid,
 387                           struct obd_connect_data *data, void *localdata)
 388 {
 389         struct client_obd *cli = &obd->u.cli;
 390         struct obd_import *imp = cli->cl_import;
 391         struct obd_export *exp;
 392         struct obd_connect_data *ocd;
 393         struct ldlm_namespace *to_be_freed = NULL;
 394         int rc;
 395         ENTRY;
 396
 397         down_write(&cli->cl_sem);
 398         rc = class_connect(dlm_handle, obd, cluuid);
 399         if (rc)
 400                 GOTO(out_sem, rc);
 401
 402         cli->cl_conn_count++;
 403         if (cli->cl_conn_count > 1)
 404                 GOTO(out_sem, rc);
 405         exp = class_conn2export(dlm_handle);
 406
 407         if (obd->obd_namespace != NULL)
 408                 CERROR("already have namespace!\n");
 409         obd->obd_namespace = ldlm_namespace_new(obd, obd->obd_name,
 410                                                 LDLM_NAMESPACE_CLIENT,
 411                                                 LDLM_NAMESPACE_GREEDY);
 412         if (obd->obd_namespace == NULL)
 413                 GOTO(out_disco, rc = -ENOMEM);
 414
 415         imp->imp_dlm_handle = *dlm_handle;
 416         rc = ptlrpc_init_import(imp);
 417         if (rc != 0)
 418                 GOTO(out_ldlm, rc);
 419
 420         ocd = &imp->imp_connect_data;
 421         if (data) {
 422                 *ocd = *data;
 423                 imp->imp_connect_flags_orig = data->ocd_connect_flags;
 424         }
 425
 426         rc = ptlrpc_connect_import(imp, NULL);
 427         if (rc != 0) {
 428                 LASSERT (imp->imp_state == LUSTRE_IMP_DISCON);
 429                 GOTO(out_ldlm, rc);
 430         }
 431         LASSERT(exp->exp_connection);
 432
 433         if (data) {
 434                 LASSERTF((ocd->ocd_connect_flags & data->ocd_connect_flags) ==
 435                          ocd->ocd_connect_flags, "old "LPX64", new "LPX64"\n",
 436                          data->ocd_connect_flags, ocd->ocd_connect_flags);
 437                 data->ocd_connect_flags = ocd->ocd_connect_flags;
 438         }
 439
 440         ptlrpc_pinger_add_import(imp);
 441
 442         EXIT;
 443
 444         if (rc) {
 445 out_ldlm:
 446                 ldlm_namespace_free_prior(obd->obd_namespace, imp, 0);
 447                 to_be_freed = obd->obd_namespace;
 448                 obd->obd_namespace = NULL;
 449 out_disco:
 450                 cli->cl_conn_count--;
 451                 class_disconnect(exp);
 452         } else {
 453                 class_export_put(exp);
 454         }
 455 out_sem:
 456         up_write(&cli->cl_sem);
 457         if (to_be_freed)
 458                 ldlm_namespace_free_post(to_be_freed);
 459
 460         return rc;
 461 }
 462
 463 int client_disconnect_export(struct obd_export *exp)
 464 {
 465         struct obd_device *obd = class_exp2obd(exp);
 466         struct client_obd *cli;
 467         struct obd_import *imp;
 468         int rc = 0, err;
 469         struct ldlm_namespace *to_be_freed = NULL;
 470         ENTRY;
 471
 472         if (!obd) {
 473                 CERROR("invalid export for disconnect: exp %p cookie "LPX64"\n",
 474                        exp, exp ? exp->exp_handle.h_cookie : -1);
 475                 RETURN(-EINVAL);
 476         }
 477
 478         cli = &obd->u.cli;
 479         imp = cli->cl_import;
 480
 481         down_write(&cli->cl_sem);
 482         if (!cli->cl_conn_count) {
 483                 CERROR("disconnecting disconnected device (%s)\n",
 484                        obd->obd_name);
 485                 GOTO(out_sem, rc = -EINVAL);
 486         }
 487
 488         cli->cl_conn_count--;
 489         if (cli->cl_conn_count)
 490                 GOTO(out_no_disconnect, rc = 0);
 491
 492         /* Mark import deactivated now, so we don't try to reconnect if any
 493          * of the cleanup RPCs fails (e.g. ldlm cancel, etc).  We don't
 494          * fully deactivate the import, or that would drop all requests. */
 495         spin_lock(&imp->imp_lock);
 496         imp->imp_deactive = 1;
 497         spin_unlock(&imp->imp_lock);
 498
 499         /* Some non-replayable imports (MDS's OSCs) are pinged, so just
 500          * delete it regardless.  (It's safe to delete an import that was
 501          * never added.) */
 502         (void)ptlrpc_pinger_del_import(imp);
 503
 504         if (obd->obd_namespace != NULL) {
 505                 /* obd_force == local only */
 506                 ldlm_cli_cancel_unused(obd->obd_namespace, NULL,
 507                                        obd->obd_force ? LDLM_FL_LOCAL_ONLY:0,
 508                                        NULL);
 509                 ldlm_namespace_free_prior(obd->obd_namespace, imp, obd->obd_force);
 510                 to_be_freed = obd->obd_namespace;
 511         }
 512
 513         rc = ptlrpc_disconnect_import(imp, 0);
 514
 515         ptlrpc_invalidate_import(imp);
 516         /* set obd_namespace to NULL only after invalidate, because we can have
 517          * some connect requests in flight, and his need store a connect flags
 518          * in obd_namespace. bug 14260 */
 519         obd->obd_namespace = NULL;
 520
 521         ptlrpc_free_rq_pool(imp->imp_rq_pool);
 522         destroy_import(imp);
 523         cli->cl_import = NULL;
 524
 525         EXIT;
 526  out_no_disconnect:
 527         err = class_disconnect(exp);
 528         if (!rc && err)
 529                 rc = err;
 530  out_sem:
 531         up_write(&cli->cl_sem);
 532         if (to_be_freed)
 533                 ldlm_namespace_free_post(to_be_freed);
 534
 535         RETURN(rc);
 536 }
 537
 538 /* --------------------------------------------------------------------------
 539  * from old lib/target.c
 540  * -------------------------------------------------------------------------- */
 541
 542 int target_handle_reconnect(struct lustre_handle *conn, struct obd_export *exp,
 543                             struct obd_uuid *cluuid, int initial_conn)
 544 {
 545         ENTRY;
 546         if (exp->exp_connection && exp->exp_imp_reverse && !initial_conn) {
 547                 struct lustre_handle *hdl;
 548                 hdl = &exp->exp_imp_reverse->imp_remote_handle;
 549                 /* Might be a re-connect after a partition. */
 550                 if (!memcmp(&conn->cookie, &hdl->cookie, sizeof conn->cookie)) {
 551                         CWARN("%s: %s reconnecting\n", exp->exp_obd->obd_name,
 552                               cluuid->uuid);
 553                         conn->cookie = exp->exp_handle.h_cookie;
 554                         /* target_handle_connect() treats EALREADY and
 555                          * -EALREADY differently.  EALREADY means we are
 556                          * doing a valid reconnect from the same client. */
 557                         RETURN(EALREADY);
 558                 } else {
 559                         CERROR("%s reconnecting from %s, "
 560                                "handle mismatch (ours "LPX64", theirs "
 561                                LPX64")\n", cluuid->uuid,
 562                                exp->exp_connection->c_remote_uuid.uuid,
 563                                hdl->cookie, conn->cookie);
 564                         memset(conn, 0, sizeof *conn);
 565                         /* target_handle_connect() treats EALREADY and
 566                          * -EALREADY differently.  -EALREADY is an error
 567                          * (same UUID, different handle). */
 568                         RETURN(-EALREADY);
 569                 }
 570         }
 571
 572         conn->cookie = exp->exp_handle.h_cookie;
 573         CDEBUG(D_HA, "connect export for UUID '%s' at %p, cookie "LPX64"\n",
 574                cluuid->uuid, exp, conn->cookie);
 575         RETURN(0);
 576 }
 577
 578 void target_client_add_cb(struct obd_device *obd, __u64 transno, void *cb_data,
 579                           int error)
 580 {
 581         struct obd_export *exp = cb_data;
 582
 583         CDEBUG(D_RPCTRACE, "%s: committing for initial connect of %s\n",
 584                obd->obd_name, exp->exp_client_uuid.uuid);
 585
 586         spin_lock(&exp->exp_lock);
 587         exp->exp_need_sync = 0;
 588         spin_unlock(&exp->exp_lock);
 589 }
 590 EXPORT_SYMBOL(target_client_add_cb);
 591 static void
 592 target_start_and_reset_recovery_timer(struct obd_device *obd,
 593                                       struct ptlrpc_request *req,
 594                                       int new_client);
 595
 596 int target_handle_connect(struct ptlrpc_request *req)
 597 {
 598         struct obd_device *target, *targref = NULL;
 599         struct obd_export *export = NULL;
 600         struct obd_import *revimp;
 601         struct lustre_handle conn;
 602         struct lustre_handle *tmp;
 603         struct obd_uuid tgtuuid;
 604         struct obd_uuid cluuid;
 605         struct obd_uuid remote_uuid;
 606         char *str;
 607         int rc = 0;
 608         int initial_conn = 0;
 609         struct obd_connect_data *data, *tmpdata;
 610         lnet_nid_t *client_nid = NULL;
 611         ENTRY;
 612
 613         OBD_RACE(OBD_FAIL_TGT_CONN_RACE);
 614
 615         str = req_capsule_client_get(&req->rq_pill, &RMF_TGTUUID);
 616         if (str == NULL) {
 617                 DEBUG_REQ(D_ERROR, req, "bad target UUID for connect");
 618                 GOTO(out, rc = -EINVAL);
 619         }
 620
 621         obd_str2uuid(&tgtuuid, str);
 622         target = class_uuid2obd(&tgtuuid);
 623         if (!target)
 624                 target = class_name2obd(str);
 625
 626         if (!target || target->obd_stopping || !target->obd_set_up) {
 627                 LCONSOLE_ERROR_MSG(0x137, "UUID '%s' is not available "
 628                                    " for connect (%s)\n", str,
 629                                    !target ? "no target" :
 630                                    (target->obd_stopping ? "stopping" :
 631                                    "not set up"));
 632                 GOTO(out, rc = -ENODEV);
 633         }
 634
 635         if (target->obd_no_conn) {
 636                 LCONSOLE_WARN("%s: temporarily refusing client connection "
 637                               "from %s\n", target->obd_name,
 638                               libcfs_nid2str(req->rq_peer.nid));
 639                 GOTO(out, rc = -EAGAIN);
 640         }
 641
 642         /* Make sure the target isn't cleaned up while we're here. Yes,
 643            there's still a race between the above check and our incref here.
 644            Really, class_uuid2obd should take the ref. */
 645         targref = class_incref(target, __FUNCTION__, cfs_current());
 646
 647
 648         str = req_capsule_client_get(&req->rq_pill, &RMF_CLUUID);
 649         if (str == NULL) {
 650                 DEBUG_REQ(D_ERROR, req, "bad client UUID for connect");
 651                 GOTO(out, rc = -EINVAL);
 652         }
 653
 654         obd_str2uuid(&cluuid, str);
 655
 656         /* XXX extract a nettype and format accordingly */
 657         switch (sizeof(lnet_nid_t)) {
 658                 /* NB the casts only avoid compiler warnings */
 659         case 8:
 660                 snprintf(remote_uuid.uuid, sizeof remote_uuid,
 661                          "NET_"LPX64"_UUID", (__u64)req->rq_peer.nid);
 662                 break;
 663         case 4:
 664                 snprintf(remote_uuid.uuid, sizeof remote_uuid,
 665                          "NET_%x_UUID", (__u32)req->rq_peer.nid);
 666                 break;
 667         default:
 668                 LBUG();
 669         }
 670
 671         tmp = req_capsule_client_get(&req->rq_pill, &RMF_CONN);
 672         if (tmp == NULL)
 673                 GOTO(out, rc = -EPROTO);
 674
 675         conn = *tmp;
 676
 677         data = req_capsule_client_get(&req->rq_pill, &RMF_CONNECT_DATA);
 678         if (!data)
 679                 GOTO(out, rc = -EPROTO);
 680
 681         rc = req_capsule_server_pack(&req->rq_pill);
 682         if (rc)
 683                 GOTO(out, rc);
 684
 685         if (lustre_msg_get_op_flags(req->rq_reqmsg) & MSG_CONNECT_LIBCLIENT) {
 686                 if (!data) {
 687                         DEBUG_REQ(D_WARNING, req, "Refusing old (unversioned) "
 688                                   "libclient connection attempt");
 689                         GOTO(out, rc = -EPROTO);
 690                 } else if (data->ocd_version < LUSTRE_VERSION_CODE -
 691                                                LUSTRE_VERSION_ALLOWED_OFFSET ||
 692                            data->ocd_version > LUSTRE_VERSION_CODE +
 693                                                LUSTRE_VERSION_ALLOWED_OFFSET) {
 694                         DEBUG_REQ(D_WARNING, req, "Refusing %s (%d.%d.%d.%d) "
 695                                   "libclient connection attempt",
 696                                   data->ocd_version < LUSTRE_VERSION_CODE ?
 697                                   "old" : "new",
 698                                   OBD_OCD_VERSION_MAJOR(data->ocd_version),
 699                                   OBD_OCD_VERSION_MINOR(data->ocd_version),
 700                                   OBD_OCD_VERSION_PATCH(data->ocd_version),
 701                                   OBD_OCD_VERSION_FIX(data->ocd_version));
 702                         data = req_capsule_server_sized_get(&req->rq_pill,
 703                                                         &RMF_CONNECT_DATA,
 704                                     offsetof(typeof(*data), ocd_version) +
 705                                              sizeof(data->ocd_version));
 706                         if (data) {
 707                                 data->ocd_connect_flags = OBD_CONNECT_VERSION;
 708                                 data->ocd_version = LUSTRE_VERSION_CODE;
 709                         }
 710                         GOTO(out, rc = -EPROTO);
 711                 }
 712         }
 713
 714         if (lustre_msg_get_op_flags(req->rq_reqmsg) & MSG_CONNECT_INITIAL)
 715                 initial_conn = 1;
 716
 717         /* lctl gets a backstage, all-access pass. */
 718         if (obd_uuid_equals(&cluuid, &target->obd_uuid))
 719                 goto dont_check_exports;
 720
 721         spin_lock(&target->obd_dev_lock);
 722         export = lustre_hash_lookup(target->obd_uuid_hash, &cluuid);
 723
 724         if (export != NULL && export->exp_connecting) { /* bug 9635, et. al. */
 725                 CWARN("%s: exp %p already connecting\n",
 726                       export->exp_obd->obd_name, export);
 727                 class_export_put(export);
 728                 export = NULL;
 729                 rc = -EALREADY;
 730         } else if (export != NULL && export->exp_connection != NULL &&
 731                    req->rq_peer.nid != export->exp_connection->c_peer.nid) {
 732                 /* make darn sure this is coming from the same peer
 733                  * if the UUIDs matched */
 734                   CWARN("%s: cookie %s seen on new NID %s when "
 735                           "existing NID %s is already connected\n",
 736                         target->obd_name, cluuid.uuid,
 737                   libcfs_nid2str(req->rq_peer.nid),
 738                   libcfs_nid2str(export->exp_connection->c_peer.nid));
 739                   class_export_put(export);
 740                   export = NULL;
 741                   rc = -EALREADY;
 742         } else if (export != NULL) {
 743                 spin_lock(&export->exp_lock);
 744                 export->exp_connecting = 1;
 745                 spin_unlock(&export->exp_lock);
 746                 class_export_put(export);
 747                 spin_unlock(&target->obd_dev_lock);
 748                 LASSERT(export->exp_obd == target);
 749
 750                 rc = target_handle_reconnect(&conn, export, &cluuid, initial_conn);
 751         }
 752
 753         /* If we found an export, we already unlocked. */
 754         if (!export) {
 755                 spin_unlock(&target->obd_dev_lock);
 756                 OBD_FAIL_TIMEOUT(OBD_FAIL_TGT_DELAY_CONNECT, 2 * obd_timeout);
 757         } else if (req->rq_export == NULL &&
 758                    atomic_read(&export->exp_rpc_count) > 0) {
 759                 CWARN("%s: refuse connection from %s/%s to 0x%p/%d\n",
 760                       target->obd_name, cluuid.uuid,
 761                       libcfs_nid2str(req->rq_peer.nid),
 762                       export, atomic_read(&export->exp_refcount));
 763                 GOTO(out, rc = -EBUSY);
 764         } else if (req->rq_export != NULL &&
 765                    (atomic_read(&export->exp_rpc_count) > 1)) {
 766                 CWARN("%s: refuse reconnection from %s@%s to 0x%p/%d\n",
 767                       target->obd_name, cluuid.uuid,
 768                       libcfs_nid2str(req->rq_peer.nid),
 769                       export, atomic_read(&export->exp_rpc_count));
 770                 GOTO(out, rc = -EBUSY);
 771         } else if (lustre_msg_get_conn_cnt(req->rq_reqmsg) == 1 &&
 772                    !initial_conn) {
 773                 CERROR("%s: NID %s (%s) reconnected with 1 conn_cnt; "
 774                        "cookies not random?\n", target->obd_name,
 775                        libcfs_nid2str(req->rq_peer.nid), cluuid.uuid);
 776                 GOTO(out, rc = -EALREADY);
 777         } else {
 778                 OBD_FAIL_TIMEOUT(OBD_FAIL_TGT_DELAY_RECONNECT, 2 * obd_timeout);
 779                 if (req->rq_export == NULL && initial_conn)
 780                        export->exp_last_request_time =
 781                                max(export->exp_last_request_time,
 782                                    (time_t)cfs_time_current_sec());
 783         }
 784
 785         if (rc < 0) {
 786                 GOTO(out, rc);
 787         }
 788
 789         CWARN("%s: connection from %s@%s %st"LPU64" exp %p cur %ld last %ld\n",
 790                target->obd_name, cluuid.uuid, libcfs_nid2str(req->rq_peer.nid),
 791               target->obd_recovering ? "recovering/" : "", data->ocd_transno,
 792               export, (long)cfs_time_current_sec(),
 793               export ? (long)export->exp_last_request_time : 0);
 794
 795         /* Tell the client if we're in recovery. */
 796         if (target->obd_recovering) {
 797                 lustre_msg_add_op_flags(req->rq_repmsg, MSG_CONNECT_RECOVERING);
 798                 /* If this is the first time a client connects,
 799                    reset the recovery timer */
 800                 if (rc == 0)
 801                         target_start_and_reset_recovery_timer(target, req,
 802                                                               !export);
 803         }
 804
 805         /* We want to handle EALREADY but *not* -EALREADY from
 806          * target_handle_reconnect(), return reconnection state in a flag */
 807         if (rc == EALREADY) {
 808                 lustre_msg_add_op_flags(req->rq_repmsg, MSG_CONNECT_RECONNECT);
 809                 rc = 0;
 810         } else {
 811                 LASSERT(rc == 0);
 812         }
 813
 814         /* Tell the client if we support replayable requests */
 815         if (target->obd_replayable)
 816                 lustre_msg_add_op_flags(req->rq_repmsg, MSG_CONNECT_REPLAYABLE);
 817         client_nid = &req->rq_peer.nid;
 818
 819         if (export == NULL) {
 820                 if (target->obd_recovering) {
 821                         cfs_time_t t;
 822
 823                         t = cfs_timer_deadline(&target->obd_recovery_timer);
 824                         t = cfs_time_sub(t, cfs_time_current());
 825                         CERROR("%s: denying connection for new client %s (%s): "
 826                                "%d clients in recovery for "CFS_TIME_T"s\n",
 827                                target->obd_name,
 828                                libcfs_nid2str(req->rq_peer.nid), cluuid.uuid,
 829                                target->obd_recoverable_clients,
 830                                cfs_duration_sec(t));
 831                         rc = -EBUSY;
 832                 } else {
 833 dont_check_exports:
 834                         rc = obd_connect(req->rq_svc_thread->t_env,
 835                                          &conn, target, &cluuid, data,
 836                                          client_nid);
 837                 }
 838         } else {
 839                 rc = obd_reconnect(req->rq_svc_thread->t_env,
 840                                    export, target, &cluuid, data);
 841         }
 842         if (rc)
 843                 GOTO(out, rc);
 844         /* Return only the parts of obd_connect_data that we understand, so the
 845          * client knows that we don't understand the rest. */
 846         if (data) {
 847                  tmpdata = req_capsule_server_get(&req->rq_pill,
 848                                                   &RMF_CONNECT_DATA);
 849                   //data->ocd_connect_flags &= OBD_CONNECT_SUPPORTED;
 850                  *tmpdata = *data;
 851         }
 852
 853         /* If all else goes well, this is our RPC return code. */
 854         req->rq_status = 0;
 855
 856         lustre_msg_set_handle(req->rq_repmsg, &conn);
 857
 858         /* ownership of this export ref transfers to the request AFTER we
 859          * drop any previous reference the request had, but we don't want
 860          * that to go to zero before we get our new export reference. */
 861         export = class_conn2export(&conn);
 862         if (!export) {
 863                 DEBUG_REQ(D_ERROR, req, "Missing export!");
 864                 GOTO(out, rc = -ENODEV);
 865         }
 866
 867         /* If the client and the server are the same node, we will already
 868          * have an export that really points to the client's DLM export,
 869          * because we have a shared handles table.
 870          *
 871          * XXX this will go away when shaver stops sending the "connect" handle
 872          * in the real "remote handle" field of the request --phik 24 Apr 2003
 873          */
 874         if (req->rq_export != NULL)
 875                 class_export_put(req->rq_export);
 876
 877         req->rq_export = export;
 878
 879         spin_lock(&export->exp_lock);
 880         if (initial_conn) {
 881                 lustre_msg_set_conn_cnt(req->rq_repmsg, export->exp_conn_cnt + 1);
 882         } else if (export->exp_conn_cnt >= lustre_msg_get_conn_cnt(req->rq_reqmsg)) {
 883                 spin_unlock(&export->exp_lock);
 884                 CERROR("%s: %s already connected at higher conn_cnt: %d > %d\n",
 885                        cluuid.uuid, libcfs_nid2str(req->rq_peer.nid),
 886                        export->exp_conn_cnt,
 887                        lustre_msg_get_conn_cnt(req->rq_reqmsg));
 888
 889                 GOTO(out, rc = -EALREADY);
 890         }
 891         export->exp_conn_cnt = lustre_msg_get_conn_cnt(req->rq_reqmsg);
 892
 893         /* request from liblustre?  Don't evict it for not pinging. */
 894         if (lustre_msg_get_op_flags(req->rq_reqmsg) & MSG_CONNECT_LIBCLIENT) {
 895                 export->exp_libclient = 1;
 896                 spin_unlock(&export->exp_lock);
 897
 898                 spin_lock(&target->obd_dev_lock);
 899                 list_del_init(&export->exp_obd_chain_timed);
 900                 spin_unlock(&target->obd_dev_lock);
 901         } else {
 902                 spin_unlock(&export->exp_lock);
 903         }
 904
 905         if (export->exp_connection != NULL)
 906                 ptlrpc_connection_put(export->exp_connection);
 907         export->exp_connection = ptlrpc_connection_get(req->rq_peer,
 908                                                        req->rq_self,
 909                                                        &remote_uuid);
 910
 911         spin_lock(&target->obd_dev_lock);
 912
 913         /* Export might be hashed already, e.g. if this is reconnect */
 914         if (hlist_unhashed(&export->exp_nid_hash))
 915                 lustre_hash_add(export->exp_obd->obd_nid_hash,
 916                                 &export->exp_connection->c_peer.nid,
 917                                 &export->exp_nid_hash);
 918
 919         spin_unlock(&target->obd_dev_lock);
 920
 921         spin_lock_bh(&target->obd_processing_task_lock);
 922         if (target->obd_recovering && !export->exp_in_recovery) {
 923                 spin_lock(&export->exp_lock);
 924                 export->exp_in_recovery = 1;
 925                 export->exp_req_replay_needed = 1;
 926                 export->exp_lock_replay_needed = 1;
 927                 spin_unlock(&export->exp_lock);
 928                 if ((lustre_msg_get_op_flags(req->rq_reqmsg) & MSG_CONNECT_TRANSNO)
 929                      && (data->ocd_transno == 0))
 930                         CWARN("Connect with zero transno!\n");
 931
 932                 if ((lustre_msg_get_op_flags(req->rq_reqmsg) & MSG_CONNECT_TRANSNO)
 933                      && data->ocd_transno < target->obd_next_recovery_transno)
 934                         target->obd_next_recovery_transno = data->ocd_transno;
 935                 target->obd_connected_clients++;
 936                 /* each connected client is counted as recoverable */
 937                 target->obd_recoverable_clients++;
 938                 atomic_inc(&target->obd_req_replay_clients);
 939                 atomic_inc(&target->obd_lock_replay_clients);
 940                 if (target->obd_connected_clients ==
 941                     target->obd_max_recoverable_clients)
 942                         wake_up(&target->obd_next_transno_waitq);
 943         }
 944         spin_unlock_bh(&target->obd_processing_task_lock);
 945         tmp = req_capsule_client_get(&req->rq_pill, &RMF_CONN);
 946         conn = *tmp;
 947
 948         if (export->exp_imp_reverse != NULL) {
 949                 /* destroyed import can be still referenced in ctxt */
 950                 obd_set_info_async(export, sizeof(KEY_REVIMP_UPD),
 951                                    KEY_REVIMP_UPD, 0, NULL, NULL);
 952
 953                 /* in some recovery senarios, previous ctx init rpc handled
 954                  * in sptlrpc_target_export_check() might be used to install
 955                  * a reverse ctx in this reverse import, and later OBD_CONNECT
 956                  * using the same gss ctx could reach here and following new
 957                  * reverse import. note all reverse ctx in new/old import are
 958                  * actually based on the same gss ctx. so we invalidate ctx
 959                  * here before destroy import, otherwise flush old import will
 960                  * lead to remote reverse ctx be destroied, thus the reverse
 961                  * ctx of new import will lost its peer.
 962                  * there might be a better way to deal with this???
 963                  */
 964                 sptlrpc_import_inval_all_ctx(export->exp_imp_reverse);
 965
 966                 destroy_import(export->exp_imp_reverse);
 967         }
 968
 969         /* for the rest part, we return -ENOTCONN in case of errors
 970          * in order to let client initialize connection again.
 971          */
 972         revimp = export->exp_imp_reverse = class_new_import(target);
 973         if (!revimp) {
 974                 CERROR("fail to alloc new reverse import.\n");
 975                 GOTO(out, rc = -ENOTCONN);
 976         }
 977
 978         revimp->imp_connection = ptlrpc_connection_addref(export->exp_connection);
 979         revimp->imp_client = &export->exp_obd->obd_ldlm_client;
 980         revimp->imp_remote_handle = conn;
 981         revimp->imp_dlm_fake = 1;
 982         revimp->imp_state = LUSTRE_IMP_FULL;
 983
 984         /* unknown versions will be caught in
 985          * ptlrpc_handle_server_req_in->lustre_unpack_msg() */
 986         revimp->imp_msg_magic = req->rq_reqmsg->lm_magic;
 987
 988         if ((export->exp_connect_flags & OBD_CONNECT_AT) &&
 989             (revimp->imp_msg_magic != LUSTRE_MSG_MAGIC_V1))
 990                 revimp->imp_msghdr_flags |= MSGHDR_AT_SUPPORT;
 991         else
 992                 revimp->imp_msghdr_flags &= ~MSGHDR_AT_SUPPORT;
 993
 994         rc = sptlrpc_import_sec_adapt(revimp, req->rq_svc_ctx,
 995                                       req->rq_flvr.sf_rpc);
 996         if (rc) {
 997                 CERROR("Failed to get sec for reverse import: %d\n", rc);
 998                 export->exp_imp_reverse = NULL;
 999                 class_destroy_import(revimp);
1000         }
1001
1002         class_import_put(revimp);
1003 out:
1004         if (export) {
1005                 spin_lock(&export->exp_lock);
1006                 export->exp_connecting = 0;
1007                 spin_unlock(&export->exp_lock);
1008         }
1009         if (targref)
1010                 class_decref(targref, __FUNCTION__, cfs_current());
1011         if (rc)
1012                 req->rq_status = rc;
1013         RETURN(rc);
1014 }
1015
1016 int target_handle_disconnect(struct ptlrpc_request *req)
1017 {
1018         int rc;
1019         ENTRY;
1020
1021         rc = req_capsule_server_pack(&req->rq_pill);
1022         if (rc)
1023                 RETURN(rc);
1024
1025         /* keep the rq_export around so we can send the reply */
1026         req->rq_status = obd_disconnect(class_export_get(req->rq_export));
1027
1028         RETURN(0);
1029 }
1030
1031 void target_destroy_export(struct obd_export *exp)
1032 {
1033         /* exports created from last_rcvd data, and "fake"
1034            exports created by lctl don't have an import */
1035         if (exp->exp_imp_reverse != NULL)
1036                 destroy_import(exp->exp_imp_reverse);
1037
1038         /* We cancel locks at disconnect time, but this will catch any locks
1039          * granted in a race with recovery-induced disconnect. */
1040         if (exp->exp_obd->obd_namespace != NULL)
1041                 ldlm_cancel_locks_for_export(exp);
1042 }
1043
1044 /*
1045  * Recovery functions
1046  */
1047
1048 struct ptlrpc_request *ptlrpc_clone_req( struct ptlrpc_request *orig_req)
1049 {
1050         struct ptlrpc_request *copy_req;
1051         struct lustre_msg *copy_reqmsg;
1052         struct ptlrpc_user_desc *udesc = NULL;
1053
1054         OBD_ALLOC_PTR(copy_req);
1055         if (!copy_req)
1056                 return NULL;
1057         OBD_ALLOC(copy_reqmsg, orig_req->rq_reqlen);
1058         if (!copy_reqmsg){
1059                 OBD_FREE_PTR(copy_req);
1060                 return NULL;
1061         }
1062
1063         if (orig_req->rq_user_desc) {
1064                 int ngroups = orig_req->rq_user_desc->pud_ngroups;
1065
1066                 OBD_ALLOC(udesc, sptlrpc_user_desc_size(ngroups));
1067                 if (!udesc) {
1068                         OBD_FREE(copy_reqmsg, orig_req->rq_reqlen);
1069                         OBD_FREE_PTR(copy_req);
1070                         return NULL;
1071                 }
1072                 memcpy(udesc, orig_req->rq_user_desc,
1073                        sptlrpc_user_desc_size(ngroups));
1074         }
1075
1076         *copy_req = *orig_req;
1077         memcpy(copy_reqmsg, orig_req->rq_reqmsg, orig_req->rq_reqlen);
1078         copy_req->rq_reqmsg = copy_reqmsg;
1079         copy_req->rq_user_desc = udesc;
1080
1081         class_export_rpc_get(copy_req->rq_export);
1082         CFS_INIT_LIST_HEAD(&copy_req->rq_list);
1083         CFS_INIT_LIST_HEAD(&copy_req->rq_replay_list);
1084         sptlrpc_svc_ctx_addref(copy_req);
1085
1086         if (copy_req->rq_reply_state) {
1087                 /* the copied req takes over the reply state */
1088                 orig_req->rq_reply_state = NULL;
1089                 /* to catch further access */
1090                 orig_req->rq_repmsg = NULL;
1091                 orig_req->rq_replen = 0;
1092         }
1093
1094         return copy_req;
1095 }
1096
1097 void ptlrpc_free_clone(struct ptlrpc_request *req)
1098 {
1099         LASSERT(list_empty(&req->rq_replay_list));
1100
1101         ptlrpc_req_drop_rs(req);
1102         sptlrpc_svc_ctx_decref(req);
1103         class_export_rpc_put(req->rq_export);
1104         list_del(&req->rq_list);
1105
1106         if (req->rq_user_desc) {
1107                 int ngroups = req->rq_user_desc->pud_ngroups;
1108                 OBD_FREE(req->rq_user_desc, sptlrpc_user_desc_size(ngroups));
1109         }
1110         OBD_FREE(req->rq_reqmsg, req->rq_reqlen);
1111         OBD_FREE_PTR(req);
1112 }
1113
1114 static int target_exp_enqueue_req_replay(struct ptlrpc_request *req)
1115 {
1116         __u64                  transno = lustre_msg_get_transno(req->rq_reqmsg);
1117         struct obd_export     *exp = req->rq_export;
1118         struct ptlrpc_request *reqiter;
1119         int                    dup = 0;
1120
1121         LASSERT(exp);
1122
1123         spin_lock(&exp->exp_lock);
1124         list_for_each_entry(reqiter, &exp->exp_req_replay_queue,
1125                             rq_replay_list) {
1126                 if (lustre_msg_get_transno(reqiter->rq_reqmsg) == transno) {
1127                         dup = 1;
1128                         break;
1129                 }
1130         }
1131
1132         if (dup) {
1133                 /* we expect it with RESENT and REPLAY flags */
1134                 if ((lustre_msg_get_flags(req->rq_reqmsg) &
1135                      (MSG_RESENT | MSG_REPLAY)) != (MSG_RESENT | MSG_REPLAY))
1136                         CERROR("invalid flags %x of resent replay\n",
1137                                lustre_msg_get_flags(req->rq_reqmsg));
1138         } else {
1139                 list_add_tail(&req->rq_replay_list, &exp->exp_req_replay_queue);
1140         }
1141
1142         spin_unlock(&exp->exp_lock);
1143         return dup;
1144 }
1145
1146 static void target_exp_dequeue_req_replay(struct ptlrpc_request *req)
1147 {
1148         LASSERT(!list_empty(&req->rq_replay_list));
1149         LASSERT(req->rq_export);
1150
1151         spin_lock(&req->rq_export->exp_lock);
1152         list_del_init(&req->rq_replay_list);
1153         spin_unlock(&req->rq_export->exp_lock);
1154 }
1155
1156 #ifdef __KERNEL__
1157 static void target_finish_recovery(struct obd_device *obd)
1158 {
1159         ENTRY;
1160         LCONSOLE_INFO("%s: sending delayed replies to recovered clients\n",
1161                       obd->obd_name);
1162
1163         ldlm_reprocess_all_ns(obd->obd_namespace);
1164
1165         /* when recovery finished, cleanup orphans on mds and ost */
1166         if (OBT(obd) && OBP(obd, postrecov)) {
1167                 int rc = OBP(obd, postrecov)(obd);
1168                 LCONSOLE_WARN("%s: recovery %s: rc %d\n", obd->obd_name,
1169                               rc < 0 ? "failed" : "complete", rc);
1170         }
1171
1172         obd->obd_recovery_end = cfs_time_current_sec();
1173         EXIT;
1174 }
1175
1176 static void abort_req_replay_queue(struct obd_device *obd)
1177 {
1178         struct ptlrpc_request *req, *n;
1179
1180         list_for_each_entry_safe(req, n, &obd->obd_req_replay_queue, rq_list) {
1181                 DEBUG_REQ(D_WARNING, req, "aborted:");
1182                 req->rq_status = -ENOTCONN;
1183                 if (ptlrpc_error(req)) {
1184                         DEBUG_REQ(D_ERROR, req,
1185                                   "failed abort_req_reply; skipping");
1186                 }
1187                 target_exp_dequeue_req_replay(req);
1188                 ptlrpc_free_clone(req);
1189         }
1190 }
1191
1192 static void abort_lock_replay_queue(struct obd_device *obd)
1193 {
1194         struct ptlrpc_request *req, *n;
1195
1196         list_for_each_entry_safe(req, n, &obd->obd_lock_replay_queue, rq_list){
1197                 DEBUG_REQ(D_ERROR, req, "aborted:");
1198                 req->rq_status = -ENOTCONN;
1199                 if (ptlrpc_error(req)) {
1200                         DEBUG_REQ(D_ERROR, req,
1201                                   "failed abort_lock_reply; skipping");
1202                 }
1203                 ptlrpc_free_clone(req);
1204         }
1205 }
1206 #endif
1207
1208 /* Called from a cleanup function if the device is being cleaned up
1209    forcefully.  The exports should all have been disconnected already,
1210    the only thing left to do is
1211      - clear the recovery flags
1212      - cancel the timer
1213      - free queued requests and replies, but don't send replies
1214    Because the obd_stopping flag is set, no new requests should be received.
1215
1216 */
1217 void target_cleanup_recovery(struct obd_device *obd)
1218 {
1219         struct ptlrpc_request *req, *n;
1220         ENTRY;
1221
1222         LASSERT(obd->obd_stopping);
1223
1224         spin_lock_bh(&obd->obd_processing_task_lock);
1225         if (!obd->obd_recovering) {
1226                 spin_unlock_bh(&obd->obd_processing_task_lock);
1227                 EXIT;
1228                 return;
1229         }
1230         obd->obd_recovering = obd->obd_abort_recovery = 0;
1231         target_cancel_recovery_timer(obd);
1232         spin_unlock_bh(&obd->obd_processing_task_lock);
1233
1234         list_for_each_entry_safe(req, n, &obd->obd_req_replay_queue, rq_list) {
1235                 LASSERT (req->rq_reply_state == 0);
1236                 target_exp_dequeue_req_replay(req);
1237                 ptlrpc_free_clone(req);
1238         }
1239         list_for_each_entry_safe(req, n, &obd->obd_lock_replay_queue, rq_list){
1240                 LASSERT (req->rq_reply_state == 0);
1241                 ptlrpc_free_clone(req);
1242         }
1243         list_for_each_entry_safe(req, n, &obd->obd_final_req_queue, rq_list) {
1244                 LASSERT (req->rq_reply_state == 0);
1245                 ptlrpc_free_clone(req);
1246         }
1247
1248         EXIT;
1249 }
1250
1251 /* obd_processing_task_lock should be held */
1252 void target_cancel_recovery_timer(struct obd_device *obd)
1253 {
1254         CDEBUG(D_HA, "%s: cancel recovery timer\n", obd->obd_name);
1255         cfs_timer_disarm(&obd->obd_recovery_timer);
1256 }
1257
1258 /* extend = 1 means require at least "duration" seconds left in the timer,
1259    extend = 0 means set the total duration (start_recovery_timer) */
1260 static void reset_recovery_timer(struct obd_device *obd, int duration,
1261                                  int extend)
1262 {
1263         cfs_time_t now = cfs_time_current_sec();
1264         cfs_duration_t left;
1265
1266         spin_lock_bh(&obd->obd_processing_task_lock);
1267         if (!obd->obd_recovering || obd->obd_abort_recovery) {
1268                 spin_unlock_bh(&obd->obd_processing_task_lock);
1269                 return;
1270         }
1271
1272         left = cfs_time_sub(obd->obd_recovery_end, now);
1273
1274         if (extend && (duration > left))
1275                 obd->obd_recovery_timeout += duration - left;
1276         else if (!extend && (duration > obd->obd_recovery_timeout))
1277                 /* Track the client's largest expected replay time */
1278                 obd->obd_recovery_timeout = duration;
1279 #ifdef CRAY_XT3
1280         /*
1281          * If total recovery time already exceed the
1282          * obd_recovery_max_time, then CRAY XT3 will
1283          * abort the recovery
1284          */
1285         if(obd->obd_recovery_timeout > obd->obd_recovery_max_time)
1286                 obd->obd_recovery_timeout = obd->obd_recovery_max_time;
1287 #endif
1288         obd->obd_recovery_end = obd->obd_recovery_start +
1289                                 obd->obd_recovery_timeout;
1290         if (!cfs_timer_is_armed(&obd->obd_recovery_timer) ||
1291             cfs_time_before(now, obd->obd_recovery_end)) {
1292                 left = cfs_time_sub(obd->obd_recovery_end, now);
1293                 cfs_timer_arm(&obd->obd_recovery_timer, cfs_time_shift(left));
1294         }
1295         spin_unlock_bh(&obd->obd_processing_task_lock);
1296         CDEBUG(D_HA, "%s: recovery timer will expire in %u seconds\n",
1297                obd->obd_name, (unsigned)left);
1298 }
1299
1300 static void check_and_start_recovery_timer(struct obd_device *obd)
1301 {
1302         spin_lock_bh(&obd->obd_processing_task_lock);
1303         if (cfs_timer_is_armed(&obd->obd_recovery_timer)) {
1304                 spin_unlock_bh(&obd->obd_processing_task_lock);
1305                 return;
1306         }
1307         CWARN("%s: starting recovery timer\n", obd->obd_name);
1308         obd->obd_recovery_start = cfs_time_current_sec();
1309         /* minimum */
1310         obd->obd_recovery_timeout = OBD_RECOVERY_FACTOR * obd_timeout;
1311         spin_unlock_bh(&obd->obd_processing_task_lock);
1312
1313         reset_recovery_timer(obd, obd->obd_recovery_timeout, 0);
1314 }
1315
1316 /* Reset the timer with each new client connection */
1317 /*
1318  * This timer is actually reconnect_timer, which is for making sure
1319  * the total recovery window is at least as big as my reconnect
1320  * attempt timing. So the initial recovery time_out will be set to
1321  * OBD_RECOVERY_FACTOR * obd_timeout. If the timeout coming
1322  * from client is bigger than this, then the recovery time_out will
1323  * be extend to make sure the client could be reconnected, in the
1324  * process, the timeout from the new client should be ignored.
1325  */
1326
1327 static void
1328 target_start_and_reset_recovery_timer(struct obd_device *obd,
1329                                       struct ptlrpc_request *req,
1330                                       int new_client)
1331 {
1332         int req_timeout = OBD_RECOVERY_FACTOR *
1333                           lustre_msg_get_timeout(req->rq_reqmsg);
1334
1335         check_and_start_recovery_timer(obd);
1336
1337         if (req_timeout > obd->obd_recovery_timeout && !new_client)
1338                 reset_recovery_timer(obd, req_timeout, 0);
1339 }
1340
1341 #ifdef __KERNEL__
1342 static int check_for_next_transno(struct obd_device *obd)
1343 {
1344         struct ptlrpc_request *req = NULL;
1345         int wake_up = 0, connected, completed, queue_len, max;
1346         __u64 next_transno, req_transno;
1347         ENTRY;
1348         spin_lock_bh(&obd->obd_processing_task_lock);
1349
1350         if (!list_empty(&obd->obd_req_replay_queue)) {
1351                 req = list_entry(obd->obd_req_replay_queue.next,
1352                                  struct ptlrpc_request, rq_list);
1353                 req_transno = lustre_msg_get_transno(req->rq_reqmsg);
1354         } else {
1355                 req_transno = 0;
1356         }
1357
1358         max = obd->obd_max_recoverable_clients;
1359         connected = obd->obd_connected_clients;
1360         completed = connected - obd->obd_recoverable_clients;
1361         queue_len = obd->obd_requests_queued_for_recovery;
1362         next_transno = obd->obd_next_recovery_transno;
1363
1364         CDEBUG(D_HA, "max: %d, connected: %d, completed: %d, queue_len: %d, "
1365                "req_transno: "LPU64", next_transno: "LPU64"\n",
1366                max, connected, completed, queue_len, req_transno, next_transno);
1367
1368         if (obd->obd_abort_recovery) {
1369                 CDEBUG(D_HA, "waking for aborted recovery\n");
1370                 wake_up = 1;
1371         } else if (atomic_read(&obd->obd_req_replay_clients) == 0) {
1372                 CDEBUG(D_HA, "waking for completed recovery\n");
1373                 wake_up = 1;
1374         } else if (req_transno == next_transno) {
1375                 CDEBUG(D_HA, "waking for next ("LPD64")\n", next_transno);
1376                 wake_up = 1;
1377         } else if (queue_len + completed == max) {
1378                 /* handle gaps occured due to lost reply. It is allowed gaps
1379                  * because all clients are connected and there will be resend
1380                  * for missed transaction */
1381                 LASSERTF(req_transno >= next_transno,
1382                          "req_transno: "LPU64", next_transno: "LPU64"\n",
1383                          req_transno, next_transno);
1384
1385                 CDEBUG(req_transno > obd->obd_last_committed ? D_ERROR : D_HA,
1386                        "waking for skipped transno (skip: "LPD64
1387                        ", ql: %d, comp: %d, conn: %d, next: "LPD64")\n",
1388                        next_transno, queue_len, completed, connected, req_transno);
1389                 obd->obd_next_recovery_transno = req_transno;
1390                 wake_up = 1;
1391         } else if (queue_len == atomic_read(&obd->obd_req_replay_clients)) {
1392                 /* some clients haven't connected in time, but we can try
1393                  * to replay requests that demand on already committed ones
1394                  * also, we can replay first non-committed transation */
1395                 LASSERT(req_transno != 0);
1396                 if (req_transno == obd->obd_last_committed + 1) {
1397                         obd->obd_next_recovery_transno = req_transno;
1398                 } else if (req_transno > obd->obd_last_committed) {
1399                         /* can't continue recovery: have no needed transno */
1400                         obd->obd_abort_recovery = 1;
1401                         CDEBUG(D_ERROR, "abort due to missed clients. max: %d, "
1402                                "connected: %d, completed: %d, queue_len: %d, "
1403                                "req_transno: "LPU64", next_transno: "LPU64"\n",
1404                                max, connected, completed, queue_len,
1405                                req_transno, next_transno);
1406                 }
1407                 wake_up = 1;
1408         }
1409
1410         spin_unlock_bh(&obd->obd_processing_task_lock);
1411         return wake_up;
1412 }
1413
1414 static struct ptlrpc_request *target_next_replay_req(struct obd_device *obd)
1415 {
1416         struct l_wait_info lwi = { 0 };
1417         struct ptlrpc_request *req;
1418
1419         CDEBUG(D_HA, "Waiting for transno "LPD64"\n",
1420                obd->obd_next_recovery_transno);
1421         l_wait_event(obd->obd_next_transno_waitq,
1422                      check_for_next_transno(obd), &lwi);
1423
1424         spin_lock_bh(&obd->obd_processing_task_lock);
1425         if (obd->obd_abort_recovery) {
1426                 req = NULL;
1427         } else if (!list_empty(&obd->obd_req_replay_queue)) {
1428                 req = list_entry(obd->obd_req_replay_queue.next,
1429                                  struct ptlrpc_request, rq_list);
1430                 target_exp_dequeue_req_replay(req);
1431                 list_del_init(&req->rq_list);
1432                 obd->obd_requests_queued_for_recovery--;
1433         } else {
1434                 req = NULL;
1435         }
1436         spin_unlock_bh(&obd->obd_processing_task_lock);
1437         RETURN(req);
1438 }
1439
1440 static int check_for_next_lock(struct obd_device *obd)
1441 {
1442         struct ptlrpc_request *req = NULL;
1443         int wake_up = 0;
1444
1445         spin_lock_bh(&obd->obd_processing_task_lock);
1446         if (!list_empty(&obd->obd_lock_replay_queue)) {
1447                 req = list_entry(obd->obd_lock_replay_queue.next,
1448                                  struct ptlrpc_request, rq_list);
1449                 CDEBUG(D_HA, "waking for next lock\n");
1450                 wake_up = 1;
1451         } else if (atomic_read(&obd->obd_lock_replay_clients) == 0) {
1452                 CDEBUG(D_HA, "waking for completed lock replay\n");
1453                 wake_up = 1;
1454         } else if (obd->obd_abort_recovery) {
1455                 CDEBUG(D_HA, "waking for aborted recovery\n");
1456                 wake_up = 1;
1457         }
1458         spin_unlock_bh(&obd->obd_processing_task_lock);
1459
1460         return wake_up;
1461 }
1462
1463 static struct ptlrpc_request *target_next_replay_lock(struct obd_device *obd)
1464 {
1465         struct l_wait_info lwi = { 0 };
1466         struct ptlrpc_request *req;
1467
1468         CDEBUG(D_HA, "Waiting for lock\n");
1469         l_wait_event(obd->obd_next_transno_waitq,
1470                      check_for_next_lock(obd), &lwi);
1471
1472         spin_lock_bh(&obd->obd_processing_task_lock);
1473         if (obd->obd_abort_recovery) {
1474                 req = NULL;
1475         } else if (!list_empty(&obd->obd_lock_replay_queue)) {
1476                 req = list_entry(obd->obd_lock_replay_queue.next,
1477                                  struct ptlrpc_request, rq_list);
1478                 list_del_init(&req->rq_list);
1479         } else {
1480                 req = NULL;
1481         }
1482         spin_unlock_bh(&obd->obd_processing_task_lock);
1483         return req;
1484 }
1485
1486 static struct ptlrpc_request *target_next_final_ping(struct obd_device *obd)
1487 {
1488         struct ptlrpc_request *req;
1489
1490         spin_lock_bh(&obd->obd_processing_task_lock);
1491         if (!list_empty(&obd->obd_final_req_queue)) {
1492                 req = list_entry(obd->obd_final_req_queue.next,
1493                                  struct ptlrpc_request, rq_list);
1494                 list_del_init(&req->rq_list);
1495         } else {
1496                 req = NULL;
1497         }
1498         spin_unlock_bh(&obd->obd_processing_task_lock);
1499         return req;
1500 }
1501
1502 static inline int req_replay_done(struct obd_export *exp)
1503 {
1504         return (exp->exp_req_replay_needed == 0);
1505 }
1506
1507 static inline int lock_replay_done(struct obd_export *exp)
1508 {
1509         return (exp->exp_lock_replay_needed == 0);
1510 }
1511
1512 static inline int connect_done(struct obd_export *exp)
1513 {
1514         return (exp->exp_in_recovery != 0);
1515 }
1516
1517 static int check_for_clients(struct obd_device *obd)
1518 {
1519         if (obd->obd_abort_recovery)
1520                 return 1;
1521         LASSERT(obd->obd_connected_clients <= obd->obd_max_recoverable_clients);
1522         if (obd->obd_no_conn == 0 &&
1523             obd->obd_connected_clients == obd->obd_max_recoverable_clients)
1524                 return 1;
1525         return 0;
1526 }
1527
1528 static int handle_recovery_req(struct ptlrpc_thread *thread,
1529                                struct ptlrpc_request *req,
1530                                svc_handler_t handler)
1531 {
1532         int rc;
1533         ENTRY;
1534
1535         rc = lu_context_init(&req->rq_session, LCT_SESSION);
1536         if (rc) {
1537                 CERROR("Failure to initialize session: %d\n", rc);
1538                 return rc;
1539         }
1540         req->rq_session.lc_thread = thread;
1541         lu_context_enter(&req->rq_session);
1542         req->rq_svc_thread = thread;
1543         req->rq_svc_thread->t_env->le_ses = &req->rq_session;
1544
1545         /* thread context */
1546         lu_context_enter(&thread->t_env->le_ctx);
1547         (void)handler(req);
1548         lu_context_exit(&thread->t_env->le_ctx);
1549
1550         lu_context_exit(&req->rq_session);
1551         lu_context_fini(&req->rq_session);
1552         /* don't reset timer for final stage */
1553         if (!req_replay_done(req->rq_export) ||
1554             !lock_replay_done(req->rq_export))
1555                 reset_recovery_timer(class_exp2obd(req->rq_export),
1556                        OBD_RECOVERY_FACTOR * AT_OFF ? obd_timeout :
1557                        at_get(&req->rq_rqbd->rqbd_service->srv_at_estimate), 1);
1558         ptlrpc_free_clone(req);
1559         RETURN(0);
1560 }
1561
1562 static void resume_recovery_timer(struct obd_device *obd)
1563 {
1564         /* to be safe, make it at least OBD_RECOVERY_FACTOR * obd_timeout */
1565         reset_recovery_timer(obd, OBD_RECOVERY_FACTOR * obd_timeout, 1);
1566 }
1567
1568 static int target_recovery_thread(void *arg)
1569 {
1570         struct obd_device *obd = arg;
1571         struct ptlrpc_request *req;
1572         struct target_recovery_data *trd = &obd->obd_recovery_data;
1573         struct l_wait_info lwi = { 0 };
1574         unsigned long delta;
1575         unsigned long flags;
1576         struct lu_env env;
1577         struct ptlrpc_thread fake_svc_thread, *thread = &fake_svc_thread;
1578         __u32 recov_ctx_tags = LCT_MD_THREAD;
1579         int rc = 0;
1580         ENTRY;
1581
1582         cfs_daemonize("tgt_recov");
1583
1584         SIGNAL_MASK_LOCK(current, flags);
1585         sigfillset(&current->blocked);
1586         RECALC_SIGPENDING;
1587         SIGNAL_MASK_UNLOCK(current, flags);
1588
1589         rc = lu_context_init(&env.le_ctx, recov_ctx_tags);
1590         if (rc)
1591                 RETURN(rc);
1592
1593         thread->t_env = &env;
1594         env.le_ctx.lc_thread = thread;
1595
1596         CERROR("%s: started recovery thread pid %d\n", obd->obd_name,
1597                current->pid);
1598         trd->trd_processing_task = current->pid;
1599
1600         obd->obd_recovering = 1;
1601         complete(&trd->trd_starting);
1602
1603         /* first of all, we have to know the first transno to replay */
1604         obd->obd_abort_recovery = 0;
1605         l_wait_event(obd->obd_next_transno_waitq,
1606                      check_for_clients(obd), &lwi);
1607
1608         spin_lock_bh(&obd->obd_processing_task_lock);
1609         target_cancel_recovery_timer(obd);
1610         spin_unlock_bh(&obd->obd_processing_task_lock);
1611
1612         /* If some clients haven't connected in time, evict them */
1613         if (obd->obd_abort_recovery) {
1614                 CWARN("Some clients haven't connect in time (%d/%d),"
1615                        "evict them\n", obd->obd_connected_clients,
1616                        obd->obd_max_recoverable_clients);
1617                 obd->obd_abort_recovery = obd->obd_stopping;
1618                 class_disconnect_stale_exports(obd, connect_done);
1619         }
1620         /* next stage: replay requests */
1621         delta = jiffies;
1622         obd->obd_req_replaying = 1;
1623         CDEBUG(D_INFO, "1: request replay stage - %d clients from t"LPU64"\n",
1624               atomic_read(&obd->obd_req_replay_clients),
1625               obd->obd_next_recovery_transno);
1626         resume_recovery_timer(obd);
1627         while ((req = target_next_replay_req(obd))) {
1628                 LASSERT(trd->trd_processing_task == current->pid);
1629                 DEBUG_REQ(D_HA, req, "processing t"LPD64" from %s",
1630                           lustre_msg_get_transno(req->rq_reqmsg),
1631                           libcfs_nid2str(req->rq_peer.nid));
1632                 handle_recovery_req(thread, req,
1633                                     trd->trd_recovery_handler);
1634                 obd->obd_replayed_requests++;
1635                 spin_lock_bh(&obd->obd_processing_task_lock);
1636                 obd->obd_next_recovery_transno++;
1637                 spin_unlock_bh(&obd->obd_processing_task_lock);
1638         }
1639
1640         spin_lock_bh(&obd->obd_processing_task_lock);
1641         target_cancel_recovery_timer(obd);
1642         spin_unlock_bh(&obd->obd_processing_task_lock);
1643
1644         /* If some clients haven't replayed requests in time, evict them */
1645         if (obd->obd_abort_recovery) {
1646                 CDEBUG(D_ERROR, "req replay timed out, aborting ...\n");
1647                 obd->obd_abort_recovery = obd->obd_stopping;
1648                 class_disconnect_stale_exports(obd, req_replay_done);
1649                 abort_req_replay_queue(obd);
1650         }
1651
1652         /* The second stage: replay locks */
1653         CDEBUG(D_INFO, "2: lock replay stage - %d clients\n",
1654                atomic_read(&obd->obd_lock_replay_clients));
1655         resume_recovery_timer(obd);
1656         while ((req = target_next_replay_lock(obd))) {
1657                 LASSERT(trd->trd_processing_task == current->pid);
1658                 DEBUG_REQ(D_HA|D_WARNING, req, "processing lock from %s: ",
1659                           libcfs_nid2str(req->rq_peer.nid));
1660                 handle_recovery_req(thread, req,
1661                                     trd->trd_recovery_handler);
1662                 obd->obd_replayed_locks++;
1663         }
1664
1665         spin_lock_bh(&obd->obd_processing_task_lock);
1666         target_cancel_recovery_timer(obd);
1667         spin_unlock_bh(&obd->obd_processing_task_lock);
1668         /* If some clients haven't replayed requests in time, evict them */
1669         if (obd->obd_abort_recovery) {
1670                 int stale;
1671                 CERROR("lock replay timed out, aborting ...\n");
1672                 obd->obd_abort_recovery = obd->obd_stopping;
1673                 stale = class_disconnect_stale_exports(obd, lock_replay_done);
1674                 abort_lock_replay_queue(obd);
1675         }
1676
1677         /* We drop recoverying flag to forward all new requests
1678          * to regular mds_handle() since now */
1679         spin_lock_bh(&obd->obd_processing_task_lock);
1680         obd->obd_recovering = obd->obd_abort_recovery = 0;
1681         spin_unlock_bh(&obd->obd_processing_task_lock);
1682         /* The third stage: reply on final pings */
1683         CDEBUG(D_INFO, "3: final stage - process recovery completion pings\n");
1684         while ((req = target_next_final_ping(obd))) {
1685                 LASSERT(trd->trd_processing_task == current->pid);
1686                 DEBUG_REQ(D_HA, req, "processing final ping from %s: ",
1687                           libcfs_nid2str(req->rq_peer.nid));
1688                 handle_recovery_req(thread, req,
1689                                     trd->trd_recovery_handler);
1690         }
1691
1692         delta = (jiffies - delta) / HZ;
1693         CDEBUG(D_INFO,"4: recovery completed in %lus - %d/%d reqs/locks\n",
1694               delta, obd->obd_replayed_requests, obd->obd_replayed_locks);
1695         LASSERT(atomic_read(&obd->obd_req_replay_clients) == 0);
1696         LASSERT(atomic_read(&obd->obd_lock_replay_clients) == 0);
1697         if (delta > obd_timeout * 2) {
1698                 CWARN("too long recovery - read logs\n");
1699                 libcfs_debug_dumplog();
1700         }
1701
1702         target_finish_recovery(obd);
1703
1704         lu_context_fini(&env.le_ctx);
1705         trd->trd_processing_task = 0;
1706         complete(&trd->trd_finishing);
1707         RETURN(rc);
1708 }
1709
1710 int target_start_recovery_thread(struct obd_device *obd, svc_handler_t handler)
1711 {
1712         int rc = 0;
1713         struct target_recovery_data *trd = &obd->obd_recovery_data;
1714
1715         memset(trd, 0, sizeof(*trd));
1716         init_completion(&trd->trd_starting);
1717         init_completion(&trd->trd_finishing);
1718         trd->trd_recovery_handler = handler;
1719
1720         if (kernel_thread(target_recovery_thread, obd, 0) > 0) {
1721                 wait_for_completion(&trd->trd_starting);
1722                 LASSERT(obd->obd_recovering != 0);
1723         } else
1724                 rc = -ECHILD;
1725
1726         return rc;
1727 }
1728
1729 void target_stop_recovery_thread(struct obd_device *obd)
1730 {
1731         spin_lock_bh(&obd->obd_processing_task_lock);
1732         if (obd->obd_recovery_data.trd_processing_task > 0) {
1733                 struct target_recovery_data *trd = &obd->obd_recovery_data;
1734                 CERROR("%s: Aborting recovery\n", obd->obd_name);
1735                 obd->obd_abort_recovery = 1;
1736                 wake_up(&obd->obd_next_transno_waitq);
1737                 spin_unlock_bh(&obd->obd_processing_task_lock);
1738                 wait_for_completion(&trd->trd_finishing);
1739         } else {
1740                 spin_unlock_bh(&obd->obd_processing_task_lock);
1741         }
1742 }
1743
1744 void target_recovery_fini(struct obd_device *obd)
1745 {
1746         class_disconnect_exports(obd);
1747         target_stop_recovery_thread(obd);
1748         target_cleanup_recovery(obd);
1749 }
1750 EXPORT_SYMBOL(target_recovery_fini);
1751
1752 static void target_recovery_expired(unsigned long castmeharder)
1753 {
1754         struct obd_device *obd = (struct obd_device *)castmeharder;
1755         LCONSOLE_WARN("%s: recovery timed out; %d clients never reconnected "
1756                       "after %lds (%d clients did)\n",
1757                       obd->obd_name, obd->obd_recoverable_clients,
1758                       cfs_time_current_sec()- obd->obd_recovery_start,
1759                       obd->obd_connected_clients);
1760         spin_lock_bh(&obd->obd_processing_task_lock);
1761         if (obd->obd_recovering)
1762                 obd->obd_abort_recovery = 1;
1763         cfs_waitq_signal(&obd->obd_next_transno_waitq);
1764         spin_unlock_bh(&obd->obd_processing_task_lock);
1765 }
1766
1767 void target_recovery_init(struct obd_device *obd, svc_handler_t handler)
1768 {
1769         if (obd->obd_max_recoverable_clients == 0)
1770                 return;
1771
1772         CWARN("RECOVERY: service %s, %d recoverable clients, "
1773               "last_transno "LPU64"\n", obd->obd_name,
1774               obd->obd_max_recoverable_clients, obd->obd_last_committed);
1775         obd->obd_next_recovery_transno = obd->obd_last_committed + 1;
1776         obd->obd_recovery_start = 0;
1777         obd->obd_recovery_end = 0;
1778         obd->obd_recovery_timeout = OBD_RECOVERY_FACTOR * obd_timeout;
1779         /* bz13079: this should be set to desired value for ost but not for mds */
1780         obd->obd_recovery_max_time = OBD_RECOVERY_MAX_TIME;
1781         cfs_timer_init(&obd->obd_recovery_timer, target_recovery_expired, obd);
1782         target_start_recovery_thread(obd, handler);
1783 }
1784 EXPORT_SYMBOL(target_recovery_init);
1785
1786 #endif
1787
1788 int target_process_req_flags(struct obd_device *obd, struct ptlrpc_request *req)
1789 {
1790         struct obd_export *exp = req->rq_export;
1791         LASSERT(exp != NULL);
1792         if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REQ_REPLAY_DONE) {
1793                 /* client declares he's ready to replay locks */
1794                 spin_lock_bh(&obd->obd_processing_task_lock);
1795                 if (exp->exp_req_replay_needed) {
1796                         LASSERT(atomic_read(&obd->obd_req_replay_clients) > 0);
1797                         spin_lock(&exp->exp_lock);
1798                         exp->exp_req_replay_needed = 0;
1799                         spin_unlock(&exp->exp_lock);
1800                         atomic_dec(&obd->obd_req_replay_clients);
1801                         LASSERT(obd->obd_recoverable_clients > 0);
1802                         obd->obd_recoverable_clients--;
1803                         if (atomic_read(&obd->obd_req_replay_clients) == 0)
1804                                 CDEBUG(D_HA, "all clients have replayed reqs\n");
1805                         wake_up(&obd->obd_next_transno_waitq);
1806                 }
1807                 spin_unlock_bh(&obd->obd_processing_task_lock);
1808         }
1809         if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_LOCK_REPLAY_DONE) {
1810                 /* client declares he's ready to complete recovery
1811                  * so, we put the request on th final queue */
1812                 spin_lock_bh(&obd->obd_processing_task_lock);
1813                 if (exp->exp_lock_replay_needed) {
1814                         LASSERT(atomic_read(&obd->obd_lock_replay_clients) > 0);
1815                         spin_lock(&exp->exp_lock);
1816                         exp->exp_lock_replay_needed = 0;
1817                         spin_unlock(&exp->exp_lock);
1818                         atomic_dec(&obd->obd_lock_replay_clients);
1819                         if (atomic_read(&obd->obd_lock_replay_clients) == 0)
1820                                 CDEBUG(D_HA, "all clients have replayed locks\n");
1821                         wake_up(&obd->obd_next_transno_waitq);
1822                 }
1823                 spin_unlock_bh(&obd->obd_processing_task_lock);
1824         }
1825
1826         return 0;
1827 }
1828
1829 int target_queue_recovery_request(struct ptlrpc_request *req,
1830                                   struct obd_device *obd)
1831 {
1832         struct list_head *tmp;
1833         int inserted = 0;
1834         __u64 transno = lustre_msg_get_transno(req->rq_reqmsg);
1835
1836         ENTRY;
1837
1838         if (obd->obd_recovery_data.trd_processing_task == cfs_curproc_pid()) {
1839                 /* Processing the queue right now, don't re-add. */
1840                 RETURN(1);
1841         }
1842
1843         target_process_req_flags(obd, req);
1844
1845         if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_LOCK_REPLAY_DONE) {
1846                 /* client declares he's ready to complete recovery
1847                  * so, we put the request on th final queue */
1848                 req = ptlrpc_clone_req(req);
1849                 if (req == NULL)
1850                         RETURN(-ENOMEM);
1851                 DEBUG_REQ(D_HA, req, "queue final req");
1852                 spin_lock_bh(&obd->obd_processing_task_lock);
1853                 if (obd->obd_recovering)
1854                         list_add_tail(&req->rq_list, &obd->obd_final_req_queue);
1855                 else {
1856                         spin_unlock_bh(&obd->obd_processing_task_lock);
1857                         ptlrpc_free_clone(req);
1858                         if (obd->obd_stopping) {
1859                                 RETURN(-ENOTCONN);
1860                         } else {
1861                                 RETURN(1);
1862                         }
1863                 }
1864                 spin_unlock_bh(&obd->obd_processing_task_lock);
1865                 RETURN(0);
1866         }
1867         if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REQ_REPLAY_DONE) {
1868                 /* client declares he's ready to replay locks */
1869                 req = ptlrpc_clone_req(req);
1870                 if (req == NULL)
1871                         RETURN(-ENOMEM);
1872                 DEBUG_REQ(D_HA, req, "queue lock replay req");
1873                 spin_lock_bh(&obd->obd_processing_task_lock);
1874                 LASSERT(obd->obd_recovering);
1875                 /* usually due to recovery abort */
1876                 if (!req->rq_export->exp_in_recovery) {
1877                         spin_unlock_bh(&obd->obd_processing_task_lock);
1878                         ptlrpc_free_clone(req);
1879                         RETURN(-ENOTCONN);
1880                 }
1881                 LASSERT(req->rq_export->exp_lock_replay_needed);
1882                 list_add_tail(&req->rq_list, &obd->obd_lock_replay_queue);
1883                 spin_unlock_bh(&obd->obd_processing_task_lock);
1884                 wake_up(&obd->obd_next_transno_waitq);
1885                 RETURN(0);
1886         }
1887
1888         /* CAVEAT EMPTOR: The incoming request message has been swabbed
1889          * (i.e. buflens etc are in my own byte order), but type-dependent
1890          * buffers (eg mds_body, ost_body etc) have NOT been swabbed. */
1891
1892         if (!transno) {
1893                 CFS_INIT_LIST_HEAD(&req->rq_list);
1894                 DEBUG_REQ(D_HA, req, "not queueing");
1895                 RETURN(1);
1896         }
1897
1898         spin_lock_bh(&obd->obd_processing_task_lock);
1899
1900         /* If we're processing the queue, we want don't want to queue this
1901          * message.
1902          *
1903          * Also, if this request has a transno less than the one we're waiting
1904          * for, we should process it now.  It could (and currently always will)
1905          * be an open request for a descriptor that was opened some time ago.
1906          *
1907          * Also, a resent, replayed request that has already been
1908          * handled will pass through here and be processed immediately.
1909          */
1910         CWARN("Next recovery transno: "LPU64", current: "LPU64", replaying: %i\n",
1911               obd->obd_next_recovery_transno, transno, obd->obd_req_replaying);
1912         if (transno < obd->obd_next_recovery_transno && obd->obd_req_replaying) {
1913                 /* Processing the queue right now, don't re-add. */
1914                 LASSERT(list_empty(&req->rq_list));
1915                 spin_unlock_bh(&obd->obd_processing_task_lock);
1916                 RETURN(1);
1917         }
1918         spin_unlock_bh(&obd->obd_processing_task_lock);
1919
1920         if (OBD_FAIL_CHECK(OBD_FAIL_TGT_REPLAY_DROP))
1921                 RETURN(0);
1922
1923         req = ptlrpc_clone_req(req);
1924         if (req == NULL)
1925                 RETURN(-ENOMEM);
1926
1927         spin_lock_bh(&obd->obd_processing_task_lock);
1928         LASSERT(obd->obd_recovering);
1929         if (!req->rq_export->exp_in_recovery) {
1930                 spin_unlock_bh(&obd->obd_processing_task_lock);
1931                 ptlrpc_free_clone(req);
1932                 RETURN(-ENOTCONN);
1933         }
1934         LASSERT(req->rq_export->exp_req_replay_needed);
1935
1936         if (target_exp_enqueue_req_replay(req)) {
1937                 spin_unlock_bh(&obd->obd_processing_task_lock);
1938                 DEBUG_REQ(D_ERROR, req, "dropping resent queued req");
1939                 ptlrpc_free_clone(req);
1940                 RETURN(0);
1941         }
1942
1943         /* XXX O(n^2) */
1944         list_for_each(tmp, &obd->obd_req_replay_queue) {
1945                 struct ptlrpc_request *reqiter =
1946                         list_entry(tmp, struct ptlrpc_request, rq_list);
1947
1948                 if (lustre_msg_get_transno(reqiter->rq_reqmsg) > transno) {
1949                         list_add_tail(&req->rq_list, &reqiter->rq_list);
1950                         inserted = 1;
1951                         break;
1952                 }
1953
1954                 if (unlikely(lustre_msg_get_transno(reqiter->rq_reqmsg) ==
1955                              transno)) {
1956                         DEBUG_REQ(D_ERROR, req, "dropping replay: transno "
1957                                   "has been claimed by another client");
1958                         spin_unlock_bh(&obd->obd_processing_task_lock);
1959                         target_exp_dequeue_req_replay(req);
1960                         ptlrpc_free_clone(req);
1961                         RETURN(0);
1962                 }
1963         }
1964
1965         if (!inserted)
1966                 list_add_tail(&req->rq_list, &obd->obd_req_replay_queue);
1967
1968         obd->obd_requests_queued_for_recovery++;
1969         wake_up(&obd->obd_next_transno_waitq);
1970         spin_unlock_bh(&obd->obd_processing_task_lock);
1971         RETURN(0);
1972 }
1973
1974 struct obd_device * target_req2obd(struct ptlrpc_request *req)
1975 {
1976         return req->rq_export->exp_obd;
1977 }
1978
1979 static inline struct ldlm_pool *ldlm_exp2pl(struct obd_export *exp)
1980 {
1981         LASSERT(exp != NULL);
1982         return &exp->exp_obd->obd_namespace->ns_pool;
1983 }
1984
1985 /**
1986  * Packs current SLV and Limit into \a req.
1987  */
1988 int target_pack_pool_reply(struct ptlrpc_request *req)
1989 {
1990         struct obd_device *obd;
1991         ENTRY;
1992
1993         /*
1994          * Check that we still have all structures alive as this may
1995          * be some late rpc in shutdown time.
1996          */
1997         if (unlikely(!req->rq_export || !req->rq_export->exp_obd ||
1998                      !exp_connect_lru_resize(req->rq_export))) {
1999                 lustre_msg_set_slv(req->rq_repmsg, 0);
2000                 lustre_msg_set_limit(req->rq_repmsg, 0);
2001                 RETURN(0);
2002         }
2003
2004         /*
2005          * OBD is alive here as export is alive, which we checked above.
2006          */
2007         obd = req->rq_export->exp_obd;
2008
2009         read_lock(&obd->obd_pool_lock);
2010         lustre_msg_set_slv(req->rq_repmsg, obd->obd_pool_slv);
2011         lustre_msg_set_limit(req->rq_repmsg, obd->obd_pool_limit);
2012         read_unlock(&obd->obd_pool_lock);
2013
2014         RETURN(0);
2015 }
2016
2017 int target_send_reply_msg(struct ptlrpc_request *req, int rc, int fail_id)
2018 {
2019         if (OBD_FAIL_CHECK_ORSET(fail_id & ~OBD_FAIL_ONCE, OBD_FAIL_ONCE)) {
2020                 DEBUG_REQ(D_ERROR, req, "dropping reply");
2021                 return (-ECOMM);
2022         }
2023
2024         if (unlikely(rc)) {
2025                 DEBUG_REQ(D_ERROR, req, "processing error (%d)", rc);
2026                 req->rq_status = rc;
2027                 return (ptlrpc_send_error(req, 1));
2028         } else {
2029                 DEBUG_REQ(D_NET, req, "sending reply");
2030         }
2031
2032         return (ptlrpc_send_reply(req, PTLRPC_REPLY_MAYBE_DIFFICULT));
2033 }
2034
2035 void target_send_reply(struct ptlrpc_request *req, int rc, int fail_id)
2036 {
2037         int                        netrc;
2038         struct ptlrpc_reply_state *rs;
2039         struct obd_device         *obd;
2040         struct obd_export         *exp;
2041         struct ptlrpc_service     *svc;
2042
2043         if (req->rq_no_reply)
2044                 return;
2045
2046         svc = req->rq_rqbd->rqbd_service;
2047         rs = req->rq_reply_state;
2048         if (rs == NULL || !rs->rs_difficult) {
2049                 /* no notifiers */
2050                 target_send_reply_msg (req, rc, fail_id);
2051                 return;
2052         }
2053
2054         /* must be an export if locks saved */
2055         LASSERT (req->rq_export != NULL);
2056         /* req/reply consistent */
2057         LASSERT (rs->rs_service == svc);
2058
2059         /* "fresh" reply */
2060         LASSERT (!rs->rs_scheduled);
2061         LASSERT (!rs->rs_scheduled_ever);
2062         LASSERT (!rs->rs_handled);
2063         LASSERT (!rs->rs_on_net);
2064         LASSERT (rs->rs_export == NULL);
2065         LASSERT (list_empty(&rs->rs_obd_list));
2066         LASSERT (list_empty(&rs->rs_exp_list));
2067
2068         exp = class_export_get (req->rq_export);
2069         obd = exp->exp_obd;
2070
2071         /* disable reply scheduling onto srv_reply_queue while I'm setting up */
2072         rs->rs_scheduled = 1;
2073         rs->rs_on_net    = 1;
2074         rs->rs_xid       = req->rq_xid;
2075         rs->rs_transno   = req->rq_transno;
2076         rs->rs_export    = exp;
2077
2078         spin_lock(&obd->obd_uncommitted_replies_lock);
2079
2080         if (rs->rs_transno > obd->obd_last_committed) {
2081                 /* not committed already */
2082                 list_add_tail (&rs->rs_obd_list,
2083                                &obd->obd_uncommitted_replies);
2084         }
2085
2086         spin_unlock (&obd->obd_uncommitted_replies_lock);
2087         spin_lock (&exp->exp_lock);
2088
2089         list_add_tail (&rs->rs_exp_list, &exp->exp_outstanding_replies);
2090
2091         spin_unlock(&exp->exp_lock);
2092
2093         netrc = target_send_reply_msg (req, rc, fail_id);
2094
2095         spin_lock(&svc->srv_lock);
2096
2097         svc->srv_n_difficult_replies++;
2098
2099         if (netrc != 0) {
2100                 /* error sending: reply is off the net.  Also we need +1
2101                  * reply ref until ptlrpc_server_handle_reply() is done
2102                  * with the reply state (if the send was successful, there
2103                  * would have been +1 ref for the net, which
2104                  * reply_out_callback leaves alone) */
2105                 rs->rs_on_net = 0;
2106                 ptlrpc_rs_addref(rs);
2107                 atomic_inc (&svc->srv_outstanding_replies);
2108         }
2109
2110         if (!rs->rs_on_net ||                   /* some notifier */
2111             list_empty(&rs->rs_exp_list) ||     /* completed already */
2112             list_empty(&rs->rs_obd_list)) {
2113                 list_add_tail (&rs->rs_list, &svc->srv_reply_queue);
2114                 cfs_waitq_signal (&svc->srv_waitq);
2115         } else {
2116                 list_add (&rs->rs_list, &svc->srv_active_replies);
2117                 rs->rs_scheduled = 0;           /* allow notifier to schedule */
2118         }
2119
2120         spin_unlock(&svc->srv_lock);
2121 }
2122
2123 int target_handle_ping(struct ptlrpc_request *req)
2124 {
2125         obd_ping(req->rq_export);
2126         return req_capsule_server_pack(&req->rq_pill);
2127 }
2128
2129 void target_committed_to_req(struct ptlrpc_request *req)
2130 {
2131         struct obd_device *obd;
2132
2133         if (req == NULL || req->rq_export == NULL)
2134                 return;
2135
2136         obd = req->rq_export->exp_obd;
2137         if (obd == NULL)
2138                 return;
2139
2140         if (!obd->obd_no_transno && req->rq_repmsg != NULL)
2141                 lustre_msg_set_last_committed(req->rq_repmsg,
2142                                               obd->obd_last_committed);
2143         else
2144                 DEBUG_REQ(D_IOCTL, req, "not sending last_committed update (%d/"
2145                           "%d)", obd->obd_no_transno, req->rq_repmsg == NULL);
2146
2147         CDEBUG(D_INFO, "last_committed "LPU64", transno "LPU64", xid "LPU64"\n",
2148                obd->obd_last_committed, req->rq_transno, req->rq_xid);
2149 }
2150
2151 EXPORT_SYMBOL(target_committed_to_req);
2152
2153 #ifdef HAVE_QUOTA_SUPPORT
2154 int target_handle_qc_callback(struct ptlrpc_request *req)
2155 {
2156         struct obd_quotactl *oqctl;
2157         struct client_obd *cli = &req->rq_export->exp_obd->u.cli;
2158
2159         oqctl = req_capsule_client_get(&req->rq_pill, &RMF_OBD_QUOTACTL);
2160         if (oqctl == NULL)
2161                 RETURN(-EPROTO);
2162
2163         cli->cl_qchk_stat = oqctl->qc_stat;
2164
2165         return 0;
2166 }
2167
2168 int target_handle_dqacq_callback(struct ptlrpc_request *req)
2169 {
2170 #ifdef __KERNEL__
2171         struct obd_device *obd = req->rq_export->exp_obd;
2172         struct obd_device *master_obd;
2173         struct lustre_quota_ctxt *qctxt;
2174         struct qunit_data *qdata;
2175         void* rep;
2176         struct qunit_data_old *qdata_old;
2177         int rc = 0;
2178         ENTRY;
2179
2180         rc = req_capsule_server_pack(&req->rq_pill);
2181         if (rc) {
2182                 CERROR("packing reply failed!: rc = %d\n", rc);
2183                 RETURN(rc);
2184         }
2185
2186         LASSERT(req->rq_export);
2187
2188         /* fixed for bug10707 */
2189         if ((req->rq_export->exp_connect_flags & OBD_CONNECT_QUOTA64) &&
2190             !OBD_FAIL_CHECK(OBD_FAIL_QUOTA_QD_COUNT_32BIT)) {
2191                 CDEBUG(D_QUOTA, "qd_count is 64bit!\n");
2192                 rep = req_capsule_server_get(&req->rq_pill,
2193                                              &RMF_QUNIT_DATA);
2194                 LASSERT(rep);
2195                 qdata = req_capsule_client_swab_get(&req->rq_pill,
2196                                                     &RMF_QUNIT_DATA,
2197                                           (void*)lustre_swab_qdata);
2198         } else {
2199                 CDEBUG(D_QUOTA, "qd_count is 32bit!\n");
2200                 rep = req_capsule_server_get(&req->rq_pill, &RMF_QUNIT_DATA);
2201                 LASSERT(rep);
2202                 qdata_old = req_capsule_client_swab_get(&req->rq_pill,
2203                                                         &RMF_QUNIT_DATA,
2204                                            (void*)lustre_swab_qdata_old);
2205                 qdata = lustre_quota_old_to_new(qdata_old);
2206         }
2207
2208         if (qdata == NULL)
2209                 RETURN(-EPROTO);
2210
2211         /* we use the observer */
2212         LASSERT(obd->obd_observer && obd->obd_observer->obd_observer);
2213         master_obd = obd->obd_observer->obd_observer;
2214         qctxt = &master_obd->u.obt.obt_qctxt;
2215
2216         LASSERT(qctxt->lqc_handler);
2217         rc = qctxt->lqc_handler(master_obd, qdata,
2218                                 lustre_msg_get_opc(req->rq_reqmsg));
2219         if (rc && rc != -EDQUOT)
2220                 CDEBUG(rc == -EBUSY  ? D_QUOTA : D_ERROR,
2221                        "dqacq failed! (rc:%d)\n", rc);
2222
2223         /* the qd_count might be changed in lqc_handler */
2224         if ((req->rq_export->exp_connect_flags & OBD_CONNECT_QUOTA64) &&
2225             !OBD_FAIL_CHECK(OBD_FAIL_QUOTA_QD_COUNT_32BIT)) {
2226                 memcpy(rep, qdata, sizeof(*qdata));
2227         } else {
2228                 qdata_old = lustre_quota_new_to_old(qdata);
2229                 memcpy(rep, qdata_old, sizeof(*qdata_old));
2230         }
2231         req->rq_status = rc;
2232         rc = ptlrpc_reply(req);
2233
2234         RETURN(rc);
2235 #else
2236         return 0;
2237 #endif /* !__KERNEL__ */
2238 }
2239 #endif /* HAVE_QUOTA_SUPPORT */
2240
2241 ldlm_mode_t lck_compat_array[] = {
2242         [LCK_EX] LCK_COMPAT_EX,
2243         [LCK_PW] LCK_COMPAT_PW,
2244         [LCK_PR] LCK_COMPAT_PR,
2245         [LCK_CW] LCK_COMPAT_CW,
2246         [LCK_CR] LCK_COMPAT_CR,
2247         [LCK_NL] LCK_COMPAT_NL,
2248         [LCK_GROUP] LCK_COMPAT_GROUP
2249 };
2250
2251 /**
2252  * Rather arbitrary mapping from LDLM error codes to errno values. This should
2253  * not escape to the user level.
2254  */
2255 int ldlm_error2errno(ldlm_error_t error)
2256 {
2257         int result;
2258
2259         switch (error) {
2260         case ELDLM_OK:
2261                 result = 0;
2262                 break;
2263         case ELDLM_LOCK_CHANGED:
2264                 result = -ESTALE;
2265                 break;
2266         case ELDLM_LOCK_ABORTED:
2267                 result = -ENAVAIL;
2268                 break;
2269         case ELDLM_LOCK_REPLACED:
2270                 result = -ESRCH;
2271                 break;
2272         case ELDLM_NO_LOCK_DATA:
2273                 result = -ENOENT;
2274                 break;
2275         case ELDLM_NAMESPACE_EXISTS:
2276                 result = -EEXIST;
2277                 break;
2278         case ELDLM_BAD_NAMESPACE:
2279                 result = -EBADF;
2280                 break;
2281         default:
2282                 if (((int)error) < 0)  /* cast to signed type */
2283                         result = error; /* as ldlm_error_t can be unsigned */
2284                 else {
2285                         CERROR("Invalid DLM result code: %i\n", error);
2286                         result = -EPROTO;
2287                 }
2288         }
2289         return result;
2290 }
2291 EXPORT_SYMBOL(ldlm_error2errno);
2292
2293 /**
2294  * Dual to ldlm_error2errno(): maps errno values back to ldlm_error_t.
2295  */
2296 ldlm_error_t ldlm_errno2error(int err_no)
2297 {
2298         int error;
2299
2300         switch (err_no) {
2301         case 0:
2302                 error = ELDLM_OK;
2303                 break;
2304         case -ESTALE:
2305                 error = ELDLM_LOCK_CHANGED;
2306                 break;
2307         case -ENAVAIL:
2308                 error = ELDLM_LOCK_ABORTED;
2309                 break;
2310         case -ESRCH:
2311                 error = ELDLM_LOCK_REPLACED;
2312                 break;
2313         case -ENOENT:
2314                 error = ELDLM_NO_LOCK_DATA;
2315                 break;
2316         case -EEXIST:
2317                 error = ELDLM_NAMESPACE_EXISTS;
2318                 break;
2319         case -EBADF:
2320                 error = ELDLM_BAD_NAMESPACE;
2321                 break;
2322         default:
2323                 error = err_no;
2324         }
2325         return error;
2326 }
2327 EXPORT_SYMBOL(ldlm_errno2error);
2328