lustre/ldlm/ldlm_lib.c

   1 /* -*- mode: c; c-basic-offset: 8; indent-tabs-mode: nil; -*-
   2  * vim:expandtab:shiftwidth=8:tabstop=8:
   3  *
   4  *  Copyright (c) 2003 Cluster File Systems, Inc.
   5  *
   6  *   This file is part of the Lustre file system, http://www.lustre.org
   7  *   Lustre is a trademark of Cluster File Systems, Inc.
   8  *
   9  *   You may have signed or agreed to another license before downloading
  10  *   this software.  If so, you are bound by the terms and conditions
  11  *   of that agreement, and the following does not apply to you.  See the
  12  *   LICENSE file included with this distribution for more information.
  13  *
  14  *   If you did not agree to a different license, then this copy of Lustre
  15  *   is open source software; you can redistribute it and/or modify it
  16  *   under the terms of version 2 of the GNU General Public License as
  17  *   published by the Free Software Foundation.
  18  *
  19  *   In either case, Lustre is distributed in the hope that it will be
  20  *   useful, but WITHOUT ANY WARRANTY; without even the implied warranty
  21  *   of MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
  22  *   license text for more details.
  23  */
  24
  25 #ifndef EXPORT_SYMTAB
  26 # define EXPORT_SYMTAB
  27 #endif
  28 #define DEBUG_SUBSYSTEM S_LDLM
  29
  30 #ifdef __KERNEL__
  31 # include <libcfs/libcfs.h>
  32 #else
  33 # include <liblustre.h>
  34 #endif
  35 #include <obd.h>
  36 #include <lustre_mds.h>
  37 #include <lustre_dlm.h>
  38 #include <lustre_net.h>
  39 #include <lustre_sec.h>
  40
  41 /* @priority: if non-zero, move the selected to the list head
  42  * @create: if zero, only search in existed connections
  43  */
  44 static int import_set_conn(struct obd_import *imp, struct obd_uuid *uuid,
  45                            int priority, int create)
  46 {
  47         struct ptlrpc_connection *ptlrpc_conn;
  48         struct obd_import_conn *imp_conn = NULL, *item;
  49         int rc = 0;
  50         ENTRY;
  51
  52         if (!create && !priority) {
  53                 CDEBUG(D_HA, "Nothing to do\n");
  54                 RETURN(-EINVAL);
  55         }
  56
  57         ptlrpc_conn = ptlrpc_uuid_to_connection(uuid);
  58         if (!ptlrpc_conn) {
  59                 CDEBUG(D_HA, "can't find connection %s\n", uuid->uuid);
  60                 RETURN (-ENOENT);
  61         }
  62
  63         if (create) {
  64                 OBD_ALLOC(imp_conn, sizeof(*imp_conn));
  65                 if (!imp_conn) {
  66                         GOTO(out_put, rc = -ENOMEM);
  67                 }
  68         }
  69
  70         spin_lock(&imp->imp_lock);
  71         list_for_each_entry(item, &imp->imp_conn_list, oic_item) {
  72                 if (obd_uuid_equals(uuid, &item->oic_uuid)) {
  73                         if (priority) {
  74                                 list_del(&item->oic_item);
  75                                 list_add(&item->oic_item, &imp->imp_conn_list);
  76                                 item->oic_last_attempt = 0;
  77                         }
  78                         CDEBUG(D_HA, "imp %p@%s: found existing conn %s%s\n",
  79                                imp, imp->imp_obd->obd_name, uuid->uuid,
  80                                (priority ? ", moved to head" : ""));
  81                         spin_unlock(&imp->imp_lock);
  82                         GOTO(out_free, rc = 0);
  83                 }
  84         }
  85         /* not found */
  86         if (create) {
  87                 imp_conn->oic_conn = ptlrpc_conn;
  88                 imp_conn->oic_uuid = *uuid;
  89                 item->oic_last_attempt = 0;
  90                 if (priority)
  91                         list_add(&imp_conn->oic_item, &imp->imp_conn_list);
  92                 else
  93                         list_add_tail(&imp_conn->oic_item, &imp->imp_conn_list);
  94                 CDEBUG(D_HA, "imp %p@%s: add connection %s at %s\n",
  95                        imp, imp->imp_obd->obd_name, uuid->uuid,
  96                        (priority ? "head" : "tail"));
  97         } else {
  98                 spin_unlock(&imp->imp_lock);
  99                 GOTO(out_free, rc = -ENOENT);
 100         }
 101
 102         spin_unlock(&imp->imp_lock);
 103         RETURN(0);
 104 out_free:
 105         if (imp_conn)
 106                 OBD_FREE(imp_conn, sizeof(*imp_conn));
 107 out_put:
 108         ptlrpc_put_connection(ptlrpc_conn);
 109         RETURN(rc);
 110 }
 111
 112 int import_set_conn_priority(struct obd_import *imp, struct obd_uuid *uuid)
 113 {
 114         return import_set_conn(imp, uuid, 1, 0);
 115 }
 116
 117 int client_import_add_conn(struct obd_import *imp, struct obd_uuid *uuid,
 118                            int priority)
 119 {
 120         return import_set_conn(imp, uuid, priority, 1);
 121 }
 122
 123 int client_import_del_conn(struct obd_import *imp, struct obd_uuid *uuid)
 124 {
 125         struct obd_import_conn *imp_conn;
 126         struct obd_export *dlmexp;
 127         int rc = -ENOENT;
 128         ENTRY;
 129
 130         spin_lock(&imp->imp_lock);
 131         if (list_empty(&imp->imp_conn_list)) {
 132                 LASSERT(!imp->imp_connection);
 133                 GOTO(out, rc);
 134         }
 135
 136         list_for_each_entry(imp_conn, &imp->imp_conn_list, oic_item) {
 137                 if (!obd_uuid_equals(uuid, &imp_conn->oic_uuid))
 138                         continue;
 139                 LASSERT(imp_conn->oic_conn);
 140
 141                 /* is current conn? */
 142                 if (imp_conn == imp->imp_conn_current) {
 143                         LASSERT(imp_conn->oic_conn == imp->imp_connection);
 144
 145                         if (imp->imp_state != LUSTRE_IMP_CLOSED &&
 146                             imp->imp_state != LUSTRE_IMP_DISCON) {
 147                                 CERROR("can't remove current connection\n");
 148                                 GOTO(out, rc = -EBUSY);
 149                         }
 150
 151                         ptlrpc_put_connection(imp->imp_connection);
 152                         imp->imp_connection = NULL;
 153
 154                         dlmexp = class_conn2export(&imp->imp_dlm_handle);
 155                         if (dlmexp && dlmexp->exp_connection) {
 156                                 LASSERT(dlmexp->exp_connection ==
 157                                         imp_conn->oic_conn);
 158                                 ptlrpc_put_connection(dlmexp->exp_connection);
 159                                 dlmexp->exp_connection = NULL;
 160                         }
 161                 }
 162
 163                 list_del(&imp_conn->oic_item);
 164                 ptlrpc_put_connection(imp_conn->oic_conn);
 165                 OBD_FREE(imp_conn, sizeof(*imp_conn));
 166                 CDEBUG(D_HA, "imp %p@%s: remove connection %s\n",
 167                        imp, imp->imp_obd->obd_name, uuid->uuid);
 168                 rc = 0;
 169                 break;
 170         }
 171 out:
 172         spin_unlock(&imp->imp_lock);
 173         if (rc == -ENOENT)
 174                 CERROR("connection %s not found\n", uuid->uuid);
 175         RETURN(rc);
 176 }
 177
 178 static void destroy_import(struct obd_import *imp)
 179 {
 180         /* drop security policy instance after all rpc finished/aborted
 181          * to let all busy credentials be released.
 182          */
 183         class_import_get(imp);
 184         class_destroy_import(imp);
 185         sptlrpc_import_put_sec(imp);
 186         class_import_put(imp);
 187 }
 188
 189 /* configure an RPC client OBD device
 190  *
 191  * lcfg parameters:
 192  * 1 - client UUID
 193  * 2 - server UUID
 194  * 3 - inactive-on-startup
 195  */
 196 int client_obd_setup(struct obd_device *obddev, struct lustre_cfg *lcfg)
 197 {
 198         struct client_obd *cli = &obddev->u.cli;
 199         struct obd_import *imp;
 200         struct obd_uuid server_uuid;
 201         int rq_portal, rp_portal, connect_op;
 202         char *name = obddev->obd_type->typ_name;
 203         int rc;
 204         ENTRY;
 205
 206         /* In a more perfect world, we would hang a ptlrpc_client off of
 207          * obd_type and just use the values from there. */
 208         if (!strcmp(name, LUSTRE_OSC_NAME)) {
 209                 rq_portal = OST_REQUEST_PORTAL;
 210                 rp_portal = OSC_REPLY_PORTAL;
 211                 connect_op = OST_CONNECT;
 212         } else if (!strcmp(name, LUSTRE_MDC_NAME)) {
 213                 rq_portal = MDS_REQUEST_PORTAL;
 214                 rp_portal = MDC_REPLY_PORTAL;
 215                 connect_op = MDS_CONNECT;
 216         } else if (!strcmp(name, LUSTRE_MGC_NAME)) {
 217                 rq_portal = MGS_REQUEST_PORTAL;
 218                 rp_portal = MGC_REPLY_PORTAL;
 219                 connect_op = MGS_CONNECT;
 220         } else {
 221                 CERROR("unknown client OBD type \"%s\", can't setup\n",
 222                        name);
 223                 RETURN(-EINVAL);
 224         }
 225
 226         if (LUSTRE_CFG_BUFLEN(lcfg, 1) < 1) {
 227                 CERROR("requires a TARGET UUID\n");
 228                 RETURN(-EINVAL);
 229         }
 230
 231         if (LUSTRE_CFG_BUFLEN(lcfg, 1) > 37) {
 232                 CERROR("client UUID must be less than 38 characters\n");
 233                 RETURN(-EINVAL);
 234         }
 235
 236         if (LUSTRE_CFG_BUFLEN(lcfg, 2) < 1) {
 237                 CERROR("setup requires a SERVER UUID\n");
 238                 RETURN(-EINVAL);
 239         }
 240
 241         if (LUSTRE_CFG_BUFLEN(lcfg, 2) > 37) {
 242                 CERROR("target UUID must be less than 38 characters\n");
 243                 RETURN(-EINVAL);
 244         }
 245
 246         sema_init(&cli->cl_sem, 1);
 247         sema_init(&cli->cl_mgc_sem, 1);
 248         cli->cl_sec_conf.sfc_rpc_flavor = SPTLRPC_FLVR_NULL;
 249         cli->cl_sec_conf.sfc_bulk_csum = BULK_CSUM_ALG_NULL;
 250         cli->cl_sec_conf.sfc_bulk_priv = BULK_PRIV_ALG_NULL;
 251         cli->cl_sec_conf.sfc_flags = 0;
 252         cli->cl_conn_count = 0;
 253         memcpy(server_uuid.uuid, lustre_cfg_buf(lcfg, 2),
 254                min_t(unsigned int, LUSTRE_CFG_BUFLEN(lcfg, 2),
 255                      sizeof(server_uuid)));
 256
 257         cli->cl_dirty = 0;
 258         cli->cl_avail_grant = 0;
 259         /* FIXME: should limit this for the sum of all cl_dirty_max */
 260         cli->cl_dirty_max = OSC_MAX_DIRTY_DEFAULT * 1024 * 1024;
 261         if (cli->cl_dirty_max >> CFS_PAGE_SHIFT > num_physpages / 8)
 262                 cli->cl_dirty_max = num_physpages << (CFS_PAGE_SHIFT - 3);
 263         CFS_INIT_LIST_HEAD(&cli->cl_cache_waiters);
 264         CFS_INIT_LIST_HEAD(&cli->cl_loi_ready_list);
 265         CFS_INIT_LIST_HEAD(&cli->cl_loi_write_list);
 266         CFS_INIT_LIST_HEAD(&cli->cl_loi_read_list);
 267         client_obd_list_lock_init(&cli->cl_loi_list_lock);
 268         cli->cl_r_in_flight = 0;
 269         cli->cl_w_in_flight = 0;
 270
 271         spin_lock_init(&cli->cl_read_rpc_hist.oh_lock);
 272         spin_lock_init(&cli->cl_write_rpc_hist.oh_lock);
 273         spin_lock_init(&cli->cl_read_page_hist.oh_lock);
 274         spin_lock_init(&cli->cl_write_page_hist.oh_lock);
 275         spin_lock_init(&cli->cl_read_offset_hist.oh_lock);
 276         spin_lock_init(&cli->cl_write_offset_hist.oh_lock);
 277
 278         /* This value may be changed at connect time in
 279            ptlrpc_connect_interpret. */
 280         cli->cl_max_pages_per_rpc = min((int)PTLRPC_MAX_BRW_PAGES,
 281                                         (int)(1024 * 1024 >> CFS_PAGE_SHIFT));
 282
 283         if (!strcmp(name, LUSTRE_MDC_NAME)) {
 284                 cli->cl_max_rpcs_in_flight = MDC_MAX_RIF_DEFAULT;
 285         } else if (num_physpages >> (20 - CFS_PAGE_SHIFT) <= 128 /* MB */) {
 286                 cli->cl_max_rpcs_in_flight = 2;
 287         } else if (num_physpages >> (20 - CFS_PAGE_SHIFT) <= 256 /* MB */) {
 288                 cli->cl_max_rpcs_in_flight = 3;
 289         } else if (num_physpages >> (20 - CFS_PAGE_SHIFT) <= 512 /* MB */) {
 290                 cli->cl_max_rpcs_in_flight = 4;
 291         } else {
 292                 cli->cl_max_rpcs_in_flight = OSC_MAX_RIF_DEFAULT;
 293         }
 294
 295         rc = ldlm_get_ref();
 296         if (rc) {
 297                 CERROR("ldlm_get_ref failed: %d\n", rc);
 298                 GOTO(err, rc);
 299         }
 300
 301         ptlrpc_init_client(rq_portal, rp_portal, name,
 302                            &obddev->obd_ldlm_client);
 303
 304         imp = class_new_import(obddev);
 305         if (imp == NULL)
 306                 GOTO(err_ldlm, rc = -ENOENT);
 307         imp->imp_client = &obddev->obd_ldlm_client;
 308         imp->imp_connect_op = connect_op;
 309         imp->imp_initial_recov = 1;
 310         imp->imp_initial_recov_bk = 0;
 311         CFS_INIT_LIST_HEAD(&imp->imp_pinger_chain);
 312         memcpy(cli->cl_target_uuid.uuid, lustre_cfg_buf(lcfg, 1),
 313                LUSTRE_CFG_BUFLEN(lcfg, 1));
 314         class_import_put(imp);
 315
 316         rc = client_import_add_conn(imp, &server_uuid, 1);
 317         if (rc) {
 318                 CERROR("can't add initial connection\n");
 319                 GOTO(err_import, rc);
 320         }
 321
 322         cli->cl_import = imp;
 323         /* cli->cl_max_mds_{easize,cookiesize} updated by mdc_init_ea_size() */
 324         cli->cl_max_mds_easize = sizeof(struct lov_mds_md);
 325         cli->cl_max_mds_cookiesize = sizeof(struct llog_cookie);
 326
 327         if (LUSTRE_CFG_BUFLEN(lcfg, 3) > 0) {
 328                 if (!strcmp(lustre_cfg_string(lcfg, 3), "inactive")) {
 329                         CDEBUG(D_HA, "marking %s %s->%s as inactive\n",
 330                                name, obddev->obd_name,
 331                                cli->cl_target_uuid.uuid);
 332                         spin_lock(&imp->imp_lock);
 333                         imp->imp_invalid = 1;
 334                         spin_unlock(&imp->imp_lock);
 335                 }
 336         }
 337
 338         cli->cl_qchk_stat = CL_NOT_QUOTACHECKED;
 339
 340         RETURN(rc);
 341
 342 err_import:
 343         class_destroy_import(imp);
 344 err_ldlm:
 345         ldlm_put_ref(0);
 346 err:
 347         RETURN(rc);
 348
 349 }
 350
 351 int client_obd_cleanup(struct obd_device *obddev)
 352 {
 353         ENTRY;
 354         ldlm_put_ref(obddev->obd_force);
 355
 356         RETURN(0);
 357 }
 358
 359 /* ->o_connect() method for client side (OSC and MDC and MGC) */
 360 int client_connect_import(const struct lu_env *env,
 361                           struct lustre_handle *dlm_handle,
 362                           struct obd_device *obd, struct obd_uuid *cluuid,
 363                           struct obd_connect_data *data)
 364 {
 365         struct client_obd *cli = &obd->u.cli;
 366         struct obd_import *imp = cli->cl_import;
 367         struct obd_export *exp;
 368         struct obd_connect_data *ocd;
 369         int rc;
 370         ENTRY;
 371
 372         mutex_down(&cli->cl_sem);
 373         rc = class_connect(dlm_handle, obd, cluuid);
 374         if (rc)
 375                 GOTO(out_sem, rc);
 376
 377         cli->cl_conn_count++;
 378         if (cli->cl_conn_count > 1)
 379                 GOTO(out_sem, rc);
 380         exp = class_conn2export(dlm_handle);
 381
 382         if (obd->obd_namespace != NULL)
 383                 CERROR("already have namespace!\n");
 384         obd->obd_namespace = ldlm_namespace_new(obd->obd_name,
 385                                                 LDLM_NAMESPACE_CLIENT);
 386         if (obd->obd_namespace == NULL)
 387                 GOTO(out_disco, rc = -ENOMEM);
 388
 389         imp->imp_dlm_handle = *dlm_handle;
 390         rc = ptlrpc_init_import(imp);
 391         if (rc != 0)
 392                 GOTO(out_ldlm, rc);
 393
 394         rc = sptlrpc_import_get_sec(imp, NULL, cli->cl_sec_conf.sfc_rpc_flavor,
 395                                     cli->cl_sec_conf.sfc_flags);
 396         if (rc)
 397                 GOTO(out_ldlm, rc);
 398
 399         ocd = &imp->imp_connect_data;
 400         if (data) {
 401                 *ocd = *data;
 402                 imp->imp_connect_flags_orig = data->ocd_connect_flags;
 403         }
 404
 405         rc = ptlrpc_connect_import(imp, NULL);
 406         if (rc != 0) {
 407                 LASSERT (imp->imp_state == LUSTRE_IMP_DISCON);
 408                 GOTO(out_ldlm, rc);
 409         }
 410         LASSERT(exp->exp_connection);
 411
 412         if (data) {
 413                 LASSERT((ocd->ocd_connect_flags & data->ocd_connect_flags) ==
 414                         ocd->ocd_connect_flags);
 415                 data->ocd_connect_flags = ocd->ocd_connect_flags;
 416         }
 417
 418         ptlrpc_pinger_add_import(imp);
 419
 420         EXIT;
 421
 422         if (rc) {
 423 out_ldlm:
 424                 ldlm_namespace_free(obd->obd_namespace, 0);
 425                 obd->obd_namespace = NULL;
 426 out_disco:
 427                 cli->cl_conn_count--;
 428                 class_disconnect(exp);
 429         } else {
 430                 class_export_put(exp);
 431         }
 432 out_sem:
 433         mutex_up(&cli->cl_sem);
 434         return rc;
 435 }
 436
 437 int client_disconnect_export(struct obd_export *exp)
 438 {
 439         struct obd_device *obd = class_exp2obd(exp);
 440         struct client_obd *cli;
 441         struct obd_import *imp;
 442         int rc = 0, err;
 443         ENTRY;
 444
 445         if (!obd) {
 446                 CERROR("invalid export for disconnect: exp %p cookie "LPX64"\n",
 447                        exp, exp ? exp->exp_handle.h_cookie : -1);
 448                 RETURN(-EINVAL);
 449         }
 450
 451         cli = &obd->u.cli;
 452         imp = cli->cl_import;
 453
 454         mutex_down(&cli->cl_sem);
 455         if (!cli->cl_conn_count) {
 456                 CERROR("disconnecting disconnected device (%s)\n",
 457                        obd->obd_name);
 458                 GOTO(out_sem, rc = -EINVAL);
 459         }
 460
 461         cli->cl_conn_count--;
 462         if (cli->cl_conn_count)
 463                 GOTO(out_no_disconnect, rc = 0);
 464
 465         /* Mark import deactivated now, so we don't try to reconnect if any
 466          * of the cleanup RPCs fails (e.g. ldlm cancel, etc).  We don't
 467          * fully deactivate the import, or that would drop all requests. */
 468         spin_lock(&imp->imp_lock);
 469         imp->imp_deactive = 1;
 470         spin_unlock(&imp->imp_lock);
 471
 472         /* Some non-replayable imports (MDS's OSCs) are pinged, so just
 473          * delete it regardless.  (It's safe to delete an import that was
 474          * never added.) */
 475         (void)ptlrpc_pinger_del_import(imp);
 476
 477         if (obd->obd_namespace != NULL) {
 478                 /* obd_force == local only */
 479                 ldlm_cli_cancel_unused(obd->obd_namespace, NULL,
 480                                        obd->obd_force ? LDLM_FL_LOCAL_ONLY:0,
 481                                        NULL);
 482                 ldlm_namespace_free(obd->obd_namespace, obd->obd_force);
 483                 obd->obd_namespace = NULL;
 484         }
 485
 486         if (!obd->obd_force)
 487                 rc = ptlrpc_disconnect_import(imp, 0);
 488
 489         ptlrpc_invalidate_import(imp);
 490         ptlrpc_free_rq_pool(imp->imp_rq_pool);
 491         destroy_import(imp);
 492         cli->cl_import = NULL;
 493
 494         EXIT;
 495  out_no_disconnect:
 496         err = class_disconnect(exp);
 497         if (!rc && err)
 498                 rc = err;
 499  out_sem:
 500         mutex_up(&cli->cl_sem);
 501         RETURN(rc);
 502 }
 503
 504 /* --------------------------------------------------------------------------
 505  * from old lib/target.c
 506  * -------------------------------------------------------------------------- */
 507
 508 int target_handle_reconnect(struct lustre_handle *conn, struct obd_export *exp,
 509                             struct obd_uuid *cluuid, int initial_conn)
 510 {
 511         ENTRY;
 512         if (exp->exp_connection && exp->exp_imp_reverse && !initial_conn) {
 513                 struct lustre_handle *hdl;
 514                 hdl = &exp->exp_imp_reverse->imp_remote_handle;
 515                 /* Might be a re-connect after a partition. */
 516                 if (!memcmp(&conn->cookie, &hdl->cookie, sizeof conn->cookie)) {
 517                         CWARN("%s: %s reconnecting\n", exp->exp_obd->obd_name,
 518                               cluuid->uuid);
 519                         conn->cookie = exp->exp_handle.h_cookie;
 520                         /* target_handle_connect() treats EALREADY and
 521                          * -EALREADY differently.  EALREADY means we are
 522                          * doing a valid reconnect from the same client. */
 523                         RETURN(EALREADY);
 524                 } else {
 525                         CERROR("%s reconnecting from %s, "
 526                                "handle mismatch (ours "LPX64", theirs "
 527                                LPX64")\n", cluuid->uuid,
 528                                exp->exp_connection->c_remote_uuid.uuid,
 529                                hdl->cookie, conn->cookie);
 530                         memset(conn, 0, sizeof *conn);
 531                         /* target_handle_connect() treats EALREADY and
 532                          * -EALREADY differently.  -EALREADY is an error
 533                          * (same UUID, different handle). */
 534                         RETURN(-EALREADY);
 535                 }
 536         }
 537
 538         conn->cookie = exp->exp_handle.h_cookie;
 539         CDEBUG(D_HA, "connect export for UUID '%s' at %p, cookie "LPX64"\n",
 540                cluuid->uuid, exp, conn->cookie);
 541         RETURN(0);
 542 }
 543
 544 void target_client_add_cb(struct obd_device *obd, __u64 transno, void *cb_data,
 545                           int error)
 546 {
 547         struct obd_export *exp = cb_data;
 548
 549         CDEBUG(D_HA, "%s: committing for initial connect of %s\n",
 550                obd->obd_name, exp->exp_client_uuid.uuid);
 551
 552         spin_lock(&exp->exp_lock);
 553         exp->exp_need_sync = 0;
 554         spin_unlock(&exp->exp_lock);
 555 }
 556 EXPORT_SYMBOL(target_client_add_cb);
 557
 558 int target_handle_connect(struct ptlrpc_request *req)
 559 {
 560         struct obd_device *target, *targref = NULL;
 561         struct obd_export *export = NULL;
 562         struct obd_import *revimp;
 563         struct lustre_handle conn;
 564         struct obd_uuid tgtuuid;
 565         struct obd_uuid cluuid;
 566         struct obd_uuid remote_uuid;
 567         struct list_head *p;
 568         char *str, *tmp;
 569         int rc = 0;
 570         int initial_conn = 0;
 571         struct obd_connect_data *data;
 572         int size[2] = { sizeof(struct ptlrpc_body), sizeof(*data) };
 573         ENTRY;
 574
 575         OBD_RACE(OBD_FAIL_TGT_CONN_RACE);
 576
 577         LASSERT_REQSWAB(req, REQ_REC_OFF);
 578         str = lustre_msg_string(req->rq_reqmsg, REQ_REC_OFF, sizeof(tgtuuid)-1);
 579         if (str == NULL) {
 580                 DEBUG_REQ(D_ERROR, req, "bad target UUID for connect");
 581                 GOTO(out, rc = -EINVAL);
 582         }
 583
 584         obd_str2uuid (&tgtuuid, str);
 585         target = class_uuid2obd(&tgtuuid);
 586         if (!target)
 587                 target = class_name2obd(str);
 588
 589         if (!target || target->obd_stopping || !target->obd_set_up) {
 590                 LCONSOLE_ERROR_MSG(0x137, "UUID '%s' is not available "
 591                                    " for connect (%s)\n", str,
 592                                    !target ? "no target" :
 593                                    (target->obd_stopping ? "stopping" :
 594                                    "not set up"));
 595                 GOTO(out, rc = -ENODEV);
 596         }
 597
 598         if (target->obd_no_conn) {
 599                 LCONSOLE_WARN("%s: temporarily refusing client connection "
 600                               "from %s\n", target->obd_name,
 601                               libcfs_nid2str(req->rq_peer.nid));
 602                 GOTO(out, rc = -EAGAIN);
 603         }
 604
 605         /* Make sure the target isn't cleaned up while we're here. Yes,
 606            there's still a race between the above check and our incref here.
 607            Really, class_uuid2obd should take the ref. */
 608         targref = class_incref(target);
 609
 610         LASSERT_REQSWAB(req, REQ_REC_OFF + 1);
 611         str = lustre_msg_string(req->rq_reqmsg, REQ_REC_OFF + 1,
 612                                 sizeof(cluuid) - 1);
 613         if (str == NULL) {
 614                 DEBUG_REQ(D_ERROR, req, "bad client UUID for connect");
 615                 GOTO(out, rc = -EINVAL);
 616         }
 617
 618         obd_str2uuid (&cluuid, str);
 619
 620         /* XXX extract a nettype and format accordingly */
 621         switch (sizeof(lnet_nid_t)) {
 622                 /* NB the casts only avoid compiler warnings */
 623         case 8:
 624                 snprintf(remote_uuid.uuid, sizeof remote_uuid,
 625                          "NET_"LPX64"_UUID", (__u64)req->rq_peer.nid);
 626                 break;
 627         case 4:
 628                 snprintf(remote_uuid.uuid, sizeof remote_uuid,
 629                          "NET_%x_UUID", (__u32)req->rq_peer.nid);
 630                 break;
 631         default:
 632                 LBUG();
 633         }
 634
 635         tmp = lustre_msg_buf(req->rq_reqmsg, REQ_REC_OFF + 2, sizeof conn);
 636         if (tmp == NULL)
 637                 GOTO(out, rc = -EPROTO);
 638
 639         memcpy(&conn, tmp, sizeof conn);
 640
 641         data = lustre_swab_reqbuf(req, REQ_REC_OFF + 3, sizeof(*data),
 642                                   lustre_swab_connect);
 643
 644         if (!data)
 645                 GOTO(out, rc = -EPROTO);
 646
 647         rc = lustre_pack_reply(req, 2, size, NULL);
 648         if (rc)
 649                 GOTO(out, rc);
 650
 651         if (lustre_msg_get_op_flags(req->rq_reqmsg) & MSG_CONNECT_LIBCLIENT) {
 652                 if (!data) {
 653                         DEBUG_REQ(D_WARNING, req, "Refusing old (unversioned) "
 654                                   "libclient connection attempt");
 655                         GOTO(out, rc = -EPROTO);
 656                 } else if (data->ocd_version < LUSTRE_VERSION_CODE -
 657                                                LUSTRE_VERSION_ALLOWED_OFFSET ||
 658                            data->ocd_version > LUSTRE_VERSION_CODE +
 659                                                LUSTRE_VERSION_ALLOWED_OFFSET) {
 660                         DEBUG_REQ(D_WARNING, req, "Refusing %s (%d.%d.%d.%d) "
 661                                   "libclient connection attempt",
 662                                   data->ocd_version < LUSTRE_VERSION_CODE ?
 663                                   "old" : "new",
 664                                   OBD_OCD_VERSION_MAJOR(data->ocd_version),
 665                                   OBD_OCD_VERSION_MINOR(data->ocd_version),
 666                                   OBD_OCD_VERSION_PATCH(data->ocd_version),
 667                                   OBD_OCD_VERSION_FIX(data->ocd_version));
 668                         data = lustre_msg_buf(req->rq_repmsg, REPLY_REC_OFF,
 669                                               offsetof(typeof(*data),
 670                                                        ocd_version) +
 671                                               sizeof(data->ocd_version));
 672                         if (data) {
 673                                 data->ocd_connect_flags = OBD_CONNECT_VERSION;
 674                                 data->ocd_version = LUSTRE_VERSION_CODE;
 675                         }
 676                         GOTO(out, rc = -EPROTO);
 677                 }
 678         }
 679
 680         if (lustre_msg_get_op_flags(req->rq_reqmsg) & MSG_CONNECT_INITIAL)
 681                 initial_conn = 1;
 682
 683         /* lctl gets a backstage, all-access pass. */
 684         if (obd_uuid_equals(&cluuid, &target->obd_uuid))
 685                 goto dont_check_exports;
 686
 687         spin_lock(&target->obd_dev_lock);
 688         list_for_each(p, &target->obd_exports) {
 689                 export = list_entry(p, struct obd_export, exp_obd_chain);
 690                 if (obd_uuid_equals(&cluuid, &export->exp_client_uuid)) {
 691                         if (export->exp_connecting) { /* bug 9635, et. al. */
 692                                 CWARN("%s: exp %p already connecting\n",
 693                                       export->exp_obd->obd_name, export);
 694                                 export = NULL;
 695                                 rc = -EALREADY;
 696                                 break;
 697                         }
 698
 699                         /* make darn sure this is coming from the same peer
 700                          * if the UUIDs matched */
 701                         if ((export->exp_connection != NULL) &&
 702                             (strcmp(libcfs_nid2str(req->rq_peer.nid),
 703                                     libcfs_nid2str(export->exp_connection->c_peer.nid)))) {
 704                                 CWARN("%s: cookie %s seen on new NID %s when "
 705                                       "existing NID %s is already connected\n",
 706                                       target->obd_name, cluuid.uuid,
 707                                       libcfs_nid2str(req->rq_peer.nid),
 708                                       libcfs_nid2str(export->exp_connection->c_peer.nid));
 709                                 export = NULL;
 710                                 rc = -EALREADY;
 711                                 break;
 712                         }
 713
 714                         spin_lock(&export->exp_lock);
 715                         export->exp_connecting = 1;
 716                         spin_unlock(&export->exp_lock);
 717                         spin_unlock(&target->obd_dev_lock);
 718                         LASSERT(export->exp_obd == target);
 719
 720                         rc = target_handle_reconnect(&conn, export, &cluuid,
 721                                                      initial_conn);
 722                         break;
 723                 }
 724                 export = NULL;
 725         }
 726
 727         /* If we found an export, we already unlocked. */
 728         if (!export) {
 729                 spin_unlock(&target->obd_dev_lock);
 730                 OBD_FAIL_TIMEOUT(OBD_FAIL_TGT_DELAY_CONNECT, 2 * obd_timeout);
 731         } else if (req->rq_export == NULL &&
 732                    atomic_read(&export->exp_rpc_count) > 0) {
 733                 CWARN("%s: refuse connection from %s/%s to 0x%p/%d\n",
 734                       target->obd_name, cluuid.uuid,
 735                       libcfs_nid2str(req->rq_peer.nid),
 736                       export, atomic_read(&export->exp_refcount));
 737                 GOTO(out, rc = -EBUSY);
 738         } else if (req->rq_export != NULL &&
 739                    (atomic_read(&export->exp_rpc_count) > 1)) {
 740                 CWARN("%s: refuse reconnection from %s@%s to 0x%p/%d\n",
 741                       target->obd_name, cluuid.uuid,
 742                       libcfs_nid2str(req->rq_peer.nid),
 743                       export, atomic_read(&export->exp_rpc_count));
 744                 GOTO(out, rc = -EBUSY);
 745         } else if (lustre_msg_get_conn_cnt(req->rq_reqmsg) == 1 &&
 746                    !initial_conn) {
 747                 CERROR("%s: NID %s (%s) reconnected with 1 conn_cnt; "
 748                        "cookies not random?\n", target->obd_name,
 749                        libcfs_nid2str(req->rq_peer.nid), cluuid.uuid);
 750                 GOTO(out, rc = -EALREADY);
 751         } else {
 752                 OBD_FAIL_TIMEOUT(OBD_FAIL_TGT_DELAY_RECONNECT, 2 * obd_timeout);
 753                 if (req->rq_export == NULL && initial_conn)
 754                        export->exp_last_request_time =
 755                                max(export->exp_last_request_time,
 756                                    (time_t)CURRENT_SECONDS);
 757         }
 758
 759         /* We want to handle EALREADY but *not* -EALREADY from
 760          * target_handle_reconnect(), return reconnection state in a flag */
 761         if (rc == EALREADY) {
 762                 lustre_msg_add_op_flags(req->rq_repmsg, MSG_CONNECT_RECONNECT);
 763                 rc = 0;
 764         } else if (rc) {
 765                 GOTO(out, rc);
 766         }
 767         /* Tell the client if we're in recovery. */
 768         /* If this is the first client, start the recovery timer */
 769         CWARN("%s: connection from %s@%s %st"LPU64" exp %p cur %ld last %ld\n",
 770                target->obd_name, cluuid.uuid, libcfs_nid2str(req->rq_peer.nid),
 771               target->obd_recovering ? "recovering/" : "", data->ocd_transno,
 772               export, (long)CURRENT_SECONDS,
 773               export ? (long)export->exp_last_request_time : 0);
 774
 775
 776         if (target->obd_recovering) {
 777                 lustre_msg_add_op_flags(req->rq_repmsg, MSG_CONNECT_RECOVERING);
 778                 target_start_recovery_timer(target);
 779         }
 780
 781         /* Tell the client if we support replayable requests */
 782         if (target->obd_replayable)
 783                 lustre_msg_add_op_flags(req->rq_repmsg, MSG_CONNECT_REPLAYABLE);
 784
 785         if (export == NULL) {
 786                 if (target->obd_recovering) {
 787                         CERROR("%s: denying connection for new client %s (%s): "
 788                                "%d clients in recovery for %lds\n",
 789                                target->obd_name,
 790                                libcfs_nid2str(req->rq_peer.nid), cluuid.uuid,
 791                                target->obd_recoverable_clients,
 792                                cfs_duration_sec(cfs_time_sub(cfs_timer_deadline(&target->obd_recovery_timer),
 793                                                              cfs_time_current())));
 794                         rc = -EBUSY;
 795                 } else {
 796 dont_check_exports:
 797                         rc = obd_connect(req->rq_svc_thread->t_env,
 798                                          &conn, target, &cluuid, data);
 799                 }
 800         } else {
 801                 rc = obd_reconnect(export, target, &cluuid, data);
 802         }
 803         if (rc)
 804                 GOTO(out, rc);
 805         /* Return only the parts of obd_connect_data that we understand, so the
 806          * client knows that we don't understand the rest. */
 807         if (data)
 808                 memcpy(lustre_msg_buf(req->rq_repmsg, REPLY_REC_OFF,
 809                                       sizeof(*data)),
 810                        data, sizeof(*data));
 811
 812         /* If all else goes well, this is our RPC return code. */
 813         req->rq_status = 0;
 814
 815         lustre_msg_set_handle(req->rq_repmsg, &conn);
 816
 817         /* ownership of this export ref transfers to the request AFTER we
 818          * drop any previous reference the request had, but we don't want
 819          * that to go to zero before we get our new export reference. */
 820         export = class_conn2export(&conn);
 821         if (!export) {
 822                 DEBUG_REQ(D_ERROR, req, "Missing export!");
 823                 GOTO(out, rc = -ENODEV);
 824         }
 825
 826         /* If the client and the server are the same node, we will already
 827          * have an export that really points to the client's DLM export,
 828          * because we have a shared handles table.
 829          *
 830          * XXX this will go away when shaver stops sending the "connect" handle
 831          * in the real "remote handle" field of the request --phik 24 Apr 2003
 832          */
 833         if (req->rq_export != NULL)
 834                 class_export_put(req->rq_export);
 835
 836         req->rq_export = export;
 837
 838         spin_lock(&export->exp_lock);
 839         if (initial_conn) {
 840                 lustre_msg_set_conn_cnt(req->rq_repmsg, export->exp_conn_cnt + 1);
 841         } else if (export->exp_conn_cnt >= lustre_msg_get_conn_cnt(req->rq_reqmsg)) {
 842                 spin_unlock(&export->exp_lock);
 843                 CERROR("%s: %s already connected at higher conn_cnt: %d > %d\n",
 844                        cluuid.uuid, libcfs_nid2str(req->rq_peer.nid),
 845                        export->exp_conn_cnt,
 846                        lustre_msg_get_conn_cnt(req->rq_reqmsg));
 847
 848                 GOTO(out, rc = -EALREADY);
 849         }
 850         export->exp_conn_cnt = lustre_msg_get_conn_cnt(req->rq_reqmsg);
 851
 852         /* request from liblustre?  Don't evict it for not pinging. */
 853         if (lustre_msg_get_op_flags(req->rq_reqmsg) & MSG_CONNECT_LIBCLIENT) {
 854                 export->exp_libclient = 1;
 855                 spin_unlock(&export->exp_lock);
 856
 857                 spin_lock(&target->obd_dev_lock);
 858                 list_del_init(&export->exp_obd_chain_timed);
 859                 spin_unlock(&target->obd_dev_lock);
 860         } else {
 861                 spin_unlock(&export->exp_lock);
 862         }
 863
 864         if (export->exp_connection != NULL)
 865                 ptlrpc_put_connection(export->exp_connection);
 866         export->exp_connection = ptlrpc_get_connection(req->rq_peer,
 867                                                        req->rq_self,
 868                                                        &remote_uuid);
 869
 870         if (lustre_msg_get_op_flags(req->rq_repmsg) & MSG_CONNECT_RECONNECT) {
 871                 LASSERT(export->exp_imp_reverse);
 872                 sptlrpc_svc_install_rvs_ctx(export->exp_imp_reverse,
 873                                             req->rq_svc_ctx);
 874                 GOTO(out, rc = 0);
 875         }
 876
 877         spin_lock_bh(&target->obd_processing_task_lock);
 878         if (target->obd_recovering && !export->exp_in_recovery) {
 879                 spin_lock(&export->exp_lock);
 880                 export->exp_in_recovery = 1;
 881                 export->exp_req_replay_needed = 1;
 882                 export->exp_lock_replay_needed = 1;
 883                 spin_unlock(&export->exp_lock);
 884                 if ((lustre_msg_get_op_flags(req->rq_reqmsg) & MSG_CONNECT_TRANSNO)
 885                      && data->ocd_transno < target->obd_next_recovery_transno)
 886                         target->obd_next_recovery_transno = data->ocd_transno;
 887                 target->obd_connected_clients++;
 888                 /* each connected client is counted as recoverable */
 889                 target->obd_recoverable_clients++;
 890                 atomic_inc(&target->obd_req_replay_clients);
 891                 atomic_inc(&target->obd_lock_replay_clients);
 892                 if (target->obd_connected_clients ==
 893                     target->obd_max_recoverable_clients)
 894                         wake_up(&target->obd_next_transno_waitq);
 895         }
 896         spin_unlock_bh(&target->obd_processing_task_lock);
 897         memcpy(&conn,
 898                lustre_msg_buf(req->rq_reqmsg, REQ_REC_OFF + 2, sizeof conn),
 899                sizeof conn);
 900
 901         if (export->exp_imp_reverse != NULL) {
 902                 /* destroyed import can be still referenced in ctxt */
 903                 obd_set_info_async(export, strlen(KEY_REVIMP_UPD),
 904                                    KEY_REVIMP_UPD, 0, NULL, NULL);
 905                 destroy_import(export->exp_imp_reverse);
 906         }
 907
 908         /* for the rest part, we return -ENOTCONN in case of errors
 909          * in order to let client initialize connection again.
 910          */
 911         revimp = export->exp_imp_reverse = class_new_import(target);
 912         if (!revimp) {
 913                 CERROR("fail to alloc new reverse import.\n");
 914                 GOTO(out, rc = -ENOTCONN);
 915         }
 916
 917         revimp->imp_connection = ptlrpc_connection_addref(export->exp_connection);
 918         revimp->imp_client = &export->exp_obd->obd_ldlm_client;
 919         revimp->imp_remote_handle = conn;
 920         revimp->imp_dlm_fake = 1;
 921         revimp->imp_state = LUSTRE_IMP_FULL;
 922
 923         if (lustre_msg_get_op_flags(req->rq_reqmsg) & MSG_CONNECT_NEXT_VER) {
 924                 revimp->imp_msg_magic = LUSTRE_MSG_MAGIC_V2;
 925                 lustre_msg_add_op_flags(req->rq_repmsg, MSG_CONNECT_NEXT_VER);
 926         }
 927
 928         rc = sptlrpc_import_get_sec(revimp, req->rq_svc_ctx,
 929                                     req->rq_sec_flavor, 0);
 930         if (rc) {
 931                 CERROR("Failed to get sec for reverse import: %d\n", rc);
 932                 export->exp_imp_reverse = NULL;
 933                 class_destroy_import(revimp);
 934         }
 935
 936         class_import_put(revimp);
 937 out:
 938         if (export) {
 939                 spin_lock(&export->exp_lock);
 940                 export->exp_connecting = 0;
 941                 spin_unlock(&export->exp_lock);
 942         }
 943         if (targref)
 944                 class_decref(targref);
 945         if (rc)
 946                 req->rq_status = rc;
 947         RETURN(rc);
 948 }
 949
 950 int target_handle_disconnect(struct ptlrpc_request *req)
 951 {
 952         int rc;
 953         ENTRY;
 954
 955         rc = lustre_pack_reply(req, 1, NULL, NULL);
 956         if (rc)
 957                 RETURN(rc);
 958
 959         /* keep the rq_export around so we can send the reply */
 960         req->rq_status = obd_disconnect(class_export_get(req->rq_export));
 961
 962         RETURN(0);
 963 }
 964
 965 void target_destroy_export(struct obd_export *exp)
 966 {
 967         /* exports created from last_rcvd data, and "fake"
 968            exports created by lctl don't have an import */
 969         if (exp->exp_imp_reverse != NULL)
 970                 destroy_import(exp->exp_imp_reverse);
 971
 972         /* We cancel locks at disconnect time, but this will catch any locks
 973          * granted in a race with recovery-induced disconnect. */
 974         if (exp->exp_obd->obd_namespace != NULL)
 975                 ldlm_cancel_locks_for_export(exp);
 976 }
 977
 978 /*
 979  * Recovery functions
 980  */
 981
 982 struct ptlrpc_request *ptlrpc_clone_req( struct ptlrpc_request *orig_req)
 983 {
 984         struct ptlrpc_request *copy_req;
 985         struct lustre_msg *copy_reqmsg;
 986         struct ptlrpc_user_desc *udesc = NULL;
 987
 988         OBD_ALLOC_PTR(copy_req);
 989         if (!copy_req)
 990                 return NULL;
 991         OBD_ALLOC(copy_reqmsg, orig_req->rq_reqlen);
 992         if (!copy_reqmsg){
 993                 OBD_FREE_PTR(copy_req);
 994                 return NULL;
 995         }
 996
 997         if (orig_req->rq_user_desc) {
 998                 int ngroups = orig_req->rq_user_desc->pud_ngroups;
 999
1000                 OBD_ALLOC(udesc, sptlrpc_user_desc_size(ngroups));
1001                 if (!udesc) {
1002                         OBD_FREE(copy_reqmsg, orig_req->rq_reqlen);
1003                         OBD_FREE_PTR(copy_req);
1004                         return NULL;
1005                 }
1006                 memcpy(udesc, orig_req->rq_user_desc,
1007                        sptlrpc_user_desc_size(ngroups));
1008         }
1009
1010         *copy_req = *orig_req;
1011         memcpy(copy_reqmsg, orig_req->rq_reqmsg, orig_req->rq_reqlen);
1012         copy_req->rq_reqmsg = copy_reqmsg;
1013         copy_req->rq_user_desc = udesc;
1014
1015         class_export_get(copy_req->rq_export);
1016         CFS_INIT_LIST_HEAD(&copy_req->rq_list);
1017         sptlrpc_svc_ctx_addref(copy_req);
1018
1019         if (copy_req->rq_reply_state) {
1020                 /* the copied req takes over the reply state */
1021                 orig_req->rq_reply_state = NULL;
1022                 /* to catch further access */
1023                 orig_req->rq_repmsg = NULL;
1024                 orig_req->rq_replen = 0;
1025         }
1026
1027         return copy_req;
1028 }
1029
1030 void ptlrpc_free_clone( struct ptlrpc_request *req)
1031 {
1032         if (req->rq_reply_state) {
1033                 ptlrpc_rs_decref(req->rq_reply_state);
1034                 req->rq_reply_state = NULL;
1035         }
1036
1037         sptlrpc_svc_ctx_decref(req);
1038         class_export_put(req->rq_export);
1039         list_del(&req->rq_list);
1040
1041         if (req->rq_user_desc) {
1042                 int ngroups = req->rq_user_desc->pud_ngroups;
1043                 OBD_FREE(req->rq_user_desc, sptlrpc_user_desc_size(ngroups));
1044         }
1045         OBD_FREE(req->rq_reqmsg, req->rq_reqlen);
1046         OBD_FREE_PTR(req);
1047 }
1048
1049 static void target_finish_recovery(struct obd_device *obd)
1050 {
1051         ENTRY;
1052         LCONSOLE_INFO("%s: sending delayed replies to recovered clients\n",
1053                       obd->obd_name);
1054
1055         ldlm_reprocess_all_ns(obd->obd_namespace);
1056
1057         /* when recovery finished, cleanup orphans on mds and ost */
1058         if (OBT(obd) && OBP(obd, postrecov)) {
1059                 int rc = OBP(obd, postrecov)(obd);
1060                 LCONSOLE_WARN("%s: recovery %s: rc %d\n", obd->obd_name,
1061                               rc < 0 ? "failed" : "complete", rc);
1062         }
1063
1064         obd->obd_recovery_end = CURRENT_SECONDS;
1065         EXIT;
1066 }
1067
1068 static void abort_req_replay_queue(struct obd_device *obd)
1069 {
1070         struct ptlrpc_request *req, *n;
1071
1072         list_for_each_entry_safe(req, n, &obd->obd_req_replay_queue, rq_list) {
1073                 DEBUG_REQ(D_WARNING, req, "aborted:");
1074                 req->rq_status = -ENOTCONN;
1075                 if (ptlrpc_error(req)) {
1076                         DEBUG_REQ(D_ERROR, req,
1077                                   "failed abort_req_reply; skipping");
1078                 }
1079                 ptlrpc_free_clone(req);
1080         }
1081 }
1082
1083 static void abort_lock_replay_queue(struct obd_device *obd)
1084 {
1085         struct ptlrpc_request *req, *n;
1086
1087         list_for_each_entry_safe(req, n, &obd->obd_lock_replay_queue, rq_list){
1088                 DEBUG_REQ(D_ERROR, req, "aborted:");
1089                 req->rq_status = -ENOTCONN;
1090                 if (ptlrpc_error(req)) {
1091                         DEBUG_REQ(D_ERROR, req,
1092                                   "failed abort_lock_reply; skipping");
1093                 }
1094                 ptlrpc_free_clone(req);
1095         }
1096 }
1097
1098 /* Called from a cleanup function if the device is being cleaned up
1099    forcefully.  The exports should all have been disconnected already,
1100    the only thing left to do is
1101      - clear the recovery flags
1102      - cancel the timer
1103      - free queued requests and replies, but don't send replies
1104    Because the obd_stopping flag is set, no new requests should be received.
1105
1106 */
1107 void target_cleanup_recovery(struct obd_device *obd)
1108 {
1109         struct ptlrpc_request *req, *n;
1110         ENTRY;
1111
1112         LASSERT(obd->obd_stopping);
1113
1114         spin_lock_bh(&obd->obd_processing_task_lock);
1115         if (!obd->obd_recovering) {
1116                 spin_unlock_bh(&obd->obd_processing_task_lock);
1117                 EXIT;
1118                 return;
1119         }
1120         obd->obd_recovering = obd->obd_abort_recovery = 0;
1121         target_cancel_recovery_timer(obd);
1122         spin_unlock_bh(&obd->obd_processing_task_lock);
1123
1124         list_for_each_entry_safe(req, n, &obd->obd_req_replay_queue, rq_list) {
1125                 LASSERT (req->rq_reply_state == 0);
1126                 ptlrpc_free_clone(req);
1127         }
1128         list_for_each_entry_safe(req, n, &obd->obd_lock_replay_queue, rq_list){
1129                 LASSERT (req->rq_reply_state == 0);
1130                 ptlrpc_free_clone(req);
1131         }
1132         list_for_each_entry_safe(req, n, &obd->obd_final_req_queue, rq_list) {
1133                 LASSERT (req->rq_reply_state == 0);
1134                 ptlrpc_free_clone(req);
1135         }
1136
1137         EXIT;
1138 }
1139
1140 static void target_recovery_expired(unsigned long castmeharder)
1141 {
1142         struct obd_device *obd = (struct obd_device *)castmeharder;
1143         CERROR("%s: recovery timed out, aborting\n", obd->obd_name);
1144         spin_lock_bh(&obd->obd_processing_task_lock);
1145         if (obd->obd_recovering)
1146                 obd->obd_abort_recovery = 1;
1147         cfs_waitq_signal(&obd->obd_next_transno_waitq);
1148         spin_unlock_bh(&obd->obd_processing_task_lock);
1149 }
1150
1151
1152 /* obd_processing_task_lock should be held */
1153 void target_cancel_recovery_timer(struct obd_device *obd)
1154 {
1155         CDEBUG(D_HA, "%s: cancel recovery timer\n", obd->obd_name);
1156         cfs_timer_disarm(&obd->obd_recovery_timer);
1157 }
1158
1159 static void reset_recovery_timer(struct obd_device *obd)
1160 {
1161         spin_lock_bh(&obd->obd_processing_task_lock);
1162         if (!obd->obd_recovering) {
1163                 spin_unlock_bh(&obd->obd_processing_task_lock);
1164                 return;
1165         }
1166         cfs_timer_arm(&obd->obd_recovery_timer,
1167                       cfs_time_shift(OBD_RECOVERY_TIMEOUT));
1168         spin_unlock_bh(&obd->obd_processing_task_lock);
1169         CDEBUG(D_HA, "%s: timer will expire in %u seconds\n", obd->obd_name,
1170                OBD_RECOVERY_TIMEOUT);
1171         /* Only used for lprocfs_status */
1172         obd->obd_recovery_end = CURRENT_SECONDS + OBD_RECOVERY_TIMEOUT;
1173 }
1174
1175
1176 /* Only start it the first time called */
1177 void target_start_recovery_timer(struct obd_device *obd)
1178 {
1179         spin_lock_bh(&obd->obd_processing_task_lock);
1180         if (obd->obd_recovery_handler
1181             || timer_pending(&obd->obd_recovery_timer)) {
1182                 spin_unlock_bh(&obd->obd_processing_task_lock);
1183                 return;
1184         }
1185         CWARN("%s: starting recovery timer (%us)\n", obd->obd_name,
1186               OBD_RECOVERY_TIMEOUT);
1187         cfs_timer_init(&obd->obd_recovery_timer, target_recovery_expired, obd);
1188         spin_unlock_bh(&obd->obd_processing_task_lock);
1189
1190         reset_recovery_timer(obd);
1191 }
1192
1193 #ifdef __KERNEL__
1194 static int check_for_next_transno(struct obd_device *obd)
1195 {
1196         struct ptlrpc_request *req = NULL;
1197         int wake_up = 0, connected, completed, queue_len, max;
1198         __u64 next_transno, req_transno;
1199         ENTRY;
1200         spin_lock_bh(&obd->obd_processing_task_lock);
1201
1202         if (!list_empty(&obd->obd_req_replay_queue)) {
1203                 req = list_entry(obd->obd_req_replay_queue.next,
1204                                  struct ptlrpc_request, rq_list);
1205                 req_transno = lustre_msg_get_transno(req->rq_reqmsg);
1206         } else {
1207                 req_transno = 0;
1208         }
1209
1210         max = obd->obd_max_recoverable_clients;
1211         connected = obd->obd_connected_clients;
1212         completed = connected - obd->obd_recoverable_clients;
1213         queue_len = obd->obd_requests_queued_for_recovery;
1214         next_transno = obd->obd_next_recovery_transno;
1215
1216         CDEBUG(D_HA, "max: %d, connected: %d, completed: %d, queue_len: %d, "
1217                "req_transno: "LPU64", next_transno: "LPU64"\n",
1218                max, connected, completed, queue_len, req_transno, next_transno);
1219
1220         if (obd->obd_abort_recovery) {
1221                 CDEBUG(D_HA, "waking for aborted recovery\n");
1222                 wake_up = 1;
1223         } else if (atomic_read(&obd->obd_req_replay_clients) == 0) {
1224                 CDEBUG(D_HA, "waking for completed recovery\n");
1225                 wake_up = 1;
1226         } else if (req_transno == next_transno) {
1227                 CDEBUG(D_HA, "waking for next ("LPD64")\n", next_transno);
1228                 wake_up = 1;
1229         } else if (queue_len + completed == max) {
1230                 /* handle gaps occured due to lost reply. It is allowed gaps
1231                  * because all clients are connected and there will be resend
1232                  * for missed transaction */
1233                 LASSERTF(req_transno >= next_transno,
1234                          "req_transno: "LPU64", next_transno: "LPU64"\n",
1235                          req_transno, next_transno);
1236
1237                 CDEBUG(req_transno > obd->obd_last_committed ? D_ERROR : D_HA,
1238                        "waking for skipped transno (skip: "LPD64
1239                        ", ql: %d, comp: %d, conn: %d, next: "LPD64")\n",
1240                        next_transno, queue_len, completed, connected, req_transno);
1241                 obd->obd_next_recovery_transno = req_transno;
1242                 wake_up = 1;
1243         } else if (queue_len == atomic_read(&obd->obd_req_replay_clients)) {
1244                 /* some clients haven't connected in time, but we can try
1245                  * to replay requests that demand on already committed ones
1246                  * also, we can replay first non-committed transation */
1247                 LASSERT(req_transno != 0);
1248                 if (req_transno == obd->obd_last_committed + 1) {
1249                         obd->obd_next_recovery_transno = req_transno;
1250                 } else if (req_transno > obd->obd_last_committed) {
1251                         /* can't continue recovery: have no needed transno */
1252                         obd->obd_abort_recovery = 1;
1253                         CDEBUG(D_ERROR, "abort due to missed clients. max: %d, "
1254                                "connected: %d, completed: %d, queue_len: %d, "
1255                                "req_transno: "LPU64", next_transno: "LPU64"\n",
1256                                max, connected, completed, queue_len,
1257                                req_transno, next_transno);
1258                 }
1259                 wake_up = 1;
1260         }
1261
1262         spin_unlock_bh(&obd->obd_processing_task_lock);
1263         return wake_up;
1264 }
1265
1266 static struct ptlrpc_request *target_next_replay_req(struct obd_device *obd)
1267 {
1268         struct l_wait_info lwi = { 0 };
1269         struct ptlrpc_request *req;
1270
1271         CDEBUG(D_HA, "Waiting for transno "LPD64"\n",
1272                obd->obd_next_recovery_transno);
1273         l_wait_event(obd->obd_next_transno_waitq,
1274                      check_for_next_transno(obd), &lwi);
1275
1276         spin_lock_bh(&obd->obd_processing_task_lock);
1277         if (obd->obd_abort_recovery) {
1278                 req = NULL;
1279         } else if (!list_empty(&obd->obd_req_replay_queue)) {
1280                 req = list_entry(obd->obd_req_replay_queue.next,
1281                                  struct ptlrpc_request, rq_list);
1282                 list_del_init(&req->rq_list);
1283                 obd->obd_requests_queued_for_recovery--;
1284         } else {
1285                 req = NULL;
1286         }
1287         spin_unlock_bh(&obd->obd_processing_task_lock);
1288         RETURN(req);
1289 }
1290
1291 static int check_for_next_lock(struct obd_device *obd)
1292 {
1293         struct ptlrpc_request *req = NULL;
1294         int wake_up = 0;
1295
1296         spin_lock_bh(&obd->obd_processing_task_lock);
1297         if (!list_empty(&obd->obd_lock_replay_queue)) {
1298                 req = list_entry(obd->obd_lock_replay_queue.next,
1299                                  struct ptlrpc_request, rq_list);
1300                 CDEBUG(D_HA, "waking for next lock\n");
1301                 wake_up = 1;
1302         } else if (atomic_read(&obd->obd_lock_replay_clients) == 0) {
1303                 CDEBUG(D_HA, "waking for completed lock replay\n");
1304                 wake_up = 1;
1305         } else if (obd->obd_abort_recovery) {
1306                 CDEBUG(D_HA, "waking for aborted recovery\n");
1307                 wake_up = 1;
1308         }
1309         spin_unlock_bh(&obd->obd_processing_task_lock);
1310
1311         return wake_up;
1312 }
1313
1314 static struct ptlrpc_request *target_next_replay_lock(struct obd_device *obd)
1315 {
1316         struct l_wait_info lwi = { 0 };
1317         struct ptlrpc_request *req;
1318
1319         CDEBUG(D_HA, "Waiting for lock\n");
1320         l_wait_event(obd->obd_next_transno_waitq,
1321                      check_for_next_lock(obd), &lwi);
1322
1323         spin_lock_bh(&obd->obd_processing_task_lock);
1324         if (obd->obd_abort_recovery) {
1325                 req = NULL;
1326         } else if (!list_empty(&obd->obd_lock_replay_queue)) {
1327                 req = list_entry(obd->obd_lock_replay_queue.next,
1328                                  struct ptlrpc_request, rq_list);
1329                 list_del_init(&req->rq_list);
1330         } else {
1331                 req = NULL;
1332         }
1333         spin_unlock_bh(&obd->obd_processing_task_lock);
1334         return req;
1335 }
1336
1337 static struct ptlrpc_request *target_next_final_ping(struct obd_device *obd)
1338 {
1339         struct ptlrpc_request *req;
1340
1341         spin_lock_bh(&obd->obd_processing_task_lock);
1342         if (!list_empty(&obd->obd_final_req_queue)) {
1343                 req = list_entry(obd->obd_final_req_queue.next,
1344                                  struct ptlrpc_request, rq_list);
1345                 list_del_init(&req->rq_list);
1346         } else {
1347                 req = NULL;
1348         }
1349         spin_unlock_bh(&obd->obd_processing_task_lock);
1350         return req;
1351 }
1352
1353 static inline int req_replay_done(struct obd_export *exp)
1354 {
1355         return (exp->exp_req_replay_needed == 0);
1356 }
1357
1358 static inline int lock_replay_done(struct obd_export *exp)
1359 {
1360         return (exp->exp_lock_replay_needed == 0);
1361 }
1362
1363 static inline int connect_done(struct obd_export *exp)
1364 {
1365         return (exp->exp_in_recovery != 0);
1366 }
1367
1368 static int check_for_clients(struct obd_device *obd)
1369 {
1370         if (obd->obd_abort_recovery)
1371                 return 1;
1372         LASSERT(obd->obd_connected_clients <= obd->obd_max_recoverable_clients);
1373         if (obd->obd_no_conn == 0 &&
1374             obd->obd_connected_clients == obd->obd_max_recoverable_clients)
1375                 return 1;
1376         return 0;
1377 }
1378
1379 static int handle_recovery_req(struct ptlrpc_thread *thread,
1380                                struct ptlrpc_request *req,
1381                                svc_handler_t handler)
1382 {
1383         int rc;
1384         ENTRY;
1385
1386         rc = lu_context_init(&req->rq_session, LCT_SESSION);
1387         if (rc) {
1388                 CERROR("Failure to initialize session: %d\n", rc);
1389                 return rc;
1390         }
1391         req->rq_session.lc_thread = thread;
1392         lu_context_enter(&req->rq_session);
1393         req->rq_svc_thread = thread;
1394         req->rq_svc_thread->t_env->le_ses = &req->rq_session;
1395
1396         (void)handler(req);
1397
1398         lu_context_exit(&req->rq_session);
1399         lu_context_fini(&req->rq_session);
1400         /* don't reset timer for final stage */
1401         if (!req_replay_done(req->rq_export) ||
1402             !lock_replay_done(req->rq_export))
1403                 reset_recovery_timer(class_exp2obd(req->rq_export));
1404         ptlrpc_free_clone(req);
1405         RETURN(0);
1406 }
1407
1408 static int target_recovery_thread(void *arg)
1409 {
1410         struct obd_device *obd = arg;
1411         struct ptlrpc_request *req;
1412         struct target_recovery_data *trd = &obd->obd_recovery_data;
1413         struct l_wait_info lwi = { 0 };
1414         unsigned long delta;
1415         unsigned long flags;
1416         struct lu_env env;
1417         struct ptlrpc_thread fake_svc_thread, *thread = &fake_svc_thread;
1418         __u32 recov_ctx_tags = LCT_MD_THREAD;
1419         int rc = 0;
1420         ENTRY;
1421
1422         cfs_daemonize("tgt_recov");
1423
1424         SIGNAL_MASK_LOCK(current, flags);
1425         sigfillset(&current->blocked);
1426         RECALC_SIGPENDING;
1427         SIGNAL_MASK_UNLOCK(current, flags);
1428
1429         rc = lu_context_init(&env.le_ctx, recov_ctx_tags);
1430         if (rc)
1431                 RETURN(rc);
1432
1433         thread->t_env = &env;
1434         env.le_ctx.lc_thread = thread;
1435
1436         CERROR("%s: started recovery thread pid %d\n", obd->obd_name,
1437                current->pid);
1438         trd->trd_processing_task = current->pid;
1439
1440         obd->obd_recovering = 1;
1441         complete(&trd->trd_starting);
1442
1443         /* first of all, we have to know the first transno to replay */
1444         obd->obd_abort_recovery = 0;
1445         l_wait_event(obd->obd_next_transno_waitq,
1446                      check_for_clients(obd), &lwi);
1447
1448         spin_lock_bh(&obd->obd_processing_task_lock);
1449         target_cancel_recovery_timer(obd);
1450         spin_unlock_bh(&obd->obd_processing_task_lock);
1451
1452         /* If some clients haven't connected in time, evict them */
1453         if (obd->obd_abort_recovery) {
1454                 CWARN("Some clients haven't connect in time (%d/%d),"
1455                        "evict them\n", obd->obd_connected_clients,
1456                        obd->obd_max_recoverable_clients);
1457                 obd->obd_abort_recovery = obd->obd_stopping;
1458                 class_disconnect_stale_exports(obd, connect_done);
1459         }
1460         /* next stage: replay requests */
1461         delta = jiffies;
1462         obd->obd_req_replaying = 1;
1463         CDEBUG(D_INFO, "1: request replay stage - %d clients from t"LPU64"\n",
1464               atomic_read(&obd->obd_req_replay_clients),
1465               obd->obd_next_recovery_transno);
1466         while ((req = target_next_replay_req(obd))) {
1467                 LASSERT(trd->trd_processing_task == current->pid);
1468                 DEBUG_REQ(D_HA, req, "processing t"LPD64" from %s",
1469                           lustre_msg_get_transno(req->rq_reqmsg),
1470                           libcfs_nid2str(req->rq_peer.nid));
1471
1472                 handle_recovery_req(thread, req,
1473                                     trd->trd_recovery_handler);
1474                 obd->obd_replayed_requests++;
1475                 spin_lock_bh(&obd->obd_processing_task_lock);
1476                 obd->obd_next_recovery_transno++;
1477                 spin_unlock_bh(&obd->obd_processing_task_lock);
1478         }
1479
1480         spin_lock_bh(&obd->obd_processing_task_lock);
1481         target_cancel_recovery_timer(obd);
1482         spin_unlock_bh(&obd->obd_processing_task_lock);
1483
1484         /* If some clients haven't replayed requests in time, evict them */
1485         if (obd->obd_abort_recovery) {
1486                 CDEBUG(D_ERROR, "req replay timed out, aborting ...\n");
1487                 obd->obd_abort_recovery = obd->obd_stopping;
1488                 class_disconnect_stale_exports(obd, req_replay_done);
1489                 abort_req_replay_queue(obd);
1490         }
1491         /* The second stage: replay locks */
1492         CDEBUG(D_INFO, "2: lock replay stage - %d clients\n",
1493                atomic_read(&obd->obd_lock_replay_clients));
1494         while ((req = target_next_replay_lock(obd))) {
1495                 LASSERT(trd->trd_processing_task == current->pid);
1496                 DEBUG_REQ(D_HA|D_WARNING, req, "processing lock from %s: ",
1497                           libcfs_nid2str(req->rq_peer.nid));
1498                 handle_recovery_req(thread, req,
1499                                     trd->trd_recovery_handler);
1500                 obd->obd_replayed_locks++;
1501         }
1502
1503         spin_lock_bh(&obd->obd_processing_task_lock);
1504         target_cancel_recovery_timer(obd);
1505         spin_unlock_bh(&obd->obd_processing_task_lock);
1506         /* If some clients haven't replayed requests in time, evict them */
1507         if (obd->obd_abort_recovery) {
1508                 int stale;
1509                 CERROR("lock replay timed out, aborting ...\n");
1510                 obd->obd_abort_recovery = obd->obd_stopping;
1511                 stale = class_disconnect_stale_exports(obd, lock_replay_done);
1512                 abort_lock_replay_queue(obd);
1513         }
1514
1515         /* We drop recoverying flag to forward all new requests
1516          * to regular mds_handle() since now */
1517         spin_lock_bh(&obd->obd_processing_task_lock);
1518         obd->obd_recovering = obd->obd_abort_recovery = 0;
1519         spin_unlock_bh(&obd->obd_processing_task_lock);
1520         /* The third stage: reply on final pings */
1521         CDEBUG(D_INFO, "3: final stage - process recovery completion pings\n");
1522         while ((req = target_next_final_ping(obd))) {
1523                 LASSERT(trd->trd_processing_task == current->pid);
1524                 DEBUG_REQ(D_HA, req, "processing final ping from %s: ",
1525                           libcfs_nid2str(req->rq_peer.nid));
1526                 handle_recovery_req(thread, req,
1527                                     trd->trd_recovery_handler);
1528         }
1529
1530         delta = (jiffies - delta) / HZ;
1531         CDEBUG(D_INFO,"4: recovery completed in %lus - %d/%d reqs/locks\n",
1532               delta, obd->obd_replayed_requests, obd->obd_replayed_locks);
1533         LASSERT(atomic_read(&obd->obd_req_replay_clients) == 0);
1534         LASSERT(atomic_read(&obd->obd_lock_replay_clients) == 0);
1535         if (delta > obd_timeout * 2) {
1536                 CWARN("too long recovery - read logs\n");
1537                 libcfs_debug_dumplog();
1538         }
1539
1540         target_finish_recovery(obd);
1541
1542         lu_env_fini(&env);
1543         trd->trd_processing_task = 0;
1544         complete(&trd->trd_finishing);
1545         RETURN(rc);
1546 }
1547
1548 int target_start_recovery_thread(struct obd_device *obd, svc_handler_t handler)
1549 {
1550         int rc = 0;
1551         struct target_recovery_data *trd = &obd->obd_recovery_data;
1552
1553         memset(trd, 0, sizeof(*trd));
1554         init_completion(&trd->trd_starting);
1555         init_completion(&trd->trd_finishing);
1556         trd->trd_recovery_handler = handler;
1557
1558         if (kernel_thread(target_recovery_thread, obd, 0) > 0) {
1559                 wait_for_completion(&trd->trd_starting);
1560                 LASSERT(obd->obd_recovering != 0);
1561         } else
1562                 rc = -ECHILD;
1563
1564         return rc;
1565 }
1566
1567 void target_stop_recovery_thread(struct obd_device *obd)
1568 {
1569         spin_lock_bh(&obd->obd_processing_task_lock);
1570         if (obd->obd_recovery_data.trd_processing_task > 0) {
1571                 struct target_recovery_data *trd = &obd->obd_recovery_data;
1572                 CERROR("%s: Aborting recovery\n", obd->obd_name);
1573                 obd->obd_abort_recovery = 1;
1574                 wake_up(&obd->obd_next_transno_waitq);
1575                 spin_unlock_bh(&obd->obd_processing_task_lock);
1576                 wait_for_completion(&trd->trd_finishing);
1577         } else {
1578                 spin_unlock_bh(&obd->obd_processing_task_lock);
1579         }
1580 }
1581
1582 void target_recovery_fini(struct obd_device *obd)
1583 {
1584         class_disconnect_exports(obd);
1585         target_stop_recovery_thread(obd);
1586         target_cleanup_recovery(obd);
1587 }
1588 EXPORT_SYMBOL(target_recovery_fini);
1589
1590 void target_recovery_init(struct obd_device *obd, svc_handler_t handler)
1591 {
1592         if (obd->obd_max_recoverable_clients == 0)
1593                 return;
1594
1595         CWARN("RECOVERY: service %s, %d recoverable clients, "
1596               "last_transno "LPU64"\n", obd->obd_name,
1597               obd->obd_max_recoverable_clients, obd->obd_last_committed);
1598         obd->obd_next_recovery_transno = obd->obd_last_committed + 1;
1599         target_start_recovery_thread(obd, handler);
1600         obd->obd_recovery_start = CURRENT_SECONDS;
1601         /* Only used for lprocfs_status */
1602         obd->obd_recovery_end = obd->obd_recovery_start + OBD_RECOVERY_TIMEOUT;
1603 }
1604 EXPORT_SYMBOL(target_recovery_init);
1605
1606 #endif
1607
1608 int target_process_req_flags(struct obd_device *obd, struct ptlrpc_request *req)
1609 {
1610         struct obd_export *exp = req->rq_export;
1611         LASSERT(exp != NULL);
1612         if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REQ_REPLAY_DONE) {
1613                 /* client declares he's ready to replay locks */
1614                 spin_lock_bh(&obd->obd_processing_task_lock);
1615                 if (exp->exp_req_replay_needed) {
1616                         LASSERT(atomic_read(&obd->obd_req_replay_clients) > 0);
1617                         spin_lock(&exp->exp_lock);
1618                         exp->exp_req_replay_needed = 0;
1619                         spin_unlock(&exp->exp_lock);
1620                         atomic_dec(&obd->obd_req_replay_clients);
1621                         LASSERT(obd->obd_recoverable_clients > 0);
1622                         obd->obd_recoverable_clients--;
1623                         if (atomic_read(&obd->obd_req_replay_clients) == 0)
1624                                 CDEBUG(D_HA, "all clients have replayed reqs\n");
1625                         wake_up(&obd->obd_next_transno_waitq);
1626                 }
1627                 spin_unlock_bh(&obd->obd_processing_task_lock);
1628         }
1629         if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_LOCK_REPLAY_DONE) {
1630                 /* client declares he's ready to complete recovery
1631                  * so, we put the request on th final queue */
1632                 spin_lock_bh(&obd->obd_processing_task_lock);
1633                 if (exp->exp_lock_replay_needed) {
1634                         LASSERT(atomic_read(&obd->obd_lock_replay_clients) > 0);
1635                         spin_lock(&exp->exp_lock);
1636                         exp->exp_lock_replay_needed = 0;
1637                         spin_unlock(&exp->exp_lock);
1638                         atomic_dec(&obd->obd_lock_replay_clients);
1639                         if (atomic_read(&obd->obd_lock_replay_clients) == 0)
1640                                 CDEBUG(D_HA, "all clients have replayed locks\n");
1641                         wake_up(&obd->obd_next_transno_waitq);
1642                 }
1643                 spin_unlock_bh(&obd->obd_processing_task_lock);
1644         }
1645
1646         return 0;
1647 }
1648
1649 int target_queue_recovery_request(struct ptlrpc_request *req,
1650                                   struct obd_device *obd)
1651 {
1652         struct list_head *tmp;
1653         int inserted = 0;
1654         __u64 transno = lustre_msg_get_transno(req->rq_reqmsg);
1655
1656         ENTRY;
1657
1658         if (obd->obd_recovery_data.trd_processing_task == cfs_curproc_pid()) {
1659                 /* Processing the queue right now, don't re-add. */
1660                 RETURN(1);
1661         }
1662
1663         target_process_req_flags(obd, req);
1664
1665         if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_LOCK_REPLAY_DONE) {
1666                 /* client declares he's ready to complete recovery
1667                  * so, we put the request on th final queue */
1668                 req = ptlrpc_clone_req(req);
1669                 if (req == NULL)
1670                         RETURN(-ENOMEM);
1671                 DEBUG_REQ(D_HA, req, "queue final req");
1672                 spin_lock_bh(&obd->obd_processing_task_lock);
1673                 if (obd->obd_recovering)
1674                         list_add_tail(&req->rq_list, &obd->obd_final_req_queue);
1675                 else {
1676                         spin_unlock_bh(&obd->obd_processing_task_lock);
1677                         ptlrpc_free_clone(req);
1678                         if (obd->obd_stopping) {
1679                                 RETURN(-ENOTCONN);
1680                         } else {
1681                                 RETURN(1);
1682                         }
1683                 }
1684                 spin_unlock_bh(&obd->obd_processing_task_lock);
1685                 RETURN(0);
1686         }
1687         if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REQ_REPLAY_DONE) {
1688                 /* client declares he's ready to replay locks */
1689                 req = ptlrpc_clone_req(req);
1690                 if (req == NULL)
1691                         RETURN(-ENOMEM);
1692                 DEBUG_REQ(D_HA, req, "queue lock replay req");
1693                 spin_lock_bh(&obd->obd_processing_task_lock);
1694                 LASSERT(obd->obd_recovering);
1695                 /* usually due to recovery abort */
1696                 if (!req->rq_export->exp_in_recovery) {
1697                         spin_unlock_bh(&obd->obd_processing_task_lock);
1698                         ptlrpc_free_clone(req);
1699                         RETURN(-ENOTCONN);
1700                 }
1701                 LASSERT(req->rq_export->exp_lock_replay_needed);
1702                 list_add_tail(&req->rq_list, &obd->obd_lock_replay_queue);
1703                 spin_unlock_bh(&obd->obd_processing_task_lock);
1704                 wake_up(&obd->obd_next_transno_waitq);
1705                 RETURN(0);
1706         }
1707
1708         /* CAVEAT EMPTOR: The incoming request message has been swabbed
1709          * (i.e. buflens etc are in my own byte order), but type-dependent
1710          * buffers (eg mds_body, ost_body etc) have NOT been swabbed. */
1711
1712         if (!transno) {
1713                 CFS_INIT_LIST_HEAD(&req->rq_list);
1714                 DEBUG_REQ(D_HA, req, "not queueing");
1715                 RETURN(1);
1716         }
1717
1718         spin_lock_bh(&obd->obd_processing_task_lock);
1719
1720         /* If we're processing the queue, we want don't want to queue this
1721          * message.
1722          *
1723          * Also, if this request has a transno less than the one we're waiting
1724          * for, we should process it now.  It could (and currently always will)
1725          * be an open request for a descriptor that was opened some time ago.
1726          *
1727          * Also, a resent, replayed request that has already been
1728          * handled will pass through here and be processed immediately.
1729          */
1730         CWARN("Next recovery transno: "LPU64", current: "LPU64", replaying: %i\n",
1731               obd->obd_next_recovery_transno, transno, obd->obd_req_replaying);
1732         if (transno < obd->obd_next_recovery_transno && obd->obd_req_replaying) {
1733                 /* Processing the queue right now, don't re-add. */
1734                 LASSERT(list_empty(&req->rq_list));
1735                 spin_unlock_bh(&obd->obd_processing_task_lock);
1736                 RETURN(1);
1737         }
1738         spin_unlock_bh(&obd->obd_processing_task_lock);
1739
1740         /* A resent, replayed request that is still on the queue; just drop it.
1741            The queued request will handle this. */
1742         if ((lustre_msg_get_flags(req->rq_reqmsg) & (MSG_RESENT|MSG_REPLAY)) ==
1743             (MSG_RESENT | MSG_REPLAY)) {
1744                 DEBUG_REQ(D_ERROR, req, "dropping resent queued req");
1745                 RETURN(0);
1746         }
1747
1748         req = ptlrpc_clone_req(req);
1749         if (req == NULL)
1750                 RETURN(-ENOMEM);
1751
1752         spin_lock_bh(&obd->obd_processing_task_lock);
1753         LASSERT(obd->obd_recovering);
1754         if (!req->rq_export->exp_in_recovery) {
1755                 spin_unlock_bh(&obd->obd_processing_task_lock);
1756                 ptlrpc_free_clone(req);
1757                 RETURN(-ENOTCONN);
1758         }
1759         LASSERT(req->rq_export->exp_req_replay_needed);
1760
1761         /* XXX O(n^2) */
1762         list_for_each(tmp, &obd->obd_req_replay_queue) {
1763                 struct ptlrpc_request *reqiter =
1764                         list_entry(tmp, struct ptlrpc_request, rq_list);
1765
1766                 if (lustre_msg_get_transno(reqiter->rq_reqmsg) > transno) {
1767                         list_add_tail(&req->rq_list, &reqiter->rq_list);
1768                         inserted = 1;
1769                         break;
1770                 }
1771         }
1772
1773         if (!inserted)
1774                 list_add_tail(&req->rq_list, &obd->obd_req_replay_queue);
1775
1776         obd->obd_requests_queued_for_recovery++;
1777         wake_up(&obd->obd_next_transno_waitq);
1778         spin_unlock_bh(&obd->obd_processing_task_lock);
1779         RETURN(0);
1780
1781 }
1782
1783 struct obd_device * target_req2obd(struct ptlrpc_request *req)
1784 {
1785         return req->rq_export->exp_obd;
1786 }
1787
1788 int target_send_reply_msg(struct ptlrpc_request *req, int rc, int fail_id)
1789 {
1790         if (OBD_FAIL_CHECK(fail_id | OBD_FAIL_ONCE)) {
1791                 obd_fail_loc |= OBD_FAIL_ONCE | OBD_FAILED;
1792                 DEBUG_REQ(D_ERROR, req, "dropping reply");
1793                 return (-ECOMM);
1794         }
1795
1796         if (unlikely(rc)) {
1797                 DEBUG_REQ(D_ERROR, req, "processing error (%d)", rc);
1798                 req->rq_status = rc;
1799                 return (ptlrpc_error(req));
1800         } else {
1801                 DEBUG_REQ(D_NET, req, "sending reply");
1802         }
1803
1804         return (ptlrpc_send_reply(req, 1));
1805 }
1806
1807 void target_send_reply(struct ptlrpc_request *req, int rc, int fail_id)
1808 {
1809         int                        netrc;
1810         struct ptlrpc_reply_state *rs;
1811         struct obd_device         *obd;
1812         struct obd_export         *exp;
1813         struct ptlrpc_service     *svc;
1814
1815         svc = req->rq_rqbd->rqbd_service;
1816         rs = req->rq_reply_state;
1817         if (rs == NULL || !rs->rs_difficult) {
1818                 /* no notifiers */
1819                 target_send_reply_msg (req, rc, fail_id);
1820                 return;
1821         }
1822
1823         /* must be an export if locks saved */
1824         LASSERT (req->rq_export != NULL);
1825         /* req/reply consistent */
1826         LASSERT (rs->rs_service == svc);
1827
1828         /* "fresh" reply */
1829         LASSERT (!rs->rs_scheduled);
1830         LASSERT (!rs->rs_scheduled_ever);
1831         LASSERT (!rs->rs_handled);
1832         LASSERT (!rs->rs_on_net);
1833         LASSERT (rs->rs_export == NULL);
1834         LASSERT (list_empty(&rs->rs_obd_list));
1835         LASSERT (list_empty(&rs->rs_exp_list));
1836
1837         exp = class_export_get (req->rq_export);
1838         obd = exp->exp_obd;
1839
1840         /* disable reply scheduling onto srv_reply_queue while I'm setting up */
1841         rs->rs_scheduled = 1;
1842         rs->rs_on_net    = 1;
1843         rs->rs_xid       = req->rq_xid;
1844         rs->rs_transno   = req->rq_transno;
1845         rs->rs_export    = exp;
1846
1847         spin_lock(&obd->obd_uncommitted_replies_lock);
1848
1849         if (rs->rs_transno > obd->obd_last_committed) {
1850                 /* not committed already */
1851                 list_add_tail (&rs->rs_obd_list,
1852                                &obd->obd_uncommitted_replies);
1853         }
1854
1855         spin_unlock (&obd->obd_uncommitted_replies_lock);
1856         spin_lock (&exp->exp_lock);
1857
1858         list_add_tail (&rs->rs_exp_list, &exp->exp_outstanding_replies);
1859
1860         spin_unlock(&exp->exp_lock);
1861
1862         netrc = target_send_reply_msg (req, rc, fail_id);
1863
1864         spin_lock(&svc->srv_lock);
1865
1866         svc->srv_n_difficult_replies++;
1867
1868         if (netrc != 0) {
1869                 /* error sending: reply is off the net.  Also we need +1
1870                  * reply ref until ptlrpc_server_handle_reply() is done
1871                  * with the reply state (if the send was successful, there
1872                  * would have been +1 ref for the net, which
1873                  * reply_out_callback leaves alone) */
1874                 rs->rs_on_net = 0;
1875                 ptlrpc_rs_addref(rs);
1876                 atomic_inc (&svc->srv_outstanding_replies);
1877         }
1878
1879         if (!rs->rs_on_net ||                   /* some notifier */
1880             list_empty(&rs->rs_exp_list) ||     /* completed already */
1881             list_empty(&rs->rs_obd_list)) {
1882                 list_add_tail (&rs->rs_list, &svc->srv_reply_queue);
1883                 cfs_waitq_signal (&svc->srv_waitq);
1884         } else {
1885                 list_add (&rs->rs_list, &svc->srv_active_replies);
1886                 rs->rs_scheduled = 0;           /* allow notifier to schedule */
1887         }
1888
1889         spin_unlock(&svc->srv_lock);
1890 }
1891
1892 int target_handle_ping(struct ptlrpc_request *req)
1893 {
1894         obd_ping(req->rq_export);
1895         return lustre_pack_reply(req, 1, NULL, NULL);
1896 }
1897
1898 void target_committed_to_req(struct ptlrpc_request *req)
1899 {
1900         struct obd_device *obd;
1901
1902         if (req == NULL || req->rq_export == NULL)
1903                 return;
1904
1905         obd = req->rq_export->exp_obd;
1906         if (obd == NULL)
1907                 return;
1908
1909         if (!obd->obd_no_transno && req->rq_repmsg != NULL)
1910                 lustre_msg_set_last_committed(req->rq_repmsg,
1911                                               obd->obd_last_committed);
1912         else
1913                 DEBUG_REQ(D_IOCTL, req, "not sending last_committed update");
1914
1915         CDEBUG(D_INFO, "last_committed "LPU64", transno "LPU64", xid "LPU64"\n",
1916                obd->obd_last_committed, req->rq_transno, req->rq_xid);
1917 }
1918
1919 EXPORT_SYMBOL(target_committed_to_req);
1920
1921 #ifdef HAVE_QUOTA_SUPPORT
1922 int target_handle_qc_callback(struct ptlrpc_request *req)
1923 {
1924         struct obd_quotactl *oqctl;
1925         struct client_obd *cli = &req->rq_export->exp_obd->u.cli;
1926
1927         oqctl = lustre_swab_reqbuf(req, REQ_REC_OFF, sizeof(*oqctl),
1928                                    lustre_swab_obd_quotactl);
1929         if (oqctl == NULL) {
1930                 CERROR("Can't unpack obd_quotactl\n");
1931                 RETURN(-EPROTO);
1932         }
1933
1934         cli->cl_qchk_stat = oqctl->qc_stat;
1935
1936         return 0;
1937 }
1938
1939 int target_handle_dqacq_callback(struct ptlrpc_request *req)
1940 {
1941 #ifdef __KERNEL__
1942         struct obd_device *obd = req->rq_export->exp_obd;
1943         struct obd_device *master_obd;
1944         struct lustre_quota_ctxt *qctxt;
1945         struct qunit_data *qdata;
1946         void* rep;
1947         struct qunit_data_old *qdata_old;
1948         int rc = 0;
1949         int repsize[2] = { sizeof(struct ptlrpc_body),
1950                            sizeof(struct qunit_data) };
1951         ENTRY;
1952
1953         rc = lustre_pack_reply(req, 2, repsize, NULL);
1954         if (rc) {
1955                 CERROR("packing reply failed!: rc = %d\n", rc);
1956                 RETURN(rc);
1957         }
1958         LASSERT(req->rq_export);
1959
1960         /* fixed for bug10707 */
1961         if ((req->rq_export->exp_connect_flags & OBD_CONNECT_QUOTA64) &&
1962             !OBD_FAIL_CHECK(OBD_FAIL_QUOTA_QD_COUNT_32BIT)) {
1963                 CDEBUG(D_QUOTA, "qd_count is 64bit!\n");
1964                 rep = lustre_msg_buf(req->rq_repmsg, REPLY_REC_OFF,
1965                                      sizeof(struct qunit_data));
1966                 LASSERT(rep);
1967                 qdata = lustre_swab_reqbuf(req, REQ_REC_OFF, sizeof(*qdata),
1968                                            lustre_swab_qdata);
1969         } else {
1970                 CDEBUG(D_QUOTA, "qd_count is 32bit!\n");
1971                 rep = lustre_msg_buf(req->rq_repmsg, REPLY_REC_OFF,
1972                                      sizeof(struct qunit_data_old));
1973                 LASSERT(rep);
1974                 qdata_old = lustre_swab_reqbuf(req, REQ_REC_OFF, sizeof(*qdata_old),
1975                                                lustre_swab_qdata_old);
1976                 qdata = lustre_quota_old_to_new(qdata_old);
1977         }
1978
1979         if (qdata == NULL) {
1980                 CERROR("Can't unpack qunit_data\n");
1981                 RETURN(-EPROTO);
1982         }
1983
1984         /* we use the observer */
1985         LASSERT(obd->obd_observer && obd->obd_observer->obd_observer);
1986         master_obd = obd->obd_observer->obd_observer;
1987         qctxt = &master_obd->u.obt.obt_qctxt;
1988
1989         LASSERT(qctxt->lqc_handler);
1990         rc = qctxt->lqc_handler(master_obd, qdata,
1991                                 lustre_msg_get_opc(req->rq_reqmsg));
1992         if (rc && rc != -EDQUOT)
1993                 CDEBUG(rc == -EBUSY  ? D_QUOTA : D_ERROR,
1994                        "dqacq failed! (rc:%d)\n", rc);
1995
1996         /* the qd_count might be changed in lqc_handler */
1997         if ((req->rq_export->exp_connect_flags & OBD_CONNECT_QUOTA64) &&
1998             !OBD_FAIL_CHECK(OBD_FAIL_QUOTA_QD_COUNT_32BIT)) {
1999                 memcpy(rep,qdata,sizeof(*qdata));
2000         } else {
2001                 qdata_old = lustre_quota_new_to_old(qdata);
2002                 memcpy(rep,qdata_old,sizeof(*qdata_old));
2003         }
2004         req->rq_status = rc;
2005         rc = ptlrpc_reply(req);
2006
2007         RETURN(rc);
2008 #else
2009         return 0;
2010 #endif /* !__KERNEL__ */
2011 }
2012 #endif /* HAVE_QUOTA_SUPPORT */
2013
2014 ldlm_mode_t lck_compat_array[] = {
2015         [LCK_EX] LCK_COMPAT_EX,
2016         [LCK_PW] LCK_COMPAT_PW,
2017         [LCK_PR] LCK_COMPAT_PR,
2018         [LCK_CW] LCK_COMPAT_CW,
2019         [LCK_CR] LCK_COMPAT_CR,
2020         [LCK_NL] LCK_COMPAT_NL,
2021         [LCK_GROUP] LCK_COMPAT_GROUP
2022 };