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 #include "ldlm_internal.h"
  41
  42
  43 /* @priority: if non-zero, move the selected to the list head
  44  * @create: if zero, only search in existed connections
  45  */
  46 static int import_set_conn(struct obd_import *imp, struct obd_uuid *uuid,
  47                            int priority, int create)
  48 {
  49         struct ptlrpc_connection *ptlrpc_conn;
  50         struct obd_import_conn *imp_conn = NULL, *item;
  51         int rc = 0;
  52         ENTRY;
  53
  54         if (!create && !priority) {
  55                 CDEBUG(D_HA, "Nothing to do\n");
  56                 RETURN(-EINVAL);
  57         }
  58
  59         ptlrpc_conn = ptlrpc_uuid_to_connection(uuid);
  60         if (!ptlrpc_conn) {
  61                 CDEBUG(D_HA, "can't find connection %s\n", uuid->uuid);
  62                 RETURN (-ENOENT);
  63         }
  64
  65         if (create) {
  66                 OBD_ALLOC(imp_conn, sizeof(*imp_conn));
  67                 if (!imp_conn) {
  68                         GOTO(out_put, rc = -ENOMEM);
  69                 }
  70         }
  71
  72         spin_lock(&imp->imp_lock);
  73         list_for_each_entry(item, &imp->imp_conn_list, oic_item) {
  74                 if (obd_uuid_equals(uuid, &item->oic_uuid)) {
  75                         if (priority) {
  76                                 list_del(&item->oic_item);
  77                                 list_add(&item->oic_item, &imp->imp_conn_list);
  78                                 item->oic_last_attempt = 0;
  79                         }
  80                         CDEBUG(D_HA, "imp %p@%s: found existing conn %s%s\n",
  81                                imp, imp->imp_obd->obd_name, uuid->uuid,
  82                                (priority ? ", moved to head" : ""));
  83                         spin_unlock(&imp->imp_lock);
  84                         GOTO(out_free, rc = 0);
  85                 }
  86         }
  87         /* not found */
  88         if (create) {
  89                 imp_conn->oic_conn = ptlrpc_conn;
  90                 imp_conn->oic_uuid = *uuid;
  91                 item->oic_last_attempt = 0;
  92                 if (priority)
  93                         list_add(&imp_conn->oic_item, &imp->imp_conn_list);
  94                 else
  95                         list_add_tail(&imp_conn->oic_item, &imp->imp_conn_list);
  96                 CDEBUG(D_HA, "imp %p@%s: add connection %s at %s\n",
  97                        imp, imp->imp_obd->obd_name, uuid->uuid,
  98                        (priority ? "head" : "tail"));
  99         } else {
 100                 spin_unlock(&imp->imp_lock);
 101                 GOTO(out_free, rc = -ENOENT);
 102         }
 103
 104         spin_unlock(&imp->imp_lock);
 105         RETURN(0);
 106 out_free:
 107         if (imp_conn)
 108                 OBD_FREE(imp_conn, sizeof(*imp_conn));
 109 out_put:
 110         ptlrpc_put_connection(ptlrpc_conn);
 111         RETURN(rc);
 112 }
 113
 114 int import_set_conn_priority(struct obd_import *imp, struct obd_uuid *uuid)
 115 {
 116         return import_set_conn(imp, uuid, 1, 0);
 117 }
 118
 119 int client_import_add_conn(struct obd_import *imp, struct obd_uuid *uuid,
 120                            int priority)
 121 {
 122         return import_set_conn(imp, uuid, priority, 1);
 123 }
 124
 125 int client_import_del_conn(struct obd_import *imp, struct obd_uuid *uuid)
 126 {
 127         struct obd_import_conn *imp_conn;
 128         struct obd_export *dlmexp;
 129         int rc = -ENOENT;
 130         ENTRY;
 131
 132         spin_lock(&imp->imp_lock);
 133         if (list_empty(&imp->imp_conn_list)) {
 134                 LASSERT(!imp->imp_connection);
 135                 GOTO(out, rc);
 136         }
 137
 138         list_for_each_entry(imp_conn, &imp->imp_conn_list, oic_item) {
 139                 if (!obd_uuid_equals(uuid, &imp_conn->oic_uuid))
 140                         continue;
 141                 LASSERT(imp_conn->oic_conn);
 142
 143                 /* is current conn? */
 144                 if (imp_conn == imp->imp_conn_current) {
 145                         LASSERT(imp_conn->oic_conn == imp->imp_connection);
 146
 147                         if (imp->imp_state != LUSTRE_IMP_CLOSED &&
 148                             imp->imp_state != LUSTRE_IMP_DISCON) {
 149                                 CERROR("can't remove current connection\n");
 150                                 GOTO(out, rc = -EBUSY);
 151                         }
 152
 153                         ptlrpc_put_connection(imp->imp_connection);
 154                         imp->imp_connection = NULL;
 155
 156                         dlmexp = class_conn2export(&imp->imp_dlm_handle);
 157                         if (dlmexp && dlmexp->exp_connection) {
 158                                 LASSERT(dlmexp->exp_connection ==
 159                                         imp_conn->oic_conn);
 160                                 ptlrpc_put_connection(dlmexp->exp_connection);
 161                                 dlmexp->exp_connection = NULL;
 162                         }
 163                 }
 164
 165                 list_del(&imp_conn->oic_item);
 166                 ptlrpc_put_connection(imp_conn->oic_conn);
 167                 OBD_FREE(imp_conn, sizeof(*imp_conn));
 168                 CDEBUG(D_HA, "imp %p@%s: remove connection %s\n",
 169                        imp, imp->imp_obd->obd_name, uuid->uuid);
 170                 rc = 0;
 171                 break;
 172         }
 173 out:
 174         spin_unlock(&imp->imp_lock);
 175         if (rc == -ENOENT)
 176                 CERROR("connection %s not found\n", uuid->uuid);
 177         RETURN(rc);
 178 }
 179
 180 static void destroy_import(struct obd_import *imp)
 181 {
 182         /* drop security policy instance after all rpc finished/aborted
 183          * to let all busy contexts be released. */
 184         class_import_get(imp);
 185         class_destroy_import(imp);
 186         sptlrpc_import_sec_put(imp);
 187         class_import_put(imp);
 188 }
 189
 190 /* configure an RPC client OBD device
 191  *
 192  * lcfg parameters:
 193  * 1 - client UUID
 194  * 2 - server UUID
 195  * 3 - inactive-on-startup
 196  */
 197 int client_obd_setup(struct obd_device *obddev, struct lustre_cfg *lcfg)
 198 {
 199         struct client_obd *cli = &obddev->u.cli;
 200         struct obd_import *imp;
 201         struct obd_uuid server_uuid;
 202         int rq_portal, rp_portal, connect_op;
 203         char *name = obddev->obd_type->typ_name;
 204         int rc;
 205         ENTRY;
 206
 207         /* In a more perfect world, we would hang a ptlrpc_client off of
 208          * obd_type and just use the values from there. */
 209         if (!strcmp(name, LUSTRE_OSC_NAME)) {
 210                 rq_portal = OST_REQUEST_PORTAL;
 211                 rp_portal = OSC_REPLY_PORTAL;
 212                 connect_op = OST_CONNECT;
 213         } else if (!strcmp(name, LUSTRE_MDC_NAME)) {
 214                 rq_portal = MDS_REQUEST_PORTAL;
 215                 rp_portal = MDC_REPLY_PORTAL;
 216                 connect_op = MDS_CONNECT;
 217         } else if (!strcmp(name, LUSTRE_MGC_NAME)) {
 218                 rq_portal = MGS_REQUEST_PORTAL;
 219                 rp_portal = MGC_REPLY_PORTAL;
 220                 connect_op = MGS_CONNECT;
 221         } else {
 222                 CERROR("unknown client OBD type \"%s\", can't setup\n",
 223                        name);
 224                 RETURN(-EINVAL);
 225         }
 226
 227         if (LUSTRE_CFG_BUFLEN(lcfg, 1) < 1) {
 228                 CERROR("requires a TARGET UUID\n");
 229                 RETURN(-EINVAL);
 230         }
 231
 232         if (LUSTRE_CFG_BUFLEN(lcfg, 1) > 37) {
 233                 CERROR("client UUID must be less than 38 characters\n");
 234                 RETURN(-EINVAL);
 235         }
 236
 237         if (LUSTRE_CFG_BUFLEN(lcfg, 2) < 1) {
 238                 CERROR("setup requires a SERVER UUID\n");
 239                 RETURN(-EINVAL);
 240         }
 241
 242         if (LUSTRE_CFG_BUFLEN(lcfg, 2) > 37) {
 243                 CERROR("target UUID must be less than 38 characters\n");
 244                 RETURN(-EINVAL);
 245         }
 246
 247         sema_init(&cli->cl_sem, 1);
 248         sema_init(&cli->cl_mgc_sem, 1);
 249         sptlrpc_rule_set_init(&cli->cl_sptlrpc_rset);
 250         cli->cl_sec_part = LUSTRE_SP_ANY;
 251         cli->cl_conn_count = 0;
 252         memcpy(server_uuid.uuid, lustre_cfg_buf(lcfg, 2),
 253                min_t(unsigned int, LUSTRE_CFG_BUFLEN(lcfg, 2),
 254                      sizeof(server_uuid)));
 255
 256         cli->cl_dirty = 0;
 257         cli->cl_avail_grant = 0;
 258         /* FIXME: should limit this for the sum of all cl_dirty_max */
 259         cli->cl_dirty_max = OSC_MAX_DIRTY_DEFAULT * 1024 * 1024;
 260         if (cli->cl_dirty_max >> CFS_PAGE_SHIFT > num_physpages / 8)
 261                 cli->cl_dirty_max = num_physpages << (CFS_PAGE_SHIFT - 3);
 262         CFS_INIT_LIST_HEAD(&cli->cl_cache_waiters);
 263         CFS_INIT_LIST_HEAD(&cli->cl_loi_ready_list);
 264         CFS_INIT_LIST_HEAD(&cli->cl_loi_write_list);
 265         CFS_INIT_LIST_HEAD(&cli->cl_loi_read_list);
 266         client_obd_list_lock_init(&cli->cl_loi_list_lock);
 267         cli->cl_r_in_flight = 0;
 268         cli->cl_w_in_flight = 0;
 269
 270         spin_lock_init(&cli->cl_read_rpc_hist.oh_lock);
 271         spin_lock_init(&cli->cl_write_rpc_hist.oh_lock);
 272         spin_lock_init(&cli->cl_read_page_hist.oh_lock);
 273         spin_lock_init(&cli->cl_write_page_hist.oh_lock);
 274         spin_lock_init(&cli->cl_read_offset_hist.oh_lock);
 275         spin_lock_init(&cli->cl_write_offset_hist.oh_lock);
 276         cfs_waitq_init(&cli->cl_destroy_waitq);
 277         atomic_set(&cli->cl_destroy_in_flight, 0);
 278 #ifdef ENABLE_CHECKSUM
 279         cli->cl_checksum = 1;
 280 #endif
 281         atomic_set(&cli->cl_resends, OSC_DEFAULT_RESENDS);
 282
 283         /* This value may be changed at connect time in
 284            ptlrpc_connect_interpret. */
 285         cli->cl_max_pages_per_rpc = min((int)PTLRPC_MAX_BRW_PAGES,
 286                                         (int)(1024 * 1024 >> CFS_PAGE_SHIFT));
 287
 288         if (!strcmp(name, LUSTRE_MDC_NAME)) {
 289                 cli->cl_max_rpcs_in_flight = MDC_MAX_RIF_DEFAULT;
 290         } else if (num_physpages >> (20 - CFS_PAGE_SHIFT) <= 128 /* MB */) {
 291                 cli->cl_max_rpcs_in_flight = 2;
 292         } else if (num_physpages >> (20 - CFS_PAGE_SHIFT) <= 256 /* MB */) {
 293                 cli->cl_max_rpcs_in_flight = 3;
 294         } else if (num_physpages >> (20 - CFS_PAGE_SHIFT) <= 512 /* MB */) {
 295                 cli->cl_max_rpcs_in_flight = 4;
 296         } else {
 297                 cli->cl_max_rpcs_in_flight = OSC_MAX_RIF_DEFAULT;
 298         }
 299
 300         rc = ldlm_get_ref();
 301         if (rc) {
 302                 CERROR("ldlm_get_ref failed: %d\n", rc);
 303                 GOTO(err, rc);
 304         }
 305
 306         ptlrpc_init_client(rq_portal, rp_portal, name,
 307                            &obddev->obd_ldlm_client);
 308
 309         imp = class_new_import(obddev);
 310         if (imp == NULL)
 311                 GOTO(err_ldlm, rc = -ENOENT);
 312         imp->imp_client = &obddev->obd_ldlm_client;
 313         imp->imp_connect_op = connect_op;
 314         imp->imp_initial_recov = 1;
 315         imp->imp_initial_recov_bk = 0;
 316         CFS_INIT_LIST_HEAD(&imp->imp_pinger_chain);
 317         memcpy(cli->cl_target_uuid.uuid, lustre_cfg_buf(lcfg, 1),
 318                LUSTRE_CFG_BUFLEN(lcfg, 1));
 319         class_import_put(imp);
 320
 321         rc = client_import_add_conn(imp, &server_uuid, 1);
 322         if (rc) {
 323                 CERROR("can't add initial connection\n");
 324                 GOTO(err_import, rc);
 325         }
 326
 327         cli->cl_import = imp;
 328         /* cli->cl_max_mds_{easize,cookiesize} updated by mdc_init_ea_size() */
 329         cli->cl_max_mds_easize = sizeof(struct lov_mds_md);
 330         cli->cl_max_mds_cookiesize = sizeof(struct llog_cookie);
 331
 332         if (LUSTRE_CFG_BUFLEN(lcfg, 3) > 0) {
 333                 if (!strcmp(lustre_cfg_string(lcfg, 3), "inactive")) {
 334                         CDEBUG(D_HA, "marking %s %s->%s as inactive\n",
 335                                name, obddev->obd_name,
 336                                cli->cl_target_uuid.uuid);
 337                         spin_lock(&imp->imp_lock);
 338                         imp->imp_invalid = 1;
 339                         spin_unlock(&imp->imp_lock);
 340                 }
 341         }
 342
 343         cli->cl_qchk_stat = CL_NOT_QUOTACHECKED;
 344
 345         RETURN(rc);
 346
 347 err_import:
 348         class_destroy_import(imp);
 349 err_ldlm:
 350         ldlm_put_ref(0);
 351 err:
 352         RETURN(rc);
 353
 354 }
 355
 356 int client_obd_cleanup(struct obd_device *obddev)
 357 {
 358         ENTRY;
 359         sptlrpc_rule_set_free(&obddev->u.cli.cl_sptlrpc_rset);
 360         ldlm_put_ref(obddev->obd_force);
 361         RETURN(0);
 362 }
 363
 364 /* ->o_connect() method for client side (OSC and MDC and MGC) */
 365 int client_connect_import(const struct lu_env *env,
 366                           struct lustre_handle *dlm_handle,
 367                           struct obd_device *obd, struct obd_uuid *cluuid,
 368                           struct obd_connect_data *data)
 369 {
 370         struct client_obd *cli = &obd->u.cli;
 371         struct obd_import *imp = cli->cl_import;
 372         struct obd_export *exp;
 373         struct obd_connect_data *ocd;
 374         struct ldlm_namespace *to_be_freed = NULL;
 375         int rc;
 376         ENTRY;
 377
 378         mutex_down(&cli->cl_sem);
 379         rc = class_connect(dlm_handle, obd, cluuid);
 380         if (rc)
 381                 GOTO(out_sem, rc);
 382
 383         cli->cl_conn_count++;
 384         if (cli->cl_conn_count > 1)
 385                 GOTO(out_sem, rc);
 386         exp = class_conn2export(dlm_handle);
 387
 388         if (obd->obd_namespace != NULL)
 389                 CERROR("already have namespace!\n");
 390         obd->obd_namespace = ldlm_namespace_new(obd->obd_name,
 391                                                 LDLM_NAMESPACE_CLIENT,
 392                                                 LDLM_NAMESPACE_GREEDY);
 393         if (obd->obd_namespace == NULL)
 394                 GOTO(out_disco, rc = -ENOMEM);
 395
 396         imp->imp_dlm_handle = *dlm_handle;
 397         rc = ptlrpc_init_import(imp);
 398         if (rc != 0)
 399                 GOTO(out_ldlm, rc);
 400
 401         ocd = &imp->imp_connect_data;
 402         if (data) {
 403                 *ocd = *data;
 404                 imp->imp_connect_flags_orig = data->ocd_connect_flags;
 405         }
 406
 407         rc = ptlrpc_connect_import(imp, NULL);
 408         if (rc != 0) {
 409                 LASSERT (imp->imp_state == LUSTRE_IMP_DISCON);
 410                 GOTO(out_ldlm, rc);
 411         }
 412         LASSERT(exp->exp_connection);
 413
 414         if (data) {
 415                 LASSERT((ocd->ocd_connect_flags & data->ocd_connect_flags) ==
 416                         ocd->ocd_connect_flags);
 417                 data->ocd_connect_flags = ocd->ocd_connect_flags;
 418         }
 419
 420         ptlrpc_pinger_add_import(imp);
 421
 422         EXIT;
 423
 424         if (rc) {
 425 out_ldlm:
 426                 ldlm_namespace_free_prior(obd->obd_namespace);
 427                 to_be_freed = obd->obd_namespace;
 428                 obd->obd_namespace = NULL;
 429 out_disco:
 430                 cli->cl_conn_count--;
 431                 class_disconnect(exp);
 432         } else {
 433                 class_export_put(exp);
 434         }
 435 out_sem:
 436         mutex_up(&cli->cl_sem);
 437         if (to_be_freed)
 438                 ldlm_namespace_free_post(to_be_freed, 1);
 439
 440         return rc;
 441 }
 442
 443 int client_disconnect_export(struct obd_export *exp)
 444 {
 445         struct obd_device *obd = class_exp2obd(exp);
 446         struct client_obd *cli;
 447         struct obd_import *imp;
 448         int rc = 0, err;
 449         struct ldlm_namespace *to_be_freed = NULL;
 450         ENTRY;
 451
 452         if (!obd) {
 453                 CERROR("invalid export for disconnect: exp %p cookie "LPX64"\n",
 454                        exp, exp ? exp->exp_handle.h_cookie : -1);
 455                 RETURN(-EINVAL);
 456         }
 457
 458         cli = &obd->u.cli;
 459         imp = cli->cl_import;
 460
 461         mutex_down(&cli->cl_sem);
 462         if (!cli->cl_conn_count) {
 463                 CERROR("disconnecting disconnected device (%s)\n",
 464                        obd->obd_name);
 465                 GOTO(out_sem, rc = -EINVAL);
 466         }
 467
 468         cli->cl_conn_count--;
 469         if (cli->cl_conn_count)
 470                 GOTO(out_no_disconnect, rc = 0);
 471
 472         /* Mark import deactivated now, so we don't try to reconnect if any
 473          * of the cleanup RPCs fails (e.g. ldlm cancel, etc).  We don't
 474          * fully deactivate the import, or that would drop all requests. */
 475         spin_lock(&imp->imp_lock);
 476         imp->imp_deactive = 1;
 477         spin_unlock(&imp->imp_lock);
 478
 479         /* Some non-replayable imports (MDS's OSCs) are pinged, so just
 480          * delete it regardless.  (It's safe to delete an import that was
 481          * never added.) */
 482         (void)ptlrpc_pinger_del_import(imp);
 483
 484         if (obd->obd_namespace != NULL) {
 485                 /* obd_force == local only */
 486                 ldlm_cli_cancel_unused(obd->obd_namespace, NULL,
 487                                        obd->obd_force ? LDLM_FL_LOCAL_ONLY:0,
 488                                        NULL);
 489                 ldlm_namespace_free_prior(obd->obd_namespace);
 490                 to_be_freed = obd->obd_namespace;
 491         }
 492
 493         rc = ptlrpc_disconnect_import(imp, 0);
 494
 495         ptlrpc_invalidate_import(imp);
 496         /* set obd_namespace to NULL only after invalidate, because we can have
 497          * some connect requests in flight, and his need store a connect flags
 498          * in obd_namespace. bug 14260 */
 499         obd->obd_namespace = NULL;
 500
 501         ptlrpc_free_rq_pool(imp->imp_rq_pool);
 502         destroy_import(imp);
 503         cli->cl_import = NULL;
 504
 505         EXIT;
 506  out_no_disconnect:
 507         err = class_disconnect(exp);
 508         if (!rc && err)
 509                 rc = err;
 510  out_sem:
 511         mutex_up(&cli->cl_sem);
 512         if (to_be_freed)
 513                 ldlm_namespace_free_post(to_be_freed, obd->obd_force);
 514
 515         RETURN(rc);
 516 }
 517
 518 /* --------------------------------------------------------------------------
 519  * from old lib/target.c
 520  * -------------------------------------------------------------------------- */
 521
 522 int target_handle_reconnect(struct lustre_handle *conn, struct obd_export *exp,
 523                             struct obd_uuid *cluuid, int initial_conn)
 524 {
 525         ENTRY;
 526         if (exp->exp_connection && exp->exp_imp_reverse && !initial_conn) {
 527                 struct lustre_handle *hdl;
 528                 hdl = &exp->exp_imp_reverse->imp_remote_handle;
 529                 /* Might be a re-connect after a partition. */
 530                 if (!memcmp(&conn->cookie, &hdl->cookie, sizeof conn->cookie)) {
 531                         CWARN("%s: %s reconnecting\n", exp->exp_obd->obd_name,
 532                               cluuid->uuid);
 533                         conn->cookie = exp->exp_handle.h_cookie;
 534                         /* target_handle_connect() treats EALREADY and
 535                          * -EALREADY differently.  EALREADY means we are
 536                          * doing a valid reconnect from the same client. */
 537                         RETURN(EALREADY);
 538                 } else {
 539                         CERROR("%s reconnecting from %s, "
 540                                "handle mismatch (ours "LPX64", theirs "
 541                                LPX64")\n", cluuid->uuid,
 542                                exp->exp_connection->c_remote_uuid.uuid,
 543                                hdl->cookie, conn->cookie);
 544                         memset(conn, 0, sizeof *conn);
 545                         /* target_handle_connect() treats EALREADY and
 546                          * -EALREADY differently.  -EALREADY is an error
 547                          * (same UUID, different handle). */
 548                         RETURN(-EALREADY);
 549                 }
 550         }
 551
 552         conn->cookie = exp->exp_handle.h_cookie;
 553         CDEBUG(D_HA, "connect export for UUID '%s' at %p, cookie "LPX64"\n",
 554                cluuid->uuid, exp, conn->cookie);
 555         RETURN(0);
 556 }
 557
 558 void target_client_add_cb(struct obd_device *obd, __u64 transno, void *cb_data,
 559                           int error)
 560 {
 561         struct obd_export *exp = cb_data;
 562
 563         CDEBUG(D_RPCTRACE, "%s: committing for initial connect of %s\n",
 564                obd->obd_name, exp->exp_client_uuid.uuid);
 565
 566         spin_lock(&exp->exp_lock);
 567         exp->exp_need_sync = 0;
 568         spin_unlock(&exp->exp_lock);
 569 }
 570 EXPORT_SYMBOL(target_client_add_cb);
 571
 572 int target_handle_connect(struct ptlrpc_request *req)
 573 {
 574         struct obd_device *target, *targref = NULL;
 575         struct obd_export *export = NULL;
 576         struct obd_import *revimp;
 577         struct lustre_handle conn;
 578         struct lustre_handle *tmp;
 579         struct obd_uuid tgtuuid;
 580         struct obd_uuid cluuid;
 581         struct obd_uuid remote_uuid;
 582         char *str;
 583         int rc = 0;
 584         int initial_conn = 0;
 585         struct obd_connect_data *data, *tmpdata;
 586         ENTRY;
 587
 588         OBD_RACE(OBD_FAIL_TGT_CONN_RACE);
 589
 590         str = req_capsule_client_get(&req->rq_pill, &RMF_TGTUUID);
 591         if (str == NULL) {
 592                 DEBUG_REQ(D_ERROR, req, "bad target UUID for connect");
 593                 GOTO(out, rc = -EINVAL);
 594         }
 595
 596         obd_str2uuid(&tgtuuid, str);
 597         target = class_uuid2obd(&tgtuuid);
 598         if (!target)
 599                 target = class_name2obd(str);
 600
 601         if (!target || target->obd_stopping || !target->obd_set_up) {
 602                 LCONSOLE_ERROR_MSG(0x137, "UUID '%s' is not available "
 603                                    " for connect (%s)\n", str,
 604                                    !target ? "no target" :
 605                                    (target->obd_stopping ? "stopping" :
 606                                    "not set up"));
 607                 GOTO(out, rc = -ENODEV);
 608         }
 609
 610         if (target->obd_no_conn) {
 611                 LCONSOLE_WARN("%s: temporarily refusing client connection "
 612                               "from %s\n", target->obd_name,
 613                               libcfs_nid2str(req->rq_peer.nid));
 614                 GOTO(out, rc = -EAGAIN);
 615         }
 616
 617         /* Make sure the target isn't cleaned up while we're here. Yes,
 618            there's still a race between the above check and our incref here.
 619            Really, class_uuid2obd should take the ref. */
 620         targref = class_incref(target);
 621
 622
 623         str = req_capsule_client_get(&req->rq_pill, &RMF_CLUUID);
 624         if (str == NULL) {
 625                 DEBUG_REQ(D_ERROR, req, "bad client UUID for connect");
 626                 GOTO(out, rc = -EINVAL);
 627         }
 628
 629         obd_str2uuid(&cluuid, str);
 630
 631         /* XXX extract a nettype and format accordingly */
 632         switch (sizeof(lnet_nid_t)) {
 633                 /* NB the casts only avoid compiler warnings */
 634         case 8:
 635                 snprintf(remote_uuid.uuid, sizeof remote_uuid,
 636                          "NET_"LPX64"_UUID", (__u64)req->rq_peer.nid);
 637                 break;
 638         case 4:
 639                 snprintf(remote_uuid.uuid, sizeof remote_uuid,
 640                          "NET_%x_UUID", (__u32)req->rq_peer.nid);
 641                 break;
 642         default:
 643                 LBUG();
 644         }
 645
 646         tmp = req_capsule_client_get(&req->rq_pill, &RMF_CONN);
 647         if (tmp == NULL)
 648                 GOTO(out, rc = -EPROTO);
 649
 650         conn = *tmp;
 651
 652         data = req_capsule_client_get(&req->rq_pill, &RMF_CONNECT_DATA);
 653         if (!data)
 654                 GOTO(out, rc = -EPROTO);
 655
 656         rc = req_capsule_server_pack(&req->rq_pill);
 657         if (rc)
 658                 GOTO(out, rc);
 659
 660         if (lustre_msg_get_op_flags(req->rq_reqmsg) & MSG_CONNECT_LIBCLIENT) {
 661                 if (!data) {
 662                         DEBUG_REQ(D_WARNING, req, "Refusing old (unversioned) "
 663                                   "libclient connection attempt");
 664                         GOTO(out, rc = -EPROTO);
 665                 } else if (data->ocd_version < LUSTRE_VERSION_CODE -
 666                                                LUSTRE_VERSION_ALLOWED_OFFSET ||
 667                            data->ocd_version > LUSTRE_VERSION_CODE +
 668                                                LUSTRE_VERSION_ALLOWED_OFFSET) {
 669                         DEBUG_REQ(D_WARNING, req, "Refusing %s (%d.%d.%d.%d) "
 670                                   "libclient connection attempt",
 671                                   data->ocd_version < LUSTRE_VERSION_CODE ?
 672                                   "old" : "new",
 673                                   OBD_OCD_VERSION_MAJOR(data->ocd_version),
 674                                   OBD_OCD_VERSION_MINOR(data->ocd_version),
 675                                   OBD_OCD_VERSION_PATCH(data->ocd_version),
 676                                   OBD_OCD_VERSION_FIX(data->ocd_version));
 677                         data = req_capsule_server_sized_get(&req->rq_pill,
 678                                                         &RMF_CONNECT_DATA,
 679                                     offsetof(typeof(*data), ocd_version) +
 680                                              sizeof(data->ocd_version));
 681                         if (data) {
 682                                 data->ocd_connect_flags = OBD_CONNECT_VERSION;
 683                                 data->ocd_version = LUSTRE_VERSION_CODE;
 684                         }
 685                         GOTO(out, rc = -EPROTO);
 686                 }
 687         }
 688
 689         if (lustre_msg_get_op_flags(req->rq_reqmsg) & MSG_CONNECT_INITIAL)
 690                 initial_conn = 1;
 691
 692         /* lctl gets a backstage, all-access pass. */
 693         if (obd_uuid_equals(&cluuid, &target->obd_uuid))
 694                 goto dont_check_exports;
 695
 696         spin_lock(&target->obd_dev_lock);
 697         export = lustre_hash_get_object_by_key(target->obd_uuid_hash_body, &cluuid);
 698
 699         if (export != NULL && export->exp_connecting) { /* bug 9635, et. al. */
 700                 CWARN("%s: exp %p already connecting\n",
 701                       export->exp_obd->obd_name, export);
 702                 class_export_put(export);
 703                 export = NULL;
 704                 rc = -EALREADY;
 705         } else if (export != NULL && export->exp_connection != NULL &&
 706                    req->rq_peer.nid != export->exp_connection->c_peer.nid) {
 707                 /* make darn sure this is coming from the same peer
 708                  * if the UUIDs matched */
 709                   CWARN("%s: cookie %s seen on new NID %s when "
 710                           "existing NID %s is already connected\n",
 711                         target->obd_name, cluuid.uuid,
 712                   libcfs_nid2str(req->rq_peer.nid),
 713                   libcfs_nid2str(export->exp_connection->c_peer.nid));
 714                   class_export_put(export);
 715                   export = NULL;
 716                   rc = -EALREADY;
 717         } else if (export != NULL) {
 718                 spin_lock(&export->exp_lock);
 719                 export->exp_connecting = 1;
 720                 spin_unlock(&export->exp_lock);
 721                 class_export_put(export);
 722                 spin_unlock(&target->obd_dev_lock);
 723                 LASSERT(export->exp_obd == target);
 724
 725                 rc = target_handle_reconnect(&conn, export, &cluuid, initial_conn);
 726         }
 727
 728         /* If we found an export, we already unlocked. */
 729         if (!export) {
 730                 spin_unlock(&target->obd_dev_lock);
 731                 OBD_FAIL_TIMEOUT(OBD_FAIL_TGT_DELAY_CONNECT, 2 * obd_timeout);
 732         } else if (req->rq_export == NULL &&
 733                    atomic_read(&export->exp_rpc_count) > 0) {
 734                 CWARN("%s: refuse connection from %s/%s to 0x%p/%d\n",
 735                       target->obd_name, cluuid.uuid,
 736                       libcfs_nid2str(req->rq_peer.nid),
 737                       export, atomic_read(&export->exp_refcount));
 738                 GOTO(out, rc = -EBUSY);
 739         } else if (req->rq_export != NULL &&
 740                    (atomic_read(&export->exp_rpc_count) > 1)) {
 741                 CWARN("%s: refuse reconnection from %s@%s to 0x%p/%d\n",
 742                       target->obd_name, cluuid.uuid,
 743                       libcfs_nid2str(req->rq_peer.nid),
 744                       export, atomic_read(&export->exp_rpc_count));
 745                 GOTO(out, rc = -EBUSY);
 746         } else if (lustre_msg_get_conn_cnt(req->rq_reqmsg) == 1 &&
 747                    !initial_conn) {
 748                 CERROR("%s: NID %s (%s) reconnected with 1 conn_cnt; "
 749                        "cookies not random?\n", target->obd_name,
 750                        libcfs_nid2str(req->rq_peer.nid), cluuid.uuid);
 751                 GOTO(out, rc = -EALREADY);
 752         } else {
 753                 OBD_FAIL_TIMEOUT(OBD_FAIL_TGT_DELAY_RECONNECT, 2 * obd_timeout);
 754                 if (req->rq_export == NULL && initial_conn)
 755                        export->exp_last_request_time =
 756                                max(export->exp_last_request_time,
 757                                    (time_t)CURRENT_SECONDS);
 758         }
 759
 760         /* We want to handle EALREADY but *not* -EALREADY from
 761          * target_handle_reconnect(), return reconnection state in a flag */
 762         if (rc == EALREADY) {
 763                 lustre_msg_add_op_flags(req->rq_repmsg, MSG_CONNECT_RECONNECT);
 764                 rc = 0;
 765         } else if (rc) {
 766                 GOTO(out, rc);
 767         }
 768         /* Tell the client if we're in recovery. */
 769         /* If this is the first client, start the recovery timer */
 770         CWARN("%s: connection from %s@%s %st"LPU64" exp %p cur %ld last %ld\n",
 771                target->obd_name, cluuid.uuid, libcfs_nid2str(req->rq_peer.nid),
 772               target->obd_recovering ? "recovering/" : "", data->ocd_transno,
 773               export, (long)CURRENT_SECONDS,
 774               export ? (long)export->exp_last_request_time : 0);
 775
 776
 777         if (target->obd_recovering) {
 778                 lustre_msg_add_op_flags(req->rq_repmsg, MSG_CONNECT_RECOVERING);
 779                 target_start_recovery_timer(target);
 780         }
 781
 782         /* Tell the client if we support replayable requests */
 783         if (target->obd_replayable)
 784                 lustre_msg_add_op_flags(req->rq_repmsg, MSG_CONNECT_REPLAYABLE);
 785
 786         if (export == NULL) {
 787                 if (target->obd_recovering) {
 788                         cfs_time_t t;
 789
 790                         t = cfs_timer_deadline(&target->obd_recovery_timer);
 791                         t = cfs_time_sub(t, cfs_time_current());
 792                         CERROR("%s: denying connection for new client %s (%s): "
 793                                "%d clients in recovery for %lds\n",
 794                                target->obd_name,
 795                                libcfs_nid2str(req->rq_peer.nid), cluuid.uuid,
 796                                target->obd_recoverable_clients,
 797                                cfs_duration_sec(t));
 798                         rc = -EBUSY;
 799                 } else {
 800 dont_check_exports:
 801                         rc = obd_connect(req->rq_svc_thread->t_env,
 802                                          &conn, target, &cluuid, data);
 803                 }
 804         } else {
 805                 rc = obd_reconnect(req->rq_svc_thread->t_env,
 806                                    export, target, &cluuid, data);
 807         }
 808         if (rc)
 809                 GOTO(out, rc);
 810         /* Return only the parts of obd_connect_data that we understand, so the
 811          * client knows that we don't understand the rest. */
 812         if (data) {
 813                  tmpdata = req_capsule_server_get(&req->rq_pill,
 814                                                   &RMF_CONNECT_DATA);
 815                   //data->ocd_connect_flags &= OBD_CONNECT_SUPPORTED;
 816                  *tmpdata = *data;
 817         }
 818
 819         /* If all else goes well, this is our RPC return code. */
 820         req->rq_status = 0;
 821
 822         lustre_msg_set_handle(req->rq_repmsg, &conn);
 823
 824         /* ownership of this export ref transfers to the request AFTER we
 825          * drop any previous reference the request had, but we don't want
 826          * that to go to zero before we get our new export reference. */
 827         export = class_conn2export(&conn);
 828         if (!export) {
 829                 DEBUG_REQ(D_ERROR, req, "Missing export!");
 830                 GOTO(out, rc = -ENODEV);
 831         }
 832
 833         /* If the client and the server are the same node, we will already
 834          * have an export that really points to the client's DLM export,
 835          * because we have a shared handles table.
 836          *
 837          * XXX this will go away when shaver stops sending the "connect" handle
 838          * in the real "remote handle" field of the request --phik 24 Apr 2003
 839          */
 840         if (req->rq_export != NULL)
 841                 class_export_put(req->rq_export);
 842
 843         req->rq_export = export;
 844
 845         spin_lock(&export->exp_lock);
 846         if (initial_conn) {
 847                 lustre_msg_set_conn_cnt(req->rq_repmsg, export->exp_conn_cnt + 1);
 848         } else if (export->exp_conn_cnt >= lustre_msg_get_conn_cnt(req->rq_reqmsg)) {
 849                 spin_unlock(&export->exp_lock);
 850                 CERROR("%s: %s already connected at higher conn_cnt: %d > %d\n",
 851                        cluuid.uuid, libcfs_nid2str(req->rq_peer.nid),
 852                        export->exp_conn_cnt,
 853                        lustre_msg_get_conn_cnt(req->rq_reqmsg));
 854
 855                 GOTO(out, rc = -EALREADY);
 856         }
 857         export->exp_conn_cnt = lustre_msg_get_conn_cnt(req->rq_reqmsg);
 858
 859         /* request from liblustre?  Don't evict it for not pinging. */
 860         if (lustre_msg_get_op_flags(req->rq_reqmsg) & MSG_CONNECT_LIBCLIENT) {
 861                 export->exp_libclient = 1;
 862                 spin_unlock(&export->exp_lock);
 863
 864                 spin_lock(&target->obd_dev_lock);
 865                 list_del_init(&export->exp_obd_chain_timed);
 866                 spin_unlock(&target->obd_dev_lock);
 867         } else {
 868                 spin_unlock(&export->exp_lock);
 869         }
 870
 871         if (export->exp_connection != NULL)
 872                 ptlrpc_put_connection(export->exp_connection);
 873         export->exp_connection = ptlrpc_get_connection(req->rq_peer,
 874                                                        req->rq_self,
 875                                                        &remote_uuid);
 876
 877         spin_lock(&target->obd_dev_lock);
 878         /* Export might be hashed already, e.g. if this is reconnect */
 879         if (hlist_unhashed(&export->exp_nid_hash))
 880                 lustre_hash_additem(export->exp_obd->obd_nid_hash_body,
 881                                     &export->exp_connection->c_peer.nid,
 882                                     &export->exp_nid_hash);
 883         spin_unlock(&target->obd_dev_lock);
 884
 885         spin_lock_bh(&target->obd_processing_task_lock);
 886         if (target->obd_recovering && !export->exp_in_recovery) {
 887                 spin_lock(&export->exp_lock);
 888                 export->exp_in_recovery = 1;
 889                 export->exp_req_replay_needed = 1;
 890                 export->exp_lock_replay_needed = 1;
 891                 spin_unlock(&export->exp_lock);
 892                 if ((lustre_msg_get_op_flags(req->rq_reqmsg) & MSG_CONNECT_TRANSNO)
 893                      && (data->ocd_transno == 0))
 894                         CWARN("Connect with zero transno!\n");
 895
 896                 if ((lustre_msg_get_op_flags(req->rq_reqmsg) & MSG_CONNECT_TRANSNO)
 897                      && data->ocd_transno < target->obd_next_recovery_transno)
 898                         target->obd_next_recovery_transno = data->ocd_transno;
 899                 target->obd_connected_clients++;
 900                 /* each connected client is counted as recoverable */
 901                 target->obd_recoverable_clients++;
 902                 atomic_inc(&target->obd_req_replay_clients);
 903                 atomic_inc(&target->obd_lock_replay_clients);
 904                 if (target->obd_connected_clients ==
 905                     target->obd_max_recoverable_clients)
 906                         wake_up(&target->obd_next_transno_waitq);
 907         }
 908         spin_unlock_bh(&target->obd_processing_task_lock);
 909         tmp = req_capsule_client_get(&req->rq_pill, &RMF_CONN);
 910         conn = *tmp;
 911
 912         if (export->exp_imp_reverse != NULL) {
 913                 /* destroyed import can be still referenced in ctxt */
 914                 obd_set_info_async(export, strlen(KEY_REVIMP_UPD),
 915                                    KEY_REVIMP_UPD, 0, NULL, NULL);
 916
 917                 /* in some recovery senarios, previous ctx init rpc handled
 918                  * in sptlrpc_target_export_check() might be used to install
 919                  * a reverse ctx in this reverse import, and later OBD_CONNECT
 920                  * using the same gss ctx could reach here and following new
 921                  * reverse import. note all reverse ctx in new/old import are
 922                  * actually based on the same gss ctx. so we invalidate ctx
 923                  * here before destroy import, otherwise flush old import will
 924                  * lead to remote reverse ctx be destroied, thus the reverse
 925                  * ctx of new import will lost its peer.
 926                  * there might be a better way to deal with this???
 927                  */
 928                 sptlrpc_import_inval_all_ctx(export->exp_imp_reverse);
 929
 930                 destroy_import(export->exp_imp_reverse);
 931         }
 932
 933         /* for the rest part, we return -ENOTCONN in case of errors
 934          * in order to let client initialize connection again.
 935          */
 936         revimp = export->exp_imp_reverse = class_new_import(target);
 937         if (!revimp) {
 938                 CERROR("fail to alloc new reverse import.\n");
 939                 GOTO(out, rc = -ENOTCONN);
 940         }
 941
 942         revimp->imp_connection = ptlrpc_connection_addref(export->exp_connection);
 943         revimp->imp_client = &export->exp_obd->obd_ldlm_client;
 944         revimp->imp_remote_handle = conn;
 945         revimp->imp_dlm_fake = 1;
 946         revimp->imp_state = LUSTRE_IMP_FULL;
 947
 948         if (lustre_msg_get_op_flags(req->rq_reqmsg) & MSG_CONNECT_NEXT_VER) {
 949                 revimp->imp_msg_magic = LUSTRE_MSG_MAGIC_V2;
 950                 lustre_msg_add_op_flags(req->rq_repmsg, MSG_CONNECT_NEXT_VER);
 951         }
 952
 953         rc = sptlrpc_import_sec_adapt(revimp, req->rq_svc_ctx,
 954                                       req->rq_flvr.sf_rpc);
 955         if (rc) {
 956                 CERROR("Failed to get sec for reverse import: %d\n", rc);
 957                 export->exp_imp_reverse = NULL;
 958                 class_destroy_import(revimp);
 959         }
 960
 961         class_import_put(revimp);
 962 out:
 963         if (export) {
 964                 spin_lock(&export->exp_lock);
 965                 export->exp_connecting = 0;
 966                 spin_unlock(&export->exp_lock);
 967         }
 968         if (targref)
 969                 class_decref(targref);
 970         if (rc)
 971                 req->rq_status = rc;
 972         RETURN(rc);
 973 }
 974
 975 int target_handle_disconnect(struct ptlrpc_request *req)
 976 {
 977         int rc;
 978         ENTRY;
 979
 980         rc = req_capsule_server_pack(&req->rq_pill);
 981         if (rc)
 982                 RETURN(rc);
 983
 984         /* keep the rq_export around so we can send the reply */
 985         req->rq_status = obd_disconnect(class_export_get(req->rq_export));
 986
 987         RETURN(0);
 988 }
 989
 990 void target_destroy_export(struct obd_export *exp)
 991 {
 992         /* exports created from last_rcvd data, and "fake"
 993            exports created by lctl don't have an import */
 994         if (exp->exp_imp_reverse != NULL)
 995                 destroy_import(exp->exp_imp_reverse);
 996
 997         /* We cancel locks at disconnect time, but this will catch any locks
 998          * granted in a race with recovery-induced disconnect. */
 999         if (exp->exp_obd->obd_namespace != NULL)
1000                 ldlm_cancel_locks_for_export(exp);
1001 }
1002
1003 /*
1004  * Recovery functions
1005  */
1006
1007 struct ptlrpc_request *ptlrpc_clone_req( struct ptlrpc_request *orig_req)
1008 {
1009         struct ptlrpc_request *copy_req;
1010         struct lustre_msg *copy_reqmsg;
1011         struct ptlrpc_user_desc *udesc = NULL;
1012
1013         OBD_ALLOC_PTR(copy_req);
1014         if (!copy_req)
1015                 return NULL;
1016         OBD_ALLOC(copy_reqmsg, orig_req->rq_reqlen);
1017         if (!copy_reqmsg){
1018                 OBD_FREE_PTR(copy_req);
1019                 return NULL;
1020         }
1021
1022         if (orig_req->rq_user_desc) {
1023                 int ngroups = orig_req->rq_user_desc->pud_ngroups;
1024
1025                 OBD_ALLOC(udesc, sptlrpc_user_desc_size(ngroups));
1026                 if (!udesc) {
1027                         OBD_FREE(copy_reqmsg, orig_req->rq_reqlen);
1028                         OBD_FREE_PTR(copy_req);
1029                         return NULL;
1030                 }
1031                 memcpy(udesc, orig_req->rq_user_desc,
1032                        sptlrpc_user_desc_size(ngroups));
1033         }
1034
1035         *copy_req = *orig_req;
1036         memcpy(copy_reqmsg, orig_req->rq_reqmsg, orig_req->rq_reqlen);
1037         copy_req->rq_reqmsg = copy_reqmsg;
1038         copy_req->rq_user_desc = udesc;
1039
1040         class_export_get(copy_req->rq_export);
1041         CFS_INIT_LIST_HEAD(&copy_req->rq_list);
1042         sptlrpc_svc_ctx_addref(copy_req);
1043
1044         if (copy_req->rq_reply_state) {
1045                 /* the copied req takes over the reply state */
1046                 orig_req->rq_reply_state = NULL;
1047                 /* to catch further access */
1048                 orig_req->rq_repmsg = NULL;
1049                 orig_req->rq_replen = 0;
1050         }
1051
1052         return copy_req;
1053 }
1054
1055 void ptlrpc_free_clone( struct ptlrpc_request *req)
1056 {
1057         if (req->rq_reply_state) {
1058                 ptlrpc_rs_decref(req->rq_reply_state);
1059                 req->rq_reply_state = NULL;
1060         }
1061
1062         sptlrpc_svc_ctx_decref(req);
1063         class_export_put(req->rq_export);
1064         list_del(&req->rq_list);
1065
1066         if (req->rq_user_desc) {
1067                 int ngroups = req->rq_user_desc->pud_ngroups;
1068                 OBD_FREE(req->rq_user_desc, sptlrpc_user_desc_size(ngroups));
1069         }
1070         OBD_FREE(req->rq_reqmsg, req->rq_reqlen);
1071         OBD_FREE_PTR(req);
1072 }
1073
1074 #ifdef __KERNEL__
1075 static void target_finish_recovery(struct obd_device *obd)
1076 {
1077         ENTRY;
1078         LCONSOLE_INFO("%s: sending delayed replies to recovered clients\n",
1079                       obd->obd_name);
1080
1081         ldlm_reprocess_all_ns(obd->obd_namespace);
1082
1083         /* when recovery finished, cleanup orphans on mds and ost */
1084         if (OBT(obd) && OBP(obd, postrecov)) {
1085                 int rc = OBP(obd, postrecov)(obd);
1086                 LCONSOLE_WARN("%s: recovery %s: rc %d\n", obd->obd_name,
1087                               rc < 0 ? "failed" : "complete", rc);
1088         }
1089
1090         obd->obd_recovery_end = CURRENT_SECONDS;
1091         EXIT;
1092 }
1093
1094 static void abort_req_replay_queue(struct obd_device *obd)
1095 {
1096         struct ptlrpc_request *req, *n;
1097
1098         list_for_each_entry_safe(req, n, &obd->obd_req_replay_queue, rq_list) {
1099                 DEBUG_REQ(D_WARNING, req, "aborted:");
1100                 req->rq_status = -ENOTCONN;
1101                 if (ptlrpc_error(req)) {
1102                         DEBUG_REQ(D_ERROR, req,
1103                                   "failed abort_req_reply; skipping");
1104                 }
1105                 ptlrpc_free_clone(req);
1106         }
1107 }
1108
1109 static void abort_lock_replay_queue(struct obd_device *obd)
1110 {
1111         struct ptlrpc_request *req, *n;
1112
1113         list_for_each_entry_safe(req, n, &obd->obd_lock_replay_queue, rq_list){
1114                 DEBUG_REQ(D_ERROR, req, "aborted:");
1115                 req->rq_status = -ENOTCONN;
1116                 if (ptlrpc_error(req)) {
1117                         DEBUG_REQ(D_ERROR, req,
1118                                   "failed abort_lock_reply; skipping");
1119                 }
1120                 ptlrpc_free_clone(req);
1121         }
1122 }
1123 #endif
1124
1125 /* Called from a cleanup function if the device is being cleaned up
1126    forcefully.  The exports should all have been disconnected already,
1127    the only thing left to do is
1128      - clear the recovery flags
1129      - cancel the timer
1130      - free queued requests and replies, but don't send replies
1131    Because the obd_stopping flag is set, no new requests should be received.
1132
1133 */
1134 void target_cleanup_recovery(struct obd_device *obd)
1135 {
1136         struct ptlrpc_request *req, *n;
1137         ENTRY;
1138
1139         LASSERT(obd->obd_stopping);
1140
1141         spin_lock_bh(&obd->obd_processing_task_lock);
1142         if (!obd->obd_recovering) {
1143                 spin_unlock_bh(&obd->obd_processing_task_lock);
1144                 EXIT;
1145                 return;
1146         }
1147         obd->obd_recovering = obd->obd_abort_recovery = 0;
1148         target_cancel_recovery_timer(obd);
1149         spin_unlock_bh(&obd->obd_processing_task_lock);
1150
1151         list_for_each_entry_safe(req, n, &obd->obd_req_replay_queue, rq_list) {
1152                 LASSERT (req->rq_reply_state == 0);
1153                 ptlrpc_free_clone(req);
1154         }
1155         list_for_each_entry_safe(req, n, &obd->obd_lock_replay_queue, rq_list){
1156                 LASSERT (req->rq_reply_state == 0);
1157                 ptlrpc_free_clone(req);
1158         }
1159         list_for_each_entry_safe(req, n, &obd->obd_final_req_queue, rq_list) {
1160                 LASSERT (req->rq_reply_state == 0);
1161                 ptlrpc_free_clone(req);
1162         }
1163
1164         EXIT;
1165 }
1166
1167 static void target_recovery_expired(unsigned long castmeharder)
1168 {
1169         struct obd_device *obd = (struct obd_device *)castmeharder;
1170         CERROR("%s: recovery timed out, aborting\n", obd->obd_name);
1171         spin_lock_bh(&obd->obd_processing_task_lock);
1172         if (obd->obd_recovering)
1173                 obd->obd_abort_recovery = 1;
1174         cfs_waitq_signal(&obd->obd_next_transno_waitq);
1175         spin_unlock_bh(&obd->obd_processing_task_lock);
1176 }
1177
1178
1179 /* obd_processing_task_lock should be held */
1180 void target_cancel_recovery_timer(struct obd_device *obd)
1181 {
1182         CDEBUG(D_HA, "%s: cancel recovery timer\n", obd->obd_name);
1183         cfs_timer_disarm(&obd->obd_recovery_timer);
1184 }
1185
1186 static void reset_recovery_timer(struct obd_device *obd)
1187 {
1188         time_t timeout_shift = OBD_RECOVERY_TIMEOUT;
1189         spin_lock_bh(&obd->obd_processing_task_lock);
1190         if (!obd->obd_recovering) {
1191                 spin_unlock_bh(&obd->obd_processing_task_lock);
1192                 return;
1193         }
1194         if (cfs_time_current_sec() + OBD_RECOVERY_TIMEOUT >
1195             obd->obd_recovery_start + obd->obd_recovery_max_time)
1196                 timeout_shift = obd->obd_recovery_start +
1197                         obd->obd_recovery_max_time - cfs_time_current_sec();
1198         cfs_timer_arm(&obd->obd_recovery_timer, cfs_time_shift(timeout_shift));
1199         spin_unlock_bh(&obd->obd_processing_task_lock);
1200         CDEBUG(D_HA, "%s: timer will expire in %u seconds\n", obd->obd_name,
1201                (unsigned int)timeout_shift);
1202         /* Only used for lprocfs_status */
1203         obd->obd_recovery_end = CURRENT_SECONDS + timeout_shift;
1204 }
1205
1206
1207 /* Only start it the first time called */
1208 void target_start_recovery_timer(struct obd_device *obd)
1209 {
1210         spin_lock_bh(&obd->obd_processing_task_lock);
1211         if (obd->obd_recovery_handler
1212             || timer_pending((struct timer_list *)&obd->obd_recovery_timer)) {
1213                 spin_unlock_bh(&obd->obd_processing_task_lock);
1214                 return;
1215         }
1216         CWARN("%s: starting recovery timer (%us)\n", obd->obd_name,
1217               OBD_RECOVERY_TIMEOUT);
1218         cfs_timer_init(&obd->obd_recovery_timer, target_recovery_expired, obd);
1219         spin_unlock_bh(&obd->obd_processing_task_lock);
1220
1221         reset_recovery_timer(obd);
1222 }
1223
1224 #ifdef __KERNEL__
1225 static int check_for_next_transno(struct obd_device *obd)
1226 {
1227         struct ptlrpc_request *req = NULL;
1228         int wake_up = 0, connected, completed, queue_len, max;
1229         __u64 next_transno, req_transno;
1230         ENTRY;
1231         spin_lock_bh(&obd->obd_processing_task_lock);
1232
1233         if (!list_empty(&obd->obd_req_replay_queue)) {
1234                 req = list_entry(obd->obd_req_replay_queue.next,
1235                                  struct ptlrpc_request, rq_list);
1236                 req_transno = lustre_msg_get_transno(req->rq_reqmsg);
1237         } else {
1238                 req_transno = 0;
1239         }
1240
1241         max = obd->obd_max_recoverable_clients;
1242         connected = obd->obd_connected_clients;
1243         completed = connected - obd->obd_recoverable_clients;
1244         queue_len = obd->obd_requests_queued_for_recovery;
1245         next_transno = obd->obd_next_recovery_transno;
1246
1247         CDEBUG(D_HA, "max: %d, connected: %d, completed: %d, queue_len: %d, "
1248                "req_transno: "LPU64", next_transno: "LPU64"\n",
1249                max, connected, completed, queue_len, req_transno, next_transno);
1250
1251         if (obd->obd_abort_recovery) {
1252                 CDEBUG(D_HA, "waking for aborted recovery\n");
1253                 wake_up = 1;
1254         } else if (atomic_read(&obd->obd_req_replay_clients) == 0) {
1255                 CDEBUG(D_HA, "waking for completed recovery\n");
1256                 wake_up = 1;
1257         } else if (req_transno == next_transno) {
1258                 CDEBUG(D_HA, "waking for next ("LPD64")\n", next_transno);
1259                 wake_up = 1;
1260         } else if (queue_len + completed == max) {
1261                 /* handle gaps occured due to lost reply. It is allowed gaps
1262                  * because all clients are connected and there will be resend
1263                  * for missed transaction */
1264                 LASSERTF(req_transno >= next_transno,
1265                          "req_transno: "LPU64", next_transno: "LPU64"\n",
1266                          req_transno, next_transno);
1267
1268                 CDEBUG(req_transno > obd->obd_last_committed ? D_ERROR : D_HA,
1269                        "waking for skipped transno (skip: "LPD64
1270                        ", ql: %d, comp: %d, conn: %d, next: "LPD64")\n",
1271                        next_transno, queue_len, completed, connected, req_transno);
1272                 obd->obd_next_recovery_transno = req_transno;
1273                 wake_up = 1;
1274         } else if (queue_len == atomic_read(&obd->obd_req_replay_clients)) {
1275                 /* some clients haven't connected in time, but we can try
1276                  * to replay requests that demand on already committed ones
1277                  * also, we can replay first non-committed transation */
1278                 LASSERT(req_transno != 0);
1279                 if (req_transno == obd->obd_last_committed + 1) {
1280                         obd->obd_next_recovery_transno = req_transno;
1281                 } else if (req_transno > obd->obd_last_committed) {
1282                         /* can't continue recovery: have no needed transno */
1283                         obd->obd_abort_recovery = 1;
1284                         CDEBUG(D_ERROR, "abort due to missed clients. max: %d, "
1285                                "connected: %d, completed: %d, queue_len: %d, "
1286                                "req_transno: "LPU64", next_transno: "LPU64"\n",
1287                                max, connected, completed, queue_len,
1288                                req_transno, next_transno);
1289                 }
1290                 wake_up = 1;
1291         }
1292
1293         spin_unlock_bh(&obd->obd_processing_task_lock);
1294         return wake_up;
1295 }
1296
1297 static struct ptlrpc_request *target_next_replay_req(struct obd_device *obd)
1298 {
1299         struct l_wait_info lwi = { 0 };
1300         struct ptlrpc_request *req;
1301
1302         CDEBUG(D_HA, "Waiting for transno "LPD64"\n",
1303                obd->obd_next_recovery_transno);
1304         l_wait_event(obd->obd_next_transno_waitq,
1305                      check_for_next_transno(obd), &lwi);
1306
1307         spin_lock_bh(&obd->obd_processing_task_lock);
1308         if (obd->obd_abort_recovery) {
1309                 req = NULL;
1310         } else if (!list_empty(&obd->obd_req_replay_queue)) {
1311                 req = list_entry(obd->obd_req_replay_queue.next,
1312                                  struct ptlrpc_request, rq_list);
1313                 list_del_init(&req->rq_list);
1314                 obd->obd_requests_queued_for_recovery--;
1315         } else {
1316                 req = NULL;
1317         }
1318         spin_unlock_bh(&obd->obd_processing_task_lock);
1319         RETURN(req);
1320 }
1321
1322 static int check_for_next_lock(struct obd_device *obd)
1323 {
1324         struct ptlrpc_request *req = NULL;
1325         int wake_up = 0;
1326
1327         spin_lock_bh(&obd->obd_processing_task_lock);
1328         if (!list_empty(&obd->obd_lock_replay_queue)) {
1329                 req = list_entry(obd->obd_lock_replay_queue.next,
1330                                  struct ptlrpc_request, rq_list);
1331                 CDEBUG(D_HA, "waking for next lock\n");
1332                 wake_up = 1;
1333         } else if (atomic_read(&obd->obd_lock_replay_clients) == 0) {
1334                 CDEBUG(D_HA, "waking for completed lock replay\n");
1335                 wake_up = 1;
1336         } else if (obd->obd_abort_recovery) {
1337                 CDEBUG(D_HA, "waking for aborted recovery\n");
1338                 wake_up = 1;
1339         }
1340         spin_unlock_bh(&obd->obd_processing_task_lock);
1341
1342         return wake_up;
1343 }
1344
1345 static struct ptlrpc_request *target_next_replay_lock(struct obd_device *obd)
1346 {
1347         struct l_wait_info lwi = { 0 };
1348         struct ptlrpc_request *req;
1349
1350         CDEBUG(D_HA, "Waiting for lock\n");
1351         l_wait_event(obd->obd_next_transno_waitq,
1352                      check_for_next_lock(obd), &lwi);
1353
1354         spin_lock_bh(&obd->obd_processing_task_lock);
1355         if (obd->obd_abort_recovery) {
1356                 req = NULL;
1357         } else if (!list_empty(&obd->obd_lock_replay_queue)) {
1358                 req = list_entry(obd->obd_lock_replay_queue.next,
1359                                  struct ptlrpc_request, rq_list);
1360                 list_del_init(&req->rq_list);
1361         } else {
1362                 req = NULL;
1363         }
1364         spin_unlock_bh(&obd->obd_processing_task_lock);
1365         return req;
1366 }
1367
1368 static struct ptlrpc_request *target_next_final_ping(struct obd_device *obd)
1369 {
1370         struct ptlrpc_request *req;
1371
1372         spin_lock_bh(&obd->obd_processing_task_lock);
1373         if (!list_empty(&obd->obd_final_req_queue)) {
1374                 req = list_entry(obd->obd_final_req_queue.next,
1375                                  struct ptlrpc_request, rq_list);
1376                 list_del_init(&req->rq_list);
1377         } else {
1378                 req = NULL;
1379         }
1380         spin_unlock_bh(&obd->obd_processing_task_lock);
1381         return req;
1382 }
1383
1384 static inline int req_replay_done(struct obd_export *exp)
1385 {
1386         return (exp->exp_req_replay_needed == 0);
1387 }
1388
1389 static inline int lock_replay_done(struct obd_export *exp)
1390 {
1391         return (exp->exp_lock_replay_needed == 0);
1392 }
1393
1394 static inline int connect_done(struct obd_export *exp)
1395 {
1396         return (exp->exp_in_recovery != 0);
1397 }
1398
1399 static int check_for_clients(struct obd_device *obd)
1400 {
1401         if (obd->obd_abort_recovery)
1402                 return 1;
1403         LASSERT(obd->obd_connected_clients <= obd->obd_max_recoverable_clients);
1404         if (obd->obd_no_conn == 0 &&
1405             obd->obd_connected_clients == obd->obd_max_recoverable_clients)
1406                 return 1;
1407         return 0;
1408 }
1409
1410 static int handle_recovery_req(struct ptlrpc_thread *thread,
1411                                struct ptlrpc_request *req,
1412                                svc_handler_t handler)
1413 {
1414         int rc;
1415         ENTRY;
1416
1417         rc = lu_context_init(&req->rq_session, LCT_SESSION);
1418         if (rc) {
1419                 CERROR("Failure to initialize session: %d\n", rc);
1420                 return rc;
1421         }
1422         req->rq_session.lc_thread = thread;
1423         lu_context_enter(&req->rq_session);
1424         req->rq_svc_thread = thread;
1425         req->rq_svc_thread->t_env->le_ses = &req->rq_session;
1426
1427         /* thread context */
1428         lu_context_enter(&thread->t_env->le_ctx);
1429         (void)handler(req);
1430         lu_context_exit(&thread->t_env->le_ctx);
1431
1432         lu_context_exit(&req->rq_session);
1433         lu_context_fini(&req->rq_session);
1434         /* don't reset timer for final stage */
1435         if (!req_replay_done(req->rq_export) ||
1436             !lock_replay_done(req->rq_export))
1437                 reset_recovery_timer(class_exp2obd(req->rq_export));
1438         ptlrpc_free_clone(req);
1439         RETURN(0);
1440 }
1441
1442 static int target_recovery_thread(void *arg)
1443 {
1444         struct obd_device *obd = arg;
1445         struct ptlrpc_request *req;
1446         struct target_recovery_data *trd = &obd->obd_recovery_data;
1447         struct l_wait_info lwi = { 0 };
1448         unsigned long delta;
1449         unsigned long flags;
1450         struct lu_env env;
1451         struct ptlrpc_thread fake_svc_thread, *thread = &fake_svc_thread;
1452         __u32 recov_ctx_tags = LCT_MD_THREAD;
1453         int rc = 0;
1454         ENTRY;
1455
1456         cfs_daemonize("tgt_recov");
1457
1458         SIGNAL_MASK_LOCK(current, flags);
1459         sigfillset(&current->blocked);
1460         RECALC_SIGPENDING;
1461         SIGNAL_MASK_UNLOCK(current, flags);
1462
1463         rc = lu_context_init(&env.le_ctx, recov_ctx_tags);
1464         if (rc)
1465                 RETURN(rc);
1466
1467         thread->t_env = &env;
1468         env.le_ctx.lc_thread = thread;
1469
1470         CERROR("%s: started recovery thread pid %d\n", obd->obd_name,
1471                current->pid);
1472         trd->trd_processing_task = current->pid;
1473
1474         obd->obd_recovering = 1;
1475         complete(&trd->trd_starting);
1476
1477         /* first of all, we have to know the first transno to replay */
1478         obd->obd_abort_recovery = 0;
1479         l_wait_event(obd->obd_next_transno_waitq,
1480                      check_for_clients(obd), &lwi);
1481
1482         spin_lock_bh(&obd->obd_processing_task_lock);
1483         target_cancel_recovery_timer(obd);
1484         spin_unlock_bh(&obd->obd_processing_task_lock);
1485
1486         /* If some clients haven't connected in time, evict them */
1487         if (obd->obd_abort_recovery) {
1488                 CWARN("Some clients haven't connect in time (%d/%d),"
1489                        "evict them\n", obd->obd_connected_clients,
1490                        obd->obd_max_recoverable_clients);
1491                 obd->obd_abort_recovery = obd->obd_stopping;
1492                 class_disconnect_stale_exports(obd, connect_done);
1493         }
1494         /* next stage: replay requests */
1495         delta = jiffies;
1496         obd->obd_req_replaying = 1;
1497         CDEBUG(D_INFO, "1: request replay stage - %d clients from t"LPU64"\n",
1498               atomic_read(&obd->obd_req_replay_clients),
1499               obd->obd_next_recovery_transno);
1500         while ((req = target_next_replay_req(obd))) {
1501                 LASSERT(trd->trd_processing_task == current->pid);
1502                 DEBUG_REQ(D_HA, req, "processing t"LPD64" from %s",
1503                           lustre_msg_get_transno(req->rq_reqmsg),
1504                           libcfs_nid2str(req->rq_peer.nid));
1505                 handle_recovery_req(thread, req,
1506                                     trd->trd_recovery_handler);
1507                 obd->obd_replayed_requests++;
1508                 spin_lock_bh(&obd->obd_processing_task_lock);
1509                 obd->obd_next_recovery_transno++;
1510                 spin_unlock_bh(&obd->obd_processing_task_lock);
1511         }
1512
1513         spin_lock_bh(&obd->obd_processing_task_lock);
1514         target_cancel_recovery_timer(obd);
1515         spin_unlock_bh(&obd->obd_processing_task_lock);
1516
1517         /* If some clients haven't replayed requests in time, evict them */
1518         if (obd->obd_abort_recovery) {
1519                 CDEBUG(D_ERROR, "req replay timed out, aborting ...\n");
1520                 obd->obd_abort_recovery = obd->obd_stopping;
1521                 class_disconnect_stale_exports(obd, req_replay_done);
1522                 abort_req_replay_queue(obd);
1523         }
1524         /* The second stage: replay locks */
1525         CDEBUG(D_INFO, "2: lock replay stage - %d clients\n",
1526                atomic_read(&obd->obd_lock_replay_clients));
1527         while ((req = target_next_replay_lock(obd))) {
1528                 LASSERT(trd->trd_processing_task == current->pid);
1529                 DEBUG_REQ(D_HA|D_WARNING, req, "processing lock from %s: ",
1530                           libcfs_nid2str(req->rq_peer.nid));
1531                 handle_recovery_req(thread, req,
1532                                     trd->trd_recovery_handler);
1533                 obd->obd_replayed_locks++;
1534         }
1535
1536         spin_lock_bh(&obd->obd_processing_task_lock);
1537         target_cancel_recovery_timer(obd);
1538         spin_unlock_bh(&obd->obd_processing_task_lock);
1539         /* If some clients haven't replayed requests in time, evict them */
1540         if (obd->obd_abort_recovery) {
1541                 int stale;
1542                 CERROR("lock replay timed out, aborting ...\n");
1543                 obd->obd_abort_recovery = obd->obd_stopping;
1544                 stale = class_disconnect_stale_exports(obd, lock_replay_done);
1545                 abort_lock_replay_queue(obd);
1546         }
1547
1548         /* We drop recoverying flag to forward all new requests
1549          * to regular mds_handle() since now */
1550         spin_lock_bh(&obd->obd_processing_task_lock);
1551         obd->obd_recovering = obd->obd_abort_recovery = 0;
1552         spin_unlock_bh(&obd->obd_processing_task_lock);
1553         /* The third stage: reply on final pings */
1554         CDEBUG(D_INFO, "3: final stage - process recovery completion pings\n");
1555         while ((req = target_next_final_ping(obd))) {
1556                 LASSERT(trd->trd_processing_task == current->pid);
1557                 DEBUG_REQ(D_HA, req, "processing final ping from %s: ",
1558                           libcfs_nid2str(req->rq_peer.nid));
1559                 handle_recovery_req(thread, req,
1560                                     trd->trd_recovery_handler);
1561         }
1562
1563         delta = (jiffies - delta) / HZ;
1564         CDEBUG(D_INFO,"4: recovery completed in %lus - %d/%d reqs/locks\n",
1565               delta, obd->obd_replayed_requests, obd->obd_replayed_locks);
1566         LASSERT(atomic_read(&obd->obd_req_replay_clients) == 0);
1567         LASSERT(atomic_read(&obd->obd_lock_replay_clients) == 0);
1568         if (delta > obd_timeout * 2) {
1569                 CWARN("too long recovery - read logs\n");
1570                 libcfs_debug_dumplog();
1571         }
1572
1573         target_finish_recovery(obd);
1574
1575         lu_context_fini(&env.le_ctx);
1576         trd->trd_processing_task = 0;
1577         complete(&trd->trd_finishing);
1578         RETURN(rc);
1579 }
1580
1581 int target_start_recovery_thread(struct obd_device *obd, svc_handler_t handler)
1582 {
1583         int rc = 0;
1584         struct target_recovery_data *trd = &obd->obd_recovery_data;
1585
1586         memset(trd, 0, sizeof(*trd));
1587         init_completion(&trd->trd_starting);
1588         init_completion(&trd->trd_finishing);
1589         trd->trd_recovery_handler = handler;
1590
1591         if (kernel_thread(target_recovery_thread, obd, 0) > 0) {
1592                 wait_for_completion(&trd->trd_starting);
1593                 LASSERT(obd->obd_recovering != 0);
1594         } else
1595                 rc = -ECHILD;
1596
1597         return rc;
1598 }
1599
1600 void target_stop_recovery_thread(struct obd_device *obd)
1601 {
1602         spin_lock_bh(&obd->obd_processing_task_lock);
1603         if (obd->obd_recovery_data.trd_processing_task > 0) {
1604                 struct target_recovery_data *trd = &obd->obd_recovery_data;
1605                 CERROR("%s: Aborting recovery\n", obd->obd_name);
1606                 obd->obd_abort_recovery = 1;
1607                 wake_up(&obd->obd_next_transno_waitq);
1608                 spin_unlock_bh(&obd->obd_processing_task_lock);
1609                 wait_for_completion(&trd->trd_finishing);
1610         } else {
1611                 spin_unlock_bh(&obd->obd_processing_task_lock);
1612         }
1613 }
1614
1615 void target_recovery_fini(struct obd_device *obd)
1616 {
1617         class_disconnect_exports(obd);
1618         target_stop_recovery_thread(obd);
1619         target_cleanup_recovery(obd);
1620 }
1621 EXPORT_SYMBOL(target_recovery_fini);
1622
1623 void target_recovery_init(struct obd_device *obd, svc_handler_t handler)
1624 {
1625         if (obd->obd_max_recoverable_clients == 0)
1626                 return;
1627
1628         CWARN("RECOVERY: service %s, %d recoverable clients, "
1629               "last_transno "LPU64"\n", obd->obd_name,
1630               obd->obd_max_recoverable_clients, obd->obd_last_committed);
1631         obd->obd_next_recovery_transno = obd->obd_last_committed + 1;
1632         target_start_recovery_thread(obd, handler);
1633         obd->obd_recovery_start = CURRENT_SECONDS;
1634         /* Only used for lprocfs_status */
1635         obd->obd_recovery_end = obd->obd_recovery_start + OBD_RECOVERY_TIMEOUT;
1636         /* bz13079: this should be set to desired value for ost but not for mds */
1637         obd->obd_recovery_max_time = OBD_RECOVERY_MAX_TIME;
1638 }
1639 EXPORT_SYMBOL(target_recovery_init);
1640
1641 #endif
1642
1643 int target_process_req_flags(struct obd_device *obd, struct ptlrpc_request *req)
1644 {
1645         struct obd_export *exp = req->rq_export;
1646         LASSERT(exp != NULL);
1647         if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REQ_REPLAY_DONE) {
1648                 /* client declares he's ready to replay locks */
1649                 spin_lock_bh(&obd->obd_processing_task_lock);
1650                 if (exp->exp_req_replay_needed) {
1651                         LASSERT(atomic_read(&obd->obd_req_replay_clients) > 0);
1652                         spin_lock(&exp->exp_lock);
1653                         exp->exp_req_replay_needed = 0;
1654                         spin_unlock(&exp->exp_lock);
1655                         atomic_dec(&obd->obd_req_replay_clients);
1656                         LASSERT(obd->obd_recoverable_clients > 0);
1657                         obd->obd_recoverable_clients--;
1658                         if (atomic_read(&obd->obd_req_replay_clients) == 0)
1659                                 CDEBUG(D_HA, "all clients have replayed reqs\n");
1660                         wake_up(&obd->obd_next_transno_waitq);
1661                 }
1662                 spin_unlock_bh(&obd->obd_processing_task_lock);
1663         }
1664         if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_LOCK_REPLAY_DONE) {
1665                 /* client declares he's ready to complete recovery
1666                  * so, we put the request on th final queue */
1667                 spin_lock_bh(&obd->obd_processing_task_lock);
1668                 if (exp->exp_lock_replay_needed) {
1669                         LASSERT(atomic_read(&obd->obd_lock_replay_clients) > 0);
1670                         spin_lock(&exp->exp_lock);
1671                         exp->exp_lock_replay_needed = 0;
1672                         spin_unlock(&exp->exp_lock);
1673                         atomic_dec(&obd->obd_lock_replay_clients);
1674                         if (atomic_read(&obd->obd_lock_replay_clients) == 0)
1675                                 CDEBUG(D_HA, "all clients have replayed locks\n");
1676                         wake_up(&obd->obd_next_transno_waitq);
1677                 }
1678                 spin_unlock_bh(&obd->obd_processing_task_lock);
1679         }
1680
1681         return 0;
1682 }
1683
1684 int target_queue_recovery_request(struct ptlrpc_request *req,
1685                                   struct obd_device *obd)
1686 {
1687         struct list_head *tmp;
1688         int inserted = 0;
1689         __u64 transno = lustre_msg_get_transno(req->rq_reqmsg);
1690
1691         ENTRY;
1692
1693         if (obd->obd_recovery_data.trd_processing_task == cfs_curproc_pid()) {
1694                 /* Processing the queue right now, don't re-add. */
1695                 RETURN(1);
1696         }
1697
1698         target_process_req_flags(obd, req);
1699
1700         if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_LOCK_REPLAY_DONE) {
1701                 /* client declares he's ready to complete recovery
1702                  * so, we put the request on th final queue */
1703                 req = ptlrpc_clone_req(req);
1704                 if (req == NULL)
1705                         RETURN(-ENOMEM);
1706                 DEBUG_REQ(D_HA, req, "queue final req");
1707                 spin_lock_bh(&obd->obd_processing_task_lock);
1708                 if (obd->obd_recovering)
1709                         list_add_tail(&req->rq_list, &obd->obd_final_req_queue);
1710                 else {
1711                         spin_unlock_bh(&obd->obd_processing_task_lock);
1712                         ptlrpc_free_clone(req);
1713                         if (obd->obd_stopping) {
1714                                 RETURN(-ENOTCONN);
1715                         } else {
1716                                 RETURN(1);
1717                         }
1718                 }
1719                 spin_unlock_bh(&obd->obd_processing_task_lock);
1720                 RETURN(0);
1721         }
1722         if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REQ_REPLAY_DONE) {
1723                 /* client declares he's ready to replay locks */
1724                 req = ptlrpc_clone_req(req);
1725                 if (req == NULL)
1726                         RETURN(-ENOMEM);
1727                 DEBUG_REQ(D_HA, req, "queue lock replay req");
1728                 spin_lock_bh(&obd->obd_processing_task_lock);
1729                 LASSERT(obd->obd_recovering);
1730                 /* usually due to recovery abort */
1731                 if (!req->rq_export->exp_in_recovery) {
1732                         spin_unlock_bh(&obd->obd_processing_task_lock);
1733                         ptlrpc_free_clone(req);
1734                         RETURN(-ENOTCONN);
1735                 }
1736                 LASSERT(req->rq_export->exp_lock_replay_needed);
1737                 list_add_tail(&req->rq_list, &obd->obd_lock_replay_queue);
1738                 spin_unlock_bh(&obd->obd_processing_task_lock);
1739                 wake_up(&obd->obd_next_transno_waitq);
1740                 RETURN(0);
1741         }
1742
1743         /* CAVEAT EMPTOR: The incoming request message has been swabbed
1744          * (i.e. buflens etc are in my own byte order), but type-dependent
1745          * buffers (eg mds_body, ost_body etc) have NOT been swabbed. */
1746
1747         if (!transno) {
1748                 CFS_INIT_LIST_HEAD(&req->rq_list);
1749                 DEBUG_REQ(D_HA, req, "not queueing");
1750                 RETURN(1);
1751         }
1752
1753         spin_lock_bh(&obd->obd_processing_task_lock);
1754
1755         /* If we're processing the queue, we want don't want to queue this
1756          * message.
1757          *
1758          * Also, if this request has a transno less than the one we're waiting
1759          * for, we should process it now.  It could (and currently always will)
1760          * be an open request for a descriptor that was opened some time ago.
1761          *
1762          * Also, a resent, replayed request that has already been
1763          * handled will pass through here and be processed immediately.
1764          */
1765         CWARN("Next recovery transno: "LPU64", current: "LPU64", replaying: %i\n",
1766               obd->obd_next_recovery_transno, transno, obd->obd_req_replaying);
1767         if (transno < obd->obd_next_recovery_transno && obd->obd_req_replaying) {
1768                 /* Processing the queue right now, don't re-add. */
1769                 LASSERT(list_empty(&req->rq_list));
1770                 spin_unlock_bh(&obd->obd_processing_task_lock);
1771                 RETURN(1);
1772         }
1773         spin_unlock_bh(&obd->obd_processing_task_lock);
1774
1775         /* A resent, replayed request that is still on the queue; just drop it.
1776            The queued request will handle this. */
1777         if ((lustre_msg_get_flags(req->rq_reqmsg) & (MSG_RESENT|MSG_REPLAY)) ==
1778             (MSG_RESENT | MSG_REPLAY)) {
1779                 DEBUG_REQ(D_ERROR, req, "dropping resent queued req");
1780                 RETURN(0);
1781         }
1782
1783         req = ptlrpc_clone_req(req);
1784         if (req == NULL)
1785                 RETURN(-ENOMEM);
1786
1787         spin_lock_bh(&obd->obd_processing_task_lock);
1788         LASSERT(obd->obd_recovering);
1789         if (!req->rq_export->exp_in_recovery) {
1790                 spin_unlock_bh(&obd->obd_processing_task_lock);
1791                 ptlrpc_free_clone(req);
1792                 RETURN(-ENOTCONN);
1793         }
1794         LASSERT(req->rq_export->exp_req_replay_needed);
1795
1796         /* XXX O(n^2) */
1797         list_for_each(tmp, &obd->obd_req_replay_queue) {
1798                 struct ptlrpc_request *reqiter =
1799                         list_entry(tmp, struct ptlrpc_request, rq_list);
1800
1801                 if (lustre_msg_get_transno(reqiter->rq_reqmsg) > transno) {
1802                         list_add_tail(&req->rq_list, &reqiter->rq_list);
1803                         inserted = 1;
1804                         break;
1805                 }
1806         }
1807
1808         if (!inserted)
1809                 list_add_tail(&req->rq_list, &obd->obd_req_replay_queue);
1810
1811         obd->obd_requests_queued_for_recovery++;
1812         wake_up(&obd->obd_next_transno_waitq);
1813         spin_unlock_bh(&obd->obd_processing_task_lock);
1814         RETURN(0);
1815
1816 }
1817
1818 struct obd_device * target_req2obd(struct ptlrpc_request *req)
1819 {
1820         return req->rq_export->exp_obd;
1821 }
1822
1823 static inline struct ldlm_pool *ldlm_exp2pl(struct obd_export *exp)
1824 {
1825         LASSERT(exp != NULL);
1826         return &exp->exp_obd->obd_namespace->ns_pool;
1827 }
1828
1829 int target_pack_pool_reply(struct ptlrpc_request *req)
1830 {
1831         struct ldlm_pool *pl;
1832         ENTRY;
1833
1834         if (!req->rq_export || !req->rq_export->exp_obd ||
1835             !req->rq_export->exp_obd->obd_namespace ||
1836             !exp_connect_lru_resize(req->rq_export)) {
1837                 lustre_msg_set_slv(req->rq_repmsg, 0);
1838                 lustre_msg_set_limit(req->rq_repmsg, 0);
1839                 RETURN(0);
1840         }
1841
1842         pl = ldlm_exp2pl(req->rq_export);
1843
1844         spin_lock(&pl->pl_lock);
1845         LASSERT(ldlm_pool_get_slv(pl) != 0 && ldlm_pool_get_limit(pl) != 0);
1846         lustre_msg_set_slv(req->rq_repmsg, ldlm_pool_get_slv(pl));
1847         lustre_msg_set_limit(req->rq_repmsg, ldlm_pool_get_limit(pl));
1848         spin_unlock(&pl->pl_lock);
1849
1850         RETURN(0);
1851 }
1852
1853 int target_send_reply_msg(struct ptlrpc_request *req, int rc, int fail_id)
1854 {
1855         if (OBD_FAIL_CHECK_ORSET(fail_id & ~OBD_FAIL_ONCE, OBD_FAIL_ONCE)) {
1856                 DEBUG_REQ(D_ERROR, req, "dropping reply");
1857                 return (-ECOMM);
1858         }
1859
1860         if (unlikely(rc)) {
1861                 DEBUG_REQ(D_ERROR, req, "processing error (%d)", rc);
1862                 req->rq_status = rc;
1863                 return (ptlrpc_error(req));
1864         } else {
1865                 DEBUG_REQ(D_NET, req, "sending reply");
1866         }
1867
1868         return (ptlrpc_send_reply(req, 1));
1869 }
1870
1871 void target_send_reply(struct ptlrpc_request *req, int rc, int fail_id)
1872 {
1873         int                        netrc;
1874         struct ptlrpc_reply_state *rs;
1875         struct obd_device         *obd;
1876         struct obd_export         *exp;
1877         struct ptlrpc_service     *svc;
1878
1879         if (req->rq_no_reply)
1880                 return;
1881
1882         svc = req->rq_rqbd->rqbd_service;
1883         rs = req->rq_reply_state;
1884         if (rs == NULL || !rs->rs_difficult) {
1885                 /* no notifiers */
1886                 target_send_reply_msg (req, rc, fail_id);
1887                 return;
1888         }
1889
1890         /* must be an export if locks saved */
1891         LASSERT (req->rq_export != NULL);
1892         /* req/reply consistent */
1893         LASSERT (rs->rs_service == svc);
1894
1895         /* "fresh" reply */
1896         LASSERT (!rs->rs_scheduled);
1897         LASSERT (!rs->rs_scheduled_ever);
1898         LASSERT (!rs->rs_handled);
1899         LASSERT (!rs->rs_on_net);
1900         LASSERT (rs->rs_export == NULL);
1901         LASSERT (list_empty(&rs->rs_obd_list));
1902         LASSERT (list_empty(&rs->rs_exp_list));
1903
1904         exp = class_export_get (req->rq_export);
1905         obd = exp->exp_obd;
1906
1907         /* disable reply scheduling onto srv_reply_queue while I'm setting up */
1908         rs->rs_scheduled = 1;
1909         rs->rs_on_net    = 1;
1910         rs->rs_xid       = req->rq_xid;
1911         rs->rs_transno   = req->rq_transno;
1912         rs->rs_export    = exp;
1913
1914         spin_lock(&obd->obd_uncommitted_replies_lock);
1915
1916         if (rs->rs_transno > obd->obd_last_committed) {
1917                 /* not committed already */
1918                 list_add_tail (&rs->rs_obd_list,
1919                                &obd->obd_uncommitted_replies);
1920         }
1921
1922         spin_unlock (&obd->obd_uncommitted_replies_lock);
1923         spin_lock (&exp->exp_lock);
1924
1925         list_add_tail (&rs->rs_exp_list, &exp->exp_outstanding_replies);
1926
1927         spin_unlock(&exp->exp_lock);
1928
1929         netrc = target_send_reply_msg (req, rc, fail_id);
1930
1931         spin_lock(&svc->srv_lock);
1932
1933         svc->srv_n_difficult_replies++;
1934
1935         if (netrc != 0) {
1936                 /* error sending: reply is off the net.  Also we need +1
1937                  * reply ref until ptlrpc_server_handle_reply() is done
1938                  * with the reply state (if the send was successful, there
1939                  * would have been +1 ref for the net, which
1940                  * reply_out_callback leaves alone) */
1941                 rs->rs_on_net = 0;
1942                 ptlrpc_rs_addref(rs);
1943                 atomic_inc (&svc->srv_outstanding_replies);
1944         }
1945
1946         if (!rs->rs_on_net ||                   /* some notifier */
1947             list_empty(&rs->rs_exp_list) ||     /* completed already */
1948             list_empty(&rs->rs_obd_list)) {
1949                 list_add_tail (&rs->rs_list, &svc->srv_reply_queue);
1950                 cfs_waitq_signal (&svc->srv_waitq);
1951         } else {
1952                 list_add (&rs->rs_list, &svc->srv_active_replies);
1953                 rs->rs_scheduled = 0;           /* allow notifier to schedule */
1954         }
1955
1956         spin_unlock(&svc->srv_lock);
1957 }
1958
1959 int target_handle_ping(struct ptlrpc_request *req)
1960 {
1961         obd_ping(req->rq_export);
1962         return req_capsule_server_pack(&req->rq_pill);
1963 }
1964
1965 void target_committed_to_req(struct ptlrpc_request *req)
1966 {
1967         struct obd_device *obd;
1968
1969         if (req == NULL || req->rq_export == NULL)
1970                 return;
1971
1972         obd = req->rq_export->exp_obd;
1973         if (obd == NULL)
1974                 return;
1975
1976         if (!obd->obd_no_transno && req->rq_repmsg != NULL)
1977                 lustre_msg_set_last_committed(req->rq_repmsg,
1978                                               obd->obd_last_committed);
1979         else
1980                 DEBUG_REQ(D_IOCTL, req, "not sending last_committed update");
1981
1982         CDEBUG(D_INFO, "last_committed "LPU64", transno "LPU64", xid "LPU64"\n",
1983                obd->obd_last_committed, req->rq_transno, req->rq_xid);
1984 }
1985
1986 EXPORT_SYMBOL(target_committed_to_req);
1987
1988 #ifdef HAVE_QUOTA_SUPPORT
1989 int target_handle_qc_callback(struct ptlrpc_request *req)
1990 {
1991         struct obd_quotactl *oqctl;
1992         struct client_obd *cli = &req->rq_export->exp_obd->u.cli;
1993
1994         oqctl = req_capsule_client_get(&req->rq_pill, &RMF_OBD_QUOTACTL);
1995         if (oqctl == NULL)
1996                 RETURN(-EPROTO);
1997
1998         cli->cl_qchk_stat = oqctl->qc_stat;
1999
2000         return 0;
2001 }
2002
2003 int target_handle_dqacq_callback(struct ptlrpc_request *req)
2004 {
2005 #ifdef __KERNEL__
2006         struct obd_device *obd = req->rq_export->exp_obd;
2007         struct obd_device *master_obd;
2008         struct lustre_quota_ctxt *qctxt;
2009         struct qunit_data *qdata;
2010         void* rep;
2011         struct qunit_data_old *qdata_old;
2012         int rc = 0;
2013         ENTRY;
2014
2015         rc = req_capsule_server_pack(&req->rq_pill);
2016         if (rc) {
2017                 CERROR("packing reply failed!: rc = %d\n", rc);
2018                 RETURN(rc);
2019         }
2020
2021         LASSERT(req->rq_export);
2022
2023         /* fixed for bug10707 */
2024         if ((req->rq_export->exp_connect_flags & OBD_CONNECT_QUOTA64) &&
2025             !OBD_FAIL_CHECK(OBD_FAIL_QUOTA_QD_COUNT_32BIT)) {
2026                 CDEBUG(D_QUOTA, "qd_count is 64bit!\n");
2027                 rep = req_capsule_server_get(&req->rq_pill,
2028                                              &RMF_QUNIT_DATA);
2029                 LASSERT(rep);
2030                 qdata = req_capsule_client_swab_get(&req->rq_pill,
2031                                                     &RMF_QUNIT_DATA,
2032                                           (void*)lustre_swab_qdata);
2033         } else {
2034                 CDEBUG(D_QUOTA, "qd_count is 32bit!\n");
2035                 rep = req_capsule_server_get(&req->rq_pill, &RMF_QUNIT_DATA);
2036                 LASSERT(rep);
2037                 qdata_old = req_capsule_client_swab_get(&req->rq_pill,
2038                                                         &RMF_QUNIT_DATA,
2039                                            (void*)lustre_swab_qdata_old);
2040                 qdata = lustre_quota_old_to_new(qdata_old);
2041         }
2042
2043         if (qdata == NULL)
2044                 RETURN(-EPROTO);
2045
2046         /* we use the observer */
2047         LASSERT(obd->obd_observer && obd->obd_observer->obd_observer);
2048         master_obd = obd->obd_observer->obd_observer;
2049         qctxt = &master_obd->u.obt.obt_qctxt;
2050
2051         LASSERT(qctxt->lqc_handler);
2052         rc = qctxt->lqc_handler(master_obd, qdata,
2053                                 lustre_msg_get_opc(req->rq_reqmsg));
2054         if (rc && rc != -EDQUOT)
2055                 CDEBUG(rc == -EBUSY  ? D_QUOTA : D_ERROR,
2056                        "dqacq failed! (rc:%d)\n", rc);
2057
2058         /* the qd_count might be changed in lqc_handler */
2059         if ((req->rq_export->exp_connect_flags & OBD_CONNECT_QUOTA64) &&
2060             !OBD_FAIL_CHECK(OBD_FAIL_QUOTA_QD_COUNT_32BIT)) {
2061                 memcpy(rep, qdata, sizeof(*qdata));
2062         } else {
2063                 qdata_old = lustre_quota_new_to_old(qdata);
2064                 memcpy(rep, qdata_old, sizeof(*qdata_old));
2065         }
2066         req->rq_status = rc;
2067         rc = ptlrpc_reply(req);
2068
2069         RETURN(rc);
2070 #else
2071         return 0;
2072 #endif /* !__KERNEL__ */
2073 }
2074 #endif /* HAVE_QUOTA_SUPPORT */
2075
2076 ldlm_mode_t lck_compat_array[] = {
2077         [LCK_EX] LCK_COMPAT_EX,
2078         [LCK_PW] LCK_COMPAT_PW,
2079         [LCK_PR] LCK_COMPAT_PR,
2080         [LCK_CW] LCK_COMPAT_CW,
2081         [LCK_CR] LCK_COMPAT_CR,
2082         [LCK_NL] LCK_COMPAT_NL,
2083         [LCK_GROUP] LCK_COMPAT_GROUP
2084 };