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         /* Turn on checksumming by default. */
 280         cli->cl_checksum = 1;
 281         /*
 282          * The supported checksum types will be worked out at connect time
 283          * Set cl_chksum* to CRC32 for now to avoid returning screwed info
 284          * through procfs.
 285          */
 286         cli->cl_cksum_type = cli->cl_supp_cksum_types = OBD_CKSUM_CRC32;
 287 #endif
 288         atomic_set(&cli->cl_resends, OSC_DEFAULT_RESENDS);
 289
 290         /* This value may be changed at connect time in
 291            ptlrpc_connect_interpret. */
 292         cli->cl_max_pages_per_rpc = min((int)PTLRPC_MAX_BRW_PAGES,
 293                                         (int)(1024 * 1024 >> CFS_PAGE_SHIFT));
 294
 295         if (!strcmp(name, LUSTRE_MDC_NAME)) {
 296                 cli->cl_max_rpcs_in_flight = MDC_MAX_RIF_DEFAULT;
 297         } else if (num_physpages >> (20 - CFS_PAGE_SHIFT) <= 128 /* MB */) {
 298                 cli->cl_max_rpcs_in_flight = 2;
 299         } else if (num_physpages >> (20 - CFS_PAGE_SHIFT) <= 256 /* MB */) {
 300                 cli->cl_max_rpcs_in_flight = 3;
 301         } else if (num_physpages >> (20 - CFS_PAGE_SHIFT) <= 512 /* MB */) {
 302                 cli->cl_max_rpcs_in_flight = 4;
 303         } else {
 304                 cli->cl_max_rpcs_in_flight = OSC_MAX_RIF_DEFAULT;
 305         }
 306
 307         rc = ldlm_get_ref();
 308         if (rc) {
 309                 CERROR("ldlm_get_ref failed: %d\n", rc);
 310                 GOTO(err, rc);
 311         }
 312
 313         ptlrpc_init_client(rq_portal, rp_portal, name,
 314                            &obddev->obd_ldlm_client);
 315
 316         imp = class_new_import(obddev);
 317         if (imp == NULL)
 318                 GOTO(err_ldlm, rc = -ENOENT);
 319         imp->imp_client = &obddev->obd_ldlm_client;
 320         imp->imp_connect_op = connect_op;
 321         imp->imp_initial_recov = 1;
 322         imp->imp_initial_recov_bk = 0;
 323         CFS_INIT_LIST_HEAD(&imp->imp_pinger_chain);
 324         memcpy(cli->cl_target_uuid.uuid, lustre_cfg_buf(lcfg, 1),
 325                LUSTRE_CFG_BUFLEN(lcfg, 1));
 326         class_import_put(imp);
 327
 328         rc = client_import_add_conn(imp, &server_uuid, 1);
 329         if (rc) {
 330                 CERROR("can't add initial connection\n");
 331                 GOTO(err_import, rc);
 332         }
 333
 334         cli->cl_import = imp;
 335         /* cli->cl_max_mds_{easize,cookiesize} updated by mdc_init_ea_size() */
 336         cli->cl_max_mds_easize = sizeof(struct lov_mds_md);
 337         cli->cl_max_mds_cookiesize = sizeof(struct llog_cookie);
 338
 339         if (LUSTRE_CFG_BUFLEN(lcfg, 3) > 0) {
 340                 if (!strcmp(lustre_cfg_string(lcfg, 3), "inactive")) {
 341                         CDEBUG(D_HA, "marking %s %s->%s as inactive\n",
 342                                name, obddev->obd_name,
 343                                cli->cl_target_uuid.uuid);
 344                         spin_lock(&imp->imp_lock);
 345                         imp->imp_invalid = 1;
 346                         spin_unlock(&imp->imp_lock);
 347                 }
 348         }
 349
 350         cli->cl_qchk_stat = CL_NOT_QUOTACHECKED;
 351
 352         RETURN(rc);
 353
 354 err_import:
 355         class_destroy_import(imp);
 356 err_ldlm:
 357         ldlm_put_ref(0);
 358 err:
 359         RETURN(rc);
 360
 361 }
 362
 363 int client_obd_cleanup(struct obd_device *obddev)
 364 {
 365         ENTRY;
 366         sptlrpc_rule_set_free(&obddev->u.cli.cl_sptlrpc_rset);
 367         ldlm_put_ref(obddev->obd_force);
 368         RETURN(0);
 369 }
 370
 371 /* ->o_connect() method for client side (OSC and MDC and MGC) */
 372 int client_connect_import(const struct lu_env *env,
 373                           struct lustre_handle *dlm_handle,
 374                           struct obd_device *obd, struct obd_uuid *cluuid,
 375                           struct obd_connect_data *data)
 376 {
 377         struct client_obd *cli = &obd->u.cli;
 378         struct obd_import *imp = cli->cl_import;
 379         struct obd_export *exp;
 380         struct obd_connect_data *ocd;
 381         struct ldlm_namespace *to_be_freed = NULL;
 382         int rc;
 383         ENTRY;
 384
 385         mutex_down(&cli->cl_sem);
 386         rc = class_connect(dlm_handle, obd, cluuid);
 387         if (rc)
 388                 GOTO(out_sem, rc);
 389
 390         cli->cl_conn_count++;
 391         if (cli->cl_conn_count > 1)
 392                 GOTO(out_sem, rc);
 393         exp = class_conn2export(dlm_handle);
 394
 395         if (obd->obd_namespace != NULL)
 396                 CERROR("already have namespace!\n");
 397         obd->obd_namespace = ldlm_namespace_new(obd->obd_name,
 398                                                 LDLM_NAMESPACE_CLIENT,
 399                                                 LDLM_NAMESPACE_GREEDY);
 400         if (obd->obd_namespace == NULL)
 401                 GOTO(out_disco, rc = -ENOMEM);
 402
 403         imp->imp_dlm_handle = *dlm_handle;
 404         rc = ptlrpc_init_import(imp);
 405         if (rc != 0)
 406                 GOTO(out_ldlm, rc);
 407
 408         ocd = &imp->imp_connect_data;
 409         if (data) {
 410                 *ocd = *data;
 411                 imp->imp_connect_flags_orig = data->ocd_connect_flags;
 412         }
 413
 414         rc = ptlrpc_connect_import(imp, NULL);
 415         if (rc != 0) {
 416                 LASSERT (imp->imp_state == LUSTRE_IMP_DISCON);
 417                 GOTO(out_ldlm, rc);
 418         }
 419         LASSERT(exp->exp_connection);
 420
 421         if (data) {
 422                 LASSERTF((ocd->ocd_connect_flags & data->ocd_connect_flags) ==
 423                          ocd->ocd_connect_flags, "old "LPX64", new "LPX64"\n",
 424                          data->ocd_connect_flags, ocd->ocd_connect_flags);
 425                 data->ocd_connect_flags = ocd->ocd_connect_flags;
 426         }
 427
 428         ptlrpc_pinger_add_import(imp);
 429
 430         EXIT;
 431
 432         if (rc) {
 433 out_ldlm:
 434                 ldlm_namespace_free_prior(obd->obd_namespace);
 435                 to_be_freed = obd->obd_namespace;
 436                 obd->obd_namespace = NULL;
 437 out_disco:
 438                 cli->cl_conn_count--;
 439                 class_disconnect(exp);
 440         } else {
 441                 class_export_put(exp);
 442         }
 443 out_sem:
 444         mutex_up(&cli->cl_sem);
 445         if (to_be_freed)
 446                 ldlm_namespace_free_post(to_be_freed, 1);
 447
 448         return rc;
 449 }
 450
 451 int client_disconnect_export(struct obd_export *exp)
 452 {
 453         struct obd_device *obd = class_exp2obd(exp);
 454         struct client_obd *cli;
 455         struct obd_import *imp;
 456         int rc = 0, err;
 457         struct ldlm_namespace *to_be_freed = NULL;
 458         ENTRY;
 459
 460         if (!obd) {
 461                 CERROR("invalid export for disconnect: exp %p cookie "LPX64"\n",
 462                        exp, exp ? exp->exp_handle.h_cookie : -1);
 463                 RETURN(-EINVAL);
 464         }
 465
 466         cli = &obd->u.cli;
 467         imp = cli->cl_import;
 468
 469         mutex_down(&cli->cl_sem);
 470         if (!cli->cl_conn_count) {
 471                 CERROR("disconnecting disconnected device (%s)\n",
 472                        obd->obd_name);
 473                 GOTO(out_sem, rc = -EINVAL);
 474         }
 475
 476         cli->cl_conn_count--;
 477         if (cli->cl_conn_count)
 478                 GOTO(out_no_disconnect, rc = 0);
 479
 480         /* Mark import deactivated now, so we don't try to reconnect if any
 481          * of the cleanup RPCs fails (e.g. ldlm cancel, etc).  We don't
 482          * fully deactivate the import, or that would drop all requests. */
 483         spin_lock(&imp->imp_lock);
 484         imp->imp_deactive = 1;
 485         spin_unlock(&imp->imp_lock);
 486
 487         /* Some non-replayable imports (MDS's OSCs) are pinged, so just
 488          * delete it regardless.  (It's safe to delete an import that was
 489          * never added.) */
 490         (void)ptlrpc_pinger_del_import(imp);
 491
 492         if (obd->obd_namespace != NULL) {
 493                 /* obd_force == local only */
 494                 ldlm_cli_cancel_unused(obd->obd_namespace, NULL,
 495                                        obd->obd_force ? LDLM_FL_LOCAL_ONLY:0,
 496                                        NULL);
 497                 ldlm_namespace_free_prior(obd->obd_namespace);
 498                 to_be_freed = obd->obd_namespace;
 499         }
 500
 501         rc = ptlrpc_disconnect_import(imp, 0);
 502
 503         ptlrpc_invalidate_import(imp);
 504         /* set obd_namespace to NULL only after invalidate, because we can have
 505          * some connect requests in flight, and his need store a connect flags
 506          * in obd_namespace. bug 14260 */
 507         obd->obd_namespace = NULL;
 508
 509         ptlrpc_free_rq_pool(imp->imp_rq_pool);
 510         destroy_import(imp);
 511         cli->cl_import = NULL;
 512
 513         EXIT;
 514  out_no_disconnect:
 515         err = class_disconnect(exp);
 516         if (!rc && err)
 517                 rc = err;
 518  out_sem:
 519         mutex_up(&cli->cl_sem);
 520         if (to_be_freed)
 521                 ldlm_namespace_free_post(to_be_freed, obd->obd_force);
 522
 523         RETURN(rc);
 524 }
 525
 526 /* --------------------------------------------------------------------------
 527  * from old lib/target.c
 528  * -------------------------------------------------------------------------- */
 529
 530 int target_handle_reconnect(struct lustre_handle *conn, struct obd_export *exp,
 531                             struct obd_uuid *cluuid, int initial_conn)
 532 {
 533         ENTRY;
 534         if (exp->exp_connection && exp->exp_imp_reverse && !initial_conn) {
 535                 struct lustre_handle *hdl;
 536                 hdl = &exp->exp_imp_reverse->imp_remote_handle;
 537                 /* Might be a re-connect after a partition. */
 538                 if (!memcmp(&conn->cookie, &hdl->cookie, sizeof conn->cookie)) {
 539                         CWARN("%s: %s reconnecting\n", exp->exp_obd->obd_name,
 540                               cluuid->uuid);
 541                         conn->cookie = exp->exp_handle.h_cookie;
 542                         /* target_handle_connect() treats EALREADY and
 543                          * -EALREADY differently.  EALREADY means we are
 544                          * doing a valid reconnect from the same client. */
 545                         RETURN(EALREADY);
 546                 } else {
 547                         CERROR("%s reconnecting from %s, "
 548                                "handle mismatch (ours "LPX64", theirs "
 549                                LPX64")\n", cluuid->uuid,
 550                                exp->exp_connection->c_remote_uuid.uuid,
 551                                hdl->cookie, conn->cookie);
 552                         memset(conn, 0, sizeof *conn);
 553                         /* target_handle_connect() treats EALREADY and
 554                          * -EALREADY differently.  -EALREADY is an error
 555                          * (same UUID, different handle). */
 556                         RETURN(-EALREADY);
 557                 }
 558         }
 559
 560         conn->cookie = exp->exp_handle.h_cookie;
 561         CDEBUG(D_HA, "connect export for UUID '%s' at %p, cookie "LPX64"\n",
 562                cluuid->uuid, exp, conn->cookie);
 563         RETURN(0);
 564 }
 565
 566 void target_client_add_cb(struct obd_device *obd, __u64 transno, void *cb_data,
 567                           int error)
 568 {
 569         struct obd_export *exp = cb_data;
 570
 571         CDEBUG(D_RPCTRACE, "%s: committing for initial connect of %s\n",
 572                obd->obd_name, exp->exp_client_uuid.uuid);
 573
 574         spin_lock(&exp->exp_lock);
 575         exp->exp_need_sync = 0;
 576         spin_unlock(&exp->exp_lock);
 577 }
 578 EXPORT_SYMBOL(target_client_add_cb);
 579
 580 int target_handle_connect(struct ptlrpc_request *req)
 581 {
 582         struct obd_device *target, *targref = NULL;
 583         struct obd_export *export = NULL;
 584         struct obd_import *revimp;
 585         struct lustre_handle conn;
 586         struct lustre_handle *tmp;
 587         struct obd_uuid tgtuuid;
 588         struct obd_uuid cluuid;
 589         struct obd_uuid remote_uuid;
 590         char *str;
 591         int rc = 0;
 592         int initial_conn = 0;
 593         struct obd_connect_data *data, *tmpdata;
 594         ENTRY;
 595
 596         OBD_RACE(OBD_FAIL_TGT_CONN_RACE);
 597
 598         str = req_capsule_client_get(&req->rq_pill, &RMF_TGTUUID);
 599         if (str == NULL) {
 600                 DEBUG_REQ(D_ERROR, req, "bad target UUID for connect");
 601                 GOTO(out, rc = -EINVAL);
 602         }
 603
 604         obd_str2uuid(&tgtuuid, str);
 605         target = class_uuid2obd(&tgtuuid);
 606         if (!target)
 607                 target = class_name2obd(str);
 608
 609         if (!target || target->obd_stopping || !target->obd_set_up) {
 610                 LCONSOLE_ERROR_MSG(0x137, "UUID '%s' is not available "
 611                                    " for connect (%s)\n", str,
 612                                    !target ? "no target" :
 613                                    (target->obd_stopping ? "stopping" :
 614                                    "not set up"));
 615                 GOTO(out, rc = -ENODEV);
 616         }
 617
 618         if (target->obd_no_conn) {
 619                 LCONSOLE_WARN("%s: temporarily refusing client connection "
 620                               "from %s\n", target->obd_name,
 621                               libcfs_nid2str(req->rq_peer.nid));
 622                 GOTO(out, rc = -EAGAIN);
 623         }
 624
 625         /* Make sure the target isn't cleaned up while we're here. Yes,
 626            there's still a race between the above check and our incref here.
 627            Really, class_uuid2obd should take the ref. */
 628         targref = class_incref(target);
 629
 630
 631         str = req_capsule_client_get(&req->rq_pill, &RMF_CLUUID);
 632         if (str == NULL) {
 633                 DEBUG_REQ(D_ERROR, req, "bad client UUID for connect");
 634                 GOTO(out, rc = -EINVAL);
 635         }
 636
 637         obd_str2uuid(&cluuid, str);
 638
 639         /* XXX extract a nettype and format accordingly */
 640         switch (sizeof(lnet_nid_t)) {
 641                 /* NB the casts only avoid compiler warnings */
 642         case 8:
 643                 snprintf(remote_uuid.uuid, sizeof remote_uuid,
 644                          "NET_"LPX64"_UUID", (__u64)req->rq_peer.nid);
 645                 break;
 646         case 4:
 647                 snprintf(remote_uuid.uuid, sizeof remote_uuid,
 648                          "NET_%x_UUID", (__u32)req->rq_peer.nid);
 649                 break;
 650         default:
 651                 LBUG();
 652         }
 653
 654         tmp = req_capsule_client_get(&req->rq_pill, &RMF_CONN);
 655         if (tmp == NULL)
 656                 GOTO(out, rc = -EPROTO);
 657
 658         conn = *tmp;
 659
 660         data = req_capsule_client_get(&req->rq_pill, &RMF_CONNECT_DATA);
 661         if (!data)
 662                 GOTO(out, rc = -EPROTO);
 663
 664         rc = req_capsule_server_pack(&req->rq_pill);
 665         if (rc)
 666                 GOTO(out, rc);
 667
 668         if (lustre_msg_get_op_flags(req->rq_reqmsg) & MSG_CONNECT_LIBCLIENT) {
 669                 if (!data) {
 670                         DEBUG_REQ(D_WARNING, req, "Refusing old (unversioned) "
 671                                   "libclient connection attempt");
 672                         GOTO(out, rc = -EPROTO);
 673                 } else if (data->ocd_version < LUSTRE_VERSION_CODE -
 674                                                LUSTRE_VERSION_ALLOWED_OFFSET ||
 675                            data->ocd_version > LUSTRE_VERSION_CODE +
 676                                                LUSTRE_VERSION_ALLOWED_OFFSET) {
 677                         DEBUG_REQ(D_WARNING, req, "Refusing %s (%d.%d.%d.%d) "
 678                                   "libclient connection attempt",
 679                                   data->ocd_version < LUSTRE_VERSION_CODE ?
 680                                   "old" : "new",
 681                                   OBD_OCD_VERSION_MAJOR(data->ocd_version),
 682                                   OBD_OCD_VERSION_MINOR(data->ocd_version),
 683                                   OBD_OCD_VERSION_PATCH(data->ocd_version),
 684                                   OBD_OCD_VERSION_FIX(data->ocd_version));
 685                         data = req_capsule_server_sized_get(&req->rq_pill,
 686                                                         &RMF_CONNECT_DATA,
 687                                     offsetof(typeof(*data), ocd_version) +
 688                                              sizeof(data->ocd_version));
 689                         if (data) {
 690                                 data->ocd_connect_flags = OBD_CONNECT_VERSION;
 691                                 data->ocd_version = LUSTRE_VERSION_CODE;
 692                         }
 693                         GOTO(out, rc = -EPROTO);
 694                 }
 695         }
 696
 697         if (lustre_msg_get_op_flags(req->rq_reqmsg) & MSG_CONNECT_INITIAL)
 698                 initial_conn = 1;
 699
 700         /* lctl gets a backstage, all-access pass. */
 701         if (obd_uuid_equals(&cluuid, &target->obd_uuid))
 702                 goto dont_check_exports;
 703
 704         spin_lock(&target->obd_dev_lock);
 705         export = lustre_hash_get_object_by_key(target->obd_uuid_hash_body, &cluuid);
 706
 707         if (export != NULL && export->exp_connecting) { /* bug 9635, et. al. */
 708                 CWARN("%s: exp %p already connecting\n",
 709                       export->exp_obd->obd_name, export);
 710                 class_export_put(export);
 711                 export = NULL;
 712                 rc = -EALREADY;
 713         } else if (export != NULL && export->exp_connection != NULL &&
 714                    req->rq_peer.nid != export->exp_connection->c_peer.nid) {
 715                 /* make darn sure this is coming from the same peer
 716                  * if the UUIDs matched */
 717                   CWARN("%s: cookie %s seen on new NID %s when "
 718                           "existing NID %s is already connected\n",
 719                         target->obd_name, cluuid.uuid,
 720                   libcfs_nid2str(req->rq_peer.nid),
 721                   libcfs_nid2str(export->exp_connection->c_peer.nid));
 722                   class_export_put(export);
 723                   export = NULL;
 724                   rc = -EALREADY;
 725         } else if (export != NULL) {
 726                 spin_lock(&export->exp_lock);
 727                 export->exp_connecting = 1;
 728                 spin_unlock(&export->exp_lock);
 729                 class_export_put(export);
 730                 spin_unlock(&target->obd_dev_lock);
 731                 LASSERT(export->exp_obd == target);
 732
 733                 rc = target_handle_reconnect(&conn, export, &cluuid, initial_conn);
 734         }
 735
 736         /* If we found an export, we already unlocked. */
 737         if (!export) {
 738                 spin_unlock(&target->obd_dev_lock);
 739                 OBD_FAIL_TIMEOUT(OBD_FAIL_TGT_DELAY_CONNECT, 2 * obd_timeout);
 740         } else if (req->rq_export == NULL &&
 741                    atomic_read(&export->exp_rpc_count) > 0) {
 742                 CWARN("%s: refuse connection from %s/%s to 0x%p/%d\n",
 743                       target->obd_name, cluuid.uuid,
 744                       libcfs_nid2str(req->rq_peer.nid),
 745                       export, atomic_read(&export->exp_refcount));
 746                 GOTO(out, rc = -EBUSY);
 747         } else if (req->rq_export != NULL &&
 748                    (atomic_read(&export->exp_rpc_count) > 1)) {
 749                 CWARN("%s: refuse reconnection from %s@%s to 0x%p/%d\n",
 750                       target->obd_name, cluuid.uuid,
 751                       libcfs_nid2str(req->rq_peer.nid),
 752                       export, atomic_read(&export->exp_rpc_count));
 753                 GOTO(out, rc = -EBUSY);
 754         } else if (lustre_msg_get_conn_cnt(req->rq_reqmsg) == 1 &&
 755                    !initial_conn) {
 756                 CERROR("%s: NID %s (%s) reconnected with 1 conn_cnt; "
 757                        "cookies not random?\n", target->obd_name,
 758                        libcfs_nid2str(req->rq_peer.nid), cluuid.uuid);
 759                 GOTO(out, rc = -EALREADY);
 760         } else {
 761                 OBD_FAIL_TIMEOUT(OBD_FAIL_TGT_DELAY_RECONNECT, 2 * obd_timeout);
 762                 if (req->rq_export == NULL && initial_conn)
 763                        export->exp_last_request_time =
 764                                max(export->exp_last_request_time,
 765                                    (time_t)CURRENT_SECONDS);
 766         }
 767
 768         /* We want to handle EALREADY but *not* -EALREADY from
 769          * target_handle_reconnect(), return reconnection state in a flag */
 770         if (rc == EALREADY) {
 771                 lustre_msg_add_op_flags(req->rq_repmsg, MSG_CONNECT_RECONNECT);
 772                 rc = 0;
 773         } else if (rc) {
 774                 GOTO(out, rc);
 775         }
 776         /* Tell the client if we're in recovery. */
 777         /* If this is the first client, start the recovery timer */
 778         CWARN("%s: connection from %s@%s %st"LPU64" exp %p cur %ld last %ld\n",
 779                target->obd_name, cluuid.uuid, libcfs_nid2str(req->rq_peer.nid),
 780               target->obd_recovering ? "recovering/" : "", data->ocd_transno,
 781               export, (long)CURRENT_SECONDS,
 782               export ? (long)export->exp_last_request_time : 0);
 783
 784
 785         if (target->obd_recovering) {
 786                 lustre_msg_add_op_flags(req->rq_repmsg, MSG_CONNECT_RECOVERING);
 787                 target_start_recovery_timer(target);
 788         }
 789
 790         /* Tell the client if we support replayable requests */
 791         if (target->obd_replayable)
 792                 lustre_msg_add_op_flags(req->rq_repmsg, MSG_CONNECT_REPLAYABLE);
 793
 794         if (export == NULL) {
 795                 if (target->obd_recovering) {
 796                         cfs_time_t t;
 797
 798                         t = cfs_timer_deadline(&target->obd_recovery_timer);
 799                         t = cfs_time_sub(t, cfs_time_current());
 800                         CERROR("%s: denying connection for new client %s (%s): "
 801                                "%d clients in recovery for %lds\n",
 802                                target->obd_name,
 803                                libcfs_nid2str(req->rq_peer.nid), cluuid.uuid,
 804                                target->obd_recoverable_clients,
 805                                cfs_duration_sec(t));
 806                         rc = -EBUSY;
 807                 } else {
 808 dont_check_exports:
 809                         rc = obd_connect(req->rq_svc_thread->t_env,
 810                                          &conn, target, &cluuid, data);
 811                 }
 812         } else {
 813                 rc = obd_reconnect(req->rq_svc_thread->t_env,
 814                                    export, target, &cluuid, data);
 815         }
 816         if (rc)
 817                 GOTO(out, rc);
 818         /* Return only the parts of obd_connect_data that we understand, so the
 819          * client knows that we don't understand the rest. */
 820         if (data) {
 821                  tmpdata = req_capsule_server_get(&req->rq_pill,
 822                                                   &RMF_CONNECT_DATA);
 823                   //data->ocd_connect_flags &= OBD_CONNECT_SUPPORTED;
 824                  *tmpdata = *data;
 825         }
 826
 827         /* If all else goes well, this is our RPC return code. */
 828         req->rq_status = 0;
 829
 830         lustre_msg_set_handle(req->rq_repmsg, &conn);
 831
 832         /* ownership of this export ref transfers to the request AFTER we
 833          * drop any previous reference the request had, but we don't want
 834          * that to go to zero before we get our new export reference. */
 835         export = class_conn2export(&conn);
 836         if (!export) {
 837                 DEBUG_REQ(D_ERROR, req, "Missing export!");
 838                 GOTO(out, rc = -ENODEV);
 839         }
 840
 841         /* If the client and the server are the same node, we will already
 842          * have an export that really points to the client's DLM export,
 843          * because we have a shared handles table.
 844          *
 845          * XXX this will go away when shaver stops sending the "connect" handle
 846          * in the real "remote handle" field of the request --phik 24 Apr 2003
 847          */
 848         if (req->rq_export != NULL)
 849                 class_export_put(req->rq_export);
 850
 851         req->rq_export = export;
 852
 853         spin_lock(&export->exp_lock);
 854         if (initial_conn) {
 855                 lustre_msg_set_conn_cnt(req->rq_repmsg, export->exp_conn_cnt + 1);
 856         } else if (export->exp_conn_cnt >= lustre_msg_get_conn_cnt(req->rq_reqmsg)) {
 857                 spin_unlock(&export->exp_lock);
 858                 CERROR("%s: %s already connected at higher conn_cnt: %d > %d\n",
 859                        cluuid.uuid, libcfs_nid2str(req->rq_peer.nid),
 860                        export->exp_conn_cnt,
 861                        lustre_msg_get_conn_cnt(req->rq_reqmsg));
 862
 863                 GOTO(out, rc = -EALREADY);
 864         }
 865         export->exp_conn_cnt = lustre_msg_get_conn_cnt(req->rq_reqmsg);
 866
 867         /* request from liblustre?  Don't evict it for not pinging. */
 868         if (lustre_msg_get_op_flags(req->rq_reqmsg) & MSG_CONNECT_LIBCLIENT) {
 869                 export->exp_libclient = 1;
 870                 spin_unlock(&export->exp_lock);
 871
 872                 spin_lock(&target->obd_dev_lock);
 873                 list_del_init(&export->exp_obd_chain_timed);
 874                 spin_unlock(&target->obd_dev_lock);
 875         } else {
 876                 spin_unlock(&export->exp_lock);
 877         }
 878
 879         if (export->exp_connection != NULL)
 880                 ptlrpc_put_connection(export->exp_connection);
 881         export->exp_connection = ptlrpc_get_connection(req->rq_peer,
 882                                                        req->rq_self,
 883                                                        &remote_uuid);
 884
 885         spin_lock(&target->obd_dev_lock);
 886         /* Export might be hashed already, e.g. if this is reconnect */
 887         if (hlist_unhashed(&export->exp_nid_hash))
 888                 lustre_hash_additem(export->exp_obd->obd_nid_hash_body,
 889                                     &export->exp_connection->c_peer.nid,
 890                                     &export->exp_nid_hash);
 891         spin_unlock(&target->obd_dev_lock);
 892
 893         spin_lock_bh(&target->obd_processing_task_lock);
 894         if (target->obd_recovering && !export->exp_in_recovery) {
 895                 spin_lock(&export->exp_lock);
 896                 export->exp_in_recovery = 1;
 897                 export->exp_req_replay_needed = 1;
 898                 export->exp_lock_replay_needed = 1;
 899                 spin_unlock(&export->exp_lock);
 900                 if ((lustre_msg_get_op_flags(req->rq_reqmsg) & MSG_CONNECT_TRANSNO)
 901                      && (data->ocd_transno == 0))
 902                         CWARN("Connect with zero transno!\n");
 903
 904                 if ((lustre_msg_get_op_flags(req->rq_reqmsg) & MSG_CONNECT_TRANSNO)
 905                      && data->ocd_transno < target->obd_next_recovery_transno)
 906                         target->obd_next_recovery_transno = data->ocd_transno;
 907                 target->obd_connected_clients++;
 908                 /* each connected client is counted as recoverable */
 909                 target->obd_recoverable_clients++;
 910                 atomic_inc(&target->obd_req_replay_clients);
 911                 atomic_inc(&target->obd_lock_replay_clients);
 912                 if (target->obd_connected_clients ==
 913                     target->obd_max_recoverable_clients)
 914                         wake_up(&target->obd_next_transno_waitq);
 915         }
 916         spin_unlock_bh(&target->obd_processing_task_lock);
 917         tmp = req_capsule_client_get(&req->rq_pill, &RMF_CONN);
 918         conn = *tmp;
 919
 920         if (export->exp_imp_reverse != NULL) {
 921                 /* destroyed import can be still referenced in ctxt */
 922                 obd_set_info_async(export, strlen(KEY_REVIMP_UPD),
 923                                    KEY_REVIMP_UPD, 0, NULL, NULL);
 924
 925                 /* in some recovery senarios, previous ctx init rpc handled
 926                  * in sptlrpc_target_export_check() might be used to install
 927                  * a reverse ctx in this reverse import, and later OBD_CONNECT
 928                  * using the same gss ctx could reach here and following new
 929                  * reverse import. note all reverse ctx in new/old import are
 930                  * actually based on the same gss ctx. so we invalidate ctx
 931                  * here before destroy import, otherwise flush old import will
 932                  * lead to remote reverse ctx be destroied, thus the reverse
 933                  * ctx of new import will lost its peer.
 934                  * there might be a better way to deal with this???
 935                  */
 936                 sptlrpc_import_inval_all_ctx(export->exp_imp_reverse);
 937
 938                 destroy_import(export->exp_imp_reverse);
 939         }
 940
 941         /* for the rest part, we return -ENOTCONN in case of errors
 942          * in order to let client initialize connection again.
 943          */
 944         revimp = export->exp_imp_reverse = class_new_import(target);
 945         if (!revimp) {
 946                 CERROR("fail to alloc new reverse import.\n");
 947                 GOTO(out, rc = -ENOTCONN);
 948         }
 949
 950         revimp->imp_connection = ptlrpc_connection_addref(export->exp_connection);
 951         revimp->imp_client = &export->exp_obd->obd_ldlm_client;
 952         revimp->imp_remote_handle = conn;
 953         revimp->imp_dlm_fake = 1;
 954         revimp->imp_state = LUSTRE_IMP_FULL;
 955
 956         if (req->rq_reqmsg->lm_magic == LUSTRE_MSG_MAGIC_V1 &&
 957             lustre_msg_get_op_flags(req->rq_reqmsg) & MSG_CONNECT_NEXT_VER) {
 958                 revimp->imp_msg_magic = LUSTRE_MSG_MAGIC_V2;
 959                 lustre_msg_add_op_flags(req->rq_repmsg, MSG_CONNECT_NEXT_VER);
 960         } else {
 961                 /* unknown magic has been filtered out by unpack */
 962                 revimp->imp_msg_magic = req->rq_reqmsg->lm_magic;
 963         }
 964
 965         rc = sptlrpc_import_sec_adapt(revimp, req->rq_svc_ctx,
 966                                       req->rq_flvr.sf_rpc);
 967         if (rc) {
 968                 CERROR("Failed to get sec for reverse import: %d\n", rc);
 969                 export->exp_imp_reverse = NULL;
 970                 class_destroy_import(revimp);
 971         }
 972
 973         class_import_put(revimp);
 974 out:
 975         if (export) {
 976                 spin_lock(&export->exp_lock);
 977                 export->exp_connecting = 0;
 978                 spin_unlock(&export->exp_lock);
 979         }
 980         if (targref)
 981                 class_decref(targref);
 982         if (rc)
 983                 req->rq_status = rc;
 984         RETURN(rc);
 985 }
 986
 987 int target_handle_disconnect(struct ptlrpc_request *req)
 988 {
 989         int rc;
 990         ENTRY;
 991
 992         rc = req_capsule_server_pack(&req->rq_pill);
 993         if (rc)
 994                 RETURN(rc);
 995
 996         /* keep the rq_export around so we can send the reply */
 997         req->rq_status = obd_disconnect(class_export_get(req->rq_export));
 998
 999         RETURN(0);
1000 }
1001
1002 void target_destroy_export(struct obd_export *exp)
1003 {
1004         /* exports created from last_rcvd data, and "fake"
1005            exports created by lctl don't have an import */
1006         if (exp->exp_imp_reverse != NULL)
1007                 destroy_import(exp->exp_imp_reverse);
1008
1009         /* We cancel locks at disconnect time, but this will catch any locks
1010          * granted in a race with recovery-induced disconnect. */
1011         if (exp->exp_obd->obd_namespace != NULL)
1012                 ldlm_cancel_locks_for_export(exp);
1013 }
1014
1015 /*
1016  * Recovery functions
1017  */
1018
1019 struct ptlrpc_request *ptlrpc_clone_req( struct ptlrpc_request *orig_req)
1020 {
1021         struct ptlrpc_request *copy_req;
1022         struct lustre_msg *copy_reqmsg;
1023         struct ptlrpc_user_desc *udesc = NULL;
1024
1025         OBD_ALLOC_PTR(copy_req);
1026         if (!copy_req)
1027                 return NULL;
1028         OBD_ALLOC(copy_reqmsg, orig_req->rq_reqlen);
1029         if (!copy_reqmsg){
1030                 OBD_FREE_PTR(copy_req);
1031                 return NULL;
1032         }
1033
1034         if (orig_req->rq_user_desc) {
1035                 int ngroups = orig_req->rq_user_desc->pud_ngroups;
1036
1037                 OBD_ALLOC(udesc, sptlrpc_user_desc_size(ngroups));
1038                 if (!udesc) {
1039                         OBD_FREE(copy_reqmsg, orig_req->rq_reqlen);
1040                         OBD_FREE_PTR(copy_req);
1041                         return NULL;
1042                 }
1043                 memcpy(udesc, orig_req->rq_user_desc,
1044                        sptlrpc_user_desc_size(ngroups));
1045         }
1046
1047         *copy_req = *orig_req;
1048         memcpy(copy_reqmsg, orig_req->rq_reqmsg, orig_req->rq_reqlen);
1049         copy_req->rq_reqmsg = copy_reqmsg;
1050         copy_req->rq_user_desc = udesc;
1051
1052         class_export_get(copy_req->rq_export);
1053         CFS_INIT_LIST_HEAD(&copy_req->rq_list);
1054         sptlrpc_svc_ctx_addref(copy_req);
1055
1056         if (copy_req->rq_reply_state) {
1057                 /* the copied req takes over the reply state */
1058                 orig_req->rq_reply_state = NULL;
1059                 /* to catch further access */
1060                 orig_req->rq_repmsg = NULL;
1061                 orig_req->rq_replen = 0;
1062         }
1063
1064         return copy_req;
1065 }
1066
1067 void ptlrpc_free_clone( struct ptlrpc_request *req)
1068 {
1069         if (req->rq_reply_state) {
1070                 ptlrpc_rs_decref(req->rq_reply_state);
1071                 req->rq_reply_state = NULL;
1072         }
1073
1074         sptlrpc_svc_ctx_decref(req);
1075         class_export_put(req->rq_export);
1076         list_del(&req->rq_list);
1077
1078         if (req->rq_user_desc) {
1079                 int ngroups = req->rq_user_desc->pud_ngroups;
1080                 OBD_FREE(req->rq_user_desc, sptlrpc_user_desc_size(ngroups));
1081         }
1082         OBD_FREE(req->rq_reqmsg, req->rq_reqlen);
1083         OBD_FREE_PTR(req);
1084 }
1085
1086 #ifdef __KERNEL__
1087 static void target_finish_recovery(struct obd_device *obd)
1088 {
1089         ENTRY;
1090         LCONSOLE_INFO("%s: sending delayed replies to recovered clients\n",
1091                       obd->obd_name);
1092
1093         ldlm_reprocess_all_ns(obd->obd_namespace);
1094
1095         /* when recovery finished, cleanup orphans on mds and ost */
1096         if (OBT(obd) && OBP(obd, postrecov)) {
1097                 int rc = OBP(obd, postrecov)(obd);
1098                 LCONSOLE_WARN("%s: recovery %s: rc %d\n", obd->obd_name,
1099                               rc < 0 ? "failed" : "complete", rc);
1100         }
1101
1102         obd->obd_recovery_end = CURRENT_SECONDS;
1103         EXIT;
1104 }
1105
1106 static void abort_req_replay_queue(struct obd_device *obd)
1107 {
1108         struct ptlrpc_request *req, *n;
1109
1110         list_for_each_entry_safe(req, n, &obd->obd_req_replay_queue, rq_list) {
1111                 DEBUG_REQ(D_WARNING, req, "aborted:");
1112                 req->rq_status = -ENOTCONN;
1113                 if (ptlrpc_error(req)) {
1114                         DEBUG_REQ(D_ERROR, req,
1115                                   "failed abort_req_reply; skipping");
1116                 }
1117                 ptlrpc_free_clone(req);
1118         }
1119 }
1120
1121 static void abort_lock_replay_queue(struct obd_device *obd)
1122 {
1123         struct ptlrpc_request *req, *n;
1124
1125         list_for_each_entry_safe(req, n, &obd->obd_lock_replay_queue, rq_list){
1126                 DEBUG_REQ(D_ERROR, req, "aborted:");
1127                 req->rq_status = -ENOTCONN;
1128                 if (ptlrpc_error(req)) {
1129                         DEBUG_REQ(D_ERROR, req,
1130                                   "failed abort_lock_reply; skipping");
1131                 }
1132                 ptlrpc_free_clone(req);
1133         }
1134 }
1135 #endif
1136
1137 /* Called from a cleanup function if the device is being cleaned up
1138    forcefully.  The exports should all have been disconnected already,
1139    the only thing left to do is
1140      - clear the recovery flags
1141      - cancel the timer
1142      - free queued requests and replies, but don't send replies
1143    Because the obd_stopping flag is set, no new requests should be received.
1144
1145 */
1146 void target_cleanup_recovery(struct obd_device *obd)
1147 {
1148         struct ptlrpc_request *req, *n;
1149         ENTRY;
1150
1151         LASSERT(obd->obd_stopping);
1152
1153         spin_lock_bh(&obd->obd_processing_task_lock);
1154         if (!obd->obd_recovering) {
1155                 spin_unlock_bh(&obd->obd_processing_task_lock);
1156                 EXIT;
1157                 return;
1158         }
1159         obd->obd_recovering = obd->obd_abort_recovery = 0;
1160         target_cancel_recovery_timer(obd);
1161         spin_unlock_bh(&obd->obd_processing_task_lock);
1162
1163         list_for_each_entry_safe(req, n, &obd->obd_req_replay_queue, rq_list) {
1164                 LASSERT (req->rq_reply_state == 0);
1165                 ptlrpc_free_clone(req);
1166         }
1167         list_for_each_entry_safe(req, n, &obd->obd_lock_replay_queue, rq_list){
1168                 LASSERT (req->rq_reply_state == 0);
1169                 ptlrpc_free_clone(req);
1170         }
1171         list_for_each_entry_safe(req, n, &obd->obd_final_req_queue, rq_list) {
1172                 LASSERT (req->rq_reply_state == 0);
1173                 ptlrpc_free_clone(req);
1174         }
1175
1176         EXIT;
1177 }
1178
1179 static void target_recovery_expired(unsigned long castmeharder)
1180 {
1181         struct obd_device *obd = (struct obd_device *)castmeharder;
1182         CERROR("%s: recovery timed out, aborting\n", obd->obd_name);
1183         spin_lock_bh(&obd->obd_processing_task_lock);
1184         if (obd->obd_recovering)
1185                 obd->obd_abort_recovery = 1;
1186         cfs_waitq_signal(&obd->obd_next_transno_waitq);
1187         spin_unlock_bh(&obd->obd_processing_task_lock);
1188 }
1189
1190
1191 /* obd_processing_task_lock should be held */
1192 void target_cancel_recovery_timer(struct obd_device *obd)
1193 {
1194         CDEBUG(D_HA, "%s: cancel recovery timer\n", obd->obd_name);
1195         cfs_timer_disarm(&obd->obd_recovery_timer);
1196 }
1197
1198 static void reset_recovery_timer(struct obd_device *obd)
1199 {
1200         time_t timeout_shift = OBD_RECOVERY_TIMEOUT;
1201         spin_lock_bh(&obd->obd_processing_task_lock);
1202         if (!obd->obd_recovering) {
1203                 spin_unlock_bh(&obd->obd_processing_task_lock);
1204                 return;
1205         }
1206         if (cfs_time_current_sec() + OBD_RECOVERY_TIMEOUT >
1207             obd->obd_recovery_start + obd->obd_recovery_max_time)
1208                 timeout_shift = obd->obd_recovery_start +
1209                         obd->obd_recovery_max_time - cfs_time_current_sec();
1210         cfs_timer_arm(&obd->obd_recovery_timer, cfs_time_shift(timeout_shift));
1211         spin_unlock_bh(&obd->obd_processing_task_lock);
1212         CDEBUG(D_HA, "%s: timer will expire in %u seconds\n", obd->obd_name,
1213                (unsigned int)timeout_shift);
1214         /* Only used for lprocfs_status */
1215         obd->obd_recovery_end = CURRENT_SECONDS + timeout_shift;
1216 }
1217
1218
1219 /* Only start it the first time called */
1220 void target_start_recovery_timer(struct obd_device *obd)
1221 {
1222         spin_lock_bh(&obd->obd_processing_task_lock);
1223         if (obd->obd_recovery_handler
1224             || timer_pending((struct timer_list *)&obd->obd_recovery_timer)) {
1225                 spin_unlock_bh(&obd->obd_processing_task_lock);
1226                 return;
1227         }
1228         CWARN("%s: starting recovery timer (%us)\n", obd->obd_name,
1229               OBD_RECOVERY_TIMEOUT);
1230         cfs_timer_init(&obd->obd_recovery_timer, target_recovery_expired, obd);
1231         spin_unlock_bh(&obd->obd_processing_task_lock);
1232
1233         reset_recovery_timer(obd);
1234 }
1235
1236 #ifdef __KERNEL__
1237 static int check_for_next_transno(struct obd_device *obd)
1238 {
1239         struct ptlrpc_request *req = NULL;
1240         int wake_up = 0, connected, completed, queue_len, max;
1241         __u64 next_transno, req_transno;
1242         ENTRY;
1243         spin_lock_bh(&obd->obd_processing_task_lock);
1244
1245         if (!list_empty(&obd->obd_req_replay_queue)) {
1246                 req = list_entry(obd->obd_req_replay_queue.next,
1247                                  struct ptlrpc_request, rq_list);
1248                 req_transno = lustre_msg_get_transno(req->rq_reqmsg);
1249         } else {
1250                 req_transno = 0;
1251         }
1252
1253         max = obd->obd_max_recoverable_clients;
1254         connected = obd->obd_connected_clients;
1255         completed = connected - obd->obd_recoverable_clients;
1256         queue_len = obd->obd_requests_queued_for_recovery;
1257         next_transno = obd->obd_next_recovery_transno;
1258
1259         CDEBUG(D_HA, "max: %d, connected: %d, completed: %d, queue_len: %d, "
1260                "req_transno: "LPU64", next_transno: "LPU64"\n",
1261                max, connected, completed, queue_len, req_transno, next_transno);
1262
1263         if (obd->obd_abort_recovery) {
1264                 CDEBUG(D_HA, "waking for aborted recovery\n");
1265                 wake_up = 1;
1266         } else if (atomic_read(&obd->obd_req_replay_clients) == 0) {
1267                 CDEBUG(D_HA, "waking for completed recovery\n");
1268                 wake_up = 1;
1269         } else if (req_transno == next_transno) {
1270                 CDEBUG(D_HA, "waking for next ("LPD64")\n", next_transno);
1271                 wake_up = 1;
1272         } else if (queue_len + completed == max) {
1273                 /* handle gaps occured due to lost reply. It is allowed gaps
1274                  * because all clients are connected and there will be resend
1275                  * for missed transaction */
1276                 LASSERTF(req_transno >= next_transno,
1277                          "req_transno: "LPU64", next_transno: "LPU64"\n",
1278                          req_transno, next_transno);
1279
1280                 CDEBUG(req_transno > obd->obd_last_committed ? D_ERROR : D_HA,
1281                        "waking for skipped transno (skip: "LPD64
1282                        ", ql: %d, comp: %d, conn: %d, next: "LPD64")\n",
1283                        next_transno, queue_len, completed, connected, req_transno);
1284                 obd->obd_next_recovery_transno = req_transno;
1285                 wake_up = 1;
1286         } else if (queue_len == atomic_read(&obd->obd_req_replay_clients)) {
1287                 /* some clients haven't connected in time, but we can try
1288                  * to replay requests that demand on already committed ones
1289                  * also, we can replay first non-committed transation */
1290                 LASSERT(req_transno != 0);
1291                 if (req_transno == obd->obd_last_committed + 1) {
1292                         obd->obd_next_recovery_transno = req_transno;
1293                 } else if (req_transno > obd->obd_last_committed) {
1294                         /* can't continue recovery: have no needed transno */
1295                         obd->obd_abort_recovery = 1;
1296                         CDEBUG(D_ERROR, "abort due to missed clients. max: %d, "
1297                                "connected: %d, completed: %d, queue_len: %d, "
1298                                "req_transno: "LPU64", next_transno: "LPU64"\n",
1299                                max, connected, completed, queue_len,
1300                                req_transno, next_transno);
1301                 }
1302                 wake_up = 1;
1303         }
1304
1305         spin_unlock_bh(&obd->obd_processing_task_lock);
1306         return wake_up;
1307 }
1308
1309 static struct ptlrpc_request *target_next_replay_req(struct obd_device *obd)
1310 {
1311         struct l_wait_info lwi = { 0 };
1312         struct ptlrpc_request *req;
1313
1314         CDEBUG(D_HA, "Waiting for transno "LPD64"\n",
1315                obd->obd_next_recovery_transno);
1316         l_wait_event(obd->obd_next_transno_waitq,
1317                      check_for_next_transno(obd), &lwi);
1318
1319         spin_lock_bh(&obd->obd_processing_task_lock);
1320         if (obd->obd_abort_recovery) {
1321                 req = NULL;
1322         } else if (!list_empty(&obd->obd_req_replay_queue)) {
1323                 req = list_entry(obd->obd_req_replay_queue.next,
1324                                  struct ptlrpc_request, rq_list);
1325                 list_del_init(&req->rq_list);
1326                 obd->obd_requests_queued_for_recovery--;
1327         } else {
1328                 req = NULL;
1329         }
1330         spin_unlock_bh(&obd->obd_processing_task_lock);
1331         RETURN(req);
1332 }
1333
1334 static int check_for_next_lock(struct obd_device *obd)
1335 {
1336         struct ptlrpc_request *req = NULL;
1337         int wake_up = 0;
1338
1339         spin_lock_bh(&obd->obd_processing_task_lock);
1340         if (!list_empty(&obd->obd_lock_replay_queue)) {
1341                 req = list_entry(obd->obd_lock_replay_queue.next,
1342                                  struct ptlrpc_request, rq_list);
1343                 CDEBUG(D_HA, "waking for next lock\n");
1344                 wake_up = 1;
1345         } else if (atomic_read(&obd->obd_lock_replay_clients) == 0) {
1346                 CDEBUG(D_HA, "waking for completed lock replay\n");
1347                 wake_up = 1;
1348         } else if (obd->obd_abort_recovery) {
1349                 CDEBUG(D_HA, "waking for aborted recovery\n");
1350                 wake_up = 1;
1351         }
1352         spin_unlock_bh(&obd->obd_processing_task_lock);
1353
1354         return wake_up;
1355 }
1356
1357 static struct ptlrpc_request *target_next_replay_lock(struct obd_device *obd)
1358 {
1359         struct l_wait_info lwi = { 0 };
1360         struct ptlrpc_request *req;
1361
1362         CDEBUG(D_HA, "Waiting for lock\n");
1363         l_wait_event(obd->obd_next_transno_waitq,
1364                      check_for_next_lock(obd), &lwi);
1365
1366         spin_lock_bh(&obd->obd_processing_task_lock);
1367         if (obd->obd_abort_recovery) {
1368                 req = NULL;
1369         } else if (!list_empty(&obd->obd_lock_replay_queue)) {
1370                 req = list_entry(obd->obd_lock_replay_queue.next,
1371                                  struct ptlrpc_request, rq_list);
1372                 list_del_init(&req->rq_list);
1373         } else {
1374                 req = NULL;
1375         }
1376         spin_unlock_bh(&obd->obd_processing_task_lock);
1377         return req;
1378 }
1379
1380 static struct ptlrpc_request *target_next_final_ping(struct obd_device *obd)
1381 {
1382         struct ptlrpc_request *req;
1383
1384         spin_lock_bh(&obd->obd_processing_task_lock);
1385         if (!list_empty(&obd->obd_final_req_queue)) {
1386                 req = list_entry(obd->obd_final_req_queue.next,
1387                                  struct ptlrpc_request, rq_list);
1388                 list_del_init(&req->rq_list);
1389         } else {
1390                 req = NULL;
1391         }
1392         spin_unlock_bh(&obd->obd_processing_task_lock);
1393         return req;
1394 }
1395
1396 static inline int req_replay_done(struct obd_export *exp)
1397 {
1398         return (exp->exp_req_replay_needed == 0);
1399 }
1400
1401 static inline int lock_replay_done(struct obd_export *exp)
1402 {
1403         return (exp->exp_lock_replay_needed == 0);
1404 }
1405
1406 static inline int connect_done(struct obd_export *exp)
1407 {
1408         return (exp->exp_in_recovery != 0);
1409 }
1410
1411 static int check_for_clients(struct obd_device *obd)
1412 {
1413         if (obd->obd_abort_recovery)
1414                 return 1;
1415         LASSERT(obd->obd_connected_clients <= obd->obd_max_recoverable_clients);
1416         if (obd->obd_no_conn == 0 &&
1417             obd->obd_connected_clients == obd->obd_max_recoverable_clients)
1418                 return 1;
1419         return 0;
1420 }
1421
1422 static int handle_recovery_req(struct ptlrpc_thread *thread,
1423                                struct ptlrpc_request *req,
1424                                svc_handler_t handler)
1425 {
1426         int rc;
1427         ENTRY;
1428
1429         rc = lu_context_init(&req->rq_session, LCT_SESSION);
1430         if (rc) {
1431                 CERROR("Failure to initialize session: %d\n", rc);
1432                 return rc;
1433         }
1434         req->rq_session.lc_thread = thread;
1435         lu_context_enter(&req->rq_session);
1436         req->rq_svc_thread = thread;
1437         req->rq_svc_thread->t_env->le_ses = &req->rq_session;
1438
1439         /* thread context */
1440         lu_context_enter(&thread->t_env->le_ctx);
1441         (void)handler(req);
1442         lu_context_exit(&thread->t_env->le_ctx);
1443
1444         lu_context_exit(&req->rq_session);
1445         lu_context_fini(&req->rq_session);
1446         /* don't reset timer for final stage */
1447         if (!req_replay_done(req->rq_export) ||
1448             !lock_replay_done(req->rq_export))
1449                 reset_recovery_timer(class_exp2obd(req->rq_export));
1450         ptlrpc_free_clone(req);
1451         RETURN(0);
1452 }
1453
1454 static int target_recovery_thread(void *arg)
1455 {
1456         struct obd_device *obd = arg;
1457         struct ptlrpc_request *req;
1458         struct target_recovery_data *trd = &obd->obd_recovery_data;
1459         struct l_wait_info lwi = { 0 };
1460         unsigned long delta;
1461         unsigned long flags;
1462         struct lu_env env;
1463         struct ptlrpc_thread fake_svc_thread, *thread = &fake_svc_thread;
1464         __u32 recov_ctx_tags = LCT_MD_THREAD;
1465         int rc = 0;
1466         ENTRY;
1467
1468         cfs_daemonize("tgt_recov");
1469
1470         SIGNAL_MASK_LOCK(current, flags);
1471         sigfillset(&current->blocked);
1472         RECALC_SIGPENDING;
1473         SIGNAL_MASK_UNLOCK(current, flags);
1474
1475         rc = lu_context_init(&env.le_ctx, recov_ctx_tags);
1476         if (rc)
1477                 RETURN(rc);
1478
1479         thread->t_env = &env;
1480         env.le_ctx.lc_thread = thread;
1481
1482         CERROR("%s: started recovery thread pid %d\n", obd->obd_name,
1483                current->pid);
1484         trd->trd_processing_task = current->pid;
1485
1486         obd->obd_recovering = 1;
1487         complete(&trd->trd_starting);
1488
1489         /* first of all, we have to know the first transno to replay */
1490         obd->obd_abort_recovery = 0;
1491         l_wait_event(obd->obd_next_transno_waitq,
1492                      check_for_clients(obd), &lwi);
1493
1494         spin_lock_bh(&obd->obd_processing_task_lock);
1495         target_cancel_recovery_timer(obd);
1496         spin_unlock_bh(&obd->obd_processing_task_lock);
1497
1498         /* If some clients haven't connected in time, evict them */
1499         if (obd->obd_abort_recovery) {
1500                 CWARN("Some clients haven't connect in time (%d/%d),"
1501                        "evict them\n", obd->obd_connected_clients,
1502                        obd->obd_max_recoverable_clients);
1503                 obd->obd_abort_recovery = obd->obd_stopping;
1504                 class_disconnect_stale_exports(obd, connect_done);
1505         }
1506         /* next stage: replay requests */
1507         delta = jiffies;
1508         obd->obd_req_replaying = 1;
1509         CDEBUG(D_INFO, "1: request replay stage - %d clients from t"LPU64"\n",
1510               atomic_read(&obd->obd_req_replay_clients),
1511               obd->obd_next_recovery_transno);
1512         while ((req = target_next_replay_req(obd))) {
1513                 LASSERT(trd->trd_processing_task == current->pid);
1514                 DEBUG_REQ(D_HA, req, "processing t"LPD64" from %s",
1515                           lustre_msg_get_transno(req->rq_reqmsg),
1516                           libcfs_nid2str(req->rq_peer.nid));
1517                 handle_recovery_req(thread, req,
1518                                     trd->trd_recovery_handler);
1519                 obd->obd_replayed_requests++;
1520                 spin_lock_bh(&obd->obd_processing_task_lock);
1521                 obd->obd_next_recovery_transno++;
1522                 spin_unlock_bh(&obd->obd_processing_task_lock);
1523         }
1524
1525         spin_lock_bh(&obd->obd_processing_task_lock);
1526         target_cancel_recovery_timer(obd);
1527         spin_unlock_bh(&obd->obd_processing_task_lock);
1528
1529         /* If some clients haven't replayed requests in time, evict them */
1530         if (obd->obd_abort_recovery) {
1531                 CDEBUG(D_ERROR, "req replay timed out, aborting ...\n");
1532                 obd->obd_abort_recovery = obd->obd_stopping;
1533                 class_disconnect_stale_exports(obd, req_replay_done);
1534                 abort_req_replay_queue(obd);
1535         }
1536         /* The second stage: replay locks */
1537         CDEBUG(D_INFO, "2: lock replay stage - %d clients\n",
1538                atomic_read(&obd->obd_lock_replay_clients));
1539         while ((req = target_next_replay_lock(obd))) {
1540                 LASSERT(trd->trd_processing_task == current->pid);
1541                 DEBUG_REQ(D_HA|D_WARNING, req, "processing lock from %s: ",
1542                           libcfs_nid2str(req->rq_peer.nid));
1543                 handle_recovery_req(thread, req,
1544                                     trd->trd_recovery_handler);
1545                 obd->obd_replayed_locks++;
1546         }
1547
1548         spin_lock_bh(&obd->obd_processing_task_lock);
1549         target_cancel_recovery_timer(obd);
1550         spin_unlock_bh(&obd->obd_processing_task_lock);
1551         /* If some clients haven't replayed requests in time, evict them */
1552         if (obd->obd_abort_recovery) {
1553                 int stale;
1554                 CERROR("lock replay timed out, aborting ...\n");
1555                 obd->obd_abort_recovery = obd->obd_stopping;
1556                 stale = class_disconnect_stale_exports(obd, lock_replay_done);
1557                 abort_lock_replay_queue(obd);
1558         }
1559
1560         /* We drop recoverying flag to forward all new requests
1561          * to regular mds_handle() since now */
1562         spin_lock_bh(&obd->obd_processing_task_lock);
1563         obd->obd_recovering = obd->obd_abort_recovery = 0;
1564         spin_unlock_bh(&obd->obd_processing_task_lock);
1565         /* The third stage: reply on final pings */
1566         CDEBUG(D_INFO, "3: final stage - process recovery completion pings\n");
1567         while ((req = target_next_final_ping(obd))) {
1568                 LASSERT(trd->trd_processing_task == current->pid);
1569                 DEBUG_REQ(D_HA, req, "processing final ping from %s: ",
1570                           libcfs_nid2str(req->rq_peer.nid));
1571                 handle_recovery_req(thread, req,
1572                                     trd->trd_recovery_handler);
1573         }
1574
1575         delta = (jiffies - delta) / HZ;
1576         CDEBUG(D_INFO,"4: recovery completed in %lus - %d/%d reqs/locks\n",
1577               delta, obd->obd_replayed_requests, obd->obd_replayed_locks);
1578         LASSERT(atomic_read(&obd->obd_req_replay_clients) == 0);
1579         LASSERT(atomic_read(&obd->obd_lock_replay_clients) == 0);
1580         if (delta > obd_timeout * 2) {
1581                 CWARN("too long recovery - read logs\n");
1582                 libcfs_debug_dumplog();
1583         }
1584
1585         target_finish_recovery(obd);
1586
1587         lu_context_fini(&env.le_ctx);
1588         trd->trd_processing_task = 0;
1589         complete(&trd->trd_finishing);
1590         RETURN(rc);
1591 }
1592
1593 int target_start_recovery_thread(struct obd_device *obd, svc_handler_t handler)
1594 {
1595         int rc = 0;
1596         struct target_recovery_data *trd = &obd->obd_recovery_data;
1597
1598         memset(trd, 0, sizeof(*trd));
1599         init_completion(&trd->trd_starting);
1600         init_completion(&trd->trd_finishing);
1601         trd->trd_recovery_handler = handler;
1602
1603         if (kernel_thread(target_recovery_thread, obd, 0) > 0) {
1604                 wait_for_completion(&trd->trd_starting);
1605                 LASSERT(obd->obd_recovering != 0);
1606         } else
1607                 rc = -ECHILD;
1608
1609         return rc;
1610 }
1611
1612 void target_stop_recovery_thread(struct obd_device *obd)
1613 {
1614         spin_lock_bh(&obd->obd_processing_task_lock);
1615         if (obd->obd_recovery_data.trd_processing_task > 0) {
1616                 struct target_recovery_data *trd = &obd->obd_recovery_data;
1617                 CERROR("%s: Aborting recovery\n", obd->obd_name);
1618                 obd->obd_abort_recovery = 1;
1619                 wake_up(&obd->obd_next_transno_waitq);
1620                 spin_unlock_bh(&obd->obd_processing_task_lock);
1621                 wait_for_completion(&trd->trd_finishing);
1622         } else {
1623                 spin_unlock_bh(&obd->obd_processing_task_lock);
1624         }
1625 }
1626
1627 void target_recovery_fini(struct obd_device *obd)
1628 {
1629         class_disconnect_exports(obd);
1630         target_stop_recovery_thread(obd);
1631         target_cleanup_recovery(obd);
1632 }
1633 EXPORT_SYMBOL(target_recovery_fini);
1634
1635 void target_recovery_init(struct obd_device *obd, svc_handler_t handler)
1636 {
1637         if (obd->obd_max_recoverable_clients == 0)
1638                 return;
1639
1640         CWARN("RECOVERY: service %s, %d recoverable clients, "
1641               "last_transno "LPU64"\n", obd->obd_name,
1642               obd->obd_max_recoverable_clients, obd->obd_last_committed);
1643         obd->obd_next_recovery_transno = obd->obd_last_committed + 1;
1644         target_start_recovery_thread(obd, handler);
1645         obd->obd_recovery_start = CURRENT_SECONDS;
1646         /* Only used for lprocfs_status */
1647         obd->obd_recovery_end = obd->obd_recovery_start + OBD_RECOVERY_TIMEOUT;
1648         /* bz13079: this should be set to desired value for ost but not for mds */
1649         obd->obd_recovery_max_time = OBD_RECOVERY_MAX_TIME;
1650 }
1651 EXPORT_SYMBOL(target_recovery_init);
1652
1653 #endif
1654
1655 int target_process_req_flags(struct obd_device *obd, struct ptlrpc_request *req)
1656 {
1657         struct obd_export *exp = req->rq_export;
1658         LASSERT(exp != NULL);
1659         if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REQ_REPLAY_DONE) {
1660                 /* client declares he's ready to replay locks */
1661                 spin_lock_bh(&obd->obd_processing_task_lock);
1662                 if (exp->exp_req_replay_needed) {
1663                         LASSERT(atomic_read(&obd->obd_req_replay_clients) > 0);
1664                         spin_lock(&exp->exp_lock);
1665                         exp->exp_req_replay_needed = 0;
1666                         spin_unlock(&exp->exp_lock);
1667                         atomic_dec(&obd->obd_req_replay_clients);
1668                         LASSERT(obd->obd_recoverable_clients > 0);
1669                         obd->obd_recoverable_clients--;
1670                         if (atomic_read(&obd->obd_req_replay_clients) == 0)
1671                                 CDEBUG(D_HA, "all clients have replayed reqs\n");
1672                         wake_up(&obd->obd_next_transno_waitq);
1673                 }
1674                 spin_unlock_bh(&obd->obd_processing_task_lock);
1675         }
1676         if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_LOCK_REPLAY_DONE) {
1677                 /* client declares he's ready to complete recovery
1678                  * so, we put the request on th final queue */
1679                 spin_lock_bh(&obd->obd_processing_task_lock);
1680                 if (exp->exp_lock_replay_needed) {
1681                         LASSERT(atomic_read(&obd->obd_lock_replay_clients) > 0);
1682                         spin_lock(&exp->exp_lock);
1683                         exp->exp_lock_replay_needed = 0;
1684                         spin_unlock(&exp->exp_lock);
1685                         atomic_dec(&obd->obd_lock_replay_clients);
1686                         if (atomic_read(&obd->obd_lock_replay_clients) == 0)
1687                                 CDEBUG(D_HA, "all clients have replayed locks\n");
1688                         wake_up(&obd->obd_next_transno_waitq);
1689                 }
1690                 spin_unlock_bh(&obd->obd_processing_task_lock);
1691         }
1692
1693         return 0;
1694 }
1695
1696 int target_queue_recovery_request(struct ptlrpc_request *req,
1697                                   struct obd_device *obd)
1698 {
1699         struct list_head *tmp;
1700         int inserted = 0;
1701         __u64 transno = lustre_msg_get_transno(req->rq_reqmsg);
1702
1703         ENTRY;
1704
1705         if (obd->obd_recovery_data.trd_processing_task == cfs_curproc_pid()) {
1706                 /* Processing the queue right now, don't re-add. */
1707                 RETURN(1);
1708         }
1709
1710         target_process_req_flags(obd, req);
1711
1712         if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_LOCK_REPLAY_DONE) {
1713                 /* client declares he's ready to complete recovery
1714                  * so, we put the request on th final queue */
1715                 req = ptlrpc_clone_req(req);
1716                 if (req == NULL)
1717                         RETURN(-ENOMEM);
1718                 DEBUG_REQ(D_HA, req, "queue final req");
1719                 spin_lock_bh(&obd->obd_processing_task_lock);
1720                 if (obd->obd_recovering)
1721                         list_add_tail(&req->rq_list, &obd->obd_final_req_queue);
1722                 else {
1723                         spin_unlock_bh(&obd->obd_processing_task_lock);
1724                         ptlrpc_free_clone(req);
1725                         if (obd->obd_stopping) {
1726                                 RETURN(-ENOTCONN);
1727                         } else {
1728                                 RETURN(1);
1729                         }
1730                 }
1731                 spin_unlock_bh(&obd->obd_processing_task_lock);
1732                 RETURN(0);
1733         }
1734         if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REQ_REPLAY_DONE) {
1735                 /* client declares he's ready to replay locks */
1736                 req = ptlrpc_clone_req(req);
1737                 if (req == NULL)
1738                         RETURN(-ENOMEM);
1739                 DEBUG_REQ(D_HA, req, "queue lock replay req");
1740                 spin_lock_bh(&obd->obd_processing_task_lock);
1741                 LASSERT(obd->obd_recovering);
1742                 /* usually due to recovery abort */
1743                 if (!req->rq_export->exp_in_recovery) {
1744                         spin_unlock_bh(&obd->obd_processing_task_lock);
1745                         ptlrpc_free_clone(req);
1746                         RETURN(-ENOTCONN);
1747                 }
1748                 LASSERT(req->rq_export->exp_lock_replay_needed);
1749                 list_add_tail(&req->rq_list, &obd->obd_lock_replay_queue);
1750                 spin_unlock_bh(&obd->obd_processing_task_lock);
1751                 wake_up(&obd->obd_next_transno_waitq);
1752                 RETURN(0);
1753         }
1754
1755         /* CAVEAT EMPTOR: The incoming request message has been swabbed
1756          * (i.e. buflens etc are in my own byte order), but type-dependent
1757          * buffers (eg mds_body, ost_body etc) have NOT been swabbed. */
1758
1759         if (!transno) {
1760                 CFS_INIT_LIST_HEAD(&req->rq_list);
1761                 DEBUG_REQ(D_HA, req, "not queueing");
1762                 RETURN(1);
1763         }
1764
1765         spin_lock_bh(&obd->obd_processing_task_lock);
1766
1767         /* If we're processing the queue, we want don't want to queue this
1768          * message.
1769          *
1770          * Also, if this request has a transno less than the one we're waiting
1771          * for, we should process it now.  It could (and currently always will)
1772          * be an open request for a descriptor that was opened some time ago.
1773          *
1774          * Also, a resent, replayed request that has already been
1775          * handled will pass through here and be processed immediately.
1776          */
1777         CWARN("Next recovery transno: "LPU64", current: "LPU64", replaying: %i\n",
1778               obd->obd_next_recovery_transno, transno, obd->obd_req_replaying);
1779         if (transno < obd->obd_next_recovery_transno && obd->obd_req_replaying) {
1780                 /* Processing the queue right now, don't re-add. */
1781                 LASSERT(list_empty(&req->rq_list));
1782                 spin_unlock_bh(&obd->obd_processing_task_lock);
1783                 RETURN(1);
1784         }
1785         spin_unlock_bh(&obd->obd_processing_task_lock);
1786
1787         /* A resent, replayed request that is still on the queue; just drop it.
1788            The queued request will handle this. */
1789         if ((lustre_msg_get_flags(req->rq_reqmsg) & (MSG_RESENT|MSG_REPLAY)) ==
1790             (MSG_RESENT | MSG_REPLAY)) {
1791                 DEBUG_REQ(D_ERROR, req, "dropping resent queued req");
1792                 RETURN(0);
1793         }
1794
1795         req = ptlrpc_clone_req(req);
1796         if (req == NULL)
1797                 RETURN(-ENOMEM);
1798
1799         spin_lock_bh(&obd->obd_processing_task_lock);
1800         LASSERT(obd->obd_recovering);
1801         if (!req->rq_export->exp_in_recovery) {
1802                 spin_unlock_bh(&obd->obd_processing_task_lock);
1803                 ptlrpc_free_clone(req);
1804                 RETURN(-ENOTCONN);
1805         }
1806         LASSERT(req->rq_export->exp_req_replay_needed);
1807
1808         /* XXX O(n^2) */
1809         list_for_each(tmp, &obd->obd_req_replay_queue) {
1810                 struct ptlrpc_request *reqiter =
1811                         list_entry(tmp, struct ptlrpc_request, rq_list);
1812
1813                 if (lustre_msg_get_transno(reqiter->rq_reqmsg) > transno) {
1814                         list_add_tail(&req->rq_list, &reqiter->rq_list);
1815                         inserted = 1;
1816                         break;
1817                 }
1818         }
1819
1820         if (!inserted)
1821                 list_add_tail(&req->rq_list, &obd->obd_req_replay_queue);
1822
1823         obd->obd_requests_queued_for_recovery++;
1824         wake_up(&obd->obd_next_transno_waitq);
1825         spin_unlock_bh(&obd->obd_processing_task_lock);
1826         RETURN(0);
1827
1828 }
1829
1830 struct obd_device * target_req2obd(struct ptlrpc_request *req)
1831 {
1832         return req->rq_export->exp_obd;
1833 }
1834
1835 static inline struct ldlm_pool *ldlm_exp2pl(struct obd_export *exp)
1836 {
1837         LASSERT(exp != NULL);
1838         return &exp->exp_obd->obd_namespace->ns_pool;
1839 }
1840
1841 int target_pack_pool_reply(struct ptlrpc_request *req)
1842 {
1843         struct ldlm_pool *pl;
1844         ENTRY;
1845
1846         if (!req->rq_export || !req->rq_export->exp_obd ||
1847             !exp_connect_lru_resize(req->rq_export)) {
1848                 lustre_msg_set_slv(req->rq_repmsg, 0);
1849                 lustre_msg_set_limit(req->rq_repmsg, 0);
1850                 RETURN(0);
1851         }
1852
1853         pl = ldlm_exp2pl(req->rq_export);
1854
1855         spin_lock(&pl->pl_lock);
1856         LASSERT(ldlm_pool_get_slv(pl) != 0 && ldlm_pool_get_limit(pl) != 0);
1857         lustre_msg_set_slv(req->rq_repmsg, ldlm_pool_get_slv(pl));
1858         lustre_msg_set_limit(req->rq_repmsg, ldlm_pool_get_limit(pl));
1859         spin_unlock(&pl->pl_lock);
1860
1861         RETURN(0);
1862 }
1863
1864 int target_send_reply_msg(struct ptlrpc_request *req, int rc, int fail_id)
1865 {
1866         if (OBD_FAIL_CHECK_ORSET(fail_id & ~OBD_FAIL_ONCE, OBD_FAIL_ONCE)) {
1867                 DEBUG_REQ(D_ERROR, req, "dropping reply");
1868                 return (-ECOMM);
1869         }
1870
1871         if (unlikely(rc)) {
1872                 DEBUG_REQ(D_ERROR, req, "processing error (%d)", rc);
1873                 req->rq_status = rc;
1874                 return (ptlrpc_error(req));
1875         } else {
1876                 DEBUG_REQ(D_NET, req, "sending reply");
1877         }
1878
1879         return (ptlrpc_send_reply(req, 1));
1880 }
1881
1882 void target_send_reply(struct ptlrpc_request *req, int rc, int fail_id)
1883 {
1884         int                        netrc;
1885         struct ptlrpc_reply_state *rs;
1886         struct obd_device         *obd;
1887         struct obd_export         *exp;
1888         struct ptlrpc_service     *svc;
1889
1890         if (req->rq_no_reply)
1891                 return;
1892
1893         svc = req->rq_rqbd->rqbd_service;
1894         rs = req->rq_reply_state;
1895         if (rs == NULL || !rs->rs_difficult) {
1896                 /* no notifiers */
1897                 target_send_reply_msg (req, rc, fail_id);
1898                 return;
1899         }
1900
1901         /* must be an export if locks saved */
1902         LASSERT (req->rq_export != NULL);
1903         /* req/reply consistent */
1904         LASSERT (rs->rs_service == svc);
1905
1906         /* "fresh" reply */
1907         LASSERT (!rs->rs_scheduled);
1908         LASSERT (!rs->rs_scheduled_ever);
1909         LASSERT (!rs->rs_handled);
1910         LASSERT (!rs->rs_on_net);
1911         LASSERT (rs->rs_export == NULL);
1912         LASSERT (list_empty(&rs->rs_obd_list));
1913         LASSERT (list_empty(&rs->rs_exp_list));
1914
1915         exp = class_export_get (req->rq_export);
1916         obd = exp->exp_obd;
1917
1918         /* disable reply scheduling onto srv_reply_queue while I'm setting up */
1919         rs->rs_scheduled = 1;
1920         rs->rs_on_net    = 1;
1921         rs->rs_xid       = req->rq_xid;
1922         rs->rs_transno   = req->rq_transno;
1923         rs->rs_export    = exp;
1924
1925         spin_lock(&obd->obd_uncommitted_replies_lock);
1926
1927         if (rs->rs_transno > obd->obd_last_committed) {
1928                 /* not committed already */
1929                 list_add_tail (&rs->rs_obd_list,
1930                                &obd->obd_uncommitted_replies);
1931         }
1932
1933         spin_unlock (&obd->obd_uncommitted_replies_lock);
1934         spin_lock (&exp->exp_lock);
1935
1936         list_add_tail (&rs->rs_exp_list, &exp->exp_outstanding_replies);
1937
1938         spin_unlock(&exp->exp_lock);
1939
1940         netrc = target_send_reply_msg (req, rc, fail_id);
1941
1942         spin_lock(&svc->srv_lock);
1943
1944         svc->srv_n_difficult_replies++;
1945
1946         if (netrc != 0) {
1947                 /* error sending: reply is off the net.  Also we need +1
1948                  * reply ref until ptlrpc_server_handle_reply() is done
1949                  * with the reply state (if the send was successful, there
1950                  * would have been +1 ref for the net, which
1951                  * reply_out_callback leaves alone) */
1952                 rs->rs_on_net = 0;
1953                 ptlrpc_rs_addref(rs);
1954                 atomic_inc (&svc->srv_outstanding_replies);
1955         }
1956
1957         if (!rs->rs_on_net ||                   /* some notifier */
1958             list_empty(&rs->rs_exp_list) ||     /* completed already */
1959             list_empty(&rs->rs_obd_list)) {
1960                 list_add_tail (&rs->rs_list, &svc->srv_reply_queue);
1961                 cfs_waitq_signal (&svc->srv_waitq);
1962         } else {
1963                 list_add (&rs->rs_list, &svc->srv_active_replies);
1964                 rs->rs_scheduled = 0;           /* allow notifier to schedule */
1965         }
1966
1967         spin_unlock(&svc->srv_lock);
1968 }
1969
1970 int target_handle_ping(struct ptlrpc_request *req)
1971 {
1972         obd_ping(req->rq_export);
1973         return req_capsule_server_pack(&req->rq_pill);
1974 }
1975
1976 void target_committed_to_req(struct ptlrpc_request *req)
1977 {
1978         struct obd_device *obd;
1979
1980         if (req == NULL || req->rq_export == NULL)
1981                 return;
1982
1983         obd = req->rq_export->exp_obd;
1984         if (obd == NULL)
1985                 return;
1986
1987         if (!obd->obd_no_transno && req->rq_repmsg != NULL)
1988                 lustre_msg_set_last_committed(req->rq_repmsg,
1989                                               obd->obd_last_committed);
1990         else
1991                 DEBUG_REQ(D_IOCTL, req, "not sending last_committed update");
1992
1993         CDEBUG(D_INFO, "last_committed "LPU64", transno "LPU64", xid "LPU64"\n",
1994                obd->obd_last_committed, req->rq_transno, req->rq_xid);
1995 }
1996
1997 EXPORT_SYMBOL(target_committed_to_req);
1998
1999 #ifdef HAVE_QUOTA_SUPPORT
2000 int target_handle_qc_callback(struct ptlrpc_request *req)
2001 {
2002         struct obd_quotactl *oqctl;
2003         struct client_obd *cli = &req->rq_export->exp_obd->u.cli;
2004
2005         oqctl = req_capsule_client_get(&req->rq_pill, &RMF_OBD_QUOTACTL);
2006         if (oqctl == NULL)
2007                 RETURN(-EPROTO);
2008
2009         cli->cl_qchk_stat = oqctl->qc_stat;
2010
2011         return 0;
2012 }
2013
2014 int target_handle_dqacq_callback(struct ptlrpc_request *req)
2015 {
2016 #ifdef __KERNEL__
2017         struct obd_device *obd = req->rq_export->exp_obd;
2018         struct obd_device *master_obd;
2019         struct lustre_quota_ctxt *qctxt;
2020         struct qunit_data *qdata;
2021         void* rep;
2022         struct qunit_data_old *qdata_old;
2023         int rc = 0;
2024         ENTRY;
2025
2026         rc = req_capsule_server_pack(&req->rq_pill);
2027         if (rc) {
2028                 CERROR("packing reply failed!: rc = %d\n", rc);
2029                 RETURN(rc);
2030         }
2031
2032         LASSERT(req->rq_export);
2033
2034         /* fixed for bug10707 */
2035         if ((req->rq_export->exp_connect_flags & OBD_CONNECT_QUOTA64) &&
2036             !OBD_FAIL_CHECK(OBD_FAIL_QUOTA_QD_COUNT_32BIT)) {
2037                 CDEBUG(D_QUOTA, "qd_count is 64bit!\n");
2038                 rep = req_capsule_server_get(&req->rq_pill,
2039                                              &RMF_QUNIT_DATA);
2040                 LASSERT(rep);
2041                 qdata = req_capsule_client_swab_get(&req->rq_pill,
2042                                                     &RMF_QUNIT_DATA,
2043                                           (void*)lustre_swab_qdata);
2044         } else {
2045                 CDEBUG(D_QUOTA, "qd_count is 32bit!\n");
2046                 rep = req_capsule_server_get(&req->rq_pill, &RMF_QUNIT_DATA);
2047                 LASSERT(rep);
2048                 qdata_old = req_capsule_client_swab_get(&req->rq_pill,
2049                                                         &RMF_QUNIT_DATA,
2050                                            (void*)lustre_swab_qdata_old);
2051                 qdata = lustre_quota_old_to_new(qdata_old);
2052         }
2053
2054         if (qdata == NULL)
2055                 RETURN(-EPROTO);
2056
2057         /* we use the observer */
2058         LASSERT(obd->obd_observer && obd->obd_observer->obd_observer);
2059         master_obd = obd->obd_observer->obd_observer;
2060         qctxt = &master_obd->u.obt.obt_qctxt;
2061
2062         LASSERT(qctxt->lqc_handler);
2063         rc = qctxt->lqc_handler(master_obd, qdata,
2064                                 lustre_msg_get_opc(req->rq_reqmsg));
2065         if (rc && rc != -EDQUOT)
2066                 CDEBUG(rc == -EBUSY  ? D_QUOTA : D_ERROR,
2067                        "dqacq failed! (rc:%d)\n", rc);
2068
2069         /* the qd_count might be changed in lqc_handler */
2070         if ((req->rq_export->exp_connect_flags & OBD_CONNECT_QUOTA64) &&
2071             !OBD_FAIL_CHECK(OBD_FAIL_QUOTA_QD_COUNT_32BIT)) {
2072                 memcpy(rep, qdata, sizeof(*qdata));
2073         } else {
2074                 qdata_old = lustre_quota_new_to_old(qdata);
2075                 memcpy(rep, qdata_old, sizeof(*qdata_old));
2076         }
2077         req->rq_status = rc;
2078         rc = ptlrpc_reply(req);
2079
2080         RETURN(rc);
2081 #else
2082         return 0;
2083 #endif /* !__KERNEL__ */
2084 }
2085 #endif /* HAVE_QUOTA_SUPPORT */
2086
2087 ldlm_mode_t lck_compat_array[] = {
2088         [LCK_EX] LCK_COMPAT_EX,
2089         [LCK_PW] LCK_COMPAT_PW,
2090         [LCK_PR] LCK_COMPAT_PR,
2091         [LCK_CW] LCK_COMPAT_CW,
2092         [LCK_CR] LCK_COMPAT_CR,
2093         [LCK_NL] LCK_COMPAT_NL,
2094         [LCK_GROUP] LCK_COMPAT_GROUP
2095 };