lustre/ldlm/ldlm_lib.c

   1 /* -*- mode: c; c-basic-offset: 8; indent-tabs-mode: nil; -*-
   2  * vim:expandtab:shiftwidth=8:tabstop=8:
   3  *
   4  * GPL HEADER START
   5  *
   6  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
   7  *
   8  * This program is free software; you can redistribute it and/or modify
   9  * it under the terms of the GNU General Public License version 2 only,
  10  * as published by the Free Software Foundation.
  11  *
  12  * This program is distributed in the hope that it will be useful, but
  13  * WITHOUT ANY WARRANTY; without even the implied warranty of
  14  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  15  * General Public License version 2 for more details (a copy is included
  16  * in the LICENSE file that accompanied this code).
  17  *
  18  * You should have received a copy of the GNU General Public License
  19  * version 2 along with this program; If not, see
  20  * http://www.sun.com/software/products/lustre/docs/GPLv2.pdf
  21  *
  22  * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
  23  * CA 95054 USA or visit www.sun.com if you need additional information or
  24  * have any questions.
  25  *
  26  * GPL HEADER END
  27  */
  28 /*
  29  * Copyright  2008 Sun Microsystems, Inc. All rights reserved
  30  * Use is subject to license terms.
  31  */
  32 /*
  33  * This file is part of Lustre, http://www.lustre.org/
  34  * Lustre is a trademark of Sun Microsystems, Inc.
  35  */
  36
  37 #ifndef EXPORT_SYMTAB
  38 # define EXPORT_SYMTAB
  39 #endif
  40 #define DEBUG_SUBSYSTEM S_LDLM
  41
  42 #ifdef __KERNEL__
  43 # include <libcfs/libcfs.h>
  44 #else
  45 # include <liblustre.h>
  46 #endif
  47 #include <obd.h>
  48 #include <lustre_mds.h>
  49 #include <lustre_dlm.h>
  50 #include <lustre_net.h>
  51 #include <lustre_sec.h>
  52 #include "ldlm_internal.h"
  53
  54 /* @priority: if non-zero, move the selected to the list head
  55  * @create: if zero, only search in existed connections
  56  */
  57 static int import_set_conn(struct obd_import *imp, struct obd_uuid *uuid,
  58                            int priority, int create)
  59 {
  60         struct ptlrpc_connection *ptlrpc_conn;
  61         struct obd_import_conn *imp_conn = NULL, *item;
  62         int rc = 0;
  63         ENTRY;
  64
  65         if (!create && !priority) {
  66                 CDEBUG(D_HA, "Nothing to do\n");
  67                 RETURN(-EINVAL);
  68         }
  69
  70         ptlrpc_conn = ptlrpc_uuid_to_connection(uuid);
  71         if (!ptlrpc_conn) {
  72                 CDEBUG(D_HA, "can't find connection %s\n", uuid->uuid);
  73                 RETURN (-ENOENT);
  74         }
  75
  76         if (create) {
  77                 OBD_ALLOC(imp_conn, sizeof(*imp_conn));
  78                 if (!imp_conn) {
  79                         GOTO(out_put, rc = -ENOMEM);
  80                 }
  81         }
  82
  83         spin_lock(&imp->imp_lock);
  84         list_for_each_entry(item, &imp->imp_conn_list, oic_item) {
  85                 if (obd_uuid_equals(uuid, &item->oic_uuid)) {
  86                         if (priority) {
  87                                 list_del(&item->oic_item);
  88                                 list_add(&item->oic_item, &imp->imp_conn_list);
  89                                 item->oic_last_attempt = 0;
  90                         }
  91                         CDEBUG(D_HA, "imp %p@%s: found existing conn %s%s\n",
  92                                imp, imp->imp_obd->obd_name, uuid->uuid,
  93                                (priority ? ", moved to head" : ""));
  94                         spin_unlock(&imp->imp_lock);
  95                         GOTO(out_free, rc = 0);
  96                 }
  97         }
  98         /* not found */
  99         if (create) {
 100                 imp_conn->oic_conn = ptlrpc_conn;
 101                 imp_conn->oic_uuid = *uuid;
 102                 imp_conn->oic_last_attempt = 0;
 103                 if (priority)
 104                         list_add(&imp_conn->oic_item, &imp->imp_conn_list);
 105                 else
 106                         list_add_tail(&imp_conn->oic_item, &imp->imp_conn_list);
 107                 CDEBUG(D_HA, "imp %p@%s: add connection %s at %s\n",
 108                        imp, imp->imp_obd->obd_name, uuid->uuid,
 109                        (priority ? "head" : "tail"));
 110         } else {
 111                 spin_unlock(&imp->imp_lock);
 112                 GOTO(out_free, rc = -ENOENT);
 113         }
 114
 115         spin_unlock(&imp->imp_lock);
 116         RETURN(0);
 117 out_free:
 118         if (imp_conn)
 119                 OBD_FREE(imp_conn, sizeof(*imp_conn));
 120 out_put:
 121         ptlrpc_connection_put(ptlrpc_conn);
 122         RETURN(rc);
 123 }
 124
 125 int import_set_conn_priority(struct obd_import *imp, struct obd_uuid *uuid)
 126 {
 127         return import_set_conn(imp, uuid, 1, 0);
 128 }
 129
 130 int client_import_add_conn(struct obd_import *imp, struct obd_uuid *uuid,
 131                            int priority)
 132 {
 133         return import_set_conn(imp, uuid, priority, 1);
 134 }
 135
 136 int client_import_del_conn(struct obd_import *imp, struct obd_uuid *uuid)
 137 {
 138         struct obd_import_conn *imp_conn;
 139         struct obd_export *dlmexp;
 140         int rc = -ENOENT;
 141         ENTRY;
 142
 143         spin_lock(&imp->imp_lock);
 144         if (list_empty(&imp->imp_conn_list)) {
 145                 LASSERT(!imp->imp_connection);
 146                 GOTO(out, rc);
 147         }
 148
 149         list_for_each_entry(imp_conn, &imp->imp_conn_list, oic_item) {
 150                 if (!obd_uuid_equals(uuid, &imp_conn->oic_uuid))
 151                         continue;
 152                 LASSERT(imp_conn->oic_conn);
 153
 154                 /* is current conn? */
 155                 if (imp_conn == imp->imp_conn_current) {
 156                         LASSERT(imp_conn->oic_conn == imp->imp_connection);
 157
 158                         if (imp->imp_state != LUSTRE_IMP_CLOSED &&
 159                             imp->imp_state != LUSTRE_IMP_DISCON) {
 160                                 CERROR("can't remove current connection\n");
 161                                 GOTO(out, rc = -EBUSY);
 162                         }
 163
 164                         ptlrpc_connection_put(imp->imp_connection);
 165                         imp->imp_connection = NULL;
 166
 167                         dlmexp = class_conn2export(&imp->imp_dlm_handle);
 168                         if (dlmexp && dlmexp->exp_connection) {
 169                                 LASSERT(dlmexp->exp_connection ==
 170                                         imp_conn->oic_conn);
 171                                 ptlrpc_connection_put(dlmexp->exp_connection);
 172                                 dlmexp->exp_connection = NULL;
 173                         }
 174                 }
 175
 176                 list_del(&imp_conn->oic_item);
 177                 ptlrpc_connection_put(imp_conn->oic_conn);
 178                 OBD_FREE(imp_conn, sizeof(*imp_conn));
 179                 CDEBUG(D_HA, "imp %p@%s: remove connection %s\n",
 180                        imp, imp->imp_obd->obd_name, uuid->uuid);
 181                 rc = 0;
 182                 break;
 183         }
 184 out:
 185         spin_unlock(&imp->imp_lock);
 186         if (rc == -ENOENT)
 187                 CERROR("connection %s not found\n", uuid->uuid);
 188         RETURN(rc);
 189 }
 190
 191 void client_destroy_import(struct obd_import *imp)
 192 {
 193         /* drop security policy instance after all rpc finished/aborted
 194          * to let all busy contexts be released. */
 195         class_import_get(imp);
 196         class_destroy_import(imp);
 197         sptlrpc_import_sec_put(imp);
 198         class_import_put(imp);
 199 }
 200
 201 /* configure an RPC client OBD device
 202  *
 203  * lcfg parameters:
 204  * 1 - client UUID
 205  * 2 - server UUID
 206  * 3 - inactive-on-startup
 207  */
 208 int client_obd_setup(struct obd_device *obddev, struct lustre_cfg *lcfg)
 209 {
 210         struct client_obd *cli = &obddev->u.cli;
 211         struct obd_import *imp;
 212         struct obd_uuid server_uuid;
 213         int rq_portal, rp_portal, connect_op;
 214         char *name = obddev->obd_type->typ_name;
 215         int rc;
 216         ENTRY;
 217
 218         /* In a more perfect world, we would hang a ptlrpc_client off of
 219          * obd_type and just use the values from there. */
 220         if (!strcmp(name, LUSTRE_OSC_NAME)) {
 221                 rq_portal = OST_REQUEST_PORTAL;
 222                 rp_portal = OSC_REPLY_PORTAL;
 223                 connect_op = OST_CONNECT;
 224                 cli->cl_sp_me = LUSTRE_SP_CLI;
 225                 cli->cl_sp_to = LUSTRE_SP_OST;
 226         } else if (!strcmp(name, LUSTRE_MDC_NAME)) {
 227                 rq_portal = MDS_REQUEST_PORTAL;
 228                 rp_portal = MDC_REPLY_PORTAL;
 229                 connect_op = MDS_CONNECT;
 230                 cli->cl_sp_me = LUSTRE_SP_CLI;
 231                 cli->cl_sp_to = LUSTRE_SP_MDT;
 232         } else if (!strcmp(name, LUSTRE_MGC_NAME)) {
 233                 rq_portal = MGS_REQUEST_PORTAL;
 234                 rp_portal = MGC_REPLY_PORTAL;
 235                 connect_op = MGS_CONNECT;
 236                 cli->cl_sp_me = LUSTRE_SP_MGC;
 237                 cli->cl_sp_to = LUSTRE_SP_MGS;
 238                 cli->cl_flvr_mgc.sf_rpc = SPTLRPC_FLVR_INVALID;
 239         } else {
 240                 CERROR("unknown client OBD type \"%s\", can't setup\n",
 241                        name);
 242                 RETURN(-EINVAL);
 243         }
 244
 245         if (LUSTRE_CFG_BUFLEN(lcfg, 1) < 1) {
 246                 CERROR("requires a TARGET UUID\n");
 247                 RETURN(-EINVAL);
 248         }
 249
 250         if (LUSTRE_CFG_BUFLEN(lcfg, 1) > 37) {
 251                 CERROR("client UUID must be less than 38 characters\n");
 252                 RETURN(-EINVAL);
 253         }
 254
 255         if (LUSTRE_CFG_BUFLEN(lcfg, 2) < 1) {
 256                 CERROR("setup requires a SERVER UUID\n");
 257                 RETURN(-EINVAL);
 258         }
 259
 260         if (LUSTRE_CFG_BUFLEN(lcfg, 2) > 37) {
 261                 CERROR("target UUID must be less than 38 characters\n");
 262                 RETURN(-EINVAL);
 263         }
 264
 265         init_rwsem(&cli->cl_sem);
 266         sema_init(&cli->cl_mgc_sem, 1);
 267         cli->cl_conn_count = 0;
 268         memcpy(server_uuid.uuid, lustre_cfg_buf(lcfg, 2),
 269                min_t(unsigned int, LUSTRE_CFG_BUFLEN(lcfg, 2),
 270                      sizeof(server_uuid)));
 271
 272         cli->cl_dirty = 0;
 273         cli->cl_avail_grant = 0;
 274         /* FIXME: should limit this for the sum of all cl_dirty_max */
 275         cli->cl_dirty_max = OSC_MAX_DIRTY_DEFAULT * 1024 * 1024;
 276         if (cli->cl_dirty_max >> CFS_PAGE_SHIFT > num_physpages / 8)
 277                 cli->cl_dirty_max = num_physpages << (CFS_PAGE_SHIFT - 3);
 278         CFS_INIT_LIST_HEAD(&cli->cl_cache_waiters);
 279         CFS_INIT_LIST_HEAD(&cli->cl_loi_ready_list);
 280         CFS_INIT_LIST_HEAD(&cli->cl_loi_hp_ready_list);
 281         CFS_INIT_LIST_HEAD(&cli->cl_loi_write_list);
 282         CFS_INIT_LIST_HEAD(&cli->cl_loi_read_list);
 283         client_obd_list_lock_init(&cli->cl_loi_list_lock);
 284         cli->cl_r_in_flight = 0;
 285         cli->cl_w_in_flight = 0;
 286
 287         spin_lock_init(&cli->cl_read_rpc_hist.oh_lock);
 288         spin_lock_init(&cli->cl_write_rpc_hist.oh_lock);
 289         spin_lock_init(&cli->cl_read_page_hist.oh_lock);
 290         spin_lock_init(&cli->cl_write_page_hist.oh_lock);
 291         spin_lock_init(&cli->cl_read_offset_hist.oh_lock);
 292         spin_lock_init(&cli->cl_write_offset_hist.oh_lock);
 293         cfs_waitq_init(&cli->cl_destroy_waitq);
 294         atomic_set(&cli->cl_destroy_in_flight, 0);
 295 #ifdef ENABLE_CHECKSUM
 296         /* Turn on checksumming by default. */
 297         cli->cl_checksum = 1;
 298         /*
 299          * The supported checksum types will be worked out at connect time
 300          * Set cl_chksum* to CRC32 for now to avoid returning screwed info
 301          * through procfs.
 302          */
 303         cli->cl_cksum_type = cli->cl_supp_cksum_types = OBD_CKSUM_CRC32;
 304 #endif
 305         atomic_set(&cli->cl_resends, OSC_DEFAULT_RESENDS);
 306         atomic_set(&cli->cl_quota_resends, CLIENT_QUOTA_DEFAULT_RESENDS);
 307
 308         /* This value may be changed at connect time in
 309            ptlrpc_connect_interpret. */
 310         cli->cl_max_pages_per_rpc = min((int)PTLRPC_MAX_BRW_PAGES,
 311                                         (int)(1024 * 1024 >> CFS_PAGE_SHIFT));
 312
 313         if (!strcmp(name, LUSTRE_MDC_NAME)) {
 314                 cli->cl_max_rpcs_in_flight = MDC_MAX_RIF_DEFAULT;
 315         } else if (num_physpages >> (20 - CFS_PAGE_SHIFT) <= 128 /* MB */) {
 316                 cli->cl_max_rpcs_in_flight = 2;
 317         } else if (num_physpages >> (20 - CFS_PAGE_SHIFT) <= 256 /* MB */) {
 318                 cli->cl_max_rpcs_in_flight = 3;
 319         } else if (num_physpages >> (20 - CFS_PAGE_SHIFT) <= 512 /* MB */) {
 320                 cli->cl_max_rpcs_in_flight = 4;
 321         } else {
 322                 cli->cl_max_rpcs_in_flight = OSC_MAX_RIF_DEFAULT;
 323         }
 324
 325         rc = ldlm_get_ref();
 326         if (rc) {
 327                 CERROR("ldlm_get_ref failed: %d\n", rc);
 328                 GOTO(err, rc);
 329         }
 330
 331         ptlrpc_init_client(rq_portal, rp_portal, name,
 332                            &obddev->obd_ldlm_client);
 333
 334         imp = class_new_import(obddev);
 335         if (imp == NULL)
 336                 GOTO(err_ldlm, rc = -ENOENT);
 337         imp->imp_client = &obddev->obd_ldlm_client;
 338         imp->imp_connect_op = connect_op;
 339         imp->imp_initial_recov = 1;
 340         imp->imp_initial_recov_bk = 0;
 341         CFS_INIT_LIST_HEAD(&imp->imp_pinger_chain);
 342         memcpy(cli->cl_target_uuid.uuid, lustre_cfg_buf(lcfg, 1),
 343                LUSTRE_CFG_BUFLEN(lcfg, 1));
 344         class_import_put(imp);
 345
 346         rc = client_import_add_conn(imp, &server_uuid, 1);
 347         if (rc) {
 348                 CERROR("can't add initial connection\n");
 349                 GOTO(err_import, rc);
 350         }
 351
 352         cli->cl_import = imp;
 353         /* cli->cl_max_mds_{easize,cookiesize} updated by mdc_init_ea_size() */
 354         cli->cl_max_mds_easize = sizeof(struct lov_mds_md_v3);
 355         cli->cl_max_mds_cookiesize = sizeof(struct llog_cookie);
 356
 357         if (LUSTRE_CFG_BUFLEN(lcfg, 3) > 0) {
 358                 if (!strcmp(lustre_cfg_string(lcfg, 3), "inactive")) {
 359                         CDEBUG(D_HA, "marking %s %s->%s as inactive\n",
 360                                name, obddev->obd_name,
 361                                cli->cl_target_uuid.uuid);
 362                         spin_lock(&imp->imp_lock);
 363                         imp->imp_deactive = 1;
 364                         spin_unlock(&imp->imp_lock);
 365                 }
 366         }
 367
 368         obddev->obd_namespace = ldlm_namespace_new(obddev, obddev->obd_name,
 369                                                    LDLM_NAMESPACE_CLIENT,
 370                                                    LDLM_NAMESPACE_GREEDY);
 371         if (obddev->obd_namespace == NULL) {
 372                 CERROR("Unable to create client namespace - %s\n",
 373                        obddev->obd_name);
 374                 GOTO(err_import, rc = -ENOMEM);
 375         }
 376
 377         cli->cl_qchk_stat = CL_NOT_QUOTACHECKED;
 378
 379         RETURN(rc);
 380
 381 err_import:
 382         class_destroy_import(imp);
 383 err_ldlm:
 384         ldlm_put_ref();
 385 err:
 386         RETURN(rc);
 387
 388 }
 389
 390 int client_obd_cleanup(struct obd_device *obddev)
 391 {
 392         ENTRY;
 393
 394         ldlm_namespace_free_post(obddev->obd_namespace);
 395         obddev->obd_namespace = NULL;
 396
 397         ldlm_put_ref();
 398         RETURN(0);
 399 }
 400
 401 /* ->o_connect() method for client side (OSC and MDC and MGC) */
 402 int client_connect_import(const struct lu_env *env,
 403                           struct obd_export **exp,
 404                           struct obd_device *obd, struct obd_uuid *cluuid,
 405                           struct obd_connect_data *data, void *localdata)
 406 {
 407         struct client_obd *cli = &obd->u.cli;
 408         struct obd_import *imp = cli->cl_import;
 409         struct obd_connect_data *ocd;
 410         struct lustre_handle conn = { 0 };
 411         int rc;
 412         ENTRY;
 413
 414         *exp = NULL;
 415         down_write(&cli->cl_sem);
 416         if (cli->cl_conn_count > 0 )
 417                 GOTO(out_sem, rc = -EALREADY);
 418
 419         rc = class_connect(&conn, obd, cluuid);
 420         if (rc)
 421                 GOTO(out_sem, rc);
 422
 423         cli->cl_conn_count++;
 424         *exp = class_conn2export(&conn);
 425
 426         LASSERT(obd->obd_namespace);
 427
 428         imp->imp_dlm_handle = conn;
 429         rc = ptlrpc_init_import(imp);
 430         if (rc != 0)
 431                 GOTO(out_ldlm, rc);
 432
 433         ocd = &imp->imp_connect_data;
 434         if (data) {
 435                 *ocd = *data;
 436                 imp->imp_connect_flags_orig = data->ocd_connect_flags;
 437         }
 438
 439         rc = ptlrpc_connect_import(imp, NULL);
 440         if (rc != 0) {
 441                 LASSERT (imp->imp_state == LUSTRE_IMP_DISCON);
 442                 GOTO(out_ldlm, rc);
 443         }
 444         LASSERT((*exp)->exp_connection);
 445
 446         if (data) {
 447                 LASSERTF((ocd->ocd_connect_flags & data->ocd_connect_flags) ==
 448                          ocd->ocd_connect_flags, "old "LPX64", new "LPX64"\n",
 449                          data->ocd_connect_flags, ocd->ocd_connect_flags);
 450                 data->ocd_connect_flags = ocd->ocd_connect_flags;
 451         }
 452
 453         ptlrpc_pinger_add_import(imp);
 454
 455         EXIT;
 456
 457         if (rc) {
 458 out_ldlm:
 459                 cli->cl_conn_count--;
 460                 class_disconnect(*exp);
 461                 *exp = NULL;
 462         }
 463 out_sem:
 464         up_write(&cli->cl_sem);
 465
 466         return rc;
 467 }
 468
 469 int client_disconnect_export(struct obd_export *exp)
 470 {
 471         struct obd_device *obd = class_exp2obd(exp);
 472         struct client_obd *cli;
 473         struct obd_import *imp;
 474         int rc = 0, err;
 475         ENTRY;
 476
 477         if (!obd) {
 478                 CERROR("invalid export for disconnect: exp %p cookie "LPX64"\n",
 479                        exp, exp ? exp->exp_handle.h_cookie : -1);
 480                 RETURN(-EINVAL);
 481         }
 482
 483         cli = &obd->u.cli;
 484         imp = cli->cl_import;
 485
 486         down_write(&cli->cl_sem);
 487         CDEBUG(D_INFO, "disconnect %s - %d\n", obd->obd_name,
 488                cli->cl_conn_count);
 489
 490         if (!cli->cl_conn_count) {
 491                 CERROR("disconnecting disconnected device (%s)\n",
 492                        obd->obd_name);
 493                 GOTO(out_disconnect, rc = -EINVAL);
 494         }
 495
 496         cli->cl_conn_count--;
 497         if (cli->cl_conn_count)
 498                 GOTO(out_disconnect, rc = 0);
 499
 500         /* Mark import deactivated now, so we don't try to reconnect if any
 501          * of the cleanup RPCs fails (e.g. ldlm cancel, etc).  We don't
 502          * fully deactivate the import, or that would drop all requests. */
 503         spin_lock(&imp->imp_lock);
 504         imp->imp_deactive = 1;
 505         spin_unlock(&imp->imp_lock);
 506
 507         /* Some non-replayable imports (MDS's OSCs) are pinged, so just
 508          * delete it regardless.  (It's safe to delete an import that was
 509          * never added.) */
 510         (void)ptlrpc_pinger_del_import(imp);
 511
 512         if (obd->obd_namespace != NULL) {
 513                 /* obd_force == local only */
 514                 ldlm_cli_cancel_unused(obd->obd_namespace, NULL,
 515                                        obd->obd_force ? LDLM_FL_LOCAL_ONLY:0,
 516                                        NULL);
 517                 ldlm_namespace_free_prior(obd->obd_namespace, imp, obd->obd_force);
 518         }
 519
 520         /*
 521          * there's no necessary to hold sem during diconnecting an import,
 522          * and actually it may cause deadlock in gss.
 523          */
 524         up_write(&cli->cl_sem);
 525         rc = ptlrpc_disconnect_import(imp, 0);
 526         down_write(&cli->cl_sem);
 527
 528         ptlrpc_invalidate_import(imp);
 529
 530         if (imp->imp_rq_pool) {
 531                 ptlrpc_free_rq_pool(imp->imp_rq_pool);
 532                 imp->imp_rq_pool = NULL;
 533         }
 534         client_destroy_import(imp);
 535         cli->cl_import = NULL;
 536
 537         EXIT;
 538
 539  out_disconnect:
 540         /* use server style - class_disconnect should be always called for
 541          * o_disconnect */
 542         err = class_disconnect(exp);
 543         if (!rc && err)
 544                 rc = err;
 545
 546         up_write(&cli->cl_sem);
 547
 548         RETURN(rc);
 549 }
 550
 551 /* --------------------------------------------------------------------------
 552  * from old lib/target.c
 553  * -------------------------------------------------------------------------- */
 554
 555 static int target_handle_reconnect(struct lustre_handle *conn,
 556                                    struct obd_export *exp,
 557                                    struct obd_uuid *cluuid)
 558 {
 559         ENTRY;
 560         if (exp->exp_connection && exp->exp_imp_reverse) {
 561                 struct lustre_handle *hdl;
 562                 hdl = &exp->exp_imp_reverse->imp_remote_handle;
 563                 /* Might be a re-connect after a partition. */
 564                 if (!memcmp(&conn->cookie, &hdl->cookie, sizeof conn->cookie)) {
 565                         CWARN("%s: %s reconnecting\n", exp->exp_obd->obd_name,
 566                               cluuid->uuid);
 567                         conn->cookie = exp->exp_handle.h_cookie;
 568                         /* target_handle_connect() treats EALREADY and
 569                          * -EALREADY differently.  EALREADY means we are
 570                          * doing a valid reconnect from the same client. */
 571                         RETURN(EALREADY);
 572                 } else {
 573                         CERROR("%s reconnecting from %s, "
 574                                "handle mismatch (ours "LPX64", theirs "
 575                                LPX64")\n", cluuid->uuid,
 576                                exp->exp_connection->c_remote_uuid.uuid,
 577                                hdl->cookie, conn->cookie);
 578                         memset(conn, 0, sizeof *conn);
 579                         /* target_handle_connect() treats EALREADY and
 580                          * -EALREADY differently.  -EALREADY is an error
 581                          * (same UUID, different handle). */
 582                         RETURN(-EALREADY);
 583                 }
 584         }
 585
 586         conn->cookie = exp->exp_handle.h_cookie;
 587         CDEBUG(D_HA, "connect export for UUID '%s' at %p, cookie "LPX64"\n",
 588                cluuid->uuid, exp, conn->cookie);
 589         RETURN(0);
 590 }
 591
 592 void target_client_add_cb(struct obd_device *obd, __u64 transno, void *cb_data,
 593                           int error)
 594 {
 595         struct obd_export *exp = cb_data;
 596
 597         CDEBUG(D_RPCTRACE, "%s: committing for initial connect of %s\n",
 598                obd->obd_name, exp->exp_client_uuid.uuid);
 599
 600         spin_lock(&exp->exp_lock);
 601         exp->exp_need_sync = 0;
 602         spin_unlock(&exp->exp_lock);
 603 }
 604 EXPORT_SYMBOL(target_client_add_cb);
 605 static void
 606 target_start_and_reset_recovery_timer(struct obd_device *obd,
 607                                       struct ptlrpc_request *req,
 608                                       int new_client);
 609
 610 int target_handle_connect(struct ptlrpc_request *req)
 611 {
 612         struct obd_device *target, *targref = NULL;
 613         struct obd_export *export = NULL;
 614         struct obd_import *revimp;
 615         struct lustre_handle conn;
 616         struct lustre_handle *tmp;
 617         struct obd_uuid tgtuuid;
 618         struct obd_uuid cluuid;
 619         struct obd_uuid remote_uuid;
 620         char *str;
 621         int rc = 0;
 622         int mds_conn = 0;
 623         struct obd_connect_data *data, *tmpdata;
 624         lnet_nid_t *client_nid = NULL;
 625         ENTRY;
 626
 627         OBD_RACE(OBD_FAIL_TGT_CONN_RACE);
 628
 629         str = req_capsule_client_get(&req->rq_pill, &RMF_TGTUUID);
 630         if (str == NULL) {
 631                 DEBUG_REQ(D_ERROR, req, "bad target UUID for connect");
 632                 GOTO(out, rc = -EINVAL);
 633         }
 634
 635         obd_str2uuid(&tgtuuid, str);
 636         target = class_uuid2obd(&tgtuuid);
 637         if (!target)
 638                 target = class_name2obd(str);
 639
 640         if (!target || target->obd_stopping || !target->obd_set_up) {
 641                 LCONSOLE_ERROR_MSG(0x137, "UUID '%s' is not available "
 642                                    " for connect (%s)\n", str,
 643                                    !target ? "no target" :
 644                                    (target->obd_stopping ? "stopping" :
 645                                    "not set up"));
 646                 GOTO(out, rc = -ENODEV);
 647         }
 648
 649         if (target->obd_no_conn) {
 650                 LCONSOLE_WARN("%s: temporarily refusing client connection "
 651                               "from %s\n", target->obd_name,
 652                               libcfs_nid2str(req->rq_peer.nid));
 653                 GOTO(out, rc = -EAGAIN);
 654         }
 655
 656         /* Make sure the target isn't cleaned up while we're here. Yes,
 657            there's still a race between the above check and our incref here.
 658            Really, class_uuid2obd should take the ref. */
 659         targref = class_incref(target, __FUNCTION__, cfs_current());
 660
 661
 662         str = req_capsule_client_get(&req->rq_pill, &RMF_CLUUID);
 663         if (str == NULL) {
 664                 DEBUG_REQ(D_ERROR, req, "bad client UUID for connect");
 665                 GOTO(out, rc = -EINVAL);
 666         }
 667
 668         obd_str2uuid(&cluuid, str);
 669
 670         /* XXX extract a nettype and format accordingly */
 671         switch (sizeof(lnet_nid_t)) {
 672                 /* NB the casts only avoid compiler warnings */
 673         case 8:
 674                 snprintf(remote_uuid.uuid, sizeof remote_uuid,
 675                          "NET_"LPX64"_UUID", (__u64)req->rq_peer.nid);
 676                 break;
 677         case 4:
 678                 snprintf(remote_uuid.uuid, sizeof remote_uuid,
 679                          "NET_%x_UUID", (__u32)req->rq_peer.nid);
 680                 break;
 681         default:
 682                 LBUG();
 683         }
 684
 685         tmp = req_capsule_client_get(&req->rq_pill, &RMF_CONN);
 686         if (tmp == NULL)
 687                 GOTO(out, rc = -EPROTO);
 688
 689         conn = *tmp;
 690
 691         data = req_capsule_client_get(&req->rq_pill, &RMF_CONNECT_DATA);
 692         if (!data)
 693                 GOTO(out, rc = -EPROTO);
 694
 695         rc = req_capsule_server_pack(&req->rq_pill);
 696         if (rc)
 697                 GOTO(out, rc);
 698
 699         if (lustre_msg_get_op_flags(req->rq_reqmsg) & MSG_CONNECT_LIBCLIENT) {
 700                 if (!data) {
 701                         DEBUG_REQ(D_WARNING, req, "Refusing old (unversioned) "
 702                                   "libclient connection attempt");
 703                         GOTO(out, rc = -EPROTO);
 704                 } else if (data->ocd_version < LUSTRE_VERSION_CODE -
 705                                                LUSTRE_VERSION_ALLOWED_OFFSET ||
 706                            data->ocd_version > LUSTRE_VERSION_CODE +
 707                                                LUSTRE_VERSION_ALLOWED_OFFSET) {
 708                         DEBUG_REQ(D_WARNING, req, "Refusing %s (%d.%d.%d.%d) "
 709                                   "libclient connection attempt",
 710                                   data->ocd_version < LUSTRE_VERSION_CODE ?
 711                                   "old" : "new",
 712                                   OBD_OCD_VERSION_MAJOR(data->ocd_version),
 713                                   OBD_OCD_VERSION_MINOR(data->ocd_version),
 714                                   OBD_OCD_VERSION_PATCH(data->ocd_version),
 715                                   OBD_OCD_VERSION_FIX(data->ocd_version));
 716                         data = req_capsule_server_sized_get(&req->rq_pill,
 717                                                         &RMF_CONNECT_DATA,
 718                                     offsetof(typeof(*data), ocd_version) +
 719                                              sizeof(data->ocd_version));
 720                         if (data) {
 721                                 data->ocd_connect_flags = OBD_CONNECT_VERSION;
 722                                 data->ocd_version = LUSTRE_VERSION_CODE;
 723                         }
 724                         GOTO(out, rc = -EPROTO);
 725                 }
 726         }
 727
 728         if ((lustre_msg_get_op_flags(req->rq_reqmsg) & MSG_CONNECT_INITIAL) &&
 729             (data->ocd_connect_flags & OBD_CONNECT_MDS))
 730                 mds_conn = 1;
 731
 732         /* lctl gets a backstage, all-access pass. */
 733         if (obd_uuid_equals(&cluuid, &target->obd_uuid))
 734                 goto dont_check_exports;
 735
 736         export = lustre_hash_lookup(target->obd_uuid_hash, &cluuid);
 737         if (!export)
 738                 goto no_export;
 739
 740         /* we've found an export in the hash */
 741         if (export->exp_connecting) { /* bug 9635, et. al. */
 742                 CWARN("%s: exp %p already connecting\n",
 743                       export->exp_obd->obd_name, export);
 744                 class_export_put(export);
 745                 export = NULL;
 746                 rc = -EALREADY;
 747         } else if (mds_conn && export->exp_connection) {
 748                 if (req->rq_peer.nid != export->exp_connection->c_peer.nid)
 749                         /* mds reconnected after failover */
 750                         CWARN("%s: received MDS connection from NID %s,"
 751                               " removing former export from NID %s\n",
 752                             target->obd_name, libcfs_nid2str(req->rq_peer.nid),
 753                             libcfs_nid2str(export->exp_connection->c_peer.nid));
 754                 else
 755                         /* new mds connection from the same nid */
 756                         CWARN("%s: received new MDS connection from NID %s,"
 757                               " removing former export from same NID\n",
 758                             target->obd_name, libcfs_nid2str(req->rq_peer.nid));
 759                 class_fail_export(export);
 760                 class_export_put(export);
 761                 export = NULL;
 762                 rc = 0;
 763         } else if (export->exp_connection != NULL &&
 764                    req->rq_peer.nid != export->exp_connection->c_peer.nid &&
 765                    (lustre_msg_get_op_flags(req->rq_reqmsg) &
 766                     MSG_CONNECT_INITIAL)) {
 767                 /* in mds failover we have static uuid but nid can be
 768                  * changed*/
 769                 CWARN("%s: cookie %s seen on new NID %s when "
 770                       "existing NID %s is already connected\n",
 771                       target->obd_name, cluuid.uuid,
 772                       libcfs_nid2str(req->rq_peer.nid),
 773                       libcfs_nid2str(export->exp_connection->c_peer.nid));
 774                 rc = -EALREADY;
 775                 class_export_put(export);
 776                 export = NULL;
 777         } else {
 778                 spin_lock(&export->exp_lock);
 779                 export->exp_connecting = 1;
 780                 spin_unlock(&export->exp_lock);
 781                 class_export_put(export);
 782                 LASSERT(export->exp_obd == target);
 783
 784                 rc = target_handle_reconnect(&conn, export, &cluuid);
 785         }
 786
 787         /* If we found an export, we already unlocked. */
 788         if (!export) {
 789 no_export:
 790                 OBD_FAIL_TIMEOUT(OBD_FAIL_TGT_DELAY_CONNECT, 2 * obd_timeout);
 791         } else if (req->rq_export == NULL &&
 792                    atomic_read(&export->exp_rpc_count) > 0) {
 793                 CWARN("%s: refuse connection from %s/%s to 0x%p/%d\n",
 794                       target->obd_name, cluuid.uuid,
 795                       libcfs_nid2str(req->rq_peer.nid),
 796                       export, atomic_read(&export->exp_refcount));
 797                 GOTO(out, rc = -EBUSY);
 798         } else if (req->rq_export != NULL &&
 799                    (atomic_read(&export->exp_rpc_count) > 1)) {
 800                 /* the current connect rpc has increased exp_rpc_count */
 801                 CWARN("%s: refuse reconnection from %s@%s to 0x%p/%d\n",
 802                       target->obd_name, cluuid.uuid,
 803                       libcfs_nid2str(req->rq_peer.nid),
 804                       export, atomic_read(&export->exp_rpc_count) - 1);
 805                 spin_lock(&export->exp_lock);
 806                 if (req->rq_export->exp_conn_cnt <
 807                     lustre_msg_get_conn_cnt(req->rq_reqmsg))
 808                         /* try to abort active requests */
 809                         req->rq_export->exp_abort_active_req = 1;
 810                 spin_unlock(&export->exp_lock);
 811                 GOTO(out, rc = -EBUSY);
 812         } else if (lustre_msg_get_conn_cnt(req->rq_reqmsg) == 1) {
 813                 CERROR("%s: NID %s (%s) reconnected with 1 conn_cnt; "
 814                        "cookies not random?\n", target->obd_name,
 815                        libcfs_nid2str(req->rq_peer.nid), cluuid.uuid);
 816                 GOTO(out, rc = -EALREADY);
 817         } else {
 818                 OBD_FAIL_TIMEOUT(OBD_FAIL_TGT_DELAY_RECONNECT, 2 * obd_timeout);
 819         }
 820
 821         if (rc < 0) {
 822                 GOTO(out, rc);
 823         }
 824
 825         CWARN("%s: connection from %s@%s %st"LPU64" exp %p cur %ld last %ld\n",
 826                target->obd_name, cluuid.uuid, libcfs_nid2str(req->rq_peer.nid),
 827               target->obd_recovering ? "recovering/" : "", data->ocd_transno,
 828               export, (long)cfs_time_current_sec(),
 829               export ? (long)export->exp_last_request_time : 0);
 830
 831         /* Tell the client if we're in recovery. */
 832         if (target->obd_recovering) {
 833                 lustre_msg_add_op_flags(req->rq_repmsg, MSG_CONNECT_RECOVERING);
 834                 /* If this is the first time a client connects,
 835                    reset the recovery timer */
 836                 if (rc == 0)
 837                         target_start_and_reset_recovery_timer(target, req,
 838                                                               !export);
 839         }
 840
 841         /* We want to handle EALREADY but *not* -EALREADY from
 842          * target_handle_reconnect(), return reconnection state in a flag */
 843         if (rc == EALREADY) {
 844                 lustre_msg_add_op_flags(req->rq_repmsg, MSG_CONNECT_RECONNECT);
 845                 rc = 0;
 846         } else {
 847                 LASSERT(rc == 0);
 848         }
 849
 850         /* Tell the client if we support replayable requests */
 851         if (target->obd_replayable)
 852                 lustre_msg_add_op_flags(req->rq_repmsg, MSG_CONNECT_REPLAYABLE);
 853         client_nid = &req->rq_peer.nid;
 854
 855         if (export == NULL) {
 856                 if (target->obd_recovering) {
 857                         cfs_time_t t;
 858
 859                         t = cfs_timer_deadline(&target->obd_recovery_timer);
 860                         t = cfs_time_sub(t, cfs_time_current());
 861                         CERROR("%s: denying connection for new client %s (%s): "
 862                                "%d clients in recovery for "CFS_TIME_T"s\n",
 863                                target->obd_name,
 864                                libcfs_nid2str(req->rq_peer.nid), cluuid.uuid,
 865                                atomic_read(&target->obd_lock_replay_clients),
 866                                cfs_duration_sec(t));
 867                         rc = -EBUSY;
 868                 } else {
 869 dont_check_exports:
 870                         rc = obd_connect(req->rq_svc_thread->t_env,
 871                                          &export, target, &cluuid, data,
 872                                          client_nid);
 873                         if (rc == 0)
 874                                 conn.cookie = export->exp_handle.h_cookie;
 875                 }
 876         } else {
 877                 rc = obd_reconnect(req->rq_svc_thread->t_env,
 878                                    export, target, &cluuid, data, client_nid);
 879                 if (rc == 0)
 880                         /* prevous done via class_conn2export */
 881                         class_export_get(export);
 882         }
 883         if (rc)
 884                 GOTO(out, rc);
 885         /* Return only the parts of obd_connect_data that we understand, so the
 886          * client knows that we don't understand the rest. */
 887         if (data) {
 888                  tmpdata = req_capsule_server_get(&req->rq_pill,
 889                                                   &RMF_CONNECT_DATA);
 890                   //data->ocd_connect_flags &= OBD_CONNECT_SUPPORTED;
 891                  *tmpdata = *data;
 892         }
 893
 894         /* If all else goes well, this is our RPC return code. */
 895         req->rq_status = 0;
 896
 897         lustre_msg_set_handle(req->rq_repmsg, &conn);
 898
 899         /* If the client and the server are the same node, we will already
 900          * have an export that really points to the client's DLM export,
 901          * because we have a shared handles table.
 902          *
 903          * XXX this will go away when shaver stops sending the "connect" handle
 904          * in the real "remote handle" field of the request --phik 24 Apr 2003
 905          */
 906         if (req->rq_export != NULL)
 907                 class_export_put(req->rq_export);
 908
 909         req->rq_export = export;
 910
 911         spin_lock(&export->exp_lock);
 912         if (export->exp_conn_cnt >= lustre_msg_get_conn_cnt(req->rq_reqmsg)) {
 913                 spin_unlock(&export->exp_lock);
 914                 CERROR("%s: %s already connected at higher conn_cnt: %d > %d\n",
 915                        cluuid.uuid, libcfs_nid2str(req->rq_peer.nid),
 916                        export->exp_conn_cnt,
 917                        lustre_msg_get_conn_cnt(req->rq_reqmsg));
 918
 919                 GOTO(out, rc = -EALREADY);
 920         }
 921         export->exp_conn_cnt = lustre_msg_get_conn_cnt(req->rq_reqmsg);
 922         export->exp_abort_active_req = 0;
 923
 924         /* request from liblustre?  Don't evict it for not pinging. */
 925         if (lustre_msg_get_op_flags(req->rq_reqmsg) & MSG_CONNECT_LIBCLIENT) {
 926                 export->exp_libclient = 1;
 927                 spin_unlock(&export->exp_lock);
 928
 929                 spin_lock(&target->obd_dev_lock);
 930                 list_del_init(&export->exp_obd_chain_timed);
 931                 spin_unlock(&target->obd_dev_lock);
 932         } else {
 933                 spin_unlock(&export->exp_lock);
 934         }
 935
 936         if (export->exp_connection != NULL) {
 937                 /* Check to see if connection came from another NID */
 938                 if ((export->exp_connection->c_peer.nid != req->rq_peer.nid) &&
 939                     !hlist_unhashed(&export->exp_nid_hash))
 940                         lustre_hash_del(export->exp_obd->obd_nid_hash,
 941                                         &export->exp_connection->c_peer.nid,
 942                                         &export->exp_nid_hash);
 943
 944                 ptlrpc_connection_put(export->exp_connection);
 945         }
 946
 947         export->exp_connection = ptlrpc_connection_get(req->rq_peer,
 948                                                        req->rq_self,
 949                                                        &remote_uuid);
 950         if (hlist_unhashed(&export->exp_nid_hash)) {
 951                 lustre_hash_add_unique(export->exp_obd->obd_nid_hash,
 952                                        &export->exp_connection->c_peer.nid,
 953                                        &export->exp_nid_hash);
 954         }
 955
 956         spin_lock_bh(&target->obd_processing_task_lock);
 957         if (target->obd_recovering && !export->exp_in_recovery) {
 958                 spin_lock(&export->exp_lock);
 959                 export->exp_in_recovery = 1;
 960                 export->exp_req_replay_needed = 1;
 961                 export->exp_lock_replay_needed = 1;
 962                 spin_unlock(&export->exp_lock);
 963                 if ((lustre_msg_get_op_flags(req->rq_reqmsg) & MSG_CONNECT_TRANSNO)
 964                      && (data->ocd_transno == 0))
 965                         CWARN("Connect with zero transno!\n");
 966
 967                 if ((lustre_msg_get_op_flags(req->rq_reqmsg) & MSG_CONNECT_TRANSNO)
 968                      && data->ocd_transno < target->obd_next_recovery_transno)
 969                         target->obd_next_recovery_transno = data->ocd_transno;
 970                 target->obd_connected_clients++;
 971                 atomic_inc(&target->obd_req_replay_clients);
 972                 atomic_inc(&target->obd_lock_replay_clients);
 973                 if (target->obd_connected_clients ==
 974                     target->obd_max_recoverable_clients)
 975                         cfs_waitq_signal(&target->obd_next_transno_waitq);
 976         }
 977         spin_unlock_bh(&target->obd_processing_task_lock);
 978         tmp = req_capsule_client_get(&req->rq_pill, &RMF_CONN);
 979         conn = *tmp;
 980
 981         if (export->exp_imp_reverse != NULL) {
 982                 /* destroyed import can be still referenced in ctxt */
 983                 obd_set_info_async(export, sizeof(KEY_REVIMP_UPD),
 984                                    KEY_REVIMP_UPD, 0, NULL, NULL);
 985
 986                 client_destroy_import(export->exp_imp_reverse);
 987         }
 988
 989         /* for the rest part, we return -ENOTCONN in case of errors
 990          * in order to let client initialize connection again.
 991          */
 992         revimp = export->exp_imp_reverse = class_new_import(target);
 993         if (!revimp) {
 994                 CERROR("fail to alloc new reverse import.\n");
 995                 GOTO(out, rc = -ENOTCONN);
 996         }
 997
 998         revimp->imp_connection = ptlrpc_connection_addref(export->exp_connection);
 999         revimp->imp_client = &export->exp_obd->obd_ldlm_client;
1000         revimp->imp_remote_handle = conn;
1001         revimp->imp_dlm_fake = 1;
1002         revimp->imp_state = LUSTRE_IMP_FULL;
1003
1004         /* unknown versions will be caught in
1005          * ptlrpc_handle_server_req_in->lustre_unpack_msg() */
1006         revimp->imp_msg_magic = req->rq_reqmsg->lm_magic;
1007
1008         if ((export->exp_connect_flags & OBD_CONNECT_AT) &&
1009             (revimp->imp_msg_magic != LUSTRE_MSG_MAGIC_V1))
1010                 revimp->imp_msghdr_flags |= MSGHDR_AT_SUPPORT;
1011         else
1012                 revimp->imp_msghdr_flags &= ~MSGHDR_AT_SUPPORT;
1013
1014         rc = sptlrpc_import_sec_adapt(revimp, req->rq_svc_ctx, &req->rq_flvr);
1015         if (rc) {
1016                 CERROR("Failed to get sec for reverse import: %d\n", rc);
1017                 export->exp_imp_reverse = NULL;
1018                 class_destroy_import(revimp);
1019         }
1020
1021         class_import_put(revimp);
1022 out:
1023         if (export) {
1024                 spin_lock(&export->exp_lock);
1025                 export->exp_connecting = 0;
1026                 spin_unlock(&export->exp_lock);
1027         }
1028         if (targref)
1029                 class_decref(targref, __FUNCTION__, cfs_current());
1030         if (rc)
1031                 req->rq_status = rc;
1032         RETURN(rc);
1033 }
1034
1035 int target_handle_disconnect(struct ptlrpc_request *req)
1036 {
1037         int rc;
1038         ENTRY;
1039
1040         rc = req_capsule_server_pack(&req->rq_pill);
1041         if (rc)
1042                 RETURN(rc);
1043
1044         /* keep the rq_export around so we can send the reply */
1045         req->rq_status = obd_disconnect(class_export_get(req->rq_export));
1046
1047         RETURN(0);
1048 }
1049
1050 void target_destroy_export(struct obd_export *exp)
1051 {
1052         /* exports created from last_rcvd data, and "fake"
1053            exports created by lctl don't have an import */
1054         if (exp->exp_imp_reverse != NULL)
1055                 client_destroy_import(exp->exp_imp_reverse);
1056
1057         LASSERT(atomic_read(&exp->exp_locks_count) == 0);
1058         LASSERT(atomic_read(&exp->exp_rpc_count) == 0);
1059         LASSERT(atomic_read(&exp->exp_cb_count) == 0);
1060         LASSERT(atomic_read(&exp->exp_replay_count) == 0);
1061 }
1062
1063 /*
1064  * Recovery functions
1065  */
1066
1067 struct ptlrpc_request *ptlrpc_clone_req( struct ptlrpc_request *orig_req)
1068 {
1069         struct ptlrpc_request *copy_req;
1070         struct lustre_msg *copy_reqmsg;
1071         struct ptlrpc_user_desc *udesc = NULL;
1072
1073         OBD_ALLOC_PTR(copy_req);
1074         if (!copy_req)
1075                 return NULL;
1076         OBD_ALLOC(copy_reqmsg, orig_req->rq_reqlen);
1077         if (!copy_reqmsg){
1078                 OBD_FREE_PTR(copy_req);
1079                 return NULL;
1080         }
1081
1082         if (orig_req->rq_user_desc) {
1083                 int ngroups = orig_req->rq_user_desc->pud_ngroups;
1084
1085                 OBD_ALLOC(udesc, sptlrpc_user_desc_size(ngroups));
1086                 if (!udesc) {
1087                         OBD_FREE(copy_reqmsg, orig_req->rq_reqlen);
1088                         OBD_FREE_PTR(copy_req);
1089                         return NULL;
1090                 }
1091                 memcpy(udesc, orig_req->rq_user_desc,
1092                        sptlrpc_user_desc_size(ngroups));
1093         }
1094
1095         *copy_req = *orig_req;
1096         memcpy(copy_reqmsg, orig_req->rq_reqmsg, orig_req->rq_reqlen);
1097         copy_req->rq_reqmsg = copy_reqmsg;
1098         copy_req->rq_user_desc = udesc;
1099
1100         class_export_rpc_get(copy_req->rq_export);
1101         CFS_INIT_LIST_HEAD(&copy_req->rq_list);
1102         CFS_INIT_LIST_HEAD(&copy_req->rq_replay_list);
1103         sptlrpc_svc_ctx_addref(copy_req);
1104
1105         if (copy_req->rq_reply_state) {
1106                 /* the copied req takes over the reply state */
1107                 orig_req->rq_reply_state = NULL;
1108                 /* to catch further access */
1109                 orig_req->rq_repmsg = NULL;
1110                 orig_req->rq_replen = 0;
1111         }
1112         /** let export know it has replays to be handled */
1113         atomic_inc(&copy_req->rq_export->exp_replay_count);
1114         return copy_req;
1115 }
1116
1117 void ptlrpc_free_clone(struct ptlrpc_request *req)
1118 {
1119         LASSERT(list_empty(&req->rq_replay_list));
1120
1121         ptlrpc_req_drop_rs(req);
1122         sptlrpc_svc_ctx_decref(req);
1123         LASSERT(atomic_read(&req->rq_export->exp_replay_count) > 0);
1124         atomic_dec(&req->rq_export->exp_replay_count);
1125         class_export_rpc_put(req->rq_export);
1126         list_del_init(&req->rq_list);
1127
1128         if (req->rq_user_desc) {
1129                 int ngroups = req->rq_user_desc->pud_ngroups;
1130                 OBD_FREE(req->rq_user_desc, sptlrpc_user_desc_size(ngroups));
1131         }
1132         OBD_FREE(req->rq_reqmsg, req->rq_reqlen);
1133         OBD_FREE_PTR(req);
1134 }
1135
1136 static int target_exp_enqueue_req_replay(struct ptlrpc_request *req)
1137 {
1138         __u64                  transno = lustre_msg_get_transno(req->rq_reqmsg);
1139         struct obd_export     *exp = req->rq_export;
1140         struct ptlrpc_request *reqiter;
1141         int                    dup = 0;
1142
1143         LASSERT(exp);
1144
1145         spin_lock(&exp->exp_lock);
1146         list_for_each_entry(reqiter, &exp->exp_req_replay_queue,
1147                             rq_replay_list) {
1148                 if (lustre_msg_get_transno(reqiter->rq_reqmsg) == transno) {
1149                         dup = 1;
1150                         break;
1151                 }
1152         }
1153
1154         if (dup) {
1155                 /* we expect it with RESENT and REPLAY flags */
1156                 if ((lustre_msg_get_flags(req->rq_reqmsg) &
1157                      (MSG_RESENT | MSG_REPLAY)) != (MSG_RESENT | MSG_REPLAY))
1158                         CERROR("invalid flags %x of resent replay\n",
1159                                lustre_msg_get_flags(req->rq_reqmsg));
1160         } else {
1161                 list_add_tail(&req->rq_replay_list, &exp->exp_req_replay_queue);
1162         }
1163
1164         spin_unlock(&exp->exp_lock);
1165         return dup;
1166 }
1167
1168 static void target_exp_dequeue_req_replay(struct ptlrpc_request *req)
1169 {
1170         LASSERT(!list_empty(&req->rq_replay_list));
1171         LASSERT(req->rq_export);
1172
1173         spin_lock(&req->rq_export->exp_lock);
1174         list_del_init(&req->rq_replay_list);
1175         spin_unlock(&req->rq_export->exp_lock);
1176 }
1177
1178 #ifdef __KERNEL__
1179 static void target_finish_recovery(struct obd_device *obd)
1180 {
1181         ENTRY;
1182         LCONSOLE_INFO("%s: sending delayed replies to recovered clients\n",
1183                       obd->obd_name);
1184
1185         ldlm_reprocess_all_ns(obd->obd_namespace);
1186         spin_lock_bh(&obd->obd_processing_task_lock);
1187         if (!list_empty(&obd->obd_req_replay_queue) ||
1188             !list_empty(&obd->obd_lock_replay_queue) ||
1189             !list_empty(&obd->obd_final_req_queue)) {
1190                 CERROR("%s: Recovery queues ( %s%s%s) are not empty\n",
1191                        obd->obd_name,
1192                        list_empty(&obd->obd_req_replay_queue) ? "" : "req ",
1193                        list_empty(&obd->obd_lock_replay_queue) ? "" : "lock ",
1194                        list_empty(&obd->obd_final_req_queue) ? "" : "final ");
1195                 spin_unlock_bh(&obd->obd_processing_task_lock);
1196                 LBUG();
1197         }
1198         spin_unlock_bh(&obd->obd_processing_task_lock);
1199
1200         obd->obd_recovery_end = cfs_time_current_sec();
1201
1202         /* when recovery finished, cleanup orphans on mds and ost */
1203         if (OBT(obd) && OBP(obd, postrecov)) {
1204                 int rc = OBP(obd, postrecov)(obd);
1205                 if (rc < 0)
1206                         LCONSOLE_WARN("%s: Post recovery failed, rc %d\n",
1207                                       obd->obd_name, rc);
1208         }
1209         EXIT;
1210 }
1211
1212 static void abort_req_replay_queue(struct obd_device *obd)
1213 {
1214         struct ptlrpc_request *req, *n;
1215         struct list_head abort_list;
1216
1217         CFS_INIT_LIST_HEAD(&abort_list);
1218         spin_lock_bh(&obd->obd_processing_task_lock);
1219         list_splice_init(&obd->obd_req_replay_queue, &abort_list);
1220         spin_unlock_bh(&obd->obd_processing_task_lock);
1221         list_for_each_entry_safe(req, n, &abort_list, rq_list) {
1222                 DEBUG_REQ(D_WARNING, req, "aborted:");
1223                 req->rq_status = -ENOTCONN;
1224                 if (ptlrpc_error(req)) {
1225                         DEBUG_REQ(D_ERROR, req,
1226                                   "failed abort_req_reply; skipping");
1227                 }
1228                 target_exp_dequeue_req_replay(req);
1229                 ptlrpc_free_clone(req);
1230         }
1231 }
1232
1233 static void abort_lock_replay_queue(struct obd_device *obd)
1234 {
1235         struct ptlrpc_request *req, *n;
1236         struct list_head abort_list;
1237
1238         CFS_INIT_LIST_HEAD(&abort_list);
1239         spin_lock_bh(&obd->obd_processing_task_lock);
1240         list_splice_init(&obd->obd_lock_replay_queue, &abort_list);
1241         spin_unlock_bh(&obd->obd_processing_task_lock);
1242         list_for_each_entry_safe(req, n, &abort_list, rq_list){
1243                 DEBUG_REQ(D_ERROR, req, "aborted:");
1244                 req->rq_status = -ENOTCONN;
1245                 if (ptlrpc_error(req)) {
1246                         DEBUG_REQ(D_ERROR, req,
1247                                   "failed abort_lock_reply; skipping");
1248                 }
1249                 ptlrpc_free_clone(req);
1250         }
1251 }
1252 #endif
1253
1254 /* Called from a cleanup function if the device is being cleaned up
1255    forcefully.  The exports should all have been disconnected already,
1256    the only thing left to do is
1257      - clear the recovery flags
1258      - cancel the timer
1259      - free queued requests and replies, but don't send replies
1260    Because the obd_stopping flag is set, no new requests should be received.
1261
1262 */
1263 void target_cleanup_recovery(struct obd_device *obd)
1264 {
1265         struct ptlrpc_request *req, *n;
1266         struct list_head clean_list;
1267         ENTRY;
1268
1269         CFS_INIT_LIST_HEAD(&clean_list);
1270         spin_lock_bh(&obd->obd_processing_task_lock);
1271         if (!obd->obd_recovering) {
1272                 spin_unlock_bh(&obd->obd_processing_task_lock);
1273                 EXIT;
1274                 return;
1275         }
1276         obd->obd_recovering = obd->obd_abort_recovery = 0;
1277         target_cancel_recovery_timer(obd);
1278
1279         list_splice_init(&obd->obd_req_replay_queue, &clean_list);
1280         spin_unlock_bh(&obd->obd_processing_task_lock);
1281
1282         list_for_each_entry_safe(req, n, &clean_list, rq_list) {
1283                 LASSERT(req->rq_reply_state == 0);
1284                 target_exp_dequeue_req_replay(req);
1285                 ptlrpc_free_clone(req);
1286         }
1287
1288         spin_lock_bh(&obd->obd_processing_task_lock);
1289         list_splice_init(&obd->obd_lock_replay_queue, &clean_list);
1290         list_splice_init(&obd->obd_final_req_queue, &clean_list);
1291         spin_unlock_bh(&obd->obd_processing_task_lock);
1292
1293         list_for_each_entry_safe(req, n, &clean_list, rq_list){
1294                 LASSERT(req->rq_reply_state == 0);
1295                 ptlrpc_free_clone(req);
1296         }
1297
1298         EXIT;
1299 }
1300
1301 /* obd_processing_task_lock should be held */
1302 void target_cancel_recovery_timer(struct obd_device *obd)
1303 {
1304         CDEBUG(D_HA, "%s: cancel recovery timer\n", obd->obd_name);
1305         cfs_timer_disarm(&obd->obd_recovery_timer);
1306 }
1307
1308 /* extend = 1 means require at least "duration" seconds left in the timer,
1309    extend = 0 means set the total duration (start_recovery_timer) */
1310 static void reset_recovery_timer(struct obd_device *obd, int duration,
1311                                  int extend)
1312 {
1313         cfs_time_t now = cfs_time_current_sec();
1314         cfs_duration_t left;
1315
1316         spin_lock_bh(&obd->obd_processing_task_lock);
1317         if (!obd->obd_recovering || obd->obd_abort_recovery) {
1318                 spin_unlock_bh(&obd->obd_processing_task_lock);
1319                 return;
1320         }
1321
1322         left = cfs_time_sub(obd->obd_recovery_end, now);
1323
1324         if (extend && (duration > left))
1325                 obd->obd_recovery_timeout += duration - left;
1326         else if (!extend && (duration > obd->obd_recovery_timeout))
1327                 /* Track the client's largest expected replay time */
1328                 obd->obd_recovery_timeout = duration;
1329 #ifdef CRAY_XT3
1330         /*
1331          * If total recovery time already exceed the
1332          * obd_recovery_max_time, then CRAY XT3 will
1333          * abort the recovery
1334          */
1335         if(obd->obd_recovery_timeout > obd->obd_recovery_max_time)
1336                 obd->obd_recovery_timeout = obd->obd_recovery_max_time;
1337 #endif
1338         obd->obd_recovery_end = obd->obd_recovery_start +
1339                                 obd->obd_recovery_timeout;
1340         if (!cfs_timer_is_armed(&obd->obd_recovery_timer) ||
1341             cfs_time_before(now, obd->obd_recovery_end)) {
1342                 left = cfs_time_sub(obd->obd_recovery_end, now);
1343                 cfs_timer_arm(&obd->obd_recovery_timer, cfs_time_shift(left));
1344         }
1345         spin_unlock_bh(&obd->obd_processing_task_lock);
1346         CDEBUG(D_HA, "%s: recovery timer will expire in %u seconds\n",
1347                obd->obd_name, (unsigned)left);
1348 }
1349
1350 static void check_and_start_recovery_timer(struct obd_device *obd)
1351 {
1352         spin_lock_bh(&obd->obd_processing_task_lock);
1353         if (cfs_timer_is_armed(&obd->obd_recovery_timer)) {
1354                 spin_unlock_bh(&obd->obd_processing_task_lock);
1355                 return;
1356         }
1357         CDEBUG(D_HA, "%s: starting recovery timer\n", obd->obd_name);
1358         obd->obd_recovery_start = cfs_time_current_sec();
1359         /* minimum */
1360         obd->obd_recovery_timeout = OBD_RECOVERY_FACTOR * obd_timeout;
1361         spin_unlock_bh(&obd->obd_processing_task_lock);
1362
1363         reset_recovery_timer(obd, obd->obd_recovery_timeout, 0);
1364 }
1365
1366 /* Reset the timer with each new client connection */
1367 /*
1368  * This timer is actually reconnect_timer, which is for making sure
1369  * the total recovery window is at least as big as my reconnect
1370  * attempt timing. So the initial recovery time_out will be set to
1371  * OBD_RECOVERY_FACTOR * obd_timeout. If the timeout coming
1372  * from client is bigger than this, then the recovery time_out will
1373  * be extend to make sure the client could be reconnected, in the
1374  * process, the timeout from the new client should be ignored.
1375  */
1376
1377 static void
1378 target_start_and_reset_recovery_timer(struct obd_device *obd,
1379                                       struct ptlrpc_request *req,
1380                                       int new_client)
1381 {
1382         int service_time = lustre_msg_get_service_time(req->rq_reqmsg);
1383
1384         if (!new_client && service_time)
1385                 /* Teach server about old server's estimates, as first guess
1386                  * at how long new requests will take. */
1387                 at_add(&req->rq_rqbd->rqbd_service->srv_at_estimate,
1388                        service_time);
1389
1390         check_and_start_recovery_timer(obd);
1391
1392         /* convert the service time to rpc timeout,
1393          * reuse service_time to limit stack usage */
1394         service_time = at_est2timeout(service_time);
1395
1396         /* We expect other clients to timeout within service_time, then try
1397          * to reconnect, then try the failover server.  The max delay between
1398          * connect attempts is SWITCH_MAX + SWITCH_INC + INITIAL */
1399         service_time += 2 * (CONNECTION_SWITCH_MAX + CONNECTION_SWITCH_INC +
1400                              INITIAL_CONNECT_TIMEOUT);
1401         if (service_time > obd->obd_recovery_timeout && !new_client)
1402                 reset_recovery_timer(obd, service_time, 0);
1403 }
1404
1405 #ifdef __KERNEL__
1406
1407 /** Health checking routines */
1408 static inline int exp_connect_healthy(struct obd_export *exp)
1409 {
1410         return (exp->exp_in_recovery);
1411 }
1412
1413 /** if export done req_replay or has replay in queue */
1414 static inline int exp_req_replay_healthy(struct obd_export *exp)
1415 {
1416         return (!exp->exp_req_replay_needed ||
1417                 atomic_read(&exp->exp_replay_count) > 0);
1418 }
1419 /** if export done lock_replay or has replay in queue */
1420 static inline int exp_lock_replay_healthy(struct obd_export *exp)
1421 {
1422         return (!exp->exp_lock_replay_needed ||
1423                 atomic_read(&exp->exp_replay_count) > 0);
1424 }
1425
1426 static inline int exp_vbr_healthy(struct obd_export *exp)
1427 {
1428         return (!exp->exp_vbr_failed);
1429 }
1430
1431 static inline int exp_finished(struct obd_export *exp)
1432 {
1433         return (exp->exp_in_recovery && !exp->exp_lock_replay_needed);
1434 }
1435
1436 /** Checking routines for recovery */
1437 static int check_for_clients(struct obd_device *obd)
1438 {
1439         if (obd->obd_abort_recovery || obd->obd_recovery_expired)
1440                 return 1;
1441         LASSERT(obd->obd_connected_clients <= obd->obd_max_recoverable_clients);
1442         if (obd->obd_no_conn == 0 &&
1443             obd->obd_connected_clients + obd->obd_stale_clients ==
1444             obd->obd_max_recoverable_clients)
1445                 return 1;
1446         return 0;
1447 }
1448
1449 static int check_for_next_transno(struct obd_device *obd)
1450 {
1451         struct ptlrpc_request *req = NULL;
1452         int wake_up = 0, connected, completed, queue_len;
1453         __u64 next_transno, req_transno;
1454         ENTRY;
1455         spin_lock_bh(&obd->obd_processing_task_lock);
1456
1457         if (!list_empty(&obd->obd_req_replay_queue)) {
1458                 req = list_entry(obd->obd_req_replay_queue.next,
1459                                  struct ptlrpc_request, rq_list);
1460                 req_transno = lustre_msg_get_transno(req->rq_reqmsg);
1461         } else {
1462                 req_transno = 0;
1463         }
1464
1465         connected = obd->obd_connected_clients;
1466         completed = connected - atomic_read(&obd->obd_req_replay_clients);
1467         queue_len = obd->obd_requests_queued_for_recovery;
1468         next_transno = obd->obd_next_recovery_transno;
1469
1470         CDEBUG(D_HA, "max: %d, connected: %d, completed: %d, queue_len: %d, "
1471                "req_transno: "LPU64", next_transno: "LPU64"\n",
1472                obd->obd_max_recoverable_clients, connected, completed,
1473                queue_len, req_transno, next_transno);
1474
1475         if (obd->obd_abort_recovery) {
1476                 CDEBUG(D_HA, "waking for aborted recovery\n");
1477                 wake_up = 1;
1478         } else if (obd->obd_recovery_expired) {
1479                 CDEBUG(D_HA, "waking for expired recovery\n");
1480                 wake_up = 1;
1481         } else if (atomic_read(&obd->obd_req_replay_clients) == 0) {
1482                 CDEBUG(D_HA, "waking for completed recovery\n");
1483                 wake_up = 1;
1484         } else if (req_transno == next_transno) {
1485                 CDEBUG(D_HA, "waking for next ("LPD64")\n", next_transno);
1486                 wake_up = 1;
1487         } else if (queue_len == atomic_read(&obd->obd_req_replay_clients)) {
1488                 int d_lvl = D_HA;
1489                 /** handle gaps occured due to lost reply or VBR */
1490                 LASSERTF(req_transno >= next_transno,
1491                          "req_transno: "LPU64", next_transno: "LPU64"\n",
1492                          req_transno, next_transno);
1493                 if (req_transno > obd->obd_last_committed &&
1494                     !obd->obd_version_recov)
1495                         d_lvl = D_ERROR;
1496                 CDEBUG(d_lvl,
1497                        "%s: waking for gap in transno, VBR is %s (skip: "
1498                        LPD64", ql: %d, comp: %d, conn: %d, next: "LPD64
1499                        ", last_committed: "LPD64")\n",
1500                        obd->obd_name, obd->obd_version_recov ? "ON" : "OFF",
1501                        next_transno, queue_len, completed, connected,
1502                        req_transno, obd->obd_last_committed);
1503                 obd->obd_next_recovery_transno = req_transno;
1504                 wake_up = 1;
1505         } else if (OBD_FAIL_CHECK(OBD_FAIL_MDS_RECOVERY_ACCEPTS_GAPS)) {
1506                 CDEBUG(D_HA, "accepting transno gaps is explicitly allowed"
1507                        " by fail_lock, waking up ("LPD64")\n", next_transno);
1508                 obd->obd_next_recovery_transno = req_transno;
1509                 wake_up = 1;
1510         }
1511         spin_unlock_bh(&obd->obd_processing_task_lock);
1512         return wake_up;
1513 }
1514
1515 static int check_for_next_lock(struct obd_device *obd)
1516 {
1517         int wake_up = 0;
1518
1519         spin_lock_bh(&obd->obd_processing_task_lock);
1520         if (!list_empty(&obd->obd_lock_replay_queue)) {
1521                 CDEBUG(D_HA, "waking for next lock\n");
1522                 wake_up = 1;
1523         } else if (atomic_read(&obd->obd_lock_replay_clients) == 0) {
1524                 CDEBUG(D_HA, "waking for completed lock replay\n");
1525                 wake_up = 1;
1526         } else if (obd->obd_abort_recovery) {
1527                 CDEBUG(D_HA, "waking for aborted recovery\n");
1528                 wake_up = 1;
1529         } else if (obd->obd_recovery_expired) {
1530                 CDEBUG(D_HA, "waking for expired recovery\n");
1531                 wake_up = 1;
1532         }
1533         spin_unlock_bh(&obd->obd_processing_task_lock);
1534
1535         return wake_up;
1536 }
1537
1538 /**
1539  * wait for recovery events,
1540  * check its status with help of check_routine
1541  * evict dead clients via health_check
1542  */
1543 static int target_recovery_overseer(struct obd_device *obd,
1544                                     int (*check_routine)(struct obd_device *),
1545                                     int (*health_check)(struct obd_export *))
1546 {
1547         int abort = 0, expired = 1;
1548
1549         do {
1550                 cfs_wait_event(obd->obd_next_transno_waitq, check_routine(obd));
1551                 spin_lock_bh(&obd->obd_processing_task_lock);
1552                 abort = obd->obd_abort_recovery;
1553                 expired = obd->obd_recovery_expired;
1554                 obd->obd_recovery_expired = 0;
1555                 spin_unlock_bh(&obd->obd_processing_task_lock);
1556                 if (abort) {
1557                         CWARN("recovery is aborted, evict exports in recovery\n");
1558                         /** evict exports which didn't finish recovery yet */
1559                         class_disconnect_stale_exports(obd, exp_finished);
1560                 } else if (expired) {
1561                         /** If some clients died being recovered, evict them */
1562                         CDEBUG(D_WARNING, "recovery is timed out, evict stale exports\n");
1563                         /** evict cexports with no replay in queue, they are stalled */
1564                         class_disconnect_stale_exports(obd, health_check);
1565                         /** continue with VBR */
1566                         spin_lock_bh(&obd->obd_processing_task_lock);
1567                         obd->obd_version_recov = 1;
1568                         spin_unlock_bh(&obd->obd_processing_task_lock);
1569                         /**
1570                          * reset timer, recovery will proceed with versions now,
1571                          * timeout is set just to handle reconnection delays
1572                          */
1573                         reset_recovery_timer(obd, RECONNECT_DELAY_MAX, 1);
1574                         /** Wait for recovery events again, after evicting bad clients */
1575                 }
1576         } while (!abort && expired);
1577
1578         return abort;
1579 }
1580
1581 static struct ptlrpc_request *target_next_replay_req(struct obd_device *obd)
1582 {
1583         struct ptlrpc_request *req = NULL;
1584         ENTRY;
1585
1586         CDEBUG(D_HA, "Waiting for transno "LPD64"\n",
1587                obd->obd_next_recovery_transno);
1588
1589         if (target_recovery_overseer(obd, check_for_next_transno,
1590                                      exp_req_replay_healthy)) {
1591                 abort_req_replay_queue(obd);
1592                 abort_lock_replay_queue(obd);
1593         }
1594
1595         spin_lock_bh(&obd->obd_processing_task_lock);
1596         if (!list_empty(&obd->obd_req_replay_queue)) {
1597                 req = list_entry(obd->obd_req_replay_queue.next,
1598                                  struct ptlrpc_request, rq_list);
1599                 list_del_init(&req->rq_list);
1600                 obd->obd_requests_queued_for_recovery--;
1601                 spin_unlock_bh(&obd->obd_processing_task_lock);
1602         } else {
1603                 spin_unlock_bh(&obd->obd_processing_task_lock);
1604                 LASSERT(list_empty(&obd->obd_req_replay_queue));
1605                 LASSERT(atomic_read(&obd->obd_req_replay_clients) == 0);
1606                 /** evict exports failed VBR */
1607                 class_disconnect_stale_exports(obd, exp_vbr_healthy);
1608         }
1609         RETURN(req);
1610 }
1611
1612 static struct ptlrpc_request *target_next_replay_lock(struct obd_device *obd)
1613 {
1614         struct ptlrpc_request *req = NULL;
1615
1616         CDEBUG(D_HA, "Waiting for lock\n");
1617         if (target_recovery_overseer(obd, check_for_next_lock,
1618                                      exp_lock_replay_healthy))
1619                 abort_lock_replay_queue(obd);
1620
1621         spin_lock_bh(&obd->obd_processing_task_lock);
1622         if (!list_empty(&obd->obd_lock_replay_queue)) {
1623                 req = list_entry(obd->obd_lock_replay_queue.next,
1624                                  struct ptlrpc_request, rq_list);
1625                 list_del_init(&req->rq_list);
1626                 spin_unlock_bh(&obd->obd_processing_task_lock);
1627         } else {
1628                 spin_unlock_bh(&obd->obd_processing_task_lock);
1629                 LASSERT(list_empty(&obd->obd_lock_replay_queue));
1630                 LASSERT(atomic_read(&obd->obd_lock_replay_clients) == 0);
1631                 /** evict exports failed VBR */
1632                 class_disconnect_stale_exports(obd, exp_vbr_healthy);
1633         }
1634         return req;
1635 }
1636
1637 static struct ptlrpc_request *target_next_final_ping(struct obd_device *obd)
1638 {
1639         struct ptlrpc_request *req = NULL;
1640
1641         spin_lock_bh(&obd->obd_processing_task_lock);
1642         if (!list_empty(&obd->obd_final_req_queue)) {
1643                 req = list_entry(obd->obd_final_req_queue.next,
1644                                  struct ptlrpc_request, rq_list);
1645                 list_del_init(&req->rq_list);
1646                 if (req->rq_export->exp_in_recovery) {
1647                         spin_lock(&req->rq_export->exp_lock);
1648                         req->rq_export->exp_in_recovery = 0;
1649                         spin_unlock(&req->rq_export->exp_lock);
1650                 }
1651         }
1652         spin_unlock_bh(&obd->obd_processing_task_lock);
1653         return req;
1654 }
1655
1656 static int handle_recovery_req(struct ptlrpc_thread *thread,
1657                                struct ptlrpc_request *req,
1658                                svc_handler_t handler)
1659 {
1660         int rc;
1661         ENTRY;
1662
1663         rc = lu_context_init(&req->rq_session, LCT_SESSION);
1664         if (rc) {
1665                 CERROR("Failure to initialize session: %d\n", rc);
1666                 GOTO(free_clone, rc);
1667         }
1668         /**
1669          * export can be evicted during recovery, no need to handle replays for
1670          * it after that, discard such request silently
1671          */
1672         if (req->rq_export->exp_disconnected)
1673                 GOTO(free_clone, rc);
1674
1675         req->rq_session.lc_thread = thread;
1676         lu_context_enter(&req->rq_session);
1677         req->rq_svc_thread = thread;
1678         req->rq_svc_thread->t_env->le_ses = &req->rq_session;
1679
1680         /* thread context */
1681         lu_context_enter(&thread->t_env->le_ctx);
1682         (void)handler(req);
1683         lu_context_exit(&thread->t_env->le_ctx);
1684
1685         lu_context_exit(&req->rq_session);
1686         lu_context_fini(&req->rq_session);
1687         /* don't reset timer for final stage */
1688         if (!exp_finished(req->rq_export)) {
1689                 /**
1690                  * XXX: until bug 18948 is fixed (enable AT for request copy)
1691                  * the client may reconnect during recovery so we may need to
1692                  * wait RECONNECT_DELAY_MAX after each replay instead of
1693                  * at_get(&req->rq_rqbd->rqbd_service->srv_at_estimate);
1694                  */
1695                  reset_recovery_timer(class_exp2obd(req->rq_export), AT_OFF ?
1696                                       obd_timeout : RECONNECT_DELAY_MAX, 1);
1697         }
1698         /**
1699          * bz18031: increase next_recovery_transno before ptlrpc_free_clone()
1700          * will drop exp_rpc reference
1701          */
1702         if (req->rq_export->exp_req_replay_needed) {
1703                 spin_lock_bh(&req->rq_export->exp_obd->obd_processing_task_lock);
1704                 req->rq_export->exp_obd->obd_next_recovery_transno++;
1705                 spin_unlock_bh(&req->rq_export->exp_obd->obd_processing_task_lock);
1706                 target_exp_dequeue_req_replay(req);
1707         }
1708 free_clone:
1709         ptlrpc_free_clone(req);
1710         RETURN(0);
1711 }
1712
1713 static int target_recovery_thread(void *arg)
1714 {
1715         struct lu_target *lut = arg;
1716         struct obd_device *obd = lut->lut_obd;
1717         struct ptlrpc_request *req;
1718         struct target_recovery_data *trd = &obd->obd_recovery_data;
1719         unsigned long delta;
1720         unsigned long flags;
1721         struct lu_env env;
1722         struct ptlrpc_thread fake_svc_thread, *thread = &fake_svc_thread;
1723         int rc = 0;
1724         ENTRY;
1725
1726         cfs_daemonize("tgt_recov");
1727
1728         SIGNAL_MASK_LOCK(current, flags);
1729         sigfillset(&current->blocked);
1730         RECALC_SIGPENDING;
1731         SIGNAL_MASK_UNLOCK(current, flags);
1732
1733         rc = lu_context_init(&env.le_ctx, LCT_MD_THREAD);
1734         if (rc)
1735                 RETURN(rc);
1736
1737         thread->t_env = &env;
1738         env.le_ctx.lc_thread = thread;
1739
1740         CERROR("%s: started recovery thread pid %d\n", obd->obd_name,
1741                cfs_curproc_pid());
1742         trd->trd_processing_task = cfs_curproc_pid();
1743
1744         obd->obd_recovering = 1;
1745         complete(&trd->trd_starting);
1746
1747         /* first of all, we have to know the first transno to replay */
1748         if (target_recovery_overseer(obd, check_for_clients,
1749                                      exp_connect_healthy)) {
1750                 abort_req_replay_queue(obd);
1751                 abort_lock_replay_queue(obd);
1752         }
1753
1754         /* next stage: replay requests */
1755         delta = jiffies;
1756         obd->obd_req_replaying = 1;
1757         CDEBUG(D_INFO, "1: request replay stage - %d clients from t"LPU64"\n",
1758                atomic_read(&obd->obd_req_replay_clients),
1759                obd->obd_next_recovery_transno);
1760         while ((req = target_next_replay_req(obd))) {
1761                 LASSERT(trd->trd_processing_task == cfs_curproc_pid());
1762                 DEBUG_REQ(D_HA, req, "processing t"LPD64" from %s",
1763                           lustre_msg_get_transno(req->rq_reqmsg),
1764                           libcfs_nid2str(req->rq_peer.nid));
1765                 handle_recovery_req(thread, req,
1766                                     trd->trd_recovery_handler);
1767                 obd->obd_replayed_requests++;
1768         }
1769
1770         /**
1771          * The second stage: replay locks
1772          */
1773         CDEBUG(D_INFO, "2: lock replay stage - %d clients\n",
1774                atomic_read(&obd->obd_lock_replay_clients));
1775         while ((req = target_next_replay_lock(obd))) {
1776                 LASSERT(trd->trd_processing_task == cfs_curproc_pid());
1777                 DEBUG_REQ(D_HA, req, "processing lock from %s: ",
1778                           libcfs_nid2str(req->rq_peer.nid));
1779                 handle_recovery_req(thread, req,
1780                                     trd->trd_recovery_handler);
1781                 obd->obd_replayed_locks++;
1782         }
1783
1784         /**
1785          * The third stage: reply on final pings, at this moment all clients
1786          * must have request in final queue
1787          */
1788         CDEBUG(D_INFO, "3: final stage - process recovery completion pings\n");
1789         /** Update server last boot epoch */
1790         lut_boot_epoch_update(lut);
1791         /* We drop recoverying flag to forward all new requests
1792          * to regular mds_handle() since now */
1793         spin_lock_bh(&obd->obd_processing_task_lock);
1794         obd->obd_recovering = obd->obd_abort_recovery = 0;
1795         target_cancel_recovery_timer(obd);
1796         spin_unlock_bh(&obd->obd_processing_task_lock);
1797         while ((req = target_next_final_ping(obd))) {
1798                 LASSERT(trd->trd_processing_task == cfs_curproc_pid());
1799                 DEBUG_REQ(D_HA, req, "processing final ping from %s: ",
1800                           libcfs_nid2str(req->rq_peer.nid));
1801                 handle_recovery_req(thread, req,
1802                                     trd->trd_recovery_handler);
1803         }
1804
1805         delta = (jiffies - delta) / HZ;
1806         CDEBUG(D_INFO,"4: recovery completed in %lus - %d/%d reqs/locks\n",
1807               delta, obd->obd_replayed_requests, obd->obd_replayed_locks);
1808         if (delta > obd_timeout * OBD_RECOVERY_FACTOR) {
1809                 CWARN("too long recovery - read logs\n");
1810                 libcfs_debug_dumplog();
1811         }
1812
1813         target_finish_recovery(obd);
1814
1815         lu_context_fini(&env.le_ctx);
1816         trd->trd_processing_task = 0;
1817         complete(&trd->trd_finishing);
1818         RETURN(rc);
1819 }
1820
1821 static int target_start_recovery_thread(struct lu_target *lut,
1822                                         svc_handler_t handler)
1823 {
1824         struct obd_device *obd = lut->lut_obd;
1825         int rc = 0;
1826         struct target_recovery_data *trd = &obd->obd_recovery_data;
1827
1828         memset(trd, 0, sizeof(*trd));
1829         init_completion(&trd->trd_starting);
1830         init_completion(&trd->trd_finishing);
1831         trd->trd_recovery_handler = handler;
1832
1833         if (kernel_thread(target_recovery_thread, lut, 0) > 0) {
1834                 wait_for_completion(&trd->trd_starting);
1835                 LASSERT(obd->obd_recovering != 0);
1836         } else
1837                 rc = -ECHILD;
1838
1839         return rc;
1840 }
1841
1842 void target_stop_recovery_thread(struct obd_device *obd)
1843 {
1844         spin_lock_bh(&obd->obd_processing_task_lock);
1845         if (obd->obd_recovery_data.trd_processing_task > 0) {
1846                 struct target_recovery_data *trd = &obd->obd_recovery_data;
1847                 /** recovery can be done but postrecovery is not yet */
1848                 if (obd->obd_recovering) {
1849                         CERROR("%s: Aborting recovery\n", obd->obd_name);
1850                         obd->obd_abort_recovery = 1;
1851                         cfs_waitq_signal(&obd->obd_next_transno_waitq);
1852                 }
1853                 spin_unlock_bh(&obd->obd_processing_task_lock);
1854                 wait_for_completion(&trd->trd_finishing);
1855         } else {
1856                 spin_unlock_bh(&obd->obd_processing_task_lock);
1857         }
1858 }
1859
1860 void target_recovery_fini(struct obd_device *obd)
1861 {
1862         class_disconnect_exports(obd);
1863         target_stop_recovery_thread(obd);
1864         target_cleanup_recovery(obd);
1865 }
1866 EXPORT_SYMBOL(target_recovery_fini);
1867
1868 static void target_recovery_expired(unsigned long castmeharder)
1869 {
1870         struct obd_device *obd = (struct obd_device *)castmeharder;
1871         CDEBUG(D_HA, "%s: recovery timed out; %d clients are still in recovery"
1872                " after %lds (%d clients connected)\n",
1873                obd->obd_name, atomic_read(&obd->obd_lock_replay_clients),
1874                cfs_time_current_sec()- obd->obd_recovery_start,
1875                obd->obd_connected_clients);
1876
1877         spin_lock_bh(&obd->obd_processing_task_lock);
1878         obd->obd_recovery_expired = 1;
1879         cfs_waitq_signal(&obd->obd_next_transno_waitq);
1880         spin_unlock_bh(&obd->obd_processing_task_lock);
1881 }
1882
1883 void target_recovery_init(struct lu_target *lut, svc_handler_t handler)
1884 {
1885         struct obd_device *obd = lut->lut_obd;
1886         if (obd->obd_max_recoverable_clients == 0) {
1887                 /** Update server last boot epoch */
1888                 lut_boot_epoch_update(lut);
1889                 return;
1890         }
1891
1892         CWARN("RECOVERY: service %s, %d recoverable clients, "
1893               "last_transno "LPU64"\n", obd->obd_name,
1894               obd->obd_max_recoverable_clients, obd->obd_last_committed);
1895         LASSERT(obd->obd_stopping == 0);
1896         obd->obd_next_recovery_transno = obd->obd_last_committed + 1;
1897         obd->obd_recovery_start = 0;
1898         obd->obd_recovery_end = 0;
1899         obd->obd_recovery_timeout = OBD_RECOVERY_FACTOR * obd_timeout;
1900         /* bz13079: this should be set to desired value for ost but not for mds */
1901         obd->obd_recovery_max_time = OBD_RECOVERY_MAX_TIME;
1902         cfs_timer_init(&obd->obd_recovery_timer, target_recovery_expired, obd);
1903         target_start_recovery_thread(lut, handler);
1904 }
1905 EXPORT_SYMBOL(target_recovery_init);
1906
1907 #endif
1908
1909 static int target_process_req_flags(struct obd_device *obd,
1910                                     struct ptlrpc_request *req)
1911 {
1912         struct obd_export *exp = req->rq_export;
1913         LASSERT(exp != NULL);
1914         if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REQ_REPLAY_DONE) {
1915                 /* client declares he's ready to replay locks */
1916                 spin_lock_bh(&obd->obd_processing_task_lock);
1917                 if (exp->exp_req_replay_needed) {
1918                         LASSERT(atomic_read(&obd->obd_req_replay_clients) > 0);
1919                         spin_lock(&exp->exp_lock);
1920                         exp->exp_req_replay_needed = 0;
1921                         spin_unlock(&exp->exp_lock);
1922                         atomic_dec(&obd->obd_req_replay_clients);
1923                 }
1924                 spin_unlock_bh(&obd->obd_processing_task_lock);
1925         }
1926         if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_LOCK_REPLAY_DONE) {
1927                 /* client declares he's ready to complete recovery
1928                  * so, we put the request on th final queue */
1929                 spin_lock_bh(&obd->obd_processing_task_lock);
1930                 if (exp->exp_lock_replay_needed) {
1931                         LASSERT(atomic_read(&obd->obd_lock_replay_clients) > 0);
1932                         spin_lock(&exp->exp_lock);
1933                         exp->exp_lock_replay_needed = 0;
1934                         spin_unlock(&exp->exp_lock);
1935                         atomic_dec(&obd->obd_lock_replay_clients);
1936                 }
1937                 spin_unlock_bh(&obd->obd_processing_task_lock);
1938         }
1939
1940         return 0;
1941 }
1942
1943 int target_queue_recovery_request(struct ptlrpc_request *req,
1944                                   struct obd_device *obd)
1945 {
1946         struct list_head *tmp;
1947         int inserted = 0;
1948         __u64 transno = lustre_msg_get_transno(req->rq_reqmsg);
1949         ENTRY;
1950
1951         if (obd->obd_recovery_data.trd_processing_task == cfs_curproc_pid()) {
1952                 /* Processing the queue right now, don't re-add. */
1953                 RETURN(1);
1954         }
1955
1956         target_process_req_flags(obd, req);
1957
1958         if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_LOCK_REPLAY_DONE) {
1959                 /* client declares he's ready to complete recovery
1960                  * so, we put the request on th final queue */
1961                 req = ptlrpc_clone_req(req);
1962                 if (req == NULL)
1963                         RETURN(-ENOMEM);
1964                 DEBUG_REQ(D_HA, req, "queue final req");
1965                 spin_lock_bh(&obd->obd_processing_task_lock);
1966                 cfs_waitq_signal(&obd->obd_next_transno_waitq);
1967                 if (obd->obd_recovering)
1968                         list_add_tail(&req->rq_list, &obd->obd_final_req_queue);
1969                 else {
1970                         spin_unlock_bh(&obd->obd_processing_task_lock);
1971                         ptlrpc_free_clone(req);
1972                         if (obd->obd_stopping) {
1973                                 RETURN(-ENOTCONN);
1974                         } else {
1975                                 RETURN(1);
1976                         }
1977                 }
1978                 spin_unlock_bh(&obd->obd_processing_task_lock);
1979                 RETURN(0);
1980         }
1981         if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REQ_REPLAY_DONE) {
1982                 /* client declares he's ready to replay locks */
1983                 req = ptlrpc_clone_req(req);
1984                 if (req == NULL)
1985                         RETURN(-ENOMEM);
1986                 DEBUG_REQ(D_HA, req, "queue lock replay req");
1987                 spin_lock_bh(&obd->obd_processing_task_lock);
1988                 cfs_waitq_signal(&obd->obd_next_transno_waitq);
1989                 LASSERT(obd->obd_recovering);
1990                 /* usually due to recovery abort */
1991                 if (!req->rq_export->exp_in_recovery) {
1992                         spin_unlock_bh(&obd->obd_processing_task_lock);
1993                         ptlrpc_free_clone(req);
1994                         RETURN(-ENOTCONN);
1995                 }
1996                 LASSERT(req->rq_export->exp_lock_replay_needed);
1997                 list_add_tail(&req->rq_list, &obd->obd_lock_replay_queue);
1998                 spin_unlock_bh(&obd->obd_processing_task_lock);
1999                 RETURN(0);
2000         }
2001
2002         /* CAVEAT EMPTOR: The incoming request message has been swabbed
2003          * (i.e. buflens etc are in my own byte order), but type-dependent
2004          * buffers (eg mds_body, ost_body etc) have NOT been swabbed. */
2005
2006         if (!transno) {
2007                 CFS_INIT_LIST_HEAD(&req->rq_list);
2008                 DEBUG_REQ(D_HA, req, "not queueing");
2009                 RETURN(1);
2010         }
2011
2012         spin_lock_bh(&obd->obd_processing_task_lock);
2013
2014         /* If we're processing the queue, we want don't want to queue this
2015          * message.
2016          *
2017          * Also, if this request has a transno less than the one we're waiting
2018          * for, we should process it now.  It could (and currently always will)
2019          * be an open request for a descriptor that was opened some time ago.
2020          *
2021          * Also, a resent, replayed request that has already been
2022          * handled will pass through here and be processed immediately.
2023          */
2024         CWARN("Next recovery transno: "LPU64", current: "LPU64", replaying: %i\n",
2025               obd->obd_next_recovery_transno, transno, obd->obd_req_replaying);
2026         if (transno < obd->obd_next_recovery_transno && obd->obd_req_replaying) {
2027                 /* Processing the queue right now, don't re-add. */
2028                 LASSERT(list_empty(&req->rq_list));
2029                 spin_unlock_bh(&obd->obd_processing_task_lock);
2030                 RETURN(1);
2031         }
2032         spin_unlock_bh(&obd->obd_processing_task_lock);
2033
2034         if (OBD_FAIL_CHECK(OBD_FAIL_TGT_REPLAY_DROP))
2035                 RETURN(0);
2036
2037         req = ptlrpc_clone_req(req);
2038         if (req == NULL)
2039                 RETURN(-ENOMEM);
2040
2041         spin_lock_bh(&obd->obd_processing_task_lock);
2042         LASSERT(obd->obd_recovering);
2043         if (!req->rq_export->exp_in_recovery) {
2044                 spin_unlock_bh(&obd->obd_processing_task_lock);
2045                 ptlrpc_free_clone(req);
2046                 RETURN(-ENOTCONN);
2047         }
2048         LASSERT(req->rq_export->exp_req_replay_needed);
2049
2050         if (target_exp_enqueue_req_replay(req)) {
2051                 spin_unlock_bh(&obd->obd_processing_task_lock);
2052                 DEBUG_REQ(D_ERROR, req, "dropping resent queued req");
2053                 ptlrpc_free_clone(req);
2054                 RETURN(0);
2055         }
2056
2057         /* XXX O(n^2) */
2058         list_for_each(tmp, &obd->obd_req_replay_queue) {
2059                 struct ptlrpc_request *reqiter =
2060                         list_entry(tmp, struct ptlrpc_request, rq_list);
2061
2062                 if (lustre_msg_get_transno(reqiter->rq_reqmsg) > transno) {
2063                         list_add_tail(&req->rq_list, &reqiter->rq_list);
2064                         inserted = 1;
2065                         break;
2066                 }
2067
2068                 if (unlikely(lustre_msg_get_transno(reqiter->rq_reqmsg) ==
2069                              transno)) {
2070                         DEBUG_REQ(D_ERROR, req, "dropping replay: transno "
2071                                   "has been claimed by another client");
2072                         spin_unlock_bh(&obd->obd_processing_task_lock);
2073                         target_exp_dequeue_req_replay(req);
2074                         ptlrpc_free_clone(req);
2075                         RETURN(0);
2076                 }
2077         }
2078
2079         if (!inserted)
2080                 list_add_tail(&req->rq_list, &obd->obd_req_replay_queue);
2081
2082         obd->obd_requests_queued_for_recovery++;
2083         cfs_waitq_signal(&obd->obd_next_transno_waitq);
2084         spin_unlock_bh(&obd->obd_processing_task_lock);
2085         RETURN(0);
2086 }
2087
2088 /**
2089  * Packs current SLV and Limit into \a req.
2090  */
2091 int target_pack_pool_reply(struct ptlrpc_request *req)
2092 {
2093         struct obd_device *obd;
2094         ENTRY;
2095
2096         /*
2097          * Check that we still have all structures alive as this may
2098          * be some late rpc in shutdown time.
2099          */
2100         if (unlikely(!req->rq_export || !req->rq_export->exp_obd ||
2101                      !exp_connect_lru_resize(req->rq_export))) {
2102                 lustre_msg_set_slv(req->rq_repmsg, 0);
2103                 lustre_msg_set_limit(req->rq_repmsg, 0);
2104                 RETURN(0);
2105         }
2106
2107         /*
2108          * OBD is alive here as export is alive, which we checked above.
2109          */
2110         obd = req->rq_export->exp_obd;
2111
2112         read_lock(&obd->obd_pool_lock);
2113         lustre_msg_set_slv(req->rq_repmsg, obd->obd_pool_slv);
2114         lustre_msg_set_limit(req->rq_repmsg, obd->obd_pool_limit);
2115         read_unlock(&obd->obd_pool_lock);
2116
2117         RETURN(0);
2118 }
2119
2120 int target_send_reply_msg(struct ptlrpc_request *req, int rc, int fail_id)
2121 {
2122         if (OBD_FAIL_CHECK_ORSET(fail_id & ~OBD_FAIL_ONCE, OBD_FAIL_ONCE)) {
2123                 DEBUG_REQ(D_ERROR, req, "dropping reply");
2124                 return (-ECOMM);
2125         }
2126
2127         if (unlikely(rc)) {
2128                 DEBUG_REQ(D_ERROR, req, "processing error (%d)", rc);
2129                 req->rq_status = rc;
2130                 return (ptlrpc_send_error(req, 1));
2131         } else {
2132                 DEBUG_REQ(D_NET, req, "sending reply");
2133         }
2134
2135         return (ptlrpc_send_reply(req, PTLRPC_REPLY_MAYBE_DIFFICULT));
2136 }
2137
2138 void target_send_reply(struct ptlrpc_request *req, int rc, int fail_id)
2139 {
2140         int                        netrc;
2141         struct ptlrpc_reply_state *rs;
2142         struct obd_device         *obd;
2143         struct obd_export         *exp;
2144         struct ptlrpc_service     *svc;
2145         ENTRY;
2146
2147         if (req->rq_no_reply) {
2148                 EXIT;
2149                 return;
2150         }
2151
2152         svc = req->rq_rqbd->rqbd_service;
2153         rs = req->rq_reply_state;
2154         if (rs == NULL || !rs->rs_difficult) {
2155                 /* no notifiers */
2156                 target_send_reply_msg (req, rc, fail_id);
2157                 EXIT;
2158                 return;
2159         }
2160
2161         /* must be an export if locks saved */
2162         LASSERT (req->rq_export != NULL);
2163         /* req/reply consistent */
2164         LASSERT (rs->rs_service == svc);
2165
2166         /* "fresh" reply */
2167         LASSERT (!rs->rs_scheduled);
2168         LASSERT (!rs->rs_scheduled_ever);
2169         LASSERT (!rs->rs_handled);
2170         LASSERT (!rs->rs_on_net);
2171         LASSERT (rs->rs_export == NULL);
2172         LASSERT (list_empty(&rs->rs_obd_list));
2173         LASSERT (list_empty(&rs->rs_exp_list));
2174
2175         exp = class_export_get (req->rq_export);
2176         obd = exp->exp_obd;
2177
2178         /* disable reply scheduling while I'm setting up */
2179         rs->rs_scheduled = 1;
2180         rs->rs_on_net    = 1;
2181         rs->rs_xid       = req->rq_xid;
2182         rs->rs_transno   = req->rq_transno;
2183         rs->rs_export    = exp;
2184         rs->rs_opc       = lustre_msg_get_opc(rs->rs_msg);
2185
2186         spin_lock(&exp->exp_uncommitted_replies_lock);
2187         CDEBUG(D_NET, "rs transno = "LPU64", last committed = "LPU64"\n",
2188                rs->rs_transno, exp->exp_last_committed);
2189         if (rs->rs_transno > exp->exp_last_committed) {
2190                 /* not committed already */
2191                 list_add_tail(&rs->rs_obd_list,
2192                               &exp->exp_uncommitted_replies);
2193         }
2194         spin_unlock (&exp->exp_uncommitted_replies_lock);
2195
2196         spin_lock(&exp->exp_lock);
2197         list_add_tail(&rs->rs_exp_list, &exp->exp_outstanding_replies);
2198         spin_unlock(&exp->exp_lock);
2199
2200         netrc = target_send_reply_msg (req, rc, fail_id);
2201
2202         spin_lock(&svc->srv_lock);
2203
2204         atomic_inc(&svc->srv_n_difficult_replies);
2205
2206         if (netrc != 0) {
2207                 /* error sending: reply is off the net.  Also we need +1
2208                  * reply ref until ptlrpc_handle_rs() is done
2209                  * with the reply state (if the send was successful, there
2210                  * would have been +1 ref for the net, which
2211                  * reply_out_callback leaves alone) */
2212                 rs->rs_on_net = 0;
2213                 ptlrpc_rs_addref(rs);
2214                 atomic_inc (&svc->srv_outstanding_replies);
2215         }
2216
2217         spin_lock(&rs->rs_lock);
2218         if (rs->rs_transno <= exp->exp_last_committed ||
2219             (!rs->rs_on_net && !rs->rs_no_ack) ||
2220              list_empty(&rs->rs_exp_list) ||     /* completed already */
2221              list_empty(&rs->rs_obd_list)) {
2222                 CDEBUG(D_HA, "Schedule reply immediately\n");
2223                 ptlrpc_dispatch_difficult_reply(rs);
2224         } else {
2225                 list_add (&rs->rs_list, &svc->srv_active_replies);
2226                 rs->rs_scheduled = 0;           /* allow notifier to schedule */
2227         }
2228         spin_unlock(&rs->rs_lock);
2229         spin_unlock(&svc->srv_lock);
2230         EXIT;
2231 }
2232
2233 int target_handle_ping(struct ptlrpc_request *req)
2234 {
2235         obd_ping(req->rq_export);
2236         return req_capsule_server_pack(&req->rq_pill);
2237 }
2238
2239 void target_committed_to_req(struct ptlrpc_request *req)
2240 {
2241         struct obd_export *exp = req->rq_export;
2242
2243         if (!exp->exp_obd->obd_no_transno && req->rq_repmsg != NULL)
2244                 lustre_msg_set_last_committed(req->rq_repmsg,
2245                                               exp->exp_last_committed);
2246         else
2247                 DEBUG_REQ(D_IOCTL, req, "not sending last_committed update (%d/"
2248                           "%d)", exp->exp_obd->obd_no_transno,
2249                           req->rq_repmsg == NULL);
2250
2251         CDEBUG(D_INFO, "last_committed "LPU64", transno "LPU64", xid "LPU64"\n",
2252                exp->exp_last_committed, req->rq_transno, req->rq_xid);
2253 }
2254 EXPORT_SYMBOL(target_committed_to_req);
2255
2256 int target_handle_qc_callback(struct ptlrpc_request *req)
2257 {
2258         struct obd_quotactl *oqctl;
2259         struct client_obd *cli = &req->rq_export->exp_obd->u.cli;
2260
2261         oqctl = req_capsule_client_get(&req->rq_pill, &RMF_OBD_QUOTACTL);
2262         if (oqctl == NULL) {
2263                 CERROR("Can't unpack obd_quotactl\n");
2264                 RETURN(-EPROTO);
2265         }
2266
2267         cli->cl_qchk_stat = oqctl->qc_stat;
2268
2269         return 0;
2270 }
2271
2272 #ifdef HAVE_QUOTA_SUPPORT
2273 int target_handle_dqacq_callback(struct ptlrpc_request *req)
2274 {
2275 #ifdef __KERNEL__
2276         struct obd_device *obd = req->rq_export->exp_obd;
2277         struct obd_device *master_obd;
2278         struct obd_device_target *obt;
2279         struct lustre_quota_ctxt *qctxt;
2280         struct qunit_data *qdata = NULL;
2281         int rc = 0;
2282         ENTRY;
2283
2284         if (OBD_FAIL_CHECK(OBD_FAIL_MDS_DROP_QUOTA_REQ))
2285                 RETURN(rc);
2286
2287         rc = req_capsule_server_pack(&req->rq_pill);
2288         if (rc) {
2289                 CERROR("packing reply failed!: rc = %d\n", rc);
2290                 RETURN(rc);
2291         }
2292
2293         LASSERT(req->rq_export);
2294
2295         qdata = quota_get_qdata(req, QUOTA_REQUEST, QUOTA_EXPORT);
2296         if (IS_ERR(qdata)) {
2297                 rc = PTR_ERR(qdata);
2298                 CDEBUG(D_ERROR, "Can't unpack qunit_data(rc: %d)\n", rc);
2299                 req->rq_status = rc;
2300                 GOTO(out, rc);
2301         }
2302
2303         /* we use the observer */
2304         if (!obd->obd_observer || !obd->obd_observer->obd_observer) {
2305                 CERROR("Can't find the observer, it is recovering\n");
2306                 req->rq_status = -EAGAIN;
2307                 GOTO(out, rc);
2308         }
2309
2310         master_obd = obd->obd_observer->obd_observer;
2311         obt = &master_obd->u.obt;
2312         qctxt = &obt->obt_qctxt;
2313
2314         if (!qctxt->lqc_setup || !qctxt->lqc_valid) {
2315                 /* quota_type has not been processed yet, return EAGAIN
2316                  * until we know whether or not quotas are supposed to
2317                  * be enabled */
2318                 CDEBUG(D_QUOTA, "quota_type not processed yet, return "
2319                        "-EAGAIN\n");
2320                 req->rq_status = -EAGAIN;
2321                 GOTO(out, rc);
2322         }
2323
2324         down_read(&obt->obt_rwsem);
2325         if (qctxt->lqc_lqs_hash == NULL) {
2326                 up_read(&obt->obt_rwsem);
2327                 /* quota_type has not been processed yet, return EAGAIN
2328                  * until we know whether or not quotas are supposed to
2329                  * be enabled */
2330                 CDEBUG(D_QUOTA, "quota_ctxt is not ready yet, return "
2331                        "-EAGAIN\n");
2332                 req->rq_status = -EAGAIN;
2333                 GOTO(out, rc);
2334         }
2335
2336         LASSERT(qctxt->lqc_handler);
2337         rc = qctxt->lqc_handler(master_obd, qdata,
2338                                 lustre_msg_get_opc(req->rq_reqmsg));
2339         up_read(&obt->obt_rwsem);
2340         if (rc && rc != -EDQUOT)
2341                 CDEBUG(rc == -EBUSY  ? D_QUOTA : D_ERROR,
2342                        "dqacq/dqrel failed! (rc:%d)\n", rc);
2343         req->rq_status = rc;
2344
2345         rc = quota_copy_qdata(req, qdata, QUOTA_REPLY, QUOTA_EXPORT);
2346         if (rc < 0) {
2347                 CERROR("Can't pack qunit_data(rc: %d)\n", rc);
2348                 GOTO(out, rc);
2349         }
2350
2351         /* Block the quota req. b=14840 */
2352         OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_BLOCK_QUOTA_REQ, obd_timeout);
2353         EXIT;
2354
2355 out:
2356         rc = ptlrpc_reply(req);
2357         return rc;
2358 #else
2359         return 0;
2360 #endif /* !__KERNEL__ */
2361 }
2362 #endif /* HAVE_QUOTA_SUPPORT */
2363
2364 /* Send a remote set_info_async.
2365  * This may go from client to server or server to client
2366  */
2367 int target_set_info_rpc(struct obd_import *imp, int opcode,
2368                         obd_count keylen, void *key,
2369                         obd_count vallen, void *val,
2370                         struct ptlrpc_request_set *set)
2371 {
2372         struct ptlrpc_request *req;
2373         char                  *tmp;
2374         int                    rc;
2375         ENTRY;
2376
2377         req = ptlrpc_request_alloc(imp, &RQF_OBD_SET_INFO);
2378         if (req == NULL)
2379                 RETURN(-ENOMEM);
2380
2381         req_capsule_set_size(&req->rq_pill, &RMF_SETINFO_KEY,
2382                              RCL_CLIENT, keylen);
2383         req_capsule_set_size(&req->rq_pill, &RMF_SETINFO_VAL,
2384                              RCL_CLIENT, vallen);
2385         rc = ptlrpc_request_pack(req, LUSTRE_OBD_VERSION, opcode);
2386         if (rc) {
2387                 ptlrpc_request_free(req);
2388                 RETURN(rc);
2389         }
2390
2391         tmp = req_capsule_client_get(&req->rq_pill, &RMF_SETINFO_KEY);
2392         memcpy(tmp, key, keylen);
2393         tmp = req_capsule_client_get(&req->rq_pill, &RMF_SETINFO_VAL);
2394         memcpy(tmp, val, vallen);
2395
2396         ptlrpc_request_set_replen(req);
2397
2398         if (set) {
2399                 ptlrpc_set_add_req(set, req);
2400                 ptlrpc_check_set(NULL, set);
2401         } else {
2402                 rc = ptlrpc_queue_wait(req);
2403                 ptlrpc_req_finished(req);
2404         }
2405
2406         RETURN(rc);
2407 }
2408 EXPORT_SYMBOL(target_set_info_rpc);
2409
2410
2411 ldlm_mode_t lck_compat_array[] = {
2412         [LCK_EX] LCK_COMPAT_EX,
2413         [LCK_PW] LCK_COMPAT_PW,
2414         [LCK_PR] LCK_COMPAT_PR,
2415         [LCK_CW] LCK_COMPAT_CW,
2416         [LCK_CR] LCK_COMPAT_CR,
2417         [LCK_NL] LCK_COMPAT_NL,
2418         [LCK_GROUP] LCK_COMPAT_GROUP,
2419         [LCK_COS] LCK_COMPAT_COS,
2420 };
2421
2422 /**
2423  * Rather arbitrary mapping from LDLM error codes to errno values. This should
2424  * not escape to the user level.
2425  */
2426 int ldlm_error2errno(ldlm_error_t error)
2427 {
2428         int result;
2429
2430         switch (error) {
2431         case ELDLM_OK:
2432                 result = 0;
2433                 break;
2434         case ELDLM_LOCK_CHANGED:
2435                 result = -ESTALE;
2436                 break;
2437         case ELDLM_LOCK_ABORTED:
2438                 result = -ENAVAIL;
2439                 break;
2440         case ELDLM_LOCK_REPLACED:
2441                 result = -ESRCH;
2442                 break;
2443         case ELDLM_NO_LOCK_DATA:
2444                 result = -ENOENT;
2445                 break;
2446         case ELDLM_NAMESPACE_EXISTS:
2447                 result = -EEXIST;
2448                 break;
2449         case ELDLM_BAD_NAMESPACE:
2450                 result = -EBADF;
2451                 break;
2452         default:
2453                 if (((int)error) < 0)  /* cast to signed type */
2454                         result = error; /* as ldlm_error_t can be unsigned */
2455                 else {
2456                         CERROR("Invalid DLM result code: %i\n", error);
2457                         result = -EPROTO;
2458                 }
2459         }
2460         return result;
2461 }
2462 EXPORT_SYMBOL(ldlm_error2errno);
2463
2464 /**
2465  * Dual to ldlm_error2errno(): maps errno values back to ldlm_error_t.
2466  */
2467 ldlm_error_t ldlm_errno2error(int err_no)
2468 {
2469         int error;
2470
2471         switch (err_no) {
2472         case 0:
2473                 error = ELDLM_OK;
2474                 break;
2475         case -ESTALE:
2476                 error = ELDLM_LOCK_CHANGED;
2477                 break;
2478         case -ENAVAIL:
2479                 error = ELDLM_LOCK_ABORTED;
2480                 break;
2481         case -ESRCH:
2482                 error = ELDLM_LOCK_REPLACED;
2483                 break;
2484         case -ENOENT:
2485                 error = ELDLM_NO_LOCK_DATA;
2486                 break;
2487         case -EEXIST:
2488                 error = ELDLM_NAMESPACE_EXISTS;
2489                 break;
2490         case -EBADF:
2491                 error = ELDLM_BAD_NAMESPACE;
2492                 break;
2493         default:
2494                 error = err_no;
2495         }
2496         return error;
2497 }
2498 EXPORT_SYMBOL(ldlm_errno2error);
2499