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
 307         /* This value may be changed at connect time in
 308            ptlrpc_connect_interpret. */
 309         cli->cl_max_pages_per_rpc = min((int)PTLRPC_MAX_BRW_PAGES,
 310                                         (int)(1024 * 1024 >> CFS_PAGE_SHIFT));
 311
 312         if (!strcmp(name, LUSTRE_MDC_NAME)) {
 313                 cli->cl_max_rpcs_in_flight = MDC_MAX_RIF_DEFAULT;
 314         } else if (num_physpages >> (20 - CFS_PAGE_SHIFT) <= 128 /* MB */) {
 315                 cli->cl_max_rpcs_in_flight = 2;
 316         } else if (num_physpages >> (20 - CFS_PAGE_SHIFT) <= 256 /* MB */) {
 317                 cli->cl_max_rpcs_in_flight = 3;
 318         } else if (num_physpages >> (20 - CFS_PAGE_SHIFT) <= 512 /* MB */) {
 319                 cli->cl_max_rpcs_in_flight = 4;
 320         } else {
 321                 cli->cl_max_rpcs_in_flight = OSC_MAX_RIF_DEFAULT;
 322         }
 323
 324         rc = ldlm_get_ref();
 325         if (rc) {
 326                 CERROR("ldlm_get_ref failed: %d\n", rc);
 327                 GOTO(err, rc);
 328         }
 329
 330         ptlrpc_init_client(rq_portal, rp_portal, name,
 331                            &obddev->obd_ldlm_client);
 332
 333         imp = class_new_import(obddev);
 334         if (imp == NULL)
 335                 GOTO(err_ldlm, rc = -ENOENT);
 336         imp->imp_client = &obddev->obd_ldlm_client;
 337         imp->imp_connect_op = connect_op;
 338         imp->imp_initial_recov = 1;
 339         imp->imp_initial_recov_bk = 0;
 340         CFS_INIT_LIST_HEAD(&imp->imp_pinger_chain);
 341         memcpy(cli->cl_target_uuid.uuid, lustre_cfg_buf(lcfg, 1),
 342                LUSTRE_CFG_BUFLEN(lcfg, 1));
 343         class_import_put(imp);
 344
 345         rc = client_import_add_conn(imp, &server_uuid, 1);
 346         if (rc) {
 347                 CERROR("can't add initial connection\n");
 348                 GOTO(err_import, rc);
 349         }
 350
 351         cli->cl_import = imp;
 352         /* cli->cl_max_mds_{easize,cookiesize} updated by mdc_init_ea_size() */
 353         cli->cl_max_mds_easize = sizeof(struct lov_mds_md_v3);
 354         cli->cl_max_mds_cookiesize = sizeof(struct llog_cookie);
 355
 356         if (LUSTRE_CFG_BUFLEN(lcfg, 3) > 0) {
 357                 if (!strcmp(lustre_cfg_string(lcfg, 3), "inactive")) {
 358                         CDEBUG(D_HA, "marking %s %s->%s as inactive\n",
 359                                name, obddev->obd_name,
 360                                cli->cl_target_uuid.uuid);
 361                         spin_lock(&imp->imp_lock);
 362                         imp->imp_deactive = 1;
 363                         spin_unlock(&imp->imp_lock);
 364                 }
 365         }
 366
 367         obddev->obd_namespace = ldlm_namespace_new(obddev, obddev->obd_name,
 368                                                    LDLM_NAMESPACE_CLIENT,
 369                                                    LDLM_NAMESPACE_GREEDY);
 370         if (obddev->obd_namespace == NULL) {
 371                 CERROR("Unable to create client namespace - %s\n",
 372                        obddev->obd_name);
 373                 GOTO(err_import, rc = -ENOMEM);
 374         }
 375
 376         cli->cl_qchk_stat = CL_NOT_QUOTACHECKED;
 377
 378         RETURN(rc);
 379
 380 err_import:
 381         class_destroy_import(imp);
 382 err_ldlm:
 383         ldlm_put_ref();
 384 err:
 385         RETURN(rc);
 386
 387 }
 388
 389 int client_obd_cleanup(struct obd_device *obddev)
 390 {
 391         ENTRY;
 392
 393         ldlm_namespace_free_post(obddev->obd_namespace);
 394         obddev->obd_namespace = NULL;
 395
 396         ldlm_put_ref();
 397         RETURN(0);
 398 }
 399
 400 /* ->o_connect() method for client side (OSC and MDC and MGC) */
 401 int client_connect_import(const struct lu_env *env,
 402                           struct obd_export **exp,
 403                           struct obd_device *obd, struct obd_uuid *cluuid,
 404                           struct obd_connect_data *data, void *localdata)
 405 {
 406         struct client_obd *cli = &obd->u.cli;
 407         struct obd_import *imp = cli->cl_import;
 408         struct obd_connect_data *ocd;
 409         struct lustre_handle conn = { 0 };
 410         int rc;
 411         ENTRY;
 412
 413         *exp = NULL;
 414         down_write(&cli->cl_sem);
 415         if (cli->cl_conn_count > 0 )
 416                 GOTO(out_sem, rc = -EALREADY);
 417
 418         rc = class_connect(&conn, obd, cluuid);
 419         if (rc)
 420                 GOTO(out_sem, rc);
 421
 422         cli->cl_conn_count++;
 423         *exp = class_conn2export(&conn);
 424
 425         LASSERT(obd->obd_namespace);
 426
 427         imp->imp_dlm_handle = conn;
 428         rc = ptlrpc_init_import(imp);
 429         if (rc != 0)
 430                 GOTO(out_ldlm, rc);
 431
 432         ocd = &imp->imp_connect_data;
 433         if (data) {
 434                 *ocd = *data;
 435                 imp->imp_connect_flags_orig = data->ocd_connect_flags;
 436         }
 437
 438         rc = ptlrpc_connect_import(imp, NULL);
 439         if (rc != 0) {
 440                 LASSERT (imp->imp_state == LUSTRE_IMP_DISCON);
 441                 GOTO(out_ldlm, rc);
 442         }
 443         LASSERT((*exp)->exp_connection);
 444
 445         if (data) {
 446                 LASSERTF((ocd->ocd_connect_flags & data->ocd_connect_flags) ==
 447                          ocd->ocd_connect_flags, "old "LPX64", new "LPX64"\n",
 448                          data->ocd_connect_flags, ocd->ocd_connect_flags);
 449                 data->ocd_connect_flags = ocd->ocd_connect_flags;
 450         }
 451
 452         ptlrpc_pinger_add_import(imp);
 453
 454         EXIT;
 455
 456         if (rc) {
 457 out_ldlm:
 458                 cli->cl_conn_count--;
 459                 class_disconnect(*exp);
 460                 *exp = NULL;
 461         }
 462 out_sem:
 463         up_write(&cli->cl_sem);
 464
 465         return rc;
 466 }
 467
 468 int client_disconnect_export(struct obd_export *exp)
 469 {
 470         struct obd_device *obd = class_exp2obd(exp);
 471         struct client_obd *cli;
 472         struct obd_import *imp;
 473         int rc = 0, err;
 474         ENTRY;
 475
 476         if (!obd) {
 477                 CERROR("invalid export for disconnect: exp %p cookie "LPX64"\n",
 478                        exp, exp ? exp->exp_handle.h_cookie : -1);
 479                 RETURN(-EINVAL);
 480         }
 481
 482         cli = &obd->u.cli;
 483         imp = cli->cl_import;
 484
 485         down_write(&cli->cl_sem);
 486         CDEBUG(D_INFO, "disconnect %s - %d\n", obd->obd_name,
 487                cli->cl_conn_count);
 488
 489         if (!cli->cl_conn_count) {
 490                 CERROR("disconnecting disconnected device (%s)\n",
 491                        obd->obd_name);
 492                 GOTO(out_disconnect, rc = -EINVAL);
 493         }
 494
 495         cli->cl_conn_count--;
 496         if (cli->cl_conn_count)
 497                 GOTO(out_disconnect, rc = 0);
 498
 499         /* Mark import deactivated now, so we don't try to reconnect if any
 500          * of the cleanup RPCs fails (e.g. ldlm cancel, etc).  We don't
 501          * fully deactivate the import, or that would drop all requests. */
 502         spin_lock(&imp->imp_lock);
 503         imp->imp_deactive = 1;
 504         spin_unlock(&imp->imp_lock);
 505
 506         /* Some non-replayable imports (MDS's OSCs) are pinged, so just
 507          * delete it regardless.  (It's safe to delete an import that was
 508          * never added.) */
 509         (void)ptlrpc_pinger_del_import(imp);
 510
 511         if (obd->obd_namespace != NULL) {
 512                 /* obd_force == local only */
 513                 ldlm_cli_cancel_unused(obd->obd_namespace, NULL,
 514                                        obd->obd_force ? LDLM_FL_LOCAL_ONLY:0,
 515                                        NULL);
 516                 ldlm_namespace_free_prior(obd->obd_namespace, imp, obd->obd_force);
 517         }
 518
 519         /*
 520          * there's no necessary to hold sem during diconnecting an import,
 521          * and actually it may cause deadlock in gss.
 522          */
 523         up_write(&cli->cl_sem);
 524         rc = ptlrpc_disconnect_import(imp, 0);
 525         down_write(&cli->cl_sem);
 526
 527         ptlrpc_invalidate_import(imp);
 528
 529         if (imp->imp_rq_pool) {
 530                 ptlrpc_free_rq_pool(imp->imp_rq_pool);
 531                 imp->imp_rq_pool = NULL;
 532         }
 533         client_destroy_import(imp);
 534         cli->cl_import = NULL;
 535
 536         EXIT;
 537
 538  out_disconnect:
 539         /* use server style - class_disconnect should be always called for
 540          * o_disconnect */
 541         err = class_disconnect(exp);
 542         if (!rc && err)
 543                 rc = err;
 544
 545         up_write(&cli->cl_sem);
 546
 547         RETURN(rc);
 548 }
 549
 550 /* --------------------------------------------------------------------------
 551  * from old lib/target.c
 552  * -------------------------------------------------------------------------- */
 553
 554 static int target_handle_reconnect(struct lustre_handle *conn,
 555                                    struct obd_export *exp,
 556                                    struct obd_uuid *cluuid)
 557 {
 558         ENTRY;
 559         if (exp->exp_connection && exp->exp_imp_reverse) {
 560                 struct lustre_handle *hdl;
 561                 hdl = &exp->exp_imp_reverse->imp_remote_handle;
 562                 /* Might be a re-connect after a partition. */
 563                 if (!memcmp(&conn->cookie, &hdl->cookie, sizeof conn->cookie)) {
 564                         CWARN("%s: %s reconnecting\n", exp->exp_obd->obd_name,
 565                               cluuid->uuid);
 566                         conn->cookie = exp->exp_handle.h_cookie;
 567                         /* target_handle_connect() treats EALREADY and
 568                          * -EALREADY differently.  EALREADY means we are
 569                          * doing a valid reconnect from the same client. */
 570                         RETURN(EALREADY);
 571                 } else {
 572                         CERROR("%s reconnecting from %s, "
 573                                "handle mismatch (ours "LPX64", theirs "
 574                                LPX64")\n", cluuid->uuid,
 575                                exp->exp_connection->c_remote_uuid.uuid,
 576                                hdl->cookie, conn->cookie);
 577                         memset(conn, 0, sizeof *conn);
 578                         /* target_handle_connect() treats EALREADY and
 579                          * -EALREADY differently.  -EALREADY is an error
 580                          * (same UUID, different handle). */
 581                         RETURN(-EALREADY);
 582                 }
 583         }
 584
 585         conn->cookie = exp->exp_handle.h_cookie;
 586         CDEBUG(D_HA, "connect export for UUID '%s' at %p, cookie "LPX64"\n",
 587                cluuid->uuid, exp, conn->cookie);
 588         RETURN(0);
 589 }
 590
 591 void target_client_add_cb(struct obd_device *obd, __u64 transno, void *cb_data,
 592                           int error)
 593 {
 594         struct obd_export *exp = cb_data;
 595
 596         CDEBUG(D_RPCTRACE, "%s: committing for initial connect of %s\n",
 597                obd->obd_name, exp->exp_client_uuid.uuid);
 598
 599         spin_lock(&exp->exp_lock);
 600         exp->exp_need_sync = 0;
 601         spin_unlock(&exp->exp_lock);
 602 }
 603 EXPORT_SYMBOL(target_client_add_cb);
 604 static void
 605 target_start_and_reset_recovery_timer(struct obd_device *obd,
 606                                       struct ptlrpc_request *req,
 607                                       int new_client);
 608
 609 int target_handle_connect(struct ptlrpc_request *req)
 610 {
 611         struct obd_device *target, *targref = NULL;
 612         struct obd_export *export = NULL;
 613         struct obd_import *revimp;
 614         struct lustre_handle conn;
 615         struct lustre_handle *tmp;
 616         struct obd_uuid tgtuuid;
 617         struct obd_uuid cluuid;
 618         struct obd_uuid remote_uuid;
 619         char *str;
 620         int rc = 0;
 621         int mds_conn = 0;
 622         struct obd_connect_data *data, *tmpdata;
 623         lnet_nid_t *client_nid = NULL;
 624         ENTRY;
 625
 626         OBD_RACE(OBD_FAIL_TGT_CONN_RACE);
 627
 628         str = req_capsule_client_get(&req->rq_pill, &RMF_TGTUUID);
 629         if (str == NULL) {
 630                 DEBUG_REQ(D_ERROR, req, "bad target UUID for connect");
 631                 GOTO(out, rc = -EINVAL);
 632         }
 633
 634         obd_str2uuid(&tgtuuid, str);
 635         target = class_uuid2obd(&tgtuuid);
 636         if (!target)
 637                 target = class_name2obd(str);
 638
 639         if (!target || target->obd_stopping || !target->obd_set_up) {
 640                 LCONSOLE_ERROR_MSG(0x137, "UUID '%s' is not available "
 641                                    " for connect (%s)\n", str,
 642                                    !target ? "no target" :
 643                                    (target->obd_stopping ? "stopping" :
 644                                    "not set up"));
 645                 GOTO(out, rc = -ENODEV);
 646         }
 647
 648         if (target->obd_no_conn) {
 649                 LCONSOLE_WARN("%s: temporarily refusing client connection "
 650                               "from %s\n", target->obd_name,
 651                               libcfs_nid2str(req->rq_peer.nid));
 652                 GOTO(out, rc = -EAGAIN);
 653         }
 654
 655         /* Make sure the target isn't cleaned up while we're here. Yes,
 656            there's still a race between the above check and our incref here.
 657            Really, class_uuid2obd should take the ref. */
 658         targref = class_incref(target, __FUNCTION__, cfs_current());
 659
 660
 661         str = req_capsule_client_get(&req->rq_pill, &RMF_CLUUID);
 662         if (str == NULL) {
 663                 DEBUG_REQ(D_ERROR, req, "bad client UUID for connect");
 664                 GOTO(out, rc = -EINVAL);
 665         }
 666
 667         obd_str2uuid(&cluuid, str);
 668
 669         /* XXX extract a nettype and format accordingly */
 670         switch (sizeof(lnet_nid_t)) {
 671                 /* NB the casts only avoid compiler warnings */
 672         case 8:
 673                 snprintf(remote_uuid.uuid, sizeof remote_uuid,
 674                          "NET_"LPX64"_UUID", (__u64)req->rq_peer.nid);
 675                 break;
 676         case 4:
 677                 snprintf(remote_uuid.uuid, sizeof remote_uuid,
 678                          "NET_%x_UUID", (__u32)req->rq_peer.nid);
 679                 break;
 680         default:
 681                 LBUG();
 682         }
 683
 684         tmp = req_capsule_client_get(&req->rq_pill, &RMF_CONN);
 685         if (tmp == NULL)
 686                 GOTO(out, rc = -EPROTO);
 687
 688         conn = *tmp;
 689
 690         data = req_capsule_client_get(&req->rq_pill, &RMF_CONNECT_DATA);
 691         if (!data)
 692                 GOTO(out, rc = -EPROTO);
 693
 694         rc = req_capsule_server_pack(&req->rq_pill);
 695         if (rc)
 696                 GOTO(out, rc);
 697
 698         if (lustre_msg_get_op_flags(req->rq_reqmsg) & MSG_CONNECT_LIBCLIENT) {
 699                 if (!data) {
 700                         DEBUG_REQ(D_WARNING, req, "Refusing old (unversioned) "
 701                                   "libclient connection attempt");
 702                         GOTO(out, rc = -EPROTO);
 703                 } else if (data->ocd_version < LUSTRE_VERSION_CODE -
 704                                                LUSTRE_VERSION_ALLOWED_OFFSET ||
 705                            data->ocd_version > LUSTRE_VERSION_CODE +
 706                                                LUSTRE_VERSION_ALLOWED_OFFSET) {
 707                         DEBUG_REQ(D_WARNING, req, "Refusing %s (%d.%d.%d.%d) "
 708                                   "libclient connection attempt",
 709                                   data->ocd_version < LUSTRE_VERSION_CODE ?
 710                                   "old" : "new",
 711                                   OBD_OCD_VERSION_MAJOR(data->ocd_version),
 712                                   OBD_OCD_VERSION_MINOR(data->ocd_version),
 713                                   OBD_OCD_VERSION_PATCH(data->ocd_version),
 714                                   OBD_OCD_VERSION_FIX(data->ocd_version));
 715                         data = req_capsule_server_sized_get(&req->rq_pill,
 716                                                         &RMF_CONNECT_DATA,
 717                                     offsetof(typeof(*data), ocd_version) +
 718                                              sizeof(data->ocd_version));
 719                         if (data) {
 720                                 data->ocd_connect_flags = OBD_CONNECT_VERSION;
 721                                 data->ocd_version = LUSTRE_VERSION_CODE;
 722                         }
 723                         GOTO(out, rc = -EPROTO);
 724                 }
 725         }
 726
 727         if ((lustre_msg_get_op_flags(req->rq_reqmsg) & MSG_CONNECT_INITIAL) &&
 728             (data->ocd_connect_flags & OBD_CONNECT_MDS))
 729                 mds_conn = 1;
 730
 731         /* lctl gets a backstage, all-access pass. */
 732         if (obd_uuid_equals(&cluuid, &target->obd_uuid))
 733                 goto dont_check_exports;
 734
 735         export = lustre_hash_lookup(target->obd_uuid_hash, &cluuid);
 736         if (!export)
 737                 goto no_export;
 738
 739         /* we've found an export in the hash */
 740         if (export->exp_connecting) { /* bug 9635, et. al. */
 741                 CWARN("%s: exp %p already connecting\n",
 742                       export->exp_obd->obd_name, export);
 743                 class_export_put(export);
 744                 export = NULL;
 745                 rc = -EALREADY;
 746         } else if (mds_conn && export->exp_connection) {
 747                 if (req->rq_peer.nid != export->exp_connection->c_peer.nid)
 748                         /* mds reconnected after failover */
 749                         CWARN("%s: received MDS connection from NID %s,"
 750                               " removing former export from NID %s\n",
 751                             target->obd_name, libcfs_nid2str(req->rq_peer.nid),
 752                             libcfs_nid2str(export->exp_connection->c_peer.nid));
 753                 else
 754                         /* new mds connection from the same nid */
 755                         CWARN("%s: received new MDS connection from NID %s,"
 756                               " removing former export from same NID\n",
 757                             target->obd_name, libcfs_nid2str(req->rq_peer.nid));
 758                 class_fail_export(export);
 759                 class_export_put(export);
 760                 export = NULL;
 761                 rc = 0;
 762         } else if (export->exp_connection != NULL &&
 763                    req->rq_peer.nid != export->exp_connection->c_peer.nid &&
 764                    (lustre_msg_get_op_flags(req->rq_reqmsg) &
 765                     MSG_CONNECT_INITIAL)) {
 766                 /* in mds failover we have static uuid but nid can be
 767                  * changed*/
 768                 CWARN("%s: cookie %s seen on new NID %s when "
 769                       "existing NID %s is already connected\n",
 770                       target->obd_name, cluuid.uuid,
 771                       libcfs_nid2str(req->rq_peer.nid),
 772                       libcfs_nid2str(export->exp_connection->c_peer.nid));
 773                 rc = -EALREADY;
 774                 class_export_put(export);
 775                 export = NULL;
 776         } else {
 777                 spin_lock(&export->exp_lock);
 778                 export->exp_connecting = 1;
 779                 spin_unlock(&export->exp_lock);
 780                 class_export_put(export);
 781                 LASSERT(export->exp_obd == target);
 782
 783                 rc = target_handle_reconnect(&conn, export, &cluuid);
 784         }
 785
 786         /* If we found an export, we already unlocked. */
 787         if (!export) {
 788 no_export:
 789                 OBD_FAIL_TIMEOUT(OBD_FAIL_TGT_DELAY_CONNECT, 2 * obd_timeout);
 790         } else if (req->rq_export == NULL &&
 791                    atomic_read(&export->exp_rpc_count) > 0) {
 792                 CWARN("%s: refuse connection from %s/%s to 0x%p/%d\n",
 793                       target->obd_name, cluuid.uuid,
 794                       libcfs_nid2str(req->rq_peer.nid),
 795                       export, atomic_read(&export->exp_refcount));
 796                 GOTO(out, rc = -EBUSY);
 797         } else if (req->rq_export != NULL &&
 798                    (atomic_read(&export->exp_rpc_count) > 1)) {
 799                 /* the current connect rpc has increased exp_rpc_count */
 800                 CWARN("%s: refuse reconnection from %s@%s to 0x%p/%d\n",
 801                       target->obd_name, cluuid.uuid,
 802                       libcfs_nid2str(req->rq_peer.nid),
 803                       export, atomic_read(&export->exp_rpc_count) - 1);
 804                 spin_lock(&export->exp_lock);
 805                 if (req->rq_export->exp_conn_cnt <
 806                     lustre_msg_get_conn_cnt(req->rq_reqmsg))
 807                         /* try to abort active requests */
 808                         req->rq_export->exp_abort_active_req = 1;
 809                 spin_unlock(&export->exp_lock);
 810                 GOTO(out, rc = -EBUSY);
 811         } else if (lustre_msg_get_conn_cnt(req->rq_reqmsg) == 1) {
 812                 CERROR("%s: NID %s (%s) reconnected with 1 conn_cnt; "
 813                        "cookies not random?\n", target->obd_name,
 814                        libcfs_nid2str(req->rq_peer.nid), cluuid.uuid);
 815                 GOTO(out, rc = -EALREADY);
 816         } else {
 817                 OBD_FAIL_TIMEOUT(OBD_FAIL_TGT_DELAY_RECONNECT, 2 * obd_timeout);
 818         }
 819
 820         if (rc < 0) {
 821                 GOTO(out, rc);
 822         }
 823
 824         CWARN("%s: connection from %s@%s %st"LPU64" exp %p cur %ld last %ld\n",
 825                target->obd_name, cluuid.uuid, libcfs_nid2str(req->rq_peer.nid),
 826               target->obd_recovering ? "recovering/" : "", data->ocd_transno,
 827               export, (long)cfs_time_current_sec(),
 828               export ? (long)export->exp_last_request_time : 0);
 829
 830         /* Tell the client if we're in recovery. */
 831         if (target->obd_recovering) {
 832                 lustre_msg_add_op_flags(req->rq_repmsg, MSG_CONNECT_RECOVERING);
 833                 /* If this is the first time a client connects,
 834                    reset the recovery timer */
 835                 if (rc == 0)
 836                         target_start_and_reset_recovery_timer(target, req,
 837                                                               !export);
 838         }
 839
 840         /* We want to handle EALREADY but *not* -EALREADY from
 841          * target_handle_reconnect(), return reconnection state in a flag */
 842         if (rc == EALREADY) {
 843                 lustre_msg_add_op_flags(req->rq_repmsg, MSG_CONNECT_RECONNECT);
 844                 rc = 0;
 845         } else {
 846                 LASSERT(rc == 0);
 847         }
 848
 849         /* Tell the client if we support replayable requests */
 850         if (target->obd_replayable)
 851                 lustre_msg_add_op_flags(req->rq_repmsg, MSG_CONNECT_REPLAYABLE);
 852         client_nid = &req->rq_peer.nid;
 853
 854         if (export == NULL) {
 855                 if (target->obd_recovering) {
 856                         cfs_time_t t;
 857
 858                         t = cfs_timer_deadline(&target->obd_recovery_timer);
 859                         t = cfs_time_sub(t, cfs_time_current());
 860                         CERROR("%s: denying connection for new client %s (%s): "
 861                                "%d clients in recovery for "CFS_TIME_T"s\n",
 862                                target->obd_name,
 863                                libcfs_nid2str(req->rq_peer.nid), cluuid.uuid,
 864                                atomic_read(&target->obd_lock_replay_clients),
 865                                cfs_duration_sec(t));
 866                         rc = -EBUSY;
 867                 } else {
 868 dont_check_exports:
 869                         rc = obd_connect(req->rq_svc_thread->t_env,
 870                                          &export, target, &cluuid, data,
 871                                          client_nid);
 872                         if (rc == 0)
 873                                 conn.cookie = export->exp_handle.h_cookie;
 874                 }
 875         } else {
 876                 rc = obd_reconnect(req->rq_svc_thread->t_env,
 877                                    export, target, &cluuid, data, client_nid);
 878                 if (rc == 0)
 879                         /* prevous done via class_conn2export */
 880                         class_export_get(export);
 881         }
 882         if (rc)
 883                 GOTO(out, rc);
 884         /* Return only the parts of obd_connect_data that we understand, so the
 885          * client knows that we don't understand the rest. */
 886         if (data) {
 887                  tmpdata = req_capsule_server_get(&req->rq_pill,
 888                                                   &RMF_CONNECT_DATA);
 889                   //data->ocd_connect_flags &= OBD_CONNECT_SUPPORTED;
 890                  *tmpdata = *data;
 891         }
 892
 893         /* If all else goes well, this is our RPC return code. */
 894         req->rq_status = 0;
 895
 896         lustre_msg_set_handle(req->rq_repmsg, &conn);
 897
 898         /* If the client and the server are the same node, we will already
 899          * have an export that really points to the client's DLM export,
 900          * because we have a shared handles table.
 901          *
 902          * XXX this will go away when shaver stops sending the "connect" handle
 903          * in the real "remote handle" field of the request --phik 24 Apr 2003
 904          */
 905         if (req->rq_export != NULL)
 906                 class_export_put(req->rq_export);
 907
 908         req->rq_export = export;
 909
 910         spin_lock(&export->exp_lock);
 911         if (export->exp_conn_cnt >= lustre_msg_get_conn_cnt(req->rq_reqmsg)) {
 912                 spin_unlock(&export->exp_lock);
 913                 CERROR("%s: %s already connected at higher conn_cnt: %d > %d\n",
 914                        cluuid.uuid, libcfs_nid2str(req->rq_peer.nid),
 915                        export->exp_conn_cnt,
 916                        lustre_msg_get_conn_cnt(req->rq_reqmsg));
 917
 918                 GOTO(out, rc = -EALREADY);
 919         }
 920         export->exp_conn_cnt = lustre_msg_get_conn_cnt(req->rq_reqmsg);
 921         export->exp_abort_active_req = 0;
 922
 923         /* request from liblustre?  Don't evict it for not pinging. */
 924         if (lustre_msg_get_op_flags(req->rq_reqmsg) & MSG_CONNECT_LIBCLIENT) {
 925                 export->exp_libclient = 1;
 926                 spin_unlock(&export->exp_lock);
 927
 928                 spin_lock(&target->obd_dev_lock);
 929                 list_del_init(&export->exp_obd_chain_timed);
 930                 spin_unlock(&target->obd_dev_lock);
 931         } else {
 932                 spin_unlock(&export->exp_lock);
 933         }
 934
 935         if (export->exp_connection != NULL) {
 936                 /* Check to see if connection came from another NID */
 937                 if ((export->exp_connection->c_peer.nid != req->rq_peer.nid) &&
 938                     !hlist_unhashed(&export->exp_nid_hash))
 939                         lustre_hash_del(export->exp_obd->obd_nid_hash,
 940                                         &export->exp_connection->c_peer.nid,
 941                                         &export->exp_nid_hash);
 942
 943                 ptlrpc_connection_put(export->exp_connection);
 944         }
 945
 946         export->exp_connection = ptlrpc_connection_get(req->rq_peer,
 947                                                        req->rq_self,
 948                                                        &remote_uuid);
 949         if (hlist_unhashed(&export->exp_nid_hash)) {
 950                 lustre_hash_add_unique(export->exp_obd->obd_nid_hash,
 951                                        &export->exp_connection->c_peer.nid,
 952                                        &export->exp_nid_hash);
 953         }
 954
 955         spin_lock_bh(&target->obd_processing_task_lock);
 956         if (target->obd_recovering && !export->exp_in_recovery) {
 957                 spin_lock(&export->exp_lock);
 958                 export->exp_in_recovery = 1;
 959                 export->exp_req_replay_needed = 1;
 960                 export->exp_lock_replay_needed = 1;
 961                 spin_unlock(&export->exp_lock);
 962                 if ((lustre_msg_get_op_flags(req->rq_reqmsg) & MSG_CONNECT_TRANSNO)
 963                      && (data->ocd_transno == 0))
 964                         CWARN("Connect with zero transno!\n");
 965
 966                 if ((lustre_msg_get_op_flags(req->rq_reqmsg) & MSG_CONNECT_TRANSNO)
 967                      && data->ocd_transno < target->obd_next_recovery_transno)
 968                         target->obd_next_recovery_transno = data->ocd_transno;
 969                 target->obd_connected_clients++;
 970                 atomic_inc(&target->obd_req_replay_clients);
 971                 atomic_inc(&target->obd_lock_replay_clients);
 972                 if (target->obd_connected_clients ==
 973                     target->obd_max_recoverable_clients)
 974                         cfs_waitq_signal(&target->obd_next_transno_waitq);
 975         }
 976         spin_unlock_bh(&target->obd_processing_task_lock);
 977         tmp = req_capsule_client_get(&req->rq_pill, &RMF_CONN);
 978         conn = *tmp;
 979
 980         if (export->exp_imp_reverse != NULL) {
 981                 /* destroyed import can be still referenced in ctxt */
 982                 obd_set_info_async(export, sizeof(KEY_REVIMP_UPD),
 983                                    KEY_REVIMP_UPD, 0, NULL, NULL);
 984
 985                 client_destroy_import(export->exp_imp_reverse);
 986         }
 987
 988         /* for the rest part, we return -ENOTCONN in case of errors
 989          * in order to let client initialize connection again.
 990          */
 991         revimp = export->exp_imp_reverse = class_new_import(target);
 992         if (!revimp) {
 993                 CERROR("fail to alloc new reverse import.\n");
 994                 GOTO(out, rc = -ENOTCONN);
 995         }
 996
 997         revimp->imp_connection = ptlrpc_connection_addref(export->exp_connection);
 998         revimp->imp_client = &export->exp_obd->obd_ldlm_client;
 999         revimp->imp_remote_handle = conn;
1000         revimp->imp_dlm_fake = 1;
1001         revimp->imp_state = LUSTRE_IMP_FULL;
1002
1003         /* unknown versions will be caught in
1004          * ptlrpc_handle_server_req_in->lustre_unpack_msg() */
1005         revimp->imp_msg_magic = req->rq_reqmsg->lm_magic;
1006
1007         if ((export->exp_connect_flags & OBD_CONNECT_AT) &&
1008             (revimp->imp_msg_magic != LUSTRE_MSG_MAGIC_V1))
1009                 revimp->imp_msghdr_flags |= MSGHDR_AT_SUPPORT;
1010         else
1011                 revimp->imp_msghdr_flags &= ~MSGHDR_AT_SUPPORT;
1012
1013         rc = sptlrpc_import_sec_adapt(revimp, req->rq_svc_ctx, &req->rq_flvr);
1014         if (rc) {
1015                 CERROR("Failed to get sec for reverse import: %d\n", rc);
1016                 export->exp_imp_reverse = NULL;
1017                 class_destroy_import(revimp);
1018         }
1019
1020         class_import_put(revimp);
1021 out:
1022         if (export) {
1023                 spin_lock(&export->exp_lock);
1024                 export->exp_connecting = 0;
1025                 spin_unlock(&export->exp_lock);
1026         }
1027         if (targref)
1028                 class_decref(targref, __FUNCTION__, cfs_current());
1029         if (rc)
1030                 req->rq_status = rc;
1031         RETURN(rc);
1032 }
1033
1034 int target_handle_disconnect(struct ptlrpc_request *req)
1035 {
1036         int rc;
1037         ENTRY;
1038
1039         rc = req_capsule_server_pack(&req->rq_pill);
1040         if (rc)
1041                 RETURN(rc);
1042
1043         /* keep the rq_export around so we can send the reply */
1044         req->rq_status = obd_disconnect(class_export_get(req->rq_export));
1045
1046         RETURN(0);
1047 }
1048
1049 void target_destroy_export(struct obd_export *exp)
1050 {
1051         /* exports created from last_rcvd data, and "fake"
1052            exports created by lctl don't have an import */
1053         if (exp->exp_imp_reverse != NULL)
1054                 client_destroy_import(exp->exp_imp_reverse);
1055
1056         LASSERT(atomic_read(&exp->exp_locks_count) == 0);
1057         LASSERT(atomic_read(&exp->exp_rpc_count) == 0);
1058         LASSERT(atomic_read(&exp->exp_cb_count) == 0);
1059         LASSERT(atomic_read(&exp->exp_replay_count) == 0);
1060 }
1061
1062 /*
1063  * Recovery functions
1064  */
1065
1066 struct ptlrpc_request *ptlrpc_clone_req( struct ptlrpc_request *orig_req)
1067 {
1068         struct ptlrpc_request *copy_req;
1069         struct lustre_msg *copy_reqmsg;
1070         struct ptlrpc_user_desc *udesc = NULL;
1071
1072         OBD_ALLOC_PTR(copy_req);
1073         if (!copy_req)
1074                 return NULL;
1075         OBD_ALLOC(copy_reqmsg, orig_req->rq_reqlen);
1076         if (!copy_reqmsg){
1077                 OBD_FREE_PTR(copy_req);
1078                 return NULL;
1079         }
1080
1081         if (orig_req->rq_user_desc) {
1082                 int ngroups = orig_req->rq_user_desc->pud_ngroups;
1083
1084                 OBD_ALLOC(udesc, sptlrpc_user_desc_size(ngroups));
1085                 if (!udesc) {
1086                         OBD_FREE(copy_reqmsg, orig_req->rq_reqlen);
1087                         OBD_FREE_PTR(copy_req);
1088                         return NULL;
1089                 }
1090                 memcpy(udesc, orig_req->rq_user_desc,
1091                        sptlrpc_user_desc_size(ngroups));
1092         }
1093
1094         *copy_req = *orig_req;
1095         memcpy(copy_reqmsg, orig_req->rq_reqmsg, orig_req->rq_reqlen);
1096         copy_req->rq_reqmsg = copy_reqmsg;
1097         copy_req->rq_user_desc = udesc;
1098
1099         class_export_rpc_get(copy_req->rq_export);
1100         CFS_INIT_LIST_HEAD(&copy_req->rq_list);
1101         CFS_INIT_LIST_HEAD(&copy_req->rq_replay_list);
1102         sptlrpc_svc_ctx_addref(copy_req);
1103
1104         if (copy_req->rq_reply_state) {
1105                 /* the copied req takes over the reply state */
1106                 orig_req->rq_reply_state = NULL;
1107                 /* to catch further access */
1108                 orig_req->rq_repmsg = NULL;
1109                 orig_req->rq_replen = 0;
1110         }
1111         /** let export know it has replays to be handled */
1112         atomic_inc(&copy_req->rq_export->exp_replay_count);
1113         return copy_req;
1114 }
1115
1116 void ptlrpc_free_clone(struct ptlrpc_request *req)
1117 {
1118         LASSERT(list_empty(&req->rq_replay_list));
1119
1120         ptlrpc_req_drop_rs(req);
1121         sptlrpc_svc_ctx_decref(req);
1122         LASSERT(atomic_read(&req->rq_export->exp_replay_count) > 0);
1123         atomic_dec(&req->rq_export->exp_replay_count);
1124         class_export_rpc_put(req->rq_export);
1125         list_del_init(&req->rq_list);
1126
1127         if (req->rq_user_desc) {
1128                 int ngroups = req->rq_user_desc->pud_ngroups;
1129                 OBD_FREE(req->rq_user_desc, sptlrpc_user_desc_size(ngroups));
1130         }
1131         OBD_FREE(req->rq_reqmsg, req->rq_reqlen);
1132         OBD_FREE_PTR(req);
1133 }
1134
1135 static int target_exp_enqueue_req_replay(struct ptlrpc_request *req)
1136 {
1137         __u64                  transno = lustre_msg_get_transno(req->rq_reqmsg);
1138         struct obd_export     *exp = req->rq_export;
1139         struct ptlrpc_request *reqiter;
1140         int                    dup = 0;
1141
1142         LASSERT(exp);
1143
1144         spin_lock(&exp->exp_lock);
1145         list_for_each_entry(reqiter, &exp->exp_req_replay_queue,
1146                             rq_replay_list) {
1147                 if (lustre_msg_get_transno(reqiter->rq_reqmsg) == transno) {
1148                         dup = 1;
1149                         break;
1150                 }
1151         }
1152
1153         if (dup) {
1154                 /* we expect it with RESENT and REPLAY flags */
1155                 if ((lustre_msg_get_flags(req->rq_reqmsg) &
1156                      (MSG_RESENT | MSG_REPLAY)) != (MSG_RESENT | MSG_REPLAY))
1157                         CERROR("invalid flags %x of resent replay\n",
1158                                lustre_msg_get_flags(req->rq_reqmsg));
1159         } else {
1160                 list_add_tail(&req->rq_replay_list, &exp->exp_req_replay_queue);
1161         }
1162
1163         spin_unlock(&exp->exp_lock);
1164         return dup;
1165 }
1166
1167 static void target_exp_dequeue_req_replay(struct ptlrpc_request *req)
1168 {
1169         LASSERT(!list_empty(&req->rq_replay_list));
1170         LASSERT(req->rq_export);
1171
1172         spin_lock(&req->rq_export->exp_lock);
1173         list_del_init(&req->rq_replay_list);
1174         spin_unlock(&req->rq_export->exp_lock);
1175 }
1176
1177 #ifdef __KERNEL__
1178 static void target_finish_recovery(struct obd_device *obd)
1179 {
1180         ENTRY;
1181         LCONSOLE_INFO("%s: sending delayed replies to recovered clients\n",
1182                       obd->obd_name);
1183
1184         ldlm_reprocess_all_ns(obd->obd_namespace);
1185         spin_lock_bh(&obd->obd_processing_task_lock);
1186         if (!list_empty(&obd->obd_req_replay_queue) ||
1187             !list_empty(&obd->obd_lock_replay_queue) ||
1188             !list_empty(&obd->obd_final_req_queue)) {
1189                 CERROR("%s: Recovery queues ( %s%s%s) are not empty\n",
1190                        obd->obd_name,
1191                        list_empty(&obd->obd_req_replay_queue) ? "" : "req ",
1192                        list_empty(&obd->obd_lock_replay_queue) ? "" : "lock ",
1193                        list_empty(&obd->obd_final_req_queue) ? "" : "final ");
1194                 spin_unlock_bh(&obd->obd_processing_task_lock);
1195                 LBUG();
1196         }
1197         spin_unlock_bh(&obd->obd_processing_task_lock);
1198
1199         obd->obd_recovery_end = cfs_time_current_sec();
1200
1201         /* when recovery finished, cleanup orphans on mds and ost */
1202         if (OBT(obd) && OBP(obd, postrecov)) {
1203                 int rc = OBP(obd, postrecov)(obd);
1204                 if (rc < 0)
1205                         LCONSOLE_WARN("%s: Post recovery failed, rc %d\n",
1206                                       obd->obd_name, rc);
1207         }
1208         EXIT;
1209 }
1210
1211 static void abort_req_replay_queue(struct obd_device *obd)
1212 {
1213         struct ptlrpc_request *req, *n;
1214         struct list_head abort_list;
1215
1216         CFS_INIT_LIST_HEAD(&abort_list);
1217         spin_lock_bh(&obd->obd_processing_task_lock);
1218         list_splice_init(&obd->obd_req_replay_queue, &abort_list);
1219         spin_unlock_bh(&obd->obd_processing_task_lock);
1220         list_for_each_entry_safe(req, n, &abort_list, rq_list) {
1221                 DEBUG_REQ(D_WARNING, req, "aborted:");
1222                 req->rq_status = -ENOTCONN;
1223                 if (ptlrpc_error(req)) {
1224                         DEBUG_REQ(D_ERROR, req,
1225                                   "failed abort_req_reply; skipping");
1226                 }
1227                 target_exp_dequeue_req_replay(req);
1228                 ptlrpc_free_clone(req);
1229         }
1230 }
1231
1232 static void abort_lock_replay_queue(struct obd_device *obd)
1233 {
1234         struct ptlrpc_request *req, *n;
1235         struct list_head abort_list;
1236
1237         CFS_INIT_LIST_HEAD(&abort_list);
1238         spin_lock_bh(&obd->obd_processing_task_lock);
1239         list_splice_init(&obd->obd_lock_replay_queue, &abort_list);
1240         spin_unlock_bh(&obd->obd_processing_task_lock);
1241         list_for_each_entry_safe(req, n, &abort_list, rq_list){
1242                 DEBUG_REQ(D_ERROR, req, "aborted:");
1243                 req->rq_status = -ENOTCONN;
1244                 if (ptlrpc_error(req)) {
1245                         DEBUG_REQ(D_ERROR, req,
1246                                   "failed abort_lock_reply; skipping");
1247                 }
1248                 ptlrpc_free_clone(req);
1249         }
1250 }
1251 #endif
1252
1253 /* Called from a cleanup function if the device is being cleaned up
1254    forcefully.  The exports should all have been disconnected already,
1255    the only thing left to do is
1256      - clear the recovery flags
1257      - cancel the timer
1258      - free queued requests and replies, but don't send replies
1259    Because the obd_stopping flag is set, no new requests should be received.
1260
1261 */
1262 void target_cleanup_recovery(struct obd_device *obd)
1263 {
1264         struct ptlrpc_request *req, *n;
1265         struct list_head clean_list;
1266         ENTRY;
1267
1268         LASSERT(obd->obd_stopping);
1269
1270         CFS_INIT_LIST_HEAD(&clean_list);
1271         spin_lock_bh(&obd->obd_processing_task_lock);
1272         if (!obd->obd_recovering) {
1273                 spin_unlock_bh(&obd->obd_processing_task_lock);
1274                 EXIT;
1275                 return;
1276         }
1277         obd->obd_recovering = obd->obd_abort_recovery = 0;
1278         target_cancel_recovery_timer(obd);
1279
1280         list_splice_init(&obd->obd_req_replay_queue, &clean_list);
1281         spin_unlock_bh(&obd->obd_processing_task_lock);
1282
1283         list_for_each_entry_safe(req, n, &clean_list, rq_list) {
1284                 LASSERT(req->rq_reply_state == 0);
1285                 target_exp_dequeue_req_replay(req);
1286                 ptlrpc_free_clone(req);
1287         }
1288
1289         spin_lock_bh(&obd->obd_processing_task_lock);
1290         list_splice_init(&obd->obd_lock_replay_queue, &clean_list);
1291         list_splice_init(&obd->obd_final_req_queue, &clean_list);
1292         spin_unlock_bh(&obd->obd_processing_task_lock);
1293
1294         list_for_each_entry_safe(req, n, &clean_list, rq_list){
1295                 LASSERT(req->rq_reply_state == 0);
1296                 ptlrpc_free_clone(req);
1297         }
1298
1299         EXIT;
1300 }
1301
1302 /* obd_processing_task_lock should be held */
1303 void target_cancel_recovery_timer(struct obd_device *obd)
1304 {
1305         CDEBUG(D_HA, "%s: cancel recovery timer\n", obd->obd_name);
1306         cfs_timer_disarm(&obd->obd_recovery_timer);
1307 }
1308
1309 /* extend = 1 means require at least "duration" seconds left in the timer,
1310    extend = 0 means set the total duration (start_recovery_timer) */
1311 static void reset_recovery_timer(struct obd_device *obd, int duration,
1312                                  int extend)
1313 {
1314         cfs_time_t now = cfs_time_current_sec();
1315         cfs_duration_t left;
1316
1317         spin_lock_bh(&obd->obd_processing_task_lock);
1318         if (!obd->obd_recovering || obd->obd_abort_recovery) {
1319                 spin_unlock_bh(&obd->obd_processing_task_lock);
1320                 return;
1321         }
1322
1323         left = cfs_time_sub(obd->obd_recovery_end, now);
1324
1325         if (extend && (duration > left))
1326                 obd->obd_recovery_timeout += duration - left;
1327         else if (!extend && (duration > obd->obd_recovery_timeout))
1328                 /* Track the client's largest expected replay time */
1329                 obd->obd_recovery_timeout = duration;
1330 #ifdef CRAY_XT3
1331         /*
1332          * If total recovery time already exceed the
1333          * obd_recovery_max_time, then CRAY XT3 will
1334          * abort the recovery
1335          */
1336         if(obd->obd_recovery_timeout > obd->obd_recovery_max_time)
1337                 obd->obd_recovery_timeout = obd->obd_recovery_max_time;
1338 #endif
1339         obd->obd_recovery_end = obd->obd_recovery_start +
1340                                 obd->obd_recovery_timeout;
1341         if (!cfs_timer_is_armed(&obd->obd_recovery_timer) ||
1342             cfs_time_before(now, obd->obd_recovery_end)) {
1343                 left = cfs_time_sub(obd->obd_recovery_end, now);
1344                 cfs_timer_arm(&obd->obd_recovery_timer, cfs_time_shift(left));
1345         }
1346         spin_unlock_bh(&obd->obd_processing_task_lock);
1347         CDEBUG(D_HA, "%s: recovery timer will expire in %u seconds\n",
1348                obd->obd_name, (unsigned)left);
1349 }
1350
1351 static void check_and_start_recovery_timer(struct obd_device *obd)
1352 {
1353         spin_lock_bh(&obd->obd_processing_task_lock);
1354         if (cfs_timer_is_armed(&obd->obd_recovery_timer)) {
1355                 spin_unlock_bh(&obd->obd_processing_task_lock);
1356                 return;
1357         }
1358         CDEBUG(D_HA, "%s: starting recovery timer\n", obd->obd_name);
1359         obd->obd_recovery_start = cfs_time_current_sec();
1360         /* minimum */
1361         obd->obd_recovery_timeout = OBD_RECOVERY_FACTOR * obd_timeout;
1362         spin_unlock_bh(&obd->obd_processing_task_lock);
1363
1364         reset_recovery_timer(obd, obd->obd_recovery_timeout, 0);
1365 }
1366
1367 /* Reset the timer with each new client connection */
1368 /*
1369  * This timer is actually reconnect_timer, which is for making sure
1370  * the total recovery window is at least as big as my reconnect
1371  * attempt timing. So the initial recovery time_out will be set to
1372  * OBD_RECOVERY_FACTOR * obd_timeout. If the timeout coming
1373  * from client is bigger than this, then the recovery time_out will
1374  * be extend to make sure the client could be reconnected, in the
1375  * process, the timeout from the new client should be ignored.
1376  */
1377
1378 static void
1379 target_start_and_reset_recovery_timer(struct obd_device *obd,
1380                                       struct ptlrpc_request *req,
1381                                       int new_client)
1382 {
1383         int service_time = lustre_msg_get_service_time(req->rq_reqmsg);
1384
1385         if (!new_client && service_time)
1386                 /* Teach server about old server's estimates, as first guess
1387                  * at how long new requests will take. */
1388                 at_add(&req->rq_rqbd->rqbd_service->srv_at_estimate,
1389                        service_time);
1390
1391         check_and_start_recovery_timer(obd);
1392
1393         /* convert the service time to rpc timeout,
1394          * reuse service_time to limit stack usage */
1395         service_time = at_est2timeout(service_time);
1396
1397         /* We expect other clients to timeout within service_time, then try
1398          * to reconnect, then try the failover server.  The max delay between
1399          * connect attempts is SWITCH_MAX + SWITCH_INC + INITIAL */
1400         service_time += 2 * (CONNECTION_SWITCH_MAX + CONNECTION_SWITCH_INC +
1401                              INITIAL_CONNECT_TIMEOUT);
1402         if (service_time > obd->obd_recovery_timeout && !new_client)
1403                 reset_recovery_timer(obd, service_time, 0);
1404 }
1405
1406 #ifdef __KERNEL__
1407
1408 /** Health checking routines */
1409 static inline int exp_connect_healthy(struct obd_export *exp)
1410 {
1411         return (exp->exp_in_recovery);
1412 }
1413
1414 /** if export done req_replay or has replay in queue */
1415 static inline int exp_req_replay_healthy(struct obd_export *exp)
1416 {
1417         return (!exp->exp_req_replay_needed ||
1418                 atomic_read(&exp->exp_replay_count) > 0);
1419 }
1420 /** if export done lock_replay or has replay in queue */
1421 static inline int exp_lock_replay_healthy(struct obd_export *exp)
1422 {
1423         return (!exp->exp_lock_replay_needed ||
1424                 atomic_read(&exp->exp_replay_count) > 0);
1425 }
1426
1427 static inline int exp_vbr_healthy(struct obd_export *exp)
1428 {
1429         return (!exp->exp_vbr_failed);
1430 }
1431
1432 static inline int exp_finished(struct obd_export *exp)
1433 {
1434         return (exp->exp_in_recovery && !exp->exp_lock_replay_needed);
1435 }
1436
1437 /** Checking routines for recovery */
1438 static int check_for_clients(struct obd_device *obd)
1439 {
1440         if (obd->obd_abort_recovery || obd->obd_recovery_expired)
1441                 return 1;
1442         LASSERT(obd->obd_connected_clients <= obd->obd_max_recoverable_clients);
1443         if (obd->obd_no_conn == 0 &&
1444             obd->obd_connected_clients + obd->obd_stale_clients ==
1445             obd->obd_max_recoverable_clients)
1446                 return 1;
1447         return 0;
1448 }
1449
1450 static int check_for_next_transno(struct obd_device *obd)
1451 {
1452         struct ptlrpc_request *req = NULL;
1453         int wake_up = 0, connected, completed, queue_len;
1454         __u64 next_transno, req_transno;
1455         ENTRY;
1456         spin_lock_bh(&obd->obd_processing_task_lock);
1457
1458         if (!list_empty(&obd->obd_req_replay_queue)) {
1459                 req = list_entry(obd->obd_req_replay_queue.next,
1460                                  struct ptlrpc_request, rq_list);
1461                 req_transno = lustre_msg_get_transno(req->rq_reqmsg);
1462         } else {
1463                 req_transno = 0;
1464         }
1465
1466         connected = obd->obd_connected_clients;
1467         completed = connected - atomic_read(&obd->obd_req_replay_clients);
1468         queue_len = obd->obd_requests_queued_for_recovery;
1469         next_transno = obd->obd_next_recovery_transno;
1470
1471         CDEBUG(D_HA, "max: %d, connected: %d, completed: %d, queue_len: %d, "
1472                "req_transno: "LPU64", next_transno: "LPU64"\n",
1473                obd->obd_max_recoverable_clients, connected, completed,
1474                queue_len, req_transno, next_transno);
1475
1476         if (obd->obd_abort_recovery) {
1477                 CDEBUG(D_HA, "waking for aborted recovery\n");
1478                 wake_up = 1;
1479         } else if (obd->obd_recovery_expired) {
1480                 CDEBUG(D_HA, "waking for expired recovery\n");
1481                 wake_up = 1;
1482         } else if (atomic_read(&obd->obd_req_replay_clients) == 0) {
1483                 CDEBUG(D_HA, "waking for completed recovery\n");
1484                 wake_up = 1;
1485         } else if (req_transno == next_transno) {
1486                 CDEBUG(D_HA, "waking for next ("LPD64")\n", next_transno);
1487                 wake_up = 1;
1488         } else if (queue_len == atomic_read(&obd->obd_req_replay_clients)) {
1489                 int d_lvl = D_HA;
1490                 /** handle gaps occured due to lost reply or VBR */
1491                 LASSERTF(req_transno >= next_transno,
1492                          "req_transno: "LPU64", next_transno: "LPU64"\n",
1493                          req_transno, next_transno);
1494                 if (req_transno > obd->obd_last_committed &&
1495                     !obd->obd_version_recov)
1496                         d_lvl = D_ERROR;
1497                 CDEBUG(d_lvl,
1498                        "%s: waking for gap in transno, VBR is %s (skip: "
1499                        LPD64", ql: %d, comp: %d, conn: %d, next: "LPD64
1500                        ", last_committed: "LPD64")\n",
1501                        obd->obd_name, obd->obd_version_recov ? "ON" : "OFF",
1502                        next_transno, queue_len, completed, connected,
1503                        req_transno, obd->obd_last_committed);
1504                 obd->obd_next_recovery_transno = req_transno;
1505                 wake_up = 1;
1506         } else if (OBD_FAIL_CHECK(OBD_FAIL_MDS_RECOVERY_ACCEPTS_GAPS)) {
1507                 CDEBUG(D_HA, "accepting transno gaps is explicitly allowed"
1508                        " by fail_lock, waking up ("LPD64")\n", next_transno);
1509                 obd->obd_next_recovery_transno = req_transno;
1510                 wake_up = 1;
1511         }
1512         spin_unlock_bh(&obd->obd_processing_task_lock);
1513         return wake_up;
1514 }
1515
1516 static int check_for_next_lock(struct obd_device *obd)
1517 {
1518         int wake_up = 0;
1519
1520         spin_lock_bh(&obd->obd_processing_task_lock);
1521         if (!list_empty(&obd->obd_lock_replay_queue)) {
1522                 CDEBUG(D_HA, "waking for next lock\n");
1523                 wake_up = 1;
1524         } else if (atomic_read(&obd->obd_lock_replay_clients) == 0) {
1525                 CDEBUG(D_HA, "waking for completed lock replay\n");
1526                 wake_up = 1;
1527         } else if (obd->obd_abort_recovery) {
1528                 CDEBUG(D_HA, "waking for aborted recovery\n");
1529                 wake_up = 1;
1530         } else if (obd->obd_recovery_expired) {
1531                 CDEBUG(D_HA, "waking for expired recovery\n");
1532                 wake_up = 1;
1533         }
1534         spin_unlock_bh(&obd->obd_processing_task_lock);
1535
1536         return wake_up;
1537 }
1538
1539 /**
1540  * wait for recovery events,
1541  * check its status with help of check_routine
1542  * evict dead clients via health_check
1543  */
1544 static int target_recovery_overseer(struct obd_device *obd,
1545                                     int (*check_routine)(struct obd_device *),
1546                                     int (*health_check)(struct obd_export *))
1547 {
1548         int abort = 0, expired = 1;
1549
1550         do {
1551                 cfs_wait_event(obd->obd_next_transno_waitq, check_routine(obd));
1552                 spin_lock_bh(&obd->obd_processing_task_lock);
1553                 abort = obd->obd_abort_recovery;
1554                 expired = obd->obd_recovery_expired;
1555                 obd->obd_recovery_expired = 0;
1556                 spin_unlock_bh(&obd->obd_processing_task_lock);
1557                 if (abort) {
1558                         CWARN("recovery is aborted, evict exports in recovery\n");
1559                         /** evict exports which didn't finish recovery yet */
1560                         class_disconnect_stale_exports(obd, exp_finished);
1561                 } else if (expired) {
1562                         /** If some clients died being recovered, evict them */
1563                         CDEBUG(D_WARNING, "recovery is timed out, evict stale exports\n");
1564                         /** evict cexports with no replay in queue, they are stalled */
1565                         class_disconnect_stale_exports(obd, health_check);
1566                         /** continue with VBR */
1567                         spin_lock_bh(&obd->obd_processing_task_lock);
1568                         obd->obd_version_recov = 1;
1569                         spin_unlock_bh(&obd->obd_processing_task_lock);
1570                         /**
1571                          * reset timer, recovery will proceed with versions now,
1572                          * timeout is set just to handle reconnection delays
1573                          */
1574                         reset_recovery_timer(obd, RECONNECT_DELAY_MAX, 1);
1575                         /** Wait for recovery events again, after evicting bad clients */
1576                 }
1577         } while (!abort && expired);
1578
1579         return abort;
1580 }
1581
1582 static struct ptlrpc_request *target_next_replay_req(struct obd_device *obd)
1583 {
1584         struct ptlrpc_request *req = NULL;
1585         ENTRY;
1586
1587         CDEBUG(D_HA, "Waiting for transno "LPD64"\n",
1588                obd->obd_next_recovery_transno);
1589
1590         if (target_recovery_overseer(obd, check_for_next_transno,
1591                                      exp_req_replay_healthy)) {
1592                 abort_req_replay_queue(obd);
1593                 abort_lock_replay_queue(obd);
1594         }
1595
1596         spin_lock_bh(&obd->obd_processing_task_lock);
1597         if (!list_empty(&obd->obd_req_replay_queue)) {
1598                 req = list_entry(obd->obd_req_replay_queue.next,
1599                                  struct ptlrpc_request, rq_list);
1600                 list_del_init(&req->rq_list);
1601                 obd->obd_requests_queued_for_recovery--;
1602                 spin_unlock_bh(&obd->obd_processing_task_lock);
1603         } else {
1604                 spin_unlock_bh(&obd->obd_processing_task_lock);
1605                 LASSERT(list_empty(&obd->obd_req_replay_queue));
1606                 LASSERT(atomic_read(&obd->obd_req_replay_clients) == 0);
1607                 /** evict exports failed VBR */
1608                 class_disconnect_stale_exports(obd, exp_vbr_healthy);
1609         }
1610         RETURN(req);
1611 }
1612
1613 static struct ptlrpc_request *target_next_replay_lock(struct obd_device *obd)
1614 {
1615         struct ptlrpc_request *req = NULL;
1616
1617         CDEBUG(D_HA, "Waiting for lock\n");
1618         if (target_recovery_overseer(obd, check_for_next_lock,
1619                                      exp_lock_replay_healthy))
1620                 abort_lock_replay_queue(obd);
1621
1622         spin_lock_bh(&obd->obd_processing_task_lock);
1623         if (!list_empty(&obd->obd_lock_replay_queue)) {
1624                 req = list_entry(obd->obd_lock_replay_queue.next,
1625                                  struct ptlrpc_request, rq_list);
1626                 list_del_init(&req->rq_list);
1627                 spin_unlock_bh(&obd->obd_processing_task_lock);
1628         } else {
1629                 spin_unlock_bh(&obd->obd_processing_task_lock);
1630                 LASSERT(list_empty(&obd->obd_lock_replay_queue));
1631                 LASSERT(atomic_read(&obd->obd_lock_replay_clients) == 0);
1632                 /** evict exports failed VBR */
1633                 class_disconnect_stale_exports(obd, exp_vbr_healthy);
1634         }
1635         return req;
1636 }
1637
1638 static struct ptlrpc_request *target_next_final_ping(struct obd_device *obd)
1639 {
1640         struct ptlrpc_request *req = NULL;
1641
1642         spin_lock_bh(&obd->obd_processing_task_lock);
1643         if (!list_empty(&obd->obd_final_req_queue)) {
1644                 req = list_entry(obd->obd_final_req_queue.next,
1645                                  struct ptlrpc_request, rq_list);
1646                 list_del_init(&req->rq_list);
1647                 if (req->rq_export->exp_in_recovery) {
1648                         spin_lock(&req->rq_export->exp_lock);
1649                         req->rq_export->exp_in_recovery = 0;
1650                         spin_unlock(&req->rq_export->exp_lock);
1651                 }
1652         }
1653         spin_unlock_bh(&obd->obd_processing_task_lock);
1654         return req;
1655 }
1656
1657 static int handle_recovery_req(struct ptlrpc_thread *thread,
1658                                struct ptlrpc_request *req,
1659                                svc_handler_t handler)
1660 {
1661         int rc;
1662         ENTRY;
1663
1664         rc = lu_context_init(&req->rq_session, LCT_SESSION);
1665         if (rc) {
1666                 CERROR("Failure to initialize session: %d\n", rc);
1667                 GOTO(free_clone, rc);
1668         }
1669         /**
1670          * export can be evicted during recovery, no need to handle replays for
1671          * it after that, discard such request silently
1672          */
1673         if (req->rq_export->exp_disconnected)
1674                 GOTO(free_clone, rc);
1675
1676         req->rq_session.lc_thread = thread;
1677         lu_context_enter(&req->rq_session);
1678         req->rq_svc_thread = thread;
1679         req->rq_svc_thread->t_env->le_ses = &req->rq_session;
1680
1681         /* thread context */
1682         lu_context_enter(&thread->t_env->le_ctx);
1683         (void)handler(req);
1684         lu_context_exit(&thread->t_env->le_ctx);
1685
1686         lu_context_exit(&req->rq_session);
1687         lu_context_fini(&req->rq_session);
1688         /* don't reset timer for final stage */
1689         if (!exp_finished(req->rq_export)) {
1690                 /**
1691                  * XXX: until bug 18948 is fixed (enable AT for request copy)
1692                  * the client may reconnect during recovery so we may need to
1693                  * wait RECONNECT_DELAY_MAX after each replay instead of
1694                  * at_get(&req->rq_rqbd->rqbd_service->srv_at_estimate);
1695                  */
1696                  reset_recovery_timer(class_exp2obd(req->rq_export), AT_OFF ?
1697                                       obd_timeout : RECONNECT_DELAY_MAX, 1);
1698         }
1699         /**
1700          * bz18031: increase next_recovery_transno before ptlrpc_free_clone()
1701          * will drop exp_rpc reference
1702          */
1703         if (req->rq_export->exp_req_replay_needed) {
1704                 spin_lock_bh(&req->rq_export->exp_obd->obd_processing_task_lock);
1705                 req->rq_export->exp_obd->obd_next_recovery_transno++;
1706                 spin_unlock_bh(&req->rq_export->exp_obd->obd_processing_task_lock);
1707                 target_exp_dequeue_req_replay(req);
1708         }
1709 free_clone:
1710         ptlrpc_free_clone(req);
1711         RETURN(0);
1712 }
1713
1714 static int target_recovery_thread(void *arg)
1715 {
1716         struct lu_target *lut = arg;
1717         struct obd_device *obd = lut->lut_obd;
1718         struct ptlrpc_request *req;
1719         struct target_recovery_data *trd = &obd->obd_recovery_data;
1720         unsigned long delta;
1721         unsigned long flags;
1722         struct lu_env env;
1723         struct ptlrpc_thread fake_svc_thread, *thread = &fake_svc_thread;
1724         int rc = 0;
1725         ENTRY;
1726
1727         cfs_daemonize("tgt_recov");
1728
1729         SIGNAL_MASK_LOCK(current, flags);
1730         sigfillset(&current->blocked);
1731         RECALC_SIGPENDING;
1732         SIGNAL_MASK_UNLOCK(current, flags);
1733
1734         rc = lu_context_init(&env.le_ctx, LCT_MD_THREAD);
1735         if (rc)
1736                 RETURN(rc);
1737
1738         thread->t_env = &env;
1739         env.le_ctx.lc_thread = thread;
1740
1741         CERROR("%s: started recovery thread pid %d\n", obd->obd_name,
1742                cfs_curproc_pid());
1743         trd->trd_processing_task = cfs_curproc_pid();
1744
1745         obd->obd_recovering = 1;
1746         complete(&trd->trd_starting);
1747
1748         /* first of all, we have to know the first transno to replay */
1749         if (target_recovery_overseer(obd, check_for_clients,
1750                                      exp_connect_healthy)) {
1751                 abort_req_replay_queue(obd);
1752                 abort_lock_replay_queue(obd);
1753         }
1754
1755         /* next stage: replay requests */
1756         delta = jiffies;
1757         obd->obd_req_replaying = 1;
1758         CDEBUG(D_INFO, "1: request replay stage - %d clients from t"LPU64"\n",
1759                atomic_read(&obd->obd_req_replay_clients),
1760                obd->obd_next_recovery_transno);
1761         while ((req = target_next_replay_req(obd))) {
1762                 LASSERT(trd->trd_processing_task == cfs_curproc_pid());
1763                 DEBUG_REQ(D_HA, req, "processing t"LPD64" from %s",
1764                           lustre_msg_get_transno(req->rq_reqmsg),
1765                           libcfs_nid2str(req->rq_peer.nid));
1766                 handle_recovery_req(thread, req,
1767                                     trd->trd_recovery_handler);
1768                 obd->obd_replayed_requests++;
1769         }
1770
1771         /**
1772          * The second stage: replay locks
1773          */
1774         CDEBUG(D_INFO, "2: lock replay stage - %d clients\n",
1775                atomic_read(&obd->obd_lock_replay_clients));
1776         while ((req = target_next_replay_lock(obd))) {
1777                 LASSERT(trd->trd_processing_task == cfs_curproc_pid());
1778                 DEBUG_REQ(D_HA, req, "processing lock from %s: ",
1779                           libcfs_nid2str(req->rq_peer.nid));
1780                 handle_recovery_req(thread, req,
1781                                     trd->trd_recovery_handler);
1782                 obd->obd_replayed_locks++;
1783         }
1784
1785         /**
1786          * The third stage: reply on final pings, at this moment all clients
1787          * must have request in final queue
1788          */
1789         CDEBUG(D_INFO, "3: final stage - process recovery completion pings\n");
1790         /** Update server last boot epoch */
1791         lut_boot_epoch_update(lut);
1792         /* We drop recoverying flag to forward all new requests
1793          * to regular mds_handle() since now */
1794         spin_lock_bh(&obd->obd_processing_task_lock);
1795         obd->obd_recovering = obd->obd_abort_recovery = 0;
1796         target_cancel_recovery_timer(obd);
1797         spin_unlock_bh(&obd->obd_processing_task_lock);
1798         while ((req = target_next_final_ping(obd))) {
1799                 LASSERT(trd->trd_processing_task == cfs_curproc_pid());
1800                 DEBUG_REQ(D_HA, req, "processing final ping from %s: ",
1801                           libcfs_nid2str(req->rq_peer.nid));
1802                 handle_recovery_req(thread, req,
1803                                     trd->trd_recovery_handler);
1804         }
1805
1806         delta = (jiffies - delta) / HZ;
1807         CDEBUG(D_INFO,"4: recovery completed in %lus - %d/%d reqs/locks\n",
1808               delta, obd->obd_replayed_requests, obd->obd_replayed_locks);
1809         if (delta > obd_timeout * OBD_RECOVERY_FACTOR) {
1810                 CWARN("too long recovery - read logs\n");
1811                 libcfs_debug_dumplog();
1812         }
1813
1814         target_finish_recovery(obd);
1815
1816         lu_context_fini(&env.le_ctx);
1817         trd->trd_processing_task = 0;
1818         complete(&trd->trd_finishing);
1819         RETURN(rc);
1820 }
1821
1822 static int target_start_recovery_thread(struct lu_target *lut,
1823                                         svc_handler_t handler)
1824 {
1825         struct obd_device *obd = lut->lut_obd;
1826         int rc = 0;
1827         struct target_recovery_data *trd = &obd->obd_recovery_data;
1828
1829         memset(trd, 0, sizeof(*trd));
1830         init_completion(&trd->trd_starting);
1831         init_completion(&trd->trd_finishing);
1832         trd->trd_recovery_handler = handler;
1833
1834         if (kernel_thread(target_recovery_thread, lut, 0) > 0) {
1835                 wait_for_completion(&trd->trd_starting);
1836                 LASSERT(obd->obd_recovering != 0);
1837         } else
1838                 rc = -ECHILD;
1839
1840         return rc;
1841 }
1842
1843 void target_stop_recovery_thread(struct obd_device *obd)
1844 {
1845         spin_lock_bh(&obd->obd_processing_task_lock);
1846         if (obd->obd_recovery_data.trd_processing_task > 0) {
1847                 struct target_recovery_data *trd = &obd->obd_recovery_data;
1848                 /** recovery can be done but postrecovery is not yet */
1849                 if (obd->obd_recovering) {
1850                         CERROR("%s: Aborting recovery\n", obd->obd_name);
1851                         obd->obd_abort_recovery = 1;
1852                         cfs_waitq_signal(&obd->obd_next_transno_waitq);
1853                 }
1854                 spin_unlock_bh(&obd->obd_processing_task_lock);
1855                 wait_for_completion(&trd->trd_finishing);
1856         } else {
1857                 spin_unlock_bh(&obd->obd_processing_task_lock);
1858         }
1859 }
1860
1861 void target_recovery_fini(struct obd_device *obd)
1862 {
1863         class_disconnect_exports(obd);
1864         target_stop_recovery_thread(obd);
1865         target_cleanup_recovery(obd);
1866 }
1867 EXPORT_SYMBOL(target_recovery_fini);
1868
1869 static void target_recovery_expired(unsigned long castmeharder)
1870 {
1871         struct obd_device *obd = (struct obd_device *)castmeharder;
1872         CDEBUG(D_HA, "%s: recovery timed out; %d clients are still in recovery"
1873                " after %lds (%d clients connected)\n",
1874                obd->obd_name, atomic_read(&obd->obd_lock_replay_clients),
1875                cfs_time_current_sec()- obd->obd_recovery_start,
1876                obd->obd_connected_clients);
1877
1878         spin_lock_bh(&obd->obd_processing_task_lock);
1879         obd->obd_recovery_expired = 1;
1880         cfs_waitq_signal(&obd->obd_next_transno_waitq);
1881         spin_unlock_bh(&obd->obd_processing_task_lock);
1882 }
1883
1884 void target_recovery_init(struct lu_target *lut, svc_handler_t handler)
1885 {
1886         struct obd_device *obd = lut->lut_obd;
1887         if (obd->obd_max_recoverable_clients == 0) {
1888                 /** Update server last boot epoch */
1889                 lut_boot_epoch_update(lut);
1890                 return;
1891         }
1892
1893         CWARN("RECOVERY: service %s, %d recoverable clients, "
1894               "last_transno "LPU64"\n", obd->obd_name,
1895               obd->obd_max_recoverable_clients, obd->obd_last_committed);
1896         LASSERT(obd->obd_stopping == 0);
1897         obd->obd_next_recovery_transno = obd->obd_last_committed + 1;
1898         obd->obd_recovery_start = 0;
1899         obd->obd_recovery_end = 0;
1900         obd->obd_recovery_timeout = OBD_RECOVERY_FACTOR * obd_timeout;
1901         /* bz13079: this should be set to desired value for ost but not for mds */
1902         obd->obd_recovery_max_time = OBD_RECOVERY_MAX_TIME;
1903         cfs_timer_init(&obd->obd_recovery_timer, target_recovery_expired, obd);
1904         target_start_recovery_thread(lut, handler);
1905 }
1906 EXPORT_SYMBOL(target_recovery_init);
1907
1908 #endif
1909
1910 static int target_process_req_flags(struct obd_device *obd,
1911                                     struct ptlrpc_request *req)
1912 {
1913         struct obd_export *exp = req->rq_export;
1914         LASSERT(exp != NULL);
1915         if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REQ_REPLAY_DONE) {
1916                 /* client declares he's ready to replay locks */
1917                 spin_lock_bh(&obd->obd_processing_task_lock);
1918                 if (exp->exp_req_replay_needed) {
1919                         LASSERT(atomic_read(&obd->obd_req_replay_clients) > 0);
1920                         spin_lock(&exp->exp_lock);
1921                         exp->exp_req_replay_needed = 0;
1922                         spin_unlock(&exp->exp_lock);
1923                         atomic_dec(&obd->obd_req_replay_clients);
1924                 }
1925                 spin_unlock_bh(&obd->obd_processing_task_lock);
1926         }
1927         if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_LOCK_REPLAY_DONE) {
1928                 /* client declares he's ready to complete recovery
1929                  * so, we put the request on th final queue */
1930                 spin_lock_bh(&obd->obd_processing_task_lock);
1931                 if (exp->exp_lock_replay_needed) {
1932                         LASSERT(atomic_read(&obd->obd_lock_replay_clients) > 0);
1933                         spin_lock(&exp->exp_lock);
1934                         exp->exp_lock_replay_needed = 0;
1935                         spin_unlock(&exp->exp_lock);
1936                         atomic_dec(&obd->obd_lock_replay_clients);
1937                 }
1938                 spin_unlock_bh(&obd->obd_processing_task_lock);
1939         }
1940
1941         return 0;
1942 }
1943
1944 int target_queue_recovery_request(struct ptlrpc_request *req,
1945                                   struct obd_device *obd)
1946 {
1947         struct list_head *tmp;
1948         int inserted = 0;
1949         __u64 transno = lustre_msg_get_transno(req->rq_reqmsg);
1950         ENTRY;
1951
1952         if (obd->obd_recovery_data.trd_processing_task == cfs_curproc_pid()) {
1953                 /* Processing the queue right now, don't re-add. */
1954                 RETURN(1);
1955         }
1956
1957         target_process_req_flags(obd, req);
1958
1959         if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_LOCK_REPLAY_DONE) {
1960                 /* client declares he's ready to complete recovery
1961                  * so, we put the request on th final queue */
1962                 req = ptlrpc_clone_req(req);
1963                 if (req == NULL)
1964                         RETURN(-ENOMEM);
1965                 DEBUG_REQ(D_HA, req, "queue final req");
1966                 spin_lock_bh(&obd->obd_processing_task_lock);
1967                 cfs_waitq_signal(&obd->obd_next_transno_waitq);
1968                 if (obd->obd_recovering)
1969                         list_add_tail(&req->rq_list, &obd->obd_final_req_queue);
1970                 else {
1971                         spin_unlock_bh(&obd->obd_processing_task_lock);
1972                         ptlrpc_free_clone(req);
1973                         if (obd->obd_stopping) {
1974                                 RETURN(-ENOTCONN);
1975                         } else {
1976                                 RETURN(1);
1977                         }
1978                 }
1979                 spin_unlock_bh(&obd->obd_processing_task_lock);
1980                 RETURN(0);
1981         }
1982         if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REQ_REPLAY_DONE) {
1983                 /* client declares he's ready to replay locks */
1984                 req = ptlrpc_clone_req(req);
1985                 if (req == NULL)
1986                         RETURN(-ENOMEM);
1987                 DEBUG_REQ(D_HA, req, "queue lock replay req");
1988                 spin_lock_bh(&obd->obd_processing_task_lock);
1989                 cfs_waitq_signal(&obd->obd_next_transno_waitq);
1990                 LASSERT(obd->obd_recovering);
1991                 /* usually due to recovery abort */
1992                 if (!req->rq_export->exp_in_recovery) {
1993                         spin_unlock_bh(&obd->obd_processing_task_lock);
1994                         ptlrpc_free_clone(req);
1995                         RETURN(-ENOTCONN);
1996                 }
1997                 LASSERT(req->rq_export->exp_lock_replay_needed);
1998                 list_add_tail(&req->rq_list, &obd->obd_lock_replay_queue);
1999                 spin_unlock_bh(&obd->obd_processing_task_lock);
2000                 RETURN(0);
2001         }
2002
2003         /* CAVEAT EMPTOR: The incoming request message has been swabbed
2004          * (i.e. buflens etc are in my own byte order), but type-dependent
2005          * buffers (eg mds_body, ost_body etc) have NOT been swabbed. */
2006
2007         if (!transno) {
2008                 CFS_INIT_LIST_HEAD(&req->rq_list);
2009                 DEBUG_REQ(D_HA, req, "not queueing");
2010                 RETURN(1);
2011         }
2012
2013         spin_lock_bh(&obd->obd_processing_task_lock);
2014
2015         /* If we're processing the queue, we want don't want to queue this
2016          * message.
2017          *
2018          * Also, if this request has a transno less than the one we're waiting
2019          * for, we should process it now.  It could (and currently always will)
2020          * be an open request for a descriptor that was opened some time ago.
2021          *
2022          * Also, a resent, replayed request that has already been
2023          * handled will pass through here and be processed immediately.
2024          */
2025         CWARN("Next recovery transno: "LPU64", current: "LPU64", replaying: %i\n",
2026               obd->obd_next_recovery_transno, transno, obd->obd_req_replaying);
2027         if (transno < obd->obd_next_recovery_transno && obd->obd_req_replaying) {
2028                 /* Processing the queue right now, don't re-add. */
2029                 LASSERT(list_empty(&req->rq_list));
2030                 spin_unlock_bh(&obd->obd_processing_task_lock);
2031                 RETURN(1);
2032         }
2033         spin_unlock_bh(&obd->obd_processing_task_lock);
2034
2035         if (OBD_FAIL_CHECK(OBD_FAIL_TGT_REPLAY_DROP))
2036                 RETURN(0);
2037
2038         req = ptlrpc_clone_req(req);
2039         if (req == NULL)
2040                 RETURN(-ENOMEM);
2041
2042         spin_lock_bh(&obd->obd_processing_task_lock);
2043         LASSERT(obd->obd_recovering);
2044         if (!req->rq_export->exp_in_recovery) {
2045                 spin_unlock_bh(&obd->obd_processing_task_lock);
2046                 ptlrpc_free_clone(req);
2047                 RETURN(-ENOTCONN);
2048         }
2049         LASSERT(req->rq_export->exp_req_replay_needed);
2050
2051         if (target_exp_enqueue_req_replay(req)) {
2052                 spin_unlock_bh(&obd->obd_processing_task_lock);
2053                 DEBUG_REQ(D_ERROR, req, "dropping resent queued req");
2054                 ptlrpc_free_clone(req);
2055                 RETURN(0);
2056         }
2057
2058         /* XXX O(n^2) */
2059         list_for_each(tmp, &obd->obd_req_replay_queue) {
2060                 struct ptlrpc_request *reqiter =
2061                         list_entry(tmp, struct ptlrpc_request, rq_list);
2062
2063                 if (lustre_msg_get_transno(reqiter->rq_reqmsg) > transno) {
2064                         list_add_tail(&req->rq_list, &reqiter->rq_list);
2065                         inserted = 1;
2066                         break;
2067                 }
2068
2069                 if (unlikely(lustre_msg_get_transno(reqiter->rq_reqmsg) ==
2070                              transno)) {
2071                         DEBUG_REQ(D_ERROR, req, "dropping replay: transno "
2072                                   "has been claimed by another client");
2073                         spin_unlock_bh(&obd->obd_processing_task_lock);
2074                         target_exp_dequeue_req_replay(req);
2075                         ptlrpc_free_clone(req);
2076                         RETURN(0);
2077                 }
2078         }
2079
2080         if (!inserted)
2081                 list_add_tail(&req->rq_list, &obd->obd_req_replay_queue);
2082
2083         obd->obd_requests_queued_for_recovery++;
2084         cfs_waitq_signal(&obd->obd_next_transno_waitq);
2085         spin_unlock_bh(&obd->obd_processing_task_lock);
2086         RETURN(0);
2087 }
2088
2089 /**
2090  * Packs current SLV and Limit into \a req.
2091  */
2092 int target_pack_pool_reply(struct ptlrpc_request *req)
2093 {
2094         struct obd_device *obd;
2095         ENTRY;
2096
2097         /*
2098          * Check that we still have all structures alive as this may
2099          * be some late rpc in shutdown time.
2100          */
2101         if (unlikely(!req->rq_export || !req->rq_export->exp_obd ||
2102                      !exp_connect_lru_resize(req->rq_export))) {
2103                 lustre_msg_set_slv(req->rq_repmsg, 0);
2104                 lustre_msg_set_limit(req->rq_repmsg, 0);
2105                 RETURN(0);
2106         }
2107
2108         /*
2109          * OBD is alive here as export is alive, which we checked above.
2110          */
2111         obd = req->rq_export->exp_obd;
2112
2113         read_lock(&obd->obd_pool_lock);
2114         lustre_msg_set_slv(req->rq_repmsg, obd->obd_pool_slv);
2115         lustre_msg_set_limit(req->rq_repmsg, obd->obd_pool_limit);
2116         read_unlock(&obd->obd_pool_lock);
2117
2118         RETURN(0);
2119 }
2120
2121 int target_send_reply_msg(struct ptlrpc_request *req, int rc, int fail_id)
2122 {
2123         if (OBD_FAIL_CHECK_ORSET(fail_id & ~OBD_FAIL_ONCE, OBD_FAIL_ONCE)) {
2124                 DEBUG_REQ(D_ERROR, req, "dropping reply");
2125                 return (-ECOMM);
2126         }
2127
2128         if (unlikely(rc)) {
2129                 DEBUG_REQ(D_ERROR, req, "processing error (%d)", rc);
2130                 req->rq_status = rc;
2131                 return (ptlrpc_send_error(req, 1));
2132         } else {
2133                 DEBUG_REQ(D_NET, req, "sending reply");
2134         }
2135
2136         return (ptlrpc_send_reply(req, PTLRPC_REPLY_MAYBE_DIFFICULT));
2137 }
2138
2139 void target_send_reply(struct ptlrpc_request *req, int rc, int fail_id)
2140 {
2141         int                        netrc;
2142         struct ptlrpc_reply_state *rs;
2143         struct obd_device         *obd;
2144         struct obd_export         *exp;
2145         struct ptlrpc_service     *svc;
2146         ENTRY;
2147
2148         if (req->rq_no_reply) {
2149                 EXIT;
2150                 return;
2151         }
2152
2153         svc = req->rq_rqbd->rqbd_service;
2154         rs = req->rq_reply_state;
2155         if (rs == NULL || !rs->rs_difficult) {
2156                 /* no notifiers */
2157                 target_send_reply_msg (req, rc, fail_id);
2158                 EXIT;
2159                 return;
2160         }
2161
2162         /* must be an export if locks saved */
2163         LASSERT (req->rq_export != NULL);
2164         /* req/reply consistent */
2165         LASSERT (rs->rs_service == svc);
2166
2167         /* "fresh" reply */
2168         LASSERT (!rs->rs_scheduled);
2169         LASSERT (!rs->rs_scheduled_ever);
2170         LASSERT (!rs->rs_handled);
2171         LASSERT (!rs->rs_on_net);
2172         LASSERT (rs->rs_export == NULL);
2173         LASSERT (list_empty(&rs->rs_obd_list));
2174         LASSERT (list_empty(&rs->rs_exp_list));
2175
2176         exp = class_export_get (req->rq_export);
2177         obd = exp->exp_obd;
2178
2179         /* disable reply scheduling onto srv_reply_queue while I'm setting up */
2180         rs->rs_scheduled = 1;
2181         rs->rs_on_net    = 1;
2182         rs->rs_xid       = req->rq_xid;
2183         rs->rs_transno   = req->rq_transno;
2184         rs->rs_export    = exp;
2185         rs->rs_opc       = lustre_msg_get_opc(rs->rs_msg);
2186
2187         spin_lock(&exp->exp_uncommitted_replies_lock);
2188         CDEBUG(D_NET, "rs transno = "LPU64", last committed = "LPU64"\n",
2189                rs->rs_transno, exp->exp_last_committed);
2190         if (rs->rs_transno > exp->exp_last_committed) {
2191                 /* not committed already */
2192                 list_add_tail(&rs->rs_obd_list,
2193                               &exp->exp_uncommitted_replies);
2194         }
2195         spin_unlock (&exp->exp_uncommitted_replies_lock);
2196
2197         spin_lock(&exp->exp_lock);
2198         list_add_tail(&rs->rs_exp_list, &exp->exp_outstanding_replies);
2199         spin_unlock(&exp->exp_lock);
2200
2201         netrc = target_send_reply_msg (req, rc, fail_id);
2202
2203         spin_lock(&svc->srv_lock);
2204
2205         atomic_inc(&svc->srv_n_difficult_replies);
2206
2207         if (netrc != 0) {
2208                 /* error sending: reply is off the net.  Also we need +1
2209                  * reply ref until ptlrpc_server_handle_reply() is done
2210                  * with the reply state (if the send was successful, there
2211                  * would have been +1 ref for the net, which
2212                  * reply_out_callback leaves alone) */
2213                 rs->rs_on_net = 0;
2214                 ptlrpc_rs_addref(rs);
2215                 atomic_inc (&svc->srv_outstanding_replies);
2216         }
2217
2218         spin_lock(&rs->rs_lock);
2219         if (rs->rs_transno <= exp->exp_last_committed ||
2220             (!rs->rs_on_net && !rs->rs_no_ack) ||
2221              list_empty(&rs->rs_exp_list) ||     /* completed already */
2222              list_empty(&rs->rs_obd_list)) {
2223                 CDEBUG(D_HA, "Schedule reply immediately\n");
2224                 ptlrpc_dispatch_difficult_reply(rs);
2225         } else {
2226                 list_add (&rs->rs_list, &svc->srv_active_replies);
2227                 rs->rs_scheduled = 0;           /* allow notifier to schedule */
2228         }
2229         spin_unlock(&rs->rs_lock);
2230         spin_unlock(&svc->srv_lock);
2231         EXIT;
2232 }
2233
2234 int target_handle_ping(struct ptlrpc_request *req)
2235 {
2236         obd_ping(req->rq_export);
2237         return req_capsule_server_pack(&req->rq_pill);
2238 }
2239
2240 void target_committed_to_req(struct ptlrpc_request *req)
2241 {
2242         struct obd_export *exp = req->rq_export;
2243
2244         if (!exp->exp_obd->obd_no_transno && req->rq_repmsg != NULL)
2245                 lustre_msg_set_last_committed(req->rq_repmsg,
2246                                               exp->exp_last_committed);
2247         else
2248                 DEBUG_REQ(D_IOCTL, req, "not sending last_committed update (%d/"
2249                           "%d)", exp->exp_obd->obd_no_transno,
2250                           req->rq_repmsg == NULL);
2251
2252         CDEBUG(D_INFO, "last_committed "LPU64", transno "LPU64", xid "LPU64"\n",
2253                exp->exp_last_committed, req->rq_transno, req->rq_xid);
2254 }
2255 EXPORT_SYMBOL(target_committed_to_req);
2256
2257 int target_handle_qc_callback(struct ptlrpc_request *req)
2258 {
2259         struct obd_quotactl *oqctl;
2260         struct client_obd *cli = &req->rq_export->exp_obd->u.cli;
2261
2262         oqctl = req_capsule_client_get(&req->rq_pill, &RMF_OBD_QUOTACTL);
2263         if (oqctl == NULL) {
2264                 CERROR("Can't unpack obd_quotactl\n");
2265                 RETURN(-EPROTO);
2266         }
2267
2268         cli->cl_qchk_stat = oqctl->qc_stat;
2269
2270         return 0;
2271 }
2272
2273 #ifdef HAVE_QUOTA_SUPPORT
2274 int target_handle_dqacq_callback(struct ptlrpc_request *req)
2275 {
2276 #ifdef __KERNEL__
2277         struct obd_device *obd = req->rq_export->exp_obd;
2278         struct obd_device *master_obd;
2279         struct obd_device_target *obt;
2280         struct lustre_quota_ctxt *qctxt;
2281         struct qunit_data *qdata = NULL;
2282         int rc = 0;
2283         ENTRY;
2284
2285         if (OBD_FAIL_CHECK(OBD_FAIL_MDS_DROP_QUOTA_REQ))
2286                 RETURN(rc);
2287
2288         rc = req_capsule_server_pack(&req->rq_pill);
2289         if (rc) {
2290                 CERROR("packing reply failed!: rc = %d\n", rc);
2291                 RETURN(rc);
2292         }
2293
2294         LASSERT(req->rq_export);
2295
2296         qdata = quota_get_qdata(req, QUOTA_REQUEST, QUOTA_EXPORT);
2297         if (IS_ERR(qdata)) {
2298                 rc = PTR_ERR(qdata);
2299                 CDEBUG(D_ERROR, "Can't unpack qunit_data(rc: %d)\n", rc);
2300                 req->rq_status = rc;
2301                 GOTO(out, rc);
2302         }
2303
2304         /* we use the observer */
2305         if (!obd->obd_observer || !obd->obd_observer->obd_observer) {
2306                 CERROR("Can't find the observer, it is recovering\n");
2307                 req->rq_status = -EAGAIN;
2308                 GOTO(out, rc);
2309         }
2310
2311         master_obd = obd->obd_observer->obd_observer;
2312         obt = &master_obd->u.obt;
2313         qctxt = &obt->obt_qctxt;
2314
2315         if (!qctxt->lqc_setup || !qctxt->lqc_valid) {
2316                 /* quota_type has not been processed yet, return EAGAIN
2317                  * until we know whether or not quotas are supposed to
2318                  * be enabled */
2319                 CDEBUG(D_QUOTA, "quota_type not processed yet, return "
2320                        "-EAGAIN\n");
2321                 req->rq_status = -EAGAIN;
2322                 GOTO(out, rc);
2323         }
2324
2325         down_read(&obt->obt_rwsem);
2326         if (qctxt->lqc_lqs_hash == NULL) {
2327                 up_read(&obt->obt_rwsem);
2328                 /* quota_type has not been processed yet, return EAGAIN
2329                  * until we know whether or not quotas are supposed to
2330                  * be enabled */
2331                 CDEBUG(D_QUOTA, "quota_ctxt is not ready yet, return "
2332                        "-EAGAIN\n");
2333                 req->rq_status = -EAGAIN;
2334                 GOTO(out, rc);
2335         }
2336
2337         LASSERT(qctxt->lqc_handler);
2338         rc = qctxt->lqc_handler(master_obd, qdata,
2339                                 lustre_msg_get_opc(req->rq_reqmsg));
2340         up_read(&obt->obt_rwsem);
2341         if (rc && rc != -EDQUOT)
2342                 CDEBUG(rc == -EBUSY  ? D_QUOTA : D_ERROR,
2343                        "dqacq/dqrel failed! (rc:%d)\n", rc);
2344         req->rq_status = rc;
2345
2346         rc = quota_copy_qdata(req, qdata, QUOTA_REPLY, QUOTA_EXPORT);
2347         if (rc < 0) {
2348                 CERROR("Can't pack qunit_data(rc: %d)\n", rc);
2349                 GOTO(out, rc);
2350         }
2351
2352         /* Block the quota req. b=14840 */
2353         OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_BLOCK_QUOTA_REQ, obd_timeout);
2354         EXIT;
2355
2356 out:
2357         rc = ptlrpc_reply(req);
2358         return rc;
2359 #else
2360         return 0;
2361 #endif /* !__KERNEL__ */
2362 }
2363 #endif /* HAVE_QUOTA_SUPPORT */
2364
2365 /* Send a remote set_info_async.
2366  * This may go from client to server or server to client
2367  */
2368 int target_set_info_rpc(struct obd_import *imp, int opcode,
2369                         obd_count keylen, void *key,
2370                         obd_count vallen, void *val,
2371                         struct ptlrpc_request_set *set)
2372 {
2373         struct ptlrpc_request *req;
2374         char                  *tmp;
2375         int                    rc;
2376         ENTRY;
2377
2378         req = ptlrpc_request_alloc(imp, &RQF_OBD_SET_INFO);
2379         if (req == NULL)
2380                 RETURN(-ENOMEM);
2381
2382         req_capsule_set_size(&req->rq_pill, &RMF_SETINFO_KEY,
2383                              RCL_CLIENT, keylen);
2384         req_capsule_set_size(&req->rq_pill, &RMF_SETINFO_VAL,
2385                              RCL_CLIENT, vallen);
2386         rc = ptlrpc_request_pack(req, LUSTRE_OBD_VERSION, opcode);
2387         if (rc) {
2388                 ptlrpc_request_free(req);
2389                 RETURN(rc);
2390         }
2391
2392         tmp = req_capsule_client_get(&req->rq_pill, &RMF_SETINFO_KEY);
2393         memcpy(tmp, key, keylen);
2394         tmp = req_capsule_client_get(&req->rq_pill, &RMF_SETINFO_VAL);
2395         memcpy(tmp, val, vallen);
2396
2397         ptlrpc_request_set_replen(req);
2398
2399         if (set) {
2400                 ptlrpc_set_add_req(set, req);
2401                 ptlrpc_check_set(NULL, set);
2402         } else {
2403                 rc = ptlrpc_queue_wait(req);
2404                 ptlrpc_req_finished(req);
2405         }
2406
2407         RETURN(rc);
2408 }
2409 EXPORT_SYMBOL(target_set_info_rpc);
2410
2411
2412 ldlm_mode_t lck_compat_array[] = {
2413         [LCK_EX] LCK_COMPAT_EX,
2414         [LCK_PW] LCK_COMPAT_PW,
2415         [LCK_PR] LCK_COMPAT_PR,
2416         [LCK_CW] LCK_COMPAT_CW,
2417         [LCK_CR] LCK_COMPAT_CR,
2418         [LCK_NL] LCK_COMPAT_NL,
2419         [LCK_GROUP] LCK_COMPAT_GROUP,
2420         [LCK_COS] LCK_COMPAT_COS,
2421 };
2422
2423 /**
2424  * Rather arbitrary mapping from LDLM error codes to errno values. This should
2425  * not escape to the user level.
2426  */
2427 int ldlm_error2errno(ldlm_error_t error)
2428 {
2429         int result;
2430
2431         switch (error) {
2432         case ELDLM_OK:
2433                 result = 0;
2434                 break;
2435         case ELDLM_LOCK_CHANGED:
2436                 result = -ESTALE;
2437                 break;
2438         case ELDLM_LOCK_ABORTED:
2439                 result = -ENAVAIL;
2440                 break;
2441         case ELDLM_LOCK_REPLACED:
2442                 result = -ESRCH;
2443                 break;
2444         case ELDLM_NO_LOCK_DATA:
2445                 result = -ENOENT;
2446                 break;
2447         case ELDLM_NAMESPACE_EXISTS:
2448                 result = -EEXIST;
2449                 break;
2450         case ELDLM_BAD_NAMESPACE:
2451                 result = -EBADF;
2452                 break;
2453         default:
2454                 if (((int)error) < 0)  /* cast to signed type */
2455                         result = error; /* as ldlm_error_t can be unsigned */
2456                 else {
2457                         CERROR("Invalid DLM result code: %i\n", error);
2458                         result = -EPROTO;
2459                 }
2460         }
2461         return result;
2462 }
2463 EXPORT_SYMBOL(ldlm_error2errno);
2464
2465 /**
2466  * Dual to ldlm_error2errno(): maps errno values back to ldlm_error_t.
2467  */
2468 ldlm_error_t ldlm_errno2error(int err_no)
2469 {
2470         int error;
2471
2472         switch (err_no) {
2473         case 0:
2474                 error = ELDLM_OK;
2475                 break;
2476         case -ESTALE:
2477                 error = ELDLM_LOCK_CHANGED;
2478                 break;
2479         case -ENAVAIL:
2480                 error = ELDLM_LOCK_ABORTED;
2481                 break;
2482         case -ESRCH:
2483                 error = ELDLM_LOCK_REPLACED;
2484                 break;
2485         case -ENOENT:
2486                 error = ELDLM_NO_LOCK_DATA;
2487                 break;
2488         case -EEXIST:
2489                 error = ELDLM_NAMESPACE_EXISTS;
2490                 break;
2491         case -EBADF:
2492                 error = ELDLM_BAD_NAMESPACE;
2493                 break;
2494         default:
2495                 error = err_no;
2496         }
2497         return error;
2498 }
2499 EXPORT_SYMBOL(ldlm_errno2error);
2500