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                 item->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 static void 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         } else if (!strcmp(name, LUSTRE_MDC_NAME)) {
 225                 rq_portal = MDS_REQUEST_PORTAL;
 226                 rp_portal = MDC_REPLY_PORTAL;
 227                 connect_op = MDS_CONNECT;
 228         } else if (!strcmp(name, LUSTRE_MGC_NAME)) {
 229                 rq_portal = MGS_REQUEST_PORTAL;
 230                 rp_portal = MGC_REPLY_PORTAL;
 231                 connect_op = MGS_CONNECT;
 232         } else {
 233                 CERROR("unknown client OBD type \"%s\", can't setup\n",
 234                        name);
 235                 RETURN(-EINVAL);
 236         }
 237
 238         if (LUSTRE_CFG_BUFLEN(lcfg, 1) < 1) {
 239                 CERROR("requires a TARGET UUID\n");
 240                 RETURN(-EINVAL);
 241         }
 242
 243         if (LUSTRE_CFG_BUFLEN(lcfg, 1) > 37) {
 244                 CERROR("client UUID must be less than 38 characters\n");
 245                 RETURN(-EINVAL);
 246         }
 247
 248         if (LUSTRE_CFG_BUFLEN(lcfg, 2) < 1) {
 249                 CERROR("setup requires a SERVER UUID\n");
 250                 RETURN(-EINVAL);
 251         }
 252
 253         if (LUSTRE_CFG_BUFLEN(lcfg, 2) > 37) {
 254                 CERROR("target UUID must be less than 38 characters\n");
 255                 RETURN(-EINVAL);
 256         }
 257
 258         init_rwsem(&cli->cl_sem);
 259         sema_init(&cli->cl_mgc_sem, 1);
 260         sptlrpc_rule_set_init(&cli->cl_sptlrpc_rset);
 261         cli->cl_sec_part = LUSTRE_SP_ANY;
 262         cli->cl_conn_count = 0;
 263         memcpy(server_uuid.uuid, lustre_cfg_buf(lcfg, 2),
 264                min_t(unsigned int, LUSTRE_CFG_BUFLEN(lcfg, 2),
 265                      sizeof(server_uuid)));
 266
 267         cli->cl_dirty = 0;
 268         cli->cl_avail_grant = 0;
 269         /* FIXME: should limit this for the sum of all cl_dirty_max */
 270         cli->cl_dirty_max = OSC_MAX_DIRTY_DEFAULT * 1024 * 1024;
 271         if (cli->cl_dirty_max >> CFS_PAGE_SHIFT > num_physpages / 8)
 272                 cli->cl_dirty_max = num_physpages << (CFS_PAGE_SHIFT - 3);
 273         CFS_INIT_LIST_HEAD(&cli->cl_cache_waiters);
 274         CFS_INIT_LIST_HEAD(&cli->cl_loi_ready_list);
 275         CFS_INIT_LIST_HEAD(&cli->cl_loi_write_list);
 276         CFS_INIT_LIST_HEAD(&cli->cl_loi_read_list);
 277         client_obd_list_lock_init(&cli->cl_loi_list_lock);
 278         cli->cl_r_in_flight = 0;
 279         cli->cl_w_in_flight = 0;
 280
 281         spin_lock_init(&cli->cl_read_rpc_hist.oh_lock);
 282         spin_lock_init(&cli->cl_write_rpc_hist.oh_lock);
 283         spin_lock_init(&cli->cl_read_page_hist.oh_lock);
 284         spin_lock_init(&cli->cl_write_page_hist.oh_lock);
 285         spin_lock_init(&cli->cl_read_offset_hist.oh_lock);
 286         spin_lock_init(&cli->cl_write_offset_hist.oh_lock);
 287         cfs_waitq_init(&cli->cl_destroy_waitq);
 288         atomic_set(&cli->cl_destroy_in_flight, 0);
 289 #ifdef ENABLE_CHECKSUM
 290         /* Turn on checksumming by default. */
 291         cli->cl_checksum = 1;
 292         /*
 293          * The supported checksum types will be worked out at connect time
 294          * Set cl_chksum* to CRC32 for now to avoid returning screwed info
 295          * through procfs.
 296          */
 297         cli->cl_cksum_type = cli->cl_supp_cksum_types = OBD_CKSUM_CRC32;
 298 #endif
 299         atomic_set(&cli->cl_resends, OSC_DEFAULT_RESENDS);
 300
 301         /* This value may be changed at connect time in
 302            ptlrpc_connect_interpret. */
 303         cli->cl_max_pages_per_rpc = min((int)PTLRPC_MAX_BRW_PAGES,
 304                                         (int)(1024 * 1024 >> CFS_PAGE_SHIFT));
 305
 306         if (!strcmp(name, LUSTRE_MDC_NAME)) {
 307                 cli->cl_max_rpcs_in_flight = MDC_MAX_RIF_DEFAULT;
 308         } else if (num_physpages >> (20 - CFS_PAGE_SHIFT) <= 128 /* MB */) {
 309                 cli->cl_max_rpcs_in_flight = 2;
 310         } else if (num_physpages >> (20 - CFS_PAGE_SHIFT) <= 256 /* MB */) {
 311                 cli->cl_max_rpcs_in_flight = 3;
 312         } else if (num_physpages >> (20 - CFS_PAGE_SHIFT) <= 512 /* MB */) {
 313                 cli->cl_max_rpcs_in_flight = 4;
 314         } else {
 315                 cli->cl_max_rpcs_in_flight = OSC_MAX_RIF_DEFAULT;
 316         }
 317
 318         rc = ldlm_get_ref();
 319         if (rc) {
 320                 CERROR("ldlm_get_ref failed: %d\n", rc);
 321                 GOTO(err, rc);
 322         }
 323
 324         ptlrpc_init_client(rq_portal, rp_portal, name,
 325                            &obddev->obd_ldlm_client);
 326
 327         imp = class_new_import(obddev);
 328         if (imp == NULL)
 329                 GOTO(err_ldlm, rc = -ENOENT);
 330         imp->imp_client = &obddev->obd_ldlm_client;
 331         imp->imp_connect_op = connect_op;
 332         imp->imp_initial_recov = 1;
 333         imp->imp_initial_recov_bk = 0;
 334         CFS_INIT_LIST_HEAD(&imp->imp_pinger_chain);
 335         memcpy(cli->cl_target_uuid.uuid, lustre_cfg_buf(lcfg, 1),
 336                LUSTRE_CFG_BUFLEN(lcfg, 1));
 337         class_import_put(imp);
 338
 339         rc = client_import_add_conn(imp, &server_uuid, 1);
 340         if (rc) {
 341                 CERROR("can't add initial connection\n");
 342                 GOTO(err_import, rc);
 343         }
 344
 345         cli->cl_import = imp;
 346         /* cli->cl_max_mds_{easize,cookiesize} updated by mdc_init_ea_size() */
 347         cli->cl_max_mds_easize = sizeof(struct lov_mds_md_v3);
 348         cli->cl_max_mds_cookiesize = sizeof(struct llog_cookie);
 349
 350         if (LUSTRE_CFG_BUFLEN(lcfg, 3) > 0) {
 351                 if (!strcmp(lustre_cfg_string(lcfg, 3), "inactive")) {
 352                         CDEBUG(D_HA, "marking %s %s->%s as inactive\n",
 353                                name, obddev->obd_name,
 354                                cli->cl_target_uuid.uuid);
 355                         spin_lock(&imp->imp_lock);
 356                         imp->imp_invalid = 1;
 357                         spin_unlock(&imp->imp_lock);
 358                 }
 359         }
 360
 361         cli->cl_qchk_stat = CL_NOT_QUOTACHECKED;
 362
 363         RETURN(rc);
 364
 365 err_import:
 366         class_destroy_import(imp);
 367 err_ldlm:
 368         ldlm_put_ref();
 369 err:
 370         RETURN(rc);
 371
 372 }
 373
 374 int client_obd_cleanup(struct obd_device *obddev)
 375 {
 376         ENTRY;
 377         sptlrpc_rule_set_free(&obddev->u.cli.cl_sptlrpc_rset);
 378         ldlm_put_ref();
 379         RETURN(0);
 380 }
 381
 382 /* ->o_connect() method for client side (OSC and MDC and MGC) */
 383 int client_connect_import(const struct lu_env *env,
 384                           struct lustre_handle *dlm_handle,
 385                           struct obd_device *obd, struct obd_uuid *cluuid,
 386                           struct obd_connect_data *data, void *localdata)
 387 {
 388         struct client_obd *cli = &obd->u.cli;
 389         struct obd_import *imp = cli->cl_import;
 390         struct obd_export *exp;
 391         struct obd_connect_data *ocd;
 392         struct ldlm_namespace *to_be_freed = NULL;
 393         int rc;
 394         ENTRY;
 395
 396         down_write(&cli->cl_sem);
 397         rc = class_connect(dlm_handle, obd, cluuid);
 398         if (rc)
 399                 GOTO(out_sem, rc);
 400
 401         cli->cl_conn_count++;
 402         if (cli->cl_conn_count > 1)
 403                 GOTO(out_sem, rc);
 404         exp = class_conn2export(dlm_handle);
 405
 406         if (obd->obd_namespace != NULL)
 407                 CERROR("already have namespace!\n");
 408         obd->obd_namespace = ldlm_namespace_new(obd, obd->obd_name,
 409                                                 LDLM_NAMESPACE_CLIENT,
 410                                                 LDLM_NAMESPACE_GREEDY);
 411         if (obd->obd_namespace == NULL)
 412                 GOTO(out_disco, rc = -ENOMEM);
 413
 414         imp->imp_dlm_handle = *dlm_handle;
 415         rc = ptlrpc_init_import(imp);
 416         if (rc != 0)
 417                 GOTO(out_ldlm, rc);
 418
 419         ocd = &imp->imp_connect_data;
 420         if (data) {
 421                 *ocd = *data;
 422                 imp->imp_connect_flags_orig = data->ocd_connect_flags;
 423         }
 424
 425         rc = ptlrpc_connect_import(imp, NULL);
 426         if (rc != 0) {
 427                 LASSERT (imp->imp_state == LUSTRE_IMP_DISCON);
 428                 GOTO(out_ldlm, rc);
 429         }
 430         LASSERT(exp->exp_connection);
 431
 432         if (data) {
 433                 LASSERTF((ocd->ocd_connect_flags & data->ocd_connect_flags) ==
 434                          ocd->ocd_connect_flags, "old "LPX64", new "LPX64"\n",
 435                          data->ocd_connect_flags, ocd->ocd_connect_flags);
 436                 data->ocd_connect_flags = ocd->ocd_connect_flags;
 437         }
 438
 439         ptlrpc_pinger_add_import(imp);
 440
 441         EXIT;
 442
 443         if (rc) {
 444 out_ldlm:
 445                 ldlm_namespace_free_prior(obd->obd_namespace, imp, 0);
 446                 to_be_freed = obd->obd_namespace;
 447                 obd->obd_namespace = NULL;
 448 out_disco:
 449                 cli->cl_conn_count--;
 450                 class_disconnect(exp);
 451         } else {
 452                 class_export_put(exp);
 453         }
 454 out_sem:
 455         up_write(&cli->cl_sem);
 456         if (to_be_freed)
 457                 ldlm_namespace_free_post(to_be_freed);
 458
 459         return rc;
 460 }
 461
 462 int client_disconnect_export(struct obd_export *exp)
 463 {
 464         struct obd_device *obd = class_exp2obd(exp);
 465         struct client_obd *cli;
 466         struct obd_import *imp;
 467         int rc = 0, err;
 468         struct ldlm_namespace *to_be_freed = NULL;
 469         ENTRY;
 470
 471         if (!obd) {
 472                 CERROR("invalid export for disconnect: exp %p cookie "LPX64"\n",
 473                        exp, exp ? exp->exp_handle.h_cookie : -1);
 474                 RETURN(-EINVAL);
 475         }
 476
 477         cli = &obd->u.cli;
 478         imp = cli->cl_import;
 479
 480         down_write(&cli->cl_sem);
 481         if (!cli->cl_conn_count) {
 482                 CERROR("disconnecting disconnected device (%s)\n",
 483                        obd->obd_name);
 484                 GOTO(out_sem, rc = -EINVAL);
 485         }
 486
 487         cli->cl_conn_count--;
 488         if (cli->cl_conn_count)
 489                 GOTO(out_no_disconnect, rc = 0);
 490
 491         /* Mark import deactivated now, so we don't try to reconnect if any
 492          * of the cleanup RPCs fails (e.g. ldlm cancel, etc).  We don't
 493          * fully deactivate the import, or that would drop all requests. */
 494         spin_lock(&imp->imp_lock);
 495         imp->imp_deactive = 1;
 496         spin_unlock(&imp->imp_lock);
 497
 498         /* Some non-replayable imports (MDS's OSCs) are pinged, so just
 499          * delete it regardless.  (It's safe to delete an import that was
 500          * never added.) */
 501         (void)ptlrpc_pinger_del_import(imp);
 502
 503         if (obd->obd_namespace != NULL) {
 504                 /* obd_force == local only */
 505                 ldlm_cli_cancel_unused(obd->obd_namespace, NULL,
 506                                        obd->obd_force ? LDLM_FL_LOCAL_ONLY:0,
 507                                        NULL);
 508                 ldlm_namespace_free_prior(obd->obd_namespace, imp, obd->obd_force);
 509                 to_be_freed = obd->obd_namespace;
 510         }
 511
 512         rc = ptlrpc_disconnect_import(imp, 0);
 513
 514         ptlrpc_invalidate_import(imp);
 515         /* set obd_namespace to NULL only after invalidate, because we can have
 516          * some connect requests in flight, and his need store a connect flags
 517          * in obd_namespace. bug 14260 */
 518         obd->obd_namespace = NULL;
 519
 520         if (imp->imp_rq_pool) {
 521                 ptlrpc_free_rq_pool(imp->imp_rq_pool);
 522                 imp->imp_rq_pool = NULL;
 523         }
 524         destroy_import(imp);
 525         cli->cl_import = NULL;
 526
 527         EXIT;
 528  out_no_disconnect:
 529         err = class_disconnect(exp);
 530         if (!rc && err)
 531                 rc = err;
 532  out_sem:
 533         up_write(&cli->cl_sem);
 534         if (to_be_freed)
 535                 ldlm_namespace_free_post(to_be_freed);
 536
 537         RETURN(rc);
 538 }
 539
 540 /* --------------------------------------------------------------------------
 541  * from old lib/target.c
 542  * -------------------------------------------------------------------------- */
 543
 544 int target_handle_reconnect(struct lustre_handle *conn, struct obd_export *exp,
 545                             struct obd_uuid *cluuid, int initial_conn)
 546 {
 547         ENTRY;
 548         if (exp->exp_connection && exp->exp_imp_reverse && !initial_conn) {
 549                 struct lustre_handle *hdl;
 550                 hdl = &exp->exp_imp_reverse->imp_remote_handle;
 551                 /* Might be a re-connect after a partition. */
 552                 if (!memcmp(&conn->cookie, &hdl->cookie, sizeof conn->cookie)) {
 553                         CWARN("%s: %s reconnecting\n", exp->exp_obd->obd_name,
 554                               cluuid->uuid);
 555                         conn->cookie = exp->exp_handle.h_cookie;
 556                         /* target_handle_connect() treats EALREADY and
 557                          * -EALREADY differently.  EALREADY means we are
 558                          * doing a valid reconnect from the same client. */
 559                         RETURN(EALREADY);
 560                 } else {
 561                         CERROR("%s reconnecting from %s, "
 562                                "handle mismatch (ours "LPX64", theirs "
 563                                LPX64")\n", cluuid->uuid,
 564                                exp->exp_connection->c_remote_uuid.uuid,
 565                                hdl->cookie, conn->cookie);
 566                         memset(conn, 0, sizeof *conn);
 567                         /* target_handle_connect() treats EALREADY and
 568                          * -EALREADY differently.  -EALREADY is an error
 569                          * (same UUID, different handle). */
 570                         RETURN(-EALREADY);
 571                 }
 572         }
 573
 574         conn->cookie = exp->exp_handle.h_cookie;
 575         CDEBUG(D_HA, "connect export for UUID '%s' at %p, cookie "LPX64"\n",
 576                cluuid->uuid, exp, conn->cookie);
 577         RETURN(0);
 578 }
 579
 580 void target_client_add_cb(struct obd_device *obd, __u64 transno, void *cb_data,
 581                           int error)
 582 {
 583         struct obd_export *exp = cb_data;
 584
 585         CDEBUG(D_RPCTRACE, "%s: committing for initial connect of %s\n",
 586                obd->obd_name, exp->exp_client_uuid.uuid);
 587
 588         spin_lock(&exp->exp_lock);
 589         exp->exp_need_sync = 0;
 590         spin_unlock(&exp->exp_lock);
 591 }
 592 EXPORT_SYMBOL(target_client_add_cb);
 593 static void
 594 target_start_and_reset_recovery_timer(struct obd_device *obd,
 595                                       struct ptlrpc_request *req,
 596                                       int new_client);
 597
 598 int target_handle_connect(struct ptlrpc_request *req)
 599 {
 600         struct obd_device *target, *targref = NULL;
 601         struct obd_export *export = NULL;
 602         struct obd_import *revimp;
 603         struct lustre_handle conn;
 604         struct lustre_handle *tmp;
 605         struct obd_uuid tgtuuid;
 606         struct obd_uuid cluuid;
 607         struct obd_uuid remote_uuid;
 608         char *str;
 609         int rc = 0;
 610         int initial_conn = 0;
 611         struct obd_connect_data *data, *tmpdata;
 612         lnet_nid_t *client_nid = NULL;
 613         ENTRY;
 614
 615         OBD_RACE(OBD_FAIL_TGT_CONN_RACE);
 616
 617         str = req_capsule_client_get(&req->rq_pill, &RMF_TGTUUID);
 618         if (str == NULL) {
 619                 DEBUG_REQ(D_ERROR, req, "bad target UUID for connect");
 620                 GOTO(out, rc = -EINVAL);
 621         }
 622
 623         obd_str2uuid(&tgtuuid, str);
 624         target = class_uuid2obd(&tgtuuid);
 625         if (!target)
 626                 target = class_name2obd(str);
 627
 628         if (!target || target->obd_stopping || !target->obd_set_up) {
 629                 LCONSOLE_ERROR_MSG(0x137, "UUID '%s' is not available "
 630                                    " for connect (%s)\n", str,
 631                                    !target ? "no target" :
 632                                    (target->obd_stopping ? "stopping" :
 633                                    "not set up"));
 634                 GOTO(out, rc = -ENODEV);
 635         }
 636
 637         if (target->obd_no_conn) {
 638                 LCONSOLE_WARN("%s: temporarily refusing client connection "
 639                               "from %s\n", target->obd_name,
 640                               libcfs_nid2str(req->rq_peer.nid));
 641                 GOTO(out, rc = -EAGAIN);
 642         }
 643
 644         /* Make sure the target isn't cleaned up while we're here. Yes,
 645            there's still a race between the above check and our incref here.
 646            Really, class_uuid2obd should take the ref. */
 647         targref = class_incref(target, __FUNCTION__, cfs_current());
 648
 649
 650         str = req_capsule_client_get(&req->rq_pill, &RMF_CLUUID);
 651         if (str == NULL) {
 652                 DEBUG_REQ(D_ERROR, req, "bad client UUID for connect");
 653                 GOTO(out, rc = -EINVAL);
 654         }
 655
 656         obd_str2uuid(&cluuid, str);
 657
 658         /* XXX extract a nettype and format accordingly */
 659         switch (sizeof(lnet_nid_t)) {
 660                 /* NB the casts only avoid compiler warnings */
 661         case 8:
 662                 snprintf(remote_uuid.uuid, sizeof remote_uuid,
 663                          "NET_"LPX64"_UUID", (__u64)req->rq_peer.nid);
 664                 break;
 665         case 4:
 666                 snprintf(remote_uuid.uuid, sizeof remote_uuid,
 667                          "NET_%x_UUID", (__u32)req->rq_peer.nid);
 668                 break;
 669         default:
 670                 LBUG();
 671         }
 672
 673         tmp = req_capsule_client_get(&req->rq_pill, &RMF_CONN);
 674         if (tmp == NULL)
 675                 GOTO(out, rc = -EPROTO);
 676
 677         conn = *tmp;
 678
 679         data = req_capsule_client_get(&req->rq_pill, &RMF_CONNECT_DATA);
 680         if (!data)
 681                 GOTO(out, rc = -EPROTO);
 682
 683         rc = req_capsule_server_pack(&req->rq_pill);
 684         if (rc)
 685                 GOTO(out, rc);
 686
 687         if (lustre_msg_get_op_flags(req->rq_reqmsg) & MSG_CONNECT_LIBCLIENT) {
 688                 if (!data) {
 689                         DEBUG_REQ(D_WARNING, req, "Refusing old (unversioned) "
 690                                   "libclient connection attempt");
 691                         GOTO(out, rc = -EPROTO);
 692                 } else if (data->ocd_version < LUSTRE_VERSION_CODE -
 693                                                LUSTRE_VERSION_ALLOWED_OFFSET ||
 694                            data->ocd_version > LUSTRE_VERSION_CODE +
 695                                                LUSTRE_VERSION_ALLOWED_OFFSET) {
 696                         DEBUG_REQ(D_WARNING, req, "Refusing %s (%d.%d.%d.%d) "
 697                                   "libclient connection attempt",
 698                                   data->ocd_version < LUSTRE_VERSION_CODE ?
 699                                   "old" : "new",
 700                                   OBD_OCD_VERSION_MAJOR(data->ocd_version),
 701                                   OBD_OCD_VERSION_MINOR(data->ocd_version),
 702                                   OBD_OCD_VERSION_PATCH(data->ocd_version),
 703                                   OBD_OCD_VERSION_FIX(data->ocd_version));
 704                         data = req_capsule_server_sized_get(&req->rq_pill,
 705                                                         &RMF_CONNECT_DATA,
 706                                     offsetof(typeof(*data), ocd_version) +
 707                                              sizeof(data->ocd_version));
 708                         if (data) {
 709                                 data->ocd_connect_flags = OBD_CONNECT_VERSION;
 710                                 data->ocd_version = LUSTRE_VERSION_CODE;
 711                         }
 712                         GOTO(out, rc = -EPROTO);
 713                 }
 714         }
 715
 716         if (lustre_msg_get_op_flags(req->rq_reqmsg) & MSG_CONNECT_INITIAL)
 717                 initial_conn = 1;
 718
 719         /* lctl gets a backstage, all-access pass. */
 720         if (obd_uuid_equals(&cluuid, &target->obd_uuid))
 721                 goto dont_check_exports;
 722
 723         spin_lock(&target->obd_dev_lock);
 724         export = lustre_hash_lookup(target->obd_uuid_hash, &cluuid);
 725
 726         if (export != NULL && export->exp_connecting) { /* bug 9635, et. al. */
 727                 CWARN("%s: exp %p already connecting\n",
 728                       export->exp_obd->obd_name, export);
 729                 class_export_put(export);
 730                 export = NULL;
 731                 rc = -EALREADY;
 732         } else if (export != NULL && export->exp_connection != NULL &&
 733                    req->rq_peer.nid != export->exp_connection->c_peer.nid) {
 734                 /* make darn sure this is coming from the same peer
 735                  * if the UUIDs matched */
 736                   CWARN("%s: cookie %s seen on new NID %s when "
 737                           "existing NID %s is already connected\n",
 738                         target->obd_name, cluuid.uuid,
 739                   libcfs_nid2str(req->rq_peer.nid),
 740                   libcfs_nid2str(export->exp_connection->c_peer.nid));
 741                   class_export_put(export);
 742                   export = NULL;
 743                   rc = -EALREADY;
 744         } else if (export != NULL) {
 745                 spin_lock(&export->exp_lock);
 746                 export->exp_connecting = 1;
 747                 spin_unlock(&export->exp_lock);
 748                 class_export_put(export);
 749                 spin_unlock(&target->obd_dev_lock);
 750                 LASSERT(export->exp_obd == target);
 751
 752                 rc = target_handle_reconnect(&conn, export, &cluuid, initial_conn);
 753         }
 754
 755         /* If we found an export, we already unlocked. */
 756         if (!export) {
 757                 spin_unlock(&target->obd_dev_lock);
 758                 OBD_FAIL_TIMEOUT(OBD_FAIL_TGT_DELAY_CONNECT, 2 * obd_timeout);
 759         } else if (req->rq_export == NULL &&
 760                    atomic_read(&export->exp_rpc_count) > 0) {
 761                 CWARN("%s: refuse connection from %s/%s to 0x%p/%d\n",
 762                       target->obd_name, cluuid.uuid,
 763                       libcfs_nid2str(req->rq_peer.nid),
 764                       export, atomic_read(&export->exp_refcount));
 765                 GOTO(out, rc = -EBUSY);
 766         } else if (req->rq_export != NULL &&
 767                    (atomic_read(&export->exp_rpc_count) > 1)) {
 768                 CWARN("%s: refuse reconnection from %s@%s to 0x%p/%d\n",
 769                       target->obd_name, cluuid.uuid,
 770                       libcfs_nid2str(req->rq_peer.nid),
 771                       export, atomic_read(&export->exp_rpc_count));
 772                 GOTO(out, rc = -EBUSY);
 773         } else if (lustre_msg_get_conn_cnt(req->rq_reqmsg) == 1 &&
 774                    !initial_conn) {
 775                 CERROR("%s: NID %s (%s) reconnected with 1 conn_cnt; "
 776                        "cookies not random?\n", target->obd_name,
 777                        libcfs_nid2str(req->rq_peer.nid), cluuid.uuid);
 778                 GOTO(out, rc = -EALREADY);
 779         } else {
 780                 OBD_FAIL_TIMEOUT(OBD_FAIL_TGT_DELAY_RECONNECT, 2 * obd_timeout);
 781                 if (req->rq_export == NULL && initial_conn)
 782                        export->exp_last_request_time =
 783                                max(export->exp_last_request_time,
 784                                    (time_t)cfs_time_current_sec());
 785         }
 786
 787         if (rc < 0) {
 788                 GOTO(out, rc);
 789         }
 790
 791         CWARN("%s: connection from %s@%s %st"LPU64" exp %p cur %ld last %ld\n",
 792                target->obd_name, cluuid.uuid, libcfs_nid2str(req->rq_peer.nid),
 793               target->obd_recovering ? "recovering/" : "", data->ocd_transno,
 794               export, (long)cfs_time_current_sec(),
 795               export ? (long)export->exp_last_request_time : 0);
 796
 797         /* Tell the client if we're in recovery. */
 798         if (target->obd_recovering) {
 799                 lustre_msg_add_op_flags(req->rq_repmsg, MSG_CONNECT_RECOVERING);
 800                 /* If this is the first time a client connects,
 801                    reset the recovery timer */
 802                 if (rc == 0)
 803                         target_start_and_reset_recovery_timer(target, req,
 804                                                               !export);
 805         }
 806
 807         /* We want to handle EALREADY but *not* -EALREADY from
 808          * target_handle_reconnect(), return reconnection state in a flag */
 809         if (rc == EALREADY) {
 810                 lustre_msg_add_op_flags(req->rq_repmsg, MSG_CONNECT_RECONNECT);
 811                 rc = 0;
 812         } else {
 813                 LASSERT(rc == 0);
 814         }
 815
 816         /* Tell the client if we support replayable requests */
 817         if (target->obd_replayable)
 818                 lustre_msg_add_op_flags(req->rq_repmsg, MSG_CONNECT_REPLAYABLE);
 819         client_nid = &req->rq_peer.nid;
 820
 821         if (export == NULL) {
 822                 if (target->obd_recovering) {
 823                         cfs_time_t t;
 824
 825                         t = cfs_timer_deadline(&target->obd_recovery_timer);
 826                         t = cfs_time_sub(t, cfs_time_current());
 827                         CERROR("%s: denying connection for new client %s (%s): "
 828                                "%d clients in recovery for "CFS_TIME_T"s\n",
 829                                target->obd_name,
 830                                libcfs_nid2str(req->rq_peer.nid), cluuid.uuid,
 831                                target->obd_recoverable_clients,
 832                                cfs_duration_sec(t));
 833                         rc = -EBUSY;
 834                 } else {
 835 dont_check_exports:
 836                         rc = obd_connect(req->rq_svc_thread->t_env,
 837                                          &conn, target, &cluuid, data,
 838                                          client_nid);
 839                 }
 840         } else {
 841                 rc = obd_reconnect(req->rq_svc_thread->t_env,
 842                                    export, target, &cluuid, data, client_nid);
 843         }
 844         if (rc)
 845                 GOTO(out, rc);
 846         /* Return only the parts of obd_connect_data that we understand, so the
 847          * client knows that we don't understand the rest. */
 848         if (data) {
 849                  tmpdata = req_capsule_server_get(&req->rq_pill,
 850                                                   &RMF_CONNECT_DATA);
 851                   //data->ocd_connect_flags &= OBD_CONNECT_SUPPORTED;
 852                  *tmpdata = *data;
 853         }
 854
 855         /* If all else goes well, this is our RPC return code. */
 856         req->rq_status = 0;
 857
 858         lustre_msg_set_handle(req->rq_repmsg, &conn);
 859
 860         /* ownership of this export ref transfers to the request AFTER we
 861          * drop any previous reference the request had, but we don't want
 862          * that to go to zero before we get our new export reference. */
 863         export = class_conn2export(&conn);
 864         if (!export) {
 865                 DEBUG_REQ(D_ERROR, req, "Missing export!");
 866                 GOTO(out, rc = -ENODEV);
 867         }
 868
 869         /* If the client and the server are the same node, we will already
 870          * have an export that really points to the client's DLM export,
 871          * because we have a shared handles table.
 872          *
 873          * XXX this will go away when shaver stops sending the "connect" handle
 874          * in the real "remote handle" field of the request --phik 24 Apr 2003
 875          */
 876         if (req->rq_export != NULL)
 877                 class_export_put(req->rq_export);
 878
 879         req->rq_export = export;
 880
 881         spin_lock(&export->exp_lock);
 882         if (initial_conn) {
 883                 lustre_msg_set_conn_cnt(req->rq_repmsg, export->exp_conn_cnt + 1);
 884         } else if (export->exp_conn_cnt >= lustre_msg_get_conn_cnt(req->rq_reqmsg)) {
 885                 spin_unlock(&export->exp_lock);
 886                 CERROR("%s: %s already connected at higher conn_cnt: %d > %d\n",
 887                        cluuid.uuid, libcfs_nid2str(req->rq_peer.nid),
 888                        export->exp_conn_cnt,
 889                        lustre_msg_get_conn_cnt(req->rq_reqmsg));
 890
 891                 GOTO(out, rc = -EALREADY);
 892         }
 893         export->exp_conn_cnt = lustre_msg_get_conn_cnt(req->rq_reqmsg);
 894
 895         /* request from liblustre?  Don't evict it for not pinging. */
 896         if (lustre_msg_get_op_flags(req->rq_reqmsg) & MSG_CONNECT_LIBCLIENT) {
 897                 export->exp_libclient = 1;
 898                 spin_unlock(&export->exp_lock);
 899
 900                 spin_lock(&target->obd_dev_lock);
 901                 list_del_init(&export->exp_obd_chain_timed);
 902                 spin_unlock(&target->obd_dev_lock);
 903         } else {
 904                 spin_unlock(&export->exp_lock);
 905         }
 906
 907         if (export->exp_connection != NULL)
 908                 ptlrpc_connection_put(export->exp_connection);
 909         export->exp_connection = ptlrpc_connection_get(req->rq_peer,
 910                                                        req->rq_self,
 911                                                        &remote_uuid);
 912         if (hlist_unhashed(&export->exp_nid_hash)) {
 913                 lustre_hash_add_unique(export->exp_obd->obd_nid_hash,
 914                                        &export->exp_connection->c_peer.nid,
 915                                        &export->exp_nid_hash);
 916         }
 917
 918         spin_lock_bh(&target->obd_processing_task_lock);
 919         if (target->obd_recovering && !export->exp_in_recovery) {
 920                 spin_lock(&export->exp_lock);
 921                 export->exp_in_recovery = 1;
 922                 export->exp_req_replay_needed = 1;
 923                 export->exp_lock_replay_needed = 1;
 924                 spin_unlock(&export->exp_lock);
 925                 if ((lustre_msg_get_op_flags(req->rq_reqmsg) & MSG_CONNECT_TRANSNO)
 926                      && (data->ocd_transno == 0))
 927                         CWARN("Connect with zero transno!\n");
 928
 929                 if ((lustre_msg_get_op_flags(req->rq_reqmsg) & MSG_CONNECT_TRANSNO)
 930                      && data->ocd_transno < target->obd_next_recovery_transno)
 931                         target->obd_next_recovery_transno = data->ocd_transno;
 932                 target->obd_connected_clients++;
 933                 /* each connected client is counted as recoverable */
 934                 target->obd_recoverable_clients++;
 935                 atomic_inc(&target->obd_req_replay_clients);
 936                 atomic_inc(&target->obd_lock_replay_clients);
 937                 if (target->obd_connected_clients ==
 938                     target->obd_max_recoverable_clients)
 939                         wake_up(&target->obd_next_transno_waitq);
 940         }
 941         spin_unlock_bh(&target->obd_processing_task_lock);
 942         tmp = req_capsule_client_get(&req->rq_pill, &RMF_CONN);
 943         conn = *tmp;
 944
 945         if (export->exp_imp_reverse != NULL) {
 946                 /* destroyed import can be still referenced in ctxt */
 947                 obd_set_info_async(export, sizeof(KEY_REVIMP_UPD),
 948                                    KEY_REVIMP_UPD, 0, NULL, NULL);
 949
 950                 /* in some recovery senarios, previous ctx init rpc handled
 951                  * in sptlrpc_target_export_check() might be used to install
 952                  * a reverse ctx in this reverse import, and later OBD_CONNECT
 953                  * using the same gss ctx could reach here and following new
 954                  * reverse import. note all reverse ctx in new/old import are
 955                  * actually based on the same gss ctx. so we invalidate ctx
 956                  * here before destroy import, otherwise flush old import will
 957                  * lead to remote reverse ctx be destroied, thus the reverse
 958                  * ctx of new import will lost its peer.
 959                  * there might be a better way to deal with this???
 960                  */
 961                 sptlrpc_import_inval_all_ctx(export->exp_imp_reverse);
 962
 963                 destroy_import(export->exp_imp_reverse);
 964         }
 965
 966         /* for the rest part, we return -ENOTCONN in case of errors
 967          * in order to let client initialize connection again.
 968          */
 969         revimp = export->exp_imp_reverse = class_new_import(target);
 970         if (!revimp) {
 971                 CERROR("fail to alloc new reverse import.\n");
 972                 GOTO(out, rc = -ENOTCONN);
 973         }
 974
 975         revimp->imp_connection = ptlrpc_connection_addref(export->exp_connection);
 976         revimp->imp_client = &export->exp_obd->obd_ldlm_client;
 977         revimp->imp_remote_handle = conn;
 978         revimp->imp_dlm_fake = 1;
 979         revimp->imp_state = LUSTRE_IMP_FULL;
 980
 981         /* unknown versions will be caught in
 982          * ptlrpc_handle_server_req_in->lustre_unpack_msg() */
 983         revimp->imp_msg_magic = req->rq_reqmsg->lm_magic;
 984
 985         if ((export->exp_connect_flags & OBD_CONNECT_AT) &&
 986             (revimp->imp_msg_magic != LUSTRE_MSG_MAGIC_V1))
 987                 revimp->imp_msghdr_flags |= MSGHDR_AT_SUPPORT;
 988         else
 989                 revimp->imp_msghdr_flags &= ~MSGHDR_AT_SUPPORT;
 990
 991         rc = sptlrpc_import_sec_adapt(revimp, req->rq_svc_ctx,
 992                                       req->rq_flvr.sf_rpc);
 993         if (rc) {
 994                 CERROR("Failed to get sec for reverse import: %d\n", rc);
 995                 export->exp_imp_reverse = NULL;
 996                 class_destroy_import(revimp);
 997         }
 998
 999         class_import_put(revimp);
1000 out:
1001         if (export) {
1002                 spin_lock(&export->exp_lock);
1003                 export->exp_connecting = 0;
1004                 spin_unlock(&export->exp_lock);
1005         }
1006         if (targref)
1007                 class_decref(targref, __FUNCTION__, cfs_current());
1008         if (rc)
1009                 req->rq_status = rc;
1010         RETURN(rc);
1011 }
1012
1013 int target_handle_disconnect(struct ptlrpc_request *req)
1014 {
1015         int rc;
1016         ENTRY;
1017
1018         rc = req_capsule_server_pack(&req->rq_pill);
1019         if (rc)
1020                 RETURN(rc);
1021
1022         /* keep the rq_export around so we can send the reply */
1023         req->rq_status = obd_disconnect(class_export_get(req->rq_export));
1024
1025         RETURN(0);
1026 }
1027
1028 void target_destroy_export(struct obd_export *exp)
1029 {
1030         /* exports created from last_rcvd data, and "fake"
1031            exports created by lctl don't have an import */
1032         if (exp->exp_imp_reverse != NULL)
1033                 destroy_import(exp->exp_imp_reverse);
1034
1035         /* We cancel locks at disconnect time, but this will catch any locks
1036          * granted in a race with recovery-induced disconnect. */
1037         if (exp->exp_obd->obd_namespace != NULL)
1038                 ldlm_cancel_locks_for_export(exp);
1039 }
1040
1041 /*
1042  * Recovery functions
1043  */
1044
1045 struct ptlrpc_request *ptlrpc_clone_req( struct ptlrpc_request *orig_req)
1046 {
1047         struct ptlrpc_request *copy_req;
1048         struct lustre_msg *copy_reqmsg;
1049         struct ptlrpc_user_desc *udesc = NULL;
1050
1051         OBD_ALLOC_PTR(copy_req);
1052         if (!copy_req)
1053                 return NULL;
1054         OBD_ALLOC(copy_reqmsg, orig_req->rq_reqlen);
1055         if (!copy_reqmsg){
1056                 OBD_FREE_PTR(copy_req);
1057                 return NULL;
1058         }
1059
1060         if (orig_req->rq_user_desc) {
1061                 int ngroups = orig_req->rq_user_desc->pud_ngroups;
1062
1063                 OBD_ALLOC(udesc, sptlrpc_user_desc_size(ngroups));
1064                 if (!udesc) {
1065                         OBD_FREE(copy_reqmsg, orig_req->rq_reqlen);
1066                         OBD_FREE_PTR(copy_req);
1067                         return NULL;
1068                 }
1069                 memcpy(udesc, orig_req->rq_user_desc,
1070                        sptlrpc_user_desc_size(ngroups));
1071         }
1072
1073         *copy_req = *orig_req;
1074         memcpy(copy_reqmsg, orig_req->rq_reqmsg, orig_req->rq_reqlen);
1075         copy_req->rq_reqmsg = copy_reqmsg;
1076         copy_req->rq_user_desc = udesc;
1077
1078         class_export_rpc_get(copy_req->rq_export);
1079         CFS_INIT_LIST_HEAD(&copy_req->rq_list);
1080         CFS_INIT_LIST_HEAD(&copy_req->rq_replay_list);
1081         sptlrpc_svc_ctx_addref(copy_req);
1082
1083         if (copy_req->rq_reply_state) {
1084                 /* the copied req takes over the reply state */
1085                 orig_req->rq_reply_state = NULL;
1086                 /* to catch further access */
1087                 orig_req->rq_repmsg = NULL;
1088                 orig_req->rq_replen = 0;
1089         }
1090
1091         return copy_req;
1092 }
1093
1094 void ptlrpc_free_clone(struct ptlrpc_request *req)
1095 {
1096         LASSERT(list_empty(&req->rq_replay_list));
1097
1098         ptlrpc_req_drop_rs(req);
1099         sptlrpc_svc_ctx_decref(req);
1100         class_export_rpc_put(req->rq_export);
1101         list_del_init(&req->rq_list);
1102
1103         if (req->rq_user_desc) {
1104                 int ngroups = req->rq_user_desc->pud_ngroups;
1105                 OBD_FREE(req->rq_user_desc, sptlrpc_user_desc_size(ngroups));
1106         }
1107         OBD_FREE(req->rq_reqmsg, req->rq_reqlen);
1108         OBD_FREE_PTR(req);
1109 }
1110
1111 static int target_exp_enqueue_req_replay(struct ptlrpc_request *req)
1112 {
1113         __u64                  transno = lustre_msg_get_transno(req->rq_reqmsg);
1114         struct obd_export     *exp = req->rq_export;
1115         struct ptlrpc_request *reqiter;
1116         int                    dup = 0;
1117
1118         LASSERT(exp);
1119
1120         spin_lock(&exp->exp_lock);
1121         list_for_each_entry(reqiter, &exp->exp_req_replay_queue,
1122                             rq_replay_list) {
1123                 if (lustre_msg_get_transno(reqiter->rq_reqmsg) == transno) {
1124                         dup = 1;
1125                         break;
1126                 }
1127         }
1128
1129         if (dup) {
1130                 /* we expect it with RESENT and REPLAY flags */
1131                 if ((lustre_msg_get_flags(req->rq_reqmsg) &
1132                      (MSG_RESENT | MSG_REPLAY)) != (MSG_RESENT | MSG_REPLAY))
1133                         CERROR("invalid flags %x of resent replay\n",
1134                                lustre_msg_get_flags(req->rq_reqmsg));
1135         } else {
1136                 list_add_tail(&req->rq_replay_list, &exp->exp_req_replay_queue);
1137         }
1138
1139         spin_unlock(&exp->exp_lock);
1140         return dup;
1141 }
1142
1143 static void target_exp_dequeue_req_replay(struct ptlrpc_request *req)
1144 {
1145         LASSERT(!list_empty(&req->rq_replay_list));
1146         LASSERT(req->rq_export);
1147
1148         spin_lock(&req->rq_export->exp_lock);
1149         list_del_init(&req->rq_replay_list);
1150         spin_unlock(&req->rq_export->exp_lock);
1151 }
1152
1153 #ifdef __KERNEL__
1154 static void target_finish_recovery(struct obd_device *obd)
1155 {
1156         ENTRY;
1157         LCONSOLE_INFO("%s: sending delayed replies to recovered clients\n",
1158                       obd->obd_name);
1159
1160         ldlm_reprocess_all_ns(obd->obd_namespace);
1161         spin_lock_bh(&obd->obd_processing_task_lock);
1162         if (list_empty(&obd->obd_req_replay_queue) &&
1163             list_empty(&obd->obd_lock_replay_queue) &&
1164             list_empty(&obd->obd_final_req_queue)) {
1165                 obd->obd_processing_task = 0;
1166         } else {
1167                 CERROR("%s: Recovery queues ( %s%s%s) are empty\n",
1168                        obd->obd_name,
1169                        list_empty(&obd->obd_req_replay_queue) ? "" : "req ",
1170                        list_empty(&obd->obd_lock_replay_queue) ? "" : "lock ",
1171                        list_empty(&obd->obd_final_req_queue) ? "" : "final ");
1172                 spin_unlock_bh(&obd->obd_processing_task_lock);
1173                 LBUG();
1174         }
1175         spin_unlock_bh(&obd->obd_processing_task_lock);
1176
1177         /* when recovery finished, cleanup orphans on mds and ost */
1178         if (OBT(obd) && OBP(obd, postrecov)) {
1179                 int rc = OBP(obd, postrecov)(obd);
1180                 LCONSOLE_WARN("%s: recovery %s: rc %d\n", obd->obd_name,
1181                               rc < 0 ? "failed" : "complete", rc);
1182         }
1183
1184         obd->obd_recovery_end = cfs_time_current_sec();
1185         EXIT;
1186 }
1187
1188 static void abort_req_replay_queue(struct obd_device *obd)
1189 {
1190         struct ptlrpc_request *req, *n;
1191         struct list_head abort_list;
1192
1193         CFS_INIT_LIST_HEAD(&abort_list);
1194         spin_lock_bh(&obd->obd_processing_task_lock);
1195         list_splice_init(&obd->obd_req_replay_queue, &abort_list);
1196         spin_unlock_bh(&obd->obd_processing_task_lock);
1197         list_for_each_entry_safe(req, n, &abort_list, rq_list) {
1198                 DEBUG_REQ(D_WARNING, req, "aborted:");
1199                 req->rq_status = -ENOTCONN;
1200                 if (ptlrpc_error(req)) {
1201                         DEBUG_REQ(D_ERROR, req,
1202                                   "failed abort_req_reply; skipping");
1203                 }
1204                 target_exp_dequeue_req_replay(req);
1205                 ptlrpc_free_clone(req);
1206         }
1207 }
1208
1209 static void abort_lock_replay_queue(struct obd_device *obd)
1210 {
1211         struct ptlrpc_request *req, *n;
1212         struct list_head abort_list;
1213
1214         CFS_INIT_LIST_HEAD(&abort_list);
1215         spin_lock_bh(&obd->obd_processing_task_lock);
1216         list_splice_init(&obd->obd_lock_replay_queue, &abort_list);
1217         spin_unlock_bh(&obd->obd_processing_task_lock);
1218         list_for_each_entry_safe(req, n, &obd->obd_lock_replay_queue, rq_list){
1219                 DEBUG_REQ(D_ERROR, req, "aborted:");
1220                 req->rq_status = -ENOTCONN;
1221                 if (ptlrpc_error(req)) {
1222                         DEBUG_REQ(D_ERROR, req,
1223                                   "failed abort_lock_reply; skipping");
1224                 }
1225                 ptlrpc_free_clone(req);
1226         }
1227 }
1228 #endif
1229
1230 /* Called from a cleanup function if the device is being cleaned up
1231    forcefully.  The exports should all have been disconnected already,
1232    the only thing left to do is
1233      - clear the recovery flags
1234      - cancel the timer
1235      - free queued requests and replies, but don't send replies
1236    Because the obd_stopping flag is set, no new requests should be received.
1237
1238 */
1239 void target_cleanup_recovery(struct obd_device *obd)
1240 {
1241         struct ptlrpc_request *req, *n;
1242         struct list_head clean_list;
1243         ENTRY;
1244
1245         LASSERT(obd->obd_stopping);
1246
1247         CFS_INIT_LIST_HEAD(&clean_list);
1248         spin_lock_bh(&obd->obd_processing_task_lock);
1249         if (!obd->obd_recovering) {
1250                 spin_unlock_bh(&obd->obd_processing_task_lock);
1251                 EXIT;
1252                 return;
1253         }
1254         obd->obd_recovering = obd->obd_abort_recovery = 0;
1255         target_cancel_recovery_timer(obd);
1256
1257         list_splice_init(&obd->obd_req_replay_queue, &clean_list);
1258         spin_unlock_bh(&obd->obd_processing_task_lock);
1259
1260         list_for_each_entry_safe(req, n, &clean_list, rq_list) {
1261                 LASSERT(req->rq_reply_state == 0);
1262                 target_exp_dequeue_req_replay(req);
1263                 ptlrpc_free_clone(req);
1264         }
1265
1266         spin_lock_bh(&obd->obd_processing_task_lock);
1267         list_splice_init(&obd->obd_lock_replay_queue, &clean_list);
1268         list_splice_init(&obd->obd_final_req_queue, &clean_list);
1269         spin_unlock_bh(&obd->obd_processing_task_lock);
1270
1271         list_for_each_entry_safe(req, n, &clean_list, rq_list){
1272                 LASSERT(req->rq_reply_state == 0);
1273                 ptlrpc_free_clone(req);
1274         }
1275
1276         EXIT;
1277 }
1278
1279 /* obd_processing_task_lock should be held */
1280 void target_cancel_recovery_timer(struct obd_device *obd)
1281 {
1282         CDEBUG(D_HA, "%s: cancel recovery timer\n", obd->obd_name);
1283         cfs_timer_disarm(&obd->obd_recovery_timer);
1284 }
1285
1286 /* extend = 1 means require at least "duration" seconds left in the timer,
1287    extend = 0 means set the total duration (start_recovery_timer) */
1288 static void reset_recovery_timer(struct obd_device *obd, int duration,
1289                                  int extend)
1290 {
1291         cfs_time_t now = cfs_time_current_sec();
1292         cfs_duration_t left;
1293
1294         spin_lock_bh(&obd->obd_processing_task_lock);
1295         if (!obd->obd_recovering || obd->obd_abort_recovery) {
1296                 spin_unlock_bh(&obd->obd_processing_task_lock);
1297                 return;
1298         }
1299
1300         left = cfs_time_sub(obd->obd_recovery_end, now);
1301
1302         if (extend && (duration > left))
1303                 obd->obd_recovery_timeout += duration - left;
1304         else if (!extend && (duration > obd->obd_recovery_timeout))
1305                 /* Track the client's largest expected replay time */
1306                 obd->obd_recovery_timeout = duration;
1307 #ifdef CRAY_XT3
1308         /*
1309          * If total recovery time already exceed the
1310          * obd_recovery_max_time, then CRAY XT3 will
1311          * abort the recovery
1312          */
1313         if(obd->obd_recovery_timeout > obd->obd_recovery_max_time)
1314                 obd->obd_recovery_timeout = obd->obd_recovery_max_time;
1315 #endif
1316         obd->obd_recovery_end = obd->obd_recovery_start +
1317                                 obd->obd_recovery_timeout;
1318         if (!cfs_timer_is_armed(&obd->obd_recovery_timer) ||
1319             cfs_time_before(now, obd->obd_recovery_end)) {
1320                 left = cfs_time_sub(obd->obd_recovery_end, now);
1321                 cfs_timer_arm(&obd->obd_recovery_timer, cfs_time_shift(left));
1322         }
1323         spin_unlock_bh(&obd->obd_processing_task_lock);
1324         CDEBUG(D_HA, "%s: recovery timer will expire in %u seconds\n",
1325                obd->obd_name, (unsigned)left);
1326 }
1327
1328 static void check_and_start_recovery_timer(struct obd_device *obd)
1329 {
1330         spin_lock_bh(&obd->obd_processing_task_lock);
1331         if (cfs_timer_is_armed(&obd->obd_recovery_timer)) {
1332                 spin_unlock_bh(&obd->obd_processing_task_lock);
1333                 return;
1334         }
1335         CWARN("%s: starting recovery timer\n", obd->obd_name);
1336         obd->obd_recovery_start = cfs_time_current_sec();
1337         /* minimum */
1338         obd->obd_recovery_timeout = OBD_RECOVERY_FACTOR * obd_timeout;
1339         spin_unlock_bh(&obd->obd_processing_task_lock);
1340
1341         reset_recovery_timer(obd, obd->obd_recovery_timeout, 0);
1342 }
1343
1344 /* Reset the timer with each new client connection */
1345 /*
1346  * This timer is actually reconnect_timer, which is for making sure
1347  * the total recovery window is at least as big as my reconnect
1348  * attempt timing. So the initial recovery time_out will be set to
1349  * OBD_RECOVERY_FACTOR * obd_timeout. If the timeout coming
1350  * from client is bigger than this, then the recovery time_out will
1351  * be extend to make sure the client could be reconnected, in the
1352  * process, the timeout from the new client should be ignored.
1353  */
1354
1355 static void
1356 target_start_and_reset_recovery_timer(struct obd_device *obd,
1357                                       struct ptlrpc_request *req,
1358                                       int new_client)
1359 {
1360         int req_timeout = lustre_msg_get_timeout(req->rq_reqmsg);
1361
1362         /* teach server about old server's estimates */
1363         if (!new_client)
1364                 at_add(&req->rq_rqbd->rqbd_service->srv_at_estimate,
1365                        at_timeout2est(req_timeout));
1366
1367         check_and_start_recovery_timer(obd);
1368
1369         req_timeout *= OBD_RECOVERY_FACTOR;
1370         if (req_timeout > obd->obd_recovery_timeout && !new_client)
1371                 reset_recovery_timer(obd, req_timeout, 0);
1372 }
1373
1374 #ifdef __KERNEL__
1375 static int check_for_next_transno(struct obd_device *obd)
1376 {
1377         struct ptlrpc_request *req = NULL;
1378         int wake_up = 0, connected, completed, queue_len, max;
1379         __u64 next_transno, req_transno;
1380         ENTRY;
1381         spin_lock_bh(&obd->obd_processing_task_lock);
1382
1383         if (!list_empty(&obd->obd_req_replay_queue)) {
1384                 req = list_entry(obd->obd_req_replay_queue.next,
1385                                  struct ptlrpc_request, rq_list);
1386                 req_transno = lustre_msg_get_transno(req->rq_reqmsg);
1387         } else {
1388                 req_transno = 0;
1389         }
1390
1391         max = obd->obd_max_recoverable_clients;
1392         connected = obd->obd_connected_clients;
1393         completed = connected - obd->obd_recoverable_clients;
1394         queue_len = obd->obd_requests_queued_for_recovery;
1395         next_transno = obd->obd_next_recovery_transno;
1396
1397         CDEBUG(D_HA, "max: %d, connected: %d, completed: %d, queue_len: %d, "
1398                "req_transno: "LPU64", next_transno: "LPU64"\n",
1399                max, connected, completed, queue_len, req_transno, next_transno);
1400
1401         if (obd->obd_abort_recovery) {
1402                 CDEBUG(D_HA, "waking for aborted recovery\n");
1403                 wake_up = 1;
1404         } else if (atomic_read(&obd->obd_req_replay_clients) == 0) {
1405                 CDEBUG(D_HA, "waking for completed recovery\n");
1406                 wake_up = 1;
1407         } else if (req_transno == next_transno) {
1408                 CDEBUG(D_HA, "waking for next ("LPD64")\n", next_transno);
1409                 wake_up = 1;
1410         } else if (queue_len + completed == max) {
1411                 /* handle gaps occured due to lost reply. It is allowed gaps
1412                  * because all clients are connected and there will be resend
1413                  * for missed transaction */
1414                 LASSERTF(req_transno >= next_transno,
1415                          "req_transno: "LPU64", next_transno: "LPU64"\n",
1416                          req_transno, next_transno);
1417
1418                 CDEBUG(req_transno > obd->obd_last_committed ? D_ERROR : D_HA,
1419                        "waking for skipped transno (skip: "LPD64
1420                        ", ql: %d, comp: %d, conn: %d, next: "LPD64")\n",
1421                        next_transno, queue_len, completed, connected, req_transno);
1422                 obd->obd_next_recovery_transno = req_transno;
1423                 wake_up = 1;
1424         } else if (OBD_FAIL_CHECK(OBD_FAIL_MDS_RECOVERY_ACCEPTS_GAPS)) {
1425                 CDEBUG(D_HA, "accepting transno gaps is explicitly allowed"
1426                        " by fail_lock, waking up ("LPD64")\n", next_transno);
1427                 obd->obd_next_recovery_transno = req_transno;
1428                 wake_up = 1;
1429         } else if (queue_len == atomic_read(&obd->obd_req_replay_clients)) {
1430                 /* some clients haven't connected in time, but we can try
1431                  * to replay requests that demand on already committed ones
1432                  * also, we can replay first non-committed transation */
1433                 LASSERT(req_transno != 0);
1434                 if (req_transno == obd->obd_last_committed + 1) {
1435                         obd->obd_next_recovery_transno = req_transno;
1436                 } else if (req_transno > obd->obd_last_committed) {
1437                         /* can't continue recovery: have no needed transno */
1438                         obd->obd_abort_recovery = 1;
1439                         CDEBUG(D_ERROR, "abort due to missed clients. max: %d, "
1440                                "connected: %d, completed: %d, queue_len: %d, "
1441                                "req_transno: "LPU64", next_transno: "LPU64"\n",
1442                                max, connected, completed, queue_len,
1443                                req_transno, next_transno);
1444                 }
1445                 wake_up = 1;
1446         }
1447
1448         spin_unlock_bh(&obd->obd_processing_task_lock);
1449         return wake_up;
1450 }
1451
1452 static struct ptlrpc_request *target_next_replay_req(struct obd_device *obd)
1453 {
1454         struct l_wait_info lwi = { 0 };
1455         struct ptlrpc_request *req;
1456
1457         CDEBUG(D_HA, "Waiting for transno "LPD64"\n",
1458                obd->obd_next_recovery_transno);
1459         l_wait_event(obd->obd_next_transno_waitq,
1460                      check_for_next_transno(obd), &lwi);
1461
1462         spin_lock_bh(&obd->obd_processing_task_lock);
1463         if (obd->obd_abort_recovery) {
1464                 req = NULL;
1465         } else if (!list_empty(&obd->obd_req_replay_queue)) {
1466                 req = list_entry(obd->obd_req_replay_queue.next,
1467                                  struct ptlrpc_request, rq_list);
1468                 target_exp_dequeue_req_replay(req);
1469                 list_del_init(&req->rq_list);
1470                 obd->obd_requests_queued_for_recovery--;
1471         } else {
1472                 req = NULL;
1473         }
1474         spin_unlock_bh(&obd->obd_processing_task_lock);
1475         RETURN(req);
1476 }
1477
1478 static int check_for_next_lock(struct obd_device *obd)
1479 {
1480         struct ptlrpc_request *req = NULL;
1481         int wake_up = 0;
1482
1483         spin_lock_bh(&obd->obd_processing_task_lock);
1484         if (!list_empty(&obd->obd_lock_replay_queue)) {
1485                 req = list_entry(obd->obd_lock_replay_queue.next,
1486                                  struct ptlrpc_request, rq_list);
1487                 CDEBUG(D_HA, "waking for next lock\n");
1488                 wake_up = 1;
1489         } else if (atomic_read(&obd->obd_lock_replay_clients) == 0) {
1490                 CDEBUG(D_HA, "waking for completed lock replay\n");
1491                 wake_up = 1;
1492         } else if (obd->obd_abort_recovery) {
1493                 CDEBUG(D_HA, "waking for aborted recovery\n");
1494                 wake_up = 1;
1495         }
1496         spin_unlock_bh(&obd->obd_processing_task_lock);
1497
1498         return wake_up;
1499 }
1500
1501 static struct ptlrpc_request *target_next_replay_lock(struct obd_device *obd)
1502 {
1503         struct l_wait_info lwi = { 0 };
1504         struct ptlrpc_request *req;
1505
1506         CDEBUG(D_HA, "Waiting for lock\n");
1507         l_wait_event(obd->obd_next_transno_waitq,
1508                      check_for_next_lock(obd), &lwi);
1509
1510         spin_lock_bh(&obd->obd_processing_task_lock);
1511         if (obd->obd_abort_recovery) {
1512                 req = NULL;
1513         } else if (!list_empty(&obd->obd_lock_replay_queue)) {
1514                 req = list_entry(obd->obd_lock_replay_queue.next,
1515                                  struct ptlrpc_request, rq_list);
1516                 list_del_init(&req->rq_list);
1517         } else {
1518                 req = NULL;
1519         }
1520         spin_unlock_bh(&obd->obd_processing_task_lock);
1521         return req;
1522 }
1523
1524 static struct ptlrpc_request *target_next_final_ping(struct obd_device *obd)
1525 {
1526         struct ptlrpc_request *req;
1527
1528         spin_lock_bh(&obd->obd_processing_task_lock);
1529         if (!list_empty(&obd->obd_final_req_queue)) {
1530                 req = list_entry(obd->obd_final_req_queue.next,
1531                                  struct ptlrpc_request, rq_list);
1532                 list_del_init(&req->rq_list);
1533         } else {
1534                 req = NULL;
1535         }
1536         spin_unlock_bh(&obd->obd_processing_task_lock);
1537         return req;
1538 }
1539
1540 static inline int req_replay_done(struct obd_export *exp)
1541 {
1542         return (exp->exp_req_replay_needed == 0);
1543 }
1544
1545 static inline int lock_replay_done(struct obd_export *exp)
1546 {
1547         return (exp->exp_lock_replay_needed == 0);
1548 }
1549
1550 static inline int connect_done(struct obd_export *exp)
1551 {
1552         return (exp->exp_in_recovery != 0);
1553 }
1554
1555 static int check_for_clients(struct obd_device *obd)
1556 {
1557         if (obd->obd_abort_recovery)
1558                 return 1;
1559         LASSERT(obd->obd_connected_clients <= obd->obd_max_recoverable_clients);
1560         if (obd->obd_no_conn == 0 &&
1561             obd->obd_connected_clients == obd->obd_max_recoverable_clients)
1562                 return 1;
1563         return 0;
1564 }
1565
1566 static int handle_recovery_req(struct ptlrpc_thread *thread,
1567                                struct ptlrpc_request *req,
1568                                svc_handler_t handler)
1569 {
1570         int rc;
1571         ENTRY;
1572
1573         rc = lu_context_init(&req->rq_session, LCT_SESSION);
1574         if (rc) {
1575                 CERROR("Failure to initialize session: %d\n", rc);
1576                 return rc;
1577         }
1578         req->rq_session.lc_thread = thread;
1579         lu_context_enter(&req->rq_session);
1580         req->rq_svc_thread = thread;
1581         req->rq_svc_thread->t_env->le_ses = &req->rq_session;
1582
1583         /* thread context */
1584         lu_context_enter(&thread->t_env->le_ctx);
1585         (void)handler(req);
1586         lu_context_exit(&thread->t_env->le_ctx);
1587
1588         lu_context_exit(&req->rq_session);
1589         lu_context_fini(&req->rq_session);
1590         /* don't reset timer for final stage */
1591         if (!req_replay_done(req->rq_export) ||
1592             !lock_replay_done(req->rq_export))
1593                 reset_recovery_timer(class_exp2obd(req->rq_export),
1594                        OBD_RECOVERY_FACTOR * AT_OFF ? obd_timeout :
1595                        at_get(&req->rq_rqbd->rqbd_service->srv_at_estimate), 1);
1596         ptlrpc_free_clone(req);
1597         RETURN(0);
1598 }
1599
1600 static void resume_recovery_timer(struct obd_device *obd)
1601 {
1602         /* to be safe, make it at least OBD_RECOVERY_FACTOR * obd_timeout */
1603         reset_recovery_timer(obd, OBD_RECOVERY_FACTOR * obd_timeout, 1);
1604 }
1605
1606 static int target_recovery_thread(void *arg)
1607 {
1608         struct obd_device *obd = arg;
1609         struct ptlrpc_request *req;
1610         struct target_recovery_data *trd = &obd->obd_recovery_data;
1611         struct l_wait_info lwi = { 0 };
1612         unsigned long delta;
1613         unsigned long flags;
1614         struct lu_env env;
1615         struct ptlrpc_thread fake_svc_thread, *thread = &fake_svc_thread;
1616         int rc = 0;
1617         ENTRY;
1618
1619         cfs_daemonize("tgt_recov");
1620
1621         SIGNAL_MASK_LOCK(current, flags);
1622         sigfillset(&current->blocked);
1623         RECALC_SIGPENDING;
1624         SIGNAL_MASK_UNLOCK(current, flags);
1625
1626         rc = lu_context_init(&env.le_ctx, LCT_MD_THREAD);
1627         if (rc)
1628                 RETURN(rc);
1629
1630         thread->t_env = &env;
1631         env.le_ctx.lc_thread = thread;
1632
1633         CERROR("%s: started recovery thread pid %d\n", obd->obd_name,
1634                current->pid);
1635         trd->trd_processing_task = current->pid;
1636
1637         obd->obd_recovering = 1;
1638         complete(&trd->trd_starting);
1639
1640         /* first of all, we have to know the first transno to replay */
1641         obd->obd_abort_recovery = 0;
1642         l_wait_event(obd->obd_next_transno_waitq,
1643                      check_for_clients(obd), &lwi);
1644
1645         spin_lock_bh(&obd->obd_processing_task_lock);
1646         target_cancel_recovery_timer(obd);
1647         spin_unlock_bh(&obd->obd_processing_task_lock);
1648
1649         /* If some clients haven't connected in time, evict them */
1650         if (obd->obd_abort_recovery) {
1651                 CWARN("Some clients haven't connect in time (%d/%d),"
1652                        "evict them\n", obd->obd_connected_clients,
1653                        obd->obd_max_recoverable_clients);
1654                 obd->obd_abort_recovery = obd->obd_stopping;
1655                 class_disconnect_stale_exports(obd, connect_done);
1656         }
1657         /* next stage: replay requests */
1658         delta = jiffies;
1659         obd->obd_req_replaying = 1;
1660         CDEBUG(D_INFO, "1: request replay stage - %d clients from t"LPU64"\n",
1661               atomic_read(&obd->obd_req_replay_clients),
1662               obd->obd_next_recovery_transno);
1663         resume_recovery_timer(obd);
1664         while ((req = target_next_replay_req(obd))) {
1665                 LASSERT(trd->trd_processing_task == current->pid);
1666                 DEBUG_REQ(D_HA, req, "processing t"LPD64" from %s",
1667                           lustre_msg_get_transno(req->rq_reqmsg),
1668                           libcfs_nid2str(req->rq_peer.nid));
1669                 handle_recovery_req(thread, req,
1670                                     trd->trd_recovery_handler);
1671                 obd->obd_replayed_requests++;
1672                 spin_lock_bh(&obd->obd_processing_task_lock);
1673                 obd->obd_next_recovery_transno++;
1674                 spin_unlock_bh(&obd->obd_processing_task_lock);
1675         }
1676
1677         spin_lock_bh(&obd->obd_processing_task_lock);
1678         target_cancel_recovery_timer(obd);
1679         spin_unlock_bh(&obd->obd_processing_task_lock);
1680
1681         /* If some clients haven't replayed requests in time, evict them */
1682         if (obd->obd_abort_recovery) {
1683                 CDEBUG(D_ERROR, "req replay timed out, aborting ...\n");
1684                 obd->obd_abort_recovery = obd->obd_stopping;
1685                 class_disconnect_stale_exports(obd, req_replay_done);
1686                 abort_req_replay_queue(obd);
1687         }
1688
1689         /* The second stage: replay locks */
1690         CDEBUG(D_INFO, "2: lock replay stage - %d clients\n",
1691                atomic_read(&obd->obd_lock_replay_clients));
1692         resume_recovery_timer(obd);
1693         while ((req = target_next_replay_lock(obd))) {
1694                 LASSERT(trd->trd_processing_task == current->pid);
1695                 DEBUG_REQ(D_HA|D_WARNING, req, "processing lock from %s: ",
1696                           libcfs_nid2str(req->rq_peer.nid));
1697                 handle_recovery_req(thread, req,
1698                                     trd->trd_recovery_handler);
1699                 obd->obd_replayed_locks++;
1700         }
1701
1702         spin_lock_bh(&obd->obd_processing_task_lock);
1703         target_cancel_recovery_timer(obd);
1704         spin_unlock_bh(&obd->obd_processing_task_lock);
1705         /* If some clients haven't replayed requests in time, evict them */
1706         if (obd->obd_abort_recovery) {
1707                 int stale;
1708                 CERROR("lock replay timed out, aborting ...\n");
1709                 obd->obd_abort_recovery = obd->obd_stopping;
1710                 stale = class_disconnect_stale_exports(obd, lock_replay_done);
1711                 abort_lock_replay_queue(obd);
1712         }
1713
1714         /* We drop recoverying flag to forward all new requests
1715          * to regular mds_handle() since now */
1716         spin_lock_bh(&obd->obd_processing_task_lock);
1717         obd->obd_recovering = obd->obd_abort_recovery = 0;
1718         spin_unlock_bh(&obd->obd_processing_task_lock);
1719         /* The third stage: reply on final pings */
1720         CDEBUG(D_INFO, "3: final stage - process recovery completion pings\n");
1721         while ((req = target_next_final_ping(obd))) {
1722                 LASSERT(trd->trd_processing_task == current->pid);
1723                 DEBUG_REQ(D_HA, req, "processing final ping from %s: ",
1724                           libcfs_nid2str(req->rq_peer.nid));
1725                 handle_recovery_req(thread, req,
1726                                     trd->trd_recovery_handler);
1727         }
1728
1729         delta = (jiffies - delta) / HZ;
1730         CDEBUG(D_INFO,"4: recovery completed in %lus - %d/%d reqs/locks\n",
1731               delta, obd->obd_replayed_requests, obd->obd_replayed_locks);
1732         LASSERT(atomic_read(&obd->obd_req_replay_clients) == 0);
1733         LASSERT(atomic_read(&obd->obd_lock_replay_clients) == 0);
1734         if (delta > obd_timeout * 2) {
1735                 CWARN("too long recovery - read logs\n");
1736                 libcfs_debug_dumplog();
1737         }
1738
1739         target_finish_recovery(obd);
1740
1741         lu_context_fini(&env.le_ctx);
1742         trd->trd_processing_task = 0;
1743         complete(&trd->trd_finishing);
1744         RETURN(rc);
1745 }
1746
1747 int target_start_recovery_thread(struct obd_device *obd, svc_handler_t handler)
1748 {
1749         int rc = 0;
1750         struct target_recovery_data *trd = &obd->obd_recovery_data;
1751
1752         memset(trd, 0, sizeof(*trd));
1753         init_completion(&trd->trd_starting);
1754         init_completion(&trd->trd_finishing);
1755         trd->trd_recovery_handler = handler;
1756
1757         if (kernel_thread(target_recovery_thread, obd, 0) > 0) {
1758                 wait_for_completion(&trd->trd_starting);
1759                 LASSERT(obd->obd_recovering != 0);
1760         } else
1761                 rc = -ECHILD;
1762
1763         return rc;
1764 }
1765
1766 void target_stop_recovery_thread(struct obd_device *obd)
1767 {
1768         spin_lock_bh(&obd->obd_processing_task_lock);
1769         if (obd->obd_recovery_data.trd_processing_task > 0) {
1770                 struct target_recovery_data *trd = &obd->obd_recovery_data;
1771                 CERROR("%s: Aborting recovery\n", obd->obd_name);
1772                 obd->obd_abort_recovery = 1;
1773                 wake_up(&obd->obd_next_transno_waitq);
1774                 spin_unlock_bh(&obd->obd_processing_task_lock);
1775                 wait_for_completion(&trd->trd_finishing);
1776         } else {
1777                 spin_unlock_bh(&obd->obd_processing_task_lock);
1778         }
1779 }
1780
1781 void target_recovery_fini(struct obd_device *obd)
1782 {
1783         class_disconnect_exports(obd);
1784         target_stop_recovery_thread(obd);
1785         target_cleanup_recovery(obd);
1786 }
1787 EXPORT_SYMBOL(target_recovery_fini);
1788
1789 static void target_recovery_expired(unsigned long castmeharder)
1790 {
1791         struct obd_device *obd = (struct obd_device *)castmeharder;
1792         LCONSOLE_WARN("%s: recovery timed out; %d clients never reconnected "
1793                       "after %lds (%d clients did)\n",
1794                       obd->obd_name, obd->obd_recoverable_clients,
1795                       cfs_time_current_sec()- obd->obd_recovery_start,
1796                       obd->obd_connected_clients);
1797         spin_lock_bh(&obd->obd_processing_task_lock);
1798         if (obd->obd_recovering)
1799                 obd->obd_abort_recovery = 1;
1800         cfs_waitq_signal(&obd->obd_next_transno_waitq);
1801         spin_unlock_bh(&obd->obd_processing_task_lock);
1802 }
1803
1804 void target_recovery_init(struct obd_device *obd, svc_handler_t handler)
1805 {
1806         if (obd->obd_max_recoverable_clients == 0)
1807                 return;
1808
1809         CWARN("RECOVERY: service %s, %d recoverable clients, "
1810               "last_transno "LPU64"\n", obd->obd_name,
1811               obd->obd_max_recoverable_clients, obd->obd_last_committed);
1812         obd->obd_next_recovery_transno = obd->obd_last_committed + 1;
1813         obd->obd_recovery_start = 0;
1814         obd->obd_recovery_end = 0;
1815         obd->obd_recovery_timeout = OBD_RECOVERY_FACTOR * obd_timeout;
1816         /* bz13079: this should be set to desired value for ost but not for mds */
1817         obd->obd_recovery_max_time = OBD_RECOVERY_MAX_TIME;
1818         cfs_timer_init(&obd->obd_recovery_timer, target_recovery_expired, obd);
1819         target_start_recovery_thread(obd, handler);
1820 }
1821 EXPORT_SYMBOL(target_recovery_init);
1822
1823 #endif
1824
1825 int target_process_req_flags(struct obd_device *obd, struct ptlrpc_request *req)
1826 {
1827         struct obd_export *exp = req->rq_export;
1828         LASSERT(exp != NULL);
1829         if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REQ_REPLAY_DONE) {
1830                 /* client declares he's ready to replay locks */
1831                 spin_lock_bh(&obd->obd_processing_task_lock);
1832                 if (exp->exp_req_replay_needed) {
1833                         LASSERT(atomic_read(&obd->obd_req_replay_clients) > 0);
1834                         spin_lock(&exp->exp_lock);
1835                         exp->exp_req_replay_needed = 0;
1836                         spin_unlock(&exp->exp_lock);
1837                         atomic_dec(&obd->obd_req_replay_clients);
1838                         LASSERT(obd->obd_recoverable_clients > 0);
1839                         obd->obd_recoverable_clients--;
1840                         if (atomic_read(&obd->obd_req_replay_clients) == 0)
1841                                 CDEBUG(D_HA, "all clients have replayed reqs\n");
1842                         wake_up(&obd->obd_next_transno_waitq);
1843                 }
1844                 spin_unlock_bh(&obd->obd_processing_task_lock);
1845         }
1846         if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_LOCK_REPLAY_DONE) {
1847                 /* client declares he's ready to complete recovery
1848                  * so, we put the request on th final queue */
1849                 spin_lock_bh(&obd->obd_processing_task_lock);
1850                 if (exp->exp_lock_replay_needed) {
1851                         LASSERT(atomic_read(&obd->obd_lock_replay_clients) > 0);
1852                         spin_lock(&exp->exp_lock);
1853                         exp->exp_lock_replay_needed = 0;
1854                         spin_unlock(&exp->exp_lock);
1855                         atomic_dec(&obd->obd_lock_replay_clients);
1856                         if (atomic_read(&obd->obd_lock_replay_clients) == 0)
1857                                 CDEBUG(D_HA, "all clients have replayed locks\n");
1858                         wake_up(&obd->obd_next_transno_waitq);
1859                 }
1860                 spin_unlock_bh(&obd->obd_processing_task_lock);
1861         }
1862
1863         return 0;
1864 }
1865
1866 int target_queue_recovery_request(struct ptlrpc_request *req,
1867                                   struct obd_device *obd)
1868 {
1869         struct list_head *tmp;
1870         int inserted = 0;
1871         __u64 transno = lustre_msg_get_transno(req->rq_reqmsg);
1872
1873         ENTRY;
1874
1875         if (obd->obd_recovery_data.trd_processing_task == cfs_curproc_pid()) {
1876                 /* Processing the queue right now, don't re-add. */
1877                 RETURN(1);
1878         }
1879
1880         target_process_req_flags(obd, req);
1881
1882         if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_LOCK_REPLAY_DONE) {
1883                 /* client declares he's ready to complete recovery
1884                  * so, we put the request on th final queue */
1885                 req = ptlrpc_clone_req(req);
1886                 if (req == NULL)
1887                         RETURN(-ENOMEM);
1888                 DEBUG_REQ(D_HA, req, "queue final req");
1889                 spin_lock_bh(&obd->obd_processing_task_lock);
1890                 if (obd->obd_recovering)
1891                         list_add_tail(&req->rq_list, &obd->obd_final_req_queue);
1892                 else {
1893                         spin_unlock_bh(&obd->obd_processing_task_lock);
1894                         ptlrpc_free_clone(req);
1895                         if (obd->obd_stopping) {
1896                                 RETURN(-ENOTCONN);
1897                         } else {
1898                                 RETURN(1);
1899                         }
1900                 }
1901                 spin_unlock_bh(&obd->obd_processing_task_lock);
1902                 RETURN(0);
1903         }
1904         if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REQ_REPLAY_DONE) {
1905                 /* client declares he's ready to replay locks */
1906                 req = ptlrpc_clone_req(req);
1907                 if (req == NULL)
1908                         RETURN(-ENOMEM);
1909                 DEBUG_REQ(D_HA, req, "queue lock replay req");
1910                 spin_lock_bh(&obd->obd_processing_task_lock);
1911                 LASSERT(obd->obd_recovering);
1912                 /* usually due to recovery abort */
1913                 if (!req->rq_export->exp_in_recovery) {
1914                         spin_unlock_bh(&obd->obd_processing_task_lock);
1915                         ptlrpc_free_clone(req);
1916                         RETURN(-ENOTCONN);
1917                 }
1918                 LASSERT(req->rq_export->exp_lock_replay_needed);
1919                 list_add_tail(&req->rq_list, &obd->obd_lock_replay_queue);
1920                 spin_unlock_bh(&obd->obd_processing_task_lock);
1921                 wake_up(&obd->obd_next_transno_waitq);
1922                 RETURN(0);
1923         }
1924
1925         /* CAVEAT EMPTOR: The incoming request message has been swabbed
1926          * (i.e. buflens etc are in my own byte order), but type-dependent
1927          * buffers (eg mds_body, ost_body etc) have NOT been swabbed. */
1928
1929         if (!transno) {
1930                 CFS_INIT_LIST_HEAD(&req->rq_list);
1931                 DEBUG_REQ(D_HA, req, "not queueing");
1932                 RETURN(1);
1933         }
1934
1935         spin_lock_bh(&obd->obd_processing_task_lock);
1936
1937         /* If we're processing the queue, we want don't want to queue this
1938          * message.
1939          *
1940          * Also, if this request has a transno less than the one we're waiting
1941          * for, we should process it now.  It could (and currently always will)
1942          * be an open request for a descriptor that was opened some time ago.
1943          *
1944          * Also, a resent, replayed request that has already been
1945          * handled will pass through here and be processed immediately.
1946          */
1947         CWARN("Next recovery transno: "LPU64", current: "LPU64", replaying: %i\n",
1948               obd->obd_next_recovery_transno, transno, obd->obd_req_replaying);
1949         if (transno < obd->obd_next_recovery_transno && obd->obd_req_replaying) {
1950                 /* Processing the queue right now, don't re-add. */
1951                 LASSERT(list_empty(&req->rq_list));
1952                 spin_unlock_bh(&obd->obd_processing_task_lock);
1953                 RETURN(1);
1954         }
1955         spin_unlock_bh(&obd->obd_processing_task_lock);
1956
1957         if (OBD_FAIL_CHECK(OBD_FAIL_TGT_REPLAY_DROP))
1958                 RETURN(0);
1959
1960         req = ptlrpc_clone_req(req);
1961         if (req == NULL)
1962                 RETURN(-ENOMEM);
1963
1964         spin_lock_bh(&obd->obd_processing_task_lock);
1965         LASSERT(obd->obd_recovering);
1966         if (!req->rq_export->exp_in_recovery) {
1967                 spin_unlock_bh(&obd->obd_processing_task_lock);
1968                 ptlrpc_free_clone(req);
1969                 RETURN(-ENOTCONN);
1970         }
1971         LASSERT(req->rq_export->exp_req_replay_needed);
1972
1973         if (target_exp_enqueue_req_replay(req)) {
1974                 spin_unlock_bh(&obd->obd_processing_task_lock);
1975                 DEBUG_REQ(D_ERROR, req, "dropping resent queued req");
1976                 ptlrpc_free_clone(req);
1977                 RETURN(0);
1978         }
1979
1980         /* XXX O(n^2) */
1981         list_for_each(tmp, &obd->obd_req_replay_queue) {
1982                 struct ptlrpc_request *reqiter =
1983                         list_entry(tmp, struct ptlrpc_request, rq_list);
1984
1985                 if (lustre_msg_get_transno(reqiter->rq_reqmsg) > transno) {
1986                         list_add_tail(&req->rq_list, &reqiter->rq_list);
1987                         inserted = 1;
1988                         break;
1989                 }
1990
1991                 if (unlikely(lustre_msg_get_transno(reqiter->rq_reqmsg) ==
1992                              transno)) {
1993                         DEBUG_REQ(D_ERROR, req, "dropping replay: transno "
1994                                   "has been claimed by another client");
1995                         spin_unlock_bh(&obd->obd_processing_task_lock);
1996                         target_exp_dequeue_req_replay(req);
1997                         ptlrpc_free_clone(req);
1998                         RETURN(0);
1999                 }
2000         }
2001
2002         if (!inserted)
2003                 list_add_tail(&req->rq_list, &obd->obd_req_replay_queue);
2004
2005         obd->obd_requests_queued_for_recovery++;
2006         wake_up(&obd->obd_next_transno_waitq);
2007         spin_unlock_bh(&obd->obd_processing_task_lock);
2008         RETURN(0);
2009 }
2010
2011 struct obd_device * target_req2obd(struct ptlrpc_request *req)
2012 {
2013         return req->rq_export->exp_obd;
2014 }
2015
2016 static inline struct ldlm_pool *ldlm_exp2pl(struct obd_export *exp)
2017 {
2018         LASSERT(exp != NULL);
2019         return &exp->exp_obd->obd_namespace->ns_pool;
2020 }
2021
2022 /**
2023  * Packs current SLV and Limit into \a req.
2024  */
2025 int target_pack_pool_reply(struct ptlrpc_request *req)
2026 {
2027         struct obd_device *obd;
2028         ENTRY;
2029
2030         /*
2031          * Check that we still have all structures alive as this may
2032          * be some late rpc in shutdown time.
2033          */
2034         if (unlikely(!req->rq_export || !req->rq_export->exp_obd ||
2035                      !exp_connect_lru_resize(req->rq_export))) {
2036                 lustre_msg_set_slv(req->rq_repmsg, 0);
2037                 lustre_msg_set_limit(req->rq_repmsg, 0);
2038                 RETURN(0);
2039         }
2040
2041         /*
2042          * OBD is alive here as export is alive, which we checked above.
2043          */
2044         obd = req->rq_export->exp_obd;
2045
2046         read_lock(&obd->obd_pool_lock);
2047         lustre_msg_set_slv(req->rq_repmsg, obd->obd_pool_slv);
2048         lustre_msg_set_limit(req->rq_repmsg, obd->obd_pool_limit);
2049         read_unlock(&obd->obd_pool_lock);
2050
2051         RETURN(0);
2052 }
2053
2054 int target_send_reply_msg(struct ptlrpc_request *req, int rc, int fail_id)
2055 {
2056         if (OBD_FAIL_CHECK_ORSET(fail_id & ~OBD_FAIL_ONCE, OBD_FAIL_ONCE)) {
2057                 DEBUG_REQ(D_ERROR, req, "dropping reply");
2058                 return (-ECOMM);
2059         }
2060
2061         if (unlikely(rc)) {
2062                 DEBUG_REQ(D_ERROR, req, "processing error (%d)", rc);
2063                 req->rq_status = rc;
2064                 return (ptlrpc_send_error(req, 1));
2065         } else {
2066                 DEBUG_REQ(D_NET, req, "sending reply");
2067         }
2068
2069         return (ptlrpc_send_reply(req, PTLRPC_REPLY_MAYBE_DIFFICULT));
2070 }
2071
2072 void target_send_reply(struct ptlrpc_request *req, int rc, int fail_id)
2073 {
2074         int                        netrc;
2075         struct ptlrpc_reply_state *rs;
2076         struct obd_device         *obd;
2077         struct obd_export         *exp;
2078         struct ptlrpc_service     *svc;
2079         ENTRY;
2080
2081         if (req->rq_no_reply) {
2082                 EXIT;
2083                 return;
2084         }
2085
2086         svc = req->rq_rqbd->rqbd_service;
2087         rs = req->rq_reply_state;
2088         if (rs == NULL || !rs->rs_difficult) {
2089                 /* no notifiers */
2090                 target_send_reply_msg (req, rc, fail_id);
2091                 EXIT;
2092                 return;
2093         }
2094
2095         /* must be an export if locks saved */
2096         LASSERT (req->rq_export != NULL);
2097         /* req/reply consistent */
2098         LASSERT (rs->rs_service == svc);
2099
2100         /* "fresh" reply */
2101         LASSERT (!rs->rs_scheduled);
2102         LASSERT (!rs->rs_scheduled_ever);
2103         LASSERT (!rs->rs_handled);
2104         LASSERT (!rs->rs_on_net);
2105         LASSERT (rs->rs_export == NULL);
2106         LASSERT (list_empty(&rs->rs_obd_list));
2107         LASSERT (list_empty(&rs->rs_exp_list));
2108
2109         exp = class_export_get (req->rq_export);
2110         obd = exp->exp_obd;
2111
2112         /* disable reply scheduling onto srv_reply_queue while I'm setting up */
2113         rs->rs_scheduled = 1;
2114         rs->rs_on_net    = 1;
2115         rs->rs_xid       = req->rq_xid;
2116         rs->rs_transno   = req->rq_transno;
2117         rs->rs_export    = exp;
2118
2119         spin_lock(&obd->obd_uncommitted_replies_lock);
2120
2121         CDEBUG(D_NET, "rs transno = "LPU64", last committed = "LPU64"\n",
2122                rs->rs_transno, obd->obd_last_committed);
2123         if (rs->rs_transno > obd->obd_last_committed) {
2124                 /* not committed already */
2125                 list_add_tail (&rs->rs_obd_list,
2126                                &obd->obd_uncommitted_replies);
2127         }
2128
2129         spin_unlock (&obd->obd_uncommitted_replies_lock);
2130         spin_lock (&exp->exp_lock);
2131
2132         list_add_tail (&rs->rs_exp_list, &exp->exp_outstanding_replies);
2133
2134         spin_unlock(&exp->exp_lock);
2135
2136         netrc = target_send_reply_msg (req, rc, fail_id);
2137
2138         spin_lock(&svc->srv_lock);
2139
2140         svc->srv_n_difficult_replies++;
2141
2142         if (netrc != 0) {
2143                 /* error sending: reply is off the net.  Also we need +1
2144                  * reply ref until ptlrpc_server_handle_reply() is done
2145                  * with the reply state (if the send was successful, there
2146                  * would have been +1 ref for the net, which
2147                  * reply_out_callback leaves alone) */
2148                 rs->rs_on_net = 0;
2149                 ptlrpc_rs_addref(rs);
2150                 atomic_inc (&svc->srv_outstanding_replies);
2151         }
2152
2153         if (rs->rs_transno <= obd->obd_last_committed ||
2154             (!rs->rs_on_net && !rs->rs_no_ack) ||
2155              list_empty(&rs->rs_exp_list) ||     /* completed already */
2156              list_empty(&rs->rs_obd_list)) {
2157                 CDEBUG(D_HA, "Schedule reply immediately\n");
2158                 list_add_tail (&rs->rs_list, &svc->srv_reply_queue);
2159                 cfs_waitq_signal (&svc->srv_waitq);
2160         } else {
2161                 list_add (&rs->rs_list, &svc->srv_active_replies);
2162                 rs->rs_scheduled = 0;           /* allow notifier to schedule */
2163         }
2164
2165         spin_unlock(&svc->srv_lock);
2166         EXIT;
2167 }
2168
2169 int target_handle_ping(struct ptlrpc_request *req)
2170 {
2171         obd_ping(req->rq_export);
2172         return req_capsule_server_pack(&req->rq_pill);
2173 }
2174
2175 void target_committed_to_req(struct ptlrpc_request *req)
2176 {
2177         struct obd_device *obd;
2178
2179         if (req == NULL || req->rq_export == NULL)
2180                 return;
2181
2182         obd = req->rq_export->exp_obd;
2183         if (obd == NULL)
2184                 return;
2185
2186         if (!obd->obd_no_transno && req->rq_repmsg != NULL)
2187                 lustre_msg_set_last_committed(req->rq_repmsg,
2188                                               obd->obd_last_committed);
2189         else
2190                 DEBUG_REQ(D_IOCTL, req, "not sending last_committed update (%d/"
2191                           "%d)", obd->obd_no_transno, req->rq_repmsg == NULL);
2192
2193         CDEBUG(D_INFO, "last_committed "LPU64", transno "LPU64", xid "LPU64"\n",
2194                obd->obd_last_committed, req->rq_transno, req->rq_xid);
2195 }
2196
2197 EXPORT_SYMBOL(target_committed_to_req);
2198
2199 int target_handle_qc_callback(struct ptlrpc_request *req)
2200 {
2201         struct obd_quotactl *oqctl;
2202         struct client_obd *cli = &req->rq_export->exp_obd->u.cli;
2203
2204         oqctl = req_capsule_client_get(&req->rq_pill, &RMF_OBD_QUOTACTL);
2205         if (oqctl == NULL) {
2206                 CERROR("Can't unpack obd_quotactl\n");
2207                 RETURN(-EPROTO);
2208         }
2209
2210         cli->cl_qchk_stat = oqctl->qc_stat;
2211
2212         return 0;
2213 }
2214
2215 #ifdef HAVE_QUOTA_SUPPORT
2216 int target_handle_dqacq_callback(struct ptlrpc_request *req)
2217 {
2218 #ifdef __KERNEL__
2219         struct obd_device *obd = req->rq_export->exp_obd;
2220         struct obd_device *master_obd;
2221         struct obd_device_target *obt;
2222         struct lustre_quota_ctxt *qctxt;
2223         struct qunit_data *qdata = NULL;
2224         int rc = 0;
2225         ENTRY;
2226
2227         if (OBD_FAIL_CHECK(OBD_FAIL_MDS_DROP_QUOTA_REQ))
2228                 RETURN(rc);
2229
2230         rc = req_capsule_server_pack(&req->rq_pill);
2231         if (rc) {
2232                 CERROR("packing reply failed!: rc = %d\n", rc);
2233                 RETURN(rc);
2234         }
2235
2236         LASSERT(req->rq_export);
2237
2238         OBD_ALLOC(qdata, sizeof(struct qunit_data));
2239         if (!qdata)
2240                 RETURN(-ENOMEM);
2241         rc = quota_get_qdata(req, qdata, QUOTA_REQUEST, QUOTA_EXPORT);
2242         if (rc < 0) {
2243                 CDEBUG(D_ERROR, "Can't unpack qunit_data(rc: %d)\n", rc);
2244                 GOTO(out, rc);
2245         }
2246
2247         /* we use the observer */
2248         if (!obd->obd_observer || !obd->obd_observer->obd_observer) {
2249                 CERROR("Can't find the observer, it is recovering\n");
2250                 req->rq_status = -EIO;
2251                 GOTO(send_reply, rc = -EIO);
2252         }
2253
2254         master_obd = obd->obd_observer->obd_observer;
2255         obt = &master_obd->u.obt;
2256         qctxt = &obt->obt_qctxt;
2257
2258         if (!qctxt->lqc_setup || !qctxt->lqc_valid) {
2259                 /* quota_type has not been processed yet, return EAGAIN
2260                  * until we know whether or not quotas are supposed to
2261                  * be enabled */
2262                 CDEBUG(D_QUOTA, "quota_type not processed yet, return "
2263                        "-EAGAIN\n");
2264                 req->rq_status = -EAGAIN;
2265                 rc = ptlrpc_reply(req);
2266                 GOTO(out, rc);
2267         }
2268
2269         down_read(&obt->obt_rwsem);
2270         if (qctxt->lqc_lqs_hash == NULL) {
2271                 up_read(&obt->obt_rwsem);
2272                 /* quota_type has not been processed yet, return EAGAIN
2273                  * until we know whether or not quotas are supposed to
2274                  * be enabled */
2275                 CDEBUG(D_QUOTA, "quota_ctxt is not ready yet, return "
2276                        "-EAGAIN\n");
2277                 req->rq_status = -EAGAIN;
2278                 rc = ptlrpc_reply(req);
2279                 GOTO(out, rc);
2280         }
2281
2282         LASSERT(qctxt->lqc_handler);
2283         rc = qctxt->lqc_handler(master_obd, qdata,
2284                                 lustre_msg_get_opc(req->rq_reqmsg));
2285         up_read(&obt->obt_rwsem);
2286         if (rc && rc != -EDQUOT)
2287                 CDEBUG(rc == -EBUSY  ? D_QUOTA : D_ERROR,
2288                        "dqacq failed! (rc:%d)\n", rc);
2289         req->rq_status = rc;
2290
2291         /* there are three forms of qunit(historic causes), so we need to
2292          * adjust the same form to different forms slaves needed */
2293         rc = quota_copy_qdata(req, qdata, QUOTA_REPLY, QUOTA_EXPORT);
2294         if (rc < 0) {
2295                 CDEBUG(D_ERROR, "Can't pack qunit_data(rc: %d)\n", rc);
2296                 GOTO(out, rc);
2297         }
2298
2299         /* Block the quota req. b=14840 */
2300         OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_BLOCK_QUOTA_REQ, obd_timeout);
2301 send_reply:
2302         rc = ptlrpc_reply(req);
2303 out:
2304         OBD_FREE(qdata, sizeof(struct qunit_data));
2305         RETURN(rc);
2306 #else
2307         return 0;
2308 #endif /* !__KERNEL__ */
2309 }
2310 #endif /* HAVE_QUOTA_SUPPORT */
2311
2312 ldlm_mode_t lck_compat_array[] = {
2313         [LCK_EX] LCK_COMPAT_EX,
2314         [LCK_PW] LCK_COMPAT_PW,
2315         [LCK_PR] LCK_COMPAT_PR,
2316         [LCK_CW] LCK_COMPAT_CW,
2317         [LCK_CR] LCK_COMPAT_CR,
2318         [LCK_NL] LCK_COMPAT_NL,
2319         [LCK_GROUP] LCK_COMPAT_GROUP,
2320         [LCK_COS] LCK_COMPAT_COS,
2321 };
2322
2323 /**
2324  * Rather arbitrary mapping from LDLM error codes to errno values. This should
2325  * not escape to the user level.
2326  */
2327 int ldlm_error2errno(ldlm_error_t error)
2328 {
2329         int result;
2330
2331         switch (error) {
2332         case ELDLM_OK:
2333                 result = 0;
2334                 break;
2335         case ELDLM_LOCK_CHANGED:
2336                 result = -ESTALE;
2337                 break;
2338         case ELDLM_LOCK_ABORTED:
2339                 result = -ENAVAIL;
2340                 break;
2341         case ELDLM_LOCK_REPLACED:
2342                 result = -ESRCH;
2343                 break;
2344         case ELDLM_NO_LOCK_DATA:
2345                 result = -ENOENT;
2346                 break;
2347         case ELDLM_NAMESPACE_EXISTS:
2348                 result = -EEXIST;
2349                 break;
2350         case ELDLM_BAD_NAMESPACE:
2351                 result = -EBADF;
2352                 break;
2353         default:
2354                 if (((int)error) < 0)  /* cast to signed type */
2355                         result = error; /* as ldlm_error_t can be unsigned */
2356                 else {
2357                         CERROR("Invalid DLM result code: %i\n", error);
2358                         result = -EPROTO;
2359                 }
2360         }
2361         return result;
2362 }
2363 EXPORT_SYMBOL(ldlm_error2errno);
2364
2365 /**
2366  * Dual to ldlm_error2errno(): maps errno values back to ldlm_error_t.
2367  */
2368 ldlm_error_t ldlm_errno2error(int err_no)
2369 {
2370         int error;
2371
2372         switch (err_no) {
2373         case 0:
2374                 error = ELDLM_OK;
2375                 break;
2376         case -ESTALE:
2377                 error = ELDLM_LOCK_CHANGED;
2378                 break;
2379         case -ENAVAIL:
2380                 error = ELDLM_LOCK_ABORTED;
2381                 break;
2382         case -ESRCH:
2383                 error = ELDLM_LOCK_REPLACED;
2384                 break;
2385         case -ENOENT:
2386                 error = ELDLM_NO_LOCK_DATA;
2387                 break;
2388         case -EEXIST:
2389                 error = ELDLM_NAMESPACE_EXISTS;
2390                 break;
2391         case -EBADF:
2392                 error = ELDLM_BAD_NAMESPACE;
2393                 break;
2394         default:
2395                 error = err_no;
2396         }
2397         return error;
2398 }
2399 EXPORT_SYMBOL(ldlm_errno2error);
2400