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 "ldlm_internal.h"
  52
  53 /* @priority: if non-zero, move the selected to the list head
  54  * @create: if zero, only search in existed connections
  55  */
  56 static int import_set_conn(struct obd_import *imp, struct obd_uuid *uuid,
  57                            int priority, int create)
  58 {
  59         struct ptlrpc_connection *ptlrpc_conn;
  60         struct obd_import_conn *imp_conn = NULL, *item;
  61         int rc = 0;
  62         ENTRY;
  63
  64         if (!create && !priority) {
  65                 CDEBUG(D_HA, "Nothing to do\n");
  66                 RETURN(-EINVAL);
  67         }
  68
  69         ptlrpc_conn = ptlrpc_uuid_to_connection(uuid);
  70         if (!ptlrpc_conn) {
  71                 CDEBUG(D_HA, "can't find connection %s\n", uuid->uuid);
  72                 RETURN (-ENOENT);
  73         }
  74
  75         if (create) {
  76                 OBD_ALLOC(imp_conn, sizeof(*imp_conn));
  77                 if (!imp_conn) {
  78                         GOTO(out_put, rc = -ENOMEM);
  79                 }
  80         }
  81
  82         spin_lock(&imp->imp_lock);
  83         list_for_each_entry(item, &imp->imp_conn_list, oic_item) {
  84                 if (obd_uuid_equals(uuid, &item->oic_uuid)) {
  85                         if (priority) {
  86                                 list_del(&item->oic_item);
  87                                 list_add(&item->oic_item, &imp->imp_conn_list);
  88                                 item->oic_last_attempt = 0;
  89                         }
  90                         CDEBUG(D_HA, "imp %p@%s: found existing conn %s%s\n",
  91                                imp, imp->imp_obd->obd_name, uuid->uuid,
  92                                (priority ? ", moved to head" : ""));
  93                         spin_unlock(&imp->imp_lock);
  94                         GOTO(out_free, rc = 0);
  95                 }
  96         }
  97         /* not found */
  98         if (create) {
  99                 imp_conn->oic_conn = ptlrpc_conn;
 100                 imp_conn->oic_uuid = *uuid;
 101                 item->oic_last_attempt = 0;
 102                 if (priority)
 103                         list_add(&imp_conn->oic_item, &imp->imp_conn_list);
 104                 else
 105                         list_add_tail(&imp_conn->oic_item, &imp->imp_conn_list);
 106                 CDEBUG(D_HA, "imp %p@%s: add connection %s at %s\n",
 107                        imp, imp->imp_obd->obd_name, uuid->uuid,
 108                        (priority ? "head" : "tail"));
 109         } else {
 110                 spin_unlock(&imp->imp_lock);
 111                 GOTO(out_free, rc = -ENOENT);
 112
 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 /* configure an RPC client OBD device
 192  *
 193  * lcfg parameters:
 194  * 1 - client UUID
 195  * 2 - server UUID
 196  * 3 - inactive-on-startup
 197  */
 198 int client_obd_setup(struct obd_device *obddev, obd_count len, void *buf)
 199 {
 200         struct lustre_cfg* lcfg = buf;
 201         struct client_obd *cli = &obddev->u.cli;
 202         struct obd_import *imp;
 203         struct obd_uuid server_uuid;
 204         int rq_portal, rp_portal, connect_op;
 205         char *name = obddev->obd_type->typ_name;
 206         int rc;
 207         ENTRY;
 208
 209         /* In a more perfect world, we would hang a ptlrpc_client off of
 210          * obd_type and just use the values from there. */
 211         if (!strcmp(name, LUSTRE_OSC_NAME)) {
 212 #ifdef __KERNEL__
 213                 /* Can be removed in Lustre 1.8, for compatibility only */
 214                 rq_portal = OST_IO_PORTAL;
 215 #else
 216                 rq_portal = OST_REQUEST_PORTAL;
 217 #endif
 218                 rp_portal = OSC_REPLY_PORTAL;
 219                 connect_op = OST_CONNECT;
 220         } else if (!strcmp(name, LUSTRE_MDC_NAME)) {
 221                 rq_portal = MDS_REQUEST_PORTAL;
 222                 rp_portal = MDC_REPLY_PORTAL;
 223                 connect_op = MDS_CONNECT;
 224         } else if (!strcmp(name, LUSTRE_MGC_NAME)) {
 225                 rq_portal = MGS_REQUEST_PORTAL;
 226                 rp_portal = MGC_REPLY_PORTAL;
 227                 connect_op = MGS_CONNECT;
 228         } else {
 229                 CERROR("unknown client OBD type \"%s\", can't setup\n",
 230                        name);
 231                 RETURN(-EINVAL);
 232         }
 233
 234         if (LUSTRE_CFG_BUFLEN(lcfg, 1) < 1) {
 235                 CERROR("requires a TARGET UUID\n");
 236                 RETURN(-EINVAL);
 237         }
 238
 239         if (LUSTRE_CFG_BUFLEN(lcfg, 1) > 37) {
 240                 CERROR("client UUID must be less than 38 characters\n");
 241                 RETURN(-EINVAL);
 242         }
 243
 244         if (LUSTRE_CFG_BUFLEN(lcfg, 2) < 1) {
 245                 CERROR("setup requires a SERVER UUID\n");
 246                 RETURN(-EINVAL);
 247         }
 248
 249         if (LUSTRE_CFG_BUFLEN(lcfg, 2) > 37) {
 250                 CERROR("target UUID must be less than 38 characters\n");
 251                 RETURN(-EINVAL);
 252         }
 253
 254         init_rwsem(&cli->cl_sem);
 255         sema_init(&cli->cl_mgc_sem, 1);
 256         cli->cl_conn_count = 0;
 257         memcpy(server_uuid.uuid, lustre_cfg_buf(lcfg, 2),
 258                min_t(unsigned int, LUSTRE_CFG_BUFLEN(lcfg, 2),
 259                      sizeof(server_uuid)));
 260
 261         cli->cl_dirty = 0;
 262         cli->cl_avail_grant = 0;
 263         /* FIXME: should limit this for the sum of all cl_dirty_max */
 264         cli->cl_dirty_max = OSC_MAX_DIRTY_DEFAULT * 1024 * 1024;
 265         if (cli->cl_dirty_max >> CFS_PAGE_SHIFT > num_physpages / 8)
 266                 cli->cl_dirty_max = num_physpages << (CFS_PAGE_SHIFT - 3);
 267         CFS_INIT_LIST_HEAD(&cli->cl_cache_waiters);
 268         CFS_INIT_LIST_HEAD(&cli->cl_loi_ready_list);
 269         CFS_INIT_LIST_HEAD(&cli->cl_loi_hp_ready_list);
 270         CFS_INIT_LIST_HEAD(&cli->cl_loi_write_list);
 271         CFS_INIT_LIST_HEAD(&cli->cl_loi_read_list);
 272         client_obd_list_lock_init(&cli->cl_loi_list_lock);
 273         cli->cl_r_in_flight = 0;
 274         cli->cl_w_in_flight = 0;
 275         spin_lock_init(&cli->cl_read_rpc_hist.oh_lock);
 276         spin_lock_init(&cli->cl_write_rpc_hist.oh_lock);
 277         spin_lock_init(&cli->cl_read_page_hist.oh_lock);
 278         spin_lock_init(&cli->cl_write_page_hist.oh_lock);
 279         spin_lock_init(&cli->cl_read_offset_hist.oh_lock);
 280         spin_lock_init(&cli->cl_write_offset_hist.oh_lock);
 281         cfs_waitq_init(&cli->cl_destroy_waitq);
 282         atomic_set(&cli->cl_destroy_in_flight, 0);
 283 #ifdef ENABLE_CHECKSUM
 284         /* Turn on checksumming by default. */
 285         cli->cl_checksum = 1;
 286         /*
 287          * The supported checksum types will be worked out at connect time
 288          * Set cl_chksum* to CRC32 for now to avoid returning screwed info
 289          * through procfs.
 290          */
 291         cli->cl_cksum_type = cli->cl_supp_cksum_types = OBD_CKSUM_CRC32;
 292 #endif
 293         atomic_set(&cli->cl_resends, OSC_DEFAULT_RESENDS);
 294
 295         /* This value may be changed at connect time in
 296            ptlrpc_connect_interpret. */
 297         cli->cl_max_pages_per_rpc = min((int)PTLRPC_MAX_BRW_PAGES,
 298                                         (int)(1024 * 1024 >> CFS_PAGE_SHIFT));
 299
 300         if (!strcmp(name, LUSTRE_MDC_NAME)) {
 301                 cli->cl_max_rpcs_in_flight = MDC_MAX_RIF_DEFAULT;
 302         } else if (num_physpages >> (20 - CFS_PAGE_SHIFT) <= 128 /* MB */) {
 303                 cli->cl_max_rpcs_in_flight = 2;
 304         } else if (num_physpages >> (20 - CFS_PAGE_SHIFT) <= 256 /* MB */) {
 305                 cli->cl_max_rpcs_in_flight = 3;
 306         } else if (num_physpages >> (20 - CFS_PAGE_SHIFT) <= 512 /* MB */) {
 307                 cli->cl_max_rpcs_in_flight = 4;
 308         } else {
 309                 cli->cl_max_rpcs_in_flight = OSC_MAX_RIF_DEFAULT;
 310         }
 311         rc = ldlm_get_ref();
 312         if (rc) {
 313                 CERROR("ldlm_get_ref failed: %d\n", rc);
 314                 GOTO(err, rc);
 315         }
 316
 317         ptlrpc_init_client(rq_portal, rp_portal, name,
 318                            &obddev->obd_ldlm_client);
 319
 320         imp = class_new_import(obddev);
 321         if (imp == NULL)
 322                 GOTO(err_ldlm, rc = -ENOENT);
 323         imp->imp_client = &obddev->obd_ldlm_client;
 324         imp->imp_connect_op = connect_op;
 325         imp->imp_initial_recov = 1;
 326         imp->imp_initial_recov_bk = 0;
 327         CFS_INIT_LIST_HEAD(&imp->imp_pinger_chain);
 328         memcpy(cli->cl_target_uuid.uuid, lustre_cfg_buf(lcfg, 1),
 329                LUSTRE_CFG_BUFLEN(lcfg, 1));
 330         class_import_put(imp);
 331
 332         rc = client_import_add_conn(imp, &server_uuid, 1);
 333         if (rc) {
 334                 CERROR("can't add initial connection\n");
 335                 GOTO(err_import, rc);
 336         }
 337
 338         cli->cl_import = imp;
 339         /* cli->cl_max_mds_{easize,cookiesize} updated by mdc_init_ea_size() */
 340         cli->cl_max_mds_easize = sizeof(struct lov_mds_md_v3);
 341         cli->cl_max_mds_cookiesize = sizeof(struct llog_cookie);
 342
 343         if (LUSTRE_CFG_BUFLEN(lcfg, 3) > 0) {
 344                 if (!strcmp(lustre_cfg_string(lcfg, 3), "inactive")) {
 345                         CDEBUG(D_HA, "marking %s %s->%s as inactive\n",
 346                                name, obddev->obd_name,
 347                                cli->cl_target_uuid.uuid);
 348                         spin_lock(&imp->imp_lock);
 349                         imp->imp_invalid = 1;
 350                         spin_unlock(&imp->imp_lock);
 351                 }
 352         }
 353
 354         cli->cl_qchk_stat = CL_NOT_QUOTACHECKED;
 355
 356         RETURN(rc);
 357
 358 err_import:
 359         class_destroy_import(imp);
 360 err_ldlm:
 361         ldlm_put_ref();
 362 err:
 363         RETURN(rc);
 364
 365 }
 366
 367 int client_obd_cleanup(struct obd_device *obddev)
 368 {
 369         ENTRY;
 370         ldlm_put_ref();
 371         RETURN(0);
 372 }
 373
 374 /* ->o_connect() method for client side (OSC and MDC and MGC) */
 375 int client_connect_import(struct lustre_handle *dlm_handle,
 376                           struct obd_device *obd, struct obd_uuid *cluuid,
 377                           struct obd_connect_data *data, void *localdata)
 378 {
 379         struct client_obd *cli = &obd->u.cli;
 380         struct obd_import *imp = cli->cl_import;
 381         struct obd_export *exp;
 382         struct obd_connect_data *ocd;
 383         struct ldlm_namespace *to_be_freed = NULL;
 384         int rc;
 385         ENTRY;
 386
 387         down_write(&cli->cl_sem);
 388         rc = class_connect(dlm_handle, obd, cluuid);
 389         if (rc)
 390                 GOTO(out_sem, rc);
 391
 392         cli->cl_conn_count++;
 393         if (cli->cl_conn_count > 1)
 394                 GOTO(out_sem, rc);
 395         exp = class_conn2export(dlm_handle);
 396
 397         if (obd->obd_namespace != NULL)
 398                 CERROR("already have namespace!\n");
 399         obd->obd_namespace = ldlm_namespace_new(obd, obd->obd_name,
 400                                                 LDLM_NAMESPACE_CLIENT,
 401                                                 LDLM_NAMESPACE_GREEDY);
 402         if (obd->obd_namespace == NULL)
 403                 GOTO(out_disco, rc = -ENOMEM);
 404
 405         imp->imp_dlm_handle = *dlm_handle;
 406         rc = ptlrpc_init_import(imp);
 407         if (rc != 0)
 408                 GOTO(out_ldlm, rc);
 409
 410         ocd = &imp->imp_connect_data;
 411         if (data) {
 412                 *ocd = *data;
 413                 imp->imp_connect_flags_orig = data->ocd_connect_flags;
 414         }
 415
 416         rc = ptlrpc_connect_import(imp, NULL);
 417         if (rc != 0) {
 418                 LASSERT (imp->imp_state == LUSTRE_IMP_DISCON);
 419                 GOTO(out_ldlm, rc);
 420         }
 421         LASSERT(exp->exp_connection);
 422
 423         if (data) {
 424                 LASSERT((ocd->ocd_connect_flags & data->ocd_connect_flags) ==
 425                         ocd->ocd_connect_flags);
 426                 data->ocd_connect_flags = ocd->ocd_connect_flags;
 427         }
 428
 429         ptlrpc_pinger_add_import(imp);
 430         EXIT;
 431
 432         if (rc) {
 433 out_ldlm:
 434                 ldlm_namespace_free_prior(obd->obd_namespace, imp, 0);
 435                 to_be_freed = obd->obd_namespace;
 436                 obd->obd_namespace = NULL;
 437 out_disco:
 438                 cli->cl_conn_count--;
 439                 class_disconnect(exp);
 440         } else {
 441                 class_export_put(exp);
 442         }
 443 out_sem:
 444         up_write(&cli->cl_sem);
 445         if (to_be_freed)
 446                 ldlm_namespace_free_post(to_be_freed);
 447         return rc;
 448 }
 449
 450 int client_disconnect_export(struct obd_export *exp)
 451 {
 452         struct obd_device *obd = class_exp2obd(exp);
 453         struct client_obd *cli;
 454         struct obd_import *imp;
 455         struct ldlm_namespace *to_be_freed = NULL;
 456         int rc = 0, err;
 457         ENTRY;
 458
 459         if (!obd) {
 460                 CERROR("invalid export for disconnect: exp %p cookie "LPX64"\n",
 461                        exp, exp ? exp->exp_handle.h_cookie : -1);
 462                 RETURN(-EINVAL);
 463         }
 464
 465         cli = &obd->u.cli;
 466         imp = cli->cl_import;
 467
 468         down_write(&cli->cl_sem);
 469         if (!cli->cl_conn_count) {
 470                 CERROR("disconnecting disconnected device (%s)\n",
 471                        obd->obd_name);
 472                 GOTO(out_sem, rc = -EINVAL);
 473         }
 474
 475         cli->cl_conn_count--;
 476         if (cli->cl_conn_count)
 477                 GOTO(out_no_disconnect, rc = 0);
 478
 479         /* Mark import deactivated now, so we don't try to reconnect if any
 480          * of the cleanup RPCs fails (e.g. ldlm cancel, etc).  We don't
 481          * fully deactivate the import, or that would drop all requests. */
 482         spin_lock(&imp->imp_lock);
 483         imp->imp_deactive = 1;
 484         spin_unlock(&imp->imp_lock);
 485
 486         /* Some non-replayable imports (MDS's OSCs) are pinged, so just
 487          * delete it regardless.  (It's safe to delete an import that was
 488          * never added.) */
 489         (void)ptlrpc_pinger_del_import(imp);
 490
 491         if (obd->obd_namespace != NULL) {
 492                 /* obd_force == local only */
 493                 ldlm_cli_cancel_unused(obd->obd_namespace, NULL,
 494                                        obd->obd_force ? LDLM_FL_LOCAL_ONLY:0,
 495                                        NULL);
 496                 ldlm_namespace_free_prior(obd->obd_namespace, imp,
 497                                           obd->obd_force);
 498                 to_be_freed = obd->obd_namespace;
 499         }
 500
 501         rc = ptlrpc_disconnect_import(imp, 0);
 502
 503         ptlrpc_invalidate_import(imp);
 504         /* set obd_namespace to NULL only after invalidate, because we can have
 505          * some connect requests in flight, and his need store a connect flags
 506          * in obd_namespace. bug 14260 */
 507         obd->obd_namespace = NULL;
 508
 509         if (imp->imp_rq_pool) {
 510                 ptlrpc_free_rq_pool(imp->imp_rq_pool);
 511                 imp->imp_rq_pool = NULL;
 512         }
 513         class_destroy_import(imp);
 514         cli->cl_import = NULL;
 515
 516         EXIT;
 517  out_no_disconnect:
 518         err = class_disconnect(exp);
 519         if (!rc && err)
 520                 rc = err;
 521  out_sem:
 522         up_write(&cli->cl_sem);
 523         if (to_be_freed)
 524                 ldlm_namespace_free_post(to_be_freed);
 525         RETURN(rc);
 526 }
 527
 528 /* --------------------------------------------------------------------------
 529  * from old lib/target.c
 530  * -------------------------------------------------------------------------- */
 531
 532 int target_handle_reconnect(struct lustre_handle *conn, struct obd_export *exp,
 533                             struct obd_uuid *cluuid)
 534 {
 535         ENTRY;
 536         if (exp->exp_connection && exp->exp_imp_reverse) {
 537                 struct lustre_handle *hdl;
 538                 hdl = &exp->exp_imp_reverse->imp_remote_handle;
 539                 /* Might be a re-connect after a partition. */
 540                 if (!memcmp(&conn->cookie, &hdl->cookie, sizeof conn->cookie)) {
 541                         CWARN("%s: %s reconnecting\n", exp->exp_obd->obd_name,
 542                               cluuid->uuid);
 543                         conn->cookie = exp->exp_handle.h_cookie;
 544                         /* target_handle_connect() treats EALREADY and
 545                          * -EALREADY differently.  EALREADY means we are
 546                          * doing a valid reconnect from the same client. */
 547                         RETURN(EALREADY);
 548                 } else {
 549                         CERROR("%s reconnecting from %s, "
 550                                "handle mismatch (ours "LPX64", theirs "
 551                                LPX64")\n", cluuid->uuid,
 552                                exp->exp_connection->c_remote_uuid.uuid,
 553                                hdl->cookie, conn->cookie);
 554                         memset(conn, 0, sizeof *conn);
 555                         /* target_handle_connect() treats EALREADY and
 556                          * -EALREADY differently.  -EALREADY is an error
 557                          * (same UUID, different handle). */
 558                         RETURN(-EALREADY);
 559                 }
 560         }
 561
 562         conn->cookie = exp->exp_handle.h_cookie;
 563         CDEBUG(D_HA, "connect export for UUID '%s' at %p, cookie "LPX64"\n",
 564                cluuid->uuid, exp, conn->cookie);
 565         RETURN(0);
 566 }
 567
 568 void target_client_add_cb(struct obd_device *obd, __u64 transno, void *cb_data,
 569                           int error)
 570 {
 571         struct obd_export *exp = cb_data;
 572
 573         CDEBUG(D_RPCTRACE, "%s: committing for initial connect of %s\n",
 574                obd->obd_name, exp->exp_client_uuid.uuid);
 575
 576         spin_lock(&exp->exp_lock);
 577         exp->exp_need_sync = 0;
 578         spin_unlock(&exp->exp_lock);
 579 }
 580 EXPORT_SYMBOL(target_client_add_cb);
 581
 582 static void
 583 target_start_and_reset_recovery_timer(struct obd_device *obd,
 584                                       svc_handler_t handler,
 585                                       struct ptlrpc_request *req,
 586                                       int new_client);
 587 void target_stop_recovery(void *, int);
 588 int target_recovery_check_and_stop(struct obd_device *obd)
 589 {
 590         int abort_recovery = 0;
 591
 592         spin_lock_bh(&obd->obd_processing_task_lock);
 593         abort_recovery = obd->obd_abort_recovery;
 594         spin_unlock_bh(&obd->obd_processing_task_lock);
 595         if (abort_recovery) {
 596                 target_stop_recovery(obd, 0);
 597                 return 1;
 598         }
 599         return 0;
 600 }
 601 EXPORT_SYMBOL(target_recovery_check_and_stop);
 602
 603 int target_handle_connect(struct ptlrpc_request *req, svc_handler_t handler)
 604 {
 605         struct obd_device *target, *targref = NULL;
 606         struct obd_export *export = NULL;
 607         struct obd_import *revimp;
 608         struct lustre_handle conn;
 609         struct obd_uuid tgtuuid;
 610         struct obd_uuid cluuid;
 611         struct obd_uuid remote_uuid;
 612         char *str, *tmp;
 613         int rc = 0;
 614         struct obd_connect_data *data;
 615         __u32 size[2] = { sizeof(struct ptlrpc_body), sizeof(*data) };
 616         lnet_nid_t *client_nid = NULL;
 617         ENTRY;
 618
 619         OBD_RACE(OBD_FAIL_TGT_CONN_RACE);
 620
 621         lustre_set_req_swabbed(req, REQ_REC_OFF);
 622         str = lustre_msg_string(req->rq_reqmsg, REQ_REC_OFF, sizeof(tgtuuid)-1);
 623         if (str == NULL) {
 624                 DEBUG_REQ(D_ERROR, req, "bad target UUID for connect");
 625                 GOTO(out, rc = -EINVAL);
 626         }
 627
 628         obd_str2uuid (&tgtuuid, str);
 629         target = class_uuid2obd(&tgtuuid);
 630         /* COMPAT_146 */
 631         /* old (pre 1.6) lustre_process_log tries to connect to mdsname
 632            (eg. mdsA) instead of uuid. */
 633         if (!target) {
 634                 snprintf((char *)tgtuuid.uuid, sizeof(tgtuuid), "%s_UUID", str);
 635                 target = class_uuid2obd(&tgtuuid);
 636         }
 637         if (!target)
 638                 target = class_name2obd(str);
 639         /* end COMPAT_146 */
 640
 641         if (!target || target->obd_stopping || !target->obd_set_up) {
 642                 LCONSOLE_ERROR_MSG(0x137, "UUID '%s' is not available "
 643                                    " for connect (%s)\n", str,
 644                                    !target ? "no target" :
 645                                    (target->obd_stopping ? "stopping" :
 646                                    "not set up"));
 647                 GOTO(out, rc = -ENODEV);
 648         }
 649
 650         if (target->obd_no_conn) {
 651                 LCONSOLE_WARN("%s: temporarily refusing client connection "
 652                               "from %s\n", target->obd_name,
 653                               libcfs_nid2str(req->rq_peer.nid));
 654                 GOTO(out, rc = -EAGAIN);
 655         }
 656
 657         /* Make sure the target isn't cleaned up while we're here. Yes,
 658            there's still a race between the above check and our incref here.
 659            Really, class_uuid2obd should take the ref. */
 660         targref = class_incref(target);
 661
 662         lustre_set_req_swabbed(req, REQ_REC_OFF + 1);
 663         str = lustre_msg_string(req->rq_reqmsg, REQ_REC_OFF + 1,
 664                                 sizeof(cluuid) - 1);
 665         if (str == NULL) {
 666                 DEBUG_REQ(D_ERROR, req, "bad client UUID for connect");
 667                 GOTO(out, rc = -EINVAL);
 668         }
 669
 670         obd_str2uuid (&cluuid, str);
 671
 672         /* XXX extract a nettype and format accordingly */
 673         switch (sizeof(lnet_nid_t)) {
 674                 /* NB the casts only avoid compiler warnings */
 675         case 8:
 676                 snprintf(remote_uuid.uuid, sizeof remote_uuid,
 677                          "NET_"LPX64"_UUID", (__u64)req->rq_peer.nid);
 678                 break;
 679         case 4:
 680                 snprintf(remote_uuid.uuid, sizeof remote_uuid,
 681                          "NET_%x_UUID", (__u32)req->rq_peer.nid);
 682                 break;
 683         default:
 684                 LBUG();
 685         }
 686
 687         target_recovery_check_and_stop(target);
 688
 689         tmp = lustre_msg_buf(req->rq_reqmsg, REQ_REC_OFF + 2, sizeof conn);
 690         if (tmp == NULL)
 691                 GOTO(out, rc = -EPROTO);
 692
 693         memcpy(&conn, tmp, sizeof conn);
 694
 695         data = lustre_swab_reqbuf(req, REQ_REC_OFF + 3, sizeof(*data),
 696                                   lustre_swab_connect);
 697         rc = lustre_pack_reply(req, 2, size, NULL);
 698         if (rc)
 699                 GOTO(out, rc);
 700
 701         if (lustre_msg_get_op_flags(req->rq_reqmsg) & MSG_CONNECT_LIBCLIENT) {
 702                 if (!data) {
 703                         DEBUG_REQ(D_WARNING, req, "Refusing old (unversioned) "
 704                                   "libclient connection attempt");
 705                         GOTO(out, rc = -EPROTO);
 706                 } else if (data->ocd_version < LUSTRE_VERSION_CODE -
 707                                                LUSTRE_VERSION_ALLOWED_OFFSET ||
 708                            data->ocd_version > LUSTRE_VERSION_CODE +
 709                                                LUSTRE_VERSION_ALLOWED_OFFSET) {
 710                         DEBUG_REQ(D_WARNING, req, "Refusing %s (%d.%d.%d.%d) "
 711                                   "libclient connection attempt",
 712                                   data->ocd_version < LUSTRE_VERSION_CODE ?
 713                                   "old" : "new",
 714                                   OBD_OCD_VERSION_MAJOR(data->ocd_version),
 715                                   OBD_OCD_VERSION_MINOR(data->ocd_version),
 716                                   OBD_OCD_VERSION_PATCH(data->ocd_version),
 717                                   OBD_OCD_VERSION_FIX(data->ocd_version));
 718                         data = lustre_msg_buf(req->rq_repmsg, REPLY_REC_OFF,
 719                                               offsetof(typeof(*data),
 720                                                        ocd_version) +
 721                                               sizeof(data->ocd_version));
 722                         if (data) {
 723                                 data->ocd_connect_flags = OBD_CONNECT_VERSION;
 724                                 data->ocd_version = LUSTRE_VERSION_CODE;
 725                         }
 726                         GOTO(out, rc = -EPROTO);
 727                 }
 728         }
 729
 730         /* lctl gets a backstage, all-access pass. */
 731         if (obd_uuid_equals(&cluuid, &target->obd_uuid))
 732                 goto dont_check_exports;
 733
 734         spin_lock(&target->obd_dev_lock);
 735         export = lustre_hash_lookup(target->obd_uuid_hash, &cluuid);
 736
 737         if (export != NULL && export->exp_connecting) { /* bug 9635, et. al. */
 738                 CWARN("%s: exp %p already connecting\n",
 739                       export->exp_obd->obd_name, export);
 740                 class_export_put(export);
 741                 export = NULL;
 742                 rc = -EALREADY;
 743         } else if (export != NULL && export->exp_connection != NULL &&
 744                    req->rq_peer.nid != export->exp_connection->c_peer.nid) {
 745                 /* make darn sure this is coming from the same peer
 746                  * if the UUIDs matched */
 747                 if (data && data->ocd_connect_flags & OBD_CONNECT_MDS) {
 748                         /* the MDS UUID can be reused, don't need to wait
 749                          * for the export to be evicted */
 750                         CWARN("%s: received MDS connection from a new NID %s,"
 751                               " removing former export from NID %s\n",
 752                             target->obd_name,
 753                             libcfs_nid2str(req->rq_peer.nid),
 754                             libcfs_nid2str(export->exp_connection->c_peer.nid));
 755                         class_fail_export(export);
 756                 } else {
 757                         CWARN("%s: cookie %s seen on new NID %s when "
 758                               "existing NID %s is already connected\n",
 759                             target->obd_name, cluuid.uuid,
 760                             libcfs_nid2str(req->rq_peer.nid),
 761                             libcfs_nid2str(export->exp_connection->c_peer.nid));
 762                         rc = -EALREADY;
 763                 }
 764                 class_export_put(export);
 765                 export = NULL;
 766         } else if (export != NULL && export->exp_failed) { /* bug 11327 */
 767                 CDEBUG(D_HA, "%s: exp %p evict in progress - new cookie needed "
 768                       "for connect\n", export->exp_obd->obd_name, export);
 769                 class_export_put(export);
 770                 export = NULL;
 771                 rc = -ENODEV;
 772         } else if (export != NULL && export->exp_delayed &&
 773                    !(data && data->ocd_connect_flags & OBD_CONNECT_VBR)) {
 774                 spin_unlock(&target->obd_dev_lock);
 775                 class_fail_export(export);
 776                 class_export_put(export);
 777                 export = NULL;
 778                 GOTO(out, rc = -ENODEV);
 779         } else if (export != NULL) {
 780                 spin_lock(&export->exp_lock);
 781                 export->exp_connecting = 1;
 782                 spin_unlock(&export->exp_lock);
 783                 class_export_put(export);
 784                 spin_unlock(&target->obd_dev_lock);
 785                 LASSERT(export->exp_obd == target);
 786
 787                 rc = target_handle_reconnect(&conn, export, &cluuid);
 788         }
 789
 790         /* If we found an export, we already unlocked. */
 791         if (!export) {
 792                 spin_unlock(&target->obd_dev_lock);
 793                 OBD_FAIL_TIMEOUT(OBD_FAIL_TGT_DELAY_CONNECT, 2 * obd_timeout);
 794         } else if (req->rq_export == NULL &&
 795                    atomic_read(&export->exp_rpc_count) > 0) {
 796                 CWARN("%s: refuse connection from %s/%s to 0x%p; still busy "
 797                       "with %d references\n", target->obd_name, cluuid.uuid,
 798                       libcfs_nid2str(req->rq_peer.nid),
 799                       export, atomic_read(&export->exp_refcount));
 800                 GOTO(out, rc = -EBUSY);
 801         } else if (req->rq_export != NULL &&
 802                    atomic_read(&export->exp_rpc_count) > 1) {
 803                 CWARN("%s: refuse reconnection from %s@%s to 0x%p; still busy "
 804                       "with %d active RPCs\n", target->obd_name, cluuid.uuid,
 805                       libcfs_nid2str(req->rq_peer.nid),
 806                       export, atomic_read(&export->exp_rpc_count));
 807                 GOTO(out, rc = -EBUSY);
 808         } else if (lustre_msg_get_conn_cnt(req->rq_reqmsg) == 1) {
 809                 CERROR("%s: NID %s (%s) reconnected with 1 conn_cnt; "
 810                        "cookies not random?\n", target->obd_name,
 811                        libcfs_nid2str(req->rq_peer.nid), cluuid.uuid);
 812                 GOTO(out, rc = -EALREADY);
 813         } else if (export->exp_delayed && target->obd_recovering) {
 814                 /* VBR: don't allow delayed connection during recovery */
 815                 CWARN("%s: NID %s (%s) export was already marked as delayed "
 816                       "and will wait for end of recovery\n", target->obd_name,
 817                        libcfs_nid2str(req->rq_peer.nid), cluuid.uuid);
 818                 GOTO(out, rc = -EBUSY);
 819         } else {
 820                 OBD_FAIL_TIMEOUT(OBD_FAIL_TGT_DELAY_RECONNECT, 2 * obd_timeout);
 821         }
 822
 823         if (rc < 0)
 824                 GOTO(out, rc);
 825
 826         /* Tell the client if we're in recovery. */
 827         if (target->obd_recovering) {
 828                 lustre_msg_add_op_flags(req->rq_repmsg, MSG_CONNECT_RECOVERING);
 829                 /* If this is the first time a client connects,
 830                    reset the recovery timer */
 831                 if (rc == 0)
 832                         target_start_and_reset_recovery_timer(target, handler,
 833                                                               req, !export);
 834         }
 835
 836         /* We want to handle EALREADY but *not* -EALREADY from
 837          * target_handle_reconnect(), return reconnection state in a flag */
 838         if (rc == EALREADY) {
 839                 lustre_msg_add_op_flags(req->rq_repmsg, MSG_CONNECT_RECONNECT);
 840                 rc = 0;
 841         } else {
 842                 LASSERT(rc == 0);
 843         }
 844
 845         /* Tell the client if we support replayable requests */
 846         if (target->obd_replayable)
 847                 lustre_msg_add_op_flags(req->rq_repmsg, MSG_CONNECT_REPLAYABLE);
 848         client_nid = &req->rq_peer.nid;
 849
 850         /* VBR: for delayed connections we start recovery */
 851         if (export && export->exp_delayed && !export->exp_in_recovery) {
 852                 LASSERT(!target->obd_recovering);
 853                 LASSERT(data && data->ocd_connect_flags & OBD_CONNECT_VBR);
 854                 lustre_msg_add_op_flags(req->rq_repmsg, MSG_CONNECT_DELAYED |
 855                                         MSG_CONNECT_RECOVERING);
 856                 spin_lock_bh(&target->obd_processing_task_lock);
 857                 target->obd_version_recov = 1;
 858                 spin_unlock_bh(&target->obd_processing_task_lock);
 859                 target_start_and_reset_recovery_timer(target, handler, req, 1);
 860         }
 861
 862         if (export == NULL) {
 863                 if (target->obd_recovering) {
 864                         CERROR("%s: denying connection for new client %s (%s): "
 865                                "%d clients in recovery for %lds\n",
 866                                target->obd_name,
 867                                libcfs_nid2str(req->rq_peer.nid), cluuid.uuid,
 868                                target->obd_recoverable_clients,
 869                                cfs_duration_sec(cfs_time_sub(cfs_timer_deadline(&target->obd_recovery_timer),
 870                                                              cfs_time_current())));
 871                         rc = -EBUSY;
 872                 } else {
 873  dont_check_exports:
 874                         rc = obd_connect(&conn, target, &cluuid, data,
 875                                          client_nid);
 876                 }
 877         } else {
 878                 rc = obd_reconnect(export, target, &cluuid, data, client_nid);
 879         }
 880
 881         if (rc)
 882                 GOTO(out, rc);
 883
 884         /* Return only the parts of obd_connect_data that we understand, so the
 885          * client knows that we don't understand the rest. */
 886         if (data)
 887                 memcpy(lustre_msg_buf(req->rq_repmsg, REPLY_REC_OFF,
 888                                       sizeof(*data)),
 889                        data, sizeof(*data));
 890
 891         /* If all else goes well, this is our RPC return code. */
 892         req->rq_status = 0;
 893
 894         lustre_msg_set_handle(req->rq_repmsg, &conn);
 895
 896         /* ownership of this export ref transfers to the request AFTER we
 897          * drop any previous reference the request had, but we don't want
 898          * that to go to zero before we get our new export reference. */
 899         export = class_conn2export(&conn);
 900         if (!export) {
 901                 DEBUG_REQ(D_ERROR, req, "Missing export!");
 902                 GOTO(out, rc = -ENODEV);
 903         }
 904
 905         /* If the client and the server are the same node, we will already
 906          * have an export that really points to the client's DLM export,
 907          * because we have a shared handles table.
 908          *
 909          * XXX this will go away when shaver stops sending the "connect" handle
 910          * in the real "remote handle" field of the request --phik 24 Apr 2003
 911          */
 912         if (req->rq_export != NULL)
 913                 class_export_put(req->rq_export);
 914
 915         req->rq_export = export;
 916
 917         spin_lock(&export->exp_lock);
 918         if (export->exp_conn_cnt >= lustre_msg_get_conn_cnt(req->rq_reqmsg)) {
 919                 CERROR("%s: %s already connected at higher conn_cnt: %d > %d\n",
 920                        cluuid.uuid, libcfs_nid2str(req->rq_peer.nid),
 921                        export->exp_conn_cnt,
 922                        lustre_msg_get_conn_cnt(req->rq_reqmsg));
 923
 924                 spin_unlock(&export->exp_lock);
 925                 GOTO(out, rc = -EALREADY);
 926         }
 927         export->exp_conn_cnt = lustre_msg_get_conn_cnt(req->rq_reqmsg);
 928
 929         /* request from liblustre?  Don't evict it for not pinging. */
 930         if (lustre_msg_get_op_flags(req->rq_reqmsg) & MSG_CONNECT_LIBCLIENT) {
 931                 export->exp_libclient = 1;
 932                 spin_unlock(&export->exp_lock);
 933
 934                 spin_lock(&target->obd_dev_lock);
 935                 list_del_init(&export->exp_obd_chain_timed);
 936                 spin_unlock(&target->obd_dev_lock);
 937         } else {
 938                 spin_unlock(&export->exp_lock);
 939         }
 940
 941         if (export->exp_connection != NULL)
 942                 ptlrpc_connection_put(export->exp_connection);
 943         export->exp_connection = ptlrpc_connection_get(req->rq_peer,
 944                                                        req->rq_self,
 945                                                        &remote_uuid);
 946
 947         if (hlist_unhashed(&export->exp_nid_hash)) {
 948                 lustre_hash_add_unique(export->exp_obd->obd_nid_hash,
 949                                        &export->exp_connection->c_peer.nid,
 950                                        &export->exp_nid_hash);
 951         }
 952
 953         if (lustre_msg_get_op_flags(req->rq_repmsg) & MSG_CONNECT_RECONNECT) {
 954                 revimp = class_import_get(export->exp_imp_reverse);
 955                 GOTO(set_flags, rc = 0);
 956         }
 957
 958         if (target->obd_recovering && !export->exp_in_recovery) {
 959                 spin_lock(&export->exp_lock);
 960                 export->exp_in_recovery = 1;
 961                 spin_unlock(&export->exp_lock);
 962                 target->obd_connected_clients++;
 963         }
 964         memcpy(&conn,
 965                lustre_msg_buf(req->rq_reqmsg, REQ_REC_OFF + 2, sizeof conn),
 966                sizeof conn);
 967
 968         if (export->exp_imp_reverse != NULL)
 969                 class_destroy_import(export->exp_imp_reverse);
 970         revimp = export->exp_imp_reverse = class_new_import(target);
 971         revimp->imp_connection = ptlrpc_connection_addref(export->exp_connection);
 972         revimp->imp_client = &export->exp_obd->obd_ldlm_client;
 973         revimp->imp_remote_handle = conn;
 974         revimp->imp_dlm_fake = 1;
 975         revimp->imp_state = LUSTRE_IMP_FULL;
 976
 977 set_flags:
 978         if (req->rq_reqmsg->lm_magic == LUSTRE_MSG_MAGIC_V1 &&
 979             lustre_msg_get_op_flags(req->rq_reqmsg) & MSG_CONNECT_NEXT_VER) {
 980                 revimp->imp_msg_magic = LUSTRE_MSG_MAGIC_V2;
 981                 lustre_msg_add_op_flags(req->rq_repmsg, MSG_CONNECT_NEXT_VER);
 982         } else {
 983                 /* unknown versions will be caught in
 984                  * ptlrpc_handle_server_req_in->lustre_unpack_msg() */
 985                 revimp->imp_msg_magic = req->rq_reqmsg->lm_magic;
 986         }
 987
 988         if (revimp->imp_msg_magic != LUSTRE_MSG_MAGIC_V1) {
 989                 if (export->exp_connect_flags & OBD_CONNECT_AT)
 990                         revimp->imp_msghdr_flags |= MSGHDR_AT_SUPPORT;
 991                 else
 992                         revimp->imp_msghdr_flags &= ~MSGHDR_AT_SUPPORT;
 993         }
 994
 995         class_import_put(revimp);
 996 out:
 997         if (export) {
 998                 spin_lock(&export->exp_lock);
 999                 export->exp_connecting = 0;
1000                 spin_unlock(&export->exp_lock);
1001         }
1002         if (targref)
1003                 class_decref(targref);
1004         if (rc)
1005                 req->rq_status = rc;
1006         RETURN(rc);
1007 }
1008
1009 int target_handle_disconnect(struct ptlrpc_request *req)
1010 {
1011         int rc;
1012         ENTRY;
1013
1014         rc = lustre_pack_reply(req, 1, NULL, NULL);
1015         if (rc)
1016                 RETURN(rc);
1017
1018         /* keep the rq_export around so we can send the reply */
1019         req->rq_status = obd_disconnect(class_export_get(req->rq_export));
1020         RETURN(0);
1021 }
1022
1023 void target_destroy_export(struct obd_export *exp)
1024 {
1025         /* exports created from last_rcvd data, and "fake"
1026            exports created by lctl don't have an import */
1027         if (exp->exp_imp_reverse != NULL)
1028                 class_destroy_import(exp->exp_imp_reverse);
1029
1030         /* We cancel locks at disconnect time, but this will catch any locks
1031          * granted in a race with recovery-induced disconnect. */
1032         if (exp->exp_obd->obd_namespace != NULL)
1033                 ldlm_cancel_locks_for_export(exp);
1034 }
1035
1036 /*
1037  * Recovery functions
1038  */
1039
1040 static int target_exp_enqueue_req_replay(struct ptlrpc_request *req)
1041 {
1042         __u64                  transno = lustre_msg_get_transno(req->rq_reqmsg);
1043         struct obd_export     *exp = req->rq_export;
1044         struct ptlrpc_request *reqiter;
1045         int                    dup = 0;
1046
1047         LASSERT(exp);
1048
1049         spin_lock(&exp->exp_lock);
1050         list_for_each_entry(reqiter, &exp->exp_req_replay_queue,
1051                             rq_replay_list) {
1052                 if (lustre_msg_get_transno(reqiter->rq_reqmsg) == transno) {
1053                         dup = 1;
1054                         break;
1055                 }
1056         }
1057
1058         if (dup) {
1059                 /* we expect it with RESENT and REPLAY flags */
1060                 if ((lustre_msg_get_flags(req->rq_reqmsg) &
1061                      (MSG_RESENT | MSG_REPLAY)) != (MSG_RESENT | MSG_REPLAY))
1062                         CERROR("invalid flags %x of resent replay\n",
1063                                lustre_msg_get_flags(req->rq_reqmsg));
1064         } else {
1065                 list_add_tail(&req->rq_replay_list, &exp->exp_req_replay_queue);
1066         }
1067
1068         spin_unlock(&exp->exp_lock);
1069         return dup;
1070 }
1071
1072 static void target_exp_dequeue_req_replay(struct ptlrpc_request *req)
1073 {
1074         LASSERT(!list_empty(&req->rq_replay_list));
1075         LASSERT(req->rq_export);
1076
1077         spin_lock(&req->rq_export->exp_lock);
1078         list_del_init(&req->rq_replay_list);
1079         spin_unlock(&req->rq_export->exp_lock);
1080 }
1081
1082 static void target_release_saved_req(struct ptlrpc_request *req)
1083 {
1084         ptlrpc_req_drop_rs(req);
1085         class_export_put(req->rq_export);
1086         OBD_FREE(req->rq_reqmsg, req->rq_reqlen);
1087         OBD_FREE(req, sizeof *req);
1088 }
1089
1090 static void target_send_delayed_replies(struct obd_device *obd)
1091 {
1092         struct ptlrpc_request *req, *tmp;
1093
1094         LCONSOLE_INFO("%s: sending delayed replies to recovered clients\n",
1095                       obd->obd_name);
1096
1097         list_for_each_entry_safe(req, tmp, &obd->obd_delayed_reply_queue,
1098                                  rq_list) {
1099                 list_del_init(&req->rq_list);
1100                 DEBUG_REQ(D_HA, req, "delayed:");
1101                 ptlrpc_reply(req);
1102                 target_release_saved_req(req);
1103         }
1104         obd->obd_recovery_end = cfs_time_current_sec();
1105 }
1106
1107 static void target_finish_recovery(struct obd_device *obd)
1108 {
1109         ldlm_reprocess_all_ns(obd->obd_namespace);
1110         spin_lock_bh(&obd->obd_processing_task_lock);
1111         if (list_empty(&obd->obd_recovery_queue)) {
1112                 obd->obd_processing_task = 0;
1113         } else {
1114                 spin_unlock_bh(&obd->obd_processing_task_lock);
1115                 CERROR("%s: Recovery queue isn't empty\n", obd->obd_name);
1116                 LBUG();
1117         }
1118         spin_unlock_bh(&obd->obd_processing_task_lock);
1119                 ;
1120         /* when recovery finished, cleanup orphans on mds and ost */
1121         if (OBT(obd) && OBP(obd, postrecov)) {
1122                 int rc = OBP(obd, postrecov)(obd);
1123                 LCONSOLE_WARN("%s: recovery %s: rc %d\n", obd->obd_name,
1124                               rc < 0 ? "failed" : "complete", rc);
1125         }
1126         target_send_delayed_replies(obd);
1127 }
1128
1129 static void abort_recovery_queue(struct obd_device *obd)
1130 {
1131         struct ptlrpc_request *req, *n;
1132         struct list_head abort_list;
1133         int rc;
1134
1135         CFS_INIT_LIST_HEAD(&abort_list);
1136         spin_lock_bh(&obd->obd_processing_task_lock);
1137         list_splice_init(&obd->obd_recovery_queue, &abort_list);
1138         spin_unlock_bh(&obd->obd_processing_task_lock);
1139         /* process abort list unlocked */
1140         list_for_each_entry_safe(req, n, &abort_list, rq_list) {
1141                 target_exp_dequeue_req_replay(req);
1142                 list_del_init(&req->rq_list);
1143                 DEBUG_REQ(D_ERROR, req, "aborted:");
1144                 req->rq_status = -ENOTCONN;
1145                 req->rq_type = PTL_RPC_MSG_ERR;
1146                 rc = lustre_pack_reply(req, 1, NULL, NULL);
1147                 if (rc == 0)
1148                         ptlrpc_reply(req);
1149                 else
1150                         DEBUG_REQ(D_ERROR, req,
1151                                   "packing failed for abort-reply; skipping");
1152                 target_release_saved_req(req);
1153         }
1154 }
1155
1156 /* Called from a cleanup function if the device is being cleaned up
1157    forcefully.  The exports should all have been disconnected already,
1158    the only thing left to do is
1159      - clear the recovery flags
1160      - cancel the timer
1161      - free queued requests and replies, but don't send replies
1162    Because the obd_stopping flag is set, no new requests should be received.
1163
1164 */
1165 void target_cleanup_recovery(struct obd_device *obd)
1166 {
1167         struct list_head *tmp, *n;
1168         struct ptlrpc_request *req;
1169         struct list_head clean_list;
1170         ENTRY;
1171
1172         LASSERT(obd->obd_stopping);
1173
1174         spin_lock_bh(&obd->obd_processing_task_lock);
1175         if (!obd->obd_recovering) {
1176                 spin_unlock_bh(&obd->obd_processing_task_lock);
1177                 EXIT;
1178                 return;
1179         }
1180         obd->obd_recovering = obd->obd_abort_recovery = 0;
1181         target_cancel_recovery_timer(obd);
1182         spin_unlock_bh(&obd->obd_processing_task_lock);
1183
1184         list_for_each_safe(tmp, n, &obd->obd_delayed_reply_queue) {
1185                 req = list_entry(tmp, struct ptlrpc_request, rq_list);
1186                 list_del(&req->rq_list);
1187                 target_release_saved_req(req);
1188         }
1189
1190         CFS_INIT_LIST_HEAD(&clean_list);
1191         spin_lock_bh(&obd->obd_processing_task_lock);
1192         list_splice_init(&obd->obd_recovery_queue, &clean_list);
1193         spin_unlock_bh(&obd->obd_processing_task_lock);
1194         list_for_each_safe(tmp, n, &clean_list) {
1195                 req = list_entry(tmp, struct ptlrpc_request, rq_list);
1196                 target_exp_dequeue_req_replay(req);
1197                 list_del_init(&req->rq_list);
1198                 target_release_saved_req(req);
1199         }
1200         EXIT;
1201 }
1202
1203 void target_stop_recovery(void *data, int abort)
1204 {
1205         struct obd_device *obd = data;
1206         enum obd_option flags;
1207         ENTRY;
1208
1209         spin_lock_bh(&obd->obd_processing_task_lock);
1210         if (!obd->obd_recovering) {
1211                 spin_unlock_bh(&obd->obd_processing_task_lock);
1212                 EXIT;
1213                 return;
1214         }
1215         flags = exp_flags_from_obd(obd) | OBD_OPT_ABORT_RECOV;
1216         obd->obd_recovering = 0;
1217         obd->obd_abort_recovery = 0;
1218         obd->obd_processing_task = 0;
1219         if (abort == 0)
1220                 LASSERT(obd->obd_recoverable_clients == 0);
1221
1222         target_cancel_recovery_timer(obd);
1223         spin_unlock_bh(&obd->obd_processing_task_lock);
1224
1225         if (abort) {
1226                 LCONSOLE_WARN("%s: recovery is aborted by administrative "
1227                               "request; %d clients are not recovered "
1228                               "(%d clients did)\n", obd->obd_name,
1229                               obd->obd_recoverable_clients,
1230                               obd->obd_connected_clients);
1231                 class_disconnect_stale_exports(obd, flags);
1232         }
1233         abort_recovery_queue(obd);
1234         target_finish_recovery(obd);
1235         CDEBUG(D_HA, "%s: recovery complete\n", obd_uuid2str(&obd->obd_uuid));
1236         EXIT;
1237 }
1238
1239 void target_abort_recovery(void *data)
1240 {
1241         target_stop_recovery(data, 1);
1242 }
1243
1244 static void reset_recovery_timer(struct obd_device *, int, int);
1245 static void target_recovery_expired(unsigned long castmeharder)
1246 {
1247         struct obd_device *obd = (struct obd_device *)castmeharder;
1248         int version_recov = obd->obd_version_recov;
1249         LCONSOLE_WARN("%s: recovery period over; %d clients never reconnected "
1250                       "after %lds (%d clients did)\n",
1251                       obd->obd_name, obd->obd_recoverable_clients,
1252                       cfs_time_current_sec()- obd->obd_recovery_start,
1253                       obd->obd_connected_clients);
1254
1255         /** check is fs version-capable */
1256         if (target_fs_version_capable(obd)) {
1257                 class_handle_stale_exports(obd);
1258         } else {
1259                 CWARN("Versions are not supported by ldiskfs, VBR is OFF\n");
1260                 class_disconnect_stale_exports(obd, exp_flags_from_obd(obd));
1261         }
1262         spin_lock_bh(&obd->obd_processing_task_lock);
1263         /* VBR: no clients are remained to replay, stop recovery */
1264         if (obd->obd_recovering && obd->obd_recoverable_clients == 0)
1265                 obd->obd_abort_recovery = 1;
1266         /* always check versions now */
1267         obd->obd_version_recov = 1;
1268         cfs_waitq_signal(&obd->obd_next_transno_waitq);
1269         spin_unlock_bh(&obd->obd_processing_task_lock);
1270         /* reset timer if recovery will proceed with versions now */
1271         reset_recovery_timer(obd, OBD_RECOVERY_FACTOR * obd_timeout,
1272                              !version_recov);
1273 }
1274
1275 /* obd_processing_task_lock should be held */
1276 void target_cancel_recovery_timer(struct obd_device *obd)
1277 {
1278         CDEBUG(D_HA, "%s: cancel recovery timer\n", obd->obd_name);
1279         cfs_timer_disarm(&obd->obd_recovery_timer);
1280 }
1281
1282 /* extend = 1 means require at least "duration" seconds left in the timer,
1283    extend = 0 means set the total duration (start_recovery_timer) */
1284 static void reset_recovery_timer(struct obd_device *obd, int duration,
1285                                  int extend)
1286 {
1287         cfs_time_t now = cfs_time_current_sec();
1288         cfs_duration_t left;
1289
1290         spin_lock_bh(&obd->obd_processing_task_lock);
1291         if (!obd->obd_recovering) {
1292                 spin_unlock_bh(&obd->obd_processing_task_lock);
1293                 return;
1294         }
1295
1296         left = cfs_time_sub(obd->obd_recovery_end, now);
1297
1298         if (extend && (duration > left))
1299                 obd->obd_recovery_timeout += duration - left;
1300         else if (!extend && (duration > obd->obd_recovery_timeout))
1301                 /* Track the client's largest expected replay time */
1302                 obd->obd_recovery_timeout = duration;
1303 #ifdef CRAY_XT3
1304         /*
1305          * If total recovery time already exceed the
1306          * obd_recovery_max_time, then CRAY XT3 will
1307          * abort the recovery
1308          */
1309         if(obd->obd_recovery_timeout > obd->obd_recovery_max_time)
1310                 obd->obd_recovery_timeout = obd->obd_recovery_max_time;
1311 #endif
1312         obd->obd_recovery_end = obd->obd_recovery_start +
1313                                 obd->obd_recovery_timeout;
1314         if (cfs_time_before(now, obd->obd_recovery_end)) {
1315                 left = cfs_time_sub(obd->obd_recovery_end, now);
1316                 cfs_timer_arm(&obd->obd_recovery_timer, cfs_time_shift(left));
1317         }
1318         spin_unlock_bh(&obd->obd_processing_task_lock);
1319         CDEBUG(D_HA, "%s: recovery timer will expire in %u seconds\n",
1320                obd->obd_name, (unsigned)left);
1321 }
1322
1323 static void check_and_start_recovery_timer(struct obd_device *obd,
1324                                            svc_handler_t handler)
1325 {
1326         spin_lock_bh(&obd->obd_processing_task_lock);
1327         if (obd->obd_recovery_handler) {
1328                 spin_unlock_bh(&obd->obd_processing_task_lock);
1329                 return;
1330         }
1331         CWARN("%s: starting recovery timer\n", obd->obd_name);
1332         obd->obd_recovery_start = cfs_time_current_sec();
1333         /* minimum */
1334         obd->obd_recovery_timeout = OBD_RECOVERY_FACTOR * obd_timeout;
1335         obd->obd_recovery_handler = handler;
1336         cfs_timer_init(&obd->obd_recovery_timer, target_recovery_expired, obd);
1337         spin_unlock_bh(&obd->obd_processing_task_lock);
1338
1339         reset_recovery_timer(obd, obd->obd_recovery_timeout, 0);
1340 }
1341
1342 /* Reset the timer with each new client connection */
1343 /*
1344  * This timer is actually reconnect_timer, which is for making sure
1345  * the total recovery window is at least as big as my reconnect
1346  * attempt timing. So the initial recovery time_out will be set to
1347  * OBD_RECOVERY_FACTOR * obd_timeout. If the timeout coming
1348  * from client is bigger than this, then the recovery time_out will
1349  * be extend to make sure the client could be reconnected, in the
1350  * process, the timeout from the new client should be ignored.
1351  */
1352
1353 static void
1354 target_start_and_reset_recovery_timer(struct obd_device *obd,
1355                                       svc_handler_t handler,
1356                                       struct ptlrpc_request *req,
1357                                       int new_client)
1358 {
1359         int service_time = lustre_msg_get_service_time(req->rq_reqmsg);
1360
1361         if (!new_client && service_time)
1362                 /* Teach server about old server's estimates, as first guess
1363                    at how long new requests will take. */
1364                 at_add(&req->rq_rqbd->rqbd_service->srv_at_estimate,
1365                        service_time);
1366
1367         check_and_start_recovery_timer(obd, handler);
1368
1369         /* convert the service time to rpc timeout,
1370          * reuse service_time to limit stack usage */
1371         service_time = at_est2timeout(service_time);
1372
1373         /* We expect other clients to timeout within service_time, then try
1374          * to reconnect, then try the failover server.  The max delay between
1375          * connect attempts is SWITCH_MAX + SWITCH_INC + INITIAL */
1376         service_time += 2 * (CONNECTION_SWITCH_MAX + CONNECTION_SWITCH_INC +
1377                              INITIAL_CONNECT_TIMEOUT);
1378         if (service_time > obd->obd_recovery_timeout && !new_client)
1379                 reset_recovery_timer(obd, service_time, 0);
1380 }
1381
1382 static int check_for_next_transno(struct obd_device *obd)
1383 {
1384         struct ptlrpc_request *req;
1385         int wake_up = 0, connected, completed, queue_len, max;
1386         __u64 next_transno, req_transno;
1387
1388         spin_lock_bh(&obd->obd_processing_task_lock);
1389         req = list_entry(obd->obd_recovery_queue.next,
1390                          struct ptlrpc_request, rq_list);
1391         max = obd->obd_max_recoverable_clients;
1392         req_transno = lustre_msg_get_transno(req->rq_reqmsg);
1393         connected = obd->obd_connected_clients;
1394         completed = max - obd->obd_recoverable_clients -
1395                     obd->obd_delayed_clients;
1396         queue_len = obd->obd_requests_queued_for_recovery;
1397         next_transno = obd->obd_next_recovery_transno;
1398
1399         CDEBUG(D_HA,"max: %d, connected: %d, delayed %d, completed: %d, "
1400                "queue_len: %d, req_transno: "LPU64", next_transno: "LPU64"\n",
1401                max, connected, obd->obd_delayed_clients, completed, queue_len,
1402                req_transno, next_transno);
1403         if (obd->obd_abort_recovery) {
1404                 CDEBUG(D_HA, "waking for aborted recovery\n");
1405                 wake_up = 1;
1406         } else if (!obd->obd_recovering) {
1407                 CDEBUG(D_HA, "waking for completed recovery (?)\n");
1408                 wake_up = 1;
1409         } else if (req_transno == next_transno) {
1410                 CDEBUG(D_HA, "waking for next ("LPD64")\n", next_transno);
1411                 wake_up = 1;
1412         } else if (queue_len == obd->obd_recoverable_clients) {
1413                 CDEBUG(D_ERROR,
1414                        "waking for skipped transno (skip: "LPD64
1415                        ", ql: %d, comp: %d, conn: %d, next: "LPD64")\n",
1416                        next_transno, queue_len, completed, max, req_transno);
1417                 obd->obd_next_recovery_transno = req_transno;
1418                 wake_up = 1;
1419         }
1420         spin_unlock_bh(&obd->obd_processing_task_lock);
1421         LASSERT(lustre_msg_get_transno(req->rq_reqmsg) >= next_transno);
1422         return wake_up;
1423 }
1424
1425 static void process_recovery_queue(struct obd_device *obd)
1426 {
1427         struct ptlrpc_request *req;
1428         struct l_wait_info lwi = { 0 };
1429         ENTRY;
1430
1431         for (;;) {
1432                 spin_lock_bh(&obd->obd_processing_task_lock);
1433                 LASSERT(obd->obd_processing_task == cfs_curproc_pid());
1434                 req = list_entry(obd->obd_recovery_queue.next,
1435                                  struct ptlrpc_request, rq_list);
1436
1437                 if (lustre_msg_get_transno(req->rq_reqmsg) !=
1438                     obd->obd_next_recovery_transno) {
1439                         spin_unlock_bh(&obd->obd_processing_task_lock);
1440                         CDEBUG(D_HA, "Waiting for transno "LPD64" (1st is "
1441                                LPD64", x"LPU64")\n",
1442                                obd->obd_next_recovery_transno,
1443                                lustre_msg_get_transno(req->rq_reqmsg),
1444                                req->rq_xid);
1445                         l_wait_event(obd->obd_next_transno_waitq,
1446                                      check_for_next_transno(obd), &lwi);
1447                         if (target_recovery_check_and_stop(obd))
1448                                 return;
1449                         continue;
1450                 }
1451                 target_exp_dequeue_req_replay(req);
1452                 list_del_init(&req->rq_list);
1453                 obd->obd_requests_queued_for_recovery--;
1454                 spin_unlock_bh(&obd->obd_processing_task_lock);
1455
1456                 DEBUG_REQ(D_HA, req, "processing: ");
1457                 (void)obd->obd_recovery_handler(req);
1458                 obd->obd_replayed_requests++;
1459                 /* Extend the recovery timer enough to complete the next
1460                  * replayed rpc */
1461                 reset_recovery_timer(obd, AT_OFF ? obd_timeout :
1462                        at_get(&req->rq_rqbd->rqbd_service->srv_at_estimate), 1);
1463                 /* bug 1580: decide how to properly sync() in recovery */
1464                 //mds_fsync_super(obd->u.obt.obt_sb);
1465                 class_export_put(req->rq_export);
1466                 ptlrpc_req_drop_rs(req);
1467                 OBD_FREE(req->rq_reqmsg, req->rq_reqlen);
1468                 OBD_FREE(req, sizeof *req);
1469                 OBD_RACE(OBD_FAIL_TGT_REPLAY_DELAY);
1470                 spin_lock_bh(&obd->obd_processing_task_lock);
1471                 obd->obd_next_recovery_transno++;
1472                 if (list_empty(&obd->obd_recovery_queue)) {
1473                         obd->obd_processing_task = 0;
1474                         spin_unlock_bh(&obd->obd_processing_task_lock);
1475                         break;
1476                 }
1477                 spin_unlock_bh(&obd->obd_processing_task_lock);
1478         }
1479         EXIT;
1480 }
1481
1482 int target_queue_recovery_request(struct ptlrpc_request *req,
1483                                   struct obd_device *obd)
1484 {
1485         struct list_head *tmp;
1486         int inserted = 0;
1487         __u64 transno = lustre_msg_get_transno(req->rq_reqmsg);
1488         struct ptlrpc_request *saved_req;
1489         struct lustre_msg *reqmsg;
1490         int rc = 0;
1491
1492         /* CAVEAT EMPTOR: The incoming request message has been swabbed
1493          * (i.e. buflens etc are in my own byte order), but type-dependent
1494          * buffers (eg mds_body, ost_body etc) have NOT been swabbed. */
1495
1496         if (!transno) {
1497                 CFS_INIT_LIST_HEAD(&req->rq_list);
1498                 DEBUG_REQ(D_HA, req, "not queueing");
1499                 return 1;
1500         }
1501
1502         /* XXX If I were a real man, these LBUGs would be sane cleanups. */
1503         /* XXX just like the request-dup code in queue_final_reply */
1504         OBD_ALLOC(saved_req, sizeof *saved_req);
1505         if (!saved_req)
1506                 LBUG();
1507         OBD_ALLOC(reqmsg, req->rq_reqlen);
1508         if (!reqmsg)
1509                 LBUG();
1510
1511         spin_lock_bh(&obd->obd_processing_task_lock);
1512
1513         /* If we're processing the queue, we want don't want to queue this
1514          * message.
1515          *
1516          * Also, if this request has a transno less than the one we're waiting
1517          * for, we should process it now.  It could (and currently always will)
1518          * be an open request for a descriptor that was opened some time ago.
1519          *
1520          * Also, a resent, replayed request that has already been
1521          * handled will pass through here and be processed immediately.
1522          */
1523         if (obd->obd_processing_task == cfs_curproc_pid() ||
1524             transno < obd->obd_next_recovery_transno) {
1525                 /* Processing the queue right now, don't re-add. */
1526                 LASSERT(list_empty(&req->rq_list));
1527                 spin_unlock_bh(&obd->obd_processing_task_lock);
1528                 GOTO(err_free, rc = 1);
1529         }
1530
1531         if (unlikely(OBD_FAIL_CHECK(OBD_FAIL_TGT_REPLAY_DROP))) {
1532                 spin_unlock_bh(&obd->obd_processing_task_lock);
1533                 GOTO(err_free, rc = 0);
1534         }
1535
1536         memcpy(saved_req, req, sizeof *req);
1537         memcpy(reqmsg, req->rq_reqmsg, req->rq_reqlen);
1538         req = saved_req;
1539         req->rq_reqmsg = reqmsg;
1540         class_export_get(req->rq_export);
1541         CFS_INIT_LIST_HEAD(&req->rq_list);
1542         CFS_INIT_LIST_HEAD(&req->rq_replay_list);
1543
1544         if (target_exp_enqueue_req_replay(req)) {
1545                 spin_unlock_bh(&obd->obd_processing_task_lock);
1546                 DEBUG_REQ(D_ERROR, req, "dropping resent queued req");
1547                 GOTO(err_exp, rc = 0);
1548         }
1549
1550         /* XXX O(n^2) */
1551         list_for_each(tmp, &obd->obd_recovery_queue) {
1552                 struct ptlrpc_request *reqiter =
1553                         list_entry(tmp, struct ptlrpc_request, rq_list);
1554
1555                 if (lustre_msg_get_transno(reqiter->rq_reqmsg) > transno) {
1556                         list_add_tail(&req->rq_list, &reqiter->rq_list);
1557                         inserted = 1;
1558                         break;
1559                 }
1560
1561                 if (unlikely(lustre_msg_get_transno(reqiter->rq_reqmsg) ==
1562                              transno)) {
1563                         spin_unlock_bh(&obd->obd_processing_task_lock);
1564                         DEBUG_REQ(D_ERROR, req, "dropping replay: transno "
1565                                   "has been claimed by another client");
1566                         target_exp_dequeue_req_replay(req);
1567                         GOTO(err_exp, rc = 0);
1568                 }
1569         }
1570
1571         if (!inserted) {
1572                 list_add_tail(&req->rq_list, &obd->obd_recovery_queue);
1573         }
1574
1575         obd->obd_requests_queued_for_recovery++;
1576
1577         if (obd->obd_processing_task != 0) {
1578                 /* Someone else is processing this queue, we'll leave it to
1579                  * them.
1580                  */
1581                 cfs_waitq_signal(&obd->obd_next_transno_waitq);
1582                 spin_unlock_bh(&obd->obd_processing_task_lock);
1583                 return 0;
1584         }
1585
1586         /* Nobody is processing, and we know there's (at least) one to process
1587          * now, so we'll do the honours.
1588          */
1589         obd->obd_processing_task = cfs_curproc_pid();
1590         spin_unlock_bh(&obd->obd_processing_task_lock);
1591
1592         process_recovery_queue(obd);
1593         return 0;
1594
1595 err_exp:
1596         class_export_put(req->rq_export);
1597 err_free:
1598         OBD_FREE(reqmsg, req->rq_reqlen);
1599         OBD_FREE(saved_req, sizeof(*saved_req));
1600         return rc;
1601 }
1602
1603 struct obd_device * target_req2obd(struct ptlrpc_request *req)
1604 {
1605         return req->rq_export->exp_obd;
1606 }
1607
1608 int target_queue_last_replay_reply(struct ptlrpc_request *req, int rc)
1609 {
1610         struct obd_device *obd = target_req2obd(req);
1611         struct ptlrpc_request *saved_req;
1612         struct lustre_msg *reqmsg;
1613         struct obd_export *exp = req->rq_export;
1614         int recovery_done = 0, delayed_done = 0;
1615
1616         LASSERT ((rc == 0) == req->rq_packed_final);
1617
1618         if (!req->rq_packed_final) {
1619                 /* Just like ptlrpc_error, but without the sending. */
1620                 rc = lustre_pack_reply(req, 1, NULL, NULL);
1621                 if (rc)
1622                         return rc;
1623                 req->rq_type = PTL_RPC_MSG_ERR;
1624         }
1625
1626         LASSERT(!req->rq_reply_state->rs_difficult);
1627         LASSERT(list_empty(&req->rq_list));
1628         /* XXX a bit like the request-dup code in queue_recovery_request */
1629         OBD_ALLOC(saved_req, sizeof *saved_req);
1630         if (!saved_req)
1631                 return -ENOMEM;
1632         OBD_ALLOC(reqmsg, req->rq_reqlen);
1633         if (!reqmsg) {
1634                 OBD_FREE(saved_req, sizeof *req);
1635                 return -ENOMEM;
1636         }
1637         *saved_req = *req;
1638         memcpy(reqmsg, req->rq_reqmsg, req->rq_reqlen);
1639
1640         /* Don't race cleanup */
1641         spin_lock_bh(&obd->obd_processing_task_lock);
1642         if (obd->obd_stopping) {
1643                 spin_unlock_bh(&obd->obd_processing_task_lock);
1644                 goto out_noconn;
1645         }
1646
1647         if (!exp->exp_vbr_failed) {
1648                 ptlrpc_rs_addref(req->rq_reply_state);  /* +1 ref for saved reply */
1649                 req = saved_req;
1650                 req->rq_reqmsg = reqmsg;
1651                 CFS_INIT_LIST_HEAD(&req->rq_list);
1652                 CFS_INIT_LIST_HEAD(&req->rq_replay_list);
1653                 class_export_get(exp);
1654                 list_add(&req->rq_list, &obd->obd_delayed_reply_queue);
1655         }
1656
1657         /* only count the first "replay over" request from each
1658            export */
1659         if (exp->exp_replay_needed) {
1660                 spin_lock(&exp->exp_lock);
1661                 exp->exp_replay_needed = 0;
1662                 spin_unlock(&exp->exp_lock);
1663
1664                 if (!exp->exp_delayed) {
1665                         --obd->obd_recoverable_clients;
1666                 } else {
1667                         spin_lock(&exp->exp_lock);
1668                         exp->exp_delayed = 0;
1669                         spin_unlock(&exp->exp_lock);
1670                         delayed_done = 1;
1671                         if (obd->obd_delayed_clients == 0) {
1672                                 spin_unlock_bh(&obd->obd_processing_task_lock);
1673                                 LBUG();
1674                         }
1675                         --obd->obd_delayed_clients;
1676                 }
1677         }
1678         recovery_done = (obd->obd_recoverable_clients == 0);
1679         spin_unlock_bh(&obd->obd_processing_task_lock);
1680
1681         if (delayed_done) {
1682                 /* start pinging export */
1683                 spin_lock(&obd->obd_dev_lock);
1684                 list_add_tail(&exp->exp_obd_chain_timed,
1685                               &obd->obd_exports_timed);
1686                 list_move_tail(&exp->exp_obd_chain, &obd->obd_exports);
1687                 spin_unlock(&obd->obd_dev_lock);
1688                 target_send_delayed_replies(obd);
1689         }
1690
1691         OBD_RACE(OBD_FAIL_LDLM_RECOV_CLIENTS);
1692         if (recovery_done) {
1693                 spin_lock_bh(&obd->obd_processing_task_lock);
1694                 obd->obd_recovering = 0;
1695                 obd->obd_version_recov = 0;
1696                 obd->obd_abort_recovery = 0;
1697                 target_cancel_recovery_timer(obd);
1698                 spin_unlock_bh(&obd->obd_processing_task_lock);
1699
1700                 OBD_RACE(OBD_FAIL_TGT_REPLAY_DELAY);
1701
1702                 if (!delayed_done)
1703                         target_finish_recovery(obd);
1704                 CDEBUG(D_HA, "%s: recovery complete\n",
1705                        obd_uuid2str(&obd->obd_uuid));
1706         } else {
1707                 CWARN("%s: %d recoverable clients remain\n",
1708                       obd->obd_name, obd->obd_recoverable_clients);
1709                 cfs_waitq_signal(&obd->obd_next_transno_waitq);
1710         }
1711
1712         /* VBR: disconnect export with failed recovery */
1713         if (exp->exp_vbr_failed) {
1714                 CWARN("%s: disconnect export %s\n", obd->obd_name,
1715                       exp->exp_client_uuid.uuid);
1716                 class_fail_export(exp);
1717                 OBD_FREE(reqmsg, req->rq_reqlen);
1718                 OBD_FREE(saved_req, sizeof *req);
1719                 req->rq_status = 0;
1720                 ptlrpc_send_reply(req, 0);
1721         }
1722
1723         return 1;
1724
1725 out_noconn:
1726         OBD_FREE(reqmsg, req->rq_reqlen);
1727         OBD_FREE(saved_req, sizeof *req);
1728         req->rq_status = -ENOTCONN;
1729         /* rv is ignored anyhow */
1730         return -ENOTCONN;
1731 }
1732
1733 int target_handle_reply(struct ptlrpc_request *req, int rc, int fail)
1734 {
1735         struct obd_device *obd = NULL;
1736
1737         if (req->rq_export)
1738                 obd = target_req2obd(req);
1739
1740         /* handle replay reply for version recovery */
1741         if (obd && obd->obd_version_recov &&
1742             (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY)) {
1743                 LASSERT(req->rq_repmsg);
1744                 lustre_msg_add_flags(req->rq_repmsg, MSG_VERSION_REPLAY);
1745         }
1746
1747         /* handle last replay */
1748         if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_LAST_REPLAY) {
1749                 if (obd &&
1750                     lustre_msg_get_flags(req->rq_reqmsg) & MSG_DELAY_REPLAY) {
1751                         DEBUG_REQ(D_HA, req,
1752                                   "delayed LAST_REPLAY, queuing reply");
1753                         rc = target_queue_last_replay_reply(req, rc);
1754                         LASSERT(req->rq_export->exp_delayed == 0);
1755                         return rc;
1756                 }
1757
1758                 if (obd && obd->obd_recovering) { /* normal recovery */
1759                         DEBUG_REQ(D_HA, req, "LAST_REPLAY, queuing reply");
1760                         rc = target_queue_last_replay_reply(req, rc);
1761                         return rc;
1762                 }
1763
1764                 /* Lost a race with recovery; let the error path DTRT. */
1765                 rc = req->rq_status = -ENOTCONN;
1766         }
1767         target_send_reply(req, rc, fail);
1768         return 0;
1769 }
1770
1771 static inline struct ldlm_pool *ldlm_exp2pl(struct obd_export *exp)
1772 {
1773         LASSERT(exp != NULL);
1774         return &exp->exp_obd->obd_namespace->ns_pool;
1775 }
1776
1777 int target_pack_pool_reply(struct ptlrpc_request *req)
1778 {
1779         struct obd_device *obd;
1780         ENTRY;
1781
1782         /*
1783          * Check that we still have all structures alive as this may
1784          * be some late rpc in shutdown time.
1785          */
1786         if (unlikely(!req->rq_export || !req->rq_export->exp_obd ||
1787                      !exp_connect_lru_resize(req->rq_export))) {
1788                 lustre_msg_set_slv(req->rq_repmsg, 0);
1789                 lustre_msg_set_limit(req->rq_repmsg, 0);
1790                 RETURN(0);
1791         }
1792
1793         /*
1794          * OBD is alive here as export is alive, which we checked above.
1795          */
1796         obd = req->rq_export->exp_obd;
1797
1798         read_lock(&obd->obd_pool_lock);
1799         lustre_msg_set_slv(req->rq_repmsg, obd->obd_pool_slv);
1800         lustre_msg_set_limit(req->rq_repmsg, obd->obd_pool_limit);
1801         read_unlock(&obd->obd_pool_lock);
1802
1803         RETURN(0);
1804 }
1805
1806 int
1807 target_send_reply_msg (struct ptlrpc_request *req, int rc, int fail_id)
1808 {
1809         if (OBD_FAIL_CHECK(fail_id | OBD_FAIL_ONCE)) {
1810                 obd_fail_loc |= OBD_FAIL_ONCE | OBD_FAILED;
1811                 DEBUG_REQ(D_ERROR, req, "dropping reply");
1812                 return (-ECOMM);
1813         }
1814
1815         if (rc) {
1816                 DEBUG_REQ(D_ERROR, req, "processing error (%d)", rc);
1817                 req->rq_status = rc;
1818                 return (ptlrpc_send_error(req, 1));
1819         } else {
1820                 DEBUG_REQ(D_NET, req, "sending reply");
1821         }
1822
1823         return (ptlrpc_send_reply(req, PTLRPC_REPLY_MAYBE_DIFFICULT));
1824 }
1825
1826 void
1827 target_send_reply(struct ptlrpc_request *req, int rc, int fail_id)
1828 {
1829         int                        netrc;
1830         struct ptlrpc_reply_state *rs;
1831         struct obd_device         *obd;
1832         struct obd_export         *exp;
1833         struct ptlrpc_service     *svc;
1834
1835         svc = req->rq_rqbd->rqbd_service;
1836         rs = req->rq_reply_state;
1837         if (rs == NULL || !rs->rs_difficult) {
1838                 /* no notifiers */
1839                 target_send_reply_msg (req, rc, fail_id);
1840                 return;
1841         }
1842
1843         /* must be an export if locks saved */
1844         LASSERT (req->rq_export != NULL);
1845         /* req/reply consistent */
1846         LASSERT (rs->rs_service == svc);
1847
1848         /* "fresh" reply */
1849         LASSERT (!rs->rs_scheduled);
1850         LASSERT (!rs->rs_scheduled_ever);
1851         LASSERT (!rs->rs_handled);
1852         LASSERT (!rs->rs_on_net);
1853         LASSERT (rs->rs_export == NULL);
1854         LASSERT (list_empty(&rs->rs_obd_list));
1855         LASSERT (list_empty(&rs->rs_exp_list));
1856
1857         exp = class_export_get(req->rq_export);
1858         obd = exp->exp_obd;
1859
1860         /* disable reply scheduling onto srv_reply_queue while I'm setting up */
1861         rs->rs_scheduled = 1;
1862         rs->rs_on_net    = 1;
1863         rs->rs_xid       = req->rq_xid;
1864         rs->rs_transno   = req->rq_transno;
1865         rs->rs_export    = exp;
1866
1867         spin_lock(&exp->exp_uncommitted_replies_lock);
1868
1869         /* VBR: use exp_last_committed */
1870         if (rs->rs_transno > exp->exp_last_committed) {
1871                 /* not committed already */
1872                 list_add_tail (&rs->rs_obd_list,
1873                                &exp->exp_uncommitted_replies);
1874         }
1875
1876         spin_unlock (&exp->exp_uncommitted_replies_lock);
1877         spin_lock (&exp->exp_lock);
1878
1879         list_add_tail (&rs->rs_exp_list, &exp->exp_outstanding_replies);
1880
1881         spin_unlock(&exp->exp_lock);
1882
1883         netrc = target_send_reply_msg (req, rc, fail_id);
1884
1885         spin_lock(&svc->srv_lock);
1886
1887         svc->srv_n_difficult_replies++;
1888
1889         if (netrc != 0) {
1890                 /* error sending: reply is off the net.  Also we need +1
1891                  * reply ref until ptlrpc_server_handle_reply() is done
1892                  * with the reply state (if the send was successful, there
1893                  * would have been +1 ref for the net, which
1894                  * reply_out_callback leaves alone) */
1895                 rs->rs_on_net = 0;
1896                 ptlrpc_rs_addref(rs);
1897                 atomic_inc (&svc->srv_outstanding_replies);
1898         }
1899
1900         if (!rs->rs_on_net ||                   /* some notifier */
1901             list_empty(&rs->rs_exp_list) ||     /* completed already */
1902             list_empty(&rs->rs_obd_list)) {
1903                 list_add_tail (&rs->rs_list, &svc->srv_reply_queue);
1904                 cfs_waitq_signal (&svc->srv_waitq);
1905         } else {
1906                 list_add (&rs->rs_list, &svc->srv_active_replies);
1907                 rs->rs_scheduled = 0;           /* allow notifier to schedule */
1908         }
1909
1910         spin_unlock(&svc->srv_lock);
1911 }
1912
1913 int target_handle_ping(struct ptlrpc_request *req)
1914 {
1915         if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_LAST_REPLAY &&
1916             req->rq_export->exp_in_recovery) {
1917                 spin_lock(&req->rq_export->exp_lock);
1918                 req->rq_export->exp_in_recovery = 0;
1919                 spin_unlock(&req->rq_export->exp_lock);
1920         }
1921         obd_ping(req->rq_export);
1922         return lustre_pack_reply(req, 1, NULL, NULL);
1923 }
1924
1925 void target_committed_to_req(struct ptlrpc_request *req)
1926 {
1927         struct obd_export *exp = req->rq_export;
1928         if (!exp->exp_obd->obd_no_transno && req->rq_repmsg != NULL) {
1929                 lustre_msg_set_last_committed(req->rq_repmsg,
1930                                               exp->exp_last_committed);
1931         } else {
1932                 DEBUG_REQ(D_IOCTL, req, "not sending last_committed update (%d/"
1933                           "%d)", exp->exp_obd->obd_no_transno,
1934                           req->rq_repmsg == NULL);
1935         }
1936         CDEBUG(D_INFO, "last_committed x"LPU64", this req x"LPU64"\n",
1937                exp->exp_obd->obd_last_committed, req->rq_xid);
1938 }
1939
1940 EXPORT_SYMBOL(target_committed_to_req);
1941
1942 int target_handle_qc_callback(struct ptlrpc_request *req)
1943 {
1944         struct obd_quotactl *oqctl;
1945         struct client_obd *cli = &req->rq_export->exp_obd->u.cli;
1946
1947         oqctl = lustre_swab_reqbuf(req, REQ_REC_OFF, sizeof(*oqctl),
1948                                    lustre_swab_obd_quotactl);
1949         if (oqctl == NULL) {
1950                 CERROR("Can't unpack obd_quatactl\n");
1951                 RETURN(-EPROTO);
1952         }
1953
1954         cli->cl_qchk_stat = oqctl->qc_stat;
1955
1956         return 0;
1957 }
1958
1959 #ifdef HAVE_QUOTA_SUPPORT
1960 int target_handle_dqacq_callback(struct ptlrpc_request *req)
1961 {
1962 #ifdef __KERNEL__
1963         struct obd_device *obd = req->rq_export->exp_obd;
1964         struct obd_device *master_obd;
1965         struct lustre_quota_ctxt *qctxt;
1966         struct qunit_data *qdata = NULL;
1967         int rc = 0;
1968         int repsize[2] = { sizeof(struct ptlrpc_body), 0 };
1969         ENTRY;
1970
1971         if (OBD_FAIL_CHECK_ONCE(OBD_FAIL_MDS_DROP_QUOTA_REQ))
1972                 RETURN(rc);
1973
1974         repsize[1] = quota_get_qunit_data_size(req->rq_export->
1975                                                exp_connect_flags);
1976
1977         rc = lustre_pack_reply(req, 2, repsize, NULL);
1978         if (rc)
1979                 RETURN(rc);
1980
1981         LASSERT(req->rq_export);
1982
1983         /* there are three forms of qunit(historic causes), so we need to
1984          * adjust qunits from slaves to the same form here */
1985         OBD_ALLOC(qdata, sizeof(struct qunit_data));
1986         if (!qdata)
1987                 RETURN(-ENOMEM);
1988         rc = quota_get_qdata(req, qdata, QUOTA_REQUEST, QUOTA_EXPORT);
1989         if (rc < 0) {
1990                 CDEBUG(D_ERROR, "Can't unpack qunit_data(rc: %d)\n", rc);
1991                 GOTO(out, rc);
1992         }
1993
1994         /* we use the observer */
1995         if (!obd->obd_observer || !obd->obd_observer->obd_observer) {
1996                 CERROR("Can't find the observer, it is recovering\n");
1997                 req->rq_status = -EAGAIN;
1998                 GOTO(send_reply, rc = -EAGAIN);
1999         }
2000
2001         master_obd = obd->obd_observer->obd_observer;
2002         qctxt = &master_obd->u.obt.obt_qctxt;
2003
2004         if (!qctxt->lqc_setup) {
2005                 /* quota_type has not been processed yet, return EAGAIN
2006                  * until we know whether or not quotas are supposed to
2007                  * be enabled */
2008                 CDEBUG(D_QUOTA, "quota_type not processed yet, return "
2009                                 "-EAGAIN\n");
2010                 req->rq_status = -EAGAIN;
2011                 rc = ptlrpc_reply(req);
2012                 GOTO(out, rc);
2013         }
2014
2015         LASSERT(qctxt->lqc_handler);
2016         rc = qctxt->lqc_handler(master_obd, qdata,
2017                                 lustre_msg_get_opc(req->rq_reqmsg));
2018         if (rc && rc != -EDQUOT)
2019                 CDEBUG(rc == -EBUSY  ? D_QUOTA : D_ERROR,
2020                        "dqacq failed! (rc:%d)\n", rc);
2021         req->rq_status = rc;
2022
2023         /* there are three forms of qunit(historic causes), so we need to
2024          * adjust the same form to different forms slaves needed */
2025         rc = quota_copy_qdata(req, qdata, QUOTA_REPLY, QUOTA_EXPORT);
2026         if (rc < 0) {
2027                 CDEBUG(D_ERROR, "Can't pack qunit_data(rc: %d)\n", rc);
2028                 GOTO(out, rc);
2029         }
2030
2031         /* Block the quota req. b=14840 */
2032         OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_BLOCK_QUOTA_REQ, obd_timeout);
2033  send_reply:
2034         rc = ptlrpc_reply(req);
2035 out:
2036         OBD_FREE(qdata, sizeof(struct qunit_data));
2037         RETURN(rc);
2038 #else
2039         return 0;
2040 #endif /* !__KERNEL__ */
2041 }
2042 #endif /* HAVE_QUOTA_SUPPORT */
2043
2044 ldlm_mode_t lck_compat_array[] = {
2045         [LCK_EX] LCK_COMPAT_EX,
2046         [LCK_PW] LCK_COMPAT_PW,
2047         [LCK_PR] LCK_COMPAT_PR,
2048         [LCK_CW] LCK_COMPAT_CW,
2049         [LCK_CR] LCK_COMPAT_CR,
2050         [LCK_NL] LCK_COMPAT_NL,
2051         [LCK_GROUP] LCK_COMPAT_GROUP
2052 };