lustre/ldlm/ldlm_lib.c

   1 /* -*- mode: c; c-basic-offset: 8; indent-tabs-mode: nil; -*-
   2  * vim:expandtab:shiftwidth=8:tabstop=8:
   3  *
   4  *  Copyright (c) 2003 Cluster File Systems, Inc.
   5  *
   6  *   This file is part of the Lustre file system, http://www.lustre.org
   7  *   Lustre is a trademark of Cluster File Systems, Inc.
   8  *
   9  *   You may have signed or agreed to another license before downloading
  10  *   this software.  If so, you are bound by the terms and conditions
  11  *   of that agreement, and the following does not apply to you.  See the
  12  *   LICENSE file included with this distribution for more information.
  13  *
  14  *   If you did not agree to a different license, then this copy of Lustre
  15  *   is open source software; you can redistribute it and/or modify it
  16  *   under the terms of version 2 of the GNU General Public License as
  17  *   published by the Free Software Foundation.
  18  *
  19  *   In either case, Lustre is distributed in the hope that it will be
  20  *   useful, but WITHOUT ANY WARRANTY; without even the implied warranty
  21  *   of MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
  22  *   license text for more details.
  23  */
  24
  25 #ifndef EXPORT_SYMTAB
  26 # define EXPORT_SYMTAB
  27 #endif
  28 #define DEBUG_SUBSYSTEM S_LDLM
  29
  30 #ifdef __KERNEL__
  31 # include <libcfs/libcfs.h>
  32 #else
  33 # include <liblustre.h>
  34 #endif
  35 #include <obd.h>
  36 #include <lustre_mds.h>
  37 #include <lustre_dlm.h>
  38 #include <lustre_net.h>
  39
  40 /* @priority: if non-zero, move the selected to the list head
  41  * @create: if zero, only search in existed connections
  42  */
  43 static int import_set_conn(struct obd_import *imp, struct obd_uuid *uuid,
  44                            int priority, int create)
  45 {
  46         struct ptlrpc_connection *ptlrpc_conn;
  47         struct obd_import_conn *imp_conn = NULL, *item;
  48         int rc = 0;
  49         ENTRY;
  50
  51         if (!create && !priority) {
  52                 CDEBUG(D_HA, "Nothing to do\n");
  53                 RETURN(-EINVAL);
  54         }
  55
  56         ptlrpc_conn = ptlrpc_uuid_to_connection(uuid);
  57         if (!ptlrpc_conn) {
  58                 CDEBUG(D_HA, "can't find connection %s\n", uuid->uuid);
  59                 RETURN (-ENOENT);
  60         }
  61
  62         if (create) {
  63                 OBD_ALLOC(imp_conn, sizeof(*imp_conn));
  64                 if (!imp_conn) {
  65                         GOTO(out_put, rc = -ENOMEM);
  66                 }
  67         }
  68
  69         spin_lock(&imp->imp_lock);
  70         list_for_each_entry(item, &imp->imp_conn_list, oic_item) {
  71                 if (obd_uuid_equals(uuid, &item->oic_uuid)) {
  72                         if (priority) {
  73                                 list_del(&item->oic_item);
  74                                 list_add(&item->oic_item, &imp->imp_conn_list);
  75                                 item->oic_last_attempt = 0;
  76                         }
  77                         CDEBUG(D_HA, "imp %p@%s: found existing conn %s%s\n",
  78                                imp, imp->imp_obd->obd_name, uuid->uuid,
  79                                (priority ? ", moved to head" : ""));
  80                         spin_unlock(&imp->imp_lock);
  81                         GOTO(out_free, rc = 0);
  82                 }
  83         }
  84         /* not found */
  85         if (create) {
  86                 imp_conn->oic_conn = ptlrpc_conn;
  87                 imp_conn->oic_uuid = *uuid;
  88                 item->oic_last_attempt = 0;
  89                 if (priority)
  90                         list_add(&imp_conn->oic_item, &imp->imp_conn_list);
  91                 else
  92                         list_add_tail(&imp_conn->oic_item, &imp->imp_conn_list);
  93                 CDEBUG(D_HA, "imp %p@%s: add connection %s at %s\n",
  94                        imp, imp->imp_obd->obd_name, uuid->uuid,
  95                        (priority ? "head" : "tail"));
  96         } else {
  97                 spin_unlock(&imp->imp_lock);
  98                 GOTO(out_free, rc = -ENOENT);
  99
 100         }
 101
 102         spin_unlock(&imp->imp_lock);
 103         RETURN(0);
 104 out_free:
 105         if (imp_conn)
 106                 OBD_FREE(imp_conn, sizeof(*imp_conn));
 107 out_put:
 108         ptlrpc_put_connection(ptlrpc_conn);
 109         RETURN(rc);
 110 }
 111
 112 int import_set_conn_priority(struct obd_import *imp, struct obd_uuid *uuid)
 113 {
 114         return import_set_conn(imp, uuid, 1, 0);
 115 }
 116
 117 int client_import_add_conn(struct obd_import *imp, struct obd_uuid *uuid,
 118                            int priority)
 119 {
 120         return import_set_conn(imp, uuid, priority, 1);
 121 }
 122
 123 int client_import_del_conn(struct obd_import *imp, struct obd_uuid *uuid)
 124 {
 125         struct obd_import_conn *imp_conn;
 126         struct obd_export *dlmexp;
 127         int rc = -ENOENT;
 128         ENTRY;
 129
 130         spin_lock(&imp->imp_lock);
 131         if (list_empty(&imp->imp_conn_list)) {
 132                 LASSERT(!imp->imp_connection);
 133                 GOTO(out, rc);
 134         }
 135
 136         list_for_each_entry(imp_conn, &imp->imp_conn_list, oic_item) {
 137                 if (!obd_uuid_equals(uuid, &imp_conn->oic_uuid))
 138                         continue;
 139                 LASSERT(imp_conn->oic_conn);
 140
 141                 /* is current conn? */
 142                 if (imp_conn == imp->imp_conn_current) {
 143                         LASSERT(imp_conn->oic_conn == imp->imp_connection);
 144
 145                         if (imp->imp_state != LUSTRE_IMP_CLOSED &&
 146                             imp->imp_state != LUSTRE_IMP_DISCON) {
 147                                 CERROR("can't remove current connection\n");
 148                                 GOTO(out, rc = -EBUSY);
 149                         }
 150
 151                         ptlrpc_put_connection(imp->imp_connection);
 152                         imp->imp_connection = NULL;
 153
 154                         dlmexp = class_conn2export(&imp->imp_dlm_handle);
 155                         if (dlmexp && dlmexp->exp_connection) {
 156                                 LASSERT(dlmexp->exp_connection ==
 157                                         imp_conn->oic_conn);
 158                                 ptlrpc_put_connection(dlmexp->exp_connection);
 159                                 dlmexp->exp_connection = NULL;
 160                         }
 161                 }
 162
 163                 list_del(&imp_conn->oic_item);
 164                 ptlrpc_put_connection(imp_conn->oic_conn);
 165                 OBD_FREE(imp_conn, sizeof(*imp_conn));
 166                 CDEBUG(D_HA, "imp %p@%s: remove connection %s\n",
 167                        imp, imp->imp_obd->obd_name, uuid->uuid);
 168                 rc = 0;
 169                 break;
 170         }
 171 out:
 172         spin_unlock(&imp->imp_lock);
 173         if (rc == -ENOENT)
 174                 CERROR("connection %s not found\n", uuid->uuid);
 175         RETURN(rc);
 176 }
 177
 178 /* configure an RPC client OBD device
 179  *
 180  * lcfg parameters:
 181  * 1 - client UUID
 182  * 2 - server UUID
 183  * 3 - inactive-on-startup
 184  */
 185 int client_obd_setup(struct obd_device *obddev, obd_count len, void *buf)
 186 {
 187         struct lustre_cfg* lcfg = buf;
 188         struct client_obd *cli = &obddev->u.cli;
 189         struct obd_import *imp;
 190         struct obd_uuid server_uuid;
 191         int rq_portal, rp_portal, connect_op;
 192         char *name = obddev->obd_type->typ_name;
 193         int rc;
 194         ENTRY;
 195
 196         /* In a more perfect world, we would hang a ptlrpc_client off of
 197          * obd_type and just use the values from there. */
 198         if (!strcmp(name, LUSTRE_OSC_NAME)) {
 199 #ifdef __KERNEL__
 200                 /* Can be removed in Lustre 1.8, for compatibility only */
 201                 rq_portal = OST_IO_PORTAL;
 202 #else
 203                 rq_portal = OST_REQUEST_PORTAL;
 204 #endif
 205                 rp_portal = OSC_REPLY_PORTAL;
 206                 connect_op = OST_CONNECT;
 207         } else if (!strcmp(name, LUSTRE_MDC_NAME)) {
 208                 rq_portal = MDS_REQUEST_PORTAL;
 209                 rp_portal = MDC_REPLY_PORTAL;
 210                 connect_op = MDS_CONNECT;
 211         } else if (!strcmp(name, LUSTRE_MGC_NAME)) {
 212                 rq_portal = MGS_REQUEST_PORTAL;
 213                 rp_portal = MGC_REPLY_PORTAL;
 214                 connect_op = MGS_CONNECT;
 215         } else {
 216                 CERROR("unknown client OBD type \"%s\", can't setup\n",
 217                        name);
 218                 RETURN(-EINVAL);
 219         }
 220
 221         if (LUSTRE_CFG_BUFLEN(lcfg, 1) < 1) {
 222                 CERROR("requires a TARGET UUID\n");
 223                 RETURN(-EINVAL);
 224         }
 225
 226         if (LUSTRE_CFG_BUFLEN(lcfg, 1) > 37) {
 227                 CERROR("client UUID must be less than 38 characters\n");
 228                 RETURN(-EINVAL);
 229         }
 230
 231         if (LUSTRE_CFG_BUFLEN(lcfg, 2) < 1) {
 232                 CERROR("setup requires a SERVER UUID\n");
 233                 RETURN(-EINVAL);
 234         }
 235
 236         if (LUSTRE_CFG_BUFLEN(lcfg, 2) > 37) {
 237                 CERROR("target UUID must be less than 38 characters\n");
 238                 RETURN(-EINVAL);
 239         }
 240
 241         sema_init(&cli->cl_sem, 1);
 242         sema_init(&cli->cl_mgc_sem, 1);
 243         cli->cl_conn_count = 0;
 244         memcpy(server_uuid.uuid, lustre_cfg_buf(lcfg, 2),
 245                min_t(unsigned int, LUSTRE_CFG_BUFLEN(lcfg, 2),
 246                      sizeof(server_uuid)));
 247
 248         cli->cl_dirty = 0;
 249         cli->cl_avail_grant = 0;
 250         /* FIXME: should limit this for the sum of all cl_dirty_max */
 251         cli->cl_dirty_max = OSC_MAX_DIRTY_DEFAULT * 1024 * 1024;
 252         if (cli->cl_dirty_max >> CFS_PAGE_SHIFT > num_physpages / 8)
 253                 cli->cl_dirty_max = num_physpages << (CFS_PAGE_SHIFT - 3);
 254         CFS_INIT_LIST_HEAD(&cli->cl_cache_waiters);
 255         CFS_INIT_LIST_HEAD(&cli->cl_loi_ready_list);
 256         CFS_INIT_LIST_HEAD(&cli->cl_loi_write_list);
 257         CFS_INIT_LIST_HEAD(&cli->cl_loi_read_list);
 258         client_obd_list_lock_init(&cli->cl_loi_list_lock);
 259         cli->cl_r_in_flight = 0;
 260         cli->cl_w_in_flight = 0;
 261         spin_lock_init(&cli->cl_read_rpc_hist.oh_lock);
 262         spin_lock_init(&cli->cl_write_rpc_hist.oh_lock);
 263         spin_lock_init(&cli->cl_read_page_hist.oh_lock);
 264         spin_lock_init(&cli->cl_write_page_hist.oh_lock);
 265         spin_lock_init(&cli->cl_read_offset_hist.oh_lock);
 266         spin_lock_init(&cli->cl_write_offset_hist.oh_lock);
 267
 268         /* This value may be changed at connect time in
 269            ptlrpc_connect_interpret. */
 270         cli->cl_max_pages_per_rpc = min((int)PTLRPC_MAX_BRW_PAGES,
 271                                         (int)(1024 * 1024 >> CFS_PAGE_SHIFT));
 272
 273         if (!strcmp(name, LUSTRE_MDC_NAME)) {
 274                 cli->cl_max_rpcs_in_flight = MDC_MAX_RIF_DEFAULT;
 275         } else if (num_physpages >> (20 - CFS_PAGE_SHIFT) <= 128 /* MB */) {
 276                 cli->cl_max_rpcs_in_flight = 2;
 277         } else if (num_physpages >> (20 - CFS_PAGE_SHIFT) <= 256 /* MB */) {
 278                 cli->cl_max_rpcs_in_flight = 3;
 279         } else if (num_physpages >> (20 - CFS_PAGE_SHIFT) <= 512 /* MB */) {
 280                 cli->cl_max_rpcs_in_flight = 4;
 281         } else {
 282                 cli->cl_max_rpcs_in_flight = OSC_MAX_RIF_DEFAULT;
 283         }
 284         rc = ldlm_get_ref();
 285         if (rc) {
 286                 CERROR("ldlm_get_ref failed: %d\n", rc);
 287                 GOTO(err, rc);
 288         }
 289
 290         ptlrpc_init_client(rq_portal, rp_portal, name,
 291                            &obddev->obd_ldlm_client);
 292
 293         imp = class_new_import(obddev);
 294         if (imp == NULL)
 295                 GOTO(err_ldlm, rc = -ENOENT);
 296         imp->imp_client = &obddev->obd_ldlm_client;
 297         imp->imp_connect_op = connect_op;
 298         imp->imp_initial_recov = 1;
 299         imp->imp_initial_recov_bk = 0;
 300         CFS_INIT_LIST_HEAD(&imp->imp_pinger_chain);
 301         memcpy(cli->cl_target_uuid.uuid, lustre_cfg_buf(lcfg, 1),
 302                LUSTRE_CFG_BUFLEN(lcfg, 1));
 303         class_import_put(imp);
 304
 305         rc = client_import_add_conn(imp, &server_uuid, 1);
 306         if (rc) {
 307                 CERROR("can't add initial connection\n");
 308                 GOTO(err_import, rc);
 309         }
 310
 311         cli->cl_import = imp;
 312         /* cli->cl_max_mds_{easize,cookiesize} updated by mdc_init_ea_size() */
 313         cli->cl_max_mds_easize = sizeof(struct lov_mds_md);
 314         cli->cl_max_mds_cookiesize = sizeof(struct llog_cookie);
 315
 316         if (LUSTRE_CFG_BUFLEN(lcfg, 3) > 0) {
 317                 if (!strcmp(lustre_cfg_string(lcfg, 3), "inactive")) {
 318                         CDEBUG(D_HA, "marking %s %s->%s as inactive\n",
 319                                name, obddev->obd_name,
 320                                cli->cl_target_uuid.uuid);
 321                         imp->imp_invalid = 1;
 322                 }
 323         }
 324
 325         cli->cl_qchk_stat = CL_NOT_QUOTACHECKED;
 326
 327         RETURN(rc);
 328
 329 err_import:
 330         class_destroy_import(imp);
 331 err_ldlm:
 332         ldlm_put_ref(0);
 333 err:
 334         RETURN(rc);
 335
 336 }
 337
 338 int client_obd_cleanup(struct obd_device *obddev)
 339 {
 340         ENTRY;
 341         ldlm_put_ref(obddev->obd_force);
 342
 343         RETURN(0);
 344 }
 345
 346 /* ->o_connect() method for client side (OSC and MDC and MGC) */
 347 int client_connect_import(struct lustre_handle *dlm_handle,
 348                           struct obd_device *obd, struct obd_uuid *cluuid,
 349                           struct obd_connect_data *data)
 350 {
 351         struct client_obd *cli = &obd->u.cli;
 352         struct obd_import *imp = cli->cl_import;
 353         struct obd_export *exp;
 354         struct obd_connect_data *ocd;
 355         int rc;
 356         ENTRY;
 357
 358         mutex_down(&cli->cl_sem);
 359         rc = class_connect(dlm_handle, obd, cluuid);
 360         if (rc)
 361                 GOTO(out_sem, rc);
 362
 363         cli->cl_conn_count++;
 364         if (cli->cl_conn_count > 1)
 365                 GOTO(out_sem, rc);
 366         exp = class_conn2export(dlm_handle);
 367
 368         if (obd->obd_namespace != NULL)
 369                 CERROR("already have namespace!\n");
 370         obd->obd_namespace = ldlm_namespace_new(obd->obd_name,
 371                                                 LDLM_NAMESPACE_CLIENT);
 372         if (obd->obd_namespace == NULL)
 373                 GOTO(out_disco, rc = -ENOMEM);
 374
 375         imp->imp_dlm_handle = *dlm_handle;
 376         rc = ptlrpc_init_import(imp);
 377         if (rc != 0)
 378                 GOTO(out_ldlm, rc);
 379
 380         ocd = &imp->imp_connect_data;
 381         if (data) {
 382                 *ocd = *data;
 383                 imp->imp_connect_flags_orig = data->ocd_connect_flags;
 384         }
 385
 386         rc = ptlrpc_connect_import(imp, NULL);
 387         if (rc != 0) {
 388                 LASSERT (imp->imp_state == LUSTRE_IMP_DISCON);
 389                 GOTO(out_ldlm, rc);
 390         }
 391         LASSERT(exp->exp_connection);
 392
 393         if (data) {
 394                 LASSERT((ocd->ocd_connect_flags & data->ocd_connect_flags) ==
 395                         ocd->ocd_connect_flags);
 396                 data->ocd_connect_flags = ocd->ocd_connect_flags;
 397         }
 398
 399         ptlrpc_pinger_add_import(imp);
 400         EXIT;
 401
 402         if (rc) {
 403 out_ldlm:
 404                 ldlm_namespace_free(obd->obd_namespace, 0);
 405                 obd->obd_namespace = NULL;
 406 out_disco:
 407                 cli->cl_conn_count--;
 408                 class_disconnect(exp);
 409         } else {
 410                 class_export_put(exp);
 411         }
 412 out_sem:
 413         mutex_up(&cli->cl_sem);
 414         return rc;
 415 }
 416
 417 int client_disconnect_export(struct obd_export *exp)
 418 {
 419         struct obd_device *obd = class_exp2obd(exp);
 420         struct client_obd *cli;
 421         struct obd_import *imp;
 422         int rc = 0, err;
 423         ENTRY;
 424
 425         if (!obd) {
 426                 CERROR("invalid export for disconnect: exp %p cookie "LPX64"\n",
 427                        exp, exp ? exp->exp_handle.h_cookie : -1);
 428                 RETURN(-EINVAL);
 429         }
 430
 431         cli = &obd->u.cli;
 432         imp = cli->cl_import;
 433
 434         mutex_down(&cli->cl_sem);
 435         if (!cli->cl_conn_count) {
 436                 CERROR("disconnecting disconnected device (%s)\n",
 437                        obd->obd_name);
 438                 GOTO(out_sem, rc = -EINVAL);
 439         }
 440
 441         cli->cl_conn_count--;
 442         if (cli->cl_conn_count)
 443                 GOTO(out_no_disconnect, rc = 0);
 444
 445         /* Mark import deactivated now, so we don't try to reconnect if any
 446          * of the cleanup RPCs fails (e.g. ldlm cancel, etc).  We don't
 447          * fully deactivate the import, or that would drop all requests. */
 448         imp->imp_deactive = 1;
 449
 450         /* Some non-replayable imports (MDS's OSCs) are pinged, so just
 451          * delete it regardless.  (It's safe to delete an import that was
 452          * never added.) */
 453         (void)ptlrpc_pinger_del_import(imp);
 454
 455         if (obd->obd_namespace != NULL) {
 456                 /* obd_no_recov == local only */
 457                 ldlm_cli_cancel_unused(obd->obd_namespace, NULL,
 458                                        obd->obd_no_recov ? LDLM_FL_LOCAL_ONLY:0,
 459                                        NULL);
 460                 ldlm_namespace_free(obd->obd_namespace, obd->obd_no_recov);
 461                 obd->obd_namespace = NULL;
 462         }
 463
 464         /* Yeah, obd_no_recov also (mainly) means "forced shutdown". */
 465         if (!obd->obd_no_recov)
 466                 rc = ptlrpc_disconnect_import(imp, 0);
 467
 468         ptlrpc_invalidate_import(imp);
 469         ptlrpc_free_rq_pool(imp->imp_rq_pool);
 470         class_destroy_import(imp);
 471         cli->cl_import = NULL;
 472
 473         EXIT;
 474  out_no_disconnect:
 475         err = class_disconnect(exp);
 476         if (!rc && err)
 477                 rc = err;
 478  out_sem:
 479         mutex_up(&cli->cl_sem);
 480         RETURN(rc);
 481 }
 482
 483 /* --------------------------------------------------------------------------
 484  * from old lib/target.c
 485  * -------------------------------------------------------------------------- */
 486
 487 int target_handle_reconnect(struct lustre_handle *conn, struct obd_export *exp,
 488                             struct obd_uuid *cluuid)
 489 {
 490         ENTRY;
 491         if (exp->exp_connection && exp->exp_imp_reverse) {
 492                 struct lustre_handle *hdl;
 493                 hdl = &exp->exp_imp_reverse->imp_remote_handle;
 494                 /* Might be a re-connect after a partition. */
 495                 if (!memcmp(&conn->cookie, &hdl->cookie, sizeof conn->cookie)) {
 496                         CWARN("%s: %s reconnecting\n", exp->exp_obd->obd_name,
 497                               cluuid->uuid);
 498                         conn->cookie = exp->exp_handle.h_cookie;
 499                         /* target_handle_connect() treats EALREADY and
 500                          * -EALREADY differently.  EALREADY means we are
 501                          * doing a valid reconnect from the same client. */
 502                         RETURN(EALREADY);
 503                 } else {
 504                         CERROR("%s reconnecting from %s, "
 505                                "handle mismatch (ours "LPX64", theirs "
 506                                LPX64")\n", cluuid->uuid,
 507                                exp->exp_connection->c_remote_uuid.uuid,
 508                                hdl->cookie, conn->cookie);
 509                         memset(conn, 0, sizeof *conn);
 510                         /* target_handle_connect() treats EALREADY and
 511                          * -EALREADY differently.  -EALREADY is an error
 512                          * (same UUID, different handle). */
 513                         RETURN(-EALREADY);
 514                 }
 515         }
 516
 517         conn->cookie = exp->exp_handle.h_cookie;
 518         CDEBUG(D_HA, "connect export for UUID '%s' at %p, cookie "LPX64"\n",
 519                cluuid->uuid, exp, conn->cookie);
 520         RETURN(0);
 521 }
 522
 523 void target_client_add_cb(struct obd_device *obd, __u64 transno, void *cb_data,
 524                           int error)
 525 {
 526         struct obd_export *exp = cb_data;
 527
 528         CDEBUG(D_HA, "%s: committing for initial connect of %s\n",
 529                obd->obd_name, exp->exp_client_uuid.uuid);
 530         exp->exp_need_sync = 0;
 531 }
 532 EXPORT_SYMBOL(target_client_add_cb);
 533
 534 int target_handle_connect(struct ptlrpc_request *req, svc_handler_t handler)
 535 {
 536         struct obd_device *target, *targref = NULL;
 537         struct obd_export *export = NULL;
 538         struct obd_import *revimp;
 539         struct lustre_handle conn;
 540         struct obd_uuid tgtuuid;
 541         struct obd_uuid cluuid;
 542         struct obd_uuid remote_uuid;
 543         char *str, *tmp;
 544         int rc = 0, abort_recovery;
 545         struct obd_connect_data *data;
 546         int size[2] = { sizeof(struct ptlrpc_body), sizeof(*data) };
 547         ENTRY;
 548
 549         OBD_RACE(OBD_FAIL_TGT_CONN_RACE);
 550
 551         LASSERT_REQSWAB(req, REQ_REC_OFF);
 552         str = lustre_msg_string(req->rq_reqmsg, REQ_REC_OFF, sizeof(tgtuuid)-1);
 553         if (str == NULL) {
 554                 DEBUG_REQ(D_ERROR, req, "bad target UUID for connect");
 555                 GOTO(out, rc = -EINVAL);
 556         }
 557
 558         obd_str2uuid (&tgtuuid, str);
 559         target = class_uuid2obd(&tgtuuid);
 560         /* COMPAT_146 */
 561         /* old (pre 1.6) lustre_process_log tries to connect to mdsname
 562            (eg. mdsA) instead of uuid. */
 563         if (!target) {
 564                 snprintf((char *)tgtuuid.uuid, sizeof(tgtuuid), "%s_UUID", str);
 565                 target = class_uuid2obd(&tgtuuid);
 566         }
 567         if (!target)
 568                 target = class_name2obd(str);
 569         /* end COMPAT_146 */
 570
 571         if (!target || target->obd_stopping || !target->obd_set_up) {
 572                 DEBUG_REQ(D_ERROR, req, "UUID '%s' is not available "
 573                           " for connect (%s)", str,
 574                           !target ? "no target" :
 575                           (target->obd_stopping ? "stopping" : "not set up"));
 576                 GOTO(out, rc = -ENODEV);
 577         }
 578
 579         /* Make sure the target isn't cleaned up while we're here. Yes,
 580            there's still a race between the above check and our incref here.
 581            Really, class_uuid2obd should take the ref. */
 582         targref = class_incref(target);
 583
 584         LASSERT_REQSWAB(req, REQ_REC_OFF + 1);
 585         str = lustre_msg_string(req->rq_reqmsg, REQ_REC_OFF + 1,
 586                                 sizeof(cluuid) - 1);
 587         if (str == NULL) {
 588                 DEBUG_REQ(D_ERROR, req, "bad client UUID for connect");
 589                 GOTO(out, rc = -EINVAL);
 590         }
 591
 592         obd_str2uuid (&cluuid, str);
 593
 594         /* XXX extract a nettype and format accordingly */
 595         switch (sizeof(lnet_nid_t)) {
 596                 /* NB the casts only avoid compiler warnings */
 597         case 8:
 598                 snprintf(remote_uuid.uuid, sizeof remote_uuid,
 599                          "NET_"LPX64"_UUID", (__u64)req->rq_peer.nid);
 600                 break;
 601         case 4:
 602                 snprintf(remote_uuid.uuid, sizeof remote_uuid,
 603                          "NET_%x_UUID", (__u32)req->rq_peer.nid);
 604                 break;
 605         default:
 606                 LBUG();
 607         }
 608
 609         spin_lock_bh(&target->obd_processing_task_lock);
 610         abort_recovery = target->obd_abort_recovery;
 611         spin_unlock_bh(&target->obd_processing_task_lock);
 612         if (abort_recovery)
 613                 target_abort_recovery(target);
 614
 615         tmp = lustre_msg_buf(req->rq_reqmsg, REQ_REC_OFF + 2, sizeof conn);
 616         if (tmp == NULL)
 617                 GOTO(out, rc = -EPROTO);
 618
 619         memcpy(&conn, tmp, sizeof conn);
 620
 621         data = lustre_swab_reqbuf(req, REQ_REC_OFF + 3, sizeof(*data),
 622                                   lustre_swab_connect);
 623         rc = lustre_pack_reply(req, 2, size, NULL);
 624         if (rc)
 625                 GOTO(out, rc);
 626
 627         if (lustre_msg_get_op_flags(req->rq_reqmsg) & MSG_CONNECT_LIBCLIENT) {
 628                 if (!data) {
 629                         DEBUG_REQ(D_WARNING, req, "Refusing old (unversioned) "
 630                                   "libclient connection attempt\n");
 631                         GOTO(out, rc = -EPROTO);
 632                 } else if (data->ocd_version < LUSTRE_VERSION_CODE -
 633                                                LUSTRE_VERSION_ALLOWED_OFFSET ||
 634                            data->ocd_version > LUSTRE_VERSION_CODE +
 635                                                LUSTRE_VERSION_ALLOWED_OFFSET) {
 636                         DEBUG_REQ(D_WARNING, req, "Refusing %s (%d.%d.%d.%d) "
 637                                   "libclient connection attempt\n",
 638                                   data->ocd_version < LUSTRE_VERSION_CODE ?
 639                                   "old" : "new",
 640                                   OBD_OCD_VERSION_MAJOR(data->ocd_version),
 641                                   OBD_OCD_VERSION_MINOR(data->ocd_version),
 642                                   OBD_OCD_VERSION_PATCH(data->ocd_version),
 643                                   OBD_OCD_VERSION_FIX(data->ocd_version));
 644                         data = lustre_msg_buf(req->rq_repmsg, REPLY_REC_OFF,
 645                                               offsetof(typeof(*data),
 646                                                        ocd_version) +
 647                                               sizeof(data->ocd_version));
 648                         if (data) {
 649                                 data->ocd_connect_flags = OBD_CONNECT_VERSION;
 650                                 data->ocd_version = LUSTRE_VERSION_CODE;
 651                         }
 652                         GOTO(out, rc = -EPROTO);
 653                 }
 654         }
 655
 656         /* lctl gets a backstage, all-access pass. */
 657         if (obd_uuid_equals(&cluuid, &target->obd_uuid))
 658                 goto dont_check_exports;
 659
 660         spin_lock(&target->obd_dev_lock);
 661         export = lustre_hash_get_object_by_key(target->obd_uuid_hash_body, &cluuid);
 662
 663         if (export != NULL && export->exp_connecting) { /* bug 9635, et. al. */
 664                 CWARN("%s: exp %p already connecting\n",
 665                       export->exp_obd->obd_name, export);
 666                 class_export_put(export);
 667                 export = NULL;
 668                 rc = -EALREADY;
 669         } else if (export != NULL) {
 670                 /* make darn sure this is coming from the same peer
 671                  * if the UUIDs matched */
 672               if ((export->exp_connection != NULL) &&
 673                   (strcmp(libcfs_nid2str(req->rq_peer.nid),
 674                           libcfs_nid2str(export->exp_connection->c_peer.nid)))) {
 675                           CWARN("%s: cookie %s seen on new NID %s when "
 676                                   "existing NID %s is already connected\n",
 677                           target->obd_name, cluuid.uuid,
 678                           libcfs_nid2str(req->rq_peer.nid),
 679                           libcfs_nid2str(export->exp_connection->c_peer.nid));
 680                           class_export_put(export);
 681                           export = NULL;
 682                           rc = -EALREADY;
 683                 } else {
 684                                 export->exp_connecting = 1;
 685                                 class_export_put(export);
 686                                 spin_unlock(&target->obd_dev_lock);
 687                                 LASSERT(export->exp_obd == target);
 688                                 rc = target_handle_reconnect(&conn, export, &cluuid);
 689                 }
 690         }
 691
 692         /* If we found an export, we already unlocked. */
 693         if (!export) {
 694                 spin_unlock(&target->obd_dev_lock);
 695                 OBD_FAIL_TIMEOUT(OBD_FAIL_TGT_DELAY_CONNECT, 2 * obd_timeout);
 696         } else if (req->rq_export == NULL &&
 697                    atomic_read(&export->exp_rpc_count) > 0) {
 698                 CWARN("%s: refuse connection from %s/%s to 0x%p/%d\n",
 699                       target->obd_name, cluuid.uuid,
 700                       libcfs_nid2str(req->rq_peer.nid),
 701                       export, atomic_read(&export->exp_refcount));
 702                 GOTO(out, rc = -EBUSY);
 703         } else if (req->rq_export != NULL &&
 704                    atomic_read(&export->exp_rpc_count) > 1) {
 705                 CWARN("%s: refuse reconnection from %s@%s to 0x%p/%d\n",
 706                       target->obd_name, cluuid.uuid,
 707                       libcfs_nid2str(req->rq_peer.nid),
 708                       export, atomic_read(&export->exp_rpc_count));
 709                 GOTO(out, rc = -EBUSY);
 710         } else if (lustre_msg_get_conn_cnt(req->rq_reqmsg) == 1) {
 711                 CERROR("%s: NID %s (%s) reconnected with 1 conn_cnt; "
 712                        "cookies not random?\n", target->obd_name,
 713                        libcfs_nid2str(req->rq_peer.nid), cluuid.uuid);
 714                 GOTO(out, rc = -EALREADY);
 715         } else {
 716                 OBD_FAIL_TIMEOUT(OBD_FAIL_TGT_DELAY_RECONNECT, 2 * obd_timeout);
 717         }
 718
 719         /* We want to handle EALREADY but *not* -EALREADY from
 720          * target_handle_reconnect(), return reconnection state in a flag */
 721         if (rc == EALREADY) {
 722                 lustre_msg_add_op_flags(req->rq_repmsg, MSG_CONNECT_RECONNECT);
 723                 rc = 0;
 724         } else if (rc) {
 725                 GOTO(out, rc);
 726         }
 727
 728         /* Tell the client if we're in recovery. */
 729         /* If this is the first client, start the recovery timer */
 730         if (target->obd_recovering) {
 731                 lustre_msg_add_op_flags(req->rq_repmsg, MSG_CONNECT_RECOVERING);
 732                 target_start_recovery_timer(target, handler);
 733         }
 734
 735         /* Tell the client if we support replayable requests */
 736         if (target->obd_replayable)
 737                 lustre_msg_add_op_flags(req->rq_repmsg, MSG_CONNECT_REPLAYABLE);
 738
 739         if (export == NULL) {
 740                 if (target->obd_recovering) {
 741                         CERROR("%s: denying connection for new client %s (%s): "
 742                                "%d clients in recovery for %lds\n",
 743                                target->obd_name,
 744                                libcfs_nid2str(req->rq_peer.nid), cluuid.uuid,
 745                                target->obd_recoverable_clients,
 746                                cfs_duration_sec(cfs_time_sub(cfs_timer_deadline(&target->obd_recovery_timer),
 747                                                              cfs_time_current())));
 748                         rc = -EBUSY;
 749                 } else {
 750  dont_check_exports:
 751                         rc = obd_connect(&conn, target, &cluuid, data);
 752                 }
 753         } else {
 754                 rc = obd_reconnect(export, target, &cluuid, data);
 755         }
 756
 757         if (rc)
 758                 GOTO(out, rc);
 759
 760         /* Return only the parts of obd_connect_data that we understand, so the
 761          * client knows that we don't understand the rest. */
 762         if (data)
 763                 memcpy(lustre_msg_buf(req->rq_repmsg, REPLY_REC_OFF,
 764                                       sizeof(*data)),
 765                        data, sizeof(*data));
 766
 767         /* If all else goes well, this is our RPC return code. */
 768         req->rq_status = 0;
 769
 770         lustre_msg_set_handle(req->rq_repmsg, &conn);
 771
 772         /* ownership of this export ref transfers to the request AFTER we
 773          * drop any previous reference the request had, but we don't want
 774          * that to go to zero before we get our new export reference. */
 775         export = class_conn2export(&conn);
 776         if (!export) {
 777                 DEBUG_REQ(D_ERROR, req, "Missing export!\n");
 778                 GOTO(out, rc = -ENODEV);
 779         }
 780
 781         /* If the client and the server are the same node, we will already
 782          * have an export that really points to the client's DLM export,
 783          * because we have a shared handles table.
 784          *
 785          * XXX this will go away when shaver stops sending the "connect" handle
 786          * in the real "remote handle" field of the request --phik 24 Apr 2003
 787          */
 788         if (req->rq_export != NULL)
 789                 class_export_put(req->rq_export);
 790
 791         req->rq_export = export;
 792
 793         spin_lock(&export->exp_lock);
 794         if (export->exp_conn_cnt >= lustre_msg_get_conn_cnt(req->rq_reqmsg)) {
 795                 CERROR("%s: %s already connected at higher conn_cnt: %d > %d\n",
 796                        cluuid.uuid, libcfs_nid2str(req->rq_peer.nid),
 797                        export->exp_conn_cnt,
 798                        lustre_msg_get_conn_cnt(req->rq_reqmsg));
 799
 800                 spin_unlock(&export->exp_lock);
 801                 GOTO(out, rc = -EALREADY);
 802         }
 803         export->exp_conn_cnt = lustre_msg_get_conn_cnt(req->rq_reqmsg);
 804         spin_unlock(&export->exp_lock);
 805
 806         /* request from liblustre?  Don't evict it for not pinging. */
 807         if (lustre_msg_get_op_flags(req->rq_reqmsg) & MSG_CONNECT_LIBCLIENT) {
 808                 export->exp_libclient = 1;
 809                 spin_lock(&target->obd_dev_lock);
 810                 list_del_init(&export->exp_obd_chain_timed);
 811                 spin_unlock(&target->obd_dev_lock);
 812         }
 813
 814         if (export->exp_connection != NULL)
 815                 ptlrpc_put_connection(export->exp_connection);
 816         export->exp_connection = ptlrpc_get_connection(req->rq_peer,
 817                                                        req->rq_self,
 818                                                        &remote_uuid);
 819
 820         spin_lock(&target->obd_dev_lock);
 821         lustre_hash_additem(export->exp_obd->obd_nid_hash_body,
 822                             &export->exp_connection->c_peer.nid,
 823                             &export->exp_nid_hash);
 824         spin_unlock(&target->obd_dev_lock);
 825
 826         if (lustre_msg_get_op_flags(req->rq_repmsg) & MSG_CONNECT_RECONNECT)
 827                 GOTO(out, rc = 0);
 828
 829         if (target->obd_recovering)
 830                 target->obd_connected_clients++;
 831
 832         memcpy(&conn,
 833                lustre_msg_buf(req->rq_reqmsg, REQ_REC_OFF + 2, sizeof conn),
 834                sizeof conn);
 835
 836         if (export->exp_imp_reverse != NULL)
 837                 class_destroy_import(export->exp_imp_reverse);
 838         revimp = export->exp_imp_reverse = class_new_import(target);
 839         revimp->imp_connection = ptlrpc_connection_addref(export->exp_connection);
 840         revimp->imp_client = &export->exp_obd->obd_ldlm_client;
 841         revimp->imp_remote_handle = conn;
 842         revimp->imp_dlm_fake = 1;
 843         revimp->imp_state = LUSTRE_IMP_FULL;
 844
 845         if (lustre_msg_get_op_flags(req->rq_reqmsg) & MSG_CONNECT_NEXT_VER) {
 846                 revimp->imp_msg_magic = LUSTRE_MSG_MAGIC_V2;
 847                 lustre_msg_add_op_flags(req->rq_repmsg, MSG_CONNECT_NEXT_VER);
 848         }
 849
 850         class_import_put(revimp);
 851 out:
 852         if (export)
 853                 export->exp_connecting = 0;
 854         if (targref)
 855                 class_decref(targref);
 856         if (rc)
 857                 req->rq_status = rc;
 858         RETURN(rc);
 859 }
 860
 861 int target_handle_disconnect(struct ptlrpc_request *req)
 862 {
 863         int rc;
 864         ENTRY;
 865
 866         rc = lustre_pack_reply(req, 1, NULL, NULL);
 867         if (rc)
 868                 RETURN(rc);
 869
 870         /* keep the rq_export around so we can send the reply */
 871         req->rq_status = obd_disconnect(class_export_get(req->rq_export));
 872         RETURN(0);
 873 }
 874
 875 void target_destroy_export(struct obd_export *exp)
 876 {
 877         /* exports created from last_rcvd data, and "fake"
 878            exports created by lctl don't have an import */
 879         if (exp->exp_imp_reverse != NULL)
 880                 class_destroy_import(exp->exp_imp_reverse);
 881
 882         /* We cancel locks at disconnect time, but this will catch any locks
 883          * granted in a race with recovery-induced disconnect. */
 884         if (exp->exp_obd->obd_namespace != NULL)
 885                 ldlm_cancel_locks_for_export(exp);
 886 }
 887
 888 /*
 889  * Recovery functions
 890  */
 891
 892
 893 static void target_release_saved_req(struct ptlrpc_request *req)
 894 {
 895         if (req->rq_reply_state != NULL) {
 896                 ptlrpc_rs_decref(req->rq_reply_state);
 897                 /* req->rq_reply_state = NULL; */
 898         }
 899
 900         class_export_put(req->rq_export);
 901         OBD_FREE(req->rq_reqmsg, req->rq_reqlen);
 902         OBD_FREE(req, sizeof *req);
 903 }
 904
 905 static void target_finish_recovery(struct obd_device *obd)
 906 {
 907         struct list_head *tmp, *n;
 908
 909         LCONSOLE_INFO("%s: sending delayed replies to recovered clients\n",
 910                       obd->obd_name);
 911
 912         ldlm_reprocess_all_ns(obd->obd_namespace);
 913
 914         /* when recovery finished, cleanup orphans on mds and ost */
 915         if (OBT(obd) && OBP(obd, postrecov)) {
 916                 int rc = OBP(obd, postrecov)(obd);
 917                 LCONSOLE_WARN("%s: recovery %s: rc %d\n", obd->obd_name,
 918                               rc < 0 ? "failed" : "complete", rc);
 919         }
 920
 921         list_for_each_safe(tmp, n, &obd->obd_delayed_reply_queue) {
 922                 struct ptlrpc_request *req;
 923                 req = list_entry(tmp, struct ptlrpc_request, rq_list);
 924                 list_del(&req->rq_list);
 925                 DEBUG_REQ(D_HA, req, "delayed:");
 926                 ptlrpc_reply(req);
 927                 target_release_saved_req(req);
 928         }
 929         obd->obd_recovery_end = CURRENT_SECONDS;
 930 }
 931
 932 static void abort_recovery_queue(struct obd_device *obd)
 933 {
 934         struct ptlrpc_request *req;
 935         struct list_head *tmp, *n;
 936         int rc;
 937
 938         list_for_each_safe(tmp, n, &obd->obd_recovery_queue) {
 939                 req = list_entry(tmp, struct ptlrpc_request, rq_list);
 940                 list_del(&req->rq_list);
 941                 DEBUG_REQ(D_ERROR, req, "aborted:");
 942                 req->rq_status = -ENOTCONN;
 943                 req->rq_type = PTL_RPC_MSG_ERR;
 944                 rc = lustre_pack_reply(req, 1, NULL, NULL);
 945                 if (rc == 0) {
 946                         ptlrpc_reply(req);
 947                 } else {
 948                         DEBUG_REQ(D_ERROR, req,
 949                                   "packing failed for abort-reply; skipping");
 950                 }
 951                 target_release_saved_req(req);
 952         }
 953 }
 954
 955 /* Called from a cleanup function if the device is being cleaned up
 956    forcefully.  The exports should all have been disconnected already,
 957    the only thing left to do is
 958      - clear the recovery flags
 959      - cancel the timer
 960      - free queued requests and replies, but don't send replies
 961    Because the obd_stopping flag is set, no new requests should be received.
 962
 963 */
 964 void target_cleanup_recovery(struct obd_device *obd)
 965 {
 966         struct list_head *tmp, *n;
 967         struct ptlrpc_request *req;
 968         ENTRY;
 969
 970         LASSERT(obd->obd_stopping);
 971
 972         spin_lock_bh(&obd->obd_processing_task_lock);
 973         if (!obd->obd_recovering) {
 974                 spin_unlock_bh(&obd->obd_processing_task_lock);
 975                 EXIT;
 976                 return;
 977         }
 978         obd->obd_recovering = obd->obd_abort_recovery = 0;
 979         target_cancel_recovery_timer(obd);
 980         spin_unlock_bh(&obd->obd_processing_task_lock);
 981
 982         list_for_each_safe(tmp, n, &obd->obd_delayed_reply_queue) {
 983                 req = list_entry(tmp, struct ptlrpc_request, rq_list);
 984                 list_del(&req->rq_list);
 985                 target_release_saved_req(req);
 986         }
 987
 988         list_for_each_safe(tmp, n, &obd->obd_recovery_queue) {
 989                 req = list_entry(tmp, struct ptlrpc_request, rq_list);
 990                 list_del(&req->rq_list);
 991                 target_release_saved_req(req);
 992         }
 993         EXIT;
 994 }
 995
 996 void target_abort_recovery(void *data)
 997 {
 998         struct obd_device *obd = data;
 999
1000         ENTRY;
1001         spin_lock_bh(&obd->obd_processing_task_lock);
1002         if (!obd->obd_recovering) {
1003                 spin_unlock_bh(&obd->obd_processing_task_lock);
1004                 EXIT;
1005                 return;
1006         }
1007         obd->obd_recovering = obd->obd_abort_recovery = 0;
1008         obd->obd_recoverable_clients = 0;
1009         target_cancel_recovery_timer(obd);
1010         spin_unlock_bh(&obd->obd_processing_task_lock);
1011
1012         LCONSOLE_WARN("%s: recovery period over; disconnecting unfinished "
1013                       "clients.\n", obd->obd_name);
1014         class_disconnect_stale_exports(obd);
1015         abort_recovery_queue(obd);
1016
1017         target_finish_recovery(obd);
1018         CDEBUG(D_HA, "%s: recovery complete\n", obd_uuid2str(&obd->obd_uuid));
1019         EXIT;
1020 }
1021
1022 static void target_recovery_expired(unsigned long castmeharder)
1023 {
1024         struct obd_device *obd = (struct obd_device *)castmeharder;
1025         CERROR("%s: recovery timed out, aborting\n", obd->obd_name);
1026         spin_lock_bh(&obd->obd_processing_task_lock);
1027         if (obd->obd_recovering)
1028                 obd->obd_abort_recovery = 1;
1029         cfs_waitq_signal(&obd->obd_next_transno_waitq);
1030         spin_unlock_bh(&obd->obd_processing_task_lock);
1031 }
1032
1033
1034 /* obd_processing_task_lock should be held */
1035 void target_cancel_recovery_timer(struct obd_device *obd)
1036 {
1037         CDEBUG(D_HA, "%s: cancel recovery timer\n", obd->obd_name);
1038         cfs_timer_disarm(&obd->obd_recovery_timer);
1039 }
1040
1041 static void reset_recovery_timer(struct obd_device *obd)
1042 {
1043         spin_lock_bh(&obd->obd_processing_task_lock);
1044         if (!obd->obd_recovering) {
1045                 spin_unlock_bh(&obd->obd_processing_task_lock);
1046                 return;
1047         }
1048         cfs_timer_arm(&obd->obd_recovery_timer,
1049                       cfs_time_shift(OBD_RECOVERY_TIMEOUT));
1050         spin_unlock_bh(&obd->obd_processing_task_lock);
1051         CDEBUG(D_HA, "%s: timer will expire in %u seconds\n", obd->obd_name,
1052                OBD_RECOVERY_TIMEOUT);
1053         /* Only used for lprocfs_status */
1054         obd->obd_recovery_end = CURRENT_SECONDS + OBD_RECOVERY_TIMEOUT;
1055 }
1056
1057
1058 /* Only start it the first time called */
1059 void target_start_recovery_timer(struct obd_device *obd, svc_handler_t handler)
1060 {
1061         spin_lock_bh(&obd->obd_processing_task_lock);
1062         if (obd->obd_recovery_handler) {
1063                 spin_unlock_bh(&obd->obd_processing_task_lock);
1064                 return;
1065         }
1066         CWARN("%s: starting recovery timer (%us)\n", obd->obd_name,
1067               OBD_RECOVERY_TIMEOUT);
1068         obd->obd_recovery_handler = handler;
1069         cfs_timer_init(&obd->obd_recovery_timer, target_recovery_expired, obd);
1070         spin_unlock_bh(&obd->obd_processing_task_lock);
1071
1072         reset_recovery_timer(obd);
1073 }
1074
1075 static int check_for_next_transno(struct obd_device *obd)
1076 {
1077         struct ptlrpc_request *req;
1078         int wake_up = 0, connected, completed, queue_len, max;
1079         __u64 next_transno, req_transno;
1080
1081         spin_lock_bh(&obd->obd_processing_task_lock);
1082         req = list_entry(obd->obd_recovery_queue.next,
1083                          struct ptlrpc_request, rq_list);
1084         max = obd->obd_max_recoverable_clients;
1085         req_transno = lustre_msg_get_transno(req->rq_reqmsg);
1086         connected = obd->obd_connected_clients;
1087         completed = max - obd->obd_recoverable_clients;
1088         queue_len = obd->obd_requests_queued_for_recovery;
1089         next_transno = obd->obd_next_recovery_transno;
1090
1091         CDEBUG(D_HA,"max: %d, connected: %d, completed: %d, queue_len: %d, "
1092                "req_transno: "LPU64", next_transno: "LPU64"\n",
1093                max, connected, completed, queue_len, req_transno, next_transno);
1094         if (obd->obd_abort_recovery) {
1095                 CDEBUG(D_HA, "waking for aborted recovery\n");
1096                 wake_up = 1;
1097         } else if (!obd->obd_recovering) {
1098                 CDEBUG(D_HA, "waking for completed recovery (?)\n");
1099                 wake_up = 1;
1100         } else if (req_transno == next_transno) {
1101                 CDEBUG(D_HA, "waking for next ("LPD64")\n", next_transno);
1102                 wake_up = 1;
1103         } else if (queue_len + completed == max) {
1104                 CDEBUG(D_ERROR,
1105                        "waking for skipped transno (skip: "LPD64
1106                        ", ql: %d, comp: %d, conn: %d, next: "LPD64")\n",
1107                        next_transno, queue_len, completed, max, req_transno);
1108                 obd->obd_next_recovery_transno = req_transno;
1109                 wake_up = 1;
1110         }
1111         spin_unlock_bh(&obd->obd_processing_task_lock);
1112         LASSERT(lustre_msg_get_transno(req->rq_reqmsg) >= next_transno);
1113         return wake_up;
1114 }
1115
1116 static void process_recovery_queue(struct obd_device *obd)
1117 {
1118         struct ptlrpc_request *req;
1119         int abort_recovery = 0;
1120         struct l_wait_info lwi = { 0 };
1121         ENTRY;
1122
1123         for (;;) {
1124                 spin_lock_bh(&obd->obd_processing_task_lock);
1125                 LASSERT(obd->obd_processing_task == cfs_curproc_pid());
1126                 req = list_entry(obd->obd_recovery_queue.next,
1127                                  struct ptlrpc_request, rq_list);
1128
1129                 if (lustre_msg_get_transno(req->rq_reqmsg) !=
1130                     obd->obd_next_recovery_transno) {
1131                         spin_unlock_bh(&obd->obd_processing_task_lock);
1132                         CDEBUG(D_HA, "Waiting for transno "LPD64" (1st is "
1133                                LPD64")\n",
1134                                obd->obd_next_recovery_transno,
1135                                lustre_msg_get_transno(req->rq_reqmsg));
1136                         l_wait_event(obd->obd_next_transno_waitq,
1137                                      check_for_next_transno(obd), &lwi);
1138                         spin_lock_bh(&obd->obd_processing_task_lock);
1139                         abort_recovery = obd->obd_abort_recovery;
1140                         spin_unlock_bh(&obd->obd_processing_task_lock);
1141                         if (abort_recovery) {
1142                                 target_abort_recovery(obd);
1143                                 return;
1144                         }
1145                         continue;
1146                 }
1147                 list_del_init(&req->rq_list);
1148                 obd->obd_requests_queued_for_recovery--;
1149                 spin_unlock_bh(&obd->obd_processing_task_lock);
1150
1151                 DEBUG_REQ(D_HA, req, "processing: ");
1152                 (void)obd->obd_recovery_handler(req);
1153                 obd->obd_replayed_requests++;
1154                 reset_recovery_timer(obd);
1155                 /* bug 1580: decide how to properly sync() in recovery */
1156                 //mds_fsync_super(obd->u.obt.obt_sb);
1157                 class_export_put(req->rq_export);
1158                 if (req->rq_reply_state != NULL) {
1159                         ptlrpc_rs_decref(req->rq_reply_state);
1160                         /* req->rq_reply_state = NULL; */
1161                 }
1162                 OBD_FREE(req->rq_reqmsg, req->rq_reqlen);
1163                 OBD_FREE(req, sizeof *req);
1164                 spin_lock_bh(&obd->obd_processing_task_lock);
1165                 obd->obd_next_recovery_transno++;
1166                 if (list_empty(&obd->obd_recovery_queue)) {
1167                         obd->obd_processing_task = 0;
1168                         spin_unlock_bh(&obd->obd_processing_task_lock);
1169                         break;
1170                 }
1171                 spin_unlock_bh(&obd->obd_processing_task_lock);
1172         }
1173         EXIT;
1174 }
1175
1176 int target_queue_recovery_request(struct ptlrpc_request *req,
1177                                   struct obd_device *obd)
1178 {
1179         struct list_head *tmp;
1180         int inserted = 0;
1181         __u64 transno = lustre_msg_get_transno(req->rq_reqmsg);
1182         struct ptlrpc_request *saved_req;
1183         struct lustre_msg *reqmsg;
1184
1185         /* CAVEAT EMPTOR: The incoming request message has been swabbed
1186          * (i.e. buflens etc are in my own byte order), but type-dependent
1187          * buffers (eg mds_body, ost_body etc) have NOT been swabbed. */
1188
1189         if (!transno) {
1190                 CFS_INIT_LIST_HEAD(&req->rq_list);
1191                 DEBUG_REQ(D_HA, req, "not queueing");
1192                 return 1;
1193         }
1194
1195         /* XXX If I were a real man, these LBUGs would be sane cleanups. */
1196         /* XXX just like the request-dup code in queue_final_reply */
1197         OBD_ALLOC(saved_req, sizeof *saved_req);
1198         if (!saved_req)
1199                 LBUG();
1200         OBD_ALLOC(reqmsg, req->rq_reqlen);
1201         if (!reqmsg)
1202                 LBUG();
1203
1204         spin_lock_bh(&obd->obd_processing_task_lock);
1205
1206         /* If we're processing the queue, we want don't want to queue this
1207          * message.
1208          *
1209          * Also, if this request has a transno less than the one we're waiting
1210          * for, we should process it now.  It could (and currently always will)
1211          * be an open request for a descriptor that was opened some time ago.
1212          *
1213          * Also, a resent, replayed request that has already been
1214          * handled will pass through here and be processed immediately.
1215          */
1216         if (obd->obd_processing_task == cfs_curproc_pid() ||
1217             transno < obd->obd_next_recovery_transno) {
1218                 /* Processing the queue right now, don't re-add. */
1219                 LASSERT(list_empty(&req->rq_list));
1220                 spin_unlock_bh(&obd->obd_processing_task_lock);
1221                 OBD_FREE(reqmsg, req->rq_reqlen);
1222                 OBD_FREE(saved_req, sizeof *saved_req);
1223                 return 1;
1224         }
1225
1226         /* A resent, replayed request that is still on the queue; just drop it.
1227            The queued request will handle this. */
1228         if ((lustre_msg_get_flags(req->rq_reqmsg) & (MSG_RESENT|MSG_REPLAY)) ==
1229             (MSG_RESENT | MSG_REPLAY)) {
1230                 DEBUG_REQ(D_ERROR, req, "dropping resent queued req");
1231                 spin_unlock_bh(&obd->obd_processing_task_lock);
1232                 OBD_FREE(reqmsg, req->rq_reqlen);
1233                 OBD_FREE(saved_req, sizeof *saved_req);
1234                 return 0;
1235         }
1236
1237         memcpy(saved_req, req, sizeof *req);
1238         memcpy(reqmsg, req->rq_reqmsg, req->rq_reqlen);
1239         req = saved_req;
1240         req->rq_reqmsg = reqmsg;
1241         class_export_get(req->rq_export);
1242         CFS_INIT_LIST_HEAD(&req->rq_list);
1243
1244         /* XXX O(n^2) */
1245         list_for_each(tmp, &obd->obd_recovery_queue) {
1246                 struct ptlrpc_request *reqiter =
1247                         list_entry(tmp, struct ptlrpc_request, rq_list);
1248
1249                 if (lustre_msg_get_transno(reqiter->rq_reqmsg) > transno) {
1250                         list_add_tail(&req->rq_list, &reqiter->rq_list);
1251                         inserted = 1;
1252                         break;
1253                 }
1254         }
1255
1256         if (!inserted) {
1257                 list_add_tail(&req->rq_list, &obd->obd_recovery_queue);
1258         }
1259
1260         obd->obd_requests_queued_for_recovery++;
1261
1262         if (obd->obd_processing_task != 0) {
1263                 /* Someone else is processing this queue, we'll leave it to
1264                  * them.
1265                  */
1266                 cfs_waitq_signal(&obd->obd_next_transno_waitq);
1267                 spin_unlock_bh(&obd->obd_processing_task_lock);
1268                 return 0;
1269         }
1270
1271         /* Nobody is processing, and we know there's (at least) one to process
1272          * now, so we'll do the honours.
1273          */
1274         obd->obd_processing_task = cfs_curproc_pid();
1275         spin_unlock_bh(&obd->obd_processing_task_lock);
1276
1277         process_recovery_queue(obd);
1278         return 0;
1279 }
1280
1281 struct obd_device * target_req2obd(struct ptlrpc_request *req)
1282 {
1283         return req->rq_export->exp_obd;
1284 }
1285
1286 int target_queue_final_reply(struct ptlrpc_request *req, int rc)
1287 {
1288         struct obd_device *obd = target_req2obd(req);
1289         struct ptlrpc_request *saved_req;
1290         struct lustre_msg *reqmsg;
1291         int recovery_done = 0;
1292
1293         LASSERT ((rc == 0) == (req->rq_reply_state != NULL));
1294
1295         if (rc) {
1296                 /* Just like ptlrpc_error, but without the sending. */
1297                 rc = lustre_pack_reply(req, 1, NULL, NULL);
1298                 LASSERT(rc == 0); /* XXX handle this */
1299                 req->rq_type = PTL_RPC_MSG_ERR;
1300         }
1301
1302         LASSERT (!req->rq_reply_state->rs_difficult);
1303         LASSERT(list_empty(&req->rq_list));
1304         /* XXX a bit like the request-dup code in queue_recovery_request */
1305         OBD_ALLOC(saved_req, sizeof *saved_req);
1306         if (!saved_req)
1307                 LBUG();
1308         OBD_ALLOC(reqmsg, req->rq_reqlen);
1309         if (!reqmsg)
1310                 LBUG();
1311         *saved_req = *req;
1312         memcpy(reqmsg, req->rq_reqmsg, req->rq_reqlen);
1313
1314         /* Don't race cleanup */
1315         spin_lock_bh(&obd->obd_processing_task_lock);
1316         if (obd->obd_stopping) {
1317                 spin_unlock_bh(&obd->obd_processing_task_lock);
1318                 OBD_FREE(reqmsg, req->rq_reqlen);
1319                 OBD_FREE(saved_req, sizeof *req);
1320                 req->rq_status = -ENOTCONN;
1321                 /* rv is ignored anyhow */
1322                 return -ENOTCONN;
1323         }
1324         ptlrpc_rs_addref(req->rq_reply_state);  /* +1 ref for saved reply */
1325         req = saved_req;
1326         req->rq_reqmsg = reqmsg;
1327         class_export_get(req->rq_export);
1328         list_add(&req->rq_list, &obd->obd_delayed_reply_queue);
1329
1330         /* only count the first "replay over" request from each
1331            export */
1332         if (req->rq_export->exp_replay_needed) {
1333                 --obd->obd_recoverable_clients;
1334                 req->rq_export->exp_replay_needed = 0;
1335         }
1336         recovery_done = (obd->obd_recoverable_clients == 0);
1337         spin_unlock_bh(&obd->obd_processing_task_lock);
1338
1339         OBD_RACE(OBD_FAIL_LDLM_RECOV_CLIENTS);
1340         if (recovery_done) {
1341                 spin_lock_bh(&obd->obd_processing_task_lock);
1342                 obd->obd_recovering = obd->obd_abort_recovery = 0;
1343                 target_cancel_recovery_timer(obd);
1344                 spin_unlock_bh(&obd->obd_processing_task_lock);
1345
1346                 target_finish_recovery(obd);
1347                 CDEBUG(D_HA, "%s: recovery complete\n",
1348                        obd_uuid2str(&obd->obd_uuid));
1349         } else {
1350                 CWARN("%s: %d recoverable clients remain\n",
1351                        obd->obd_name, obd->obd_recoverable_clients);
1352                 cfs_waitq_signal(&obd->obd_next_transno_waitq);
1353         }
1354
1355         return 1;
1356 }
1357
1358 int
1359 target_send_reply_msg (struct ptlrpc_request *req, int rc, int fail_id)
1360 {
1361         if (OBD_FAIL_CHECK(fail_id | OBD_FAIL_ONCE)) {
1362                 obd_fail_loc |= OBD_FAIL_ONCE | OBD_FAILED;
1363                 DEBUG_REQ(D_ERROR, req, "dropping reply");
1364                 return (-ECOMM);
1365         }
1366
1367         if (rc) {
1368                 DEBUG_REQ(D_ERROR, req, "processing error (%d)", rc);
1369                 req->rq_status = rc;
1370                 return (ptlrpc_error(req));
1371         } else {
1372                 DEBUG_REQ(D_NET, req, "sending reply");
1373         }
1374
1375         return (ptlrpc_send_reply(req, 1));
1376 }
1377
1378 void
1379 target_send_reply(struct ptlrpc_request *req, int rc, int fail_id)
1380 {
1381         int                        netrc;
1382         struct ptlrpc_reply_state *rs;
1383         struct obd_device         *obd;
1384         struct obd_export         *exp;
1385         struct ptlrpc_service     *svc;
1386
1387         svc = req->rq_rqbd->rqbd_service;
1388         rs = req->rq_reply_state;
1389         if (rs == NULL || !rs->rs_difficult) {
1390                 /* no notifiers */
1391                 target_send_reply_msg (req, rc, fail_id);
1392                 return;
1393         }
1394
1395         /* must be an export if locks saved */
1396         LASSERT (req->rq_export != NULL);
1397         /* req/reply consistent */
1398         LASSERT (rs->rs_service == svc);
1399
1400         /* "fresh" reply */
1401         LASSERT (!rs->rs_scheduled);
1402         LASSERT (!rs->rs_scheduled_ever);
1403         LASSERT (!rs->rs_handled);
1404         LASSERT (!rs->rs_on_net);
1405         LASSERT (rs->rs_export == NULL);
1406         LASSERT (list_empty(&rs->rs_obd_list));
1407         LASSERT (list_empty(&rs->rs_exp_list));
1408
1409         exp = class_export_get (req->rq_export);
1410         obd = exp->exp_obd;
1411
1412         /* disable reply scheduling onto srv_reply_queue while I'm setting up */
1413         rs->rs_scheduled = 1;
1414         rs->rs_on_net    = 1;
1415         rs->rs_xid       = req->rq_xid;
1416         rs->rs_transno   = req->rq_transno;
1417         rs->rs_export    = exp;
1418
1419         spin_lock(&obd->obd_uncommitted_replies_lock);
1420
1421         if (rs->rs_transno > obd->obd_last_committed) {
1422                 /* not committed already */
1423                 list_add_tail (&rs->rs_obd_list,
1424                                &obd->obd_uncommitted_replies);
1425         }
1426
1427         spin_unlock (&obd->obd_uncommitted_replies_lock);
1428         spin_lock (&exp->exp_lock);
1429
1430         list_add_tail (&rs->rs_exp_list, &exp->exp_outstanding_replies);
1431
1432         spin_unlock(&exp->exp_lock);
1433
1434         netrc = target_send_reply_msg (req, rc, fail_id);
1435
1436         spin_lock(&svc->srv_lock);
1437
1438         svc->srv_n_difficult_replies++;
1439
1440         if (netrc != 0) {
1441                 /* error sending: reply is off the net.  Also we need +1
1442                  * reply ref until ptlrpc_server_handle_reply() is done
1443                  * with the reply state (if the send was successful, there
1444                  * would have been +1 ref for the net, which
1445                  * reply_out_callback leaves alone) */
1446                 rs->rs_on_net = 0;
1447                 ptlrpc_rs_addref(rs);
1448                 atomic_inc (&svc->srv_outstanding_replies);
1449         }
1450
1451         if (!rs->rs_on_net ||                   /* some notifier */
1452             list_empty(&rs->rs_exp_list) ||     /* completed already */
1453             list_empty(&rs->rs_obd_list)) {
1454                 list_add_tail (&rs->rs_list, &svc->srv_reply_queue);
1455                 cfs_waitq_signal (&svc->srv_waitq);
1456         } else {
1457                 list_add (&rs->rs_list, &svc->srv_active_replies);
1458                 rs->rs_scheduled = 0;           /* allow notifier to schedule */
1459         }
1460
1461         spin_unlock(&svc->srv_lock);
1462 }
1463
1464 int target_handle_ping(struct ptlrpc_request *req)
1465 {
1466         obd_ping(req->rq_export);
1467         return lustre_pack_reply(req, 1, NULL, NULL);
1468 }
1469
1470 void target_committed_to_req(struct ptlrpc_request *req)
1471 {
1472         struct obd_device *obd = req->rq_export->exp_obd;
1473
1474         if (!obd->obd_no_transno && req->rq_repmsg != NULL)
1475                 lustre_msg_set_last_committed(req->rq_repmsg,
1476                                               obd->obd_last_committed);
1477         else
1478                 DEBUG_REQ(D_IOCTL, req, "not sending last_committed update");
1479
1480         CDEBUG(D_INFO, "last_committed "LPU64", xid "LPU64"\n",
1481                obd->obd_last_committed, req->rq_xid);
1482 }
1483
1484 EXPORT_SYMBOL(target_committed_to_req);
1485
1486 #ifdef HAVE_QUOTA_SUPPORT
1487 int target_handle_qc_callback(struct ptlrpc_request *req)
1488 {
1489         struct obd_quotactl *oqctl;
1490         struct client_obd *cli = &req->rq_export->exp_obd->u.cli;
1491
1492         oqctl = lustre_swab_reqbuf(req, REQ_REC_OFF, sizeof(*oqctl),
1493                                    lustre_swab_obd_quotactl);
1494         if (oqctl == NULL) {
1495                 CERROR("Can't unpack obd_quotactl\n");
1496                 RETURN(-EPROTO);
1497         }
1498
1499         cli->cl_qchk_stat = oqctl->qc_stat;
1500
1501         return 0;
1502 }
1503
1504 int target_handle_dqacq_callback(struct ptlrpc_request *req)
1505 {
1506 #ifdef __KERNEL__
1507         struct obd_device *obd = req->rq_export->exp_obd;
1508         struct obd_device *master_obd;
1509         struct lustre_quota_ctxt *qctxt;
1510         struct qunit_data *qdata;
1511         void* rep;
1512         struct qunit_data_old *qdata_old;
1513         int rc = 0;
1514         int repsize[2] = { sizeof(struct ptlrpc_body),
1515                            sizeof(struct qunit_data) };
1516         ENTRY;
1517
1518         rc = lustre_pack_reply(req, 2, repsize, NULL);
1519         if (rc) {
1520                 CERROR("packing reply failed!: rc = %d\n", rc);
1521                 RETURN(rc);
1522         }
1523         LASSERT(req->rq_export);
1524
1525         /* fixed for bug10707 */
1526         if ((req->rq_export->exp_connect_flags & OBD_CONNECT_QUOTA64) &&
1527             !OBD_FAIL_CHECK(OBD_FAIL_QUOTA_QD_COUNT_32BIT)) {
1528                 CDEBUG(D_QUOTA, "qd_count is 64bit!\n");
1529                 rep = lustre_msg_buf(req->rq_repmsg, REPLY_REC_OFF,
1530                                      sizeof(struct qunit_data));
1531                 LASSERT(rep);
1532                 qdata = lustre_swab_reqbuf(req, REQ_REC_OFF, sizeof(*qdata),
1533                                            lustre_swab_qdata);
1534         } else {
1535                 CDEBUG(D_QUOTA, "qd_count is 32bit!\n");
1536                 rep = lustre_msg_buf(req->rq_repmsg, REPLY_REC_OFF,
1537                                      sizeof(struct qunit_data_old));
1538                 LASSERT(rep);
1539                 qdata_old = lustre_swab_reqbuf(req, REQ_REC_OFF, sizeof(*qdata_old),
1540                                                lustre_swab_qdata_old);
1541                 qdata = lustre_quota_old_to_new(qdata_old);
1542         }
1543
1544         if (qdata == NULL) {
1545                 CERROR("Can't unpack qunit_data\n");
1546                 RETURN(-EPROTO);
1547         }
1548
1549         /* we use the observer */
1550         LASSERT(obd->obd_observer && obd->obd_observer->obd_observer);
1551         master_obd = obd->obd_observer->obd_observer;
1552         qctxt = &master_obd->u.obt.obt_qctxt;
1553
1554         LASSERT(qctxt->lqc_handler);
1555         rc = qctxt->lqc_handler(master_obd, qdata,
1556                                 lustre_msg_get_opc(req->rq_reqmsg));
1557         if (rc && rc != -EDQUOT)
1558                 CDEBUG(rc == -EBUSY  ? D_QUOTA : D_ERROR,
1559                        "dqacq failed! (rc:%d)\n", rc);
1560
1561         /* the qd_count might be changed in lqc_handler */
1562         if ((req->rq_export->exp_connect_flags & OBD_CONNECT_QUOTA64) &&
1563             !OBD_FAIL_CHECK(OBD_FAIL_QUOTA_QD_COUNT_32BIT)) {
1564                 memcpy(rep,qdata,sizeof(*qdata));
1565         } else {
1566                 qdata_old = lustre_quota_new_to_old(qdata);
1567                 memcpy(rep,qdata_old,sizeof(*qdata_old));
1568         }
1569         req->rq_status = rc;
1570         rc = ptlrpc_reply(req);
1571
1572         RETURN(rc);
1573 #else
1574         return 0;
1575 #endif /* !__KERNEL__ */
1576 }
1577 #endif /* HAVE_QUOTA_SUPPORT */
1578
1579 ldlm_mode_t lck_compat_array[] = {
1580         [LCK_EX] LCK_COMPAT_EX,
1581         [LCK_PW] LCK_COMPAT_PW,
1582         [LCK_PR] LCK_COMPAT_PR,
1583         [LCK_CW] LCK_COMPAT_CW,
1584         [LCK_CR] LCK_COMPAT_CR,
1585         [LCK_NL] LCK_COMPAT_NL,
1586         [LCK_GROUP] LCK_COMPAT_GROUP
1587 };