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
 322                         spin_lock(&imp->imp_lock);
 323                         imp->imp_invalid = 1;
 324                         spin_unlock(&imp->imp_lock);
 325                 }
 326         }
 327
 328         cli->cl_qchk_stat = CL_NOT_QUOTACHECKED;
 329
 330         RETURN(rc);
 331
 332 err_import:
 333         class_destroy_import(imp);
 334 err_ldlm:
 335         ldlm_put_ref(0);
 336 err:
 337         RETURN(rc);
 338
 339 }
 340
 341 int client_obd_cleanup(struct obd_device *obddev)
 342 {
 343         ENTRY;
 344         ldlm_put_ref(obddev->obd_force);
 345
 346         RETURN(0);
 347 }
 348
 349 /* ->o_connect() method for client side (OSC and MDC and MGC) */
 350 int client_connect_import(struct lustre_handle *dlm_handle,
 351                           struct obd_device *obd, struct obd_uuid *cluuid,
 352                           struct obd_connect_data *data)
 353 {
 354         struct client_obd *cli = &obd->u.cli;
 355         struct obd_import *imp = cli->cl_import;
 356         struct obd_export *exp;
 357         struct obd_connect_data *ocd;
 358         int rc;
 359         ENTRY;
 360
 361         mutex_down(&cli->cl_sem);
 362         rc = class_connect(dlm_handle, obd, cluuid);
 363         if (rc)
 364                 GOTO(out_sem, rc);
 365
 366         cli->cl_conn_count++;
 367         if (cli->cl_conn_count > 1)
 368                 GOTO(out_sem, rc);
 369         exp = class_conn2export(dlm_handle);
 370
 371         if (obd->obd_namespace != NULL)
 372                 CERROR("already have namespace!\n");
 373         obd->obd_namespace = ldlm_namespace_new(obd->obd_name,
 374                                                 LDLM_NAMESPACE_CLIENT);
 375         if (obd->obd_namespace == NULL)
 376                 GOTO(out_disco, rc = -ENOMEM);
 377
 378         imp->imp_dlm_handle = *dlm_handle;
 379         rc = ptlrpc_init_import(imp);
 380         if (rc != 0)
 381                 GOTO(out_ldlm, rc);
 382
 383         ocd = &imp->imp_connect_data;
 384         if (data) {
 385                 *ocd = *data;
 386                 imp->imp_connect_flags_orig = data->ocd_connect_flags;
 387         }
 388
 389         rc = ptlrpc_connect_import(imp, NULL);
 390         if (rc != 0) {
 391                 LASSERT (imp->imp_state == LUSTRE_IMP_DISCON);
 392                 GOTO(out_ldlm, rc);
 393         }
 394         LASSERT(exp->exp_connection);
 395
 396         if (data) {
 397                 LASSERT((ocd->ocd_connect_flags & data->ocd_connect_flags) ==
 398                         ocd->ocd_connect_flags);
 399                 data->ocd_connect_flags = ocd->ocd_connect_flags;
 400         }
 401
 402         ptlrpc_pinger_add_import(imp);
 403         EXIT;
 404
 405         if (rc) {
 406 out_ldlm:
 407                 ldlm_namespace_free(obd->obd_namespace, 0);
 408                 obd->obd_namespace = NULL;
 409 out_disco:
 410                 cli->cl_conn_count--;
 411                 class_disconnect(exp);
 412         } else {
 413                 class_export_put(exp);
 414         }
 415 out_sem:
 416         mutex_up(&cli->cl_sem);
 417         return rc;
 418 }
 419
 420 int client_disconnect_export(struct obd_export *exp)
 421 {
 422         struct obd_device *obd = class_exp2obd(exp);
 423         struct client_obd *cli;
 424         struct obd_import *imp;
 425         int rc = 0, err;
 426         ENTRY;
 427
 428         if (!obd) {
 429                 CERROR("invalid export for disconnect: exp %p cookie "LPX64"\n",
 430                        exp, exp ? exp->exp_handle.h_cookie : -1);
 431                 RETURN(-EINVAL);
 432         }
 433
 434         cli = &obd->u.cli;
 435         imp = cli->cl_import;
 436
 437         mutex_down(&cli->cl_sem);
 438         if (!cli->cl_conn_count) {
 439                 CERROR("disconnecting disconnected device (%s)\n",
 440                        obd->obd_name);
 441                 GOTO(out_sem, rc = -EINVAL);
 442         }
 443
 444         cli->cl_conn_count--;
 445         if (cli->cl_conn_count)
 446                 GOTO(out_no_disconnect, rc = 0);
 447
 448         /* Mark import deactivated now, so we don't try to reconnect if any
 449          * of the cleanup RPCs fails (e.g. ldlm cancel, etc).  We don't
 450          * fully deactivate the import, or that would drop all requests. */
 451         spin_lock(&imp->imp_lock);
 452         imp->imp_deactive = 1;
 453         spin_unlock(&imp->imp_lock);
 454
 455         /* Some non-replayable imports (MDS's OSCs) are pinged, so just
 456          * delete it regardless.  (It's safe to delete an import that was
 457          * never added.) */
 458         (void)ptlrpc_pinger_del_import(imp);
 459
 460         if (obd->obd_namespace != NULL) {
 461                 /* obd_no_recov == local only */
 462                 ldlm_cli_cancel_unused(obd->obd_namespace, NULL,
 463                                        obd->obd_no_recov ? LDLM_FL_LOCAL_ONLY:0,
 464                                        NULL);
 465                 ldlm_namespace_free(obd->obd_namespace, obd->obd_no_recov);
 466                 obd->obd_namespace = NULL;
 467         }
 468
 469         /* Yeah, obd_no_recov also (mainly) means "forced shutdown". */
 470         if (!obd->obd_no_recov)
 471                 rc = ptlrpc_disconnect_import(imp, 0);
 472
 473         ptlrpc_invalidate_import(imp);
 474         ptlrpc_free_rq_pool(imp->imp_rq_pool);
 475         class_destroy_import(imp);
 476         cli->cl_import = NULL;
 477
 478         EXIT;
 479  out_no_disconnect:
 480         err = class_disconnect(exp);
 481         if (!rc && err)
 482                 rc = err;
 483  out_sem:
 484         mutex_up(&cli->cl_sem);
 485         RETURN(rc);
 486 }
 487
 488 /* --------------------------------------------------------------------------
 489  * from old lib/target.c
 490  * -------------------------------------------------------------------------- */
 491
 492 int target_handle_reconnect(struct lustre_handle *conn, struct obd_export *exp,
 493                             struct obd_uuid *cluuid)
 494 {
 495         ENTRY;
 496         if (exp->exp_connection && exp->exp_imp_reverse) {
 497                 struct lustre_handle *hdl;
 498                 hdl = &exp->exp_imp_reverse->imp_remote_handle;
 499                 /* Might be a re-connect after a partition. */
 500                 if (!memcmp(&conn->cookie, &hdl->cookie, sizeof conn->cookie)) {
 501                         CWARN("%s: %s reconnecting\n", exp->exp_obd->obd_name,
 502                               cluuid->uuid);
 503                         conn->cookie = exp->exp_handle.h_cookie;
 504                         /* target_handle_connect() treats EALREADY and
 505                          * -EALREADY differently.  EALREADY means we are
 506                          * doing a valid reconnect from the same client. */
 507                         RETURN(EALREADY);
 508                 } else {
 509                         CERROR("%s reconnecting from %s, "
 510                                "handle mismatch (ours "LPX64", theirs "
 511                                LPX64")\n", cluuid->uuid,
 512                                exp->exp_connection->c_remote_uuid.uuid,
 513                                hdl->cookie, conn->cookie);
 514                         memset(conn, 0, sizeof *conn);
 515                         /* target_handle_connect() treats EALREADY and
 516                          * -EALREADY differently.  -EALREADY is an error
 517                          * (same UUID, different handle). */
 518                         RETURN(-EALREADY);
 519                 }
 520         }
 521
 522         conn->cookie = exp->exp_handle.h_cookie;
 523         CDEBUG(D_HA, "connect export for UUID '%s' at %p, cookie "LPX64"\n",
 524                cluuid->uuid, exp, conn->cookie);
 525         RETURN(0);
 526 }
 527
 528 void target_client_add_cb(struct obd_device *obd, __u64 transno, void *cb_data,
 529                           int error)
 530 {
 531         struct obd_export *exp = cb_data;
 532
 533         CDEBUG(D_HA, "%s: committing for initial connect of %s\n",
 534                obd->obd_name, exp->exp_client_uuid.uuid);
 535
 536         spin_lock(&exp->exp_lock);
 537         exp->exp_need_sync = 0;
 538         spin_unlock(&exp->exp_lock);
 539 }
 540 EXPORT_SYMBOL(target_client_add_cb);
 541
 542 int target_handle_connect(struct ptlrpc_request *req, svc_handler_t handler)
 543 {
 544         struct obd_device *target, *targref = NULL;
 545         struct obd_export *export = NULL;
 546         struct obd_import *revimp;
 547         struct lustre_handle conn;
 548         struct obd_uuid tgtuuid;
 549         struct obd_uuid cluuid;
 550         struct obd_uuid remote_uuid;
 551         struct list_head *p;
 552         char *str, *tmp;
 553         int rc = 0, abort_recovery;
 554         struct obd_connect_data *data;
 555         int size[2] = { sizeof(struct ptlrpc_body), sizeof(*data) };
 556         ENTRY;
 557
 558         OBD_RACE(OBD_FAIL_TGT_CONN_RACE);
 559
 560         LASSERT_REQSWAB(req, REQ_REC_OFF);
 561         str = lustre_msg_string(req->rq_reqmsg, REQ_REC_OFF, sizeof(tgtuuid)-1);
 562         if (str == NULL) {
 563                 DEBUG_REQ(D_ERROR, req, "bad target UUID for connect\n");
 564                 GOTO(out, rc = -EINVAL);
 565         }
 566
 567         obd_str2uuid (&tgtuuid, str);
 568         target = class_uuid2obd(&tgtuuid);
 569         /* COMPAT_146 */
 570         /* old (pre 1.6) lustre_process_log tries to connect to mdsname
 571            (eg. mdsA) instead of uuid. */
 572         if (!target) {
 573                 snprintf((char *)tgtuuid.uuid, sizeof(tgtuuid), "%s_UUID", str);
 574                 target = class_uuid2obd(&tgtuuid);
 575         }
 576         if (!target)
 577                 target = class_name2obd(str);
 578         /* end COMPAT_146 */
 579
 580         if (!target || target->obd_stopping || !target->obd_set_up) {
 581                 LCONSOLE_ERROR("UUID '%s' is not available "
 582                                " for connect (%s)\n", str,
 583                                !target ? "no target" :
 584                                (target->obd_stopping ? "stopping" :
 585                                 "not set up"));
 586                 GOTO(out, rc = -ENODEV);
 587         }
 588
 589         if (target->obd_no_conn) {
 590                 LCONSOLE_WARN("%s: temporarily refusing client connection "
 591                               "from %s\n", target->obd_name,
 592                               libcfs_nid2str(req->rq_peer.nid));
 593                 GOTO(out, rc = -EAGAIN);
 594         }
 595
 596         /* Make sure the target isn't cleaned up while we're here. Yes,
 597            there's still a race between the above check and our incref here.
 598            Really, class_uuid2obd should take the ref. */
 599         targref = class_incref(target);
 600
 601         LASSERT_REQSWAB(req, REQ_REC_OFF + 1);
 602         str = lustre_msg_string(req->rq_reqmsg, REQ_REC_OFF + 1,
 603                                 sizeof(cluuid) - 1);
 604         if (str == NULL) {
 605                 DEBUG_REQ(D_ERROR, req, "bad client UUID for connect\n");
 606                 GOTO(out, rc = -EINVAL);
 607         }
 608
 609         obd_str2uuid (&cluuid, str);
 610
 611         /* XXX extract a nettype and format accordingly */
 612         switch (sizeof(lnet_nid_t)) {
 613                 /* NB the casts only avoid compiler warnings */
 614         case 8:
 615                 snprintf(remote_uuid.uuid, sizeof remote_uuid,
 616                          "NET_"LPX64"_UUID", (__u64)req->rq_peer.nid);
 617                 break;
 618         case 4:
 619                 snprintf(remote_uuid.uuid, sizeof remote_uuid,
 620                          "NET_%x_UUID", (__u32)req->rq_peer.nid);
 621                 break;
 622         default:
 623                 LBUG();
 624         }
 625
 626         spin_lock_bh(&target->obd_processing_task_lock);
 627         abort_recovery = target->obd_abort_recovery;
 628         spin_unlock_bh(&target->obd_processing_task_lock);
 629         if (abort_recovery)
 630                 target_abort_recovery(target);
 631
 632         tmp = lustre_msg_buf(req->rq_reqmsg, REQ_REC_OFF + 2, sizeof conn);
 633         if (tmp == NULL)
 634                 GOTO(out, rc = -EPROTO);
 635
 636         memcpy(&conn, tmp, sizeof conn);
 637
 638         data = lustre_swab_reqbuf(req, REQ_REC_OFF + 3, sizeof(*data),
 639                                   lustre_swab_connect);
 640         rc = lustre_pack_reply(req, 2, size, NULL);
 641         if (rc)
 642                 GOTO(out, rc);
 643
 644         if (lustre_msg_get_op_flags(req->rq_reqmsg) & MSG_CONNECT_LIBCLIENT) {
 645                 if (!data) {
 646                         DEBUG_REQ(D_WARNING, req, "Refusing old (unversioned) "
 647                                   "libclient connection attempt\n");
 648                         GOTO(out, rc = -EPROTO);
 649                 } else if (data->ocd_version < LUSTRE_VERSION_CODE -
 650                                                LUSTRE_VERSION_ALLOWED_OFFSET ||
 651                            data->ocd_version > LUSTRE_VERSION_CODE +
 652                                                LUSTRE_VERSION_ALLOWED_OFFSET) {
 653                         DEBUG_REQ(D_WARNING, req, "Refusing %s (%d.%d.%d.%d) "
 654                                   "libclient connection attempt\n",
 655                                   data->ocd_version < LUSTRE_VERSION_CODE ?
 656                                   "old" : "new",
 657                                   OBD_OCD_VERSION_MAJOR(data->ocd_version),
 658                                   OBD_OCD_VERSION_MINOR(data->ocd_version),
 659                                   OBD_OCD_VERSION_PATCH(data->ocd_version),
 660                                   OBD_OCD_VERSION_FIX(data->ocd_version));
 661                         data = lustre_msg_buf(req->rq_repmsg, REPLY_REC_OFF,
 662                                               offsetof(typeof(*data),
 663                                                        ocd_version) +
 664                                               sizeof(data->ocd_version));
 665                         if (data) {
 666                                 data->ocd_connect_flags = OBD_CONNECT_VERSION;
 667                                 data->ocd_version = LUSTRE_VERSION_CODE;
 668                         }
 669                         GOTO(out, rc = -EPROTO);
 670                 }
 671         }
 672
 673         /* lctl gets a backstage, all-access pass. */
 674         if (obd_uuid_equals(&cluuid, &target->obd_uuid))
 675                 goto dont_check_exports;
 676
 677         spin_lock(&target->obd_dev_lock);
 678         list_for_each(p, &target->obd_exports) {
 679                 export = list_entry(p, struct obd_export, exp_obd_chain);
 680                 if (obd_uuid_equals(&cluuid, &export->exp_client_uuid)) {
 681                         if (export->exp_connecting) { /* bug 9635, et. al. */
 682                                 CWARN("%s: exp %p already connecting\n",
 683                                       export->exp_obd->obd_name, export);
 684                                 export = NULL;
 685                                 rc = -EALREADY;
 686                                 break;
 687                         }
 688
 689                         /* make darn sure this is coming from the same peer
 690                          * if the UUIDs matched */
 691                         if ((export->exp_connection != NULL) &&
 692                                         (strcmp(libcfs_nid2str(req->rq_peer.nid),
 693                                                 libcfs_nid2str(export->exp_connection->c_peer.nid)))) {
 694                                         CWARN("%s: cookie %s seen on new NID %s when "
 695                                                         "existing NID %s is already connected\n",
 696                                                         target->obd_name, cluuid.uuid,
 697                                                         libcfs_nid2str(req->rq_peer.nid),
 698                                                         libcfs_nid2str(export->exp_connection->c_peer.nid));
 699                                         export = NULL;
 700                                         rc = -EALREADY;
 701                                         break;
 702                         }
 703
 704                         spin_lock(&export->exp_lock);
 705                         export->exp_connecting = 1;
 706                         spin_unlock(&export->exp_lock);
 707                         spin_unlock(&target->obd_dev_lock);
 708                         LASSERT(export->exp_obd == target);
 709
 710                         rc = target_handle_reconnect(&conn, export, &cluuid);
 711                         break;
 712                 }
 713                 export = NULL;
 714         }
 715         /* If we found an export, we already unlocked. */
 716         if (!export) {
 717                 spin_unlock(&target->obd_dev_lock);
 718                 OBD_FAIL_TIMEOUT(OBD_FAIL_TGT_DELAY_CONNECT, 2 * obd_timeout);
 719         } else if (req->rq_export == NULL &&
 720                    atomic_read(&export->exp_rpc_count) > 0) {
 721                 CWARN("%s: refuse connection from %s/%s to 0x%p/%d\n",
 722                       target->obd_name, cluuid.uuid,
 723                       libcfs_nid2str(req->rq_peer.nid),
 724                       export, atomic_read(&export->exp_refcount));
 725                 GOTO(out, rc = -EBUSY);
 726         } else if (req->rq_export != NULL &&
 727                    atomic_read(&export->exp_rpc_count) > 1) {
 728                 CWARN("%s: refuse reconnection from %s@%s to 0x%p/%d\n",
 729                       target->obd_name, cluuid.uuid,
 730                       libcfs_nid2str(req->rq_peer.nid),
 731                       export, atomic_read(&export->exp_rpc_count));
 732                 GOTO(out, rc = -EBUSY);
 733         } else if (lustre_msg_get_conn_cnt(req->rq_reqmsg) == 1) {
 734                 CERROR("%s: NID %s (%s) reconnected with 1 conn_cnt; "
 735                        "cookies not random?\n", target->obd_name,
 736                        libcfs_nid2str(req->rq_peer.nid), cluuid.uuid);
 737                 GOTO(out, rc = -EALREADY);
 738         } else {
 739                 OBD_FAIL_TIMEOUT(OBD_FAIL_TGT_DELAY_RECONNECT, 2 * obd_timeout);
 740         }
 741
 742         /* We want to handle EALREADY but *not* -EALREADY from
 743          * target_handle_reconnect(), return reconnection state in a flag */
 744         if (rc == EALREADY) {
 745                 lustre_msg_add_op_flags(req->rq_repmsg, MSG_CONNECT_RECONNECT);
 746                 rc = 0;
 747         } else if (rc) {
 748                 GOTO(out, rc);
 749         }
 750
 751         /* Tell the client if we're in recovery. */
 752         /* If this is the first client, start the recovery timer */
 753         if (target->obd_recovering) {
 754                 lustre_msg_add_op_flags(req->rq_repmsg, MSG_CONNECT_RECOVERING);
 755                 target_start_recovery_timer(target, handler);
 756         }
 757
 758         /* Tell the client if we support replayable requests */
 759         if (target->obd_replayable)
 760                 lustre_msg_add_op_flags(req->rq_repmsg, MSG_CONNECT_REPLAYABLE);
 761
 762         if (export == NULL) {
 763                 if (target->obd_recovering) {
 764                         CERROR("%s: denying connection for new client %s (%s): "
 765                                "%d clients in recovery for %lds\n",
 766                                target->obd_name,
 767                                libcfs_nid2str(req->rq_peer.nid), cluuid.uuid,
 768                                target->obd_recoverable_clients,
 769                                cfs_duration_sec(cfs_time_sub(cfs_timer_deadline(&target->obd_recovery_timer),
 770                                                              cfs_time_current())));
 771                         rc = -EBUSY;
 772                 } else {
 773  dont_check_exports:
 774                         rc = obd_connect(&conn, target, &cluuid, data);
 775                 }
 776         } else {
 777                 rc = obd_reconnect(export, target, &cluuid, data);
 778         }
 779
 780         if (rc)
 781                 GOTO(out, rc);
 782
 783         /* Return only the parts of obd_connect_data that we understand, so the
 784          * client knows that we don't understand the rest. */
 785         if (data)
 786                 memcpy(lustre_msg_buf(req->rq_repmsg, REPLY_REC_OFF,
 787                                       sizeof(*data)),
 788                        data, sizeof(*data));
 789
 790         /* If all else goes well, this is our RPC return code. */
 791         req->rq_status = 0;
 792
 793         lustre_msg_set_handle(req->rq_repmsg, &conn);
 794
 795         /* ownership of this export ref transfers to the request AFTER we
 796          * drop any previous reference the request had, but we don't want
 797          * that to go to zero before we get our new export reference. */
 798         export = class_conn2export(&conn);
 799         if (!export) {
 800                 DEBUG_REQ(D_ERROR, req, "Missing export!\n");
 801                 GOTO(out, rc = -ENODEV);
 802         }
 803
 804         /* If the client and the server are the same node, we will already
 805          * have an export that really points to the client's DLM export,
 806          * because we have a shared handles table.
 807          *
 808          * XXX this will go away when shaver stops sending the "connect" handle
 809          * in the real "remote handle" field of the request --phik 24 Apr 2003
 810          */
 811         if (req->rq_export != NULL)
 812                 class_export_put(req->rq_export);
 813
 814         req->rq_export = export;
 815
 816         spin_lock(&export->exp_lock);
 817         if (export->exp_conn_cnt >= lustre_msg_get_conn_cnt(req->rq_reqmsg)) {
 818                 CERROR("%s: %s already connected at higher conn_cnt: %d > %d\n",
 819                        cluuid.uuid, libcfs_nid2str(req->rq_peer.nid),
 820                        export->exp_conn_cnt,
 821                        lustre_msg_get_conn_cnt(req->rq_reqmsg));
 822
 823                 spin_unlock(&export->exp_lock);
 824                 GOTO(out, rc = -EALREADY);
 825         }
 826         export->exp_conn_cnt = lustre_msg_get_conn_cnt(req->rq_reqmsg);
 827
 828         /* request from liblustre?  Don't evict it for not pinging. */
 829         if (lustre_msg_get_op_flags(req->rq_reqmsg) & MSG_CONNECT_LIBCLIENT) {
 830                 export->exp_libclient = 1;
 831                 spin_unlock(&export->exp_lock);
 832
 833                 spin_lock(&target->obd_dev_lock);
 834                 list_del_init(&export->exp_obd_chain_timed);
 835                 spin_unlock(&target->obd_dev_lock);
 836         } else {
 837                 spin_unlock(&export->exp_lock);
 838         }
 839
 840         if (export->exp_connection != NULL)
 841                 ptlrpc_put_connection(export->exp_connection);
 842         export->exp_connection = ptlrpc_get_connection(req->rq_peer,
 843                                                        req->rq_self,
 844                                                        &remote_uuid);
 845
 846         if (lustre_msg_get_op_flags(req->rq_repmsg) & MSG_CONNECT_RECONNECT)
 847                 GOTO(out, rc = 0);
 848
 849         if (target->obd_recovering)
 850                 target->obd_connected_clients++;
 851
 852         memcpy(&conn,
 853                lustre_msg_buf(req->rq_reqmsg, REQ_REC_OFF + 2, sizeof conn),
 854                sizeof conn);
 855
 856         if (export->exp_imp_reverse != NULL)
 857                 class_destroy_import(export->exp_imp_reverse);
 858         revimp = export->exp_imp_reverse = class_new_import(target);
 859         revimp->imp_connection = ptlrpc_connection_addref(export->exp_connection);
 860         revimp->imp_client = &export->exp_obd->obd_ldlm_client;
 861         revimp->imp_remote_handle = conn;
 862         revimp->imp_dlm_fake = 1;
 863         revimp->imp_state = LUSTRE_IMP_FULL;
 864
 865         if (lustre_msg_get_op_flags(req->rq_reqmsg) & MSG_CONNECT_NEXT_VER) {
 866                 revimp->imp_msg_magic = LUSTRE_MSG_MAGIC_V2;
 867                 lustre_msg_add_op_flags(req->rq_repmsg, MSG_CONNECT_NEXT_VER);
 868         }
 869
 870         class_import_put(revimp);
 871 out:
 872         if (export) {
 873                 spin_lock(&export->exp_lock);
 874                 export->exp_connecting = 0;
 875                 spin_unlock(&export->exp_lock);
 876         }
 877         if (targref)
 878                 class_decref(targref);
 879         if (rc)
 880                 req->rq_status = rc;
 881         RETURN(rc);
 882 }
 883
 884 int target_handle_disconnect(struct ptlrpc_request *req)
 885 {
 886         int rc;
 887         ENTRY;
 888
 889         rc = lustre_pack_reply(req, 1, NULL, NULL);
 890         if (rc)
 891                 RETURN(rc);
 892
 893         /* keep the rq_export around so we can send the reply */
 894         req->rq_status = obd_disconnect(class_export_get(req->rq_export));
 895         RETURN(0);
 896 }
 897
 898 void target_destroy_export(struct obd_export *exp)
 899 {
 900         /* exports created from last_rcvd data, and "fake"
 901            exports created by lctl don't have an import */
 902         if (exp->exp_imp_reverse != NULL)
 903                 class_destroy_import(exp->exp_imp_reverse);
 904
 905         /* We cancel locks at disconnect time, but this will catch any locks
 906          * granted in a race with recovery-induced disconnect. */
 907         if (exp->exp_obd->obd_namespace != NULL)
 908                 ldlm_cancel_locks_for_export(exp);
 909 }
 910
 911 /*
 912  * Recovery functions
 913  */
 914
 915
 916 static void target_release_saved_req(struct ptlrpc_request *req)
 917 {
 918         if (req->rq_reply_state != NULL) {
 919                 ptlrpc_rs_decref(req->rq_reply_state);
 920                 /* req->rq_reply_state = NULL; */
 921         }
 922
 923         class_export_put(req->rq_export);
 924         OBD_FREE(req->rq_reqmsg, req->rq_reqlen);
 925         OBD_FREE(req, sizeof *req);
 926 }
 927
 928 static void target_finish_recovery(struct obd_device *obd)
 929 {
 930         struct list_head *tmp, *n;
 931
 932         LCONSOLE_INFO("%s: sending delayed replies to recovered clients\n",
 933                       obd->obd_name);
 934
 935         ldlm_reprocess_all_ns(obd->obd_namespace);
 936
 937         /* when recovery finished, cleanup orphans on mds and ost */
 938         if (OBT(obd) && OBP(obd, postrecov)) {
 939                 int rc = OBP(obd, postrecov)(obd);
 940                 LCONSOLE_WARN("%s: recovery %s: rc %d\n", obd->obd_name,
 941                               rc < 0 ? "failed" : "complete", rc);
 942         }
 943
 944         list_for_each_safe(tmp, n, &obd->obd_delayed_reply_queue) {
 945                 struct ptlrpc_request *req;
 946                 req = list_entry(tmp, struct ptlrpc_request, rq_list);
 947                 list_del(&req->rq_list);
 948                 DEBUG_REQ(D_HA, req, "delayed:");
 949                 ptlrpc_reply(req);
 950                 target_release_saved_req(req);
 951         }
 952         obd->obd_recovery_end = CURRENT_SECONDS;
 953 }
 954
 955 static void abort_recovery_queue(struct obd_device *obd)
 956 {
 957         struct ptlrpc_request *req;
 958         struct list_head *tmp, *n;
 959         int rc;
 960
 961         list_for_each_safe(tmp, n, &obd->obd_recovery_queue) {
 962                 req = list_entry(tmp, struct ptlrpc_request, rq_list);
 963                 list_del(&req->rq_list);
 964                 DEBUG_REQ(D_ERROR, req, "aborted:");
 965                 req->rq_status = -ENOTCONN;
 966                 req->rq_type = PTL_RPC_MSG_ERR;
 967                 rc = lustre_pack_reply(req, 1, NULL, NULL);
 968                 if (rc == 0) {
 969                         ptlrpc_reply(req);
 970                 } else {
 971                         DEBUG_REQ(D_ERROR, req,
 972                                   "packing failed for abort-reply; skipping");
 973                 }
 974                 target_release_saved_req(req);
 975         }
 976 }
 977
 978 /* Called from a cleanup function if the device is being cleaned up
 979    forcefully.  The exports should all have been disconnected already,
 980    the only thing left to do is
 981      - clear the recovery flags
 982      - cancel the timer
 983      - free queued requests and replies, but don't send replies
 984    Because the obd_stopping flag is set, no new requests should be received.
 985
 986 */
 987 void target_cleanup_recovery(struct obd_device *obd)
 988 {
 989         struct list_head *tmp, *n;
 990         struct ptlrpc_request *req;
 991         ENTRY;
 992
 993         LASSERT(obd->obd_stopping);
 994
 995         spin_lock_bh(&obd->obd_processing_task_lock);
 996         if (!obd->obd_recovering) {
 997                 spin_unlock_bh(&obd->obd_processing_task_lock);
 998                 EXIT;
 999                 return;
1000         }
1001         obd->obd_recovering = obd->obd_abort_recovery = 0;
1002         target_cancel_recovery_timer(obd);
1003         spin_unlock_bh(&obd->obd_processing_task_lock);
1004
1005         list_for_each_safe(tmp, n, &obd->obd_delayed_reply_queue) {
1006                 req = list_entry(tmp, struct ptlrpc_request, rq_list);
1007                 list_del(&req->rq_list);
1008                 target_release_saved_req(req);
1009         }
1010
1011         list_for_each_safe(tmp, n, &obd->obd_recovery_queue) {
1012                 req = list_entry(tmp, struct ptlrpc_request, rq_list);
1013                 list_del(&req->rq_list);
1014                 target_release_saved_req(req);
1015         }
1016         EXIT;
1017 }
1018
1019 void target_abort_recovery(void *data)
1020 {
1021         struct obd_device *obd = data;
1022
1023         ENTRY;
1024         spin_lock_bh(&obd->obd_processing_task_lock);
1025         if (!obd->obd_recovering) {
1026                 spin_unlock_bh(&obd->obd_processing_task_lock);
1027                 EXIT;
1028                 return;
1029         }
1030         obd->obd_recovering = obd->obd_abort_recovery = 0;
1031         obd->obd_recoverable_clients = 0;
1032         target_cancel_recovery_timer(obd);
1033         spin_unlock_bh(&obd->obd_processing_task_lock);
1034
1035         LCONSOLE_WARN("%s: recovery period over; disconnecting unfinished "
1036                       "clients.\n", obd->obd_name);
1037         class_disconnect_stale_exports(obd);
1038         abort_recovery_queue(obd);
1039
1040         target_finish_recovery(obd);
1041         CDEBUG(D_HA, "%s: recovery complete\n", obd_uuid2str(&obd->obd_uuid));
1042         EXIT;
1043 }
1044
1045 static void target_recovery_expired(unsigned long castmeharder)
1046 {
1047         struct obd_device *obd = (struct obd_device *)castmeharder;
1048         CERROR("%s: recovery timed out, aborting\n", obd->obd_name);
1049         spin_lock_bh(&obd->obd_processing_task_lock);
1050         if (obd->obd_recovering)
1051                 obd->obd_abort_recovery = 1;
1052         cfs_waitq_signal(&obd->obd_next_transno_waitq);
1053         spin_unlock_bh(&obd->obd_processing_task_lock);
1054 }
1055
1056
1057 /* obd_processing_task_lock should be held */
1058 void target_cancel_recovery_timer(struct obd_device *obd)
1059 {
1060         CDEBUG(D_HA, "%s: cancel recovery timer\n", obd->obd_name);
1061         cfs_timer_disarm(&obd->obd_recovery_timer);
1062 }
1063
1064 static void reset_recovery_timer(struct obd_device *obd)
1065 {
1066         spin_lock_bh(&obd->obd_processing_task_lock);
1067         if (!obd->obd_recovering) {
1068                 spin_unlock_bh(&obd->obd_processing_task_lock);
1069                 return;
1070         }
1071         cfs_timer_arm(&obd->obd_recovery_timer,
1072                       cfs_time_shift(OBD_RECOVERY_TIMEOUT));
1073         spin_unlock_bh(&obd->obd_processing_task_lock);
1074         CDEBUG(D_HA, "%s: timer will expire in %u seconds\n", obd->obd_name,
1075                OBD_RECOVERY_TIMEOUT);
1076         /* Only used for lprocfs_status */
1077         obd->obd_recovery_end = CURRENT_SECONDS + OBD_RECOVERY_TIMEOUT;
1078 }
1079
1080
1081 /* Only start it the first time called */
1082 void target_start_recovery_timer(struct obd_device *obd, svc_handler_t handler)
1083 {
1084         spin_lock_bh(&obd->obd_processing_task_lock);
1085         if (obd->obd_recovery_handler) {
1086                 spin_unlock_bh(&obd->obd_processing_task_lock);
1087                 return;
1088         }
1089         CWARN("%s: starting recovery timer (%us)\n", obd->obd_name,
1090               OBD_RECOVERY_TIMEOUT);
1091         obd->obd_recovery_handler = handler;
1092         cfs_timer_init(&obd->obd_recovery_timer, target_recovery_expired, obd);
1093         spin_unlock_bh(&obd->obd_processing_task_lock);
1094
1095         reset_recovery_timer(obd);
1096 }
1097
1098 static int check_for_next_transno(struct obd_device *obd)
1099 {
1100         struct ptlrpc_request *req;
1101         int wake_up = 0, connected, completed, queue_len, max;
1102         __u64 next_transno, req_transno;
1103
1104         spin_lock_bh(&obd->obd_processing_task_lock);
1105         req = list_entry(obd->obd_recovery_queue.next,
1106                          struct ptlrpc_request, rq_list);
1107         max = obd->obd_max_recoverable_clients;
1108         req_transno = lustre_msg_get_transno(req->rq_reqmsg);
1109         connected = obd->obd_connected_clients;
1110         completed = max - obd->obd_recoverable_clients;
1111         queue_len = obd->obd_requests_queued_for_recovery;
1112         next_transno = obd->obd_next_recovery_transno;
1113
1114         CDEBUG(D_HA,"max: %d, connected: %d, completed: %d, queue_len: %d, "
1115                "req_transno: "LPU64", next_transno: "LPU64"\n",
1116                max, connected, completed, queue_len, req_transno, next_transno);
1117         if (obd->obd_abort_recovery) {
1118                 CDEBUG(D_HA, "waking for aborted recovery\n");
1119                 wake_up = 1;
1120         } else if (!obd->obd_recovering) {
1121                 CDEBUG(D_HA, "waking for completed recovery (?)\n");
1122                 wake_up = 1;
1123         } else if (req_transno == next_transno) {
1124                 CDEBUG(D_HA, "waking for next ("LPD64")\n", next_transno);
1125                 wake_up = 1;
1126         } else if (queue_len + completed == max) {
1127                 CDEBUG(D_ERROR,
1128                        "waking for skipped transno (skip: "LPD64
1129                        ", ql: %d, comp: %d, conn: %d, next: "LPD64")\n",
1130                        next_transno, queue_len, completed, max, req_transno);
1131                 obd->obd_next_recovery_transno = req_transno;
1132                 wake_up = 1;
1133         }
1134         spin_unlock_bh(&obd->obd_processing_task_lock);
1135         LASSERT(lustre_msg_get_transno(req->rq_reqmsg) >= next_transno);
1136         return wake_up;
1137 }
1138
1139 static void process_recovery_queue(struct obd_device *obd)
1140 {
1141         struct ptlrpc_request *req;
1142         int abort_recovery = 0;
1143         struct l_wait_info lwi = { 0 };
1144         ENTRY;
1145
1146         for (;;) {
1147                 spin_lock_bh(&obd->obd_processing_task_lock);
1148                 LASSERT(obd->obd_processing_task == cfs_curproc_pid());
1149                 req = list_entry(obd->obd_recovery_queue.next,
1150                                  struct ptlrpc_request, rq_list);
1151
1152                 if (lustre_msg_get_transno(req->rq_reqmsg) !=
1153                     obd->obd_next_recovery_transno) {
1154                         spin_unlock_bh(&obd->obd_processing_task_lock);
1155                         CDEBUG(D_HA, "Waiting for transno "LPD64" (1st is "
1156                                LPD64")\n",
1157                                obd->obd_next_recovery_transno,
1158                                lustre_msg_get_transno(req->rq_reqmsg));
1159                         l_wait_event(obd->obd_next_transno_waitq,
1160                                      check_for_next_transno(obd), &lwi);
1161                         spin_lock_bh(&obd->obd_processing_task_lock);
1162                         abort_recovery = obd->obd_abort_recovery;
1163                         spin_unlock_bh(&obd->obd_processing_task_lock);
1164                         if (abort_recovery) {
1165                                 target_abort_recovery(obd);
1166                                 return;
1167                         }
1168                         continue;
1169                 }
1170                 list_del_init(&req->rq_list);
1171                 obd->obd_requests_queued_for_recovery--;
1172                 spin_unlock_bh(&obd->obd_processing_task_lock);
1173
1174                 DEBUG_REQ(D_HA, req, "processing: ");
1175                 (void)obd->obd_recovery_handler(req);
1176                 obd->obd_replayed_requests++;
1177                 reset_recovery_timer(obd);
1178                 /* bug 1580: decide how to properly sync() in recovery */
1179                 //mds_fsync_super(obd->u.obt.obt_sb);
1180                 class_export_put(req->rq_export);
1181                 if (req->rq_reply_state != NULL) {
1182                         ptlrpc_rs_decref(req->rq_reply_state);
1183                         /* req->rq_reply_state = NULL; */
1184                 }
1185                 OBD_FREE(req->rq_reqmsg, req->rq_reqlen);
1186                 OBD_FREE(req, sizeof *req);
1187                 spin_lock_bh(&obd->obd_processing_task_lock);
1188                 obd->obd_next_recovery_transno++;
1189                 if (list_empty(&obd->obd_recovery_queue)) {
1190                         obd->obd_processing_task = 0;
1191                         spin_unlock_bh(&obd->obd_processing_task_lock);
1192                         break;
1193                 }
1194                 spin_unlock_bh(&obd->obd_processing_task_lock);
1195         }
1196         EXIT;
1197 }
1198
1199 int target_queue_recovery_request(struct ptlrpc_request *req,
1200                                   struct obd_device *obd)
1201 {
1202         struct list_head *tmp;
1203         int inserted = 0;
1204         __u64 transno = lustre_msg_get_transno(req->rq_reqmsg);
1205         struct ptlrpc_request *saved_req;
1206         struct lustre_msg *reqmsg;
1207
1208         /* CAVEAT EMPTOR: The incoming request message has been swabbed
1209          * (i.e. buflens etc are in my own byte order), but type-dependent
1210          * buffers (eg mds_body, ost_body etc) have NOT been swabbed. */
1211
1212         if (!transno) {
1213                 CFS_INIT_LIST_HEAD(&req->rq_list);
1214                 DEBUG_REQ(D_HA, req, "not queueing");
1215                 return 1;
1216         }
1217
1218         /* XXX If I were a real man, these LBUGs would be sane cleanups. */
1219         /* XXX just like the request-dup code in queue_final_reply */
1220         OBD_ALLOC(saved_req, sizeof *saved_req);
1221         if (!saved_req)
1222                 LBUG();
1223         OBD_ALLOC(reqmsg, req->rq_reqlen);
1224         if (!reqmsg)
1225                 LBUG();
1226
1227         spin_lock_bh(&obd->obd_processing_task_lock);
1228
1229         /* If we're processing the queue, we want don't want to queue this
1230          * message.
1231          *
1232          * Also, if this request has a transno less than the one we're waiting
1233          * for, we should process it now.  It could (and currently always will)
1234          * be an open request for a descriptor that was opened some time ago.
1235          *
1236          * Also, a resent, replayed request that has already been
1237          * handled will pass through here and be processed immediately.
1238          */
1239         if (obd->obd_processing_task == cfs_curproc_pid() ||
1240             transno < obd->obd_next_recovery_transno) {
1241                 /* Processing the queue right now, don't re-add. */
1242                 LASSERT(list_empty(&req->rq_list));
1243                 spin_unlock_bh(&obd->obd_processing_task_lock);
1244                 OBD_FREE(reqmsg, req->rq_reqlen);
1245                 OBD_FREE(saved_req, sizeof *saved_req);
1246                 return 1;
1247         }
1248
1249         /* A resent, replayed request that is still on the queue; just drop it.
1250            The queued request will handle this. */
1251         if ((lustre_msg_get_flags(req->rq_reqmsg) & (MSG_RESENT|MSG_REPLAY)) ==
1252             (MSG_RESENT | MSG_REPLAY)) {
1253                 DEBUG_REQ(D_ERROR, req, "dropping resent queued req");
1254                 spin_unlock_bh(&obd->obd_processing_task_lock);
1255                 OBD_FREE(reqmsg, req->rq_reqlen);
1256                 OBD_FREE(saved_req, sizeof *saved_req);
1257                 return 0;
1258         }
1259
1260         memcpy(saved_req, req, sizeof *req);
1261         memcpy(reqmsg, req->rq_reqmsg, req->rq_reqlen);
1262         req = saved_req;
1263         req->rq_reqmsg = reqmsg;
1264         class_export_get(req->rq_export);
1265         CFS_INIT_LIST_HEAD(&req->rq_list);
1266
1267         /* XXX O(n^2) */
1268         list_for_each(tmp, &obd->obd_recovery_queue) {
1269                 struct ptlrpc_request *reqiter =
1270                         list_entry(tmp, struct ptlrpc_request, rq_list);
1271
1272                 if (lustre_msg_get_transno(reqiter->rq_reqmsg) > transno) {
1273                         list_add_tail(&req->rq_list, &reqiter->rq_list);
1274                         inserted = 1;
1275                         break;
1276                 }
1277         }
1278
1279         if (!inserted) {
1280                 list_add_tail(&req->rq_list, &obd->obd_recovery_queue);
1281         }
1282
1283         obd->obd_requests_queued_for_recovery++;
1284
1285         if (obd->obd_processing_task != 0) {
1286                 /* Someone else is processing this queue, we'll leave it to
1287                  * them.
1288                  */
1289                 cfs_waitq_signal(&obd->obd_next_transno_waitq);
1290                 spin_unlock_bh(&obd->obd_processing_task_lock);
1291                 return 0;
1292         }
1293
1294         /* Nobody is processing, and we know there's (at least) one to process
1295          * now, so we'll do the honours.
1296          */
1297         obd->obd_processing_task = cfs_curproc_pid();
1298         spin_unlock_bh(&obd->obd_processing_task_lock);
1299
1300         process_recovery_queue(obd);
1301         return 0;
1302 }
1303
1304 struct obd_device * target_req2obd(struct ptlrpc_request *req)
1305 {
1306         return req->rq_export->exp_obd;
1307 }
1308
1309 int target_queue_final_reply(struct ptlrpc_request *req, int rc)
1310 {
1311         struct obd_device *obd = target_req2obd(req);
1312         struct ptlrpc_request *saved_req;
1313         struct lustre_msg *reqmsg;
1314         int recovery_done = 0;
1315
1316         LASSERT ((rc == 0) == (req->rq_reply_state != NULL));
1317
1318         if (rc) {
1319                 /* Just like ptlrpc_error, but without the sending. */
1320                 rc = lustre_pack_reply(req, 1, NULL, NULL);
1321                 LASSERT(rc == 0); /* XXX handle this */
1322                 req->rq_type = PTL_RPC_MSG_ERR;
1323         }
1324
1325         LASSERT (!req->rq_reply_state->rs_difficult);
1326         LASSERT(list_empty(&req->rq_list));
1327         /* XXX a bit like the request-dup code in queue_recovery_request */
1328         OBD_ALLOC(saved_req, sizeof *saved_req);
1329         if (!saved_req)
1330                 LBUG();
1331         OBD_ALLOC(reqmsg, req->rq_reqlen);
1332         if (!reqmsg)
1333                 LBUG();
1334         *saved_req = *req;
1335         memcpy(reqmsg, req->rq_reqmsg, req->rq_reqlen);
1336
1337         /* Don't race cleanup */
1338         spin_lock_bh(&obd->obd_processing_task_lock);
1339         if (obd->obd_stopping) {
1340                 spin_unlock_bh(&obd->obd_processing_task_lock);
1341                 OBD_FREE(reqmsg, req->rq_reqlen);
1342                 OBD_FREE(saved_req, sizeof *req);
1343                 req->rq_status = -ENOTCONN;
1344                 /* rv is ignored anyhow */
1345                 return -ENOTCONN;
1346         }
1347         ptlrpc_rs_addref(req->rq_reply_state);  /* +1 ref for saved reply */
1348         req = saved_req;
1349         req->rq_reqmsg = reqmsg;
1350         class_export_get(req->rq_export);
1351         list_add(&req->rq_list, &obd->obd_delayed_reply_queue);
1352
1353         /* only count the first "replay over" request from each
1354            export */
1355         if (req->rq_export->exp_replay_needed) {
1356                 --obd->obd_recoverable_clients;
1357
1358                 spin_lock(&req->rq_export->exp_lock);
1359                 req->rq_export->exp_replay_needed = 0;
1360                 spin_unlock(&req->rq_export->exp_lock);
1361         }
1362         recovery_done = (obd->obd_recoverable_clients == 0);
1363         spin_unlock_bh(&obd->obd_processing_task_lock);
1364
1365         OBD_RACE(OBD_FAIL_LDLM_RECOV_CLIENTS);
1366         if (recovery_done) {
1367                 spin_lock_bh(&obd->obd_processing_task_lock);
1368                 obd->obd_recovering = obd->obd_abort_recovery = 0;
1369                 target_cancel_recovery_timer(obd);
1370                 spin_unlock_bh(&obd->obd_processing_task_lock);
1371
1372                 target_finish_recovery(obd);
1373                 CDEBUG(D_HA, "%s: recovery complete\n",
1374                        obd_uuid2str(&obd->obd_uuid));
1375         } else {
1376                 CWARN("%s: %d recoverable clients remain\n",
1377                        obd->obd_name, obd->obd_recoverable_clients);
1378                 cfs_waitq_signal(&obd->obd_next_transno_waitq);
1379         }
1380
1381         return 1;
1382 }
1383
1384 int
1385 target_send_reply_msg (struct ptlrpc_request *req, int rc, int fail_id)
1386 {
1387         if (OBD_FAIL_CHECK(fail_id | OBD_FAIL_ONCE)) {
1388                 obd_fail_loc |= OBD_FAIL_ONCE | OBD_FAILED;
1389                 DEBUG_REQ(D_ERROR, req, "dropping reply");
1390                 return (-ECOMM);
1391         }
1392
1393         if (rc) {
1394                 DEBUG_REQ(D_ERROR, req, "processing error (%d)", rc);
1395                 req->rq_status = rc;
1396                 return (ptlrpc_error(req));
1397         } else {
1398                 DEBUG_REQ(D_NET, req, "sending reply");
1399         }
1400
1401         return (ptlrpc_send_reply(req, 1));
1402 }
1403
1404 void
1405 target_send_reply(struct ptlrpc_request *req, int rc, int fail_id)
1406 {
1407         int                        netrc;
1408         struct ptlrpc_reply_state *rs;
1409         struct obd_device         *obd;
1410         struct obd_export         *exp;
1411         struct ptlrpc_service     *svc;
1412
1413         svc = req->rq_rqbd->rqbd_service;
1414         rs = req->rq_reply_state;
1415         if (rs == NULL || !rs->rs_difficult) {
1416                 /* no notifiers */
1417                 target_send_reply_msg (req, rc, fail_id);
1418                 return;
1419         }
1420
1421         /* must be an export if locks saved */
1422         LASSERT (req->rq_export != NULL);
1423         /* req/reply consistent */
1424         LASSERT (rs->rs_service == svc);
1425
1426         /* "fresh" reply */
1427         LASSERT (!rs->rs_scheduled);
1428         LASSERT (!rs->rs_scheduled_ever);
1429         LASSERT (!rs->rs_handled);
1430         LASSERT (!rs->rs_on_net);
1431         LASSERT (rs->rs_export == NULL);
1432         LASSERT (list_empty(&rs->rs_obd_list));
1433         LASSERT (list_empty(&rs->rs_exp_list));
1434
1435         exp = class_export_get (req->rq_export);
1436         obd = exp->exp_obd;
1437
1438         /* disable reply scheduling onto srv_reply_queue while I'm setting up */
1439         rs->rs_scheduled = 1;
1440         rs->rs_on_net    = 1;
1441         rs->rs_xid       = req->rq_xid;
1442         rs->rs_transno   = req->rq_transno;
1443         rs->rs_export    = exp;
1444
1445         spin_lock(&obd->obd_uncommitted_replies_lock);
1446
1447         if (rs->rs_transno > obd->obd_last_committed) {
1448                 /* not committed already */
1449                 list_add_tail (&rs->rs_obd_list,
1450                                &obd->obd_uncommitted_replies);
1451         }
1452
1453         spin_unlock (&obd->obd_uncommitted_replies_lock);
1454         spin_lock (&exp->exp_lock);
1455
1456         list_add_tail (&rs->rs_exp_list, &exp->exp_outstanding_replies);
1457
1458         spin_unlock(&exp->exp_lock);
1459
1460         netrc = target_send_reply_msg (req, rc, fail_id);
1461
1462         spin_lock(&svc->srv_lock);
1463
1464         svc->srv_n_difficult_replies++;
1465
1466         if (netrc != 0) {
1467                 /* error sending: reply is off the net.  Also we need +1
1468                  * reply ref until ptlrpc_server_handle_reply() is done
1469                  * with the reply state (if the send was successful, there
1470                  * would have been +1 ref for the net, which
1471                  * reply_out_callback leaves alone) */
1472                 rs->rs_on_net = 0;
1473                 ptlrpc_rs_addref(rs);
1474                 atomic_inc (&svc->srv_outstanding_replies);
1475         }
1476
1477         if (!rs->rs_on_net ||                   /* some notifier */
1478             list_empty(&rs->rs_exp_list) ||     /* completed already */
1479             list_empty(&rs->rs_obd_list)) {
1480                 list_add_tail (&rs->rs_list, &svc->srv_reply_queue);
1481                 cfs_waitq_signal (&svc->srv_waitq);
1482         } else {
1483                 list_add (&rs->rs_list, &svc->srv_active_replies);
1484                 rs->rs_scheduled = 0;           /* allow notifier to schedule */
1485         }
1486
1487         spin_unlock(&svc->srv_lock);
1488 }
1489
1490 int target_handle_ping(struct ptlrpc_request *req)
1491 {
1492         obd_ping(req->rq_export);
1493         return lustre_pack_reply(req, 1, NULL, NULL);
1494 }
1495
1496 void target_committed_to_req(struct ptlrpc_request *req)
1497 {
1498         struct obd_device *obd = req->rq_export->exp_obd;
1499
1500         if (!obd->obd_no_transno && req->rq_repmsg != NULL)
1501                 lustre_msg_set_last_committed(req->rq_repmsg,
1502                                               obd->obd_last_committed);
1503         else
1504                 DEBUG_REQ(D_IOCTL, req, "not sending last_committed update");
1505
1506         CDEBUG(D_INFO, "last_committed "LPU64", xid "LPU64"\n",
1507                obd->obd_last_committed, req->rq_xid);
1508 }
1509
1510 EXPORT_SYMBOL(target_committed_to_req);
1511
1512 #ifdef HAVE_QUOTA_SUPPORT
1513 int target_handle_qc_callback(struct ptlrpc_request *req)
1514 {
1515         struct obd_quotactl *oqctl;
1516         struct client_obd *cli = &req->rq_export->exp_obd->u.cli;
1517
1518         oqctl = lustre_swab_reqbuf(req, REQ_REC_OFF, sizeof(*oqctl),
1519                                    lustre_swab_obd_quotactl);
1520         if (oqctl == NULL) {
1521                 CERROR("Can't unpack obd_quotactl\n");
1522                 RETURN(-EPROTO);
1523         }
1524
1525         cli->cl_qchk_stat = oqctl->qc_stat;
1526
1527         return 0;
1528 }
1529
1530 int target_handle_dqacq_callback(struct ptlrpc_request *req)
1531 {
1532 #ifdef __KERNEL__
1533         struct obd_device *obd = req->rq_export->exp_obd;
1534         struct obd_device *master_obd;
1535         struct lustre_quota_ctxt *qctxt;
1536         struct qunit_data *qdata;
1537         void* rep;
1538         struct qunit_data_old *qdata_old;
1539         int rc = 0;
1540         int repsize[2] = { sizeof(struct ptlrpc_body),
1541                            sizeof(struct qunit_data) };
1542         ENTRY;
1543
1544         rc = lustre_pack_reply(req, 2, repsize, NULL);
1545         if (rc) {
1546                 CERROR("packing reply failed!: rc = %d\n", rc);
1547                 RETURN(rc);
1548         }
1549         LASSERT(req->rq_export);
1550
1551         /* fixed for bug10707 */
1552         if ((req->rq_export->exp_connect_flags & OBD_CONNECT_QUOTA64) &&
1553             !OBD_FAIL_CHECK(OBD_FAIL_QUOTA_QD_COUNT_32BIT)) {
1554                 CDEBUG(D_QUOTA, "qd_count is 64bit!\n");
1555                 rep = lustre_msg_buf(req->rq_repmsg, REPLY_REC_OFF,
1556                                      sizeof(struct qunit_data));
1557                 LASSERT(rep);
1558                 qdata = lustre_swab_reqbuf(req, REQ_REC_OFF, sizeof(*qdata),
1559                                            lustre_swab_qdata);
1560         } else {
1561                 CDEBUG(D_QUOTA, "qd_count is 32bit!\n");
1562                 rep = lustre_msg_buf(req->rq_repmsg, REPLY_REC_OFF,
1563                                      sizeof(struct qunit_data_old));
1564                 LASSERT(rep);
1565                 qdata_old = lustre_swab_reqbuf(req, REQ_REC_OFF, sizeof(*qdata_old),
1566                                                lustre_swab_qdata_old);
1567                 qdata = lustre_quota_old_to_new(qdata_old);
1568         }
1569
1570         if (qdata == NULL) {
1571                 CERROR("Can't unpack qunit_data\n");
1572                 RETURN(-EPROTO);
1573         }
1574
1575         /* we use the observer */
1576         LASSERT(obd->obd_observer && obd->obd_observer->obd_observer);
1577         master_obd = obd->obd_observer->obd_observer;
1578         qctxt = &master_obd->u.obt.obt_qctxt;
1579
1580         LASSERT(qctxt->lqc_handler);
1581         rc = qctxt->lqc_handler(master_obd, qdata,
1582                                 lustre_msg_get_opc(req->rq_reqmsg));
1583         if (rc && rc != -EDQUOT)
1584                 CDEBUG(rc == -EBUSY  ? D_QUOTA : D_ERROR,
1585                        "dqacq failed! (rc:%d)\n", rc);
1586
1587         /* the qd_count might be changed in lqc_handler */
1588         if ((req->rq_export->exp_connect_flags & OBD_CONNECT_QUOTA64) &&
1589             !OBD_FAIL_CHECK(OBD_FAIL_QUOTA_QD_COUNT_32BIT)) {
1590                 memcpy(rep,qdata,sizeof(*qdata));
1591         } else {
1592                 qdata_old = lustre_quota_new_to_old(qdata);
1593                 memcpy(rep,qdata_old,sizeof(*qdata_old));
1594         }
1595         req->rq_status = rc;
1596         rc = ptlrpc_reply(req);
1597
1598         RETURN(rc);
1599 #else
1600         return 0;
1601 #endif /* !__KERNEL__ */
1602 }
1603 #endif /* HAVE_QUOTA_SUPPORT */
1604
1605 ldlm_mode_t lck_compat_array[] = {
1606         [LCK_EX] LCK_COMPAT_EX,
1607         [LCK_PW] LCK_COMPAT_PW,
1608         [LCK_PR] LCK_COMPAT_PR,
1609         [LCK_CW] LCK_COMPAT_CW,
1610         [LCK_CR] LCK_COMPAT_CR,
1611         [LCK_NL] LCK_COMPAT_NL,
1612         [LCK_GROUP] LCK_COMPAT_GROUP
1613 };