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 #include <lustre_sec.h>
  40 #include <lustre_ver.h>
  41
  42 /* @priority: if non-zero, move the selected to the list head
  43  * @create: if zero, only search in existed connections
  44  */
  45 static int import_set_conn(struct obd_import *imp, struct obd_uuid *uuid,
  46                            int priority, int create)
  47 {
  48         struct ptlrpc_connection *ptlrpc_conn;
  49         struct obd_import_conn *imp_conn = NULL, *item;
  50         int rc = 0;
  51         ENTRY;
  52
  53         if (!create && !priority) {
  54                 CDEBUG(D_HA, "Nothing to do\n");
  55                 RETURN(-EINVAL);
  56         }
  57
  58         ptlrpc_conn = ptlrpc_uuid_to_connection(uuid);
  59         if (!ptlrpc_conn) {
  60                 CDEBUG(D_HA, "can't find connection %s\n", uuid->uuid);
  61                 RETURN (-ENOENT);
  62         }
  63
  64         if (create) {
  65                 OBD_ALLOC(imp_conn, sizeof(*imp_conn));
  66                 if (!imp_conn) {
  67                         GOTO(out_put, rc = -ENOMEM);
  68                 }
  69         }
  70
  71         spin_lock(&imp->imp_lock);
  72         list_for_each_entry(item, &imp->imp_conn_list, oic_item) {
  73                 if (obd_uuid_equals(uuid, &item->oic_uuid)) {
  74                         if (priority) {
  75                                 list_del(&item->oic_item);
  76                                 list_add(&item->oic_item, &imp->imp_conn_list);
  77                                 item->oic_last_attempt = 0;
  78                         }
  79                         CDEBUG(D_HA, "imp %p@%s: found existing conn %s%s\n",
  80                                imp, imp->imp_obd->obd_name, uuid->uuid,
  81                                (priority ? ", moved to head" : ""));
  82                         spin_unlock(&imp->imp_lock);
  83                         GOTO(out_free, rc = 0);
  84                 }
  85         }
  86         /* not found */
  87         if (create) {
  88                 imp_conn->oic_conn = ptlrpc_conn;
  89                 imp_conn->oic_uuid = *uuid;
  90                 item->oic_last_attempt = 0;
  91                 if (priority)
  92                         list_add(&imp_conn->oic_item, &imp->imp_conn_list);
  93                 else
  94                         list_add_tail(&imp_conn->oic_item, &imp->imp_conn_list);
  95                 CDEBUG(D_HA, "imp %p@%s: add connection %s at %s\n",
  96                        imp, imp->imp_obd->obd_name, uuid->uuid,
  97                        (priority ? "head" : "tail"));
  98         } else {
  99                 spin_unlock(&imp->imp_lock);
 100                 GOTO(out_free, rc = -ENOENT);
 101
 102         }
 103
 104         spin_unlock(&imp->imp_lock);
 105         RETURN(0);
 106 out_free:
 107         if (imp_conn)
 108                 OBD_FREE(imp_conn, sizeof(*imp_conn));
 109 out_put:
 110         ptlrpc_put_connection(ptlrpc_conn);
 111         RETURN(rc);
 112 }
 113
 114 int import_set_conn_priority(struct obd_import *imp, struct obd_uuid *uuid)
 115 {
 116         return import_set_conn(imp, uuid, 1, 0);
 117 }
 118
 119 int client_import_add_conn(struct obd_import *imp, struct obd_uuid *uuid,
 120                            int priority)
 121 {
 122         return import_set_conn(imp, uuid, priority, 1);
 123 }
 124
 125 int client_import_del_conn(struct obd_import *imp, struct obd_uuid *uuid)
 126 {
 127         struct obd_import_conn *imp_conn;
 128         struct obd_export *dlmexp;
 129         int rc = -ENOENT;
 130         ENTRY;
 131
 132         spin_lock(&imp->imp_lock);
 133         if (list_empty(&imp->imp_conn_list)) {
 134                 LASSERT(!imp->imp_connection);
 135                 GOTO(out, rc);
 136         }
 137
 138         list_for_each_entry(imp_conn, &imp->imp_conn_list, oic_item) {
 139                 if (!obd_uuid_equals(uuid, &imp_conn->oic_uuid))
 140                         continue;
 141                 LASSERT(imp_conn->oic_conn);
 142
 143                 /* is current conn? */
 144                 if (imp_conn == imp->imp_conn_current) {
 145                         LASSERT(imp_conn->oic_conn == imp->imp_connection);
 146
 147                         if (imp->imp_state != LUSTRE_IMP_CLOSED &&
 148                             imp->imp_state != LUSTRE_IMP_DISCON) {
 149                                 CERROR("can't remove current connection\n");
 150                                 GOTO(out, rc = -EBUSY);
 151                         }
 152
 153                         ptlrpc_put_connection(imp->imp_connection);
 154                         imp->imp_connection = NULL;
 155
 156                         dlmexp = class_conn2export(&imp->imp_dlm_handle);
 157                         if (dlmexp && dlmexp->exp_connection) {
 158                                 LASSERT(dlmexp->exp_connection ==
 159                                         imp_conn->oic_conn);
 160                                 ptlrpc_put_connection(dlmexp->exp_connection);
 161                                 dlmexp->exp_connection = NULL;
 162                         }
 163                 }
 164
 165                 list_del(&imp_conn->oic_item);
 166                 ptlrpc_put_connection(imp_conn->oic_conn);
 167                 OBD_FREE(imp_conn, sizeof(*imp_conn));
 168                 CDEBUG(D_HA, "imp %p@%s: remove connection %s\n",
 169                        imp, imp->imp_obd->obd_name, uuid->uuid);
 170                 rc = 0;
 171                 break;
 172         }
 173 out:
 174         spin_unlock(&imp->imp_lock);
 175         if (rc == -ENOENT)
 176                 CERROR("connection %s not found\n", uuid->uuid);
 177         RETURN(rc);
 178 }
 179
 180 static
 181 void destroy_import(struct obd_import *imp)
 182 {
 183         /* drop security policy instance after all rpc finished/aborted
 184          * to let all busy credentials be released.
 185          */
 186         class_import_get(imp);
 187         class_destroy_import(imp);
 188         sptlrpc_import_put_sec(imp);
 189         class_import_put(imp);
 190 }
 191
 192 /* configure an RPC client OBD device
 193  *
 194  * lcfg parameters:
 195  * 1 - client UUID
 196  * 2 - server UUID
 197  * 3 - inactive-on-startup
 198  */
 199 int client_obd_setup(struct obd_device *obddev, struct lustre_cfg *lcfg)
 200 {
 201         struct client_obd *cli = &obddev->u.cli;
 202         struct obd_import *imp;
 203         struct obd_uuid server_uuid;
 204         int rq_portal, rp_portal, connect_op;
 205         char *name = obddev->obd_type->typ_name;
 206         int rc;
 207         ENTRY;
 208
 209         /* In a more perfect world, we would hang a ptlrpc_client off of
 210          * obd_type and just use the values from there. */
 211         if (!strcmp(name, LUSTRE_OSC_NAME)) {
 212                 rq_portal = OST_IO_PORTAL;
 213                 rp_portal = OSC_REPLY_PORTAL;
 214                 connect_op = OST_CONNECT;
 215         } else if (!strcmp(name, LUSTRE_MDC_NAME)) {
 216                 rq_portal = MDS_REQUEST_PORTAL;
 217                 rp_portal = MDC_REPLY_PORTAL;
 218                 connect_op = MDS_CONNECT;
 219         } else if (!strcmp(name, LUSTRE_MGC_NAME)) {
 220                 rq_portal = MGS_REQUEST_PORTAL;
 221                 rp_portal = MGC_REPLY_PORTAL;
 222                 connect_op = MGS_CONNECT;
 223         } else {
 224                 CERROR("unknown client OBD type \"%s\", can't setup\n",
 225                        name);
 226                 RETURN(-EINVAL);
 227         }
 228
 229         if (LUSTRE_CFG_BUFLEN(lcfg, 1) < 1) {
 230                 CERROR("requires a TARGET UUID\n");
 231                 RETURN(-EINVAL);
 232         }
 233
 234         if (LUSTRE_CFG_BUFLEN(lcfg, 1) > 37) {
 235                 CERROR("client UUID must be less than 38 characters\n");
 236                 RETURN(-EINVAL);
 237         }
 238
 239         if (LUSTRE_CFG_BUFLEN(lcfg, 2) < 1) {
 240                 CERROR("setup requires a SERVER UUID\n");
 241                 RETURN(-EINVAL);
 242         }
 243
 244         if (LUSTRE_CFG_BUFLEN(lcfg, 2) > 37) {
 245                 CERROR("target UUID must be less than 38 characters\n");
 246                 RETURN(-EINVAL);
 247         }
 248
 249         sema_init(&cli->cl_sem, 1);
 250         sema_init(&cli->cl_mgc_sem, 1);
 251         cli->cl_sec_conf.sfc_rpc_flavor = SPTLRPC_FLVR_NULL;
 252         cli->cl_sec_conf.sfc_bulk_csum = BULK_CSUM_ALG_NULL;
 253         cli->cl_sec_conf.sfc_bulk_priv = BULK_PRIV_ALG_NULL;
 254         cli->cl_sec_conf.sfc_flags = 0;
 255         cli->cl_conn_count = 0;
 256         memcpy(server_uuid.uuid, lustre_cfg_buf(lcfg, 2),
 257                min_t(unsigned int, LUSTRE_CFG_BUFLEN(lcfg, 2),
 258                      sizeof(server_uuid)));
 259
 260         cli->cl_dirty = 0;
 261         cli->cl_avail_grant = 0;
 262         /* FIXME: should limit this for the sum of all cl_dirty_max */
 263         cli->cl_dirty_max = OSC_MAX_DIRTY_DEFAULT * 1024 * 1024;
 264         if (cli->cl_dirty_max >> PAGE_SHIFT > num_physpages / 8)
 265                 cli->cl_dirty_max = num_physpages << (PAGE_SHIFT - 3);
 266         CFS_INIT_LIST_HEAD(&cli->cl_cache_waiters);
 267         CFS_INIT_LIST_HEAD(&cli->cl_loi_ready_list);
 268         CFS_INIT_LIST_HEAD(&cli->cl_loi_write_list);
 269         CFS_INIT_LIST_HEAD(&cli->cl_loi_read_list);
 270         client_obd_list_lock_init(&cli->cl_loi_list_lock);
 271         cli->cl_r_in_flight = 0;
 272         cli->cl_w_in_flight = 0;
 273
 274         spin_lock_init(&cli->cl_read_rpc_hist.oh_lock);
 275         spin_lock_init(&cli->cl_write_rpc_hist.oh_lock);
 276         spin_lock_init(&cli->cl_read_page_hist.oh_lock);
 277         spin_lock_init(&cli->cl_write_page_hist.oh_lock);
 278         spin_lock_init(&cli->cl_read_offset_hist.oh_lock);
 279         spin_lock_init(&cli->cl_write_offset_hist.oh_lock);
 280         /*
 281             cli->cl_max_pages_per_rpc = PTLRPC_MAX_BRW_PAGES;
 282             This filed value is seted at connect time by ocd_brw_size.
 283         */
 284         if (!strcmp(name, LUSTRE_MDC_NAME)) {
 285                 cli->cl_max_rpcs_in_flight = MDC_MAX_RIF_DEFAULT;
 286         } else if (num_physpages >> (20 - PAGE_SHIFT) <= 128 /* MB */) {
 287                 cli->cl_max_rpcs_in_flight = 2;
 288         } else if (num_physpages >> (20 - PAGE_SHIFT) <= 256 /* MB */) {
 289                 cli->cl_max_rpcs_in_flight = 3;
 290         } else if (num_physpages >> (20 - PAGE_SHIFT) <= 512 /* MB */) {
 291                 cli->cl_max_rpcs_in_flight = 4;
 292         } else {
 293                 cli->cl_max_rpcs_in_flight = OSC_MAX_RIF_DEFAULT;
 294         }
 295
 296         rc = ldlm_get_ref();
 297         if (rc) {
 298                 CERROR("ldlm_get_ref failed: %d\n", rc);
 299                 GOTO(err, rc);
 300         }
 301
 302         ptlrpc_init_client(rq_portal, rp_portal, name,
 303                            &obddev->obd_ldlm_client);
 304
 305         imp = class_new_import(obddev);
 306         if (imp == NULL)
 307                 GOTO(err_ldlm, rc = -ENOENT);
 308         imp->imp_client = &obddev->obd_ldlm_client;
 309         imp->imp_connect_op = connect_op;
 310         imp->imp_initial_recov = 1;
 311         imp->imp_initial_recov_bk = 0;
 312         CFS_INIT_LIST_HEAD(&imp->imp_pinger_chain);
 313         memcpy(cli->cl_target_uuid.uuid, lustre_cfg_buf(lcfg, 1),
 314                LUSTRE_CFG_BUFLEN(lcfg, 1));
 315         class_import_put(imp);
 316
 317         rc = client_import_add_conn(imp, &server_uuid, 1);
 318         if (rc) {
 319                 CERROR("can't add initial connection\n");
 320                 GOTO(err_import, rc);
 321         }
 322
 323         cli->cl_import = imp;
 324         /* cli->cl_max_mds_{easize,cookiesize} updated by mdc_init_ea_size() */
 325         cli->cl_max_mds_easize = sizeof(struct lov_mds_md);
 326         cli->cl_max_mds_cookiesize = sizeof(struct llog_cookie);
 327
 328         if (LUSTRE_CFG_BUFLEN(lcfg, 3) > 0) {
 329                 if (!strcmp(lustre_cfg_string(lcfg, 3), "inactive")) {
 330                         CDEBUG(D_HA, "marking %s %s->%s as inactive\n",
 331                                name, obddev->obd_name,
 332                                cli->cl_target_uuid.uuid);
 333                         imp->imp_invalid = 1;
 334                 }
 335         }
 336
 337         cli->cl_qchk_stat = CL_NOT_QUOTACHECKED;
 338
 339         RETURN(rc);
 340
 341 err_import:
 342         class_destroy_import(imp);
 343 err_ldlm:
 344         ldlm_put_ref(0);
 345 err:
 346         RETURN(rc);
 347
 348 }
 349
 350 int client_obd_cleanup(struct obd_device *obddev)
 351 {
 352         ENTRY;
 353         ldlm_put_ref(obddev->obd_force);
 354
 355         RETURN(0);
 356 }
 357
 358 /* ->o_connect() method for client side (OSC and MDC and MGC) */
 359 int client_connect_import(const struct lu_env *env,
 360                           struct lustre_handle *dlm_handle,
 361                           struct obd_device *obd, struct obd_uuid *cluuid,
 362                           struct obd_connect_data *data)
 363 {
 364         struct client_obd *cli = &obd->u.cli;
 365         struct obd_import *imp = cli->cl_import;
 366         struct obd_export *exp;
 367         struct obd_connect_data *ocd;
 368         int rc;
 369         ENTRY;
 370
 371         mutex_down(&cli->cl_sem);
 372         rc = class_connect(dlm_handle, obd, cluuid);
 373         if (rc)
 374                 GOTO(out_sem, rc);
 375
 376         cli->cl_conn_count++;
 377         if (cli->cl_conn_count > 1)
 378                 GOTO(out_sem, rc);
 379         exp = class_conn2export(dlm_handle);
 380
 381         if (obd->obd_namespace != NULL)
 382                 CERROR("already have namespace!\n");
 383         obd->obd_namespace = ldlm_namespace_new(obd->obd_name,
 384                                                 LDLM_NAMESPACE_CLIENT);
 385         if (obd->obd_namespace == NULL)
 386                 GOTO(out_disco, rc = -ENOMEM);
 387
 388         imp->imp_dlm_handle = *dlm_handle;
 389         rc = ptlrpc_init_import(imp);
 390         if (rc != 0)
 391                 GOTO(out_ldlm, rc);
 392
 393         rc = sptlrpc_import_get_sec(imp, NULL, cli->cl_sec_conf.sfc_rpc_flavor,
 394                                     cli->cl_sec_conf.sfc_flags);
 395         if (rc)
 396                 GOTO(out_ldlm, rc);
 397
 398         ocd = &imp->imp_connect_data;
 399         if (data) {
 400                 *ocd = *data;
 401                 imp->imp_connect_flags_orig = data->ocd_connect_flags;
 402         }
 403
 404         rc = ptlrpc_connect_import(imp, NULL);
 405         if (rc != 0) {
 406                 LASSERT (imp->imp_state == LUSTRE_IMP_DISCON);
 407                 GOTO(out_ldlm, rc);
 408         }
 409         LASSERT(exp->exp_connection);
 410
 411         if (data) {
 412                 LASSERT((ocd->ocd_connect_flags & data->ocd_connect_flags) ==
 413                         ocd->ocd_connect_flags);
 414                 data->ocd_connect_flags = ocd->ocd_connect_flags;
 415         }
 416
 417         ptlrpc_pinger_add_import(imp);
 418
 419         EXIT;
 420
 421         if (rc) {
 422 out_ldlm:
 423                 ldlm_namespace_free(obd->obd_namespace, 0);
 424                 obd->obd_namespace = NULL;
 425 out_disco:
 426                 cli->cl_conn_count--;
 427                 class_disconnect(exp);
 428         } else {
 429                 class_export_put(exp);
 430         }
 431 out_sem:
 432         mutex_up(&cli->cl_sem);
 433         return rc;
 434 }
 435
 436 int client_disconnect_export(struct obd_export *exp)
 437 {
 438         struct obd_device *obd = class_exp2obd(exp);
 439         struct client_obd *cli;
 440         struct obd_import *imp;
 441         int rc = 0, err;
 442         ENTRY;
 443
 444         if (!obd) {
 445                 CERROR("invalid export for disconnect: exp %p cookie "LPX64"\n",
 446                        exp, exp ? exp->exp_handle.h_cookie : -1);
 447                 RETURN(-EINVAL);
 448         }
 449
 450         cli = &obd->u.cli;
 451         imp = cli->cl_import;
 452
 453         mutex_down(&cli->cl_sem);
 454         if (!cli->cl_conn_count) {
 455                 CERROR("disconnecting disconnected device (%s)\n",
 456                        obd->obd_name);
 457                 GOTO(out_sem, rc = -EINVAL);
 458         }
 459
 460         cli->cl_conn_count--;
 461         if (cli->cl_conn_count)
 462                 GOTO(out_no_disconnect, rc = 0);
 463
 464         /* Mark import deactivated now, so we don't try to reconnect if any
 465          * of the cleanup RPCs fails (e.g. ldlm cancel, etc).  We don't
 466          * fully deactivate the import, or that would drop all requests. */
 467         imp->imp_deactive = 1;
 468
 469         /* Some non-replayable imports (MDS's OSCs) are pinged, so just
 470          * delete it regardless.  (It's safe to delete an import that was
 471          * never added.) */
 472         (void)ptlrpc_pinger_del_import(imp);
 473
 474         if (obd->obd_namespace != NULL) {
 475                 /* obd_no_recov == local only */
 476                 ldlm_cli_cancel_unused(obd->obd_namespace, NULL,
 477                                        obd->obd_no_recov ? LDLM_FL_LOCAL_ONLY:0,
 478                                        NULL);
 479                 ldlm_namespace_free(obd->obd_namespace, obd->obd_no_recov);
 480                 obd->obd_namespace = NULL;
 481         }
 482
 483         /* Yeah, obd_no_recov also (mainly) means "forced shutdown". */
 484         if (!obd->obd_no_recov)
 485                 rc = ptlrpc_disconnect_import(imp);
 486
 487         ptlrpc_invalidate_import(imp);
 488         ptlrpc_free_rq_pool(imp->imp_rq_pool);
 489         destroy_import(imp);
 490         cli->cl_import = NULL;
 491
 492         EXIT;
 493  out_no_disconnect:
 494         err = class_disconnect(exp);
 495         if (!rc && err)
 496                 rc = err;
 497  out_sem:
 498         mutex_up(&cli->cl_sem);
 499         RETURN(rc);
 500 }
 501
 502 /* --------------------------------------------------------------------------
 503  * from old lib/target.c
 504  * -------------------------------------------------------------------------- */
 505
 506 int target_handle_reconnect(struct lustre_handle *conn, struct obd_export *exp,
 507                             struct obd_uuid *cluuid)
 508 {
 509         ENTRY;
 510         if (exp->exp_connection && exp->exp_imp_reverse) {
 511                 struct lustre_handle *hdl;
 512                 hdl = &exp->exp_imp_reverse->imp_remote_handle;
 513                 /* Might be a re-connect after a partition. */
 514                 if (!memcmp(&conn->cookie, &hdl->cookie, sizeof conn->cookie)) {
 515                         CWARN("%s: %s reconnecting\n", exp->exp_obd->obd_name,
 516                               cluuid->uuid);
 517                         conn->cookie = exp->exp_handle.h_cookie;
 518                         /* target_handle_connect() treats EALREADY and
 519                          * -EALREADY differently.  EALREADY means we are
 520                          * doing a valid reconnect from the same client. */
 521                         RETURN(EALREADY);
 522                 } else {
 523                         CERROR("%s reconnecting from %s, "
 524                                "handle mismatch (ours "LPX64", theirs "
 525                                LPX64")\n", cluuid->uuid,
 526                                exp->exp_connection->c_remote_uuid.uuid,
 527                                hdl->cookie, conn->cookie);
 528                         memset(conn, 0, sizeof *conn);
 529                         /* target_handle_connect() treats EALREADY and
 530                          * -EALREADY differently.  -EALREADY is an error
 531                          * (same UUID, different handle). */
 532                         RETURN(-EALREADY);
 533                 }
 534         }
 535
 536         conn->cookie = exp->exp_handle.h_cookie;
 537         CDEBUG(D_HA, "connect export for UUID '%s' at %p, cookie "LPX64"\n",
 538                cluuid->uuid, exp, conn->cookie);
 539         RETURN(0);
 540 }
 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");
 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                 DEBUG_REQ(D_ERROR, req, "UUID '%s' is not available "
 582                           " for connect (%s)", str,
 583                           !target ? "no target" :
 584                           (target->obd_stopping ? "stopping" : "not set up"));
 585                 GOTO(out, rc = -ENODEV);
 586         }
 587
 588         /* Make sure the target isn't cleaned up while we're here. Yes,
 589            there's still a race between the above check and our incref here.
 590            Really, class_uuid2obd should take the ref. */
 591         targref = class_incref(target);
 592
 593         LASSERT_REQSWAB(req, REQ_REC_OFF + 1);
 594         str = lustre_msg_string(req->rq_reqmsg, REQ_REC_OFF + 1,
 595                                 sizeof(cluuid) - 1);
 596         if (str == NULL) {
 597                 DEBUG_REQ(D_ERROR, req, "bad client UUID for connect");
 598                 GOTO(out, rc = -EINVAL);
 599         }
 600
 601         obd_str2uuid (&cluuid, str);
 602
 603         /* XXX extract a nettype and format accordingly */
 604         switch (sizeof(lnet_nid_t)) {
 605                 /* NB the casts only avoid compiler warnings */
 606         case 8:
 607                 snprintf(remote_uuid.uuid, sizeof remote_uuid,
 608                          "NET_"LPX64"_UUID", (__u64)req->rq_peer.nid);
 609                 break;
 610         case 4:
 611                 snprintf(remote_uuid.uuid, sizeof remote_uuid,
 612                          "NET_%x_UUID", (__u32)req->rq_peer.nid);
 613                 break;
 614         default:
 615                 LBUG();
 616         }
 617
 618         spin_lock_bh(&target->obd_processing_task_lock);
 619         abort_recovery = target->obd_abort_recovery;
 620         spin_unlock_bh(&target->obd_processing_task_lock);
 621         if (abort_recovery)
 622                 target_abort_recovery(target);
 623
 624         tmp = lustre_msg_buf(req->rq_reqmsg, REQ_REC_OFF + 2, sizeof conn);
 625         if (tmp == NULL)
 626                 GOTO(out, rc = -EPROTO);
 627
 628         memcpy(&conn, tmp, sizeof conn);
 629
 630         data = lustre_swab_reqbuf(req, REQ_REC_OFF + 3, sizeof(*data),
 631                                   lustre_swab_connect);
 632         rc = lustre_pack_reply(req, 2, size, NULL);
 633         if (rc)
 634                 GOTO(out, rc);
 635
 636         if (lustre_msg_get_op_flags(req->rq_reqmsg) & MSG_CONNECT_LIBCLIENT) {
 637                 if (!data) {
 638                         DEBUG_REQ(D_WARNING, req, "Refusing old (unversioned) "
 639                                   "libclient connection attempt\n");
 640                         GOTO(out, rc = -EPROTO);
 641                 } else if (data->ocd_version < LUSTRE_VERSION_CODE -
 642                                                LUSTRE_VERSION_ALLOWED_OFFSET) {
 643                         DEBUG_REQ(D_WARNING, req, "Refusing old (%d.%d.%d.%d) "
 644                                   "libclient connection attempt\n",
 645                                   OBD_OCD_VERSION_MAJOR(data->ocd_version),
 646                                   OBD_OCD_VERSION_MINOR(data->ocd_version),
 647                                   OBD_OCD_VERSION_PATCH(data->ocd_version),
 648                                   OBD_OCD_VERSION_FIX(data->ocd_version));
 649                         data = lustre_msg_buf(req->rq_repmsg, REPLY_REC_OFF,
 650                                               offsetof(typeof(*data),
 651                                                        ocd_version) +
 652                                               sizeof(data->ocd_version));
 653                         if (data) {
 654                                 data->ocd_connect_flags = OBD_CONNECT_VERSION;
 655                                 data->ocd_version = LUSTRE_VERSION_CODE;
 656                         }
 657                         GOTO(out, rc = -EPROTO);
 658                 }
 659         }
 660
 661         /* lctl gets a backstage, all-access pass. */
 662         if (obd_uuid_equals(&cluuid, &target->obd_uuid))
 663                 goto dont_check_exports;
 664
 665         spin_lock(&target->obd_dev_lock);
 666         list_for_each(p, &target->obd_exports) {
 667                 export = list_entry(p, struct obd_export, exp_obd_chain);
 668                 if (obd_uuid_equals(&cluuid, &export->exp_client_uuid)) {
 669                         if (export->exp_connecting) { /* bug 9635, et. al. */
 670                                 CWARN("%s: exp %p already connecting\n",
 671                                       export->exp_obd->obd_name, export);
 672                                 export = NULL;
 673                                 rc = -EALREADY;
 674                                 break;
 675                         }
 676                         export->exp_connecting = 1;
 677                         spin_unlock(&target->obd_dev_lock);
 678                         LASSERT(export->exp_obd == target);
 679
 680                         rc = target_handle_reconnect(&conn, export, &cluuid);
 681                         break;
 682                 }
 683                 export = NULL;
 684         }
 685         /* If we found an export, we already unlocked. */
 686         if (!export) {
 687                 spin_unlock(&target->obd_dev_lock);
 688                 OBD_FAIL_TIMEOUT(OBD_FAIL_TGT_DELAY_CONNECT, 2 * obd_timeout);
 689         } else if (lustre_msg_get_conn_cnt(req->rq_reqmsg) == 1) {
 690                 CERROR("%s: NID %s (%s) reconnected with 1 conn_cnt; "
 691                        "cookies not random?\n", target->obd_name,
 692                        libcfs_nid2str(req->rq_peer.nid), cluuid.uuid);
 693                 GOTO(out, rc = -EALREADY);
 694         } else {
 695                 OBD_FAIL_TIMEOUT(OBD_FAIL_TGT_DELAY_RECONNECT, 2 * obd_timeout);
 696         }
 697
 698         /* We want to handle EALREADY but *not* -EALREADY from
 699          * target_handle_reconnect(), return reconnection state in a flag */
 700         if (rc == EALREADY) {
 701                 lustre_msg_add_op_flags(req->rq_repmsg, MSG_CONNECT_RECONNECT);
 702                 rc = 0;
 703         } else if (rc) {
 704                 GOTO(out, rc);
 705         }
 706
 707         /* Tell the client if we're in recovery. */
 708         /* If this is the first client, start the recovery timer */
 709         if (target->obd_recovering) {
 710                 lustre_msg_add_op_flags(req->rq_repmsg, MSG_CONNECT_RECOVERING);
 711                 target_start_recovery_timer(target, handler);
 712         }
 713
 714         /* Tell the client if we support replayable requests */
 715         if (target->obd_replayable)
 716                 lustre_msg_add_op_flags(req->rq_repmsg, MSG_CONNECT_REPLAYABLE);
 717
 718         if (export == NULL) {
 719                 if (target->obd_recovering) {
 720                         CERROR("%s: denying connection for new client %s (%s): "
 721                                "%d clients in recovery for %lds\n",
 722                                target->obd_name,
 723                                libcfs_nid2str(req->rq_peer.nid), cluuid.uuid,
 724                                target->obd_recoverable_clients,
 725                                cfs_duration_sec(cfs_time_sub(cfs_timer_deadline(&target->obd_recovery_timer),
 726                                                              cfs_time_current())));
 727                         rc = -EBUSY;
 728                 } else {
 729  dont_check_exports:
 730                         rc = obd_connect(req->rq_svc_thread->t_env,
 731                                          &conn, target, &cluuid, data);
 732                 }
 733         } else {
 734                 rc = obd_reconnect(export, target, &cluuid, data);
 735         }
 736
 737         if (rc)
 738                 GOTO(out, rc);
 739
 740         /* Return only the parts of obd_connect_data that we understand, so the
 741          * client knows that we don't understand the rest. */
 742         if (data)
 743                 memcpy(lustre_msg_buf(req->rq_repmsg, REPLY_REC_OFF,
 744                                       sizeof(*data)),
 745                        data, sizeof(*data));
 746
 747         /* If all else goes well, this is our RPC return code. */
 748         req->rq_status = 0;
 749
 750         lustre_msg_set_handle(req->rq_repmsg, &conn);
 751
 752         /* ownership of this export ref transfers to the request AFTER we
 753          * drop any previous reference the request had, but we don't want
 754          * that to go to zero before we get our new export reference. */
 755         export = class_conn2export(&conn);
 756         if (!export) {
 757                 DEBUG_REQ(D_ERROR, req, "Missing export!\n");
 758                 GOTO(out, rc = -ENODEV);
 759         }
 760
 761         /* If the client and the server are the same node, we will already
 762          * have an export that really points to the client's DLM export,
 763          * because we have a shared handles table.
 764          *
 765          * XXX this will go away when shaver stops sending the "connect" handle
 766          * in the real "remote handle" field of the request --phik 24 Apr 2003
 767          */
 768         if (req->rq_export != NULL)
 769                 class_export_put(req->rq_export);
 770
 771         req->rq_export = export;
 772
 773         spin_lock(&export->exp_lock);
 774         if (export->exp_conn_cnt >= lustre_msg_get_conn_cnt(req->rq_reqmsg)) {
 775                 CERROR("%s: %s already connected at higher conn_cnt: %d > %d\n",
 776                        cluuid.uuid, libcfs_nid2str(req->rq_peer.nid),
 777                        export->exp_conn_cnt,
 778                        lustre_msg_get_conn_cnt(req->rq_reqmsg));
 779
 780                 spin_unlock(&export->exp_lock);
 781                 GOTO(out, rc = -EALREADY);
 782         }
 783         export->exp_conn_cnt = lustre_msg_get_conn_cnt(req->rq_reqmsg);
 784         spin_unlock(&export->exp_lock);
 785
 786         /* request from liblustre?  Don't evict it for not pinging. */
 787         if (lustre_msg_get_op_flags(req->rq_reqmsg) & MSG_CONNECT_LIBCLIENT) {
 788                 export->exp_libclient = 1;
 789                 spin_lock(&target->obd_dev_lock);
 790                 list_del_init(&export->exp_obd_chain_timed);
 791                 spin_unlock(&target->obd_dev_lock);
 792         }
 793
 794         if (export->exp_connection != NULL)
 795                 ptlrpc_put_connection(export->exp_connection);
 796         export->exp_connection = ptlrpc_get_connection(req->rq_peer,
 797                                                        req->rq_self,
 798                                                        &remote_uuid);
 799
 800         if (lustre_msg_get_op_flags(req->rq_repmsg) & MSG_CONNECT_RECONNECT) {
 801                 LASSERT(export->exp_imp_reverse);
 802                 sptlrpc_svc_install_rvs_ctx(export->exp_imp_reverse,
 803                                             req->rq_svc_ctx);
 804                 GOTO(out, rc = 0);
 805         }
 806
 807         if (target->obd_recovering)
 808                 target->obd_connected_clients++;
 809
 810         memcpy(&conn,
 811                lustre_msg_buf(req->rq_reqmsg, REQ_REC_OFF + 2, sizeof conn),
 812                sizeof conn);
 813
 814         if (export->exp_imp_reverse != NULL)
 815                 destroy_import(export->exp_imp_reverse);
 816         revimp = export->exp_imp_reverse = class_new_import(target);
 817         revimp->imp_connection = ptlrpc_connection_addref(export->exp_connection);
 818         revimp->imp_client = &export->exp_obd->obd_ldlm_client;
 819         revimp->imp_remote_handle = conn;
 820         revimp->imp_dlm_fake = 1;
 821         revimp->imp_state = LUSTRE_IMP_FULL;
 822
 823         if (lustre_msg_get_op_flags(req->rq_reqmsg) & MSG_CONNECT_NEXT_VER) {
 824                 revimp->imp_msg_magic = LUSTRE_MSG_MAGIC_V2;
 825                 lustre_msg_add_op_flags(req->rq_repmsg, MSG_CONNECT_NEXT_VER);
 826         }
 827
 828         rc = sptlrpc_import_get_sec(revimp, req->rq_svc_ctx,
 829                                     req->rq_sec_flavor, 0);
 830         if (rc) {
 831                 CERROR("Failed to get sec for reverse import: %d\n", rc);
 832                 export->exp_imp_reverse = NULL;
 833                 class_destroy_import(revimp);
 834         }
 835
 836         class_import_put(revimp);
 837 out:
 838         if (export)
 839                 export->exp_connecting = 0;
 840         if (targref)
 841                 class_decref(targref);
 842         if (rc)
 843                 req->rq_status = rc;
 844         RETURN(rc);
 845 }
 846
 847 int target_handle_disconnect(struct ptlrpc_request *req)
 848 {
 849         int rc;
 850         ENTRY;
 851
 852         rc = lustre_pack_reply(req, 1, NULL, NULL);
 853         if (rc)
 854                 RETURN(rc);
 855
 856         /* keep the rq_export around so we can send the reply */
 857         req->rq_status = obd_disconnect(class_export_get(req->rq_export));
 858         RETURN(0);
 859 }
 860
 861 void target_destroy_export(struct obd_export *exp)
 862 {
 863         /* exports created from last_rcvd data, and "fake"
 864            exports created by lctl don't have an import */
 865         if (exp->exp_imp_reverse != NULL)
 866                 destroy_import(exp->exp_imp_reverse);
 867
 868         /* We cancel locks at disconnect time, but this will catch any locks
 869          * granted in a race with recovery-induced disconnect. */
 870         if (exp->exp_obd->obd_namespace != NULL)
 871                 ldlm_cancel_locks_for_export(exp);
 872 }
 873
 874 /*
 875  * Recovery functions
 876  */
 877
 878
 879 static
 880 struct ptlrpc_request *target_save_req(struct ptlrpc_request *src)
 881 {
 882         struct ptlrpc_request *req;
 883         struct lustre_msg *reqmsg;
 884
 885         OBD_ALLOC_PTR(req);
 886         if (!req)
 887                 return NULL;
 888
 889         OBD_ALLOC(reqmsg, src->rq_reqlen);
 890         if (!reqmsg) {
 891                 OBD_FREE_PTR(req);
 892                 return NULL;
 893         }
 894
 895         *req = *src;
 896         memcpy(reqmsg, src->rq_reqmsg, src->rq_reqlen);
 897         req->rq_reqmsg = reqmsg;
 898
 899         class_export_get(req->rq_export);
 900         CFS_INIT_LIST_HEAD(&req->rq_list);
 901         sptlrpc_svc_ctx_addref(req);
 902         if (req->rq_reply_state)
 903                 ptlrpc_rs_addref(req->rq_reply_state);
 904
 905         /* repmsg have been taken over, in privacy mode this might point to
 906          * invalid data. prevent further access on it.
 907          */
 908         src->rq_repmsg = NULL;
 909         src->rq_replen = 0;
 910
 911         return req;
 912 }
 913
 914 static
 915 void target_release_saved_req(struct ptlrpc_request *req)
 916 {
 917         if (req->rq_reply_state) {
 918                 ptlrpc_rs_decref(req->rq_reply_state);
 919                 req->rq_reply_state = NULL;
 920         }
 921         sptlrpc_svc_ctx_decref(req);
 922         class_export_put(req->rq_export);
 923
 924         OBD_FREE(req->rq_reqmsg, req->rq_reqlen);
 925         OBD_FREE_PTR(req);
 926 }
 927
 928 static void target_finish_recovery(struct obd_device *obd)
 929 {
 930         struct list_head *tmp, *n;
 931
 932         CWARN("%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                 CWARN("%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_WARNING, 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                 target_release_saved_req(req);
1181
1182                 spin_lock_bh(&obd->obd_processing_task_lock);
1183                 obd->obd_next_recovery_transno++;
1184                 if (list_empty(&obd->obd_recovery_queue)) {
1185                         obd->obd_processing_task = 0;
1186                         spin_unlock_bh(&obd->obd_processing_task_lock);
1187                         break;
1188                 }
1189                 spin_unlock_bh(&obd->obd_processing_task_lock);
1190         }
1191         EXIT;
1192 }
1193
1194 int target_queue_recovery_request(struct ptlrpc_request *req,
1195                                   struct obd_device *obd)
1196 {
1197         struct list_head *tmp;
1198         int inserted = 0;
1199         __u64 transno = lustre_msg_get_transno(req->rq_reqmsg);
1200         struct ptlrpc_request *saved_req;
1201
1202         /* CAVEAT EMPTOR: The incoming request message has been swabbed
1203          * (i.e. buflens etc are in my own byte order), but type-dependent
1204          * buffers (eg mds_body, ost_body etc) have NOT been swabbed. */
1205
1206         if (!transno) {
1207                 CFS_INIT_LIST_HEAD(&req->rq_list);
1208                 DEBUG_REQ(D_HA, req, "not queueing");
1209                 return 1;
1210         }
1211
1212         /* XXX If I were a real man, these LBUGs would be sane cleanups. */
1213         saved_req = target_save_req(req);
1214         if (!saved_req)
1215                 LBUG();
1216
1217         spin_lock_bh(&obd->obd_processing_task_lock);
1218
1219         /*
1220          * If we're processing the queue, we don't want to queue this message.
1221          *
1222          * Also, if this request has a transno less than the one we're waiting
1223          * for, we should process it now.  It could (and currently always will)
1224          * be an open request for a descriptor that was opened some time ago.
1225          *
1226          * Also, a resent, replayed request that has already been handled will
1227          * pass through here and be processed immediately.
1228          */
1229         if (obd->obd_processing_task == cfs_curproc_pid() ||
1230             transno < obd->obd_next_recovery_transno) {
1231                 /* Processing the queue right now, don't re-add. */
1232                 LASSERT(list_empty(&req->rq_list));
1233                 spin_unlock_bh(&obd->obd_processing_task_lock);
1234
1235                 target_release_saved_req(saved_req);
1236                 return 1;
1237         }
1238
1239         /* A resent, replayed request that is still on the queue; just drop it.
1240            The queued request will handle this. */
1241         if ((lustre_msg_get_flags(req->rq_reqmsg) & (MSG_RESENT|MSG_REPLAY)) ==
1242             (MSG_RESENT | MSG_REPLAY)) {
1243                 DEBUG_REQ(D_ERROR, req, "dropping resent queued req");
1244                 spin_unlock_bh(&obd->obd_processing_task_lock);
1245
1246                 target_release_saved_req(saved_req);
1247                 return 0;
1248         }
1249
1250         req = saved_req;
1251
1252         /* XXX O(n^2) */
1253         list_for_each(tmp, &obd->obd_recovery_queue) {
1254                 struct ptlrpc_request *reqiter =
1255                         list_entry(tmp, struct ptlrpc_request, rq_list);
1256
1257                 if (lustre_msg_get_transno(reqiter->rq_reqmsg) > transno) {
1258                         list_add_tail(&req->rq_list, &reqiter->rq_list);
1259                         inserted = 1;
1260                         break;
1261                 }
1262         }
1263
1264         if (!inserted) {
1265                 list_add_tail(&req->rq_list, &obd->obd_recovery_queue);
1266         }
1267
1268         obd->obd_requests_queued_for_recovery++;
1269
1270         if (obd->obd_processing_task != 0) {
1271                 /* Someone else is processing this queue, we'll leave it to
1272                  * them.
1273                  */
1274                 cfs_waitq_signal(&obd->obd_next_transno_waitq);
1275                 spin_unlock_bh(&obd->obd_processing_task_lock);
1276                 return 0;
1277         }
1278
1279         /* Nobody is processing, and we know there's (at least) one to process
1280          * now, so we'll do the honours.
1281          */
1282         obd->obd_processing_task = cfs_curproc_pid();
1283         spin_unlock_bh(&obd->obd_processing_task_lock);
1284
1285         process_recovery_queue(obd);
1286         return 0;
1287 }
1288
1289 struct obd_device * target_req2obd(struct ptlrpc_request *req)
1290 {
1291         return req->rq_export->exp_obd;
1292 }
1293
1294 int target_queue_final_reply(struct ptlrpc_request *req, int rc)
1295 {
1296         struct obd_device *obd = target_req2obd(req);
1297         struct ptlrpc_request *saved_req;
1298         int recovery_done = 0;
1299
1300         LASSERT ((rc == 0) == (req->rq_reply_state != NULL));
1301
1302         if (rc) {
1303                 /* Just like ptlrpc_error, but without the sending. */
1304                 rc = lustre_pack_reply(req, 1, NULL, NULL);
1305                 LASSERT(rc == 0); /* XXX handle this */
1306                 req->rq_type = PTL_RPC_MSG_ERR;
1307         }
1308
1309         LASSERT (!req->rq_reply_state->rs_difficult);
1310         LASSERT(list_empty(&req->rq_list));
1311
1312         saved_req = target_save_req(req);
1313         if (!saved_req)
1314                 LBUG();
1315
1316         /* Don't race cleanup */
1317         spin_lock_bh(&obd->obd_processing_task_lock);
1318         if (obd->obd_stopping) {
1319                 spin_unlock_bh(&obd->obd_processing_task_lock);
1320                 target_release_saved_req(saved_req);
1321                 req->rq_status = -ENOTCONN;
1322                 /* rv is ignored anyhow */
1323                 return -ENOTCONN;
1324         }
1325
1326         req = saved_req;
1327         list_add(&req->rq_list, &obd->obd_delayed_reply_queue);
1328
1329         /* only count the first "replay over" request from each
1330            export */
1331         if (req->rq_export->exp_replay_needed) {
1332                 --obd->obd_recoverable_clients;
1333                 req->rq_export->exp_replay_needed = 0;
1334         }
1335         recovery_done = (obd->obd_recoverable_clients == 0);
1336         spin_unlock_bh(&obd->obd_processing_task_lock);
1337
1338         OBD_RACE(OBD_FAIL_LDLM_RECOV_CLIENTS);
1339         if (recovery_done) {
1340                 spin_lock_bh(&obd->obd_processing_task_lock);
1341                 obd->obd_recovering = obd->obd_abort_recovery = 0;
1342                 target_cancel_recovery_timer(obd);
1343                 spin_unlock_bh(&obd->obd_processing_task_lock);
1344
1345                 target_finish_recovery(obd);
1346                 CDEBUG(D_HA, "%s: recovery complete\n",
1347                        obd_uuid2str(&obd->obd_uuid));
1348         } else {
1349                 CWARN("%s: %d recoverable clients remain\n",
1350                        obd->obd_name, obd->obd_recoverable_clients);
1351                 cfs_waitq_signal(&obd->obd_next_transno_waitq);
1352         }
1353
1354         return 1;
1355 }
1356
1357 int
1358 target_send_reply_msg(struct ptlrpc_request *req, int rc, int fail_id)
1359 {
1360         if (OBD_FAIL_CHECK(fail_id | OBD_FAIL_ONCE)) {
1361                 obd_fail_loc |= OBD_FAIL_ONCE | OBD_FAILED;
1362                 DEBUG_REQ(D_ERROR, req, "dropping reply");
1363                 return (-ECOMM);
1364         }
1365
1366         if (rc || req->rq_reply_state == NULL) {
1367                 if (rc == 0) {
1368                         DEBUG_REQ(D_ERROR, req, "no reply message packed");
1369                         rc = -ENOMEM;
1370                 } else
1371                         DEBUG_REQ(D_ERROR, req, "processing error (%d)", rc);
1372                 req->rq_status = rc;
1373                 return (ptlrpc_error(req));
1374         }
1375
1376         DEBUG_REQ(D_NET, req, "sending reply");
1377         return (ptlrpc_send_reply(req, 1));
1378 }
1379
1380 void
1381 target_send_reply(struct ptlrpc_request *req, int rc, int fail_id)
1382 {
1383         int                        netrc;
1384         struct ptlrpc_reply_state *rs;
1385         struct obd_device         *obd;
1386         struct obd_export         *exp;
1387         struct ptlrpc_service     *svc;
1388
1389         svc = req->rq_rqbd->rqbd_service;
1390         rs = req->rq_reply_state;
1391         if (rs == NULL || !rs->rs_difficult) {
1392                 /* no notifiers */
1393                 target_send_reply_msg(req, rc, fail_id);
1394                 return;
1395         }
1396
1397         /* must be an export if locks saved */
1398         LASSERT (req->rq_export != NULL);
1399         /* req/reply consistent */
1400         LASSERT (rs->rs_service == svc);
1401
1402         /* "fresh" reply */
1403         LASSERT (!rs->rs_scheduled);
1404         LASSERT (!rs->rs_scheduled_ever);
1405         LASSERT (!rs->rs_handled);
1406         LASSERT (!rs->rs_on_net);
1407         LASSERT (rs->rs_export == NULL);
1408         LASSERT (list_empty(&rs->rs_obd_list));
1409         LASSERT (list_empty(&rs->rs_exp_list));
1410
1411         exp = class_export_get (req->rq_export);
1412         obd = exp->exp_obd;
1413
1414         /* disable reply scheduling onto srv_reply_queue while I'm setting up */
1415         rs->rs_scheduled = 1;
1416         rs->rs_on_net    = 1;
1417         rs->rs_xid       = req->rq_xid;
1418         rs->rs_transno   = req->rq_transno;
1419         rs->rs_export    = exp;
1420
1421         spin_lock(&obd->obd_uncommitted_replies_lock);
1422
1423         if (rs->rs_transno > obd->obd_last_committed) {
1424                 /* not committed already */
1425                 list_add_tail (&rs->rs_obd_list,
1426                                &obd->obd_uncommitted_replies);
1427         }
1428
1429         spin_unlock (&obd->obd_uncommitted_replies_lock);
1430         spin_lock (&exp->exp_lock);
1431
1432         list_add_tail (&rs->rs_exp_list, &exp->exp_outstanding_replies);
1433
1434         spin_unlock(&exp->exp_lock);
1435
1436         netrc = target_send_reply_msg(req, rc, fail_id);
1437
1438         spin_lock(&svc->srv_lock);
1439
1440         svc->srv_n_difficult_replies++;
1441
1442         if (netrc != 0) {
1443                 /* error sending: reply is off the net.  Also we need +1
1444                  * reply ref until ptlrpc_server_handle_reply() is done
1445                  * with the reply state (if the send was successful, there
1446                  * would have been +1 ref for the net, which
1447                  * reply_out_callback leaves alone) */
1448                 rs->rs_on_net = 0;
1449                 ptlrpc_rs_addref(rs);
1450                 atomic_inc (&svc->srv_outstanding_replies);
1451         }
1452
1453         if (!rs->rs_on_net ||                   /* some notifier */
1454             list_empty(&rs->rs_exp_list) ||     /* completed already */
1455             list_empty(&rs->rs_obd_list)) {
1456                 list_add_tail (&rs->rs_list, &svc->srv_reply_queue);
1457                 cfs_waitq_signal (&svc->srv_waitq);
1458         } else {
1459                 list_add (&rs->rs_list, &svc->srv_active_replies);
1460                 rs->rs_scheduled = 0;           /* allow notifier to schedule */
1461         }
1462
1463         spin_unlock(&svc->srv_lock);
1464 }
1465
1466 int target_handle_ping(struct ptlrpc_request *req)
1467 {
1468         obd_ping(req->rq_export);
1469         return lustre_pack_reply(req, 1, NULL, NULL);
1470 }
1471
1472 void target_committed_to_req(struct ptlrpc_request *req)
1473 {
1474         struct obd_device *obd;
1475
1476         if (req == NULL || req->rq_export == NULL)
1477                 return;
1478
1479         obd = req->rq_export->exp_obd;
1480         if (obd == NULL)
1481                 return;
1482
1483         if (!obd->obd_no_transno && req->rq_repmsg != NULL)
1484                 lustre_msg_set_last_committed(req->rq_repmsg,
1485                                               obd->obd_last_committed);
1486         else
1487                 DEBUG_REQ(D_IOCTL, req, "not sending last_committed update");
1488
1489         CDEBUG(D_INFO, "last_committed "LPU64", xid "LPU64"\n",
1490                obd->obd_last_committed, req->rq_xid);
1491 }
1492
1493 EXPORT_SYMBOL(target_committed_to_req);
1494
1495 #ifdef HAVE_QUOTA_SUPPORT
1496 int target_handle_qc_callback(struct ptlrpc_request *req)
1497 {
1498         struct obd_quotactl *oqctl;
1499         struct client_obd *cli = &req->rq_export->exp_obd->u.cli;
1500
1501         oqctl = lustre_swab_reqbuf(req, REQ_REC_OFF, sizeof(*oqctl),
1502                                    lustre_swab_obd_quotactl);
1503         if (oqctl == NULL) {
1504                 CERROR("Can't unpack obd_quotactl\n");
1505                 RETURN(-EPROTO);
1506         }
1507
1508         cli->cl_qchk_stat = oqctl->qc_stat;
1509
1510         return 0;
1511 }
1512
1513 int target_handle_dqacq_callback(struct ptlrpc_request *req)
1514 {
1515 #ifdef __KERNEL__
1516         struct obd_device *obd = req->rq_export->exp_obd;
1517         struct obd_device *master_obd;
1518         struct lustre_quota_ctxt *qctxt;
1519         struct qunit_data *qdata;
1520         void* rep;
1521         struct qunit_data_old *qdata_old;
1522         int rc = 0;
1523         int repsize[2] = { sizeof(struct ptlrpc_body),
1524                            sizeof(struct qunit_data) };
1525         ENTRY;
1526
1527         rc = lustre_pack_reply(req, 2, repsize, NULL);
1528         if (rc) {
1529                 CERROR("packing reply failed!: rc = %d\n", rc);
1530                 RETURN(rc);
1531         }
1532         LASSERT(req->rq_export);
1533
1534         /* fixed for bug10707 */
1535         if ((req->rq_export->exp_connect_flags & OBD_CONNECT_QUOTA64) &&
1536             !OBD_FAIL_CHECK(OBD_FAIL_QUOTA_QD_COUNT_32BIT)) {
1537                 CDEBUG(D_QUOTA, "qd_count is 64bit!\n");
1538                 rep = lustre_msg_buf(req->rq_repmsg, REPLY_REC_OFF,
1539                                      sizeof(struct qunit_data));
1540                 LASSERT(rep);
1541                 qdata = lustre_swab_reqbuf(req, REQ_REC_OFF, sizeof(*qdata),
1542                                            lustre_swab_qdata);
1543         } else {
1544                 CDEBUG(D_QUOTA, "qd_count is 32bit!\n");
1545                 rep = lustre_msg_buf(req->rq_repmsg, REPLY_REC_OFF,
1546                                      sizeof(struct qunit_data_old));
1547                 LASSERT(rep);
1548                 qdata_old = lustre_swab_reqbuf(req, REQ_REC_OFF, sizeof(*qdata_old),
1549                                                lustre_swab_qdata_old);
1550                 qdata = lustre_quota_old_to_new(qdata_old);
1551         }
1552
1553         if (qdata == NULL) {
1554                 CERROR("Can't unpack qunit_data\n");
1555                 RETURN(-EPROTO);
1556         }
1557
1558         /* we use the observer */
1559         LASSERT(obd->obd_observer && obd->obd_observer->obd_observer);
1560         master_obd = obd->obd_observer->obd_observer;
1561         qctxt = &master_obd->u.obt.obt_qctxt;
1562
1563         LASSERT(qctxt->lqc_handler);
1564         rc = qctxt->lqc_handler(master_obd, qdata,
1565                                 lustre_msg_get_opc(req->rq_reqmsg));
1566         if (rc && rc != -EDQUOT)
1567                 CDEBUG(rc == -EBUSY  ? D_QUOTA : D_ERROR,
1568                        "dqacq failed! (rc:%d)\n", rc);
1569
1570         /* the qd_count might be changed in lqc_handler */
1571         if ((req->rq_export->exp_connect_flags & OBD_CONNECT_QUOTA64) &&
1572             !OBD_FAIL_CHECK(OBD_FAIL_QUOTA_QD_COUNT_32BIT)) {
1573                 memcpy(rep,qdata,sizeof(*qdata));
1574         } else {
1575                 qdata_old = lustre_quota_new_to_old(qdata);
1576                 memcpy(rep,qdata_old,sizeof(*qdata_old));
1577         }
1578         req->rq_status = rc;
1579         rc = ptlrpc_reply(req);
1580
1581         RETURN(rc);
1582 #else
1583         return 0;
1584 #endif /* !__KERNEL__ */
1585 }
1586 #endif /* HAVE_QUOTA_SUPPORT */
1587
1588 ldlm_mode_t lck_compat_array[] = {
1589         [LCK_EX] LCK_COMPAT_EX,
1590         [LCK_PW] LCK_COMPAT_PW,
1591         [LCK_PR] LCK_COMPAT_PR,
1592         [LCK_CW] LCK_COMPAT_CW,
1593         [LCK_CR] LCK_COMPAT_CR,
1594         [LCK_NL] LCK_COMPAT_NL,
1595         [LCK_GROUP] LCK_COMPAT_GROUP
1596 };