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 Lustre, http://www.lustre.org.
   7  *
   8  *   Lustre is free software; you can redistribute it and/or
   9  *   modify it under the terms of version 2 of the GNU General Public
  10  *   License as published by the Free Software Foundation.
  11  *
  12  *   Lustre is distributed in the hope that it will be useful,
  13  *   but WITHOUT ANY WARRANTY; without even the implied warranty of
  14  *   MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
  15  *   GNU General Public License for more details.
  16  *
  17  *   You should have received a copy of the GNU General Public License
  18  *   along with Lustre; if not, write to the Free Software
  19  *   Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA.
  20  */
  21
  22 #ifndef EXPORT_SYMTAB
  23 # define EXPORT_SYMTAB
  24 #endif
  25 #define DEBUG_SUBSYSTEM S_LDLM
  26
  27 #ifdef __KERNEL__
  28 # include <linux/module.h>
  29 #else
  30 # include <liblustre.h>
  31 #endif
  32 #include <linux/obd.h>
  33 #include <linux/obd_ost.h> /* for LUSTRE_OSC_NAME */
  34 #include <linux/lustre_mds.h> /* for LUSTRE_MDC_NAME */
  35 #include <linux/lustre_mgmt.h>
  36 #include <linux/lustre_dlm.h>
  37 #include <linux/lustre_net.h>
  38 #include <linux/lustre_sec.h>
  39 #include <linux/lustre_gs.h>
  40
  41 /* @priority: if non-zero, move the selected to the list head
  42  * @nocreate: if non-zero, only search in existed connections
  43  */
  44 static int import_set_conn(struct obd_import *imp, struct obd_uuid *uuid,
  45                            int priority, int nocreate)
  46 {
  47         struct ptlrpc_connection *ptlrpc_conn;
  48         struct obd_import_conn *imp_conn = NULL, *item;
  49         int rc = 0;
  50         ENTRY;
  51
  52         LASSERT(!(nocreate && !priority));
  53
  54         ptlrpc_conn = ptlrpc_uuid_to_connection(uuid);
  55         if (!ptlrpc_conn) {
  56                 CERROR("can't find connection %s\n", uuid->uuid);
  57                 RETURN (-EINVAL);
  58         }
  59
  60         if (!nocreate) {
  61                 OBD_ALLOC(imp_conn, sizeof(*imp_conn));
  62                 if (!imp_conn) {
  63                         CERROR("fail to alloc memory\n");
  64                         GOTO(out_put, rc = -ENOMEM);
  65                 }
  66         }
  67
  68         spin_lock(&imp->imp_lock);
  69         list_for_each_entry(item, &imp->imp_conn_list, oic_item) {
  70                 if (obd_uuid_equals(uuid, &item->oic_uuid)) {
  71                         if (priority) {
  72                                 list_del(&item->oic_item);
  73                                 list_add(&item->oic_item, &imp->imp_conn_list);
  74                                 item->oic_last_attempt = 0;
  75                         }
  76                         CDEBUG(D_HA, "imp %p@%s: find existed conn %s%s\n",
  77                                imp, imp->imp_obd->obd_name, uuid->uuid,
  78                                (priority ? ", move to head." : ""));
  79                         spin_unlock(&imp->imp_lock);
  80                         GOTO(out_free, rc = 0);
  81                 }
  82         }
  83         /* not found */
  84         if (!nocreate) {
  85                 imp_conn->oic_conn = ptlrpc_conn;
  86                 imp_conn->oic_uuid = *uuid;
  87                 imp_conn->oic_last_attempt = 0;
  88                 if (priority)
  89                         list_add(&imp_conn->oic_item, &imp->imp_conn_list);
  90                 else
  91                         list_add_tail(&imp_conn->oic_item, &imp->imp_conn_list);
  92                 CDEBUG(D_HA, "imp %p@%s: add connection %s at %s\n",
  93                        imp, imp->imp_obd->obd_name, uuid->uuid,
  94                        (priority ? "head" : "tail"));
  95         } else
  96                 rc = -ENOENT;
  97
  98         spin_unlock(&imp->imp_lock);
  99         RETURN(0);
 100 out_free:
 101         if (imp_conn)
 102                 OBD_FREE(imp_conn, sizeof(*imp_conn));
 103 out_put:
 104         ptlrpc_put_connection(ptlrpc_conn);
 105         RETURN(rc);
 106 }
 107
 108 int import_set_conn_priority(struct obd_import *imp, struct obd_uuid *uuid)
 109 {
 110         return import_set_conn(imp, uuid, 1, 1);
 111 }
 112
 113 int client_import_add_conn(struct obd_import *imp, struct obd_uuid *uuid,
 114                            int priority)
 115 {
 116         return import_set_conn(imp, uuid, priority, 0);
 117 }
 118
 119 int client_import_del_conn(struct obd_import *imp, struct obd_uuid *uuid)
 120 {
 121         struct obd_import_conn *imp_conn;
 122         struct obd_export *dlmexp;
 123         int rc = -ENOENT;
 124         ENTRY;
 125
 126         spin_lock(&imp->imp_lock);
 127         if (list_empty(&imp->imp_conn_list)) {
 128                 LASSERT(!imp->imp_conn_current);
 129                 LASSERT(!imp->imp_connection);
 130                 GOTO(out, rc);
 131         }
 132
 133         list_for_each_entry(imp_conn, &imp->imp_conn_list, oic_item) {
 134                 if (!obd_uuid_equals(uuid, &imp_conn->oic_uuid))
 135                         continue;
 136                 LASSERT(imp_conn->oic_conn);
 137
 138                 /* is current conn? */
 139                 if (imp_conn == imp->imp_conn_current) {
 140                         LASSERT(imp_conn->oic_conn == imp->imp_connection);
 141
 142                         if (imp->imp_state != LUSTRE_IMP_CLOSED &&
 143                             imp->imp_state != LUSTRE_IMP_DISCON) {
 144                                 CERROR("can't remove current connection\n");
 145                                 GOTO(out, rc = -EBUSY);
 146                         }
 147
 148                         ptlrpc_put_connection(imp->imp_connection);
 149                         imp->imp_connection = NULL;
 150
 151                         dlmexp = class_conn2export(&imp->imp_dlm_handle);
 152                         if (dlmexp && dlmexp->exp_connection) {
 153                                 LASSERT(dlmexp->exp_connection ==
 154                                         imp_conn->oic_conn);
 155                                 ptlrpc_put_connection(dlmexp->exp_connection);
 156                                 dlmexp->exp_connection = NULL;
 157                         }
 158                 }
 159
 160                 list_del(&imp_conn->oic_item);
 161                 ptlrpc_put_connection(imp_conn->oic_conn);
 162                 OBD_FREE(imp_conn, sizeof(*imp_conn));
 163                 CDEBUG(D_HA, "imp %p@%s: remove connection %s\n",
 164                        imp, imp->imp_obd->obd_name, uuid->uuid);
 165                 rc = 0;
 166                 break;
 167         }
 168 out:
 169         spin_unlock(&imp->imp_lock);
 170         if (rc == -ENOENT)
 171                 CERROR("connection %s not found\n", uuid->uuid);
 172         RETURN(rc);
 173 }
 174
 175 int client_obd_setup(struct obd_device *obddev, obd_count len, void *buf)
 176 {
 177         struct lustre_cfg* lcfg = buf;
 178         struct client_obd *cli = &obddev->u.cli;
 179         struct obd_import *imp;
 180         struct obd_uuid server_uuid;
 181         int rq_portal, rp_portal, connect_op;
 182         char *name = obddev->obd_type->typ_name;
 183         char *mgmt_name = NULL;
 184         int rc;
 185         ENTRY;
 186
 187         /* In a more perfect world, we would hang a ptlrpc_client off of
 188          * obd_type and just use the values from there. */
 189         if (!strcmp(name, OBD_OSC_DEVICENAME)) {
 190                 rq_portal = OST_REQUEST_PORTAL;
 191                 rp_portal = OSC_REPLY_PORTAL;
 192                 connect_op = OST_CONNECT;
 193         } else if (!strcmp(name, OBD_MDC_DEVICENAME)) {
 194                 rq_portal = MDS_REQUEST_PORTAL;
 195                 rp_portal = MDC_REPLY_PORTAL;
 196                 connect_op = MDS_CONNECT;
 197         } else if (!strcmp(name, OBD_MGMTCLI_DEVICENAME)) {
 198                 rq_portal = MGMT_REQUEST_PORTAL;
 199                 rp_portal = MGMT_REPLY_PORTAL;
 200                 connect_op = MGMT_CONNECT;
 201         } else if (!strcmp(name, LUSTRE_GKC_NAME)) {
 202                 rq_portal = GKS_REQUEST_PORTAL;
 203                 rp_portal = GKC_REPLY_PORTAL;
 204                 connect_op = GKS_CONNECT;
 205
 206         } else {
 207                 CERROR("unknown client OBD type \"%s\", can't setup\n",
 208                        name);
 209                 RETURN(-EINVAL);
 210         }
 211
 212
 213         if (LUSTRE_CFG_BUFLEN(lcfg, 1) < 1) {
 214                 CERROR("requires a TARGET UUID\n");
 215                 RETURN(-EINVAL);
 216         }
 217
 218         if (LUSTRE_CFG_BUFLEN(lcfg, 1) > 37) {
 219                 CERROR("client UUID must be less than 38 characters\n");
 220                 RETURN(-EINVAL);
 221         }
 222
 223         if (LUSTRE_CFG_BUFLEN(lcfg, 2) < 1) {
 224                 CERROR("setup requires a SERVER UUID\n");
 225                 RETURN(-EINVAL);
 226         }
 227
 228         if (LUSTRE_CFG_BUFLEN(lcfg, 2) > 37) {
 229                 CERROR("target UUID must be less than 38 characters\n");
 230                 RETURN(-EINVAL);
 231         }
 232
 233         sema_init(&cli->cl_sem, 1);
 234         cli->cl_conn_count = 0;
 235         memcpy(server_uuid.uuid,  lustre_cfg_buf(lcfg, 2),
 236                min_t(unsigned int, LUSTRE_CFG_BUFLEN(lcfg, 2),
 237                sizeof(server_uuid)));
 238
 239         cli->cl_dirty = 0;
 240         cli->cl_avail_grant = 0;
 241
 242         /* XXX: making ->cl_dirty_max twice bigger for debug purposes! */
 243         /* FIXME: should limit this for the sum of all cl_dirty_max */
 244         cli->cl_dirty_max = OSC_MAX_DIRTY_DEFAULT * 1024 * 1024;
 245         if (cli->cl_dirty_max >> PAGE_SHIFT > num_physpages / 4)
 246                 cli->cl_dirty_max = num_physpages << (PAGE_SHIFT - 2);
 247
 248         INIT_LIST_HEAD(&cli->cl_cache_waiters);
 249         INIT_LIST_HEAD(&cli->cl_loi_ready_list);
 250         INIT_LIST_HEAD(&cli->cl_loi_write_list);
 251         INIT_LIST_HEAD(&cli->cl_loi_read_list);
 252         spin_lock_init(&cli->cl_loi_list_lock);
 253         cli->cl_r_in_flight = 0;
 254         cli->cl_w_in_flight = 0;
 255         spin_lock_init(&cli->cl_read_rpc_hist.oh_lock);
 256         spin_lock_init(&cli->cl_write_rpc_hist.oh_lock);
 257         spin_lock_init(&cli->cl_read_page_hist.oh_lock);
 258         spin_lock_init(&cli->cl_write_page_hist.oh_lock);
 259
 260         memset(&cli->cl_last_write_time, 0,
 261                sizeof(cli->cl_last_write_time));
 262
 263         cli->cl_write_gap_sum = 0;
 264         cli->cl_write_gaps = 0;
 265         cli->cl_write_num = 0;
 266         cli->cl_read_num = 0;
 267         cli->cl_cache_wait_num = 0;
 268         cli->cl_cache_wait_sum = 0;
 269
 270         cli->cl_dirty_num = 0;
 271         cli->cl_dirty_sum = 0;
 272         cli->cl_dirty_av = 0;
 273         cli->cl_dirty_dmax = 0;
 274         cli->cl_dirty_dmin = 0;
 275         cli->cl_sync_rpcs = 0;
 276
 277         if (num_physpages >> (20 - PAGE_SHIFT) <= 128) { /* <= 128 MB */
 278                 cli->cl_max_pages_per_rpc = PTLRPC_MAX_BRW_PAGES / 4;
 279                 cli->cl_max_rpcs_in_flight = OSC_MAX_RIF_DEFAULT / 4;
 280 #if 0
 281         } else if (num_physpages >> (20 - PAGE_SHIFT) <= 512) { /* <= 512 MB */
 282                 cli->cl_max_pages_per_rpc = PTLRPC_MAX_BRW_PAGES / 2;
 283                 cli->cl_max_rpcs_in_flight = OSC_MAX_RIF_DEFAULT / 2;
 284 #endif
 285         } else {
 286                 cli->cl_max_pages_per_rpc = PTLRPC_MAX_BRW_PAGES;
 287                 cli->cl_max_rpcs_in_flight = OSC_MAX_RIF_DEFAULT;
 288         }
 289
 290         rc = ldlm_get_ref();
 291         if (rc) {
 292                 CERROR("ldlm_get_ref failed: %d\n", rc);
 293                 GOTO(err, rc);
 294         }
 295
 296         ptlrpc_init_client(rq_portal, rp_portal, name,
 297                            &obddev->obd_ldlm_client);
 298
 299         imp = class_new_import();
 300         if (imp == NULL)
 301                 GOTO(err_ldlm, rc = -ENOENT);
 302         imp->imp_client = &obddev->obd_ldlm_client;
 303         imp->imp_obd = obddev;
 304         imp->imp_connect_op = connect_op;
 305         imp->imp_generation = 0;
 306         imp->imp_initial_recov = 1;
 307         INIT_LIST_HEAD(&imp->imp_pinger_chain);
 308         memcpy(imp->imp_target_uuid.uuid, lustre_cfg_buf(lcfg, 1),
 309                LUSTRE_CFG_BUFLEN(lcfg, 1));
 310         class_import_put(imp);
 311
 312         rc = client_import_add_conn(imp, &server_uuid, 1);
 313         if (rc) {
 314                 CERROR("can't add initial connection\n");
 315                 GOTO(err_import, rc);
 316         }
 317
 318         cli->cl_import = imp;
 319         cli->cl_max_mds_easize = sizeof(struct lov_mds_md);
 320         cli->cl_max_mds_cookiesize = sizeof(struct llog_cookie);
 321         cli->cl_sandev = to_kdev_t(0);
 322
 323         if (LUSTRE_CFG_BUFLEN(lcfg, 3) > 0) {
 324                 if (!strcmp(lustre_cfg_string(lcfg, 3), "inactive")) {
 325                         CDEBUG(D_HA, "marking %s %s->%s as inactive\n",
 326                                name, obddev->obd_name,
 327                                imp->imp_target_uuid.uuid);
 328                         imp->imp_invalid = 1;
 329
 330                         if (LUSTRE_CFG_BUFLEN(lcfg, 4) > 0)
 331                                 mgmt_name = lustre_cfg_string(lcfg, 4);
 332                 } else {
 333                         mgmt_name = lustre_cfg_string(lcfg, 3);
 334                 }
 335         }
 336 #if 0
 337         if (mgmt_name != NULL) {
 338                 /* Register with management client if we need to. */
 339                 CDEBUG(D_HA, "%s registering with %s for events about %s\n",
 340                        obddev->obd_name, mgmt_name, server_uuid.uuid);
 341
 342                 mgmt_obd = class_name2obd(mgmt_name);
 343                 if (!mgmt_obd) {
 344                         CERROR("can't find mgmtcli %s to register\n",
 345                                mgmt_name);
 346                         GOTO(err_import, rc = -ENOSYS);
 347                 }
 348
 349                 register_f = (mgmtcli_register_for_events_t)symbol_get("mgmtcli_register_for_events");
 350                 if (!register_f) {
 351                         CERROR("can't i_m_g mgmtcli_register_for_events\n");
 352                         GOTO(err_import, rc = -ENOSYS);
 353                 }
 354
 355                 rc = register_f(mgmt_obd, obddev, &imp->imp_target_uuid);
 356                 symbol_put("mgmtcli_register_for_events");
 357
 358                 if (!rc)
 359                         cli->cl_mgmtcli_obd = mgmt_obd;
 360         }
 361 #endif
 362         RETURN(rc);
 363
 364 err_import:
 365         class_destroy_import(imp);
 366 err_ldlm:
 367         ldlm_put_ref(0);
 368 err:
 369         RETURN(rc);
 370
 371 }
 372
 373 int client_obd_cleanup(struct obd_device *obddev, int flags)
 374 {
 375         struct client_obd *cli = &obddev->u.cli;
 376         ENTRY;
 377
 378         if (!cli->cl_import)
 379                 RETURN(-EINVAL);
 380
 381         if (cli->cl_mgmtcli_obd) {
 382                 mgmtcli_deregister_for_events_t dereg_f;
 383
 384                 dereg_f = (mgmtcli_deregister_for_events_t)symbol_get("mgmtcli_deregister_for_events");
 385                 dereg_f(cli->cl_mgmtcli_obd, obddev);
 386                 symbol_put("mgmtcli_deregister_for_events");
 387         }
 388
 389         if (cli->cl_write_gaps) {
 390                 CWARN("%s: [writes num: %lu, reads num: %lu]: %lu write gaps: %lu "
 391                       "av. (usec), %lu total (usec), %lu rpcs falled back to sync\n",
 392                       obddev->obd_name, cli->cl_write_num, cli->cl_read_num,
 393                       cli->cl_write_gaps, cli->cl_write_gap_sum / cli->cl_write_gaps,
 394                       cli->cl_write_gap_sum, cli->cl_sync_rpcs);
 395         }
 396
 397         if (cli->cl_cache_wait_num) {
 398                 CWARN("%s: [cache waits num: %lu]: cache wait av. %lu (usec)\n",
 399                       obddev->obd_name, cli->cl_cache_wait_num,
 400                       cli->cl_cache_wait_sum / cli->cl_cache_wait_num);
 401         }
 402
 403         if (cli->cl_dirty_av) {
 404                 CWARN("%s: pipe loading av. %lu (b), max pipe room %lu (b), pipe "
 405                       "loading av. ratio %lu%%, pipe dirty max %lu (b), pipe dirty "
 406                       "min %lu (b), pipe loading max ratio %lu%%\n", obddev->obd_name,
 407                       cli->cl_dirty_av, cli->cl_dirty_max,
 408                       (cli->cl_dirty_av * 100) / cli->cl_dirty_max,
 409                       cli->cl_dirty_dmax, cli->cl_dirty_dmin,
 410                       (cli->cl_dirty_dmax * 100) / cli->cl_dirty_max);
 411         }
 412
 413         /* Here we try to drop the security structure after destroy import,
 414          * to avoid issue of "sleep in spinlock".
 415          */
 416         class_import_get(cli->cl_import);
 417         class_destroy_import(cli->cl_import);
 418         ptlrpcs_import_drop_sec(cli->cl_import);
 419         class_import_put(cli->cl_import);
 420         cli->cl_import = NULL;
 421
 422         ldlm_put_ref(flags & OBD_OPT_FORCE);
 423         RETURN(0);
 424 }
 425
 426 int client_connect_import(struct lustre_handle *dlm_handle,
 427                           struct obd_device *obd,
 428                           struct obd_uuid *cluuid,
 429                           struct obd_connect_data *conn_data,
 430                           unsigned long connect_flags)
 431 {
 432         struct client_obd *cli = &obd->u.cli;
 433         struct obd_import *imp = cli->cl_import;
 434         struct obd_export *exp;
 435         int rc;
 436         ENTRY;
 437
 438         down(&cli->cl_sem);
 439         rc = class_connect(dlm_handle, obd, cluuid);
 440         if (rc)
 441                 GOTO(out_sem, rc);
 442
 443         cli->cl_conn_count++;
 444         if (cli->cl_conn_count > 1)
 445                 GOTO(out_sem, rc);
 446         exp = class_conn2export(dlm_handle);
 447
 448         if (obd->obd_namespace != NULL)
 449                 CERROR("already have namespace!\n");
 450         obd->obd_namespace = ldlm_namespace_new(obd->obd_name,
 451                                                 LDLM_NAMESPACE_CLIENT);
 452         if (obd->obd_namespace == NULL)
 453                 GOTO(out_disco, rc = -ENOMEM);
 454
 455         rc = ptlrpcs_import_get_sec(imp);
 456         if (rc != 0)
 457                 GOTO(out_ldlm, rc);
 458
 459         imp->imp_dlm_handle = *dlm_handle;
 460         rc = ptlrpc_init_import(imp);
 461         if (rc != 0)
 462                 GOTO(out_ldlm, rc);
 463
 464         imp->imp_connect_flags = connect_flags;
 465         if (conn_data)
 466                 memcpy(&imp->imp_connect_data, conn_data, sizeof(*conn_data));
 467
 468         rc = ptlrpc_connect_import(imp, NULL);
 469         if (rc != 0) {
 470                 LASSERT (imp->imp_state == LUSTRE_IMP_DISCON);
 471                 GOTO(out_ldlm, rc);
 472         }
 473         LASSERT(exp->exp_connection);
 474         ptlrpc_pinger_add_import(imp);
 475         EXIT;
 476
 477         if (rc) {
 478 out_ldlm:
 479                 ldlm_namespace_free(obd->obd_namespace, 0);
 480                 obd->obd_namespace = NULL;
 481 out_disco:
 482                 cli->cl_conn_count--;
 483                 class_disconnect(exp, 0);
 484         } else {
 485                 class_export_put(exp);
 486         }
 487 out_sem:
 488         up(&cli->cl_sem);
 489         return rc;
 490 }
 491
 492 int client_disconnect_export(struct obd_export *exp, unsigned long flags)
 493 {
 494         struct obd_device *obd = class_exp2obd(exp);
 495         struct client_obd *cli = &obd->u.cli;
 496         struct obd_import *imp = cli->cl_import;
 497         int rc = 0, err;
 498         ENTRY;
 499
 500         if (!obd) {
 501                 CERROR("invalid export for disconnect: exp %p cookie "LPX64"\n",
 502                        exp, exp ? exp->exp_handle.h_cookie : -1);
 503                 RETURN(-EINVAL);
 504         }
 505
 506         down(&cli->cl_sem);
 507         if (!cli->cl_conn_count) {
 508                 CERROR("disconnecting disconnected device (%s)\n",
 509                        obd->obd_name);
 510                 GOTO(out_sem, rc = -EINVAL);
 511         }
 512
 513         cli->cl_conn_count--;
 514         if (cli->cl_conn_count)
 515                 GOTO(out_no_disconnect, rc = 0);
 516
 517         /* Some non-replayable imports (MDS's OSCs) are pinged, so just
 518          * delete it regardless.  (It's safe to delete an import that was
 519          * never added.) */
 520         (void)ptlrpc_pinger_del_import(imp);
 521
 522         if (obd->obd_namespace != NULL) {
 523                 /* obd_no_recov == local only */
 524                 ldlm_cli_cancel_unused(obd->obd_namespace, NULL,
 525                                        obd->obd_no_recov, NULL);
 526                 ldlm_namespace_free(obd->obd_namespace, obd->obd_no_recov);
 527                 obd->obd_namespace = NULL;
 528         }
 529
 530         /*
 531          * Yeah, obd_no_recov also (mainly) means "forced shutdown".
 532          */
 533         if (obd->obd_no_recov)
 534                 ptlrpc_invalidate_import(imp, 0);
 535         else
 536                 rc = ptlrpc_disconnect_import(imp);
 537
 538         EXIT;
 539  out_no_disconnect:
 540         err = class_disconnect(exp, 0);
 541         if (!rc && err)
 542                 rc = err;
 543  out_sem:
 544         up(&cli->cl_sem);
 545         RETURN(rc);
 546 }
 547
 548 /* --------------------------------------------------------------------------
 549  * from old lib/target.c
 550  * -------------------------------------------------------------------------- */
 551
 552 int target_handle_reconnect(struct lustre_handle *conn, struct obd_export *exp,
 553                             struct obd_uuid *cluuid, int initial_conn)
 554 {
 555         if (exp->exp_connection && !initial_conn) {
 556                 struct lustre_handle *hdl;
 557                 hdl = &exp->exp_imp_reverse->imp_remote_handle;
 558                 /* Might be a re-connect after a partition. */
 559                 if (!memcmp(&conn->cookie, &hdl->cookie, sizeof conn->cookie)) {
 560                         CERROR("%s reconnecting\n", cluuid->uuid);
 561                         conn->cookie = exp->exp_handle.h_cookie;
 562                         RETURN(EALREADY);
 563                 } else {
 564                         CERROR("%s reconnecting from %s, "
 565                                "handle mismatch (ours "LPX64", theirs "
 566                                LPX64")\n", cluuid->uuid,
 567                                exp->exp_connection->c_remote_uuid.uuid,
 568                                hdl->cookie, conn->cookie);
 569                         memset(conn, 0, sizeof *conn);
 570                         RETURN(-EALREADY);
 571                 }
 572         }
 573
 574         conn->cookie = exp->exp_handle.h_cookie;
 575         CDEBUG(D_INFO, "existing export for UUID '%s' at %p\n",
 576                cluuid->uuid, exp);
 577         CDEBUG(D_IOCTL,"connect: cookie "LPX64"\n", conn->cookie);
 578         RETURN(0);
 579 }
 580
 581 static inline int ptlrpc_peer_is_local(struct ptlrpc_peer *peer)
 582 {
 583         ptl_process_id_t myid;
 584
 585         PtlGetId(peer->peer_ni->pni_ni_h, &myid);
 586         return (memcmp(&peer->peer_id, &myid, sizeof(myid)) == 0);
 587 }
 588
 589 /* To check whether the p_flavor is in deny list or not
 590  * rc:
 591  *      0           not found, pass
 592  *      EPERM       found, refuse
 593  */
 594
 595 static int check_deny_list(struct list_head *head, __u32 flavor)
 596 {
 597         deny_sec_t *p_deny_sec = NULL;
 598         deny_sec_t *n_deny_sec = NULL;
 599
 600         list_for_each_entry_safe(p_deny_sec, n_deny_sec, head, list) {
 601                 if (p_deny_sec->flavor == flavor)
 602                         return -EPERM;
 603         }
 604         return 0;
 605 }
 606
 607 int target_check_deny_sec(struct obd_device *target, struct ptlrpc_request *req)
 608 {
 609         __u32 flavor;
 610         int rc = 0;
 611
 612         flavor = req->rq_req_secflvr;
 613
 614         if (!strcmp(target->obd_type->typ_name, OBD_MDS_DEVICENAME)) {
 615                 spin_lock(&target->u.mds.mds_denylist_lock);
 616                 rc = check_deny_list(&target->u.mds.mds_denylist, flavor);
 617                 spin_unlock(&target->u.mds.mds_denylist_lock);
 618         } else if (!strcmp(target->obd_type->typ_name, OBD_FILTER_DEVICENAME)) {
 619                 spin_lock(&target->u.filter.fo_denylist_lock);
 620                 rc = check_deny_list(&target->u.filter.fo_denylist, flavor);
 621                 spin_unlock(&target->u.filter.fo_denylist_lock);
 622         }
 623
 624         return rc;
 625 }
 626
 627 int target_handle_connect(struct ptlrpc_request *req)
 628 {
 629         unsigned long connect_flags = 0, *cfp;
 630         struct obd_device *target;
 631         struct obd_export *export = NULL;
 632         struct obd_import *revimp;
 633         struct lustre_handle conn;
 634         struct obd_uuid tgtuuid;
 635         struct obd_uuid cluuid;
 636         struct obd_uuid remote_uuid;
 637         struct list_head *p;
 638         struct obd_connect_data *conn_data;
 639         int conn_data_size = sizeof(*conn_data);
 640         char *str, *tmp;
 641         int rc = 0;
 642         unsigned long flags;
 643         int initial_conn = 0;
 644         char peer_str[PTL_NALFMT_SIZE];
 645         const int offset = 1;
 646         ENTRY;
 647
 648         OBD_RACE(OBD_FAIL_TGT_CONN_RACE);
 649
 650         LASSERT_REQSWAB (req, offset + 0);
 651         str = lustre_msg_string(req->rq_reqmsg, offset + 0,
 652                                 sizeof(tgtuuid) - 1);
 653         if (str == NULL) {
 654                 CERROR("bad target UUID for connect\n");
 655                 GOTO(out, rc = -EINVAL);
 656         }
 657
 658         obd_str2uuid (&tgtuuid, str);
 659         target = class_uuid2obd(&tgtuuid);
 660         if (!target)
 661                 target = class_name2obd(str);
 662
 663         if (!target || target->obd_stopping || !target->obd_set_up) {
 664                 CERROR("UUID '%s' is not available for connect from %s\n",
 665                        str, req->rq_peerstr);
 666                 GOTO(out, rc = -ENODEV);
 667         }
 668
 669         /* check the secure deny list of mds/ost */
 670         rc = target_check_deny_sec(target, req);
 671         if (rc != 0)
 672                 GOTO(out, rc);
 673
 674         LASSERT_REQSWAB (req, offset + 1);
 675         str = lustre_msg_string(req->rq_reqmsg, offset + 1, sizeof(cluuid) - 1);
 676         if (str == NULL) {
 677                 CERROR("bad client UUID for connect\n");
 678                 GOTO(out, rc = -EINVAL);
 679         }
 680
 681         obd_str2uuid (&cluuid, str);
 682
 683         /* XXX extract a nettype and format accordingly */
 684         switch (sizeof(ptl_nid_t)) {
 685                 /* NB the casts only avoid compiler warnings */
 686         case 8:
 687                 snprintf((char *)remote_uuid.uuid, sizeof(remote_uuid),
 688                          "NET_"LPX64"_UUID", (__u64)req->rq_peer.peer_id.nid);
 689                 break;
 690         case 4:
 691                 snprintf((char *)remote_uuid.uuid, sizeof(remote_uuid),
 692                          "NET_%x_UUID", (__u32)req->rq_peer.peer_id.nid);
 693                 break;
 694         default:
 695                 LBUG();
 696         }
 697
 698         tmp = lustre_msg_buf(req->rq_reqmsg, offset + 2, sizeof(conn));
 699         if (tmp == NULL)
 700                 GOTO(out, rc = -EPROTO);
 701
 702         memcpy(&conn, tmp, sizeof conn);
 703
 704         cfp = lustre_msg_buf(req->rq_reqmsg, offset + 3, sizeof(unsigned long));
 705         LASSERT(cfp != NULL);
 706         connect_flags = *cfp;
 707
 708         conn_data = lustre_swab_reqbuf(req, offset + 4, sizeof(*conn_data),
 709                                        lustre_swab_connect);
 710         if (!conn_data)
 711                 GOTO(out, rc = -EPROTO);
 712
 713         rc = lustre_pack_reply(req, 1, &conn_data_size, NULL);
 714         if (rc)
 715                 GOTO(out, rc);
 716
 717         if (lustre_msg_get_op_flags(req->rq_reqmsg) & MSG_CONNECT_INITIAL)
 718                 initial_conn = 1;
 719
 720         /* lctl gets a backstage, all-access pass. */
 721         if (obd_uuid_equals(&cluuid, &target->obd_uuid))
 722                 goto dont_check_exports;
 723
 724         spin_lock(&target->obd_dev_lock);
 725         list_for_each(p, &target->obd_exports) {
 726                 export = list_entry(p, struct obd_export, exp_obd_chain);
 727                 if (obd_uuid_equals(&cluuid, &export->exp_client_uuid)) {
 728                         spin_unlock(&target->obd_dev_lock);
 729                         LASSERT(export->exp_obd == target);
 730
 731                         rc = target_handle_reconnect(&conn, export, &cluuid,
 732                                                      initial_conn);
 733                         break;
 734                 }
 735                 export = NULL;
 736         }
 737         /* If we found an export, we already unlocked. */
 738         if (!export) {
 739                 spin_unlock(&target->obd_dev_lock);
 740         } else if (req->rq_export == NULL &&
 741                    atomic_read(&export->exp_rpc_count) > 0) {
 742                 CWARN("%s: refuse connection from %s/%s to 0x%p/%d\n",
 743                       target->obd_name, cluuid.uuid,
 744                       ptlrpc_peernid2str(&req->rq_peer, peer_str),
 745                       export, atomic_read(&export->exp_refcount));
 746                 GOTO(out, rc = -EBUSY);
 747         } else if (req->rq_export != NULL &&
 748                    atomic_read(&export->exp_rpc_count) > 1) {
 749                 CWARN("%s: refuse reconnection from %s@%s to 0x%p/%d\n",
 750                       target->obd_name, cluuid.uuid,
 751                       ptlrpc_peernid2str(&req->rq_peer, peer_str),
 752                       export, atomic_read(&export->exp_rpc_count));
 753                 GOTO(out, rc = -EBUSY);
 754         } else if (req->rq_reqmsg->conn_cnt == 1 && !initial_conn) {
 755                 CERROR("%s reconnected with 1 conn_cnt; cookies not random?\n",
 756                        cluuid.uuid);
 757                 GOTO(out, rc = -EALREADY);
 758         }
 759
 760         /* Tell the client if we're in recovery. */
 761         /* If this is the first client, start the recovery timer */
 762         CWARN("%s: connection from %s@%s/%lu %st"LPU64"\n", target->obd_name,
 763               cluuid.uuid, ptlrpc_peernid2str(&req->rq_peer, peer_str), *cfp,
 764               target->obd_recovering ? "recovering/" : "", conn_data->transno);
 765
 766         if (target->obd_recovering) {
 767                 lustre_msg_add_op_flags(req->rq_repmsg, MSG_CONNECT_RECOVERING);
 768                 target_start_recovery_timer(target);
 769         }
 770
 771 #if 0
 772         /* Tell the client if we support replayable requests */
 773         if (target->obd_replayable)
 774                 lustre_msg_add_op_flags(req->rq_repmsg, MSG_CONNECT_REPLAYABLE);
 775 #endif
 776
 777         if (export == NULL) {
 778                 if (target->obd_recovering) {
 779                         CERROR("%s denying connection for new client %s@%s: "
 780                                "%d clients in recovery for %lds\n", target->obd_name,
 781                                cluuid.uuid,
 782                                ptlrpc_peernid2str(&req->rq_peer, peer_str),
 783                                target->obd_recoverable_clients,
 784                                (target->obd_recovery_timer.expires-jiffies)/HZ);
 785                         rc = -EBUSY;
 786                 } else {
 787  dont_check_exports:
 788                         rc = obd_connect(&conn, target, &cluuid, conn_data,
 789                                          connect_flags);
 790                 }
 791         }
 792
 793         /* Return only the parts of obd_connect_data that we understand, so the
 794          * client knows that we don't understand the rest. */
 795         conn_data->ocd_connect_flags &= OBD_CONNECT_SUPPORTED;
 796         memcpy(lustre_msg_buf(req->rq_repmsg, 0, sizeof(*conn_data)), conn_data,
 797                sizeof(*conn_data));
 798
 799         /* Tell the client if we support replayable requests */
 800         if (target->obd_replayable)
 801                 lustre_msg_add_op_flags(req->rq_repmsg, MSG_CONNECT_REPLAYABLE);
 802
 803         /* If all else goes well, this is our RPC return code. */
 804         req->rq_status = 0;
 805
 806         if (rc && rc != EALREADY)
 807                 GOTO(out, rc);
 808
 809         req->rq_repmsg->handle = conn;
 810
 811         /* If the client and the server are the same node, we will already
 812          * have an export that really points to the client's DLM export,
 813          * because we have a shared handles table.
 814          *
 815          * XXX this will go away when shaver stops sending the "connect" handle
 816          * in the real "remote handle" field of the request --phik 24 Apr 2003
 817          */
 818         if (req->rq_export != NULL)
 819                 class_export_put(req->rq_export);
 820
 821         /* ownership of this export ref transfers to the request */
 822         export = req->rq_export = class_conn2export(&conn);
 823         LASSERT(export != NULL);
 824
 825         spin_lock_irqsave(&export->exp_lock, flags);
 826         if (initial_conn) {
 827                 req->rq_repmsg->conn_cnt = export->exp_conn_cnt + 1;
 828         } else if (export->exp_conn_cnt >= req->rq_reqmsg->conn_cnt) {
 829                 CERROR("%s@%s: already connected at a higher conn_cnt: %d > %d\n",
 830                        cluuid.uuid, ptlrpc_peernid2str(&req->rq_peer, peer_str),
 831                        export->exp_conn_cnt,
 832                        req->rq_reqmsg->conn_cnt);
 833                 spin_unlock_irqrestore(&export->exp_lock, flags);
 834                 GOTO(out, rc = -EALREADY);
 835         }
 836         export->exp_conn_cnt = req->rq_reqmsg->conn_cnt;
 837         spin_unlock_irqrestore(&export->exp_lock, flags);
 838
 839         /* request from liblustre? */
 840         if (lustre_msg_get_op_flags(req->rq_reqmsg) & MSG_CONNECT_LIBCLIENT)
 841                 export->exp_libclient = 1;
 842
 843         if (!(lustre_msg_get_op_flags(req->rq_reqmsg) & MSG_CONNECT_ASYNC) &&
 844             ptlrpc_peer_is_local(&req->rq_peer)) {
 845                 CWARN("%s: exp %p set sync\n", target->obd_name, export);
 846                 export->exp_sync = 1;
 847         } else {
 848                 CDEBUG(D_HA, "%s: exp %p set async\n",target->obd_name,export);
 849                 export->exp_sync = 0;
 850         }
 851
 852         if (export->exp_connection != NULL)
 853                 ptlrpc_put_connection(export->exp_connection);
 854         export->exp_connection = ptlrpc_get_connection(&req->rq_peer,
 855                                                        &remote_uuid);
 856
 857         if (rc == EALREADY) {
 858                 /* We indicate the reconnection in a flag, not an error code. */
 859                 lustre_msg_add_op_flags(req->rq_repmsg, MSG_CONNECT_RECONNECT);
 860                 GOTO(out, rc = 0);
 861         }
 862
 863         spin_lock_bh(&target->obd_processing_task_lock);
 864         if (target->obd_recovering && export->exp_connected == 0) {
 865                 __u64 t = conn_data->transno;
 866                 export->exp_connected = 1;
 867                 if ((lustre_msg_get_op_flags(req->rq_reqmsg) & MSG_CONNECT_TRANSNO)
 868                                 && t < target->obd_next_recovery_transno)
 869                         target->obd_next_recovery_transno = t;
 870                 target->obd_connected_clients++;
 871                 if (target->obd_connected_clients == target->obd_max_recoverable_clients)
 872                         wake_up(&target->obd_next_transno_waitq);
 873         }
 874         spin_unlock_bh(&target->obd_processing_task_lock);
 875
 876         memcpy(&conn, lustre_msg_buf(req->rq_reqmsg, offset + 2, sizeof(conn)),
 877                sizeof(conn));
 878
 879         if (export->exp_imp_reverse != NULL) {
 880                 /* same logic as client_obd_cleanup */
 881                 class_import_get(export->exp_imp_reverse);
 882                 class_destroy_import(export->exp_imp_reverse);
 883                 ptlrpcs_import_drop_sec(export->exp_imp_reverse);
 884                 class_import_put(export->exp_imp_reverse);
 885         }
 886
 887         /* for the rest part, we return -ENOTCONN in case of errors
 888          * in order to let client initialize connection again.
 889          */
 890         revimp = export->exp_imp_reverse = class_new_import();
 891         if (!revimp) {
 892                 CERROR("fail to alloc new reverse import.\n");
 893                 GOTO(out, rc = -ENOTCONN);
 894         }
 895
 896         revimp->imp_connection = ptlrpc_connection_addref(export->exp_connection);
 897         revimp->imp_client = &export->exp_obd->obd_ldlm_client;
 898         revimp->imp_remote_handle = conn;
 899         revimp->imp_obd = target;
 900         revimp->imp_dlm_fake = 1;
 901         revimp->imp_state = LUSTRE_IMP_FULL;
 902
 903         rc = ptlrpcs_import_get_sec(revimp);
 904         if (rc) {
 905                 CERROR("reverse import can not get sec: %d\n", rc);
 906                 class_destroy_import(revimp);
 907                 export->exp_imp_reverse = NULL;
 908                 GOTO(out, rc = -ENOTCONN);
 909         }
 910
 911         class_import_put(revimp);
 912
 913         rc = obd_connect_post(export, initial_conn, connect_flags);
 914 out:
 915         if (rc)
 916                 req->rq_status = rc;
 917         RETURN(rc);
 918 }
 919
 920 int target_handle_disconnect(struct ptlrpc_request *req)
 921 {
 922         struct obd_export *exp;
 923         int rc;
 924         ENTRY;
 925
 926         rc = lustre_pack_reply(req, 0, NULL, NULL);
 927         if (rc)
 928                 RETURN(rc);
 929
 930         /* keep the rq_export around so we can send the reply */
 931         exp = class_export_get(req->rq_export);
 932         req->rq_status = obd_disconnect(exp, 0);
 933         RETURN(0);
 934 }
 935
 936 void target_destroy_export(struct obd_export *exp)
 937 {
 938         /* exports created from last_rcvd data, and "fake"
 939            exports created by lctl don't have an import */
 940         if (exp->exp_imp_reverse != NULL) {
 941                 ptlrpcs_import_drop_sec(exp->exp_imp_reverse);
 942                 class_destroy_import(exp->exp_imp_reverse);
 943         }
 944
 945         /* We cancel locks at disconnect time, but this will catch any locks
 946          * granted in a race with recovery-induced disconnect. */
 947         if (exp->exp_obd->obd_namespace != NULL)
 948                 ldlm_cancel_locks_for_export(exp);
 949 }
 950
 951 /*
 952  * Recovery functions
 953  */
 954
 955 struct ptlrpc_request *
 956 ptlrpc_clone_req( struct ptlrpc_request *orig_req)
 957 {
 958         struct ptlrpc_request *copy_req;
 959         struct lustre_msg *copy_reqmsg;
 960
 961         OBD_ALLOC(copy_req, sizeof *copy_req);
 962         if (!copy_req)
 963                 return NULL;
 964         OBD_ALLOC(copy_reqmsg, orig_req->rq_reqlen);
 965         if (!copy_reqmsg){
 966                 OBD_FREE(copy_req, sizeof *copy_req);
 967                 return NULL;
 968         }
 969
 970         memcpy(copy_req, orig_req, sizeof *copy_req);
 971         memcpy(copy_reqmsg, orig_req->rq_reqmsg, orig_req->rq_reqlen);
 972         /* the copied req takes over the reply state and security data */
 973         orig_req->rq_reply_state = NULL;
 974         orig_req->rq_svcsec_data = NULL;
 975
 976         copy_req->rq_reqmsg = copy_reqmsg;
 977         class_export_get(copy_req->rq_export);
 978         INIT_LIST_HEAD(&copy_req->rq_list);
 979
 980         return copy_req;
 981 }
 982
 983 void ptlrpc_free_clone( struct ptlrpc_request *req)
 984 {
 985         if (req->rq_svcsec)
 986                 svcsec_cleanup_req(req);
 987
 988         class_export_put(req->rq_export);
 989         list_del(&req->rq_list);
 990         OBD_FREE(req->rq_reqmsg, req->rq_reqlen);
 991         OBD_FREE(req, sizeof *req);
 992 }
 993
 994 static void target_release_saved_req(struct ptlrpc_request *req)
 995 {
 996         if (req->rq_svcsec)
 997                 svcsec_cleanup_req(req);
 998
 999         class_export_put(req->rq_export);
1000         OBD_FREE(req->rq_reqmsg, req->rq_reqlen);
1001         OBD_FREE(req, sizeof *req);
1002 }
1003
1004 static void target_finish_recovery(struct obd_device *obd)
1005 {
1006         int rc;
1007
1008         ldlm_reprocess_all_ns(obd->obd_namespace);
1009
1010         /* when recovery finished, cleanup orphans on mds and ost */
1011         if (OBT(obd) && OBP(obd, postrecov)) {
1012                 rc = OBP(obd, postrecov)(obd);
1013                 if (rc >= 0)
1014                         CWARN("%s: all clients recovered, %d MDS "
1015                               "orphans deleted\n", obd->obd_name, rc);
1016                 else
1017                         CERROR("postrecov failed %d\n", rc);
1018         }
1019
1020         obd->obd_recovery_end = LTIME_S(CURRENT_TIME);
1021         return;
1022 }
1023
1024 static void abort_req_replay_queue(struct obd_device *obd)
1025 {
1026         struct ptlrpc_request *req;
1027         struct list_head *tmp, *n;
1028         int rc;
1029
1030         list_for_each_safe(tmp, n, &obd->obd_req_replay_queue) {
1031                 req = list_entry(tmp, struct ptlrpc_request, rq_list);
1032                 list_del(&req->rq_list);
1033                 DEBUG_REQ(D_ERROR, req, "aborted:");
1034                 req->rq_status = -ENOTCONN;
1035                 req->rq_type = PTL_RPC_MSG_ERR;
1036                 rc = lustre_pack_reply(req, 0, NULL, NULL);
1037                 if (rc == 0) {
1038                         ptlrpc_reply(req);
1039                 } else {
1040                         DEBUG_REQ(D_ERROR, req,
1041                                   "packing failed for abort-reply; skipping");
1042                 }
1043                 target_release_saved_req(req);
1044         }
1045 }
1046
1047 static void abort_lock_replay_queue(struct obd_device *obd)
1048 {
1049         struct ptlrpc_request *req;
1050         struct list_head *tmp, *n;
1051         int rc;
1052
1053         list_for_each_safe(tmp, n, &obd->obd_lock_replay_queue) {
1054                 req = list_entry(tmp, struct ptlrpc_request, rq_list);
1055                 list_del(&req->rq_list);
1056                 DEBUG_REQ(D_ERROR, req, "aborted:");
1057                 req->rq_status = -ENOTCONN;
1058                 req->rq_type = PTL_RPC_MSG_ERR;
1059                 rc = lustre_pack_reply(req, 0, NULL, NULL);
1060                 if (rc == 0) {
1061                         ptlrpc_reply(req);
1062                 } else {
1063                         DEBUG_REQ(D_ERROR, req,
1064                                   "packing failed for abort-reply; skipping");
1065                 }
1066                 target_release_saved_req(req);
1067         }
1068 }
1069
1070 /* Called from a cleanup function if the device is being cleaned up
1071    forcefully.  The exports should all have been disconnected already,
1072    the only thing left to do is
1073      - clear the recovery flags
1074      - cancel the timer
1075      - free queued requests and replies, but don't send replies
1076    Because the obd_stopping flag is set, no new requests should be received.
1077
1078 */
1079 void target_cleanup_recovery(struct obd_device *obd)
1080 {
1081         struct list_head *tmp, *n;
1082         struct ptlrpc_request *req;
1083
1084         spin_lock_bh(&obd->obd_processing_task_lock);
1085         if (!obd->obd_recovering) {
1086                 spin_unlock_bh(&obd->obd_processing_task_lock);
1087                 EXIT;
1088                 return;
1089         }
1090         obd->obd_recovering = obd->obd_abort_recovery = 0;
1091         target_cancel_recovery_timer(obd);
1092         spin_unlock_bh(&obd->obd_processing_task_lock);
1093
1094         list_for_each_safe(tmp, n, &obd->obd_req_replay_queue) {
1095                 req = list_entry(tmp, struct ptlrpc_request, rq_list);
1096                 list_del(&req->rq_list);
1097                 LASSERT (req->rq_reply_state == 0);
1098                 target_release_saved_req(req);
1099         }
1100         list_for_each_safe(tmp, n, &obd->obd_lock_replay_queue) {
1101                 req = list_entry(tmp, struct ptlrpc_request, rq_list);
1102                 list_del(&req->rq_list);
1103                 LASSERT (req->rq_reply_state == 0);
1104                 target_release_saved_req(req);
1105         }
1106         list_for_each_safe(tmp, n, &obd->obd_final_req_queue) {
1107                 req = list_entry(tmp, struct ptlrpc_request, rq_list);
1108                 list_del(&req->rq_list);
1109                 LASSERT (req->rq_reply_state == 0);
1110                 target_release_saved_req(req);
1111         }
1112 }
1113
1114 #if 0
1115 static void target_abort_recovery(void *data)
1116 {
1117         struct obd_device *obd = data;
1118
1119         LASSERT(!obd->obd_recovering);
1120
1121         class_disconnect_stale_exports(obd, 0);
1122
1123         CERROR("%s: recovery period over; disconnecting unfinished clients.\n",
1124                obd->obd_name);
1125
1126         abort_recovery_queue(obd);
1127         target_finish_recovery(obd);
1128         ptlrpc_run_recovery_over_upcall(obd);
1129 }
1130 #endif
1131
1132 static void target_recovery_expired(unsigned long castmeharder)
1133 {
1134         struct obd_device *obd = (struct obd_device *)castmeharder;
1135         spin_lock_bh(&obd->obd_processing_task_lock);
1136         if (obd->obd_recovering)
1137                 obd->obd_abort_recovery = 1;
1138
1139         wake_up(&obd->obd_next_transno_waitq);
1140         spin_unlock_bh(&obd->obd_processing_task_lock);
1141 }
1142
1143
1144 /* obd_processing_task_lock should be held */
1145 void target_cancel_recovery_timer(struct obd_device *obd)
1146 {
1147         CDEBUG(D_HA, "%s: cancel recovery timer\n", obd->obd_name);
1148         del_timer(&obd->obd_recovery_timer);
1149 }
1150
1151 #ifdef __KERNEL__
1152 static void reset_recovery_timer(struct obd_device *obd)
1153 {
1154         spin_lock_bh(&obd->obd_processing_task_lock);
1155         if (!obd->obd_recovering) {
1156                 spin_unlock_bh(&obd->obd_processing_task_lock);
1157                 return;
1158         }
1159         CDEBUG(D_HA, "timer will expire in %u seconds\n",
1160                OBD_RECOVERY_TIMEOUT / HZ);
1161         mod_timer(&obd->obd_recovery_timer, jiffies + OBD_RECOVERY_TIMEOUT);
1162         spin_unlock_bh(&obd->obd_processing_task_lock);
1163 }
1164 #endif
1165
1166 /* Only start it the first time called */
1167 void target_start_recovery_timer(struct obd_device *obd)
1168 {
1169         spin_lock_bh(&obd->obd_processing_task_lock);
1170         if (!obd->obd_recovering || timer_pending(&obd->obd_recovery_timer)) {
1171                 spin_unlock_bh(&obd->obd_processing_task_lock);
1172                 return;
1173         }
1174         CWARN("%s: starting recovery timer (%us)\n", obd->obd_name,
1175                OBD_RECOVERY_TIMEOUT / HZ);
1176         obd->obd_recovery_timer.function = target_recovery_expired;
1177         obd->obd_recovery_timer.data = (unsigned long)obd;
1178         mod_timer(&obd->obd_recovery_timer, jiffies + OBD_RECOVERY_TIMEOUT);
1179         spin_unlock_bh(&obd->obd_processing_task_lock);
1180 }
1181
1182 #ifdef __KERNEL__
1183 static int check_for_next_transno(struct obd_device *obd)
1184 {
1185         struct ptlrpc_request *req = NULL;
1186         int wake_up = 0, connected, completed, queue_len, max;
1187         __u64 next_transno, req_transno;
1188
1189         spin_lock_bh(&obd->obd_processing_task_lock);
1190         if (!list_empty(&obd->obd_req_replay_queue)) {
1191                 req = list_entry(obd->obd_req_replay_queue.next,
1192                                  struct ptlrpc_request, rq_list);
1193                 req_transno = req->rq_reqmsg->transno;
1194         } else {
1195                 req_transno = 0;
1196         }
1197
1198         max = obd->obd_max_recoverable_clients;
1199         connected = obd->obd_connected_clients;
1200         completed = max - obd->obd_recoverable_clients;
1201         queue_len = obd->obd_requests_queued_for_recovery;
1202         next_transno = obd->obd_next_recovery_transno;
1203
1204         CDEBUG(D_HA,"max: %d, connected: %d, completed: %d, queue_len: %d, "
1205                "req_transno: "LPU64", next_transno: "LPU64"\n",
1206                max, connected, completed, queue_len, req_transno, next_transno);
1207         if (obd->obd_abort_recovery) {
1208                 CDEBUG(D_HA, "waking for aborted recovery\n");
1209                 wake_up = 1;
1210         } else if (atomic_read(&obd->obd_req_replay_clients) == 0) {
1211                 CDEBUG(D_HA, "waking for completed recovery\n");
1212                 wake_up = 1;
1213         } else if (req_transno == next_transno) {
1214                 CDEBUG(D_HA, "waking for next ("LPD64")\n", next_transno);
1215                 wake_up = 1;
1216         } else if (queue_len + completed == max) {
1217                 LASSERT(req->rq_reqmsg->transno >= next_transno);
1218                 CDEBUG(req_transno > obd->obd_last_committed ? D_ERROR : D_HA,
1219                        "waking for skipped transno (skip: "LPD64
1220                        ", ql: %d, comp: %d, conn: %d, next: "LPD64")\n",
1221                        next_transno, queue_len, completed, max, req_transno);
1222                 obd->obd_next_recovery_transno = req_transno;
1223                 wake_up = 1;
1224         } else if (queue_len == atomic_read(&obd->obd_req_replay_clients)) {
1225                 /* some clients haven't connected in time, but we can try
1226                  * to replay requests that demand on already committed ones
1227                  * also, we can replay first non-committed transation */
1228                 LASSERT(req_transno != 0);
1229                 if (req_transno == obd->obd_last_committed + 1) {
1230                         obd->obd_next_recovery_transno = req_transno;
1231                 } else if (req_transno > obd->obd_last_committed) {
1232                         /* can't continue recovery: have no needed transno */
1233                         obd->obd_abort_recovery = 1;
1234                         CDEBUG(D_ERROR, "abort due to missed clients. max: %d, "
1235                                "connected: %d, completed: %d, queue_len: %d, "
1236                                "req_transno: "LPU64", next_transno: "LPU64"\n",
1237                                max, connected, completed, queue_len,
1238                                req_transno, next_transno);
1239                 }
1240                 wake_up = 1;
1241         }
1242         spin_unlock_bh(&obd->obd_processing_task_lock);
1243
1244         return wake_up;
1245 }
1246
1247 static struct ptlrpc_request *
1248 target_next_replay_req(struct obd_device *obd)
1249 {
1250         struct l_wait_info lwi = { 0 };
1251         struct ptlrpc_request *req;
1252
1253         CDEBUG(D_HA, "Waiting for transno "LPD64"\n",
1254                obd->obd_next_recovery_transno);
1255         l_wait_event(obd->obd_next_transno_waitq,
1256                      check_for_next_transno(obd), &lwi);
1257
1258         spin_lock_bh(&obd->obd_processing_task_lock);
1259         if (obd->obd_abort_recovery) {
1260                 req = NULL;
1261         } else if (!list_empty(&obd->obd_req_replay_queue)) {
1262                 req = list_entry(obd->obd_req_replay_queue.next,
1263                                  struct ptlrpc_request, rq_list);
1264                 list_del_init(&req->rq_list);
1265                 obd->obd_requests_queued_for_recovery--;
1266         } else {
1267                 req = NULL;
1268         }
1269         spin_unlock_bh(&obd->obd_processing_task_lock);
1270         return req;
1271 }
1272
1273 static int check_for_next_lock(struct obd_device *obd)
1274 {
1275         struct ptlrpc_request *req = NULL;
1276         int wake_up = 0;
1277
1278         spin_lock_bh(&obd->obd_processing_task_lock);
1279         if (!list_empty(&obd->obd_lock_replay_queue)) {
1280                 req = list_entry(obd->obd_lock_replay_queue.next,
1281                                  struct ptlrpc_request, rq_list);
1282                 CDEBUG(D_HA, "waking for next lock\n");
1283                 wake_up = 1;
1284         } else if (atomic_read(&obd->obd_lock_replay_clients) == 0) {
1285                 CDEBUG(D_HA, "waking for completed lock replay\n");
1286                 wake_up = 1;
1287         } else if (obd->obd_abort_recovery) {
1288                 CDEBUG(D_HA, "waking for aborted recovery\n");
1289                 wake_up = 1;
1290         }
1291         spin_unlock_bh(&obd->obd_processing_task_lock);
1292
1293         return wake_up;
1294 }
1295
1296 static struct ptlrpc_request *
1297 target_next_replay_lock(struct obd_device *obd)
1298 {
1299         struct l_wait_info lwi = { 0 };
1300         struct ptlrpc_request *req;
1301
1302         CDEBUG(D_HA, "Waiting for lock\n");
1303         l_wait_event(obd->obd_next_transno_waitq,
1304                      check_for_next_lock(obd), &lwi);
1305
1306         spin_lock_bh(&obd->obd_processing_task_lock);
1307         if (obd->obd_abort_recovery) {
1308                 req = NULL;
1309         } else if (!list_empty(&obd->obd_lock_replay_queue)) {
1310                 req = list_entry(obd->obd_lock_replay_queue.next,
1311                                  struct ptlrpc_request, rq_list);
1312                 list_del_init(&req->rq_list);
1313         } else {
1314                 req = NULL;
1315         }
1316         spin_unlock_bh(&obd->obd_processing_task_lock);
1317         return req;
1318 }
1319
1320 static struct ptlrpc_request *
1321 target_next_final_ping(struct obd_device *obd)
1322 {
1323         struct ptlrpc_request *req;
1324
1325         spin_lock_bh(&obd->obd_processing_task_lock);
1326         if (!list_empty(&obd->obd_final_req_queue)) {
1327                 req = list_entry(obd->obd_final_req_queue.next,
1328                                  struct ptlrpc_request, rq_list);
1329                 list_del_init(&req->rq_list);
1330         } else {
1331                 req = NULL;
1332         }
1333         spin_unlock_bh(&obd->obd_processing_task_lock);
1334         return req;
1335 }
1336
1337 static int req_replay_done(struct obd_export *exp)
1338 {
1339         if (exp->exp_req_replay_needed)
1340                 return 0;
1341         return 1;
1342 }
1343
1344 static int lock_replay_done(struct obd_export *exp)
1345 {
1346         if (exp->exp_lock_replay_needed)
1347                 return 0;
1348         return 1;
1349 }
1350
1351 static int connect_done(struct obd_export *exp)
1352 {
1353         if (exp->exp_connected)
1354                 return 1;
1355         return 0;
1356 }
1357
1358 static int check_for_clients(struct obd_device *obd)
1359 {
1360         if (obd->obd_abort_recovery)
1361                 return 1;
1362         LASSERT(obd->obd_connected_clients <= obd->obd_max_recoverable_clients);
1363         if (obd->obd_connected_clients == obd->obd_max_recoverable_clients)
1364                 return 1;
1365         return 0;
1366 }
1367
1368 static int target_recovery_thread(void *arg)
1369 {
1370         struct obd_device *obd = arg;
1371         struct ptlrpc_request *req;
1372         struct target_recovery_data *trd = &obd->obd_recovery_data;
1373         char peer_str[PTL_NALFMT_SIZE];
1374         struct l_wait_info lwi = { 0 };
1375         unsigned long delta;
1376         unsigned long flags;
1377         ENTRY;
1378
1379         kportal_daemonize("tgt-recov");
1380
1381         SIGNAL_MASK_LOCK(current, flags);
1382         sigfillset(&current->blocked);
1383         RECALC_SIGPENDING;
1384         SIGNAL_MASK_UNLOCK(current, flags);
1385
1386         CERROR("%s: started recovery thread pid %d\n", obd->obd_name,
1387                current->pid);
1388         trd->trd_processing_task = current->pid;
1389
1390         obd->obd_recovering = 1;
1391         complete(&trd->trd_starting);
1392
1393         /* first of all, we have to know the first transno to replay */
1394         obd->obd_abort_recovery = 0;
1395         l_wait_event(obd->obd_next_transno_waitq,
1396                      check_for_clients(obd), &lwi);
1397
1398         spin_lock_bh(&obd->obd_processing_task_lock);
1399         target_cancel_recovery_timer(obd);
1400         spin_unlock_bh(&obd->obd_processing_task_lock);
1401
1402         /* If some clients haven't connected in time, evict them */
1403         if (obd->obd_abort_recovery) {
1404                 int stale;
1405                 CDEBUG(D_ERROR, "few clients haven't connect in time (%d/%d),"
1406                        "evict them ...\n", obd->obd_connected_clients,
1407                        obd->obd_max_recoverable_clients);
1408                 obd->obd_abort_recovery = 0;
1409                 stale = class_disconnect_stale_exports(obd, connect_done, 0);
1410                 atomic_sub(stale, &obd->obd_req_replay_clients);
1411                 atomic_sub(stale, &obd->obd_lock_replay_clients);
1412         }
1413
1414         /* next stage: replay requests */
1415         delta = jiffies;
1416         obd->obd_req_replaying = 1;
1417         CDEBUG(D_ERROR, "1: request replay stage - %d clients from t"LPU64"\n",
1418               atomic_read(&obd->obd_req_replay_clients),
1419               obd->obd_next_recovery_transno);
1420         while ((req = target_next_replay_req(obd))) {
1421                 LASSERT(trd->trd_processing_task == current->pid);
1422                 DEBUG_REQ(D_HA, req, "processing t"LPD64" from %s",
1423                           req->rq_reqmsg->transno,
1424                           ptlrpc_peernid2str(&req->rq_peer, peer_str));
1425                 (void)trd->trd_recovery_handler(req);
1426                 obd->obd_replayed_requests++;
1427                 reset_recovery_timer(obd);
1428                 /* bug 1580: decide how to properly sync() in recovery*/
1429                 //mds_fsync_super(mds->mds_sb);
1430                 ptlrpc_free_clone(req);
1431                 spin_lock_bh(&obd->obd_processing_task_lock);
1432                 obd->obd_next_recovery_transno++;
1433                 spin_unlock_bh(&obd->obd_processing_task_lock);
1434         }
1435
1436         spin_lock_bh(&obd->obd_processing_task_lock);
1437         target_cancel_recovery_timer(obd);
1438         spin_unlock_bh(&obd->obd_processing_task_lock);
1439
1440         /* If some clients haven't replayed requests in time, evict them */
1441         if (obd->obd_abort_recovery) {
1442                 int stale;
1443                 CDEBUG(D_ERROR, "req replay timed out, aborting ...\n");
1444                 obd->obd_abort_recovery = 0;
1445                 stale = class_disconnect_stale_exports(obd, req_replay_done, 0);
1446                 atomic_sub(stale, &obd->obd_lock_replay_clients);
1447                 abort_req_replay_queue(obd);
1448                 /* XXX for debuggin tests 11 and 17 */
1449                 LBUG();
1450         }
1451
1452         /* The second stage: replay locks */
1453         CDEBUG(D_ERROR, "2: lock replay stage - %d clients\n",
1454               atomic_read(&obd->obd_lock_replay_clients));
1455         while ((req = target_next_replay_lock(obd))) {
1456                 LASSERT(trd->trd_processing_task == current->pid);
1457                 DEBUG_REQ(D_HA, req, "processing lock from %s: ",
1458                           ptlrpc_peernid2str(&req->rq_peer, peer_str));
1459                 (void)trd->trd_recovery_handler(req);
1460                 reset_recovery_timer(obd);
1461                 ptlrpc_free_clone(req);
1462                 obd->obd_replayed_locks++;
1463         }
1464
1465         spin_lock_bh(&obd->obd_processing_task_lock);
1466         target_cancel_recovery_timer(obd);
1467         spin_unlock_bh(&obd->obd_processing_task_lock);
1468
1469         /* If some clients haven't replayed requests in time, evict them */
1470         if (obd->obd_abort_recovery) {
1471                 int stale;
1472                 CERROR("lock replay timed out, aborting ...\n");
1473                 obd->obd_abort_recovery = 0;
1474                 stale = class_disconnect_stale_exports(obd, lock_replay_done, 0);
1475                 abort_lock_replay_queue(obd);
1476         }
1477
1478         /* We drop recoverying flag to forward all new requests
1479          * to regular mds_handle() since now */
1480         spin_lock_bh(&obd->obd_processing_task_lock);
1481         obd->obd_recovering = 0;
1482         spin_unlock_bh(&obd->obd_processing_task_lock);
1483
1484         /* The third stage: reply on final pings */
1485         CDEBUG(D_ERROR, "3: final stage - process recovery completion pings\n");
1486         while ((req = target_next_final_ping(obd))) {
1487                 LASSERT(trd->trd_processing_task == current->pid);
1488                 DEBUG_REQ(D_HA, req, "processing final ping from %s: ",
1489                           ptlrpc_peernid2str(&req->rq_peer, peer_str));
1490                 (void)trd->trd_recovery_handler(req);
1491                 ptlrpc_free_clone(req);
1492         }
1493
1494         delta = (jiffies - delta) / HZ;
1495         CDEBUG(D_ERROR,"4: recovery completed in %lus - %d/%d reqs/locks\n",
1496               delta, obd->obd_replayed_requests, obd->obd_replayed_locks);
1497         if (delta > obd_timeout * 2) {
1498                 CWARN("too long recovery - read logs\n");
1499                 portals_debug_dumplog();
1500         }
1501         target_finish_recovery(obd);
1502
1503         trd->trd_processing_task = 0;
1504         complete(&trd->trd_finishing);
1505         return 0;
1506 }
1507
1508 int target_start_recovery_thread(struct obd_device *obd, svc_handler_t handler)
1509 {
1510         int rc = 0;
1511         struct target_recovery_data *trd = &obd->obd_recovery_data;
1512
1513         memset(trd, 0, sizeof(*trd));
1514         init_completion(&trd->trd_starting);
1515         init_completion(&trd->trd_finishing);
1516         trd->trd_recovery_handler = handler;
1517
1518         if (kernel_thread(target_recovery_thread, obd, 0) > 0) {
1519                 wait_for_completion(&trd->trd_starting);
1520                 LASSERT(obd->obd_recovering != 0);
1521         } else
1522                 rc = -ECHILD;
1523
1524         return rc;
1525 }
1526
1527 void target_stop_recovery_thread(struct obd_device *obd)
1528 {
1529         spin_lock_bh(&obd->obd_processing_task_lock);
1530         if (obd->obd_recovery_data.trd_processing_task > 0) {
1531                 struct target_recovery_data *trd = &obd->obd_recovery_data;
1532                 CERROR("%s: aborting recovery\n", obd->obd_name);
1533                 obd->obd_abort_recovery = 1;
1534                 wake_up(&obd->obd_next_transno_waitq);
1535                 spin_unlock_bh(&obd->obd_processing_task_lock);
1536                 wait_for_completion(&trd->trd_finishing);
1537         } else {
1538                 spin_unlock_bh(&obd->obd_processing_task_lock);
1539         }
1540 }
1541 #endif
1542
1543 int target_process_req_flags(struct obd_device *obd, struct ptlrpc_request *req)
1544 {
1545         struct obd_export *exp = req->rq_export;
1546         LASSERT(exp != NULL);
1547         if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REQ_REPLAY_DONE) {
1548                 /* client declares he's ready to replay locks */
1549                 spin_lock_bh(&obd->obd_processing_task_lock);
1550                 if (exp->exp_req_replay_needed) {
1551                         LASSERT(atomic_read(&obd->obd_req_replay_clients) > 0);
1552                         exp->exp_req_replay_needed = 0;
1553                         atomic_dec(&obd->obd_req_replay_clients);
1554                         obd->obd_recoverable_clients--;
1555                         if (atomic_read(&obd->obd_req_replay_clients) == 0)
1556                                 CDEBUG(D_HA, "all clients have replayed reqs\n");
1557                         wake_up(&obd->obd_next_transno_waitq);
1558                 }
1559                 spin_unlock_bh(&obd->obd_processing_task_lock);
1560         }
1561         if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_LOCK_REPLAY_DONE) {
1562                 /* client declares he's ready to complete recovery
1563                  * so, we put the request on th final queue */
1564                 spin_lock_bh(&obd->obd_processing_task_lock);
1565                 if (exp->exp_lock_replay_needed) {
1566                         LASSERT(atomic_read(&obd->obd_lock_replay_clients) > 0);
1567                         exp->exp_lock_replay_needed = 0;
1568                         atomic_dec(&obd->obd_lock_replay_clients);
1569                         if (atomic_read(&obd->obd_lock_replay_clients) == 0)
1570                                 CDEBUG(D_HA, "all clients have replayed locks\n");
1571                         wake_up(&obd->obd_next_transno_waitq);
1572                 }
1573                 spin_unlock_bh(&obd->obd_processing_task_lock);
1574         }
1575
1576         return 0;
1577 }
1578
1579 int target_queue_recovery_request(struct ptlrpc_request *req,
1580                                   struct obd_device *obd)
1581 {
1582         struct list_head *tmp;
1583         int inserted = 0;
1584         __u64 transno = req->rq_reqmsg->transno;
1585
1586         if (obd->obd_recovery_data.trd_processing_task == current->pid) {
1587                 /* Processing the queue right now, don't re-add. */
1588                 return 1;
1589         }
1590
1591         target_process_req_flags(obd, req);
1592
1593         if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_LOCK_REPLAY_DONE) {
1594                 /* client declares he's ready to complete recovery
1595                  * so, we put the request on th final queue */
1596                 req = ptlrpc_clone_req(req);
1597                 if (req == NULL)
1598                         return -ENOMEM;
1599                 DEBUG_REQ(D_HA, req, "queue final req");
1600                 spin_lock_bh(&obd->obd_processing_task_lock);
1601                 list_add_tail(&req->rq_list, &obd->obd_final_req_queue);
1602                 spin_unlock_bh(&obd->obd_processing_task_lock);
1603                 return 0;
1604         }
1605         if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REQ_REPLAY_DONE) {
1606                 /* client declares he's ready to replay locks */
1607                 req = ptlrpc_clone_req(req);
1608                 if (req == NULL)
1609                         return -ENOMEM;
1610                 DEBUG_REQ(D_HA, req, "queue lock replay req");
1611                 spin_lock_bh(&obd->obd_processing_task_lock);
1612                 list_add_tail(&req->rq_list, &obd->obd_lock_replay_queue);
1613                 spin_unlock_bh(&obd->obd_processing_task_lock);
1614                 wake_up(&obd->obd_next_transno_waitq);
1615                 return 0;
1616         }
1617
1618
1619         /* CAVEAT EMPTOR: The incoming request message has been swabbed
1620          * (i.e. buflens etc are in my own byte order), but type-dependent
1621          * buffers (eg mds_body, ost_body etc) have NOT been swabbed. */
1622
1623         if (!transno) {
1624                 INIT_LIST_HEAD(&req->rq_list);
1625                 DEBUG_REQ(D_HA, req, "not queueing");
1626                 return 1;
1627         }
1628
1629
1630         /* If we're processing the queue, we want don't want to queue this
1631          * message.
1632          *
1633          * Also, if this request has a transno less than the one we're waiting
1634          * for, we should process it now.  It could (and currently always will)
1635          * be an open request for a descriptor that was opened some time ago.
1636          *
1637          * Also, a resent, replayed request that has already been
1638          * handled will pass through here and be processed immediately.
1639          */
1640         spin_lock_bh(&obd->obd_processing_task_lock);
1641         if (transno < obd->obd_next_recovery_transno && obd->obd_req_replaying) {
1642                 /* Processing the queue right now, don't re-add. */
1643                 LASSERT(list_empty(&req->rq_list));
1644                 spin_unlock_bh(&obd->obd_processing_task_lock);
1645                 return 1;
1646         }
1647         spin_unlock_bh(&obd->obd_processing_task_lock);
1648
1649         /* A resent, replayed request that is still on the queue; just drop it.
1650            The queued request will handle this. */
1651         if ((lustre_msg_get_flags(req->rq_reqmsg) & (MSG_RESENT | MSG_REPLAY))
1652             == (MSG_RESENT | MSG_REPLAY)) {
1653                 DEBUG_REQ(D_ERROR, req, "dropping resent queued req");
1654                 return 0;
1655         }
1656
1657         req = ptlrpc_clone_req(req);
1658         if (req == NULL)
1659                 return -ENOMEM;
1660
1661         spin_lock_bh(&obd->obd_processing_task_lock);
1662
1663         /* XXX O(n^2) */
1664         list_for_each(tmp, &obd->obd_req_replay_queue) {
1665                 struct ptlrpc_request *reqiter =
1666                         list_entry(tmp, struct ptlrpc_request, rq_list);
1667
1668                 if (reqiter->rq_reqmsg->transno > transno) {
1669                         list_add_tail(&req->rq_list, &reqiter->rq_list);
1670                         inserted = 1;
1671                         break;
1672                 }
1673         }
1674
1675         if (!inserted)
1676                 list_add_tail(&req->rq_list, &obd->obd_req_replay_queue);
1677
1678         obd->obd_requests_queued_for_recovery++;
1679         wake_up(&obd->obd_next_transno_waitq);
1680         spin_unlock_bh(&obd->obd_processing_task_lock);
1681         return 0;
1682 }
1683
1684 struct obd_device * target_req2obd(struct ptlrpc_request *req)
1685 {
1686         return req->rq_export->exp_obd;
1687 }
1688
1689 int
1690 target_send_reply_msg (struct ptlrpc_request *req, int rc, int fail_id)
1691 {
1692         if (OBD_FAIL_CHECK(fail_id | OBD_FAIL_ONCE)) {
1693                 obd_fail_loc |= OBD_FAIL_ONCE | OBD_FAILED;
1694                 DEBUG_REQ(D_ERROR, req, "dropping reply");
1695                 /* NB this does _not_ send with ACK disabled, to simulate
1696                  * sending OK, but timing out for the ACK */
1697                 if (req->rq_reply_state != NULL) {
1698                         if (!req->rq_reply_state->rs_difficult) {
1699                                 lustre_free_reply_state (req->rq_reply_state);
1700                                 req->rq_reply_state = NULL;
1701                         } else {
1702                                 struct ptlrpc_service *svc =
1703                                         req->rq_rqbd->rqbd_srv_ni->sni_service;
1704                                 atomic_inc(&svc->srv_outstanding_replies);
1705                         }
1706                 }
1707                 return (-ECOMM);
1708         }
1709
1710         if (rc) {
1711                 req->rq_status = rc;
1712                 return (ptlrpc_error(req));
1713         } else {
1714                 DEBUG_REQ(D_NET, req, "sending reply");
1715         }
1716
1717         return (ptlrpc_send_reply(req, 1));
1718 }
1719
1720 void
1721 target_send_reply(struct ptlrpc_request *req, int rc, int fail_id)
1722 {
1723         int                        netrc;
1724         unsigned long              flags;
1725         struct ptlrpc_reply_state *rs;
1726         struct obd_device         *obd;
1727         struct obd_export         *exp;
1728         struct ptlrpc_srv_ni      *sni;
1729         struct ptlrpc_service     *svc;
1730
1731         sni = req->rq_rqbd->rqbd_srv_ni;
1732         svc = sni->sni_service;
1733
1734         rs = req->rq_reply_state;
1735         if (rs == NULL || !rs->rs_difficult) {
1736                 /* The easy case; no notifiers and reply_out_callback()
1737                  * cleans up (i.e. we can't look inside rs after a
1738                  * successful send) */
1739                 netrc = target_send_reply_msg (req, rc, fail_id);
1740
1741                 LASSERT (netrc == 0 || req->rq_reply_state == NULL);
1742                 return;
1743         }
1744
1745         /* must be an export if locks saved */
1746         LASSERT (req->rq_export != NULL);
1747         /* req/reply consistent */
1748         LASSERT (rs->rs_srv_ni == sni);
1749
1750         /* "fresh" reply */
1751         LASSERT (!rs->rs_scheduled);
1752         LASSERT (!rs->rs_scheduled_ever);
1753         LASSERT (!rs->rs_handled);
1754         LASSERT (!rs->rs_on_net);
1755         LASSERT (rs->rs_export == NULL);
1756         LASSERT (list_empty(&rs->rs_obd_list));
1757         LASSERT (list_empty(&rs->rs_exp_list));
1758
1759         exp = class_export_get (req->rq_export);
1760         obd = exp->exp_obd;
1761
1762         /* disable reply scheduling onto srv_reply_queue while I'm setting up */
1763         rs->rs_scheduled = 1;
1764         rs->rs_on_net    = 1;
1765         rs->rs_xid       = req->rq_xid;
1766         rs->rs_transno   = req->rq_transno;
1767         rs->rs_export    = exp;
1768
1769         spin_lock_irqsave (&obd->obd_uncommitted_replies_lock, flags);
1770
1771         if (rs->rs_transno > obd->obd_last_committed) {
1772                 /* not committed already */
1773                 list_add_tail (&rs->rs_obd_list,
1774                                &obd->obd_uncommitted_replies);
1775         }
1776
1777         spin_unlock (&obd->obd_uncommitted_replies_lock);
1778         spin_lock (&exp->exp_lock);
1779
1780         list_add_tail (&rs->rs_exp_list, &exp->exp_outstanding_replies);
1781
1782         spin_unlock_irqrestore (&exp->exp_lock, flags);
1783
1784         netrc = target_send_reply_msg (req, rc, fail_id);
1785
1786         spin_lock_irqsave (&svc->srv_lock, flags);
1787
1788         svc->srv_n_difficult_replies++;
1789
1790         if (netrc != 0) /* error sending: reply is off the net */
1791                 rs->rs_on_net = 0;
1792
1793         if (!rs->rs_on_net ||                   /* some notifier */
1794             list_empty(&rs->rs_exp_list) ||     /* completed already */
1795             list_empty(&rs->rs_obd_list)) {
1796                 list_add_tail (&rs->rs_list, &svc->srv_reply_queue);
1797                 wake_up (&svc->srv_waitq);
1798         } else {
1799                 list_add (&rs->rs_list, &sni->sni_active_replies);
1800                 rs->rs_scheduled = 0;           /* allow notifier to schedule */
1801         }
1802
1803         spin_unlock_irqrestore (&svc->srv_lock, flags);
1804 }
1805
1806 int target_handle_ping(struct ptlrpc_request *req)
1807 {
1808         return lustre_pack_reply(req, 0, NULL, NULL);
1809 }