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 /* @priority: if non-zero, move the selected to the list head
  39  * @nocreate: if non-zero, only search in existed connections
  40  */
  41 static int import_set_conn(struct obd_import *imp, struct obd_uuid *uuid,
  42                            int priority, int nocreate)
  43 {
  44         struct ptlrpc_connection *ptlrpc_conn;
  45         struct obd_import_conn *imp_conn = NULL, *item;
  46         int rc = 0;
  47         ENTRY;
  48
  49         LASSERT(!(nocreate && !priority));
  50
  51         ptlrpc_conn = ptlrpc_uuid_to_connection(uuid);
  52         if (!ptlrpc_conn) {
  53                 CERROR("can't find connection %s\n", uuid->uuid);
  54                 RETURN (-EINVAL);
  55         }
  56
  57         if (!nocreate) {
  58                 OBD_ALLOC(imp_conn, sizeof(*imp_conn));
  59                 if (!imp_conn) {
  60                         CERROR("fail to alloc memory\n");
  61                         GOTO(out_put, rc = -ENOMEM);
  62                 }
  63         }
  64
  65         spin_lock(&imp->imp_lock);
  66         list_for_each_entry(item, &imp->imp_conn_list, oic_item) {
  67                 if (obd_uuid_equals(uuid, &item->oic_uuid)) {
  68                         if (priority) {
  69                                 list_del(&item->oic_item);
  70                                 list_add(&item->oic_item, &imp->imp_conn_list);
  71                                 item->oic_last_attempt = 0;
  72                         }
  73                         CDEBUG(D_HA, "imp %p@%s: find existed conn %s%s\n",
  74                                imp, imp->imp_obd->obd_name, uuid->uuid,
  75                                (priority ? ", move to head." : ""));
  76                         spin_unlock(&imp->imp_lock);
  77                         GOTO(out_free, rc = 0);
  78                 }
  79         }
  80         /* not found */
  81         if (!nocreate) {
  82                 imp_conn->oic_conn = ptlrpc_conn;
  83                 imp_conn->oic_uuid = *uuid;
  84                 imp_conn->oic_last_attempt = 0;
  85                 if (priority)
  86                         list_add(&imp_conn->oic_item, &imp->imp_conn_list);
  87                 else
  88                         list_add_tail(&imp_conn->oic_item, &imp->imp_conn_list);
  89                 CDEBUG(D_HA, "imp %p@%s: add connection %s at %s\n",
  90                        imp, imp->imp_obd->obd_name, uuid->uuid,
  91                        (priority ? "head" : "tail"));
  92         } else
  93                 rc = -ENOENT;
  94
  95         spin_unlock(&imp->imp_lock);
  96         RETURN(0);
  97 out_free:
  98         if (imp_conn)
  99                 OBD_FREE(imp_conn, sizeof(*imp_conn));
 100 out_put:
 101         ptlrpc_put_connection(ptlrpc_conn);
 102         RETURN(rc);
 103 }
 104
 105 int import_set_conn_priority(struct obd_import *imp, struct obd_uuid *uuid)
 106 {
 107         return import_set_conn(imp, uuid, 1, 1);
 108 }
 109
 110 int client_import_add_conn(struct obd_import *imp, struct obd_uuid *uuid,
 111                            int priority)
 112 {
 113         return import_set_conn(imp, uuid, priority, 0);
 114 }
 115
 116 int client_import_del_conn(struct obd_import *imp, struct obd_uuid *uuid)
 117 {
 118         struct obd_import_conn *imp_conn;
 119         struct obd_export *dlmexp;
 120         int rc = -ENOENT;
 121         ENTRY;
 122
 123         spin_lock(&imp->imp_lock);
 124         if (list_empty(&imp->imp_conn_list)) {
 125                 LASSERT(!imp->imp_conn_current);
 126                 LASSERT(!imp->imp_connection);
 127                 GOTO(out, rc);
 128         }
 129
 130         list_for_each_entry(imp_conn, &imp->imp_conn_list, oic_item) {
 131                 if (!obd_uuid_equals(uuid, &imp_conn->oic_uuid))
 132                         continue;
 133                 LASSERT(imp_conn->oic_conn);
 134
 135                 /* is current conn? */
 136                 if (imp_conn == imp->imp_conn_current) {
 137                         LASSERT(imp_conn->oic_conn == imp->imp_connection);
 138
 139                         if (imp->imp_state != LUSTRE_IMP_CLOSED &&
 140                             imp->imp_state != LUSTRE_IMP_DISCON) {
 141                                 CERROR("can't remove current connection\n");
 142                                 GOTO(out, rc = -EBUSY);
 143                         }
 144
 145                         ptlrpc_put_connection(imp->imp_connection);
 146                         imp->imp_connection = NULL;
 147
 148                         dlmexp = class_conn2export(&imp->imp_dlm_handle);
 149                         if (dlmexp && dlmexp->exp_connection) {
 150                                 LASSERT(dlmexp->exp_connection ==
 151                                         imp_conn->oic_conn);
 152                                 ptlrpc_put_connection(dlmexp->exp_connection);
 153                                 dlmexp->exp_connection = NULL;
 154                         }
 155                 }
 156
 157                 list_del(&imp_conn->oic_item);
 158                 ptlrpc_put_connection(imp_conn->oic_conn);
 159                 OBD_FREE(imp_conn, sizeof(*imp_conn));
 160                 CDEBUG(D_HA, "imp %p@%s: remove connection %s\n",
 161                        imp, imp->imp_obd->obd_name, uuid->uuid);
 162                 rc = 0;
 163                 break;
 164         }
 165 out:
 166         spin_unlock(&imp->imp_lock);
 167         if (rc == -ENOENT)
 168                 CERROR("connection %s not found\n", uuid->uuid);
 169         RETURN(rc);
 170 }
 171
 172 int client_obd_setup(struct obd_device *obddev, obd_count len, void *buf)
 173 {
 174         struct lustre_cfg* lcfg = buf;
 175         struct client_obd *cli = &obddev->u.cli;
 176         struct obd_import *imp;
 177         struct obd_uuid server_uuid;
 178         int rq_portal, rp_portal, connect_op;
 179         char *name = obddev->obd_type->typ_name;
 180         char *mgmt_name = NULL;
 181         int rc;
 182         struct obd_device *mgmt_obd;
 183         mgmtcli_register_for_events_t register_f;
 184         ENTRY;
 185
 186         /* In a more perfect world, we would hang a ptlrpc_client off of
 187          * obd_type and just use the values from there. */
 188         if (!strcmp(name, LUSTRE_OSC_NAME)) {
 189                 rq_portal = OST_REQUEST_PORTAL;
 190                 rp_portal = OSC_REPLY_PORTAL;
 191                 connect_op = OST_CONNECT;
 192         } else if (!strcmp(name, LUSTRE_MDC_NAME)) {
 193                 rq_portal = MDS_REQUEST_PORTAL;
 194                 rp_portal = MDC_REPLY_PORTAL;
 195                 connect_op = MDS_CONNECT;
 196         } else if (!strcmp(name, LUSTRE_MGMTCLI_NAME)) {
 197                 rq_portal = MGMT_REQUEST_PORTAL;
 198                 rp_portal = MGMT_REPLY_PORTAL;
 199                 connect_op = MGMT_CONNECT;
 200         } else {
 201                 CERROR("unknown client OBD type \"%s\", can't setup\n",
 202                        name);
 203                 RETURN(-EINVAL);
 204         }
 205
 206         if (lcfg->lcfg_inllen1 < 1) {
 207                 CERROR("requires a TARGET UUID\n");
 208                 RETURN(-EINVAL);
 209         }
 210
 211         if (lcfg->lcfg_inllen1 > 37) {
 212                 CERROR("client UUID must be less than 38 characters\n");
 213                 RETURN(-EINVAL);
 214         }
 215
 216         if (lcfg->lcfg_inllen2 < 1) {
 217                 CERROR("setup requires a SERVER UUID\n");
 218                 RETURN(-EINVAL);
 219         }
 220
 221         if (lcfg->lcfg_inllen2 > 37) {
 222                 CERROR("target UUID must be less than 38 characters\n");
 223                 RETURN(-EINVAL);
 224         }
 225
 226         sema_init(&cli->cl_sem, 1);
 227         cli->cl_conn_count = 0;
 228         memcpy(server_uuid.uuid, lcfg->lcfg_inlbuf2,
 229                min_t(unsigned int, lcfg->lcfg_inllen2, sizeof(server_uuid)));
 230
 231         cli->cl_dirty = 0;
 232         cli->cl_avail_grant = 0;
 233         /* FIXME: should limit this for the sum of all cl_dirty_max */
 234         cli->cl_dirty_max = OSC_MAX_DIRTY_DEFAULT * 1024 * 1024;
 235         if (cli->cl_dirty_max >> PAGE_SHIFT > num_physpages / 8)
 236                 cli->cl_dirty_max = num_physpages << (PAGE_SHIFT - 3);
 237         INIT_LIST_HEAD(&cli->cl_cache_waiters);
 238         INIT_LIST_HEAD(&cli->cl_loi_ready_list);
 239         INIT_LIST_HEAD(&cli->cl_loi_write_list);
 240         INIT_LIST_HEAD(&cli->cl_loi_read_list);
 241         spin_lock_init(&cli->cl_loi_list_lock);
 242         cli->cl_r_in_flight = 0;
 243         cli->cl_w_in_flight = 0;
 244         spin_lock_init(&cli->cl_read_rpc_hist.oh_lock);
 245         spin_lock_init(&cli->cl_write_rpc_hist.oh_lock);
 246         spin_lock_init(&cli->cl_read_page_hist.oh_lock);
 247         spin_lock_init(&cli->cl_write_page_hist.oh_lock);
 248
 249         if (num_physpages <= 32768) { /* <= 128 MB */
 250                 cli->cl_max_pages_per_rpc = PTLRPC_MAX_BRW_PAGES / 3;
 251                 cli->cl_max_rpcs_in_flight = OSC_MAX_RIF_DEFAULT / 3;
 252         } else {
 253                 cli->cl_max_pages_per_rpc = PTLRPC_MAX_BRW_PAGES;
 254                 cli->cl_max_rpcs_in_flight = OSC_MAX_RIF_DEFAULT;
 255         }
 256
 257
 258         rc = ldlm_get_ref();
 259         if (rc) {
 260                 CERROR("ldlm_get_ref failed: %d\n", rc);
 261                 GOTO(err, rc);
 262         }
 263
 264         ptlrpc_init_client(rq_portal, rp_portal, name,
 265                            &obddev->obd_ldlm_client);
 266
 267         imp = class_new_import();
 268         if (imp == NULL)
 269                 GOTO(err_ldlm, rc = -ENOENT);
 270         imp->imp_client = &obddev->obd_ldlm_client;
 271         imp->imp_obd = obddev;
 272         imp->imp_connect_op = connect_op;
 273         imp->imp_generation = 0;
 274         imp->imp_initial_recov = 1;
 275         INIT_LIST_HEAD(&imp->imp_pinger_chain);
 276         memcpy(imp->imp_target_uuid.uuid, lcfg->lcfg_inlbuf1,
 277               lcfg->lcfg_inllen1);
 278         class_import_put(imp);
 279
 280         rc = client_import_add_conn(imp, &server_uuid, 1);
 281         if (rc) {
 282                 CERROR("can't add initial connection\n");
 283                 GOTO(err_import, rc);
 284         }
 285
 286         cli->cl_import = imp;
 287         cli->cl_max_mds_easize = sizeof(struct lov_mds_md);
 288         cli->cl_max_mds_cookiesize = sizeof(struct llog_cookie);
 289         cli->cl_sandev = to_kdev_t(0);
 290
 291         if (lcfg->lcfg_inllen3 != 0) {
 292                 if (!strcmp(lcfg->lcfg_inlbuf3, "inactive")) {
 293                         CDEBUG(D_HA, "marking %s %s->%s as inactive\n",
 294                                name, obddev->obd_name,
 295                                imp->imp_target_uuid.uuid);
 296                         imp->imp_invalid = 1;
 297
 298                         if (lcfg->lcfg_inllen4 != 0)
 299                                 mgmt_name = lcfg->lcfg_inlbuf4;
 300                 } else {
 301                         mgmt_name = lcfg->lcfg_inlbuf3;
 302                 }
 303         }
 304
 305         if (mgmt_name != NULL) {
 306                 /* Register with management client if we need to. */
 307                 CDEBUG(D_HA, "%s registering with %s for events about %s\n",
 308                        obddev->obd_name, mgmt_name, server_uuid.uuid);
 309
 310                 mgmt_obd = class_name2obd(mgmt_name);
 311                 if (!mgmt_obd) {
 312                         CERROR("can't find mgmtcli %s to register\n",
 313                                mgmt_name);
 314                         GOTO(err_import, rc = -ENOSYS);
 315                 }
 316
 317                 register_f = inter_module_get("mgmtcli_register_for_events");
 318                 if (!register_f) {
 319                         CERROR("can't i_m_g mgmtcli_register_for_events\n");
 320                         GOTO(err_import, rc = -ENOSYS);
 321                 }
 322
 323                 rc = register_f(mgmt_obd, obddev, &imp->imp_target_uuid);
 324                 inter_module_put("mgmtcli_register_for_events");
 325
 326                 if (!rc)
 327                         cli->cl_mgmtcli_obd = mgmt_obd;
 328         }
 329
 330         RETURN(rc);
 331
 332 err_import:
 333         class_destroy_import(imp);
 334 err_ldlm:
 335         ldlm_put_ref(0);
 336 err:
 337         RETURN(rc);
 338
 339 }
 340
 341 int client_obd_cleanup(struct obd_device *obddev, int flags)
 342 {
 343         struct client_obd *cli = &obddev->u.cli;
 344
 345         if (!cli->cl_import)
 346                 RETURN(-EINVAL);
 347         if (cli->cl_mgmtcli_obd) {
 348                 mgmtcli_deregister_for_events_t dereg_f;
 349
 350                 dereg_f = inter_module_get("mgmtcli_deregister_for_events");
 351                 dereg_f(cli->cl_mgmtcli_obd, obddev);
 352                 inter_module_put("mgmtcli_deregister_for_events");
 353         }
 354         class_destroy_import(cli->cl_import);
 355         cli->cl_import = NULL;
 356
 357         ldlm_put_ref(flags & OBD_OPT_FORCE);
 358
 359         RETURN(0);
 360 }
 361
 362 int client_connect_import(struct lustre_handle *dlm_handle,
 363                           struct obd_device *obd,
 364                           struct obd_uuid *cluuid,
 365                           unsigned long connect_flags)
 366 {
 367         struct client_obd *cli = &obd->u.cli;
 368         struct obd_import *imp = cli->cl_import;
 369         struct obd_export *exp;
 370         int rc;
 371         ENTRY;
 372
 373         down(&cli->cl_sem);
 374         rc = class_connect(dlm_handle, obd, cluuid);
 375         if (rc)
 376                 GOTO(out_sem, rc);
 377
 378         cli->cl_conn_count++;
 379         if (cli->cl_conn_count > 1)
 380                 GOTO(out_sem, rc);
 381         exp = class_conn2export(dlm_handle);
 382
 383         if (obd->obd_namespace != NULL)
 384                 CERROR("already have namespace!\n");
 385         obd->obd_namespace = ldlm_namespace_new(obd->obd_name,
 386                                                 LDLM_NAMESPACE_CLIENT);
 387         if (obd->obd_namespace == NULL)
 388                 GOTO(out_disco, rc = -ENOMEM);
 389
 390         imp->imp_dlm_handle = *dlm_handle;
 391         rc = ptlrpc_init_import(imp);
 392         if (rc != 0)
 393                 GOTO(out_ldlm, rc);
 394
 395         imp->imp_connect_flags = connect_flags;
 396         rc = ptlrpc_connect_import(imp, NULL);
 397         if (rc != 0) {
 398                 LASSERT (imp->imp_state == LUSTRE_IMP_DISCON);
 399                 GOTO(out_ldlm, rc);
 400         }
 401         LASSERT(exp->exp_connection);
 402         ptlrpc_pinger_add_import(imp);
 403         EXIT;
 404
 405         if (rc) {
 406 out_ldlm:
 407                 ldlm_namespace_free(obd->obd_namespace, 0);
 408                 obd->obd_namespace = NULL;
 409 out_disco:
 410                 cli->cl_conn_count--;
 411                 class_disconnect(exp, 0);
 412         } else {
 413                 class_export_put(exp);
 414         }
 415 out_sem:
 416         up(&cli->cl_sem);
 417         return rc;
 418 }
 419
 420 int client_disconnect_export(struct obd_export *exp, unsigned long flags)
 421 {
 422         struct obd_device *obd = class_exp2obd(exp);
 423         struct client_obd *cli = &obd->u.cli;
 424         struct obd_import *imp = cli->cl_import;
 425         int rc = 0, err;
 426         ENTRY;
 427
 428         if (!obd) {
 429                 CERROR("invalid export for disconnect: exp %p cookie "LPX64"\n",
 430                        exp, exp ? exp->exp_handle.h_cookie : -1);
 431                 RETURN(-EINVAL);
 432         }
 433
 434         down(&cli->cl_sem);
 435         if (!cli->cl_conn_count) {
 436                 CERROR("disconnecting disconnected device (%s)\n",
 437                        obd->obd_name);
 438                 GOTO(out_sem, rc = -EINVAL);
 439         }
 440
 441         cli->cl_conn_count--;
 442         if (cli->cl_conn_count)
 443                 GOTO(out_no_disconnect, rc = 0);
 444
 445         /* Some non-replayable imports (MDS's OSCs) are pinged, so just
 446          * delete it regardless.  (It's safe to delete an import that was
 447          * never added.) */
 448         (void)ptlrpc_pinger_del_import(imp);
 449
 450         if (obd->obd_namespace != NULL) {
 451                 /* obd_no_recov == local only */
 452                 ldlm_cli_cancel_unused(obd->obd_namespace, NULL,
 453                                        obd->obd_no_recov, NULL);
 454                 ldlm_namespace_free(obd->obd_namespace, obd->obd_no_recov);
 455                 obd->obd_namespace = NULL;
 456         }
 457
 458         /*
 459          * Yeah, obd_no_recov also (mainly) means "forced shutdown".
 460          */
 461         if (obd->obd_no_recov)
 462                 ptlrpc_invalidate_import(imp, 0);
 463         else
 464                 rc = ptlrpc_disconnect_import(imp);
 465
 466         EXIT;
 467  out_no_disconnect:
 468         err = class_disconnect(exp, 0);
 469         if (!rc && err)
 470                 rc = err;
 471  out_sem:
 472         up(&cli->cl_sem);
 473         RETURN(rc);
 474 }
 475
 476 /* --------------------------------------------------------------------------
 477  * from old lib/target.c
 478  * -------------------------------------------------------------------------- */
 479
 480 int target_handle_reconnect(struct lustre_handle *conn, struct obd_export *exp,
 481                             struct obd_uuid *cluuid, int initial_conn)
 482 {
 483         if (exp->exp_connection && !initial_conn) {
 484                 struct lustre_handle *hdl;
 485                 hdl = &exp->exp_imp_reverse->imp_remote_handle;
 486                 /* Might be a re-connect after a partition. */
 487                 if (!memcmp(&conn->cookie, &hdl->cookie, sizeof conn->cookie)) {
 488                         CERROR("%s reconnecting\n", cluuid->uuid);
 489                         conn->cookie = exp->exp_handle.h_cookie;
 490                         RETURN(EALREADY);
 491                 } else {
 492                         CERROR("%s reconnecting from %s, "
 493                                "handle mismatch (ours "LPX64", theirs "
 494                                LPX64")\n", cluuid->uuid,
 495                                exp->exp_connection->c_remote_uuid.uuid,
 496                                hdl->cookie, conn->cookie);
 497                         memset(conn, 0, sizeof *conn);
 498                         RETURN(-EALREADY);
 499                 }
 500         }
 501
 502         conn->cookie = exp->exp_handle.h_cookie;
 503         CDEBUG(D_INFO, "existing export for UUID '%s' at %p\n",
 504                cluuid->uuid, exp);
 505         CDEBUG(D_IOCTL,"connect: cookie "LPX64"\n", conn->cookie);
 506         RETURN(0);
 507 }
 508
 509 static char nidstr[PTL_NALFMT_SIZE];
 510 int target_handle_connect(struct ptlrpc_request *req)
 511 {
 512         unsigned long connect_flags = 0, *cfp;
 513         struct obd_device *target;
 514         struct obd_export *export = NULL;
 515         struct obd_import *revimp;
 516         struct lustre_handle conn;
 517         struct obd_uuid tgtuuid;
 518         struct obd_uuid cluuid;
 519         struct obd_uuid remote_uuid;
 520         struct list_head *p;
 521         char *str, *tmp;
 522         int rc = 0;
 523         unsigned long flags;
 524         int initial_conn = 0;
 525         char peer_str[PTL_NALFMT_SIZE];
 526         ENTRY;
 527
 528         OBD_RACE(OBD_FAIL_TGT_CONN_RACE);
 529
 530         LASSERT_REQSWAB (req, 0);
 531         str = lustre_msg_string(req->rq_reqmsg, 0, sizeof(tgtuuid) - 1);
 532         if (str == NULL) {
 533                 CERROR("bad target UUID for connect\n");
 534                 GOTO(out, rc = -EINVAL);
 535         }
 536
 537         obd_str2uuid (&tgtuuid, str);
 538         target = class_uuid2obd(&tgtuuid);
 539         if (!target) {
 540                 target = class_name2obd(str);
 541         }
 542
 543         if (!target || target->obd_stopping || !target->obd_set_up) {
 544                 CERROR("UUID '%s' is not available for connect from NID %s\n",
 545                        str, ptlrpc_peernid2str(&req->rq_peer, nidstr));
 546                 GOTO(out, rc = -ENODEV);
 547         }
 548
 549         LASSERT_REQSWAB (req, 1);
 550         str = lustre_msg_string(req->rq_reqmsg, 1, sizeof(cluuid) - 1);
 551         if (str == NULL) {
 552                 CERROR("bad client UUID for connect\n");
 553                 GOTO(out, rc = -EINVAL);
 554         }
 555
 556         obd_str2uuid (&cluuid, str);
 557
 558         /* XXX extract a nettype and format accordingly */
 559         switch (sizeof(ptl_nid_t)) {
 560                 /* NB the casts only avoid compiler warnings */
 561         case 8:
 562                 snprintf(remote_uuid.uuid, sizeof remote_uuid,
 563                          "NET_"LPX64"_UUID", (__u64)req->rq_peer.peer_id.nid);
 564                 break;
 565         case 4:
 566                 snprintf(remote_uuid.uuid, sizeof remote_uuid,
 567                          "NET_%x_UUID", (__u32)req->rq_peer.peer_id.nid);
 568                 break;
 569         default:
 570                 LBUG();
 571         }
 572
 573         tmp = lustre_msg_buf(req->rq_reqmsg, 2, sizeof conn);
 574         if (tmp == NULL)
 575                 GOTO(out, rc = -EPROTO);
 576
 577         memcpy(&conn, tmp, sizeof conn);
 578
 579         cfp = lustre_msg_buf(req->rq_reqmsg, 3, sizeof(unsigned long));
 580         LASSERT(cfp != NULL);
 581         connect_flags = *cfp;
 582
 583         rc = lustre_pack_reply(req, 0, NULL, NULL);
 584         if (rc)
 585                 GOTO(out, rc);
 586
 587         if (lustre_msg_get_op_flags(req->rq_reqmsg) & MSG_CONNECT_INITIAL)
 588                 initial_conn = 1;
 589
 590         /* lctl gets a backstage, all-access pass. */
 591         if (obd_uuid_equals(&cluuid, &target->obd_uuid))
 592                 goto dont_check_exports;
 593
 594         spin_lock(&target->obd_dev_lock);
 595         list_for_each(p, &target->obd_exports) {
 596                 export = list_entry(p, struct obd_export, exp_obd_chain);
 597                 if (obd_uuid_equals(&cluuid, &export->exp_client_uuid)) {
 598                         spin_unlock(&target->obd_dev_lock);
 599                         LASSERT(export->exp_obd == target);
 600
 601                         rc = target_handle_reconnect(&conn, export, &cluuid,
 602                                                      initial_conn);
 603                         break;
 604                 }
 605                 export = NULL;
 606         }
 607         /* If we found an export, we already unlocked. */
 608         if (!export) {
 609                 spin_unlock(&target->obd_dev_lock);
 610         } else if (req->rq_export == NULL &&
 611                    atomic_read(&export->exp_rpc_count) > 0) {
 612                 CWARN("%s: refuse connection from %s/%s to 0x%p/%d\n",
 613                       target->obd_name, cluuid.uuid,
 614                       ptlrpc_peernid2str(&req->rq_peer, peer_str),
 615                       export, atomic_read(&export->exp_refcount));
 616                 GOTO(out, rc = -EBUSY);
 617         } else if (req->rq_export != NULL &&
 618                    atomic_read(&export->exp_rpc_count) > 1) {
 619                 CWARN("%s: refuse reconnection from %s@%s to 0x%p/%d\n",
 620                       target->obd_name, cluuid.uuid,
 621                       ptlrpc_peernid2str(&req->rq_peer, peer_str),
 622                       export, atomic_read(&export->exp_rpc_count));
 623                 GOTO(out, rc = -EBUSY);
 624         }
 625         else if (req->rq_reqmsg->conn_cnt == 1 && !initial_conn) {
 626                 CERROR("%s reconnected with 1 conn_cnt; cookies not random?\n",
 627                        cluuid.uuid);
 628                 GOTO(out, rc = -EALREADY);
 629         }
 630
 631         /* Tell the client if we're in recovery. */
 632         /* If this is the first client, start the recovery timer */
 633         CWARN("%s: connection from %s@%s/%lu %s\n", target->obd_name, cluuid.uuid,
 634               ptlrpc_peernid2str(&req->rq_peer, peer_str), *cfp,
 635               target->obd_recovering ? "(recovering)" : "");
 636         if (target->obd_recovering) {
 637                 lustre_msg_add_op_flags(req->rq_repmsg, MSG_CONNECT_RECOVERING);
 638                 target_start_recovery_timer(target);
 639         }
 640 #if 0
 641         /* Tell the client if we support replayable requests */
 642         if (target->obd_replayable)
 643                 lustre_msg_add_op_flags(req->rq_repmsg, MSG_CONNECT_REPLAYABLE);
 644 #endif
 645         if (export == NULL) {
 646                 if (target->obd_recovering) {
 647                         CERROR("%s denying connection for new client %s@%s: "
 648                                "%d clients in recovery for %lds\n", target->obd_name,
 649                                cluuid.uuid,
 650                                ptlrpc_peernid2str(&req->rq_peer, peer_str),
 651                                target->obd_recoverable_clients,
 652                                (target->obd_recovery_timer.expires-jiffies)/HZ);
 653                         rc = -EBUSY;
 654                 } else {
 655  dont_check_exports:
 656                         rc = obd_connect(&conn, target, &cluuid, connect_flags);
 657                 }
 658         }
 659         /* Tell the client if we support replayable requests */
 660         if (target->obd_replayable)
 661                 lustre_msg_add_op_flags(req->rq_repmsg, MSG_CONNECT_REPLAYABLE);
 662
 663         /* If all else goes well, this is our RPC return code. */
 664         req->rq_status = 0;
 665
 666         if (rc && rc != EALREADY)
 667                 GOTO(out, rc);
 668
 669         req->rq_repmsg->handle = conn;
 670
 671         /* If the client and the server are the same node, we will already
 672          * have an export that really points to the client's DLM export,
 673          * because we have a shared handles table.
 674          *
 675          * XXX this will go away when shaver stops sending the "connect" handle
 676          * in the real "remote handle" field of the request --phik 24 Apr 2003
 677          */
 678         if (req->rq_export != NULL)
 679                 class_export_put(req->rq_export);
 680
 681         /* ownership of this export ref transfers to the request */
 682         export = req->rq_export = class_conn2export(&conn);
 683         LASSERT(export != NULL);
 684
 685         spin_lock_irqsave(&export->exp_lock, flags);
 686         if (initial_conn) {
 687                 req->rq_repmsg->conn_cnt = export->exp_conn_cnt + 1;
 688         } else if (export->exp_conn_cnt >= req->rq_reqmsg->conn_cnt) {
 689                 CERROR("%s@%s: already connected at a higher conn_cnt: %d > %d\n",
 690                        cluuid.uuid, ptlrpc_peernid2str(&req->rq_peer, peer_str),
 691                        export->exp_conn_cnt,
 692                        req->rq_reqmsg->conn_cnt);
 693                 spin_unlock_irqrestore(&export->exp_lock, flags);
 694                 GOTO(out, rc = -EALREADY);
 695         }
 696         export->exp_conn_cnt = req->rq_reqmsg->conn_cnt;
 697         spin_unlock_irqrestore(&export->exp_lock, flags);
 698
 699         /* request from liblustre? */
 700         if (lustre_msg_get_op_flags(req->rq_reqmsg) & MSG_CONNECT_LIBCLIENT)
 701                 export->exp_libclient = 1;
 702
 703         if (export->exp_connection != NULL)
 704                 ptlrpc_put_connection(export->exp_connection);
 705         export->exp_connection = ptlrpc_get_connection(&req->rq_peer,
 706                                                        &remote_uuid);
 707
 708         if (rc == EALREADY) {
 709                 /* We indicate the reconnection in a flag, not an error code. */
 710                 lustre_msg_add_op_flags(req->rq_repmsg, MSG_CONNECT_RECONNECT);
 711                 GOTO(out, rc = 0);
 712         }
 713
 714         if (target->obd_recovering) {
 715                 target->obd_connected_clients++;
 716         }
 717
 718         memcpy(&conn, lustre_msg_buf(req->rq_reqmsg, 2, sizeof conn),
 719                sizeof conn);
 720
 721         if (export->exp_imp_reverse != NULL)
 722                 class_destroy_import(export->exp_imp_reverse);
 723         revimp = export->exp_imp_reverse = class_new_import();
 724         revimp->imp_connection = ptlrpc_connection_addref(export->exp_connection);
 725         revimp->imp_client = &export->exp_obd->obd_ldlm_client;
 726         revimp->imp_remote_handle = conn;
 727         revimp->imp_obd = target;
 728         revimp->imp_dlm_fake = 1;
 729         revimp->imp_state = LUSTRE_IMP_FULL;
 730         class_import_put(revimp);
 731
 732         rc = obd_connect_post(export, connect_flags);
 733 out:
 734         if (rc)
 735                 req->rq_status = rc;
 736         RETURN(rc);
 737 }
 738
 739 int target_handle_disconnect(struct ptlrpc_request *req)
 740 {
 741         struct obd_export *exp;
 742         int rc;
 743         ENTRY;
 744
 745         rc = lustre_pack_reply(req, 0, NULL, NULL);
 746         if (rc)
 747                 RETURN(rc);
 748
 749         /* keep the rq_export around so we can send the reply */
 750         exp = class_export_get(req->rq_export);
 751         req->rq_status = obd_disconnect(exp, 0);
 752         RETURN(0);
 753 }
 754
 755 void target_destroy_export(struct obd_export *exp)
 756 {
 757         /* exports created from last_rcvd data, and "fake"
 758            exports created by lctl don't have an import */
 759         if (exp->exp_imp_reverse != NULL)
 760                 class_destroy_import(exp->exp_imp_reverse);
 761
 762         /* We cancel locks at disconnect time, but this will catch any locks
 763          * granted in a race with recovery-induced disconnect. */
 764         if (exp->exp_obd->obd_namespace != NULL)
 765                 ldlm_cancel_locks_for_export(exp);
 766 }
 767
 768 /*
 769  * Recovery functions
 770  */
 771
 772 struct ptlrpc_request *
 773 ptlrpc_clone_req( struct ptlrpc_request *orig_req)
 774 {
 775         struct ptlrpc_request *copy_req;
 776         struct lustre_msg *copy_reqmsg;
 777
 778         OBD_ALLOC(copy_req, sizeof *copy_req);
 779         if (!copy_req)
 780                 return NULL;
 781         OBD_ALLOC(copy_reqmsg, orig_req->rq_reqlen);
 782         if (!copy_reqmsg){
 783                 OBD_FREE(copy_req, sizeof *copy_req);
 784                 return NULL;
 785         }
 786
 787         memcpy(copy_req, orig_req, sizeof *copy_req);
 788         memcpy(copy_reqmsg, orig_req->rq_reqmsg, orig_req->rq_reqlen);
 789         /* the copied req takes over the reply state */
 790         orig_req->rq_reply_state = NULL;
 791
 792         copy_req->rq_reqmsg = copy_reqmsg;
 793         class_export_get(copy_req->rq_export);
 794         INIT_LIST_HEAD(&copy_req->rq_list);
 795
 796         return copy_req;
 797 }
 798 void ptlrpc_free_clone( struct ptlrpc_request *req)
 799 {
 800         class_export_put(req->rq_export);
 801         list_del(&req->rq_list);
 802         OBD_FREE(req->rq_reqmsg, req->rq_reqlen);
 803         OBD_FREE(req, sizeof *req);
 804 }
 805
 806
 807
 808 static void target_release_saved_req(struct ptlrpc_request *req)
 809 {
 810         class_export_put(req->rq_export);
 811         OBD_FREE(req->rq_reqmsg, req->rq_reqlen);
 812         OBD_FREE(req, sizeof *req);
 813 }
 814
 815 #ifdef __KERNEL__
 816 static void target_finish_recovery(struct obd_device *obd)
 817 {
 818         struct list_head *tmp, *n;
 819         int rc;
 820
 821         CWARN("%s: sending delayed replies to recovered clients\n",
 822               obd->obd_name);
 823
 824         ldlm_reprocess_all_ns(obd->obd_namespace);
 825
 826         /* when recovery finished, cleanup orphans on mds and ost */
 827         if (OBT(obd) && OBP(obd, postrecov)) {
 828                 rc = OBP(obd, postrecov)(obd);
 829                 if (rc >= 0)
 830                         CWARN("%s: all clients recovered, %d MDS "
 831                               "orphans deleted\n", obd->obd_name, rc);
 832                 else
 833                         CERROR("postrecov failed %d\n", rc);
 834         }
 835
 836
 837         list_for_each_safe(tmp, n, &obd->obd_delayed_reply_queue) {
 838                 struct ptlrpc_request *req;
 839                 req = list_entry(tmp, struct ptlrpc_request, rq_list);
 840                 list_del(&req->rq_list);
 841                 DEBUG_REQ(D_ERROR, req, "delayed:");
 842                 ptlrpc_reply(req);
 843                 target_release_saved_req(req);
 844         }
 845         obd->obd_recovery_end = LTIME_S(CURRENT_TIME);
 846         return;
 847 }
 848
 849 static void abort_recovery_queue(struct obd_device *obd)
 850 {
 851         struct ptlrpc_request *req;
 852         struct list_head *tmp, *n;
 853         int rc;
 854
 855         list_for_each_safe(tmp, n, &obd->obd_recovery_queue) {
 856                 req = list_entry(tmp, struct ptlrpc_request, rq_list);
 857                 list_del(&req->rq_list);
 858                 DEBUG_REQ(D_ERROR, req, "aborted:");
 859                 req->rq_status = -ENOTCONN;
 860                 req->rq_type = PTL_RPC_MSG_ERR;
 861                 rc = lustre_pack_reply(req, 0, NULL, NULL);
 862                 if (rc == 0) {
 863                         ptlrpc_reply(req);
 864                 } else {
 865                         DEBUG_REQ(D_ERROR, req,
 866                                   "packing failed for abort-reply; skipping");
 867                 }
 868                 target_release_saved_req(req);
 869         }
 870 }
 871 #endif
 872
 873 /* Called from a cleanup function if the device is being cleaned up
 874    forcefully.  The exports should all have been disconnected already,
 875    the only thing left to do is
 876      - clear the recovery flags
 877      - cancel the timer
 878      - free queued requests and replies, but don't send replies
 879    Because the obd_stopping flag is set, no new requests should be received.
 880
 881 */
 882 void target_cleanup_recovery(struct obd_device *obd)
 883 {
 884         struct list_head *tmp, *n;
 885         struct ptlrpc_request *req;
 886
 887         spin_lock_bh(&obd->obd_processing_task_lock);
 888         if (!obd->obd_recovering) {
 889                 spin_unlock_bh(&obd->obd_processing_task_lock);
 890                 EXIT;
 891                 return;
 892         }
 893         obd->obd_recovering = obd->obd_abort_recovery = 0;
 894         target_cancel_recovery_timer(obd);
 895         spin_unlock_bh(&obd->obd_processing_task_lock);
 896
 897         list_for_each_safe(tmp, n, &obd->obd_delayed_reply_queue) {
 898                 req = list_entry(tmp, struct ptlrpc_request, rq_list);
 899                 list_del(&req->rq_list);
 900                 LASSERT (req->rq_reply_state);
 901                 lustre_free_reply_state(req->rq_reply_state);
 902                 target_release_saved_req(req);
 903         }
 904         list_for_each_safe(tmp, n, &obd->obd_recovery_queue) {
 905                 req = list_entry(tmp, struct ptlrpc_request, rq_list);
 906                 list_del(&req->rq_list);
 907                 LASSERT (req->rq_reply_state == 0);
 908                 target_release_saved_req(req);
 909          }
 910 }
 911
 912 #ifdef __KERNEL__
 913 static void target_abort_recovery(void *data)
 914 {
 915         struct obd_device *obd = data;
 916
 917         LASSERT(!obd->obd_recovering);
 918
 919         class_disconnect_stale_exports(obd, 0);
 920
 921         CERROR("%s: recovery period over; disconnecting unfinished clients.\n",
 922                obd->obd_name);
 923
 924         abort_recovery_queue(obd);
 925         target_finish_recovery(obd);
 926         ptlrpc_run_recovery_over_upcall(obd);
 927 }
 928 #endif
 929
 930 static void target_recovery_expired(unsigned long castmeharder)
 931 {
 932         struct obd_device *obd = (struct obd_device *)castmeharder;
 933         CERROR("recovery timed out, aborting\n");
 934         spin_lock_bh(&obd->obd_processing_task_lock);
 935         if (obd->obd_recovering)
 936                 obd->obd_abort_recovery = 1;
 937
 938         wake_up(&obd->obd_next_transno_waitq);
 939         spin_unlock_bh(&obd->obd_processing_task_lock);
 940 }
 941
 942
 943 /* obd_processing_task_lock should be held */
 944 void target_cancel_recovery_timer(struct obd_device *obd)
 945 {
 946         CDEBUG(D_HA, "%s: cancel recovery timer\n", obd->obd_name);
 947         del_timer(&obd->obd_recovery_timer);
 948 }
 949
 950 #ifdef __KERNEL__
 951 static void reset_recovery_timer(struct obd_device *obd)
 952 {
 953         spin_lock_bh(&obd->obd_processing_task_lock);
 954         if (!obd->obd_recovering) {
 955                 spin_unlock_bh(&obd->obd_processing_task_lock);
 956                 return;
 957         }
 958         CDEBUG(D_HA, "timer will expire in %u seconds\n",
 959                OBD_RECOVERY_TIMEOUT / HZ);
 960         mod_timer(&obd->obd_recovery_timer, jiffies + OBD_RECOVERY_TIMEOUT);
 961         spin_unlock_bh(&obd->obd_processing_task_lock);
 962 }
 963 #endif
 964
 965 /* Only start it the first time called */
 966 void target_start_recovery_timer(struct obd_device *obd)
 967 {
 968         spin_lock_bh(&obd->obd_processing_task_lock);
 969         if (!obd->obd_recovering || timer_pending(&obd->obd_recovery_timer)) {
 970                 spin_unlock_bh(&obd->obd_processing_task_lock);
 971                 return;
 972         }
 973         CWARN("%s: starting recovery timer (%us)\n", obd->obd_name,
 974                OBD_RECOVERY_TIMEOUT / HZ);
 975         obd->obd_recovery_timer.function = target_recovery_expired;
 976         obd->obd_recovery_timer.data = (unsigned long)obd;
 977         mod_timer(&obd->obd_recovery_timer, jiffies + OBD_RECOVERY_TIMEOUT);
 978         spin_unlock_bh(&obd->obd_processing_task_lock);
 979 }
 980
 981 #ifdef __KERNEL__
 982 static int check_for_next_transno(struct obd_device *obd)
 983 {
 984         struct ptlrpc_request *req = NULL;
 985         int wake_up = 0, connected, completed, queue_len, max;
 986         __u64 next_transno, req_transno;
 987
 988         spin_lock_bh(&obd->obd_processing_task_lock);
 989         if (!list_empty(&obd->obd_recovery_queue)) {
 990                 req = list_entry(obd->obd_recovery_queue.next,
 991                                  struct ptlrpc_request, rq_list);
 992                 req_transno = req->rq_reqmsg->transno;
 993         } else {
 994                 req_transno = 0;
 995         }
 996
 997         max = obd->obd_max_recoverable_clients;
 998         connected = obd->obd_connected_clients;
 999         completed = max - obd->obd_recoverable_clients;
1000         queue_len = obd->obd_requests_queued_for_recovery;
1001         next_transno = obd->obd_next_recovery_transno;
1002
1003         CDEBUG(D_HA,"max: %d, connected: %d, completed: %d, queue_len: %d, "
1004                "req_transno: "LPU64", next_transno: "LPU64"\n",
1005                max, connected, completed, queue_len, req_transno, next_transno);
1006         if (obd->obd_abort_recovery) {
1007                 CDEBUG(D_HA, "waking for aborted recovery\n");
1008                 wake_up = 1;
1009         } else if (max == completed) {
1010                 CDEBUG(D_HA, "waking for completed recovery\n");
1011                 wake_up = 1;
1012         } else if (req_transno == next_transno) {
1013                 CDEBUG(D_HA, "waking for next ("LPD64")\n", next_transno);
1014                 wake_up = 1;
1015         } else if (queue_len + completed == max) {
1016                 LASSERT(req->rq_reqmsg->transno >= next_transno);
1017                 CDEBUG(D_ERROR,
1018                        "waking for skipped transno (skip: "LPD64
1019                        ", ql: %d, comp: %d, conn: %d, next: "LPD64")\n",
1020                        next_transno, queue_len, completed, max, req_transno);
1021                 obd->obd_next_recovery_transno = req_transno;
1022                 wake_up = 1;
1023         }
1024         spin_unlock_bh(&obd->obd_processing_task_lock);
1025
1026         return wake_up;
1027 }
1028
1029 static struct ptlrpc_request *
1030 target_next_replay_req(struct obd_device *obd)
1031 {
1032         struct l_wait_info lwi = { 0 };
1033         struct ptlrpc_request *req;
1034
1035         CDEBUG(D_HA, "Waiting for transno "LPD64"\n",
1036                obd->obd_next_recovery_transno);
1037         l_wait_event(obd->obd_next_transno_waitq,
1038                      check_for_next_transno(obd), &lwi);
1039
1040         spin_lock_bh(&obd->obd_processing_task_lock);
1041         if (obd->obd_abort_recovery) {
1042                 req = NULL;
1043         } else if (!list_empty(&obd->obd_recovery_queue)) {
1044                 req = list_entry(obd->obd_recovery_queue.next,
1045                                  struct ptlrpc_request, rq_list);
1046                 list_del_init(&req->rq_list);
1047                 obd->obd_requests_queued_for_recovery--;
1048         } else {
1049                 req = NULL;
1050         }
1051         spin_unlock_bh(&obd->obd_processing_task_lock);
1052         return req;
1053 }
1054
1055 static int target_recovery_thread(void *arg)
1056 {
1057         struct obd_device *obd = arg;
1058         struct ptlrpc_request *req;
1059         struct target_recovery_data *trd = &obd->obd_recovery_data;
1060         unsigned long flags;
1061         ENTRY;
1062
1063         kportal_daemonize("tgt-recov");
1064
1065         SIGNAL_MASK_LOCK(current, flags);
1066         sigfillset(&current->blocked);
1067         RECALC_SIGPENDING;
1068         SIGNAL_MASK_UNLOCK(current, flags);
1069
1070         CERROR("%s: started recovery thread pid %d\n", obd->obd_name,
1071                current->pid);
1072         trd->trd_processing_task = current->pid;
1073
1074         obd->obd_recovering = 1;
1075         complete(&trd->trd_starting);
1076
1077         while (obd->obd_recovering) {
1078                 LASSERT(trd->trd_processing_task == current->pid);
1079                 req = target_next_replay_req(obd);
1080                 if (req != NULL) {
1081                         char peer_str[PTL_NALFMT_SIZE];
1082                         DEBUG_REQ(D_HA, req, "processing t"LPD64" from %s: ",
1083                                   req->rq_reqmsg->transno,
1084                                   ptlrpc_peernid2str(&req->rq_peer, peer_str));
1085                         (void)trd->trd_recovery_handler(req);
1086                         obd->obd_replayed_requests++;
1087                         reset_recovery_timer(obd);
1088                         /* bug 1580: decide how to properly sync() in recovery*/
1089                         //mds_fsync_super(mds->mds_sb);
1090                         ptlrpc_free_clone(req);
1091                         spin_lock_bh(&obd->obd_processing_task_lock);
1092                         obd->obd_next_recovery_transno++;
1093                         spin_unlock_bh(&obd->obd_processing_task_lock);
1094                 } else {
1095                         /* recovery is over */
1096                         spin_lock_bh(&obd->obd_processing_task_lock);
1097                         obd->obd_recovering = 0;
1098                         target_cancel_recovery_timer(obd);
1099                         if (obd->obd_abort_recovery) {
1100                                 obd->obd_abort_recovery = 0;
1101                                 spin_unlock_bh(&obd->obd_processing_task_lock);
1102                                 target_abort_recovery(obd);
1103                         } else {
1104                                 LASSERT(obd->obd_recoverable_clients == 0);
1105                                 spin_unlock_bh(&obd->obd_processing_task_lock);
1106                                 target_finish_recovery(obd);
1107                         }
1108                 }
1109         }
1110
1111         trd->trd_processing_task = 0;
1112         complete(&trd->trd_finishing);
1113         return 0;
1114 }
1115
1116 int target_start_recovery_thread(struct obd_device *obd, svc_handler_t handler)
1117 {
1118         int rc = 0;
1119         struct target_recovery_data *trd = &obd->obd_recovery_data;
1120
1121         memset(trd, 0, sizeof(*trd));
1122         init_completion(&trd->trd_starting);
1123         init_completion(&trd->trd_finishing);
1124         trd->trd_recovery_handler = handler;
1125
1126         if (kernel_thread(target_recovery_thread, obd, 0) == 0)
1127                 wait_for_completion(&trd->trd_starting);
1128         else
1129                 rc = -ECHILD;
1130
1131         return rc;
1132 }
1133
1134 void target_stop_recovery_thread(struct obd_device *obd)
1135 {
1136         spin_lock_bh(&obd->obd_processing_task_lock);
1137         if (obd->obd_recovery_data.trd_processing_task > 0) {
1138                 struct target_recovery_data *trd = &obd->obd_recovery_data;
1139                 CERROR("%s: aborting recovery\n", obd->obd_name);
1140                 obd->obd_abort_recovery = 1;
1141                 wake_up(&obd->obd_next_transno_waitq);
1142                 spin_unlock_bh(&obd->obd_processing_task_lock);
1143                 wait_for_completion(&trd->trd_finishing);
1144         } else {
1145                 spin_unlock_bh(&obd->obd_processing_task_lock);
1146         }
1147 }
1148 #endif
1149
1150 int target_queue_recovery_request(struct ptlrpc_request *req,
1151                                   struct obd_device *obd)
1152 {
1153         struct list_head *tmp;
1154         int inserted = 0;
1155         __u64 transno = req->rq_reqmsg->transno;
1156
1157         /* CAVEAT EMPTOR: The incoming request message has been swabbed
1158          * (i.e. buflens etc are in my own byte order), but type-dependent
1159          * buffers (eg mds_body, ost_body etc) have NOT been swabbed. */
1160
1161         if (!transno) {
1162                 INIT_LIST_HEAD(&req->rq_list);
1163                 DEBUG_REQ(D_HA, req, "not queueing");
1164                 return 1;
1165         }
1166
1167
1168         /* If we're processing the queue, we want don't want to queue this
1169          * message.
1170          *
1171          * Also, if this request has a transno less than the one we're waiting
1172          * for, we should process it now.  It could (and currently always will)
1173          * be an open request for a descriptor that was opened some time ago.
1174          *
1175          * Also, a resent, replayed request that has already been
1176          * handled will pass through here and be processed immediately.
1177          */
1178         spin_lock_bh(&obd->obd_processing_task_lock);
1179         if (obd->obd_recovery_data.trd_processing_task == current->pid ||
1180             transno < obd->obd_next_recovery_transno) {
1181                 /* Processing the queue right now, don't re-add. */
1182                 LASSERT(list_empty(&req->rq_list));
1183                 spin_unlock_bh(&obd->obd_processing_task_lock);
1184                 return 1;
1185         }
1186         spin_unlock_bh(&obd->obd_processing_task_lock);
1187
1188         /* A resent, replayed request that is still on the queue; just drop it.
1189            The queued request will handle this. */
1190         if ((lustre_msg_get_flags(req->rq_reqmsg) & (MSG_RESENT | MSG_REPLAY))
1191             == (MSG_RESENT | MSG_REPLAY)) {
1192                 DEBUG_REQ(D_ERROR, req, "dropping resent queued req");
1193                 return 0;
1194         }
1195
1196         req = ptlrpc_clone_req(req);
1197         if (req == NULL)
1198                 return -ENOMEM;
1199
1200         spin_lock_bh(&obd->obd_processing_task_lock);
1201
1202         /* XXX O(n^2) */
1203         list_for_each(tmp, &obd->obd_recovery_queue) {
1204                 struct ptlrpc_request *reqiter =
1205                         list_entry(tmp, struct ptlrpc_request, rq_list);
1206
1207                 if (reqiter->rq_reqmsg->transno > transno) {
1208                         list_add_tail(&req->rq_list, &reqiter->rq_list);
1209                         inserted = 1;
1210                         break;
1211                 }
1212         }
1213
1214         if (!inserted) {
1215                 list_add_tail(&req->rq_list, &obd->obd_recovery_queue);
1216         }
1217
1218         obd->obd_requests_queued_for_recovery++;
1219         wake_up(&obd->obd_next_transno_waitq);
1220         spin_unlock_bh(&obd->obd_processing_task_lock);
1221
1222         return 0;
1223 }
1224
1225 struct obd_device * target_req2obd(struct ptlrpc_request *req)
1226 {
1227         return req->rq_export->exp_obd;
1228 }
1229
1230 int target_queue_final_reply(struct ptlrpc_request *req, int rc)
1231 {
1232         struct obd_device *obd = target_req2obd(req);
1233
1234         LASSERT ((rc == 0) == (req->rq_reply_state != NULL));
1235
1236         if (rc) {
1237                 /* Just like ptlrpc_error, but without the sending. */
1238                 rc = lustre_pack_reply(req, 0, NULL, NULL);
1239                 LASSERT(rc == 0); /* XXX handle this */
1240                 req->rq_type = PTL_RPC_MSG_ERR;
1241         }
1242
1243         LASSERT (!req->rq_reply_state->rs_difficult);
1244         LASSERT(list_empty(&req->rq_list));
1245
1246         req = ptlrpc_clone_req(req);
1247
1248         spin_lock_bh(&obd->obd_processing_task_lock);
1249
1250         list_add(&req->rq_list, &obd->obd_delayed_reply_queue);
1251
1252         /* only count the first "replay over" request from each
1253            export */
1254         if (req->rq_export->exp_replay_needed) {
1255                 --obd->obd_recoverable_clients;
1256                 req->rq_export->exp_replay_needed = 0;
1257                 CWARN("%s: %d recoverable clients remain\n",
1258                       obd->obd_name, obd->obd_recoverable_clients);
1259         }
1260         wake_up(&obd->obd_next_transno_waitq);
1261         spin_unlock_bh(&obd->obd_processing_task_lock);
1262         return 1;
1263 }
1264
1265 int
1266 target_send_reply_msg (struct ptlrpc_request *req, int rc, int fail_id)
1267 {
1268         if (OBD_FAIL_CHECK(fail_id | OBD_FAIL_ONCE)) {
1269                 obd_fail_loc |= OBD_FAIL_ONCE | OBD_FAILED;
1270                 DEBUG_REQ(D_ERROR, req, "dropping reply");
1271                 /* NB this does _not_ send with ACK disabled, to simulate
1272                  * sending OK, but timing out for the ACK */
1273                 if (req->rq_reply_state != NULL) {
1274                         if (!req->rq_reply_state->rs_difficult) {
1275                                 lustre_free_reply_state (req->rq_reply_state);
1276                                 req->rq_reply_state = NULL;
1277                         } else {
1278                                 struct ptlrpc_service *svc =
1279                                         req->rq_rqbd->rqbd_srv_ni->sni_service;
1280                                 atomic_inc(&svc->srv_outstanding_replies);
1281                         }
1282                 }
1283                 return (-ECOMM);
1284         }
1285
1286         if (rc) {
1287                 req->rq_status = rc;
1288                 return (ptlrpc_error(req));
1289         } else {
1290                 DEBUG_REQ(D_NET, req, "sending reply");
1291         }
1292
1293         return (ptlrpc_send_reply(req, 1));
1294 }
1295
1296 void
1297 target_send_reply(struct ptlrpc_request *req, int rc, int fail_id)
1298 {
1299         int                        netrc;
1300         unsigned long              flags;
1301         struct ptlrpc_reply_state *rs;
1302         struct obd_device         *obd;
1303         struct obd_export         *exp;
1304         struct ptlrpc_srv_ni      *sni;
1305         struct ptlrpc_service     *svc;
1306
1307         sni = req->rq_rqbd->rqbd_srv_ni;
1308         svc = sni->sni_service;
1309
1310         rs = req->rq_reply_state;
1311         if (rs == NULL || !rs->rs_difficult) {
1312                 /* The easy case; no notifiers and reply_out_callback()
1313                  * cleans up (i.e. we can't look inside rs after a
1314                  * successful send) */
1315                 netrc = target_send_reply_msg (req, rc, fail_id);
1316
1317                 LASSERT (netrc == 0 || req->rq_reply_state == NULL);
1318                 return;
1319         }
1320
1321         /* must be an export if locks saved */
1322         LASSERT (req->rq_export != NULL);
1323         /* req/reply consistent */
1324         LASSERT (rs->rs_srv_ni == sni);
1325
1326         /* "fresh" reply */
1327         LASSERT (!rs->rs_scheduled);
1328         LASSERT (!rs->rs_scheduled_ever);
1329         LASSERT (!rs->rs_handled);
1330         LASSERT (!rs->rs_on_net);
1331         LASSERT (rs->rs_export == NULL);
1332         LASSERT (list_empty(&rs->rs_obd_list));
1333         LASSERT (list_empty(&rs->rs_exp_list));
1334
1335         exp = class_export_get (req->rq_export);
1336         obd = exp->exp_obd;
1337
1338         /* disable reply scheduling onto srv_reply_queue while I'm setting up */
1339         rs->rs_scheduled = 1;
1340         rs->rs_on_net    = 1;
1341         rs->rs_xid       = req->rq_xid;
1342         rs->rs_transno   = req->rq_transno;
1343         rs->rs_export    = exp;
1344
1345         spin_lock_irqsave (&obd->obd_uncommitted_replies_lock, flags);
1346
1347         if (rs->rs_transno > obd->obd_last_committed) {
1348                 /* not committed already */
1349                 list_add_tail (&rs->rs_obd_list,
1350                                &obd->obd_uncommitted_replies);
1351         }
1352
1353         spin_unlock (&obd->obd_uncommitted_replies_lock);
1354         spin_lock (&exp->exp_lock);
1355
1356         list_add_tail (&rs->rs_exp_list, &exp->exp_outstanding_replies);
1357
1358         spin_unlock_irqrestore (&exp->exp_lock, flags);
1359
1360         netrc = target_send_reply_msg (req, rc, fail_id);
1361
1362         spin_lock_irqsave (&svc->srv_lock, flags);
1363
1364         svc->srv_n_difficult_replies++;
1365
1366         if (netrc != 0) /* error sending: reply is off the net */
1367                 rs->rs_on_net = 0;
1368
1369         if (!rs->rs_on_net ||                   /* some notifier */
1370             list_empty(&rs->rs_exp_list) ||     /* completed already */
1371             list_empty(&rs->rs_obd_list)) {
1372                 list_add_tail (&rs->rs_list, &svc->srv_reply_queue);
1373                 wake_up (&svc->srv_waitq);
1374         } else {
1375                 list_add (&rs->rs_list, &sni->sni_active_replies);
1376                 rs->rs_scheduled = 0;           /* allow notifier to schedule */
1377         }
1378
1379         spin_unlock_irqrestore (&svc->srv_lock, flags);
1380 }
1381
1382 int target_handle_ping(struct ptlrpc_request *req)
1383 {
1384         return lustre_pack_reply(req, 0, NULL, NULL);
1385 }