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