lustre/ldlm/ldlm_lib.c

   1 /* -*- mode: c; c-basic-offset: 8; indent-tabs-mode: nil; -*-
   2  * vim:expandtab:shiftwidth=8:tabstop=8:
   3  *
   4  *  Copyright (c) 2003 Cluster File Systems, Inc.
   5  *
   6  *   This file is part of the Lustre file system, http://www.lustre.org
   7  *   Lustre is a trademark of Cluster File Systems, Inc.
   8  *
   9  *   You may have signed or agreed to another license before downloading
  10  *   this software.  If so, you are bound by the terms and conditions
  11  *   of that agreement, and the following does not apply to you.  See the
  12  *   LICENSE file included with this distribution for more information.
  13  *
  14  *   If you did not agree to a different license, then this copy of Lustre
  15  *   is open source software; you can redistribute it and/or modify it
  16  *   under the terms of version 2 of the GNU General Public License as
  17  *   published by the Free Software Foundation.
  18  *
  19  *   In either case, Lustre is distributed in the hope that it will be
  20  *   useful, but WITHOUT ANY WARRANTY; without even the implied warranty
  21  *   of MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
  22  *   license text for more details.
  23  */
  24
  25 #ifndef EXPORT_SYMTAB
  26 # define EXPORT_SYMTAB
  27 #endif
  28 #define DEBUG_SUBSYSTEM S_LDLM
  29
  30 #ifdef __KERNEL__
  31 # include <libcfs/libcfs.h>
  32 #else
  33 # include <liblustre.h>
  34 #endif
  35 #include <obd.h>
  36 #include <lustre_mds.h>
  37 #include <lustre_dlm.h>
  38 #include <lustre_net.h>
  39 #include <lustre_sec.h>
  40 #include "ldlm_internal.h"
  41
  42
  43 /* @priority: if non-zero, move the selected to the list head
  44  * @create: if zero, only search in existed connections
  45  */
  46 static int import_set_conn(struct obd_import *imp, struct obd_uuid *uuid,
  47                            int priority, int create)
  48 {
  49         struct ptlrpc_connection *ptlrpc_conn;
  50         struct obd_import_conn *imp_conn = NULL, *item;
  51         int rc = 0;
  52         ENTRY;
  53
  54         if (!create && !priority) {
  55                 CDEBUG(D_HA, "Nothing to do\n");
  56                 RETURN(-EINVAL);
  57         }
  58
  59         ptlrpc_conn = ptlrpc_uuid_to_connection(uuid);
  60         if (!ptlrpc_conn) {
  61                 CDEBUG(D_HA, "can't find connection %s\n", uuid->uuid);
  62                 RETURN (-ENOENT);
  63         }
  64
  65         if (create) {
  66                 OBD_ALLOC(imp_conn, sizeof(*imp_conn));
  67                 if (!imp_conn) {
  68                         GOTO(out_put, rc = -ENOMEM);
  69                 }
  70         }
  71
  72         spin_lock(&imp->imp_lock);
  73         list_for_each_entry(item, &imp->imp_conn_list, oic_item) {
  74                 if (obd_uuid_equals(uuid, &item->oic_uuid)) {
  75                         if (priority) {
  76                                 list_del(&item->oic_item);
  77                                 list_add(&item->oic_item, &imp->imp_conn_list);
  78                                 item->oic_last_attempt = 0;
  79                         }
  80                         CDEBUG(D_HA, "imp %p@%s: found existing conn %s%s\n",
  81                                imp, imp->imp_obd->obd_name, uuid->uuid,
  82                                (priority ? ", moved to head" : ""));
  83                         spin_unlock(&imp->imp_lock);
  84                         GOTO(out_free, rc = 0);
  85                 }
  86         }
  87         /* not found */
  88         if (create) {
  89                 imp_conn->oic_conn = ptlrpc_conn;
  90                 imp_conn->oic_uuid = *uuid;
  91                 item->oic_last_attempt = 0;
  92                 if (priority)
  93                         list_add(&imp_conn->oic_item, &imp->imp_conn_list);
  94                 else
  95                         list_add_tail(&imp_conn->oic_item, &imp->imp_conn_list);
  96                 CDEBUG(D_HA, "imp %p@%s: add connection %s at %s\n",
  97                        imp, imp->imp_obd->obd_name, uuid->uuid,
  98                        (priority ? "head" : "tail"));
  99         } else {
 100                 spin_unlock(&imp->imp_lock);
 101                 GOTO(out_free, rc = -ENOENT);
 102         }
 103
 104         spin_unlock(&imp->imp_lock);
 105         RETURN(0);
 106 out_free:
 107         if (imp_conn)
 108                 OBD_FREE(imp_conn, sizeof(*imp_conn));
 109 out_put:
 110         ptlrpc_put_connection(ptlrpc_conn);
 111         RETURN(rc);
 112 }
 113
 114 int import_set_conn_priority(struct obd_import *imp, struct obd_uuid *uuid)
 115 {
 116         return import_set_conn(imp, uuid, 1, 0);
 117 }
 118
 119 int client_import_add_conn(struct obd_import *imp, struct obd_uuid *uuid,
 120                            int priority)
 121 {
 122         return import_set_conn(imp, uuid, priority, 1);
 123 }
 124
 125 int client_import_del_conn(struct obd_import *imp, struct obd_uuid *uuid)
 126 {
 127         struct obd_import_conn *imp_conn;
 128         struct obd_export *dlmexp;
 129         int rc = -ENOENT;
 130         ENTRY;
 131
 132         spin_lock(&imp->imp_lock);
 133         if (list_empty(&imp->imp_conn_list)) {
 134                 LASSERT(!imp->imp_connection);
 135                 GOTO(out, rc);
 136         }
 137
 138         list_for_each_entry(imp_conn, &imp->imp_conn_list, oic_item) {
 139                 if (!obd_uuid_equals(uuid, &imp_conn->oic_uuid))
 140                         continue;
 141                 LASSERT(imp_conn->oic_conn);
 142
 143                 /* is current conn? */
 144                 if (imp_conn == imp->imp_conn_current) {
 145                         LASSERT(imp_conn->oic_conn == imp->imp_connection);
 146
 147                         if (imp->imp_state != LUSTRE_IMP_CLOSED &&
 148                             imp->imp_state != LUSTRE_IMP_DISCON) {
 149                                 CERROR("can't remove current connection\n");
 150                                 GOTO(out, rc = -EBUSY);
 151                         }
 152
 153                         ptlrpc_put_connection(imp->imp_connection);
 154                         imp->imp_connection = NULL;
 155
 156                         dlmexp = class_conn2export(&imp->imp_dlm_handle);
 157                         if (dlmexp && dlmexp->exp_connection) {
 158                                 LASSERT(dlmexp->exp_connection ==
 159                                         imp_conn->oic_conn);
 160                                 ptlrpc_put_connection(dlmexp->exp_connection);
 161                                 dlmexp->exp_connection = NULL;
 162                         }
 163                 }
 164
 165                 list_del(&imp_conn->oic_item);
 166                 ptlrpc_put_connection(imp_conn->oic_conn);
 167                 OBD_FREE(imp_conn, sizeof(*imp_conn));
 168                 CDEBUG(D_HA, "imp %p@%s: remove connection %s\n",
 169                        imp, imp->imp_obd->obd_name, uuid->uuid);
 170                 rc = 0;
 171                 break;
 172         }
 173 out:
 174         spin_unlock(&imp->imp_lock);
 175         if (rc == -ENOENT)
 176                 CERROR("connection %s not found\n", uuid->uuid);
 177         RETURN(rc);
 178 }
 179
 180 static void destroy_import(struct obd_import *imp)
 181 {
 182         /* drop security policy instance after all rpc finished/aborted
 183          * to let all busy credentials be released.
 184          */
 185         class_import_get(imp);
 186         class_destroy_import(imp);
 187         sptlrpc_import_put_sec(imp);
 188         class_import_put(imp);
 189 }
 190
 191 /* configure an RPC client OBD device
 192  *
 193  * lcfg parameters:
 194  * 1 - client UUID
 195  * 2 - server UUID
 196  * 3 - inactive-on-startup
 197  */
 198 int client_obd_setup(struct obd_device *obddev, struct lustre_cfg *lcfg)
 199 {
 200         struct client_obd *cli = &obddev->u.cli;
 201         struct obd_import *imp;
 202         struct obd_uuid server_uuid;
 203         int rq_portal, rp_portal, connect_op;
 204         char *name = obddev->obd_type->typ_name;
 205         int rc;
 206         ENTRY;
 207
 208         /* In a more perfect world, we would hang a ptlrpc_client off of
 209          * obd_type and just use the values from there. */
 210         if (!strcmp(name, LUSTRE_OSC_NAME)) {
 211                 rq_portal = OST_REQUEST_PORTAL;
 212                 rp_portal = OSC_REPLY_PORTAL;
 213                 connect_op = OST_CONNECT;
 214         } else if (!strcmp(name, LUSTRE_MDC_NAME)) {
 215                 rq_portal = MDS_REQUEST_PORTAL;
 216                 rp_portal = MDC_REPLY_PORTAL;
 217                 connect_op = MDS_CONNECT;
 218         } else if (!strcmp(name, LUSTRE_MGC_NAME)) {
 219                 rq_portal = MGS_REQUEST_PORTAL;
 220                 rp_portal = MGC_REPLY_PORTAL;
 221                 connect_op = MGS_CONNECT;
 222         } else {
 223                 CERROR("unknown client OBD type \"%s\", can't setup\n",
 224                        name);
 225                 RETURN(-EINVAL);
 226         }
 227
 228         if (LUSTRE_CFG_BUFLEN(lcfg, 1) < 1) {
 229                 CERROR("requires a TARGET UUID\n");
 230                 RETURN(-EINVAL);
 231         }
 232
 233         if (LUSTRE_CFG_BUFLEN(lcfg, 1) > 37) {
 234                 CERROR("client UUID must be less than 38 characters\n");
 235                 RETURN(-EINVAL);
 236         }
 237
 238         if (LUSTRE_CFG_BUFLEN(lcfg, 2) < 1) {
 239                 CERROR("setup requires a SERVER UUID\n");
 240                 RETURN(-EINVAL);
 241         }
 242
 243         if (LUSTRE_CFG_BUFLEN(lcfg, 2) > 37) {
 244                 CERROR("target UUID must be less than 38 characters\n");
 245                 RETURN(-EINVAL);
 246         }
 247
 248         sema_init(&cli->cl_sem, 1);
 249         sema_init(&cli->cl_mgc_sem, 1);
 250         cli->cl_sec_conf.sfc_rpc_flavor = SPTLRPC_FLVR_NULL;
 251         cli->cl_sec_conf.sfc_bulk_csum = BULK_CSUM_ALG_NULL;
 252         cli->cl_sec_conf.sfc_bulk_priv = BULK_PRIV_ALG_NULL;
 253         cli->cl_sec_conf.sfc_flags = 0;
 254         cli->cl_conn_count = 0;
 255         memcpy(server_uuid.uuid, lustre_cfg_buf(lcfg, 2),
 256                min_t(unsigned int, LUSTRE_CFG_BUFLEN(lcfg, 2),
 257                      sizeof(server_uuid)));
 258
 259         cli->cl_dirty = 0;
 260         cli->cl_avail_grant = 0;
 261         /* FIXME: should limit this for the sum of all cl_dirty_max */
 262         cli->cl_dirty_max = OSC_MAX_DIRTY_DEFAULT * 1024 * 1024;
 263         if (cli->cl_dirty_max >> CFS_PAGE_SHIFT > num_physpages / 8)
 264                 cli->cl_dirty_max = num_physpages << (CFS_PAGE_SHIFT - 3);
 265         CFS_INIT_LIST_HEAD(&cli->cl_cache_waiters);
 266         CFS_INIT_LIST_HEAD(&cli->cl_loi_ready_list);
 267         CFS_INIT_LIST_HEAD(&cli->cl_loi_write_list);
 268         CFS_INIT_LIST_HEAD(&cli->cl_loi_read_list);
 269         client_obd_list_lock_init(&cli->cl_loi_list_lock);
 270         cli->cl_r_in_flight = 0;
 271         cli->cl_w_in_flight = 0;
 272
 273         spin_lock_init(&cli->cl_read_rpc_hist.oh_lock);
 274         spin_lock_init(&cli->cl_write_rpc_hist.oh_lock);
 275         spin_lock_init(&cli->cl_read_page_hist.oh_lock);
 276         spin_lock_init(&cli->cl_write_page_hist.oh_lock);
 277         spin_lock_init(&cli->cl_read_offset_hist.oh_lock);
 278         spin_lock_init(&cli->cl_write_offset_hist.oh_lock);
 279         cfs_waitq_init(&cli->cl_destroy_waitq);
 280         atomic_set(&cli->cl_destroy_in_flight, 0);
 281 #ifdef ENABLE_CHECKSUM
 282         cli->cl_checksum = 1;
 283 #endif
 284         atomic_set(&cli->cl_resends, OSC_DEFAULT_RESENDS);
 285
 286         /* This value may be changed at connect time in
 287            ptlrpc_connect_interpret. */
 288         cli->cl_max_pages_per_rpc = min((int)PTLRPC_MAX_BRW_PAGES,
 289                                         (int)(1024 * 1024 >> CFS_PAGE_SHIFT));
 290
 291         if (!strcmp(name, LUSTRE_MDC_NAME)) {
 292                 cli->cl_max_rpcs_in_flight = MDC_MAX_RIF_DEFAULT;
 293         } else if (num_physpages >> (20 - CFS_PAGE_SHIFT) <= 128 /* MB */) {
 294                 cli->cl_max_rpcs_in_flight = 2;
 295         } else if (num_physpages >> (20 - CFS_PAGE_SHIFT) <= 256 /* MB */) {
 296                 cli->cl_max_rpcs_in_flight = 3;
 297         } else if (num_physpages >> (20 - CFS_PAGE_SHIFT) <= 512 /* MB */) {
 298                 cli->cl_max_rpcs_in_flight = 4;
 299         } else {
 300                 cli->cl_max_rpcs_in_flight = OSC_MAX_RIF_DEFAULT;
 301         }
 302
 303         rc = ldlm_get_ref();
 304         if (rc) {
 305                 CERROR("ldlm_get_ref failed: %d\n", rc);
 306                 GOTO(err, rc);
 307         }
 308
 309         ptlrpc_init_client(rq_portal, rp_portal, name,
 310                            &obddev->obd_ldlm_client);
 311
 312         imp = class_new_import(obddev);
 313         if (imp == NULL)
 314                 GOTO(err_ldlm, rc = -ENOENT);
 315         imp->imp_client = &obddev->obd_ldlm_client;
 316         imp->imp_connect_op = connect_op;
 317         imp->imp_initial_recov = 1;
 318         imp->imp_initial_recov_bk = 0;
 319         CFS_INIT_LIST_HEAD(&imp->imp_pinger_chain);
 320         memcpy(cli->cl_target_uuid.uuid, lustre_cfg_buf(lcfg, 1),
 321                LUSTRE_CFG_BUFLEN(lcfg, 1));
 322         class_import_put(imp);
 323
 324         rc = client_import_add_conn(imp, &server_uuid, 1);
 325         if (rc) {
 326                 CERROR("can't add initial connection\n");
 327                 GOTO(err_import, rc);
 328         }
 329
 330         cli->cl_import = imp;
 331         /* cli->cl_max_mds_{easize,cookiesize} updated by mdc_init_ea_size() */
 332         cli->cl_max_mds_easize = sizeof(struct lov_mds_md);
 333         cli->cl_max_mds_cookiesize = sizeof(struct llog_cookie);
 334
 335         if (LUSTRE_CFG_BUFLEN(lcfg, 3) > 0) {
 336                 if (!strcmp(lustre_cfg_string(lcfg, 3), "inactive")) {
 337                         CDEBUG(D_HA, "marking %s %s->%s as inactive\n",
 338                                name, obddev->obd_name,
 339                                cli->cl_target_uuid.uuid);
 340                         spin_lock(&imp->imp_lock);
 341                         imp->imp_invalid = 1;
 342                         spin_unlock(&imp->imp_lock);
 343                 }
 344         }
 345
 346         cli->cl_qchk_stat = CL_NOT_QUOTACHECKED;
 347
 348         RETURN(rc);
 349
 350 err_import:
 351         class_destroy_import(imp);
 352 err_ldlm:
 353         ldlm_put_ref(0);
 354 err:
 355         RETURN(rc);
 356
 357 }
 358
 359 int client_obd_cleanup(struct obd_device *obddev)
 360 {
 361         ENTRY;
 362         ldlm_put_ref(obddev->obd_force);
 363         RETURN(0);
 364 }
 365
 366 /* ->o_connect() method for client side (OSC and MDC and MGC) */
 367 int client_connect_import(const struct lu_env *env,
 368                           struct lustre_handle *dlm_handle,
 369                           struct obd_device *obd, struct obd_uuid *cluuid,
 370                           struct obd_connect_data *data)
 371 {
 372         struct client_obd *cli = &obd->u.cli;
 373         struct obd_import *imp = cli->cl_import;
 374         struct obd_export *exp;
 375         struct obd_connect_data *ocd;
 376         struct ldlm_namespace *to_be_freed = NULL;
 377         int rc;
 378         ENTRY;
 379
 380         mutex_down(&cli->cl_sem);
 381         rc = class_connect(dlm_handle, obd, cluuid);
 382         if (rc)
 383                 GOTO(out_sem, rc);
 384
 385         cli->cl_conn_count++;
 386         if (cli->cl_conn_count > 1)
 387                 GOTO(out_sem, rc);
 388         exp = class_conn2export(dlm_handle);
 389
 390         if (obd->obd_namespace != NULL)
 391                 CERROR("already have namespace!\n");
 392         obd->obd_namespace = ldlm_namespace_new(obd->obd_name,
 393                                                 LDLM_NAMESPACE_CLIENT,
 394                                                 LDLM_NAMESPACE_GREEDY);
 395         if (obd->obd_namespace == NULL)
 396                 GOTO(out_disco, rc = -ENOMEM);
 397
 398         imp->imp_dlm_handle = *dlm_handle;
 399         rc = ptlrpc_init_import(imp);
 400         if (rc != 0)
 401                 GOTO(out_ldlm, rc);
 402
 403         rc = sptlrpc_import_get_sec(imp, NULL, cli->cl_sec_conf.sfc_rpc_flavor,
 404                                     cli->cl_sec_conf.sfc_flags);
 405         if (rc)
 406                 GOTO(out_ldlm, rc);
 407
 408         ocd = &imp->imp_connect_data;
 409         if (data) {
 410                 *ocd = *data;
 411                 imp->imp_connect_flags_orig = data->ocd_connect_flags;
 412         }
 413
 414         rc = ptlrpc_connect_import(imp, NULL);
 415         if (rc != 0) {
 416                 LASSERT (imp->imp_state == LUSTRE_IMP_DISCON);
 417                 GOTO(out_ldlm, rc);
 418         }
 419         LASSERT(exp->exp_connection);
 420
 421         if (data) {
 422                 LASSERT((ocd->ocd_connect_flags & data->ocd_connect_flags) ==
 423                         ocd->ocd_connect_flags);
 424                 data->ocd_connect_flags = ocd->ocd_connect_flags;
 425         }
 426
 427         ptlrpc_pinger_add_import(imp);
 428
 429         EXIT;
 430
 431         if (rc) {
 432 out_ldlm:
 433                 ldlm_namespace_free_prior(obd->obd_namespace);
 434                 to_be_freed = obd->obd_namespace;
 435                 obd->obd_namespace = NULL;
 436 out_disco:
 437                 cli->cl_conn_count--;
 438                 class_disconnect(exp);
 439         } else {
 440                 class_export_put(exp);
 441         }
 442 out_sem:
 443         mutex_up(&cli->cl_sem);
 444         if (to_be_freed)
 445                 ldlm_namespace_free_post(to_be_freed, 1);
 446
 447         return rc;
 448 }
 449
 450 int client_disconnect_export(struct obd_export *exp)
 451 {
 452         struct obd_device *obd = class_exp2obd(exp);
 453         struct client_obd *cli;
 454         struct obd_import *imp;
 455         int rc = 0, err;
 456         struct ldlm_namespace *to_be_freed = NULL;
 457         ENTRY;
 458
 459         if (!obd) {
 460                 CERROR("invalid export for disconnect: exp %p cookie "LPX64"\n",
 461                        exp, exp ? exp->exp_handle.h_cookie : -1);
 462                 RETURN(-EINVAL);
 463         }
 464
 465         cli = &obd->u.cli;
 466         imp = cli->cl_import;
 467
 468         mutex_down(&cli->cl_sem);
 469         if (!cli->cl_conn_count) {
 470                 CERROR("disconnecting disconnected device (%s)\n",
 471                        obd->obd_name);
 472                 GOTO(out_sem, rc = -EINVAL);
 473         }
 474
 475         cli->cl_conn_count--;
 476         if (cli->cl_conn_count)
 477                 GOTO(out_no_disconnect, rc = 0);
 478
 479         /* Mark import deactivated now, so we don't try to reconnect if any
 480          * of the cleanup RPCs fails (e.g. ldlm cancel, etc).  We don't
 481          * fully deactivate the import, or that would drop all requests. */
 482         spin_lock(&imp->imp_lock);
 483         imp->imp_deactive = 1;
 484         spin_unlock(&imp->imp_lock);
 485
 486         /* Some non-replayable imports (MDS's OSCs) are pinged, so just
 487          * delete it regardless.  (It's safe to delete an import that was
 488          * never added.) */
 489         (void)ptlrpc_pinger_del_import(imp);
 490
 491         if (obd->obd_namespace != NULL) {
 492                 /* obd_force == local only */
 493                 ldlm_cli_cancel_unused(obd->obd_namespace, NULL,
 494                                        obd->obd_force ? LDLM_FL_LOCAL_ONLY:0,
 495                                        NULL);
 496                 ldlm_namespace_free_prior(obd->obd_namespace);
 497                 to_be_freed = obd->obd_namespace;
 498         }
 499
 500         rc = ptlrpc_disconnect_import(imp, 0);
 501
 502         ptlrpc_invalidate_import(imp);
 503         /* set obd_namespace to NULL only after invalidate, because we can have
 504          * some connect requests in flight, and his need store a connect flags
 505          * in obd_namespace. bug 14260 */
 506         obd->obd_namespace = NULL;
 507
 508         ptlrpc_free_rq_pool(imp->imp_rq_pool);
 509         destroy_import(imp);
 510         cli->cl_import = NULL;
 511
 512         EXIT;
 513  out_no_disconnect:
 514         err = class_disconnect(exp);
 515         if (!rc && err)
 516                 rc = err;
 517  out_sem:
 518         mutex_up(&cli->cl_sem);
 519         if (to_be_freed)
 520                 ldlm_namespace_free_post(to_be_freed, obd->obd_force);
 521
 522         RETURN(rc);
 523 }
 524
 525 /* --------------------------------------------------------------------------
 526  * from old lib/target.c
 527  * -------------------------------------------------------------------------- */
 528
 529 int target_handle_reconnect(struct lustre_handle *conn, struct obd_export *exp,
 530                             struct obd_uuid *cluuid, int initial_conn)
 531 {
 532         ENTRY;
 533         if (exp->exp_connection && exp->exp_imp_reverse && !initial_conn) {
 534                 struct lustre_handle *hdl;
 535                 hdl = &exp->exp_imp_reverse->imp_remote_handle;
 536                 /* Might be a re-connect after a partition. */
 537                 if (!memcmp(&conn->cookie, &hdl->cookie, sizeof conn->cookie)) {
 538                         CWARN("%s: %s reconnecting\n", exp->exp_obd->obd_name,
 539                               cluuid->uuid);
 540                         conn->cookie = exp->exp_handle.h_cookie;
 541                         /* target_handle_connect() treats EALREADY and
 542                          * -EALREADY differently.  EALREADY means we are
 543                          * doing a valid reconnect from the same client. */
 544                         RETURN(EALREADY);
 545                 } else {
 546                         CERROR("%s reconnecting from %s, "
 547                                "handle mismatch (ours "LPX64", theirs "
 548                                LPX64")\n", cluuid->uuid,
 549                                exp->exp_connection->c_remote_uuid.uuid,
 550                                hdl->cookie, conn->cookie);
 551                         memset(conn, 0, sizeof *conn);
 552                         /* target_handle_connect() treats EALREADY and
 553                          * -EALREADY differently.  -EALREADY is an error
 554                          * (same UUID, different handle). */
 555                         RETURN(-EALREADY);
 556                 }
 557         }
 558
 559         conn->cookie = exp->exp_handle.h_cookie;
 560         CDEBUG(D_HA, "connect export for UUID '%s' at %p, cookie "LPX64"\n",
 561                cluuid->uuid, exp, conn->cookie);
 562         RETURN(0);
 563 }
 564
 565 void target_client_add_cb(struct obd_device *obd, __u64 transno, void *cb_data,
 566                           int error)
 567 {
 568         struct obd_export *exp = cb_data;
 569
 570         CDEBUG(D_RPCTRACE, "%s: committing for initial connect of %s\n",
 571                obd->obd_name, exp->exp_client_uuid.uuid);
 572
 573         spin_lock(&exp->exp_lock);
 574         exp->exp_need_sync = 0;
 575         spin_unlock(&exp->exp_lock);
 576 }
 577 EXPORT_SYMBOL(target_client_add_cb);
 578
 579 int target_handle_connect(struct ptlrpc_request *req)
 580 {
 581         struct obd_device *target, *targref = NULL;
 582         struct obd_export *export = NULL;
 583         struct obd_import *revimp;
 584         struct lustre_handle conn;
 585         struct obd_uuid tgtuuid;
 586         struct obd_uuid cluuid;
 587         struct obd_uuid remote_uuid;
 588         char *str, *tmp;
 589         int rc = 0;
 590         int initial_conn = 0;
 591         struct obd_connect_data *data;
 592         int size[2] = { sizeof(struct ptlrpc_body), sizeof(*data) };
 593         ENTRY;
 594
 595         OBD_RACE(OBD_FAIL_TGT_CONN_RACE);
 596
 597         lustre_set_req_swabbed(req, REQ_REC_OFF);
 598         str = lustre_msg_string(req->rq_reqmsg, REQ_REC_OFF, sizeof(tgtuuid)-1);
 599         if (str == NULL) {
 600                 DEBUG_REQ(D_ERROR, req, "bad target UUID for connect");
 601                 GOTO(out, rc = -EINVAL);
 602         }
 603
 604         obd_str2uuid (&tgtuuid, str);
 605         target = class_uuid2obd(&tgtuuid);
 606         if (!target)
 607                 target = class_name2obd(str);
 608
 609         if (!target || target->obd_stopping || !target->obd_set_up) {
 610                 LCONSOLE_ERROR_MSG(0x137, "UUID '%s' is not available "
 611                                    " for connect (%s)\n", str,
 612                                    !target ? "no target" :
 613                                    (target->obd_stopping ? "stopping" :
 614                                    "not set up"));
 615                 GOTO(out, rc = -ENODEV);
 616         }
 617
 618         if (target->obd_no_conn) {
 619                 LCONSOLE_WARN("%s: temporarily refusing client connection "
 620                               "from %s\n", target->obd_name,
 621                               libcfs_nid2str(req->rq_peer.nid));
 622                 GOTO(out, rc = -EAGAIN);
 623         }
 624
 625         /* Make sure the target isn't cleaned up while we're here. Yes,
 626            there's still a race between the above check and our incref here.
 627            Really, class_uuid2obd should take the ref. */
 628         targref = class_incref(target);
 629
 630         lustre_set_req_swabbed(req, REQ_REC_OFF + 1);
 631         str = lustre_msg_string(req->rq_reqmsg, REQ_REC_OFF + 1,
 632                                 sizeof(cluuid) - 1);
 633         if (str == NULL) {
 634                 DEBUG_REQ(D_ERROR, req, "bad client UUID for connect");
 635                 GOTO(out, rc = -EINVAL);
 636         }
 637
 638         obd_str2uuid (&cluuid, str);
 639
 640         /* XXX extract a nettype and format accordingly */
 641         switch (sizeof(lnet_nid_t)) {
 642                 /* NB the casts only avoid compiler warnings */
 643         case 8:
 644                 snprintf(remote_uuid.uuid, sizeof remote_uuid,
 645                          "NET_"LPX64"_UUID", (__u64)req->rq_peer.nid);
 646                 break;
 647         case 4:
 648                 snprintf(remote_uuid.uuid, sizeof remote_uuid,
 649                          "NET_%x_UUID", (__u32)req->rq_peer.nid);
 650                 break;
 651         default:
 652                 LBUG();
 653         }
 654
 655         tmp = lustre_msg_buf(req->rq_reqmsg, REQ_REC_OFF + 2, sizeof conn);
 656         if (tmp == NULL)
 657                 GOTO(out, rc = -EPROTO);
 658
 659         memcpy(&conn, tmp, sizeof conn);
 660
 661         data = lustre_swab_reqbuf(req, REQ_REC_OFF + 3, sizeof(*data),
 662                                   lustre_swab_connect);
 663
 664         if (!data)
 665                 GOTO(out, rc = -EPROTO);
 666
 667         rc = lustre_pack_reply(req, 2, size, NULL);
 668         if (rc)
 669                 GOTO(out, rc);
 670
 671         if (lustre_msg_get_op_flags(req->rq_reqmsg) & MSG_CONNECT_LIBCLIENT) {
 672                 if (!data) {
 673                         DEBUG_REQ(D_WARNING, req, "Refusing old (unversioned) "
 674                                   "libclient connection attempt");
 675                         GOTO(out, rc = -EPROTO);
 676                 } else if (data->ocd_version < LUSTRE_VERSION_CODE -
 677                                                LUSTRE_VERSION_ALLOWED_OFFSET ||
 678                            data->ocd_version > LUSTRE_VERSION_CODE +
 679                                                LUSTRE_VERSION_ALLOWED_OFFSET) {
 680                         DEBUG_REQ(D_WARNING, req, "Refusing %s (%d.%d.%d.%d) "
 681                                   "libclient connection attempt",
 682                                   data->ocd_version < LUSTRE_VERSION_CODE ?
 683                                   "old" : "new",
 684                                   OBD_OCD_VERSION_MAJOR(data->ocd_version),
 685                                   OBD_OCD_VERSION_MINOR(data->ocd_version),
 686                                   OBD_OCD_VERSION_PATCH(data->ocd_version),
 687                                   OBD_OCD_VERSION_FIX(data->ocd_version));
 688                         data = lustre_msg_buf(req->rq_repmsg, REPLY_REC_OFF,
 689                                               offsetof(typeof(*data),
 690                                                        ocd_version) +
 691                                               sizeof(data->ocd_version));
 692                         if (data) {
 693                                 data->ocd_connect_flags = OBD_CONNECT_VERSION;
 694                                 data->ocd_version = LUSTRE_VERSION_CODE;
 695                         }
 696                         GOTO(out, rc = -EPROTO);
 697                 }
 698         }
 699
 700         if (lustre_msg_get_op_flags(req->rq_reqmsg) & MSG_CONNECT_INITIAL)
 701                 initial_conn = 1;
 702
 703         /* lctl gets a backstage, all-access pass. */
 704         if (obd_uuid_equals(&cluuid, &target->obd_uuid))
 705                 goto dont_check_exports;
 706
 707         spin_lock(&target->obd_dev_lock);
 708         export = lustre_hash_get_object_by_key(target->obd_uuid_hash_body, &cluuid);
 709
 710         if (export != NULL && export->exp_connecting) { /* bug 9635, et. al. */
 711                 CWARN("%s: exp %p already connecting\n",
 712                       export->exp_obd->obd_name, export);
 713                 class_export_put(export);
 714                 export = NULL;
 715                 rc = -EALREADY;
 716         } else if (export != NULL && export->exp_connection != NULL &&
 717                    req->rq_peer.nid != export->exp_connection->c_peer.nid) {
 718                 /* make darn sure this is coming from the same peer
 719                  * if the UUIDs matched */
 720                   CWARN("%s: cookie %s seen on new NID %s when "
 721                           "existing NID %s is already connected\n",
 722                         target->obd_name, cluuid.uuid,
 723                   libcfs_nid2str(req->rq_peer.nid),
 724                   libcfs_nid2str(export->exp_connection->c_peer.nid));
 725                   class_export_put(export);
 726                   export = NULL;
 727                   rc = -EALREADY;
 728         } else if (export != NULL) {
 729                 spin_lock(&export->exp_lock);
 730                 export->exp_connecting = 1;
 731                 spin_unlock(&export->exp_lock);
 732                 class_export_put(export);
 733                 spin_unlock(&target->obd_dev_lock);
 734                 LASSERT(export->exp_obd == target);
 735
 736                 rc = target_handle_reconnect(&conn, export, &cluuid, initial_conn);
 737         }
 738
 739         /* If we found an export, we already unlocked. */
 740         if (!export) {
 741                 spin_unlock(&target->obd_dev_lock);
 742                 OBD_FAIL_TIMEOUT(OBD_FAIL_TGT_DELAY_CONNECT, 2 * obd_timeout);
 743         } else if (req->rq_export == NULL &&
 744                    atomic_read(&export->exp_rpc_count) > 0) {
 745                 CWARN("%s: refuse connection from %s/%s to 0x%p/%d\n",
 746                       target->obd_name, cluuid.uuid,
 747                       libcfs_nid2str(req->rq_peer.nid),
 748                       export, atomic_read(&export->exp_refcount));
 749                 GOTO(out, rc = -EBUSY);
 750         } else if (req->rq_export != NULL &&
 751                    (atomic_read(&export->exp_rpc_count) > 1)) {
 752                 CWARN("%s: refuse reconnection from %s@%s to 0x%p/%d\n",
 753                       target->obd_name, cluuid.uuid,
 754                       libcfs_nid2str(req->rq_peer.nid),
 755                       export, atomic_read(&export->exp_rpc_count));
 756                 GOTO(out, rc = -EBUSY);
 757         } else if (lustre_msg_get_conn_cnt(req->rq_reqmsg) == 1 &&
 758                    !initial_conn) {
 759                 CERROR("%s: NID %s (%s) reconnected with 1 conn_cnt; "
 760                        "cookies not random?\n", target->obd_name,
 761                        libcfs_nid2str(req->rq_peer.nid), cluuid.uuid);
 762                 GOTO(out, rc = -EALREADY);
 763         } else {
 764                 OBD_FAIL_TIMEOUT(OBD_FAIL_TGT_DELAY_RECONNECT, 2 * obd_timeout);
 765                 if (req->rq_export == NULL && initial_conn)
 766                        export->exp_last_request_time =
 767                                max(export->exp_last_request_time,
 768                                    (time_t)CURRENT_SECONDS);
 769         }
 770
 771         /* We want to handle EALREADY but *not* -EALREADY from
 772          * target_handle_reconnect(), return reconnection state in a flag */
 773         if (rc == EALREADY) {
 774                 lustre_msg_add_op_flags(req->rq_repmsg, MSG_CONNECT_RECONNECT);
 775                 rc = 0;
 776         } else if (rc) {
 777                 GOTO(out, rc);
 778         }
 779         /* Tell the client if we're in recovery. */
 780         /* If this is the first client, start the recovery timer */
 781         CWARN("%s: connection from %s@%s %st"LPU64" exp %p cur %ld last %ld\n",
 782                target->obd_name, cluuid.uuid, libcfs_nid2str(req->rq_peer.nid),
 783               target->obd_recovering ? "recovering/" : "", data->ocd_transno,
 784               export, (long)CURRENT_SECONDS,
 785               export ? (long)export->exp_last_request_time : 0);
 786
 787
 788         if (target->obd_recovering) {
 789                 lustre_msg_add_op_flags(req->rq_repmsg, MSG_CONNECT_RECOVERING);
 790                 target_start_recovery_timer(target);
 791         }
 792
 793         /* Tell the client if we support replayable requests */
 794         if (target->obd_replayable)
 795                 lustre_msg_add_op_flags(req->rq_repmsg, MSG_CONNECT_REPLAYABLE);
 796
 797         if (export == NULL) {
 798                 if (target->obd_recovering) {
 799                         CERROR("%s: denying connection for new client %s (%s): "
 800                                "%d clients in recovery for %lds\n",
 801                                target->obd_name,
 802                                libcfs_nid2str(req->rq_peer.nid), cluuid.uuid,
 803                                target->obd_recoverable_clients,
 804                                cfs_duration_sec(cfs_time_sub(cfs_timer_deadline(&target->obd_recovery_timer),
 805                                                              cfs_time_current())));
 806                         rc = -EBUSY;
 807                 } else {
 808 dont_check_exports:
 809                         rc = obd_connect(req->rq_svc_thread->t_env,
 810                                          &conn, target, &cluuid, data);
 811                 }
 812         } else {
 813                 rc = obd_reconnect(export, target, &cluuid, data);
 814         }
 815         if (rc)
 816                 GOTO(out, rc);
 817         /* Return only the parts of obd_connect_data that we understand, so the
 818          * client knows that we don't understand the rest. */
 819         if (data)
 820                 memcpy(lustre_msg_buf(req->rq_repmsg, REPLY_REC_OFF,
 821                                       sizeof(*data)),
 822                        data, sizeof(*data));
 823
 824         /* If all else goes well, this is our RPC return code. */
 825         req->rq_status = 0;
 826
 827         lustre_msg_set_handle(req->rq_repmsg, &conn);
 828
 829         /* ownership of this export ref transfers to the request AFTER we
 830          * drop any previous reference the request had, but we don't want
 831          * that to go to zero before we get our new export reference. */
 832         export = class_conn2export(&conn);
 833         if (!export) {
 834                 DEBUG_REQ(D_ERROR, req, "Missing export!");
 835                 GOTO(out, rc = -ENODEV);
 836         }
 837
 838         /* If the client and the server are the same node, we will already
 839          * have an export that really points to the client's DLM export,
 840          * because we have a shared handles table.
 841          *
 842          * XXX this will go away when shaver stops sending the "connect" handle
 843          * in the real "remote handle" field of the request --phik 24 Apr 2003
 844          */
 845         if (req->rq_export != NULL)
 846                 class_export_put(req->rq_export);
 847
 848         req->rq_export = export;
 849
 850         spin_lock(&export->exp_lock);
 851         if (initial_conn) {
 852                 lustre_msg_set_conn_cnt(req->rq_repmsg, export->exp_conn_cnt + 1);
 853         } else if (export->exp_conn_cnt >= lustre_msg_get_conn_cnt(req->rq_reqmsg)) {
 854                 spin_unlock(&export->exp_lock);
 855                 CERROR("%s: %s already connected at higher conn_cnt: %d > %d\n",
 856                        cluuid.uuid, libcfs_nid2str(req->rq_peer.nid),
 857                        export->exp_conn_cnt,
 858                        lustre_msg_get_conn_cnt(req->rq_reqmsg));
 859
 860                 GOTO(out, rc = -EALREADY);
 861         }
 862         export->exp_conn_cnt = lustre_msg_get_conn_cnt(req->rq_reqmsg);
 863
 864         /* request from liblustre?  Don't evict it for not pinging. */
 865         if (lustre_msg_get_op_flags(req->rq_reqmsg) & MSG_CONNECT_LIBCLIENT) {
 866                 export->exp_libclient = 1;
 867                 spin_unlock(&export->exp_lock);
 868
 869                 spin_lock(&target->obd_dev_lock);
 870                 list_del_init(&export->exp_obd_chain_timed);
 871                 spin_unlock(&target->obd_dev_lock);
 872         } else {
 873                 spin_unlock(&export->exp_lock);
 874         }
 875
 876         if (export->exp_connection != NULL)
 877                 ptlrpc_put_connection(export->exp_connection);
 878         export->exp_connection = ptlrpc_get_connection(req->rq_peer,
 879                                                        req->rq_self,
 880                                                        &remote_uuid);
 881
 882         spin_lock(&target->obd_dev_lock);
 883         /* Export might be hashed already, e.g. if this is reconnect */
 884         if (hlist_unhashed(&export->exp_nid_hash))
 885                 lustre_hash_additem(export->exp_obd->obd_nid_hash_body,
 886                                     &export->exp_connection->c_peer.nid,
 887                                     &export->exp_nid_hash);
 888         spin_unlock(&target->obd_dev_lock);
 889
 890         spin_lock_bh(&target->obd_processing_task_lock);
 891         if (target->obd_recovering && !export->exp_in_recovery) {
 892                 spin_lock(&export->exp_lock);
 893                 export->exp_in_recovery = 1;
 894                 export->exp_req_replay_needed = 1;
 895                 export->exp_lock_replay_needed = 1;
 896                 spin_unlock(&export->exp_lock);
 897                 if ((lustre_msg_get_op_flags(req->rq_reqmsg) & MSG_CONNECT_TRANSNO)
 898                      && data->ocd_transno < target->obd_next_recovery_transno)
 899                         target->obd_next_recovery_transno = data->ocd_transno;
 900                 target->obd_connected_clients++;
 901                 /* each connected client is counted as recoverable */
 902                 target->obd_recoverable_clients++;
 903                 atomic_inc(&target->obd_req_replay_clients);
 904                 atomic_inc(&target->obd_lock_replay_clients);
 905                 if (target->obd_connected_clients ==
 906                     target->obd_max_recoverable_clients)
 907                         wake_up(&target->obd_next_transno_waitq);
 908         }
 909         spin_unlock_bh(&target->obd_processing_task_lock);
 910         memcpy(&conn,
 911                lustre_msg_buf(req->rq_reqmsg, REQ_REC_OFF + 2, sizeof conn),
 912                sizeof conn);
 913
 914         if (export->exp_imp_reverse != NULL) {
 915                 /* destroyed import can be still referenced in ctxt */
 916                 obd_set_info_async(export, strlen(KEY_REVIMP_UPD),
 917                                    KEY_REVIMP_UPD, 0, NULL, NULL);
 918                 destroy_import(export->exp_imp_reverse);
 919         }
 920
 921         /* for the rest part, we return -ENOTCONN in case of errors
 922          * in order to let client initialize connection again.
 923          */
 924         revimp = export->exp_imp_reverse = class_new_import(target);
 925         if (!revimp) {
 926                 CERROR("fail to alloc new reverse import.\n");
 927                 GOTO(out, rc = -ENOTCONN);
 928         }
 929
 930         revimp->imp_connection = ptlrpc_connection_addref(export->exp_connection);
 931         revimp->imp_client = &export->exp_obd->obd_ldlm_client;
 932         revimp->imp_remote_handle = conn;
 933         revimp->imp_dlm_fake = 1;
 934         revimp->imp_state = LUSTRE_IMP_FULL;
 935
 936         if (lustre_msg_get_op_flags(req->rq_reqmsg) & MSG_CONNECT_NEXT_VER) {
 937                 revimp->imp_msg_magic = LUSTRE_MSG_MAGIC_V2;
 938                 lustre_msg_add_op_flags(req->rq_repmsg, MSG_CONNECT_NEXT_VER);
 939         }
 940
 941         rc = sptlrpc_import_get_sec(revimp, req->rq_svc_ctx,
 942                                     req->rq_sec_flavor, 0);
 943         if (rc) {
 944                 CERROR("Failed to get sec for reverse import: %d\n", rc);
 945                 export->exp_imp_reverse = NULL;
 946                 class_destroy_import(revimp);
 947         }
 948
 949         class_import_put(revimp);
 950 out:
 951         if (export) {
 952                 spin_lock(&export->exp_lock);
 953                 export->exp_connecting = 0;
 954                 spin_unlock(&export->exp_lock);
 955         }
 956         if (targref)
 957                 class_decref(targref);
 958         if (rc)
 959                 req->rq_status = rc;
 960         RETURN(rc);
 961 }
 962
 963 int target_handle_disconnect(struct ptlrpc_request *req)
 964 {
 965         int rc;
 966         ENTRY;
 967
 968         rc = lustre_pack_reply(req, 1, NULL, NULL);
 969         if (rc)
 970                 RETURN(rc);
 971
 972         /* keep the rq_export around so we can send the reply */
 973         req->rq_status = obd_disconnect(class_export_get(req->rq_export));
 974
 975         RETURN(0);
 976 }
 977
 978 void target_destroy_export(struct obd_export *exp)
 979 {
 980         /* exports created from last_rcvd data, and "fake"
 981            exports created by lctl don't have an import */
 982         if (exp->exp_imp_reverse != NULL)
 983                 destroy_import(exp->exp_imp_reverse);
 984
 985         /* We cancel locks at disconnect time, but this will catch any locks
 986          * granted in a race with recovery-induced disconnect. */
 987         if (exp->exp_obd->obd_namespace != NULL)
 988                 ldlm_cancel_locks_for_export(exp);
 989 }
 990
 991 /*
 992  * Recovery functions
 993  */
 994
 995 struct ptlrpc_request *ptlrpc_clone_req( struct ptlrpc_request *orig_req)
 996 {
 997         struct ptlrpc_request *copy_req;
 998         struct lustre_msg *copy_reqmsg;
 999         struct ptlrpc_user_desc *udesc = NULL;
1000
1001         OBD_ALLOC_PTR(copy_req);
1002         if (!copy_req)
1003                 return NULL;
1004         OBD_ALLOC(copy_reqmsg, orig_req->rq_reqlen);
1005         if (!copy_reqmsg){
1006                 OBD_FREE_PTR(copy_req);
1007                 return NULL;
1008         }
1009
1010         if (orig_req->rq_user_desc) {
1011                 int ngroups = orig_req->rq_user_desc->pud_ngroups;
1012
1013                 OBD_ALLOC(udesc, sptlrpc_user_desc_size(ngroups));
1014                 if (!udesc) {
1015                         OBD_FREE(copy_reqmsg, orig_req->rq_reqlen);
1016                         OBD_FREE_PTR(copy_req);
1017                         return NULL;
1018                 }
1019                 memcpy(udesc, orig_req->rq_user_desc,
1020                        sptlrpc_user_desc_size(ngroups));
1021         }
1022
1023         *copy_req = *orig_req;
1024         memcpy(copy_reqmsg, orig_req->rq_reqmsg, orig_req->rq_reqlen);
1025         copy_req->rq_reqmsg = copy_reqmsg;
1026         copy_req->rq_user_desc = udesc;
1027
1028         class_export_get(copy_req->rq_export);
1029         CFS_INIT_LIST_HEAD(&copy_req->rq_list);
1030         sptlrpc_svc_ctx_addref(copy_req);
1031
1032         if (copy_req->rq_reply_state) {
1033                 /* the copied req takes over the reply state */
1034                 orig_req->rq_reply_state = NULL;
1035                 /* to catch further access */
1036                 orig_req->rq_repmsg = NULL;
1037                 orig_req->rq_replen = 0;
1038         }
1039
1040         return copy_req;
1041 }
1042
1043 void ptlrpc_free_clone( struct ptlrpc_request *req)
1044 {
1045         if (req->rq_reply_state) {
1046                 ptlrpc_rs_decref(req->rq_reply_state);
1047                 req->rq_reply_state = NULL;
1048         }
1049
1050         sptlrpc_svc_ctx_decref(req);
1051         class_export_put(req->rq_export);
1052         list_del(&req->rq_list);
1053
1054         if (req->rq_user_desc) {
1055                 int ngroups = req->rq_user_desc->pud_ngroups;
1056                 OBD_FREE(req->rq_user_desc, sptlrpc_user_desc_size(ngroups));
1057         }
1058         OBD_FREE(req->rq_reqmsg, req->rq_reqlen);
1059         OBD_FREE_PTR(req);
1060 }
1061
1062 #ifdef __KERNEL__
1063 static void target_finish_recovery(struct obd_device *obd)
1064 {
1065         ENTRY;
1066         LCONSOLE_INFO("%s: sending delayed replies to recovered clients\n",
1067                       obd->obd_name);
1068
1069         ldlm_reprocess_all_ns(obd->obd_namespace);
1070
1071         /* when recovery finished, cleanup orphans on mds and ost */
1072         if (OBT(obd) && OBP(obd, postrecov)) {
1073                 int rc = OBP(obd, postrecov)(obd);
1074                 LCONSOLE_WARN("%s: recovery %s: rc %d\n", obd->obd_name,
1075                               rc < 0 ? "failed" : "complete", rc);
1076         }
1077
1078         obd->obd_recovery_end = CURRENT_SECONDS;
1079         EXIT;
1080 }
1081
1082 static void abort_req_replay_queue(struct obd_device *obd)
1083 {
1084         struct ptlrpc_request *req, *n;
1085
1086         list_for_each_entry_safe(req, n, &obd->obd_req_replay_queue, rq_list) {
1087                 DEBUG_REQ(D_WARNING, req, "aborted:");
1088                 req->rq_status = -ENOTCONN;
1089                 if (ptlrpc_error(req)) {
1090                         DEBUG_REQ(D_ERROR, req,
1091                                   "failed abort_req_reply; skipping");
1092                 }
1093                 ptlrpc_free_clone(req);
1094         }
1095 }
1096
1097 static void abort_lock_replay_queue(struct obd_device *obd)
1098 {
1099         struct ptlrpc_request *req, *n;
1100
1101         list_for_each_entry_safe(req, n, &obd->obd_lock_replay_queue, rq_list){
1102                 DEBUG_REQ(D_ERROR, req, "aborted:");
1103                 req->rq_status = -ENOTCONN;
1104                 if (ptlrpc_error(req)) {
1105                         DEBUG_REQ(D_ERROR, req,
1106                                   "failed abort_lock_reply; skipping");
1107                 }
1108                 ptlrpc_free_clone(req);
1109         }
1110 }
1111 #endif
1112
1113 /* Called from a cleanup function if the device is being cleaned up
1114    forcefully.  The exports should all have been disconnected already,
1115    the only thing left to do is
1116      - clear the recovery flags
1117      - cancel the timer
1118      - free queued requests and replies, but don't send replies
1119    Because the obd_stopping flag is set, no new requests should be received.
1120
1121 */
1122 void target_cleanup_recovery(struct obd_device *obd)
1123 {
1124         struct ptlrpc_request *req, *n;
1125         ENTRY;
1126
1127         LASSERT(obd->obd_stopping);
1128
1129         spin_lock_bh(&obd->obd_processing_task_lock);
1130         if (!obd->obd_recovering) {
1131                 spin_unlock_bh(&obd->obd_processing_task_lock);
1132                 EXIT;
1133                 return;
1134         }
1135         obd->obd_recovering = obd->obd_abort_recovery = 0;
1136         target_cancel_recovery_timer(obd);
1137         spin_unlock_bh(&obd->obd_processing_task_lock);
1138
1139         list_for_each_entry_safe(req, n, &obd->obd_req_replay_queue, rq_list) {
1140                 LASSERT (req->rq_reply_state == 0);
1141                 ptlrpc_free_clone(req);
1142         }
1143         list_for_each_entry_safe(req, n, &obd->obd_lock_replay_queue, rq_list){
1144                 LASSERT (req->rq_reply_state == 0);
1145                 ptlrpc_free_clone(req);
1146         }
1147         list_for_each_entry_safe(req, n, &obd->obd_final_req_queue, rq_list) {
1148                 LASSERT (req->rq_reply_state == 0);
1149                 ptlrpc_free_clone(req);
1150         }
1151
1152         EXIT;
1153 }
1154
1155 static void target_recovery_expired(unsigned long castmeharder)
1156 {
1157         struct obd_device *obd = (struct obd_device *)castmeharder;
1158         CERROR("%s: recovery timed out, aborting\n", obd->obd_name);
1159         spin_lock_bh(&obd->obd_processing_task_lock);
1160         if (obd->obd_recovering)
1161                 obd->obd_abort_recovery = 1;
1162         cfs_waitq_signal(&obd->obd_next_transno_waitq);
1163         spin_unlock_bh(&obd->obd_processing_task_lock);
1164 }
1165
1166
1167 /* obd_processing_task_lock should be held */
1168 void target_cancel_recovery_timer(struct obd_device *obd)
1169 {
1170         CDEBUG(D_HA, "%s: cancel recovery timer\n", obd->obd_name);
1171         cfs_timer_disarm(&obd->obd_recovery_timer);
1172 }
1173
1174 static void reset_recovery_timer(struct obd_device *obd)
1175 {
1176         spin_lock_bh(&obd->obd_processing_task_lock);
1177         if (!obd->obd_recovering) {
1178                 spin_unlock_bh(&obd->obd_processing_task_lock);
1179                 return;
1180         }
1181         cfs_timer_arm(&obd->obd_recovery_timer,
1182                       cfs_time_shift(OBD_RECOVERY_TIMEOUT));
1183         spin_unlock_bh(&obd->obd_processing_task_lock);
1184         CDEBUG(D_HA, "%s: timer will expire in %u seconds\n", obd->obd_name,
1185                OBD_RECOVERY_TIMEOUT);
1186         /* Only used for lprocfs_status */
1187         obd->obd_recovery_end = CURRENT_SECONDS + OBD_RECOVERY_TIMEOUT;
1188 }
1189
1190
1191 /* Only start it the first time called */
1192 void target_start_recovery_timer(struct obd_device *obd)
1193 {
1194         spin_lock_bh(&obd->obd_processing_task_lock);
1195         if (obd->obd_recovery_handler
1196             || timer_pending((struct timer_list *)&obd->obd_recovery_timer)) {
1197                 spin_unlock_bh(&obd->obd_processing_task_lock);
1198                 return;
1199         }
1200         CWARN("%s: starting recovery timer (%us)\n", obd->obd_name,
1201               OBD_RECOVERY_TIMEOUT);
1202         cfs_timer_init(&obd->obd_recovery_timer, target_recovery_expired, obd);
1203         spin_unlock_bh(&obd->obd_processing_task_lock);
1204
1205         reset_recovery_timer(obd);
1206 }
1207
1208 #ifdef __KERNEL__
1209 static int check_for_next_transno(struct obd_device *obd)
1210 {
1211         struct ptlrpc_request *req = NULL;
1212         int wake_up = 0, connected, completed, queue_len, max;
1213         __u64 next_transno, req_transno;
1214         ENTRY;
1215         spin_lock_bh(&obd->obd_processing_task_lock);
1216
1217         if (!list_empty(&obd->obd_req_replay_queue)) {
1218                 req = list_entry(obd->obd_req_replay_queue.next,
1219                                  struct ptlrpc_request, rq_list);
1220                 req_transno = lustre_msg_get_transno(req->rq_reqmsg);
1221         } else {
1222                 req_transno = 0;
1223         }
1224
1225         max = obd->obd_max_recoverable_clients;
1226         connected = obd->obd_connected_clients;
1227         completed = connected - obd->obd_recoverable_clients;
1228         queue_len = obd->obd_requests_queued_for_recovery;
1229         next_transno = obd->obd_next_recovery_transno;
1230
1231         CDEBUG(D_HA, "max: %d, connected: %d, completed: %d, queue_len: %d, "
1232                "req_transno: "LPU64", next_transno: "LPU64"\n",
1233                max, connected, completed, queue_len, req_transno, next_transno);
1234
1235         if (obd->obd_abort_recovery) {
1236                 CDEBUG(D_HA, "waking for aborted recovery\n");
1237                 wake_up = 1;
1238         } else if (atomic_read(&obd->obd_req_replay_clients) == 0) {
1239                 CDEBUG(D_HA, "waking for completed recovery\n");
1240                 wake_up = 1;
1241         } else if (req_transno == next_transno) {
1242                 CDEBUG(D_HA, "waking for next ("LPD64")\n", next_transno);
1243                 wake_up = 1;
1244         } else if (queue_len + completed == max) {
1245                 /* handle gaps occured due to lost reply. It is allowed gaps
1246                  * because all clients are connected and there will be resend
1247                  * for missed transaction */
1248                 LASSERTF(req_transno >= next_transno,
1249                          "req_transno: "LPU64", next_transno: "LPU64"\n",
1250                          req_transno, next_transno);
1251
1252                 CDEBUG(req_transno > obd->obd_last_committed ? D_ERROR : D_HA,
1253                        "waking for skipped transno (skip: "LPD64
1254                        ", ql: %d, comp: %d, conn: %d, next: "LPD64")\n",
1255                        next_transno, queue_len, completed, connected, req_transno);
1256                 obd->obd_next_recovery_transno = req_transno;
1257                 wake_up = 1;
1258         } else if (queue_len == atomic_read(&obd->obd_req_replay_clients)) {
1259                 /* some clients haven't connected in time, but we can try
1260                  * to replay requests that demand on already committed ones
1261                  * also, we can replay first non-committed transation */
1262                 LASSERT(req_transno != 0);
1263                 if (req_transno == obd->obd_last_committed + 1) {
1264                         obd->obd_next_recovery_transno = req_transno;
1265                 } else if (req_transno > obd->obd_last_committed) {
1266                         /* can't continue recovery: have no needed transno */
1267                         obd->obd_abort_recovery = 1;
1268                         CDEBUG(D_ERROR, "abort due to missed clients. max: %d, "
1269                                "connected: %d, completed: %d, queue_len: %d, "
1270                                "req_transno: "LPU64", next_transno: "LPU64"\n",
1271                                max, connected, completed, queue_len,
1272                                req_transno, next_transno);
1273                 }
1274                 wake_up = 1;
1275         }
1276
1277         spin_unlock_bh(&obd->obd_processing_task_lock);
1278         return wake_up;
1279 }
1280
1281 static struct ptlrpc_request *target_next_replay_req(struct obd_device *obd)
1282 {
1283         struct l_wait_info lwi = { 0 };
1284         struct ptlrpc_request *req;
1285
1286         CDEBUG(D_HA, "Waiting for transno "LPD64"\n",
1287                obd->obd_next_recovery_transno);
1288         l_wait_event(obd->obd_next_transno_waitq,
1289                      check_for_next_transno(obd), &lwi);
1290
1291         spin_lock_bh(&obd->obd_processing_task_lock);
1292         if (obd->obd_abort_recovery) {
1293                 req = NULL;
1294         } else if (!list_empty(&obd->obd_req_replay_queue)) {
1295                 req = list_entry(obd->obd_req_replay_queue.next,
1296                                  struct ptlrpc_request, rq_list);
1297                 list_del_init(&req->rq_list);
1298                 obd->obd_requests_queued_for_recovery--;
1299         } else {
1300                 req = NULL;
1301         }
1302         spin_unlock_bh(&obd->obd_processing_task_lock);
1303         RETURN(req);
1304 }
1305
1306 static int check_for_next_lock(struct obd_device *obd)
1307 {
1308         struct ptlrpc_request *req = NULL;
1309         int wake_up = 0;
1310
1311         spin_lock_bh(&obd->obd_processing_task_lock);
1312         if (!list_empty(&obd->obd_lock_replay_queue)) {
1313                 req = list_entry(obd->obd_lock_replay_queue.next,
1314                                  struct ptlrpc_request, rq_list);
1315                 CDEBUG(D_HA, "waking for next lock\n");
1316                 wake_up = 1;
1317         } else if (atomic_read(&obd->obd_lock_replay_clients) == 0) {
1318                 CDEBUG(D_HA, "waking for completed lock replay\n");
1319                 wake_up = 1;
1320         } else if (obd->obd_abort_recovery) {
1321                 CDEBUG(D_HA, "waking for aborted recovery\n");
1322                 wake_up = 1;
1323         }
1324         spin_unlock_bh(&obd->obd_processing_task_lock);
1325
1326         return wake_up;
1327 }
1328
1329 static struct ptlrpc_request *target_next_replay_lock(struct obd_device *obd)
1330 {
1331         struct l_wait_info lwi = { 0 };
1332         struct ptlrpc_request *req;
1333
1334         CDEBUG(D_HA, "Waiting for lock\n");
1335         l_wait_event(obd->obd_next_transno_waitq,
1336                      check_for_next_lock(obd), &lwi);
1337
1338         spin_lock_bh(&obd->obd_processing_task_lock);
1339         if (obd->obd_abort_recovery) {
1340                 req = NULL;
1341         } else if (!list_empty(&obd->obd_lock_replay_queue)) {
1342                 req = list_entry(obd->obd_lock_replay_queue.next,
1343                                  struct ptlrpc_request, rq_list);
1344                 list_del_init(&req->rq_list);
1345         } else {
1346                 req = NULL;
1347         }
1348         spin_unlock_bh(&obd->obd_processing_task_lock);
1349         return req;
1350 }
1351
1352 static struct ptlrpc_request *target_next_final_ping(struct obd_device *obd)
1353 {
1354         struct ptlrpc_request *req;
1355
1356         spin_lock_bh(&obd->obd_processing_task_lock);
1357         if (!list_empty(&obd->obd_final_req_queue)) {
1358                 req = list_entry(obd->obd_final_req_queue.next,
1359                                  struct ptlrpc_request, rq_list);
1360                 list_del_init(&req->rq_list);
1361         } else {
1362                 req = NULL;
1363         }
1364         spin_unlock_bh(&obd->obd_processing_task_lock);
1365         return req;
1366 }
1367
1368 static inline int req_replay_done(struct obd_export *exp)
1369 {
1370         return (exp->exp_req_replay_needed == 0);
1371 }
1372
1373 static inline int lock_replay_done(struct obd_export *exp)
1374 {
1375         return (exp->exp_lock_replay_needed == 0);
1376 }
1377
1378 static inline int connect_done(struct obd_export *exp)
1379 {
1380         return (exp->exp_in_recovery != 0);
1381 }
1382
1383 static int check_for_clients(struct obd_device *obd)
1384 {
1385         if (obd->obd_abort_recovery)
1386                 return 1;
1387         LASSERT(obd->obd_connected_clients <= obd->obd_max_recoverable_clients);
1388         if (obd->obd_no_conn == 0 &&
1389             obd->obd_connected_clients == obd->obd_max_recoverable_clients)
1390                 return 1;
1391         return 0;
1392 }
1393
1394 static int handle_recovery_req(struct ptlrpc_thread *thread,
1395                                struct ptlrpc_request *req,
1396                                svc_handler_t handler)
1397 {
1398         int rc;
1399         ENTRY;
1400
1401         rc = lu_context_init(&req->rq_session, LCT_SESSION);
1402         if (rc) {
1403                 CERROR("Failure to initialize session: %d\n", rc);
1404                 return rc;
1405         }
1406         req->rq_session.lc_thread = thread;
1407         lu_context_enter(&req->rq_session);
1408         req->rq_svc_thread = thread;
1409         req->rq_svc_thread->t_env->le_ses = &req->rq_session;
1410
1411         (void)handler(req);
1412
1413         lu_context_exit(&req->rq_session);
1414         lu_context_fini(&req->rq_session);
1415         /* don't reset timer for final stage */
1416         if (!req_replay_done(req->rq_export) ||
1417             !lock_replay_done(req->rq_export))
1418                 reset_recovery_timer(class_exp2obd(req->rq_export));
1419         ptlrpc_free_clone(req);
1420         RETURN(0);
1421 }
1422
1423 static int target_recovery_thread(void *arg)
1424 {
1425         struct obd_device *obd = arg;
1426         struct ptlrpc_request *req;
1427         struct target_recovery_data *trd = &obd->obd_recovery_data;
1428         struct l_wait_info lwi = { 0 };
1429         unsigned long delta;
1430         unsigned long flags;
1431         struct lu_env env;
1432         struct ptlrpc_thread fake_svc_thread, *thread = &fake_svc_thread;
1433         __u32 recov_ctx_tags = LCT_MD_THREAD;
1434         int rc = 0;
1435         ENTRY;
1436
1437         cfs_daemonize("tgt_recov");
1438
1439         SIGNAL_MASK_LOCK(current, flags);
1440         sigfillset(&current->blocked);
1441         RECALC_SIGPENDING;
1442         SIGNAL_MASK_UNLOCK(current, flags);
1443
1444         rc = lu_context_init(&env.le_ctx, recov_ctx_tags);
1445         if (rc)
1446                 RETURN(rc);
1447
1448         thread->t_env = &env;
1449         env.le_ctx.lc_thread = thread;
1450
1451         CERROR("%s: started recovery thread pid %d\n", obd->obd_name,
1452                current->pid);
1453         trd->trd_processing_task = current->pid;
1454
1455         obd->obd_recovering = 1;
1456         complete(&trd->trd_starting);
1457
1458         /* first of all, we have to know the first transno to replay */
1459         obd->obd_abort_recovery = 0;
1460         l_wait_event(obd->obd_next_transno_waitq,
1461                      check_for_clients(obd), &lwi);
1462
1463         spin_lock_bh(&obd->obd_processing_task_lock);
1464         target_cancel_recovery_timer(obd);
1465         spin_unlock_bh(&obd->obd_processing_task_lock);
1466
1467         /* If some clients haven't connected in time, evict them */
1468         if (obd->obd_abort_recovery) {
1469                 CWARN("Some clients haven't connect in time (%d/%d),"
1470                        "evict them\n", obd->obd_connected_clients,
1471                        obd->obd_max_recoverable_clients);
1472                 obd->obd_abort_recovery = obd->obd_stopping;
1473                 class_disconnect_stale_exports(obd, connect_done);
1474         }
1475         /* next stage: replay requests */
1476         delta = jiffies;
1477         obd->obd_req_replaying = 1;
1478         CDEBUG(D_INFO, "1: request replay stage - %d clients from t"LPU64"\n",
1479               atomic_read(&obd->obd_req_replay_clients),
1480               obd->obd_next_recovery_transno);
1481         while ((req = target_next_replay_req(obd))) {
1482                 LASSERT(trd->trd_processing_task == current->pid);
1483                 DEBUG_REQ(D_HA, req, "processing t"LPD64" from %s",
1484                           lustre_msg_get_transno(req->rq_reqmsg),
1485                           libcfs_nid2str(req->rq_peer.nid));
1486
1487                 handle_recovery_req(thread, req,
1488                                     trd->trd_recovery_handler);
1489                 obd->obd_replayed_requests++;
1490                 spin_lock_bh(&obd->obd_processing_task_lock);
1491                 obd->obd_next_recovery_transno++;
1492                 spin_unlock_bh(&obd->obd_processing_task_lock);
1493         }
1494
1495         spin_lock_bh(&obd->obd_processing_task_lock);
1496         target_cancel_recovery_timer(obd);
1497         spin_unlock_bh(&obd->obd_processing_task_lock);
1498
1499         /* If some clients haven't replayed requests in time, evict them */
1500         if (obd->obd_abort_recovery) {
1501                 CDEBUG(D_ERROR, "req replay timed out, aborting ...\n");
1502                 obd->obd_abort_recovery = obd->obd_stopping;
1503                 class_disconnect_stale_exports(obd, req_replay_done);
1504                 abort_req_replay_queue(obd);
1505         }
1506         /* The second stage: replay locks */
1507         CDEBUG(D_INFO, "2: lock replay stage - %d clients\n",
1508                atomic_read(&obd->obd_lock_replay_clients));
1509         while ((req = target_next_replay_lock(obd))) {
1510                 LASSERT(trd->trd_processing_task == current->pid);
1511                 DEBUG_REQ(D_HA|D_WARNING, req, "processing lock from %s: ",
1512                           libcfs_nid2str(req->rq_peer.nid));
1513                 handle_recovery_req(thread, req,
1514                                     trd->trd_recovery_handler);
1515                 obd->obd_replayed_locks++;
1516         }
1517
1518         spin_lock_bh(&obd->obd_processing_task_lock);
1519         target_cancel_recovery_timer(obd);
1520         spin_unlock_bh(&obd->obd_processing_task_lock);
1521         /* If some clients haven't replayed requests in time, evict them */
1522         if (obd->obd_abort_recovery) {
1523                 int stale;
1524                 CERROR("lock replay timed out, aborting ...\n");
1525                 obd->obd_abort_recovery = obd->obd_stopping;
1526                 stale = class_disconnect_stale_exports(obd, lock_replay_done);
1527                 abort_lock_replay_queue(obd);
1528         }
1529
1530         /* We drop recoverying flag to forward all new requests
1531          * to regular mds_handle() since now */
1532         spin_lock_bh(&obd->obd_processing_task_lock);
1533         obd->obd_recovering = obd->obd_abort_recovery = 0;
1534         spin_unlock_bh(&obd->obd_processing_task_lock);
1535         /* The third stage: reply on final pings */
1536         CDEBUG(D_INFO, "3: final stage - process recovery completion pings\n");
1537         while ((req = target_next_final_ping(obd))) {
1538                 LASSERT(trd->trd_processing_task == current->pid);
1539                 DEBUG_REQ(D_HA, req, "processing final ping from %s: ",
1540                           libcfs_nid2str(req->rq_peer.nid));
1541                 handle_recovery_req(thread, req,
1542                                     trd->trd_recovery_handler);
1543         }
1544
1545         delta = (jiffies - delta) / HZ;
1546         CDEBUG(D_INFO,"4: recovery completed in %lus - %d/%d reqs/locks\n",
1547               delta, obd->obd_replayed_requests, obd->obd_replayed_locks);
1548         LASSERT(atomic_read(&obd->obd_req_replay_clients) == 0);
1549         LASSERT(atomic_read(&obd->obd_lock_replay_clients) == 0);
1550         if (delta > obd_timeout * 2) {
1551                 CWARN("too long recovery - read logs\n");
1552                 libcfs_debug_dumplog();
1553         }
1554
1555         target_finish_recovery(obd);
1556
1557         lu_env_fini(&env);
1558         trd->trd_processing_task = 0;
1559         complete(&trd->trd_finishing);
1560         RETURN(rc);
1561 }
1562
1563 int target_start_recovery_thread(struct obd_device *obd, svc_handler_t handler)
1564 {
1565         int rc = 0;
1566         struct target_recovery_data *trd = &obd->obd_recovery_data;
1567
1568         memset(trd, 0, sizeof(*trd));
1569         init_completion(&trd->trd_starting);
1570         init_completion(&trd->trd_finishing);
1571         trd->trd_recovery_handler = handler;
1572
1573         if (kernel_thread(target_recovery_thread, obd, 0) > 0) {
1574                 wait_for_completion(&trd->trd_starting);
1575                 LASSERT(obd->obd_recovering != 0);
1576         } else
1577                 rc = -ECHILD;
1578
1579         return rc;
1580 }
1581
1582 void target_stop_recovery_thread(struct obd_device *obd)
1583 {
1584         spin_lock_bh(&obd->obd_processing_task_lock);
1585         if (obd->obd_recovery_data.trd_processing_task > 0) {
1586                 struct target_recovery_data *trd = &obd->obd_recovery_data;
1587                 CERROR("%s: Aborting recovery\n", obd->obd_name);
1588                 obd->obd_abort_recovery = 1;
1589                 wake_up(&obd->obd_next_transno_waitq);
1590                 spin_unlock_bh(&obd->obd_processing_task_lock);
1591                 wait_for_completion(&trd->trd_finishing);
1592         } else {
1593                 spin_unlock_bh(&obd->obd_processing_task_lock);
1594         }
1595 }
1596
1597 void target_recovery_fini(struct obd_device *obd)
1598 {
1599         class_disconnect_exports(obd);
1600         target_stop_recovery_thread(obd);
1601         target_cleanup_recovery(obd);
1602 }
1603 EXPORT_SYMBOL(target_recovery_fini);
1604
1605 void target_recovery_init(struct obd_device *obd, svc_handler_t handler)
1606 {
1607         if (obd->obd_max_recoverable_clients == 0)
1608                 return;
1609
1610         CWARN("RECOVERY: service %s, %d recoverable clients, "
1611               "last_transno "LPU64"\n", obd->obd_name,
1612               obd->obd_max_recoverable_clients, obd->obd_last_committed);
1613         obd->obd_next_recovery_transno = obd->obd_last_committed + 1;
1614         target_start_recovery_thread(obd, handler);
1615         obd->obd_recovery_start = CURRENT_SECONDS;
1616         /* Only used for lprocfs_status */
1617         obd->obd_recovery_end = obd->obd_recovery_start + OBD_RECOVERY_TIMEOUT;
1618 }
1619 EXPORT_SYMBOL(target_recovery_init);
1620
1621 #endif
1622
1623 int target_process_req_flags(struct obd_device *obd, struct ptlrpc_request *req)
1624 {
1625         struct obd_export *exp = req->rq_export;
1626         LASSERT(exp != NULL);
1627         if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REQ_REPLAY_DONE) {
1628                 /* client declares he's ready to replay locks */
1629                 spin_lock_bh(&obd->obd_processing_task_lock);
1630                 if (exp->exp_req_replay_needed) {
1631                         LASSERT(atomic_read(&obd->obd_req_replay_clients) > 0);
1632                         spin_lock(&exp->exp_lock);
1633                         exp->exp_req_replay_needed = 0;
1634                         spin_unlock(&exp->exp_lock);
1635                         atomic_dec(&obd->obd_req_replay_clients);
1636                         LASSERT(obd->obd_recoverable_clients > 0);
1637                         obd->obd_recoverable_clients--;
1638                         if (atomic_read(&obd->obd_req_replay_clients) == 0)
1639                                 CDEBUG(D_HA, "all clients have replayed reqs\n");
1640                         wake_up(&obd->obd_next_transno_waitq);
1641                 }
1642                 spin_unlock_bh(&obd->obd_processing_task_lock);
1643         }
1644         if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_LOCK_REPLAY_DONE) {
1645                 /* client declares he's ready to complete recovery
1646                  * so, we put the request on th final queue */
1647                 spin_lock_bh(&obd->obd_processing_task_lock);
1648                 if (exp->exp_lock_replay_needed) {
1649                         LASSERT(atomic_read(&obd->obd_lock_replay_clients) > 0);
1650                         spin_lock(&exp->exp_lock);
1651                         exp->exp_lock_replay_needed = 0;
1652                         spin_unlock(&exp->exp_lock);
1653                         atomic_dec(&obd->obd_lock_replay_clients);
1654                         if (atomic_read(&obd->obd_lock_replay_clients) == 0)
1655                                 CDEBUG(D_HA, "all clients have replayed locks\n");
1656                         wake_up(&obd->obd_next_transno_waitq);
1657                 }
1658                 spin_unlock_bh(&obd->obd_processing_task_lock);
1659         }
1660
1661         return 0;
1662 }
1663
1664 int target_queue_recovery_request(struct ptlrpc_request *req,
1665                                   struct obd_device *obd)
1666 {
1667         struct list_head *tmp;
1668         int inserted = 0;
1669         __u64 transno = lustre_msg_get_transno(req->rq_reqmsg);
1670
1671         ENTRY;
1672
1673         if (obd->obd_recovery_data.trd_processing_task == cfs_curproc_pid()) {
1674                 /* Processing the queue right now, don't re-add. */
1675                 RETURN(1);
1676         }
1677
1678         target_process_req_flags(obd, req);
1679
1680         if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_LOCK_REPLAY_DONE) {
1681                 /* client declares he's ready to complete recovery
1682                  * so, we put the request on th final queue */
1683                 req = ptlrpc_clone_req(req);
1684                 if (req == NULL)
1685                         RETURN(-ENOMEM);
1686                 DEBUG_REQ(D_HA, req, "queue final req");
1687                 spin_lock_bh(&obd->obd_processing_task_lock);
1688                 if (obd->obd_recovering)
1689                         list_add_tail(&req->rq_list, &obd->obd_final_req_queue);
1690                 else {
1691                         spin_unlock_bh(&obd->obd_processing_task_lock);
1692                         ptlrpc_free_clone(req);
1693                         if (obd->obd_stopping) {
1694                                 RETURN(-ENOTCONN);
1695                         } else {
1696                                 RETURN(1);
1697                         }
1698                 }
1699                 spin_unlock_bh(&obd->obd_processing_task_lock);
1700                 RETURN(0);
1701         }
1702         if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REQ_REPLAY_DONE) {
1703                 /* client declares he's ready to replay locks */
1704                 req = ptlrpc_clone_req(req);
1705                 if (req == NULL)
1706                         RETURN(-ENOMEM);
1707                 DEBUG_REQ(D_HA, req, "queue lock replay req");
1708                 spin_lock_bh(&obd->obd_processing_task_lock);
1709                 LASSERT(obd->obd_recovering);
1710                 /* usually due to recovery abort */
1711                 if (!req->rq_export->exp_in_recovery) {
1712                         spin_unlock_bh(&obd->obd_processing_task_lock);
1713                         ptlrpc_free_clone(req);
1714                         RETURN(-ENOTCONN);
1715                 }
1716                 LASSERT(req->rq_export->exp_lock_replay_needed);
1717                 list_add_tail(&req->rq_list, &obd->obd_lock_replay_queue);
1718                 spin_unlock_bh(&obd->obd_processing_task_lock);
1719                 wake_up(&obd->obd_next_transno_waitq);
1720                 RETURN(0);
1721         }
1722
1723         /* CAVEAT EMPTOR: The incoming request message has been swabbed
1724          * (i.e. buflens etc are in my own byte order), but type-dependent
1725          * buffers (eg mds_body, ost_body etc) have NOT been swabbed. */
1726
1727         if (!transno) {
1728                 CFS_INIT_LIST_HEAD(&req->rq_list);
1729                 DEBUG_REQ(D_HA, req, "not queueing");
1730                 RETURN(1);
1731         }
1732
1733         spin_lock_bh(&obd->obd_processing_task_lock);
1734
1735         /* If we're processing the queue, we want don't want to queue this
1736          * message.
1737          *
1738          * Also, if this request has a transno less than the one we're waiting
1739          * for, we should process it now.  It could (and currently always will)
1740          * be an open request for a descriptor that was opened some time ago.
1741          *
1742          * Also, a resent, replayed request that has already been
1743          * handled will pass through here and be processed immediately.
1744          */
1745         CWARN("Next recovery transno: "LPU64", current: "LPU64", replaying: %i\n",
1746               obd->obd_next_recovery_transno, transno, obd->obd_req_replaying);
1747         if (transno < obd->obd_next_recovery_transno && obd->obd_req_replaying) {
1748                 /* Processing the queue right now, don't re-add. */
1749                 LASSERT(list_empty(&req->rq_list));
1750                 spin_unlock_bh(&obd->obd_processing_task_lock);
1751                 RETURN(1);
1752         }
1753         spin_unlock_bh(&obd->obd_processing_task_lock);
1754
1755         /* A resent, replayed request that is still on the queue; just drop it.
1756            The queued request will handle this. */
1757         if ((lustre_msg_get_flags(req->rq_reqmsg) & (MSG_RESENT|MSG_REPLAY)) ==
1758             (MSG_RESENT | MSG_REPLAY)) {
1759                 DEBUG_REQ(D_ERROR, req, "dropping resent queued req");
1760                 RETURN(0);
1761         }
1762
1763         req = ptlrpc_clone_req(req);
1764         if (req == NULL)
1765                 RETURN(-ENOMEM);
1766
1767         spin_lock_bh(&obd->obd_processing_task_lock);
1768         LASSERT(obd->obd_recovering);
1769         if (!req->rq_export->exp_in_recovery) {
1770                 spin_unlock_bh(&obd->obd_processing_task_lock);
1771                 ptlrpc_free_clone(req);
1772                 RETURN(-ENOTCONN);
1773         }
1774         LASSERT(req->rq_export->exp_req_replay_needed);
1775
1776         /* XXX O(n^2) */
1777         list_for_each(tmp, &obd->obd_req_replay_queue) {
1778                 struct ptlrpc_request *reqiter =
1779                         list_entry(tmp, struct ptlrpc_request, rq_list);
1780
1781                 if (lustre_msg_get_transno(reqiter->rq_reqmsg) > transno) {
1782                         list_add_tail(&req->rq_list, &reqiter->rq_list);
1783                         inserted = 1;
1784                         break;
1785                 }
1786         }
1787
1788         if (!inserted)
1789                 list_add_tail(&req->rq_list, &obd->obd_req_replay_queue);
1790
1791         obd->obd_requests_queued_for_recovery++;
1792         wake_up(&obd->obd_next_transno_waitq);
1793         spin_unlock_bh(&obd->obd_processing_task_lock);
1794         RETURN(0);
1795
1796 }
1797
1798 struct obd_device * target_req2obd(struct ptlrpc_request *req)
1799 {
1800         return req->rq_export->exp_obd;
1801 }
1802
1803 static inline struct ldlm_pool *ldlm_exp2pl(struct obd_export *exp)
1804 {
1805         LASSERT(exp != NULL);
1806         return &exp->exp_obd->obd_namespace->ns_pool;
1807 }
1808
1809 int target_pack_pool_reply(struct ptlrpc_request *req)
1810 {
1811         struct ldlm_pool *pl;
1812         ENTRY;
1813
1814         if (!req->rq_export || !exp_connect_lru_resize(req->rq_export)) {
1815                 lustre_msg_set_slv(req->rq_repmsg, 0);
1816                 lustre_msg_set_limit(req->rq_repmsg, 0);
1817                 RETURN(0);
1818         }
1819
1820         pl = ldlm_exp2pl(req->rq_export);
1821
1822         spin_lock(&pl->pl_lock);
1823         LASSERT(ldlm_pool_get_slv(pl) != 0 && ldlm_pool_get_limit(pl) != 0);
1824         lustre_msg_set_slv(req->rq_repmsg, ldlm_pool_get_slv(pl));
1825         lustre_msg_set_limit(req->rq_repmsg, ldlm_pool_get_limit(pl));
1826         spin_unlock(&pl->pl_lock);
1827
1828         RETURN(0);
1829 }
1830
1831 int target_send_reply_msg(struct ptlrpc_request *req, int rc, int fail_id)
1832 {
1833         if (OBD_FAIL_CHECK_ORSET(fail_id & ~OBD_FAIL_ONCE, OBD_FAIL_ONCE)) {
1834                 DEBUG_REQ(D_ERROR, req, "dropping reply");
1835                 return (-ECOMM);
1836         }
1837
1838         if (unlikely(rc)) {
1839                 DEBUG_REQ(D_ERROR, req, "processing error (%d)", rc);
1840                 req->rq_status = rc;
1841                 return (ptlrpc_error(req));
1842         } else {
1843                 DEBUG_REQ(D_NET, req, "sending reply");
1844         }
1845
1846         return (ptlrpc_send_reply(req, 1));
1847 }
1848
1849 void target_send_reply(struct ptlrpc_request *req, int rc, int fail_id)
1850 {
1851         int                        netrc;
1852         struct ptlrpc_reply_state *rs;
1853         struct obd_device         *obd;
1854         struct obd_export         *exp;
1855         struct ptlrpc_service     *svc;
1856
1857         if (req->rq_no_reply)
1858                 return;
1859
1860         svc = req->rq_rqbd->rqbd_service;
1861         rs = req->rq_reply_state;
1862         if (rs == NULL || !rs->rs_difficult) {
1863                 /* no notifiers */
1864                 target_send_reply_msg (req, rc, fail_id);
1865                 return;
1866         }
1867
1868         /* must be an export if locks saved */
1869         LASSERT (req->rq_export != NULL);
1870         /* req/reply consistent */
1871         LASSERT (rs->rs_service == svc);
1872
1873         /* "fresh" reply */
1874         LASSERT (!rs->rs_scheduled);
1875         LASSERT (!rs->rs_scheduled_ever);
1876         LASSERT (!rs->rs_handled);
1877         LASSERT (!rs->rs_on_net);
1878         LASSERT (rs->rs_export == NULL);
1879         LASSERT (list_empty(&rs->rs_obd_list));
1880         LASSERT (list_empty(&rs->rs_exp_list));
1881
1882         exp = class_export_get (req->rq_export);
1883         obd = exp->exp_obd;
1884
1885         /* disable reply scheduling onto srv_reply_queue while I'm setting up */
1886         rs->rs_scheduled = 1;
1887         rs->rs_on_net    = 1;
1888         rs->rs_xid       = req->rq_xid;
1889         rs->rs_transno   = req->rq_transno;
1890         rs->rs_export    = exp;
1891
1892         spin_lock(&obd->obd_uncommitted_replies_lock);
1893
1894         if (rs->rs_transno > obd->obd_last_committed) {
1895                 /* not committed already */
1896                 list_add_tail (&rs->rs_obd_list,
1897                                &obd->obd_uncommitted_replies);
1898         }
1899
1900         spin_unlock (&obd->obd_uncommitted_replies_lock);
1901         spin_lock (&exp->exp_lock);
1902
1903         list_add_tail (&rs->rs_exp_list, &exp->exp_outstanding_replies);
1904
1905         spin_unlock(&exp->exp_lock);
1906
1907         netrc = target_send_reply_msg (req, rc, fail_id);
1908
1909         spin_lock(&svc->srv_lock);
1910
1911         svc->srv_n_difficult_replies++;
1912
1913         if (netrc != 0) {
1914                 /* error sending: reply is off the net.  Also we need +1
1915                  * reply ref until ptlrpc_server_handle_reply() is done
1916                  * with the reply state (if the send was successful, there
1917                  * would have been +1 ref for the net, which
1918                  * reply_out_callback leaves alone) */
1919                 rs->rs_on_net = 0;
1920                 ptlrpc_rs_addref(rs);
1921                 atomic_inc (&svc->srv_outstanding_replies);
1922         }
1923
1924         if (!rs->rs_on_net ||                   /* some notifier */
1925             list_empty(&rs->rs_exp_list) ||     /* completed already */
1926             list_empty(&rs->rs_obd_list)) {
1927                 list_add_tail (&rs->rs_list, &svc->srv_reply_queue);
1928                 cfs_waitq_signal (&svc->srv_waitq);
1929         } else {
1930                 list_add (&rs->rs_list, &svc->srv_active_replies);
1931                 rs->rs_scheduled = 0;           /* allow notifier to schedule */
1932         }
1933
1934         spin_unlock(&svc->srv_lock);
1935 }
1936
1937 int target_handle_ping(struct ptlrpc_request *req)
1938 {
1939         obd_ping(req->rq_export);
1940         return lustre_pack_reply(req, 1, NULL, NULL);
1941 }
1942
1943 void target_committed_to_req(struct ptlrpc_request *req)
1944 {
1945         struct obd_device *obd;
1946
1947         if (req == NULL || req->rq_export == NULL)
1948                 return;
1949
1950         obd = req->rq_export->exp_obd;
1951         if (obd == NULL)
1952                 return;
1953
1954         if (!obd->obd_no_transno && req->rq_repmsg != NULL)
1955                 lustre_msg_set_last_committed(req->rq_repmsg,
1956                                               obd->obd_last_committed);
1957         else
1958                 DEBUG_REQ(D_IOCTL, req, "not sending last_committed update");
1959
1960         CDEBUG(D_INFO, "last_committed "LPU64", transno "LPU64", xid "LPU64"\n",
1961                obd->obd_last_committed, req->rq_transno, req->rq_xid);
1962 }
1963
1964 EXPORT_SYMBOL(target_committed_to_req);
1965
1966 #ifdef HAVE_QUOTA_SUPPORT
1967 int target_handle_qc_callback(struct ptlrpc_request *req)
1968 {
1969         struct obd_quotactl *oqctl;
1970         struct client_obd *cli = &req->rq_export->exp_obd->u.cli;
1971
1972         oqctl = lustre_swab_reqbuf(req, REQ_REC_OFF, sizeof(*oqctl),
1973                                    lustre_swab_obd_quotactl);
1974         if (oqctl == NULL) {
1975                 CERROR("Can't unpack obd_quotactl\n");
1976                 RETURN(-EPROTO);
1977         }
1978
1979         cli->cl_qchk_stat = oqctl->qc_stat;
1980
1981         return 0;
1982 }
1983
1984 int target_handle_dqacq_callback(struct ptlrpc_request *req)
1985 {
1986 #ifdef __KERNEL__
1987         struct obd_device *obd = req->rq_export->exp_obd;
1988         struct obd_device *master_obd;
1989         struct lustre_quota_ctxt *qctxt;
1990         struct qunit_data *qdata;
1991         void* rep;
1992         struct qunit_data_old *qdata_old;
1993         int rc = 0;
1994         int repsize[2] = { sizeof(struct ptlrpc_body),
1995                            sizeof(struct qunit_data) };
1996         ENTRY;
1997
1998         rc = lustre_pack_reply(req, 2, repsize, NULL);
1999         if (rc)
2000                 RETURN(rc);
2001
2002         LASSERT(req->rq_export);
2003
2004         /* fixed for bug10707 */
2005         if ((req->rq_export->exp_connect_flags & OBD_CONNECT_QUOTA64) &&
2006             !OBD_FAIL_CHECK(OBD_FAIL_QUOTA_QD_COUNT_32BIT)) {
2007                 CDEBUG(D_QUOTA, "qd_count is 64bit!\n");
2008                 rep = lustre_msg_buf(req->rq_repmsg, REPLY_REC_OFF,
2009                                      sizeof(struct qunit_data));
2010                 LASSERT(rep);
2011                 qdata = lustre_swab_reqbuf(req, REQ_REC_OFF, sizeof(*qdata),
2012                                            lustre_swab_qdata);
2013         } else {
2014                 CDEBUG(D_QUOTA, "qd_count is 32bit!\n");
2015                 rep = lustre_msg_buf(req->rq_repmsg, REPLY_REC_OFF,
2016                                      sizeof(struct qunit_data_old));
2017                 LASSERT(rep);
2018                 qdata_old = lustre_swab_reqbuf(req, REQ_REC_OFF, sizeof(*qdata_old),
2019                                                lustre_swab_qdata_old);
2020                 qdata = lustre_quota_old_to_new(qdata_old);
2021         }
2022
2023         if (qdata == NULL) {
2024                 CERROR("Can't unpack qunit_data\n");
2025                 RETURN(-EPROTO);
2026         }
2027
2028         /* we use the observer */
2029         LASSERT(obd->obd_observer && obd->obd_observer->obd_observer);
2030         master_obd = obd->obd_observer->obd_observer;
2031         qctxt = &master_obd->u.obt.obt_qctxt;
2032
2033         LASSERT(qctxt->lqc_handler);
2034         rc = qctxt->lqc_handler(master_obd, qdata,
2035                                 lustre_msg_get_opc(req->rq_reqmsg));
2036         if (rc && rc != -EDQUOT)
2037                 CDEBUG(rc == -EBUSY  ? D_QUOTA : D_ERROR,
2038                        "dqacq failed! (rc:%d)\n", rc);
2039
2040         /* the qd_count might be changed in lqc_handler */
2041         if ((req->rq_export->exp_connect_flags & OBD_CONNECT_QUOTA64) &&
2042             !OBD_FAIL_CHECK(OBD_FAIL_QUOTA_QD_COUNT_32BIT)) {
2043                 memcpy(rep,qdata,sizeof(*qdata));
2044         } else {
2045                 qdata_old = lustre_quota_new_to_old(qdata);
2046                 memcpy(rep,qdata_old,sizeof(*qdata_old));
2047         }
2048         req->rq_status = rc;
2049         rc = ptlrpc_reply(req);
2050
2051         RETURN(rc);
2052 #else
2053         return 0;
2054 #endif /* !__KERNEL__ */
2055 }
2056 #endif /* HAVE_QUOTA_SUPPORT */
2057
2058 ldlm_mode_t lck_compat_array[] = {
2059         [LCK_EX] LCK_COMPAT_EX,
2060         [LCK_PW] LCK_COMPAT_PW,
2061         [LCK_PR] LCK_COMPAT_PR,
2062         [LCK_CW] LCK_COMPAT_CW,
2063         [LCK_CR] LCK_COMPAT_CR,
2064         [LCK_NL] LCK_COMPAT_NL,
2065         [LCK_GROUP] LCK_COMPAT_GROUP
2066 };