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 contexts be released. */
 184         class_import_get(imp);
 185         class_destroy_import(imp);
 186         sptlrpc_import_sec_put(imp);
 187         class_import_put(imp);
 188 }
 189
 190 /* configure an RPC client OBD device
 191  *
 192  * lcfg parameters:
 193  * 1 - client UUID
 194  * 2 - server UUID
 195  * 3 - inactive-on-startup
 196  */
 197 int client_obd_setup(struct obd_device *obddev, struct lustre_cfg *lcfg)
 198 {
 199         struct client_obd *cli = &obddev->u.cli;
 200         struct obd_import *imp;
 201         struct obd_uuid server_uuid;
 202         int rq_portal, rp_portal, connect_op;
 203         char *name = obddev->obd_type->typ_name;
 204         int rc;
 205         ENTRY;
 206
 207         /* In a more perfect world, we would hang a ptlrpc_client off of
 208          * obd_type and just use the values from there. */
 209         if (!strcmp(name, LUSTRE_OSC_NAME)) {
 210                 rq_portal = OST_REQUEST_PORTAL;
 211                 rp_portal = OSC_REPLY_PORTAL;
 212                 connect_op = OST_CONNECT;
 213         } else if (!strcmp(name, LUSTRE_MDC_NAME)) {
 214                 rq_portal = MDS_REQUEST_PORTAL;
 215                 rp_portal = MDC_REPLY_PORTAL;
 216                 connect_op = MDS_CONNECT;
 217         } else if (!strcmp(name, LUSTRE_MGC_NAME)) {
 218                 rq_portal = MGS_REQUEST_PORTAL;
 219                 rp_portal = MGC_REPLY_PORTAL;
 220                 connect_op = MGS_CONNECT;
 221         } else {
 222                 CERROR("unknown client OBD type \"%s\", can't setup\n",
 223                        name);
 224                 RETURN(-EINVAL);
 225         }
 226
 227         if (LUSTRE_CFG_BUFLEN(lcfg, 1) < 1) {
 228                 CERROR("requires a TARGET UUID\n");
 229                 RETURN(-EINVAL);
 230         }
 231
 232         if (LUSTRE_CFG_BUFLEN(lcfg, 1) > 37) {
 233                 CERROR("client UUID must be less than 38 characters\n");
 234                 RETURN(-EINVAL);
 235         }
 236
 237         if (LUSTRE_CFG_BUFLEN(lcfg, 2) < 1) {
 238                 CERROR("setup requires a SERVER UUID\n");
 239                 RETURN(-EINVAL);
 240         }
 241
 242         if (LUSTRE_CFG_BUFLEN(lcfg, 2) > 37) {
 243                 CERROR("target UUID must be less than 38 characters\n");
 244                 RETURN(-EINVAL);
 245         }
 246
 247         init_rwsem(&cli->cl_sem);
 248         sema_init(&cli->cl_mgc_sem, 1);
 249         sptlrpc_rule_set_init(&cli->cl_sptlrpc_rset);
 250         cli->cl_sec_part = LUSTRE_SP_ANY;
 251         cli->cl_conn_count = 0;
 252         memcpy(server_uuid.uuid, lustre_cfg_buf(lcfg, 2),
 253                min_t(unsigned int, LUSTRE_CFG_BUFLEN(lcfg, 2),
 254                      sizeof(server_uuid)));
 255
 256         cli->cl_dirty = 0;
 257         cli->cl_avail_grant = 0;
 258         /* FIXME: should limit this for the sum of all cl_dirty_max */
 259         cli->cl_dirty_max = OSC_MAX_DIRTY_DEFAULT * 1024 * 1024;
 260         if (cli->cl_dirty_max >> CFS_PAGE_SHIFT > num_physpages / 8)
 261                 cli->cl_dirty_max = num_physpages << (CFS_PAGE_SHIFT - 3);
 262         CFS_INIT_LIST_HEAD(&cli->cl_cache_waiters);
 263         CFS_INIT_LIST_HEAD(&cli->cl_loi_ready_list);
 264         CFS_INIT_LIST_HEAD(&cli->cl_loi_write_list);
 265         CFS_INIT_LIST_HEAD(&cli->cl_loi_read_list);
 266         client_obd_list_lock_init(&cli->cl_loi_list_lock);
 267         cli->cl_r_in_flight = 0;
 268         cli->cl_w_in_flight = 0;
 269
 270         spin_lock_init(&cli->cl_read_rpc_hist.oh_lock);
 271         spin_lock_init(&cli->cl_write_rpc_hist.oh_lock);
 272         spin_lock_init(&cli->cl_read_page_hist.oh_lock);
 273         spin_lock_init(&cli->cl_write_page_hist.oh_lock);
 274         spin_lock_init(&cli->cl_read_offset_hist.oh_lock);
 275         spin_lock_init(&cli->cl_write_offset_hist.oh_lock);
 276         cfs_waitq_init(&cli->cl_destroy_waitq);
 277         atomic_set(&cli->cl_destroy_in_flight, 0);
 278 #ifdef ENABLE_CHECKSUM
 279         /* Turn on checksumming by default. */
 280         cli->cl_checksum = 1;
 281         /*
 282          * The supported checksum types will be worked out at connect time
 283          * Set cl_chksum* to CRC32 for now to avoid returning screwed info
 284          * through procfs.
 285          */
 286         cli->cl_cksum_type = cli->cl_supp_cksum_types = OBD_CKSUM_CRC32;
 287 #endif
 288         atomic_set(&cli->cl_resends, OSC_DEFAULT_RESENDS);
 289
 290         /* This value may be changed at connect time in
 291            ptlrpc_connect_interpret. */
 292         cli->cl_max_pages_per_rpc = min((int)PTLRPC_MAX_BRW_PAGES,
 293                                         (int)(1024 * 1024 >> CFS_PAGE_SHIFT));
 294
 295         if (!strcmp(name, LUSTRE_MDC_NAME)) {
 296                 cli->cl_max_rpcs_in_flight = MDC_MAX_RIF_DEFAULT;
 297         } else if (num_physpages >> (20 - CFS_PAGE_SHIFT) <= 128 /* MB */) {
 298                 cli->cl_max_rpcs_in_flight = 2;
 299         } else if (num_physpages >> (20 - CFS_PAGE_SHIFT) <= 256 /* MB */) {
 300                 cli->cl_max_rpcs_in_flight = 3;
 301         } else if (num_physpages >> (20 - CFS_PAGE_SHIFT) <= 512 /* MB */) {
 302                 cli->cl_max_rpcs_in_flight = 4;
 303         } else {
 304                 cli->cl_max_rpcs_in_flight = OSC_MAX_RIF_DEFAULT;
 305         }
 306
 307         rc = ldlm_get_ref();
 308         if (rc) {
 309                 CERROR("ldlm_get_ref failed: %d\n", rc);
 310                 GOTO(err, rc);
 311         }
 312
 313         ptlrpc_init_client(rq_portal, rp_portal, name,
 314                            &obddev->obd_ldlm_client);
 315
 316         imp = class_new_import(obddev);
 317         if (imp == NULL)
 318                 GOTO(err_ldlm, rc = -ENOENT);
 319         imp->imp_client = &obddev->obd_ldlm_client;
 320         imp->imp_connect_op = connect_op;
 321         imp->imp_initial_recov = 1;
 322         imp->imp_initial_recov_bk = 0;
 323         CFS_INIT_LIST_HEAD(&imp->imp_pinger_chain);
 324         memcpy(cli->cl_target_uuid.uuid, lustre_cfg_buf(lcfg, 1),
 325                LUSTRE_CFG_BUFLEN(lcfg, 1));
 326         class_import_put(imp);
 327
 328         rc = client_import_add_conn(imp, &server_uuid, 1);
 329         if (rc) {
 330                 CERROR("can't add initial connection\n");
 331                 GOTO(err_import, rc);
 332         }
 333
 334         cli->cl_import = imp;
 335         /* cli->cl_max_mds_{easize,cookiesize} updated by mdc_init_ea_size() */
 336         cli->cl_max_mds_easize = sizeof(struct lov_mds_md);
 337         cli->cl_max_mds_cookiesize = sizeof(struct llog_cookie);
 338
 339         if (LUSTRE_CFG_BUFLEN(lcfg, 3) > 0) {
 340                 if (!strcmp(lustre_cfg_string(lcfg, 3), "inactive")) {
 341                         CDEBUG(D_HA, "marking %s %s->%s as inactive\n",
 342                                name, obddev->obd_name,
 343                                cli->cl_target_uuid.uuid);
 344                         spin_lock(&imp->imp_lock);
 345                         imp->imp_invalid = 1;
 346                         spin_unlock(&imp->imp_lock);
 347                 }
 348         }
 349
 350         cli->cl_qchk_stat = CL_NOT_QUOTACHECKED;
 351
 352         RETURN(rc);
 353
 354 err_import:
 355         class_destroy_import(imp);
 356 err_ldlm:
 357         ldlm_put_ref();
 358 err:
 359         RETURN(rc);
 360
 361 }
 362
 363 int client_obd_cleanup(struct obd_device *obddev)
 364 {
 365         ENTRY;
 366         sptlrpc_rule_set_free(&obddev->u.cli.cl_sptlrpc_rset);
 367         ldlm_put_ref();
 368         RETURN(0);
 369 }
 370
 371 /* ->o_connect() method for client side (OSC and MDC and MGC) */
 372 int client_connect_import(const struct lu_env *env,
 373                           struct lustre_handle *dlm_handle,
 374                           struct obd_device *obd, struct obd_uuid *cluuid,
 375                           struct obd_connect_data *data, void *localdata)
 376 {
 377         struct client_obd *cli = &obd->u.cli;
 378         struct obd_import *imp = cli->cl_import;
 379         struct obd_export *exp;
 380         struct obd_connect_data *ocd;
 381         struct ldlm_namespace *to_be_freed = NULL;
 382         int rc;
 383         ENTRY;
 384
 385         down_write(&cli->cl_sem);
 386         rc = class_connect(dlm_handle, obd, cluuid);
 387         if (rc)
 388                 GOTO(out_sem, rc);
 389
 390         cli->cl_conn_count++;
 391         if (cli->cl_conn_count > 1)
 392                 GOTO(out_sem, rc);
 393         exp = class_conn2export(dlm_handle);
 394
 395         if (obd->obd_namespace != NULL)
 396                 CERROR("already have namespace!\n");
 397         obd->obd_namespace = ldlm_namespace_new(obd, obd->obd_name,
 398                                                 LDLM_NAMESPACE_CLIENT,
 399                                                 LDLM_NAMESPACE_GREEDY);
 400         if (obd->obd_namespace == NULL)
 401                 GOTO(out_disco, rc = -ENOMEM);
 402
 403         imp->imp_dlm_handle = *dlm_handle;
 404         rc = ptlrpc_init_import(imp);
 405         if (rc != 0)
 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                 LASSERTF((ocd->ocd_connect_flags & data->ocd_connect_flags) ==
 423                          ocd->ocd_connect_flags, "old "LPX64", new "LPX64"\n",
 424                          data->ocd_connect_flags, ocd->ocd_connect_flags);
 425                 data->ocd_connect_flags = ocd->ocd_connect_flags;
 426         }
 427
 428         ptlrpc_pinger_add_import(imp);
 429
 430         EXIT;
 431
 432         if (rc) {
 433 out_ldlm:
 434                 ldlm_namespace_free_prior(obd->obd_namespace, imp, 0);
 435                 to_be_freed = obd->obd_namespace;
 436                 obd->obd_namespace = NULL;
 437 out_disco:
 438                 cli->cl_conn_count--;
 439                 class_disconnect(exp);
 440         } else {
 441                 class_export_put(exp);
 442         }
 443 out_sem:
 444         up_write(&cli->cl_sem);
 445         if (to_be_freed)
 446                 ldlm_namespace_free_post(to_be_freed);
 447
 448         return rc;
 449 }
 450
 451 int client_disconnect_export(struct obd_export *exp)
 452 {
 453         struct obd_device *obd = class_exp2obd(exp);
 454         struct client_obd *cli;
 455         struct obd_import *imp;
 456         int rc = 0, err;
 457         struct ldlm_namespace *to_be_freed = NULL;
 458         ENTRY;
 459
 460         if (!obd) {
 461                 CERROR("invalid export for disconnect: exp %p cookie "LPX64"\n",
 462                        exp, exp ? exp->exp_handle.h_cookie : -1);
 463                 RETURN(-EINVAL);
 464         }
 465
 466         cli = &obd->u.cli;
 467         imp = cli->cl_import;
 468
 469         down_write(&cli->cl_sem);
 470         if (!cli->cl_conn_count) {
 471                 CERROR("disconnecting disconnected device (%s)\n",
 472                        obd->obd_name);
 473                 GOTO(out_sem, rc = -EINVAL);
 474         }
 475
 476         cli->cl_conn_count--;
 477         if (cli->cl_conn_count)
 478                 GOTO(out_no_disconnect, rc = 0);
 479
 480         /* Mark import deactivated now, so we don't try to reconnect if any
 481          * of the cleanup RPCs fails (e.g. ldlm cancel, etc).  We don't
 482          * fully deactivate the import, or that would drop all requests. */
 483         spin_lock(&imp->imp_lock);
 484         imp->imp_deactive = 1;
 485         spin_unlock(&imp->imp_lock);
 486
 487         /* Some non-replayable imports (MDS's OSCs) are pinged, so just
 488          * delete it regardless.  (It's safe to delete an import that was
 489          * never added.) */
 490         (void)ptlrpc_pinger_del_import(imp);
 491
 492         if (obd->obd_namespace != NULL) {
 493                 /* obd_force == local only */
 494                 ldlm_cli_cancel_unused(obd->obd_namespace, NULL,
 495                                        obd->obd_force ? LDLM_FL_LOCAL_ONLY:0,
 496                                        NULL);
 497                 ldlm_namespace_free_prior(obd->obd_namespace, imp, obd->obd_force);
 498                 to_be_freed = obd->obd_namespace;
 499         }
 500
 501         rc = ptlrpc_disconnect_import(imp, 0);
 502
 503         ptlrpc_invalidate_import(imp);
 504         /* set obd_namespace to NULL only after invalidate, because we can have
 505          * some connect requests in flight, and his need store a connect flags
 506          * in obd_namespace. bug 14260 */
 507         obd->obd_namespace = NULL;
 508
 509         ptlrpc_free_rq_pool(imp->imp_rq_pool);
 510         destroy_import(imp);
 511         cli->cl_import = NULL;
 512
 513         EXIT;
 514  out_no_disconnect:
 515         err = class_disconnect(exp);
 516         if (!rc && err)
 517                 rc = err;
 518  out_sem:
 519         up_write(&cli->cl_sem);
 520         if (to_be_freed)
 521                 ldlm_namespace_free_post(to_be_freed);
 522
 523         RETURN(rc);
 524 }
 525
 526 /* --------------------------------------------------------------------------
 527  * from old lib/target.c
 528  * -------------------------------------------------------------------------- */
 529
 530 int target_handle_reconnect(struct lustre_handle *conn, struct obd_export *exp,
 531                             struct obd_uuid *cluuid, int initial_conn)
 532 {
 533         ENTRY;
 534         if (exp->exp_connection && exp->exp_imp_reverse && !initial_conn) {
 535                 struct lustre_handle *hdl;
 536                 hdl = &exp->exp_imp_reverse->imp_remote_handle;
 537                 /* Might be a re-connect after a partition. */
 538                 if (!memcmp(&conn->cookie, &hdl->cookie, sizeof conn->cookie)) {
 539                         CWARN("%s: %s reconnecting\n", exp->exp_obd->obd_name,
 540                               cluuid->uuid);
 541                         conn->cookie = exp->exp_handle.h_cookie;
 542                         /* target_handle_connect() treats EALREADY and
 543                          * -EALREADY differently.  EALREADY means we are
 544                          * doing a valid reconnect from the same client. */
 545                         RETURN(EALREADY);
 546                 } else {
 547                         CERROR("%s reconnecting from %s, "
 548                                "handle mismatch (ours "LPX64", theirs "
 549                                LPX64")\n", cluuid->uuid,
 550                                exp->exp_connection->c_remote_uuid.uuid,
 551                                hdl->cookie, conn->cookie);
 552                         memset(conn, 0, sizeof *conn);
 553                         /* target_handle_connect() treats EALREADY and
 554                          * -EALREADY differently.  -EALREADY is an error
 555                          * (same UUID, different handle). */
 556                         RETURN(-EALREADY);
 557                 }
 558         }
 559
 560         conn->cookie = exp->exp_handle.h_cookie;
 561         CDEBUG(D_HA, "connect export for UUID '%s' at %p, cookie "LPX64"\n",
 562                cluuid->uuid, exp, conn->cookie);
 563         RETURN(0);
 564 }
 565
 566 void target_client_add_cb(struct obd_device *obd, __u64 transno, void *cb_data,
 567                           int error)
 568 {
 569         struct obd_export *exp = cb_data;
 570
 571         CDEBUG(D_RPCTRACE, "%s: committing for initial connect of %s\n",
 572                obd->obd_name, exp->exp_client_uuid.uuid);
 573
 574         spin_lock(&exp->exp_lock);
 575         exp->exp_need_sync = 0;
 576         spin_unlock(&exp->exp_lock);
 577 }
 578 EXPORT_SYMBOL(target_client_add_cb);
 579
 580 int target_handle_connect(struct ptlrpc_request *req)
 581 {
 582         struct obd_device *target, *targref = NULL;
 583         struct obd_export *export = NULL;
 584         struct obd_import *revimp;
 585         struct lustre_handle conn;
 586         struct lustre_handle *tmp;
 587         struct obd_uuid tgtuuid;
 588         struct obd_uuid cluuid;
 589         struct obd_uuid remote_uuid;
 590         char *str;
 591         int rc = 0;
 592         int initial_conn = 0;
 593         struct obd_connect_data *data, *tmpdata;
 594         lnet_nid_t *client_nid = NULL;
 595         ENTRY;
 596
 597         OBD_RACE(OBD_FAIL_TGT_CONN_RACE);
 598
 599         str = req_capsule_client_get(&req->rq_pill, &RMF_TGTUUID);
 600         if (str == NULL) {
 601                 DEBUG_REQ(D_ERROR, req, "bad target UUID for connect");
 602                 GOTO(out, rc = -EINVAL);
 603         }
 604
 605         obd_str2uuid(&tgtuuid, str);
 606         target = class_uuid2obd(&tgtuuid);
 607         if (!target)
 608                 target = class_name2obd(str);
 609
 610         if (!target || target->obd_stopping || !target->obd_set_up) {
 611                 LCONSOLE_ERROR_MSG(0x137, "UUID '%s' is not available "
 612                                    " for connect (%s)\n", str,
 613                                    !target ? "no target" :
 614                                    (target->obd_stopping ? "stopping" :
 615                                    "not set up"));
 616                 GOTO(out, rc = -ENODEV);
 617         }
 618
 619         if (target->obd_no_conn) {
 620                 LCONSOLE_WARN("%s: temporarily refusing client connection "
 621                               "from %s\n", target->obd_name,
 622                               libcfs_nid2str(req->rq_peer.nid));
 623                 GOTO(out, rc = -EAGAIN);
 624         }
 625
 626         /* Make sure the target isn't cleaned up while we're here. Yes,
 627            there's still a race between the above check and our incref here.
 628            Really, class_uuid2obd should take the ref. */
 629         targref = class_incref(target);
 630
 631
 632         str = req_capsule_client_get(&req->rq_pill, &RMF_CLUUID);
 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 = req_capsule_client_get(&req->rq_pill, &RMF_CONN);
 656         if (tmp == NULL)
 657                 GOTO(out, rc = -EPROTO);
 658
 659         conn = *tmp;
 660
 661         data = req_capsule_client_get(&req->rq_pill, &RMF_CONNECT_DATA);
 662         if (!data)
 663                 GOTO(out, rc = -EPROTO);
 664
 665         rc = req_capsule_server_pack(&req->rq_pill);
 666         if (rc)
 667                 GOTO(out, rc);
 668
 669         if (lustre_msg_get_op_flags(req->rq_reqmsg) & MSG_CONNECT_LIBCLIENT) {
 670                 if (!data) {
 671                         DEBUG_REQ(D_WARNING, req, "Refusing old (unversioned) "
 672                                   "libclient connection attempt");
 673                         GOTO(out, rc = -EPROTO);
 674                 } else if (data->ocd_version < LUSTRE_VERSION_CODE -
 675                                                LUSTRE_VERSION_ALLOWED_OFFSET ||
 676                            data->ocd_version > LUSTRE_VERSION_CODE +
 677                                                LUSTRE_VERSION_ALLOWED_OFFSET) {
 678                         DEBUG_REQ(D_WARNING, req, "Refusing %s (%d.%d.%d.%d) "
 679                                   "libclient connection attempt",
 680                                   data->ocd_version < LUSTRE_VERSION_CODE ?
 681                                   "old" : "new",
 682                                   OBD_OCD_VERSION_MAJOR(data->ocd_version),
 683                                   OBD_OCD_VERSION_MINOR(data->ocd_version),
 684                                   OBD_OCD_VERSION_PATCH(data->ocd_version),
 685                                   OBD_OCD_VERSION_FIX(data->ocd_version));
 686                         data = req_capsule_server_sized_get(&req->rq_pill,
 687                                                         &RMF_CONNECT_DATA,
 688                                     offsetof(typeof(*data), ocd_version) +
 689                                              sizeof(data->ocd_version));
 690                         if (data) {
 691                                 data->ocd_connect_flags = OBD_CONNECT_VERSION;
 692                                 data->ocd_version = LUSTRE_VERSION_CODE;
 693                         }
 694                         GOTO(out, rc = -EPROTO);
 695                 }
 696         }
 697
 698         if (lustre_msg_get_op_flags(req->rq_reqmsg) & MSG_CONNECT_INITIAL)
 699                 initial_conn = 1;
 700
 701         /* lctl gets a backstage, all-access pass. */
 702         if (obd_uuid_equals(&cluuid, &target->obd_uuid))
 703                 goto dont_check_exports;
 704
 705         spin_lock(&target->obd_dev_lock);
 706         export = lustre_hash_get_object_by_key(target->obd_uuid_hash_body, &cluuid);
 707
 708         if (export != NULL && export->exp_connecting) { /* bug 9635, et. al. */
 709                 CWARN("%s: exp %p already connecting\n",
 710                       export->exp_obd->obd_name, export);
 711                 class_export_put(export);
 712                 export = NULL;
 713                 rc = -EALREADY;
 714         } else if (export != NULL && export->exp_connection != NULL &&
 715                    req->rq_peer.nid != export->exp_connection->c_peer.nid) {
 716                 /* make darn sure this is coming from the same peer
 717                  * if the UUIDs matched */
 718                   CWARN("%s: cookie %s seen on new NID %s when "
 719                           "existing NID %s is already connected\n",
 720                         target->obd_name, cluuid.uuid,
 721                   libcfs_nid2str(req->rq_peer.nid),
 722                   libcfs_nid2str(export->exp_connection->c_peer.nid));
 723                   class_export_put(export);
 724                   export = NULL;
 725                   rc = -EALREADY;
 726         } else if (export != NULL) {
 727                 spin_lock(&export->exp_lock);
 728                 export->exp_connecting = 1;
 729                 spin_unlock(&export->exp_lock);
 730                 class_export_put(export);
 731                 spin_unlock(&target->obd_dev_lock);
 732                 LASSERT(export->exp_obd == target);
 733
 734                 rc = target_handle_reconnect(&conn, export, &cluuid, initial_conn);
 735         }
 736
 737         /* If we found an export, we already unlocked. */
 738         if (!export) {
 739                 spin_unlock(&target->obd_dev_lock);
 740                 OBD_FAIL_TIMEOUT(OBD_FAIL_TGT_DELAY_CONNECT, 2 * obd_timeout);
 741         } else if (req->rq_export == NULL &&
 742                    atomic_read(&export->exp_rpc_count) > 0) {
 743                 CWARN("%s: refuse connection from %s/%s to 0x%p/%d\n",
 744                       target->obd_name, cluuid.uuid,
 745                       libcfs_nid2str(req->rq_peer.nid),
 746                       export, atomic_read(&export->exp_refcount));
 747                 GOTO(out, rc = -EBUSY);
 748         } else if (req->rq_export != NULL &&
 749                    (atomic_read(&export->exp_rpc_count) > 1)) {
 750                 CWARN("%s: refuse reconnection from %s@%s to 0x%p/%d\n",
 751                       target->obd_name, cluuid.uuid,
 752                       libcfs_nid2str(req->rq_peer.nid),
 753                       export, atomic_read(&export->exp_rpc_count));
 754                 GOTO(out, rc = -EBUSY);
 755         } else if (lustre_msg_get_conn_cnt(req->rq_reqmsg) == 1 &&
 756                    !initial_conn) {
 757                 CERROR("%s: NID %s (%s) reconnected with 1 conn_cnt; "
 758                        "cookies not random?\n", target->obd_name,
 759                        libcfs_nid2str(req->rq_peer.nid), cluuid.uuid);
 760                 GOTO(out, rc = -EALREADY);
 761         } else {
 762                 OBD_FAIL_TIMEOUT(OBD_FAIL_TGT_DELAY_RECONNECT, 2 * obd_timeout);
 763                 if (req->rq_export == NULL && initial_conn)
 764                        export->exp_last_request_time =
 765                                max(export->exp_last_request_time,
 766                                    (time_t)cfs_time_current_sec());
 767         }
 768
 769         /* We want to handle EALREADY but *not* -EALREADY from
 770          * target_handle_reconnect(), return reconnection state in a flag */
 771         if (rc == EALREADY) {
 772                 lustre_msg_add_op_flags(req->rq_repmsg, MSG_CONNECT_RECONNECT);
 773                 rc = 0;
 774         } else if (rc) {
 775                 GOTO(out, rc);
 776         }
 777         /* Tell the client if we're in recovery. */
 778         /* If this is the first client, start the recovery timer */
 779         CWARN("%s: connection from %s@%s %st"LPU64" exp %p cur %ld last %ld\n",
 780                target->obd_name, cluuid.uuid, libcfs_nid2str(req->rq_peer.nid),
 781               target->obd_recovering ? "recovering/" : "", data->ocd_transno,
 782               export, (long)cfs_time_current_sec(),
 783               export ? (long)export->exp_last_request_time : 0);
 784
 785
 786         if (target->obd_recovering) {
 787                 lustre_msg_add_op_flags(req->rq_repmsg, MSG_CONNECT_RECOVERING);
 788                 target_start_recovery_timer(target);
 789         }
 790
 791         /* Tell the client if we support replayable requests */
 792         if (target->obd_replayable)
 793                 lustre_msg_add_op_flags(req->rq_repmsg, MSG_CONNECT_REPLAYABLE);
 794         client_nid = &req->rq_peer.nid;
 795
 796         if (export == NULL) {
 797                 if (target->obd_recovering) {
 798                         cfs_time_t t;
 799
 800                         t = cfs_timer_deadline(&target->obd_recovery_timer);
 801                         t = cfs_time_sub(t, cfs_time_current());
 802                         CERROR("%s: denying connection for new client %s (%s): "
 803                                "%d clients in recovery for "CFS_TIME_T"s\n",
 804                                target->obd_name,
 805                                libcfs_nid2str(req->rq_peer.nid), cluuid.uuid,
 806                                target->obd_recoverable_clients,
 807                                cfs_duration_sec(t));
 808                         rc = -EBUSY;
 809                 } else {
 810 dont_check_exports:
 811                         rc = obd_connect(req->rq_svc_thread->t_env,
 812                                          &conn, target, &cluuid, data,
 813                                          client_nid);
 814                 }
 815         } else {
 816                 rc = obd_reconnect(req->rq_svc_thread->t_env,
 817                                    export, target, &cluuid, data);
 818         }
 819         if (rc)
 820                 GOTO(out, rc);
 821         /* Return only the parts of obd_connect_data that we understand, so the
 822          * client knows that we don't understand the rest. */
 823         if (data) {
 824                  tmpdata = req_capsule_server_get(&req->rq_pill,
 825                                                   &RMF_CONNECT_DATA);
 826                   //data->ocd_connect_flags &= OBD_CONNECT_SUPPORTED;
 827                  *tmpdata = *data;
 828         }
 829
 830         /* If all else goes well, this is our RPC return code. */
 831         req->rq_status = 0;
 832
 833         lustre_msg_set_handle(req->rq_repmsg, &conn);
 834
 835         /* ownership of this export ref transfers to the request AFTER we
 836          * drop any previous reference the request had, but we don't want
 837          * that to go to zero before we get our new export reference. */
 838         export = class_conn2export(&conn);
 839         if (!export) {
 840                 DEBUG_REQ(D_ERROR, req, "Missing export!");
 841                 GOTO(out, rc = -ENODEV);
 842         }
 843
 844         /* If the client and the server are the same node, we will already
 845          * have an export that really points to the client's DLM export,
 846          * because we have a shared handles table.
 847          *
 848          * XXX this will go away when shaver stops sending the "connect" handle
 849          * in the real "remote handle" field of the request --phik 24 Apr 2003
 850          */
 851         if (req->rq_export != NULL)
 852                 class_export_put(req->rq_export);
 853
 854         req->rq_export = export;
 855
 856         spin_lock(&export->exp_lock);
 857         if (initial_conn) {
 858                 lustre_msg_set_conn_cnt(req->rq_repmsg, export->exp_conn_cnt + 1);
 859         } else if (export->exp_conn_cnt >= lustre_msg_get_conn_cnt(req->rq_reqmsg)) {
 860                 spin_unlock(&export->exp_lock);
 861                 CERROR("%s: %s already connected at higher conn_cnt: %d > %d\n",
 862                        cluuid.uuid, libcfs_nid2str(req->rq_peer.nid),
 863                        export->exp_conn_cnt,
 864                        lustre_msg_get_conn_cnt(req->rq_reqmsg));
 865
 866                 GOTO(out, rc = -EALREADY);
 867         }
 868         export->exp_conn_cnt = lustre_msg_get_conn_cnt(req->rq_reqmsg);
 869
 870         /* request from liblustre?  Don't evict it for not pinging. */
 871         if (lustre_msg_get_op_flags(req->rq_reqmsg) & MSG_CONNECT_LIBCLIENT) {
 872                 export->exp_libclient = 1;
 873                 spin_unlock(&export->exp_lock);
 874
 875                 spin_lock(&target->obd_dev_lock);
 876                 list_del_init(&export->exp_obd_chain_timed);
 877                 spin_unlock(&target->obd_dev_lock);
 878         } else {
 879                 spin_unlock(&export->exp_lock);
 880         }
 881
 882         if (export->exp_connection != NULL)
 883                 ptlrpc_put_connection(export->exp_connection);
 884         export->exp_connection = ptlrpc_get_connection(req->rq_peer,
 885                                                        req->rq_self,
 886                                                        &remote_uuid);
 887
 888         spin_lock(&target->obd_dev_lock);
 889         /* Export might be hashed already, e.g. if this is reconnect */
 890         if (hlist_unhashed(&export->exp_nid_hash))
 891                 lustre_hash_additem(export->exp_obd->obd_nid_hash_body,
 892                                     &export->exp_connection->c_peer.nid,
 893                                     &export->exp_nid_hash);
 894         spin_unlock(&target->obd_dev_lock);
 895
 896         spin_lock_bh(&target->obd_processing_task_lock);
 897         if (target->obd_recovering && !export->exp_in_recovery) {
 898                 spin_lock(&export->exp_lock);
 899                 export->exp_in_recovery = 1;
 900                 export->exp_req_replay_needed = 1;
 901                 export->exp_lock_replay_needed = 1;
 902                 spin_unlock(&export->exp_lock);
 903                 if ((lustre_msg_get_op_flags(req->rq_reqmsg) & MSG_CONNECT_TRANSNO)
 904                      && (data->ocd_transno == 0))
 905                         CWARN("Connect with zero transno!\n");
 906
 907                 if ((lustre_msg_get_op_flags(req->rq_reqmsg) & MSG_CONNECT_TRANSNO)
 908                      && data->ocd_transno < target->obd_next_recovery_transno)
 909                         target->obd_next_recovery_transno = data->ocd_transno;
 910                 target->obd_connected_clients++;
 911                 /* each connected client is counted as recoverable */
 912                 target->obd_recoverable_clients++;
 913                 atomic_inc(&target->obd_req_replay_clients);
 914                 atomic_inc(&target->obd_lock_replay_clients);
 915                 if (target->obd_connected_clients ==
 916                     target->obd_max_recoverable_clients)
 917                         wake_up(&target->obd_next_transno_waitq);
 918         }
 919         spin_unlock_bh(&target->obd_processing_task_lock);
 920         tmp = req_capsule_client_get(&req->rq_pill, &RMF_CONN);
 921         conn = *tmp;
 922
 923         if (export->exp_imp_reverse != NULL) {
 924                 /* destroyed import can be still referenced in ctxt */
 925                 obd_set_info_async(export, sizeof(KEY_REVIMP_UPD),
 926                                    KEY_REVIMP_UPD, 0, NULL, NULL);
 927
 928                 /* in some recovery senarios, previous ctx init rpc handled
 929                  * in sptlrpc_target_export_check() might be used to install
 930                  * a reverse ctx in this reverse import, and later OBD_CONNECT
 931                  * using the same gss ctx could reach here and following new
 932                  * reverse import. note all reverse ctx in new/old import are
 933                  * actually based on the same gss ctx. so we invalidate ctx
 934                  * here before destroy import, otherwise flush old import will
 935                  * lead to remote reverse ctx be destroied, thus the reverse
 936                  * ctx of new import will lost its peer.
 937                  * there might be a better way to deal with this???
 938                  */
 939                 sptlrpc_import_inval_all_ctx(export->exp_imp_reverse);
 940
 941                 destroy_import(export->exp_imp_reverse);
 942         }
 943
 944         /* for the rest part, we return -ENOTCONN in case of errors
 945          * in order to let client initialize connection again.
 946          */
 947         revimp = export->exp_imp_reverse = class_new_import(target);
 948         if (!revimp) {
 949                 CERROR("fail to alloc new reverse import.\n");
 950                 GOTO(out, rc = -ENOTCONN);
 951         }
 952
 953         revimp->imp_connection = ptlrpc_connection_addref(export->exp_connection);
 954         revimp->imp_client = &export->exp_obd->obd_ldlm_client;
 955         revimp->imp_remote_handle = conn;
 956         revimp->imp_dlm_fake = 1;
 957         revimp->imp_state = LUSTRE_IMP_FULL;
 958         revimp->imp_msg_magic = req->rq_reqmsg->lm_magic;
 959
 960         rc = sptlrpc_import_sec_adapt(revimp, req->rq_svc_ctx,
 961                                       req->rq_flvr.sf_rpc);
 962         if (rc) {
 963                 CERROR("Failed to get sec for reverse import: %d\n", rc);
 964                 export->exp_imp_reverse = NULL;
 965                 class_destroy_import(revimp);
 966         }
 967
 968         class_import_put(revimp);
 969 out:
 970         if (export) {
 971                 spin_lock(&export->exp_lock);
 972                 export->exp_connecting = 0;
 973                 spin_unlock(&export->exp_lock);
 974         }
 975         if (targref)
 976                 class_decref(targref);
 977         if (rc)
 978                 req->rq_status = rc;
 979         RETURN(rc);
 980 }
 981
 982 int target_handle_disconnect(struct ptlrpc_request *req)
 983 {
 984         int rc;
 985         ENTRY;
 986
 987         rc = req_capsule_server_pack(&req->rq_pill);
 988         if (rc)
 989                 RETURN(rc);
 990
 991         /* keep the rq_export around so we can send the reply */
 992         req->rq_status = obd_disconnect(class_export_get(req->rq_export));
 993
 994         RETURN(0);
 995 }
 996
 997 void target_destroy_export(struct obd_export *exp)
 998 {
 999         /* exports created from last_rcvd data, and "fake"
1000            exports created by lctl don't have an import */
1001         if (exp->exp_imp_reverse != NULL)
1002                 destroy_import(exp->exp_imp_reverse);
1003
1004         /* We cancel locks at disconnect time, but this will catch any locks
1005          * granted in a race with recovery-induced disconnect. */
1006         if (exp->exp_obd->obd_namespace != NULL)
1007                 ldlm_cancel_locks_for_export(exp);
1008 }
1009
1010 /*
1011  * Recovery functions
1012  */
1013
1014 struct ptlrpc_request *ptlrpc_clone_req( struct ptlrpc_request *orig_req)
1015 {
1016         struct ptlrpc_request *copy_req;
1017         struct lustre_msg *copy_reqmsg;
1018         struct ptlrpc_user_desc *udesc = NULL;
1019
1020         OBD_ALLOC_PTR(copy_req);
1021         if (!copy_req)
1022                 return NULL;
1023         OBD_ALLOC(copy_reqmsg, orig_req->rq_reqlen);
1024         if (!copy_reqmsg){
1025                 OBD_FREE_PTR(copy_req);
1026                 return NULL;
1027         }
1028
1029         if (orig_req->rq_user_desc) {
1030                 int ngroups = orig_req->rq_user_desc->pud_ngroups;
1031
1032                 OBD_ALLOC(udesc, sptlrpc_user_desc_size(ngroups));
1033                 if (!udesc) {
1034                         OBD_FREE(copy_reqmsg, orig_req->rq_reqlen);
1035                         OBD_FREE_PTR(copy_req);
1036                         return NULL;
1037                 }
1038                 memcpy(udesc, orig_req->rq_user_desc,
1039                        sptlrpc_user_desc_size(ngroups));
1040         }
1041
1042         *copy_req = *orig_req;
1043         memcpy(copy_reqmsg, orig_req->rq_reqmsg, orig_req->rq_reqlen);
1044         copy_req->rq_reqmsg = copy_reqmsg;
1045         copy_req->rq_user_desc = udesc;
1046
1047         class_export_get(copy_req->rq_export);
1048         CFS_INIT_LIST_HEAD(&copy_req->rq_list);
1049         sptlrpc_svc_ctx_addref(copy_req);
1050
1051         if (copy_req->rq_reply_state) {
1052                 /* the copied req takes over the reply state */
1053                 orig_req->rq_reply_state = NULL;
1054                 /* to catch further access */
1055                 orig_req->rq_repmsg = NULL;
1056                 orig_req->rq_replen = 0;
1057         }
1058
1059         return copy_req;
1060 }
1061
1062 void ptlrpc_free_clone( struct ptlrpc_request *req)
1063 {
1064         if (req->rq_reply_state) {
1065                 ptlrpc_rs_decref(req->rq_reply_state);
1066                 req->rq_reply_state = NULL;
1067         }
1068
1069         sptlrpc_svc_ctx_decref(req);
1070         class_export_put(req->rq_export);
1071         list_del(&req->rq_list);
1072
1073         if (req->rq_user_desc) {
1074                 int ngroups = req->rq_user_desc->pud_ngroups;
1075                 OBD_FREE(req->rq_user_desc, sptlrpc_user_desc_size(ngroups));
1076         }
1077         OBD_FREE(req->rq_reqmsg, req->rq_reqlen);
1078         OBD_FREE_PTR(req);
1079 }
1080
1081 #ifdef __KERNEL__
1082 static void target_finish_recovery(struct obd_device *obd)
1083 {
1084         ENTRY;
1085         LCONSOLE_INFO("%s: sending delayed replies to recovered clients\n",
1086                       obd->obd_name);
1087
1088         ldlm_reprocess_all_ns(obd->obd_namespace);
1089
1090         /* when recovery finished, cleanup orphans on mds and ost */
1091         if (OBT(obd) && OBP(obd, postrecov)) {
1092                 int rc = OBP(obd, postrecov)(obd);
1093                 LCONSOLE_WARN("%s: recovery %s: rc %d\n", obd->obd_name,
1094                               rc < 0 ? "failed" : "complete", rc);
1095         }
1096
1097         obd->obd_recovery_end = cfs_time_current_sec();
1098         EXIT;
1099 }
1100
1101 static void abort_req_replay_queue(struct obd_device *obd)
1102 {
1103         struct ptlrpc_request *req, *n;
1104
1105         list_for_each_entry_safe(req, n, &obd->obd_req_replay_queue, rq_list) {
1106                 DEBUG_REQ(D_WARNING, req, "aborted:");
1107                 req->rq_status = -ENOTCONN;
1108                 if (ptlrpc_error(req)) {
1109                         DEBUG_REQ(D_ERROR, req,
1110                                   "failed abort_req_reply; skipping");
1111                 }
1112                 ptlrpc_free_clone(req);
1113         }
1114 }
1115
1116 static void abort_lock_replay_queue(struct obd_device *obd)
1117 {
1118         struct ptlrpc_request *req, *n;
1119
1120         list_for_each_entry_safe(req, n, &obd->obd_lock_replay_queue, rq_list){
1121                 DEBUG_REQ(D_ERROR, req, "aborted:");
1122                 req->rq_status = -ENOTCONN;
1123                 if (ptlrpc_error(req)) {
1124                         DEBUG_REQ(D_ERROR, req,
1125                                   "failed abort_lock_reply; skipping");
1126                 }
1127                 ptlrpc_free_clone(req);
1128         }
1129 }
1130 #endif
1131
1132 /* Called from a cleanup function if the device is being cleaned up
1133    forcefully.  The exports should all have been disconnected already,
1134    the only thing left to do is
1135      - clear the recovery flags
1136      - cancel the timer
1137      - free queued requests and replies, but don't send replies
1138    Because the obd_stopping flag is set, no new requests should be received.
1139
1140 */
1141 void target_cleanup_recovery(struct obd_device *obd)
1142 {
1143         struct ptlrpc_request *req, *n;
1144         ENTRY;
1145
1146         LASSERT(obd->obd_stopping);
1147
1148         spin_lock_bh(&obd->obd_processing_task_lock);
1149         if (!obd->obd_recovering) {
1150                 spin_unlock_bh(&obd->obd_processing_task_lock);
1151                 EXIT;
1152                 return;
1153         }
1154         obd->obd_recovering = obd->obd_abort_recovery = 0;
1155         target_cancel_recovery_timer(obd);
1156         spin_unlock_bh(&obd->obd_processing_task_lock);
1157
1158         list_for_each_entry_safe(req, n, &obd->obd_req_replay_queue, rq_list) {
1159                 LASSERT (req->rq_reply_state == 0);
1160                 ptlrpc_free_clone(req);
1161         }
1162         list_for_each_entry_safe(req, n, &obd->obd_lock_replay_queue, rq_list){
1163                 LASSERT (req->rq_reply_state == 0);
1164                 ptlrpc_free_clone(req);
1165         }
1166         list_for_each_entry_safe(req, n, &obd->obd_final_req_queue, rq_list) {
1167                 LASSERT (req->rq_reply_state == 0);
1168                 ptlrpc_free_clone(req);
1169         }
1170
1171         EXIT;
1172 }
1173
1174 static void target_recovery_expired(unsigned long castmeharder)
1175 {
1176         struct obd_device *obd = (struct obd_device *)castmeharder;
1177         CERROR("%s: recovery timed out, aborting\n", obd->obd_name);
1178         spin_lock_bh(&obd->obd_processing_task_lock);
1179         if (obd->obd_recovering)
1180                 obd->obd_abort_recovery = 1;
1181         cfs_waitq_signal(&obd->obd_next_transno_waitq);
1182         spin_unlock_bh(&obd->obd_processing_task_lock);
1183 }
1184
1185
1186 /* obd_processing_task_lock should be held */
1187 void target_cancel_recovery_timer(struct obd_device *obd)
1188 {
1189         CDEBUG(D_HA, "%s: cancel recovery timer\n", obd->obd_name);
1190         cfs_timer_disarm(&obd->obd_recovery_timer);
1191 }
1192
1193 static void reset_recovery_timer(struct obd_device *obd)
1194 {
1195         time_t timeout_shift = OBD_RECOVERY_TIMEOUT;
1196         spin_lock_bh(&obd->obd_processing_task_lock);
1197         if (!obd->obd_recovering) {
1198                 spin_unlock_bh(&obd->obd_processing_task_lock);
1199                 return;
1200         }
1201         if (cfs_time_current_sec() + OBD_RECOVERY_TIMEOUT >
1202             obd->obd_recovery_start + obd->obd_recovery_max_time)
1203                 timeout_shift = obd->obd_recovery_start +
1204                         obd->obd_recovery_max_time - cfs_time_current_sec();
1205         cfs_timer_arm(&obd->obd_recovery_timer, cfs_time_shift(timeout_shift));
1206         spin_unlock_bh(&obd->obd_processing_task_lock);
1207         CDEBUG(D_HA, "%s: timer will expire in %u seconds\n", obd->obd_name,
1208                (unsigned int)timeout_shift);
1209         /* Only used for lprocfs_status */
1210         obd->obd_recovery_end = cfs_time_current_sec() + timeout_shift;
1211 }
1212
1213
1214 /* Only start it the first time called */
1215 void target_start_recovery_timer(struct obd_device *obd)
1216 {
1217         spin_lock_bh(&obd->obd_processing_task_lock);
1218         if (obd->obd_recovery_handler
1219             || timer_pending((struct timer_list *)&obd->obd_recovery_timer)) {
1220                 spin_unlock_bh(&obd->obd_processing_task_lock);
1221                 return;
1222         }
1223         CWARN("%s: starting recovery timer (%us)\n", obd->obd_name,
1224               OBD_RECOVERY_TIMEOUT);
1225         cfs_timer_init(&obd->obd_recovery_timer, target_recovery_expired, obd);
1226         spin_unlock_bh(&obd->obd_processing_task_lock);
1227
1228         reset_recovery_timer(obd);
1229 }
1230
1231 #ifdef __KERNEL__
1232 static int check_for_next_transno(struct obd_device *obd)
1233 {
1234         struct ptlrpc_request *req = NULL;
1235         int wake_up = 0, connected, completed, queue_len, max;
1236         __u64 next_transno, req_transno;
1237         ENTRY;
1238         spin_lock_bh(&obd->obd_processing_task_lock);
1239
1240         if (!list_empty(&obd->obd_req_replay_queue)) {
1241                 req = list_entry(obd->obd_req_replay_queue.next,
1242                                  struct ptlrpc_request, rq_list);
1243                 req_transno = lustre_msg_get_transno(req->rq_reqmsg);
1244         } else {
1245                 req_transno = 0;
1246         }
1247
1248         max = obd->obd_max_recoverable_clients;
1249         connected = obd->obd_connected_clients;
1250         completed = connected - obd->obd_recoverable_clients;
1251         queue_len = obd->obd_requests_queued_for_recovery;
1252         next_transno = obd->obd_next_recovery_transno;
1253
1254         CDEBUG(D_HA, "max: %d, connected: %d, completed: %d, queue_len: %d, "
1255                "req_transno: "LPU64", next_transno: "LPU64"\n",
1256                max, connected, completed, queue_len, req_transno, next_transno);
1257
1258         if (obd->obd_abort_recovery) {
1259                 CDEBUG(D_HA, "waking for aborted recovery\n");
1260                 wake_up = 1;
1261         } else if (atomic_read(&obd->obd_req_replay_clients) == 0) {
1262                 CDEBUG(D_HA, "waking for completed recovery\n");
1263                 wake_up = 1;
1264         } else if (req_transno == next_transno) {
1265                 CDEBUG(D_HA, "waking for next ("LPD64")\n", next_transno);
1266                 wake_up = 1;
1267         } else if (queue_len + completed == max) {
1268                 /* handle gaps occured due to lost reply. It is allowed gaps
1269                  * because all clients are connected and there will be resend
1270                  * for missed transaction */
1271                 LASSERTF(req_transno >= next_transno,
1272                          "req_transno: "LPU64", next_transno: "LPU64"\n",
1273                          req_transno, next_transno);
1274
1275                 CDEBUG(req_transno > obd->obd_last_committed ? D_ERROR : D_HA,
1276                        "waking for skipped transno (skip: "LPD64
1277                        ", ql: %d, comp: %d, conn: %d, next: "LPD64")\n",
1278                        next_transno, queue_len, completed, connected, req_transno);
1279                 obd->obd_next_recovery_transno = req_transno;
1280                 wake_up = 1;
1281         } else if (queue_len == atomic_read(&obd->obd_req_replay_clients)) {
1282                 /* some clients haven't connected in time, but we can try
1283                  * to replay requests that demand on already committed ones
1284                  * also, we can replay first non-committed transation */
1285                 LASSERT(req_transno != 0);
1286                 if (req_transno == obd->obd_last_committed + 1) {
1287                         obd->obd_next_recovery_transno = req_transno;
1288                 } else if (req_transno > obd->obd_last_committed) {
1289                         /* can't continue recovery: have no needed transno */
1290                         obd->obd_abort_recovery = 1;
1291                         CDEBUG(D_ERROR, "abort due to missed clients. max: %d, "
1292                                "connected: %d, completed: %d, queue_len: %d, "
1293                                "req_transno: "LPU64", next_transno: "LPU64"\n",
1294                                max, connected, completed, queue_len,
1295                                req_transno, next_transno);
1296                 }
1297                 wake_up = 1;
1298         }
1299
1300         spin_unlock_bh(&obd->obd_processing_task_lock);
1301         return wake_up;
1302 }
1303
1304 static struct ptlrpc_request *target_next_replay_req(struct obd_device *obd)
1305 {
1306         struct l_wait_info lwi = { 0 };
1307         struct ptlrpc_request *req;
1308
1309         CDEBUG(D_HA, "Waiting for transno "LPD64"\n",
1310                obd->obd_next_recovery_transno);
1311         l_wait_event(obd->obd_next_transno_waitq,
1312                      check_for_next_transno(obd), &lwi);
1313
1314         spin_lock_bh(&obd->obd_processing_task_lock);
1315         if (obd->obd_abort_recovery) {
1316                 req = NULL;
1317         } else if (!list_empty(&obd->obd_req_replay_queue)) {
1318                 req = list_entry(obd->obd_req_replay_queue.next,
1319                                  struct ptlrpc_request, rq_list);
1320                 list_del_init(&req->rq_list);
1321                 obd->obd_requests_queued_for_recovery--;
1322         } else {
1323                 req = NULL;
1324         }
1325         spin_unlock_bh(&obd->obd_processing_task_lock);
1326         RETURN(req);
1327 }
1328
1329 static int check_for_next_lock(struct obd_device *obd)
1330 {
1331         struct ptlrpc_request *req = NULL;
1332         int wake_up = 0;
1333
1334         spin_lock_bh(&obd->obd_processing_task_lock);
1335         if (!list_empty(&obd->obd_lock_replay_queue)) {
1336                 req = list_entry(obd->obd_lock_replay_queue.next,
1337                                  struct ptlrpc_request, rq_list);
1338                 CDEBUG(D_HA, "waking for next lock\n");
1339                 wake_up = 1;
1340         } else if (atomic_read(&obd->obd_lock_replay_clients) == 0) {
1341                 CDEBUG(D_HA, "waking for completed lock replay\n");
1342                 wake_up = 1;
1343         } else if (obd->obd_abort_recovery) {
1344                 CDEBUG(D_HA, "waking for aborted recovery\n");
1345                 wake_up = 1;
1346         }
1347         spin_unlock_bh(&obd->obd_processing_task_lock);
1348
1349         return wake_up;
1350 }
1351
1352 static struct ptlrpc_request *target_next_replay_lock(struct obd_device *obd)
1353 {
1354         struct l_wait_info lwi = { 0 };
1355         struct ptlrpc_request *req;
1356
1357         CDEBUG(D_HA, "Waiting for lock\n");
1358         l_wait_event(obd->obd_next_transno_waitq,
1359                      check_for_next_lock(obd), &lwi);
1360
1361         spin_lock_bh(&obd->obd_processing_task_lock);
1362         if (obd->obd_abort_recovery) {
1363                 req = NULL;
1364         } else if (!list_empty(&obd->obd_lock_replay_queue)) {
1365                 req = list_entry(obd->obd_lock_replay_queue.next,
1366                                  struct ptlrpc_request, rq_list);
1367                 list_del_init(&req->rq_list);
1368         } else {
1369                 req = NULL;
1370         }
1371         spin_unlock_bh(&obd->obd_processing_task_lock);
1372         return req;
1373 }
1374
1375 static struct ptlrpc_request *target_next_final_ping(struct obd_device *obd)
1376 {
1377         struct ptlrpc_request *req;
1378
1379         spin_lock_bh(&obd->obd_processing_task_lock);
1380         if (!list_empty(&obd->obd_final_req_queue)) {
1381                 req = list_entry(obd->obd_final_req_queue.next,
1382                                  struct ptlrpc_request, rq_list);
1383                 list_del_init(&req->rq_list);
1384         } else {
1385                 req = NULL;
1386         }
1387         spin_unlock_bh(&obd->obd_processing_task_lock);
1388         return req;
1389 }
1390
1391 static inline int req_replay_done(struct obd_export *exp)
1392 {
1393         return (exp->exp_req_replay_needed == 0);
1394 }
1395
1396 static inline int lock_replay_done(struct obd_export *exp)
1397 {
1398         return (exp->exp_lock_replay_needed == 0);
1399 }
1400
1401 static inline int connect_done(struct obd_export *exp)
1402 {
1403         return (exp->exp_in_recovery != 0);
1404 }
1405
1406 static int check_for_clients(struct obd_device *obd)
1407 {
1408         if (obd->obd_abort_recovery)
1409                 return 1;
1410         LASSERT(obd->obd_connected_clients <= obd->obd_max_recoverable_clients);
1411         if (obd->obd_no_conn == 0 &&
1412             obd->obd_connected_clients == obd->obd_max_recoverable_clients)
1413                 return 1;
1414         return 0;
1415 }
1416
1417 static int handle_recovery_req(struct ptlrpc_thread *thread,
1418                                struct ptlrpc_request *req,
1419                                svc_handler_t handler)
1420 {
1421         int rc;
1422         ENTRY;
1423
1424         rc = lu_context_init(&req->rq_session, LCT_SESSION);
1425         if (rc) {
1426                 CERROR("Failure to initialize session: %d\n", rc);
1427                 return rc;
1428         }
1429         req->rq_session.lc_thread = thread;
1430         lu_context_enter(&req->rq_session);
1431         req->rq_svc_thread = thread;
1432         req->rq_svc_thread->t_env->le_ses = &req->rq_session;
1433
1434         /* thread context */
1435         lu_context_enter(&thread->t_env->le_ctx);
1436         (void)handler(req);
1437         lu_context_exit(&thread->t_env->le_ctx);
1438
1439         lu_context_exit(&req->rq_session);
1440         lu_context_fini(&req->rq_session);
1441         /* don't reset timer for final stage */
1442         if (!req_replay_done(req->rq_export) ||
1443             !lock_replay_done(req->rq_export))
1444                 reset_recovery_timer(class_exp2obd(req->rq_export));
1445         ptlrpc_free_clone(req);
1446         RETURN(0);
1447 }
1448
1449 static int target_recovery_thread(void *arg)
1450 {
1451         struct obd_device *obd = arg;
1452         struct ptlrpc_request *req;
1453         struct target_recovery_data *trd = &obd->obd_recovery_data;
1454         struct l_wait_info lwi = { 0 };
1455         unsigned long delta;
1456         unsigned long flags;
1457         struct lu_env env;
1458         struct ptlrpc_thread fake_svc_thread, *thread = &fake_svc_thread;
1459         __u32 recov_ctx_tags = LCT_MD_THREAD;
1460         int rc = 0;
1461         ENTRY;
1462
1463         cfs_daemonize("tgt_recov");
1464
1465         SIGNAL_MASK_LOCK(current, flags);
1466         sigfillset(&current->blocked);
1467         RECALC_SIGPENDING;
1468         SIGNAL_MASK_UNLOCK(current, flags);
1469
1470         rc = lu_context_init(&env.le_ctx, recov_ctx_tags);
1471         if (rc)
1472                 RETURN(rc);
1473
1474         thread->t_env = &env;
1475         env.le_ctx.lc_thread = thread;
1476
1477         CERROR("%s: started recovery thread pid %d\n", obd->obd_name,
1478                current->pid);
1479         trd->trd_processing_task = current->pid;
1480
1481         obd->obd_recovering = 1;
1482         complete(&trd->trd_starting);
1483
1484         /* first of all, we have to know the first transno to replay */
1485         obd->obd_abort_recovery = 0;
1486         l_wait_event(obd->obd_next_transno_waitq,
1487                      check_for_clients(obd), &lwi);
1488
1489         spin_lock_bh(&obd->obd_processing_task_lock);
1490         target_cancel_recovery_timer(obd);
1491         spin_unlock_bh(&obd->obd_processing_task_lock);
1492
1493         /* If some clients haven't connected in time, evict them */
1494         if (obd->obd_abort_recovery) {
1495                 CWARN("Some clients haven't connect in time (%d/%d),"
1496                        "evict them\n", obd->obd_connected_clients,
1497                        obd->obd_max_recoverable_clients);
1498                 obd->obd_abort_recovery = obd->obd_stopping;
1499                 class_disconnect_stale_exports(obd, connect_done);
1500         }
1501         /* next stage: replay requests */
1502         delta = jiffies;
1503         obd->obd_req_replaying = 1;
1504         CDEBUG(D_INFO, "1: request replay stage - %d clients from t"LPU64"\n",
1505               atomic_read(&obd->obd_req_replay_clients),
1506               obd->obd_next_recovery_transno);
1507         while ((req = target_next_replay_req(obd))) {
1508                 LASSERT(trd->trd_processing_task == current->pid);
1509                 DEBUG_REQ(D_HA, req, "processing t"LPD64" from %s",
1510                           lustre_msg_get_transno(req->rq_reqmsg),
1511                           libcfs_nid2str(req->rq_peer.nid));
1512                 handle_recovery_req(thread, req,
1513                                     trd->trd_recovery_handler);
1514                 obd->obd_replayed_requests++;
1515                 spin_lock_bh(&obd->obd_processing_task_lock);
1516                 obd->obd_next_recovery_transno++;
1517                 spin_unlock_bh(&obd->obd_processing_task_lock);
1518         }
1519
1520         spin_lock_bh(&obd->obd_processing_task_lock);
1521         target_cancel_recovery_timer(obd);
1522         spin_unlock_bh(&obd->obd_processing_task_lock);
1523
1524         /* If some clients haven't replayed requests in time, evict them */
1525         if (obd->obd_abort_recovery) {
1526                 CDEBUG(D_ERROR, "req replay timed out, aborting ...\n");
1527                 obd->obd_abort_recovery = obd->obd_stopping;
1528                 class_disconnect_stale_exports(obd, req_replay_done);
1529                 abort_req_replay_queue(obd);
1530         }
1531         /* The second stage: replay locks */
1532         CDEBUG(D_INFO, "2: lock replay stage - %d clients\n",
1533                atomic_read(&obd->obd_lock_replay_clients));
1534         while ((req = target_next_replay_lock(obd))) {
1535                 LASSERT(trd->trd_processing_task == current->pid);
1536                 DEBUG_REQ(D_HA|D_WARNING, req, "processing lock from %s: ",
1537                           libcfs_nid2str(req->rq_peer.nid));
1538                 handle_recovery_req(thread, req,
1539                                     trd->trd_recovery_handler);
1540                 obd->obd_replayed_locks++;
1541         }
1542
1543         spin_lock_bh(&obd->obd_processing_task_lock);
1544         target_cancel_recovery_timer(obd);
1545         spin_unlock_bh(&obd->obd_processing_task_lock);
1546         /* If some clients haven't replayed requests in time, evict them */
1547         if (obd->obd_abort_recovery) {
1548                 int stale;
1549                 CERROR("lock replay timed out, aborting ...\n");
1550                 obd->obd_abort_recovery = obd->obd_stopping;
1551                 stale = class_disconnect_stale_exports(obd, lock_replay_done);
1552                 abort_lock_replay_queue(obd);
1553         }
1554
1555         /* We drop recoverying flag to forward all new requests
1556          * to regular mds_handle() since now */
1557         spin_lock_bh(&obd->obd_processing_task_lock);
1558         obd->obd_recovering = obd->obd_abort_recovery = 0;
1559         spin_unlock_bh(&obd->obd_processing_task_lock);
1560         /* The third stage: reply on final pings */
1561         CDEBUG(D_INFO, "3: final stage - process recovery completion pings\n");
1562         while ((req = target_next_final_ping(obd))) {
1563                 LASSERT(trd->trd_processing_task == current->pid);
1564                 DEBUG_REQ(D_HA, req, "processing final ping from %s: ",
1565                           libcfs_nid2str(req->rq_peer.nid));
1566                 handle_recovery_req(thread, req,
1567                                     trd->trd_recovery_handler);
1568         }
1569
1570         delta = (jiffies - delta) / HZ;
1571         CDEBUG(D_INFO,"4: recovery completed in %lus - %d/%d reqs/locks\n",
1572               delta, obd->obd_replayed_requests, obd->obd_replayed_locks);
1573         LASSERT(atomic_read(&obd->obd_req_replay_clients) == 0);
1574         LASSERT(atomic_read(&obd->obd_lock_replay_clients) == 0);
1575         if (delta > obd_timeout * 2) {
1576                 CWARN("too long recovery - read logs\n");
1577                 libcfs_debug_dumplog();
1578         }
1579
1580         target_finish_recovery(obd);
1581
1582         lu_context_fini(&env.le_ctx);
1583         trd->trd_processing_task = 0;
1584         complete(&trd->trd_finishing);
1585         RETURN(rc);
1586 }
1587
1588 int target_start_recovery_thread(struct obd_device *obd, svc_handler_t handler)
1589 {
1590         int rc = 0;
1591         struct target_recovery_data *trd = &obd->obd_recovery_data;
1592
1593         memset(trd, 0, sizeof(*trd));
1594         init_completion(&trd->trd_starting);
1595         init_completion(&trd->trd_finishing);
1596         trd->trd_recovery_handler = handler;
1597
1598         if (kernel_thread(target_recovery_thread, obd, 0) > 0) {
1599                 wait_for_completion(&trd->trd_starting);
1600                 LASSERT(obd->obd_recovering != 0);
1601         } else
1602                 rc = -ECHILD;
1603
1604         return rc;
1605 }
1606
1607 void target_stop_recovery_thread(struct obd_device *obd)
1608 {
1609         spin_lock_bh(&obd->obd_processing_task_lock);
1610         if (obd->obd_recovery_data.trd_processing_task > 0) {
1611                 struct target_recovery_data *trd = &obd->obd_recovery_data;
1612                 CERROR("%s: Aborting recovery\n", obd->obd_name);
1613                 obd->obd_abort_recovery = 1;
1614                 wake_up(&obd->obd_next_transno_waitq);
1615                 spin_unlock_bh(&obd->obd_processing_task_lock);
1616                 wait_for_completion(&trd->trd_finishing);
1617         } else {
1618                 spin_unlock_bh(&obd->obd_processing_task_lock);
1619         }
1620 }
1621
1622 void target_recovery_fini(struct obd_device *obd)
1623 {
1624         class_disconnect_exports(obd);
1625         target_stop_recovery_thread(obd);
1626         target_cleanup_recovery(obd);
1627 }
1628 EXPORT_SYMBOL(target_recovery_fini);
1629
1630 void target_recovery_init(struct obd_device *obd, svc_handler_t handler)
1631 {
1632         if (obd->obd_max_recoverable_clients == 0)
1633                 return;
1634
1635         CWARN("RECOVERY: service %s, %d recoverable clients, "
1636               "last_transno "LPU64"\n", obd->obd_name,
1637               obd->obd_max_recoverable_clients, obd->obd_last_committed);
1638         obd->obd_next_recovery_transno = obd->obd_last_committed + 1;
1639         target_start_recovery_thread(obd, handler);
1640         obd->obd_recovery_start = cfs_time_current_sec();
1641         /* Only used for lprocfs_status */
1642         obd->obd_recovery_end = obd->obd_recovery_start + OBD_RECOVERY_TIMEOUT;
1643         /* bz13079: this should be set to desired value for ost but not for mds */
1644         obd->obd_recovery_max_time = OBD_RECOVERY_MAX_TIME;
1645 }
1646 EXPORT_SYMBOL(target_recovery_init);
1647
1648 #endif
1649
1650 int target_process_req_flags(struct obd_device *obd, struct ptlrpc_request *req)
1651 {
1652         struct obd_export *exp = req->rq_export;
1653         LASSERT(exp != NULL);
1654         if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REQ_REPLAY_DONE) {
1655                 /* client declares he's ready to replay locks */
1656                 spin_lock_bh(&obd->obd_processing_task_lock);
1657                 if (exp->exp_req_replay_needed) {
1658                         LASSERT(atomic_read(&obd->obd_req_replay_clients) > 0);
1659                         spin_lock(&exp->exp_lock);
1660                         exp->exp_req_replay_needed = 0;
1661                         spin_unlock(&exp->exp_lock);
1662                         atomic_dec(&obd->obd_req_replay_clients);
1663                         LASSERT(obd->obd_recoverable_clients > 0);
1664                         obd->obd_recoverable_clients--;
1665                         if (atomic_read(&obd->obd_req_replay_clients) == 0)
1666                                 CDEBUG(D_HA, "all clients have replayed reqs\n");
1667                         wake_up(&obd->obd_next_transno_waitq);
1668                 }
1669                 spin_unlock_bh(&obd->obd_processing_task_lock);
1670         }
1671         if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_LOCK_REPLAY_DONE) {
1672                 /* client declares he's ready to complete recovery
1673                  * so, we put the request on th final queue */
1674                 spin_lock_bh(&obd->obd_processing_task_lock);
1675                 if (exp->exp_lock_replay_needed) {
1676                         LASSERT(atomic_read(&obd->obd_lock_replay_clients) > 0);
1677                         spin_lock(&exp->exp_lock);
1678                         exp->exp_lock_replay_needed = 0;
1679                         spin_unlock(&exp->exp_lock);
1680                         atomic_dec(&obd->obd_lock_replay_clients);
1681                         if (atomic_read(&obd->obd_lock_replay_clients) == 0)
1682                                 CDEBUG(D_HA, "all clients have replayed locks\n");
1683                         wake_up(&obd->obd_next_transno_waitq);
1684                 }
1685                 spin_unlock_bh(&obd->obd_processing_task_lock);
1686         }
1687
1688         return 0;
1689 }
1690
1691 int target_queue_recovery_request(struct ptlrpc_request *req,
1692                                   struct obd_device *obd)
1693 {
1694         struct list_head *tmp;
1695         int inserted = 0;
1696         __u64 transno = lustre_msg_get_transno(req->rq_reqmsg);
1697
1698         ENTRY;
1699
1700         if (obd->obd_recovery_data.trd_processing_task == cfs_curproc_pid()) {
1701                 /* Processing the queue right now, don't re-add. */
1702                 RETURN(1);
1703         }
1704
1705         target_process_req_flags(obd, req);
1706
1707         if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_LOCK_REPLAY_DONE) {
1708                 /* client declares he's ready to complete recovery
1709                  * so, we put the request on th final queue */
1710                 req = ptlrpc_clone_req(req);
1711                 if (req == NULL)
1712                         RETURN(-ENOMEM);
1713                 DEBUG_REQ(D_HA, req, "queue final req");
1714                 spin_lock_bh(&obd->obd_processing_task_lock);
1715                 if (obd->obd_recovering)
1716                         list_add_tail(&req->rq_list, &obd->obd_final_req_queue);
1717                 else {
1718                         spin_unlock_bh(&obd->obd_processing_task_lock);
1719                         ptlrpc_free_clone(req);
1720                         if (obd->obd_stopping) {
1721                                 RETURN(-ENOTCONN);
1722                         } else {
1723                                 RETURN(1);
1724                         }
1725                 }
1726                 spin_unlock_bh(&obd->obd_processing_task_lock);
1727                 RETURN(0);
1728         }
1729         if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REQ_REPLAY_DONE) {
1730                 /* client declares he's ready to replay locks */
1731                 req = ptlrpc_clone_req(req);
1732                 if (req == NULL)
1733                         RETURN(-ENOMEM);
1734                 DEBUG_REQ(D_HA, req, "queue lock replay req");
1735                 spin_lock_bh(&obd->obd_processing_task_lock);
1736                 LASSERT(obd->obd_recovering);
1737                 /* usually due to recovery abort */
1738                 if (!req->rq_export->exp_in_recovery) {
1739                         spin_unlock_bh(&obd->obd_processing_task_lock);
1740                         ptlrpc_free_clone(req);
1741                         RETURN(-ENOTCONN);
1742                 }
1743                 LASSERT(req->rq_export->exp_lock_replay_needed);
1744                 list_add_tail(&req->rq_list, &obd->obd_lock_replay_queue);
1745                 spin_unlock_bh(&obd->obd_processing_task_lock);
1746                 wake_up(&obd->obd_next_transno_waitq);
1747                 RETURN(0);
1748         }
1749
1750         /* CAVEAT EMPTOR: The incoming request message has been swabbed
1751          * (i.e. buflens etc are in my own byte order), but type-dependent
1752          * buffers (eg mds_body, ost_body etc) have NOT been swabbed. */
1753
1754         if (!transno) {
1755                 CFS_INIT_LIST_HEAD(&req->rq_list);
1756                 DEBUG_REQ(D_HA, req, "not queueing");
1757                 RETURN(1);
1758         }
1759
1760         spin_lock_bh(&obd->obd_processing_task_lock);
1761
1762         /* If we're processing the queue, we want don't want to queue this
1763          * message.
1764          *
1765          * Also, if this request has a transno less than the one we're waiting
1766          * for, we should process it now.  It could (and currently always will)
1767          * be an open request for a descriptor that was opened some time ago.
1768          *
1769          * Also, a resent, replayed request that has already been
1770          * handled will pass through here and be processed immediately.
1771          */
1772         CWARN("Next recovery transno: "LPU64", current: "LPU64", replaying: %i\n",
1773               obd->obd_next_recovery_transno, transno, obd->obd_req_replaying);
1774         if (transno < obd->obd_next_recovery_transno && obd->obd_req_replaying) {
1775                 /* Processing the queue right now, don't re-add. */
1776                 LASSERT(list_empty(&req->rq_list));
1777                 spin_unlock_bh(&obd->obd_processing_task_lock);
1778                 RETURN(1);
1779         }
1780         spin_unlock_bh(&obd->obd_processing_task_lock);
1781
1782         /* A resent, replayed request that is still on the queue; just drop it.
1783            The queued request will handle this. */
1784         if ((lustre_msg_get_flags(req->rq_reqmsg) & (MSG_RESENT|MSG_REPLAY)) ==
1785             (MSG_RESENT | MSG_REPLAY)) {
1786                 DEBUG_REQ(D_ERROR, req, "dropping resent queued req");
1787                 RETURN(0);
1788         }
1789
1790         req = ptlrpc_clone_req(req);
1791         if (req == NULL)
1792                 RETURN(-ENOMEM);
1793
1794         spin_lock_bh(&obd->obd_processing_task_lock);
1795         LASSERT(obd->obd_recovering);
1796         if (!req->rq_export->exp_in_recovery) {
1797                 spin_unlock_bh(&obd->obd_processing_task_lock);
1798                 ptlrpc_free_clone(req);
1799                 RETURN(-ENOTCONN);
1800         }
1801         LASSERT(req->rq_export->exp_req_replay_needed);
1802
1803         /* XXX O(n^2) */
1804         list_for_each(tmp, &obd->obd_req_replay_queue) {
1805                 struct ptlrpc_request *reqiter =
1806                         list_entry(tmp, struct ptlrpc_request, rq_list);
1807
1808                 if (lustre_msg_get_transno(reqiter->rq_reqmsg) > transno) {
1809                         list_add_tail(&req->rq_list, &reqiter->rq_list);
1810                         inserted = 1;
1811                         break;
1812                 }
1813         }
1814
1815         if (!inserted)
1816                 list_add_tail(&req->rq_list, &obd->obd_req_replay_queue);
1817
1818         obd->obd_requests_queued_for_recovery++;
1819         wake_up(&obd->obd_next_transno_waitq);
1820         spin_unlock_bh(&obd->obd_processing_task_lock);
1821         RETURN(0);
1822
1823 }
1824
1825 struct obd_device * target_req2obd(struct ptlrpc_request *req)
1826 {
1827         return req->rq_export->exp_obd;
1828 }
1829
1830 static inline struct ldlm_pool *ldlm_exp2pl(struct obd_export *exp)
1831 {
1832         LASSERT(exp != NULL);
1833         return &exp->exp_obd->obd_namespace->ns_pool;
1834 }
1835
1836 /**
1837  * Packs current SLV and Limit into \a req.
1838  */
1839 int target_pack_pool_reply(struct ptlrpc_request *req)
1840 {
1841         struct obd_device *obd;
1842         ENTRY;
1843
1844         /*
1845          * Check that we still have all structures alive as this may
1846          * be some late rpc in shutdown time.
1847          */
1848         if (unlikely(!req->rq_export || !req->rq_export->exp_obd ||
1849                      !exp_connect_lru_resize(req->rq_export))) {
1850                 lustre_msg_set_slv(req->rq_repmsg, 0);
1851                 lustre_msg_set_limit(req->rq_repmsg, 0);
1852                 RETURN(0);
1853         }
1854
1855         /*
1856          * OBD is alive here as export is alive, which we checked above.
1857          */
1858         obd = req->rq_export->exp_obd;
1859
1860         read_lock(&obd->obd_pool_lock);
1861         lustre_msg_set_slv(req->rq_repmsg, obd->obd_pool_slv);
1862         lustre_msg_set_limit(req->rq_repmsg, obd->obd_pool_limit);
1863         read_unlock(&obd->obd_pool_lock);
1864
1865         RETURN(0);
1866 }
1867
1868 int target_send_reply_msg(struct ptlrpc_request *req, int rc, int fail_id)
1869 {
1870         if (OBD_FAIL_CHECK_ORSET(fail_id & ~OBD_FAIL_ONCE, OBD_FAIL_ONCE)) {
1871                 DEBUG_REQ(D_ERROR, req, "dropping reply");
1872                 return (-ECOMM);
1873         }
1874
1875         if (unlikely(rc)) {
1876                 DEBUG_REQ(D_ERROR, req, "processing error (%d)", rc);
1877                 req->rq_status = rc;
1878                 return (ptlrpc_send_error(req, 1));
1879         } else {
1880                 DEBUG_REQ(D_NET, req, "sending reply");
1881         }
1882
1883         return (ptlrpc_send_reply(req, 1));
1884 }
1885
1886 void target_send_reply(struct ptlrpc_request *req, int rc, int fail_id)
1887 {
1888         int                        netrc;
1889         struct ptlrpc_reply_state *rs;
1890         struct obd_device         *obd;
1891         struct obd_export         *exp;
1892         struct ptlrpc_service     *svc;
1893
1894         if (req->rq_no_reply)
1895                 return;
1896
1897         svc = req->rq_rqbd->rqbd_service;
1898         rs = req->rq_reply_state;
1899         if (rs == NULL || !rs->rs_difficult) {
1900                 /* no notifiers */
1901                 target_send_reply_msg (req, rc, fail_id);
1902                 return;
1903         }
1904
1905         /* must be an export if locks saved */
1906         LASSERT (req->rq_export != NULL);
1907         /* req/reply consistent */
1908         LASSERT (rs->rs_service == svc);
1909
1910         /* "fresh" reply */
1911         LASSERT (!rs->rs_scheduled);
1912         LASSERT (!rs->rs_scheduled_ever);
1913         LASSERT (!rs->rs_handled);
1914         LASSERT (!rs->rs_on_net);
1915         LASSERT (rs->rs_export == NULL);
1916         LASSERT (list_empty(&rs->rs_obd_list));
1917         LASSERT (list_empty(&rs->rs_exp_list));
1918
1919         exp = class_export_get (req->rq_export);
1920         obd = exp->exp_obd;
1921
1922         /* disable reply scheduling onto srv_reply_queue while I'm setting up */
1923         rs->rs_scheduled = 1;
1924         rs->rs_on_net    = 1;
1925         rs->rs_xid       = req->rq_xid;
1926         rs->rs_transno   = req->rq_transno;
1927         rs->rs_export    = exp;
1928
1929         spin_lock(&obd->obd_uncommitted_replies_lock);
1930
1931         if (rs->rs_transno > obd->obd_last_committed) {
1932                 /* not committed already */
1933                 list_add_tail (&rs->rs_obd_list,
1934                                &obd->obd_uncommitted_replies);
1935         }
1936
1937         spin_unlock (&obd->obd_uncommitted_replies_lock);
1938         spin_lock (&exp->exp_lock);
1939
1940         list_add_tail (&rs->rs_exp_list, &exp->exp_outstanding_replies);
1941
1942         spin_unlock(&exp->exp_lock);
1943
1944         netrc = target_send_reply_msg (req, rc, fail_id);
1945
1946         spin_lock(&svc->srv_lock);
1947
1948         svc->srv_n_difficult_replies++;
1949
1950         if (netrc != 0) {
1951                 /* error sending: reply is off the net.  Also we need +1
1952                  * reply ref until ptlrpc_server_handle_reply() is done
1953                  * with the reply state (if the send was successful, there
1954                  * would have been +1 ref for the net, which
1955                  * reply_out_callback leaves alone) */
1956                 rs->rs_on_net = 0;
1957                 ptlrpc_rs_addref(rs);
1958                 atomic_inc (&svc->srv_outstanding_replies);
1959         }
1960
1961         if (!rs->rs_on_net ||                   /* some notifier */
1962             list_empty(&rs->rs_exp_list) ||     /* completed already */
1963             list_empty(&rs->rs_obd_list)) {
1964                 list_add_tail (&rs->rs_list, &svc->srv_reply_queue);
1965                 cfs_waitq_signal (&svc->srv_waitq);
1966         } else {
1967                 list_add (&rs->rs_list, &svc->srv_active_replies);
1968                 rs->rs_scheduled = 0;           /* allow notifier to schedule */
1969         }
1970
1971         spin_unlock(&svc->srv_lock);
1972 }
1973
1974 int target_handle_ping(struct ptlrpc_request *req)
1975 {
1976         obd_ping(req->rq_export);
1977         return req_capsule_server_pack(&req->rq_pill);
1978 }
1979
1980 void target_committed_to_req(struct ptlrpc_request *req)
1981 {
1982         struct obd_device *obd;
1983
1984         if (req == NULL || req->rq_export == NULL)
1985                 return;
1986
1987         obd = req->rq_export->exp_obd;
1988         if (obd == NULL)
1989                 return;
1990
1991         if (!obd->obd_no_transno && req->rq_repmsg != NULL)
1992                 lustre_msg_set_last_committed(req->rq_repmsg,
1993                                               obd->obd_last_committed);
1994         else
1995                 DEBUG_REQ(D_IOCTL, req, "not sending last_committed update");
1996
1997         CDEBUG(D_INFO, "last_committed "LPU64", transno "LPU64", xid "LPU64"\n",
1998                obd->obd_last_committed, req->rq_transno, req->rq_xid);
1999 }
2000
2001 EXPORT_SYMBOL(target_committed_to_req);
2002
2003 #ifdef HAVE_QUOTA_SUPPORT
2004 int target_handle_qc_callback(struct ptlrpc_request *req)
2005 {
2006         struct obd_quotactl *oqctl;
2007         struct client_obd *cli = &req->rq_export->exp_obd->u.cli;
2008
2009         oqctl = req_capsule_client_get(&req->rq_pill, &RMF_OBD_QUOTACTL);
2010         if (oqctl == NULL)
2011                 RETURN(-EPROTO);
2012
2013         cli->cl_qchk_stat = oqctl->qc_stat;
2014
2015         return 0;
2016 }
2017
2018 int target_handle_dqacq_callback(struct ptlrpc_request *req)
2019 {
2020 #ifdef __KERNEL__
2021         struct obd_device *obd = req->rq_export->exp_obd;
2022         struct obd_device *master_obd;
2023         struct lustre_quota_ctxt *qctxt;
2024         struct qunit_data *qdata;
2025         void* rep;
2026         struct qunit_data_old *qdata_old;
2027         int rc = 0;
2028         ENTRY;
2029
2030         rc = req_capsule_server_pack(&req->rq_pill);
2031         if (rc) {
2032                 CERROR("packing reply failed!: rc = %d\n", rc);
2033                 RETURN(rc);
2034         }
2035
2036         LASSERT(req->rq_export);
2037
2038         /* fixed for bug10707 */
2039         if ((req->rq_export->exp_connect_flags & OBD_CONNECT_QUOTA64) &&
2040             !OBD_FAIL_CHECK(OBD_FAIL_QUOTA_QD_COUNT_32BIT)) {
2041                 CDEBUG(D_QUOTA, "qd_count is 64bit!\n");
2042                 rep = req_capsule_server_get(&req->rq_pill,
2043                                              &RMF_QUNIT_DATA);
2044                 LASSERT(rep);
2045                 qdata = req_capsule_client_swab_get(&req->rq_pill,
2046                                                     &RMF_QUNIT_DATA,
2047                                           (void*)lustre_swab_qdata);
2048         } else {
2049                 CDEBUG(D_QUOTA, "qd_count is 32bit!\n");
2050                 rep = req_capsule_server_get(&req->rq_pill, &RMF_QUNIT_DATA);
2051                 LASSERT(rep);
2052                 qdata_old = req_capsule_client_swab_get(&req->rq_pill,
2053                                                         &RMF_QUNIT_DATA,
2054                                            (void*)lustre_swab_qdata_old);
2055                 qdata = lustre_quota_old_to_new(qdata_old);
2056         }
2057
2058         if (qdata == NULL)
2059                 RETURN(-EPROTO);
2060
2061         /* we use the observer */
2062         LASSERT(obd->obd_observer && obd->obd_observer->obd_observer);
2063         master_obd = obd->obd_observer->obd_observer;
2064         qctxt = &master_obd->u.obt.obt_qctxt;
2065
2066         LASSERT(qctxt->lqc_handler);
2067         rc = qctxt->lqc_handler(master_obd, qdata,
2068                                 lustre_msg_get_opc(req->rq_reqmsg));
2069         if (rc && rc != -EDQUOT)
2070                 CDEBUG(rc == -EBUSY  ? D_QUOTA : D_ERROR,
2071                        "dqacq failed! (rc:%d)\n", rc);
2072
2073         /* the qd_count might be changed in lqc_handler */
2074         if ((req->rq_export->exp_connect_flags & OBD_CONNECT_QUOTA64) &&
2075             !OBD_FAIL_CHECK(OBD_FAIL_QUOTA_QD_COUNT_32BIT)) {
2076                 memcpy(rep, qdata, sizeof(*qdata));
2077         } else {
2078                 qdata_old = lustre_quota_new_to_old(qdata);
2079                 memcpy(rep, qdata_old, sizeof(*qdata_old));
2080         }
2081         req->rq_status = rc;
2082         rc = ptlrpc_reply(req);
2083
2084         RETURN(rc);
2085 #else
2086         return 0;
2087 #endif /* !__KERNEL__ */
2088 }
2089 #endif /* HAVE_QUOTA_SUPPORT */
2090
2091 ldlm_mode_t lck_compat_array[] = {
2092         [LCK_EX] LCK_COMPAT_EX,
2093         [LCK_PW] LCK_COMPAT_PW,
2094         [LCK_PR] LCK_COMPAT_PR,
2095         [LCK_CW] LCK_COMPAT_CW,
2096         [LCK_CR] LCK_COMPAT_CR,
2097         [LCK_NL] LCK_COMPAT_NL,
2098         [LCK_GROUP] LCK_COMPAT_GROUP
2099 };