lustre/ldlm/ldlm_lib.c

   1 /*
   2  * GPL HEADER START
   3  *
   4  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
   5  *
   6  * This program is free software; you can redistribute it and/or modify
   7  * it under the terms of the GNU General Public License version 2 only,
   8  * as published by the Free Software Foundation.
   9  *
  10  * This program is distributed in the hope that it will be useful, but
  11  * WITHOUT ANY WARRANTY; without even the implied warranty of
  12  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  13  * General Public License version 2 for more details (a copy is included
  14  * in the LICENSE file that accompanied this code).
  15  *
  16  * You should have received a copy of the GNU General Public License
  17  * version 2 along with this program; If not, see
  18  * http://www.gnu.org/licenses/gpl-2.0.html
  19  *
  20  * GPL HEADER END
  21  */
  22 /*
  23  * Copyright (c) 2003, 2010, Oracle and/or its affiliates. All rights reserved.
  24  * Use is subject to license terms.
  25  *
  26  * Copyright (c) 2010, 2017, Intel Corporation.
  27  */
  28 /*
  29  * This file is part of Lustre, http://www.lustre.org/
  30  * Lustre is a trademark of Sun Microsystems, Inc.
  31  */
  32
  33 /**
  34  * This file deals with various client/target related logic including recovery.
  35  *
  36  * TODO: This code more logically belongs in the ptlrpc module than in ldlm and
  37  * should be moved.
  38  */
  39
  40 #define DEBUG_SUBSYSTEM S_LDLM
  41
  42 #include <linux/jiffies.h>
  43 #include <linux/kthread.h>
  44 #include <libcfs/libcfs.h>
  45 #include <obd.h>
  46 #include <obd_class.h>
  47 #include <lustre_dlm.h>
  48 #include <lustre_net.h>
  49 #include <lustre_sec.h>
  50 #include "ldlm_internal.h"
  51
  52 /* @priority: If non-zero, move the selected connection to the list head.
  53  * @create: If zero, only search in existing connections.
  54  */
  55 static int import_set_conn(struct obd_import *imp, struct obd_uuid *uuid,
  56                            int priority, int create)
  57 {
  58         struct ptlrpc_connection *ptlrpc_conn;
  59         struct obd_import_conn *imp_conn = NULL, *item;
  60         lnet_nid_t nid4refnet = LNET_NID_ANY;
  61         int rc = 0;
  62         ENTRY;
  63
  64         if (!create && !priority) {
  65                 CDEBUG(D_HA, "Nothing to do\n");
  66                 RETURN(-EINVAL);
  67         }
  68
  69         if (imp->imp_connection &&
  70             imp->imp_connection->c_remote_uuid.uuid[0] == 0)
  71                 /* nid4refnet is used to restrict network connections */
  72                 nid4refnet = imp->imp_connection->c_self;
  73         ptlrpc_conn = ptlrpc_uuid_to_connection(uuid, nid4refnet);
  74         if (!ptlrpc_conn) {
  75                 CDEBUG(D_HA, "can't find connection %s\n", uuid->uuid);
  76                 RETURN(-ENOENT);
  77         }
  78
  79         if (create) {
  80                 OBD_ALLOC(imp_conn, sizeof(*imp_conn));
  81                 if (!imp_conn)
  82                         GOTO(out_put, rc = -ENOMEM);
  83         }
  84
  85         spin_lock(&imp->imp_lock);
  86         list_for_each_entry(item, &imp->imp_conn_list, oic_item) {
  87                 if (obd_uuid_equals(uuid, &item->oic_uuid)) {
  88                         if (priority) {
  89                                 list_del(&item->oic_item);
  90                                 list_add(&item->oic_item,
  91                                          &imp->imp_conn_list);
  92                                 item->oic_last_attempt = 0;
  93                         }
  94                         CDEBUG(D_HA, "imp %p@%s: found existing conn %s%s\n",
  95                                imp, imp->imp_obd->obd_name, uuid->uuid,
  96                                (priority ? ", moved to head" : ""));
  97                         spin_unlock(&imp->imp_lock);
  98                         GOTO(out_free, rc = 0);
  99                 }
 100         }
 101         /* No existing import connection found for \a uuid. */
 102         if (create) {
 103                 imp_conn->oic_conn = ptlrpc_conn;
 104                 imp_conn->oic_uuid = *uuid;
 105                 imp_conn->oic_last_attempt = 0;
 106                 if (priority)
 107                         list_add(&imp_conn->oic_item, &imp->imp_conn_list);
 108                 else
 109                         list_add_tail(&imp_conn->oic_item,
 110                                       &imp->imp_conn_list);
 111                 CDEBUG(D_HA, "imp %p@%s: add connection %s at %s\n",
 112                        imp, imp->imp_obd->obd_name, uuid->uuid,
 113                        (priority ? "head" : "tail"));
 114         } else {
 115                 spin_unlock(&imp->imp_lock);
 116                 GOTO(out_free, rc = -ENOENT);
 117         }
 118
 119         spin_unlock(&imp->imp_lock);
 120         RETURN(0);
 121 out_free:
 122         if (imp_conn)
 123                 OBD_FREE(imp_conn, sizeof(*imp_conn));
 124 out_put:
 125         ptlrpc_connection_put(ptlrpc_conn);
 126         RETURN(rc);
 127 }
 128
 129 int import_set_conn_priority(struct obd_import *imp, struct obd_uuid *uuid)
 130 {
 131         return import_set_conn(imp, uuid, 1, 0);
 132 }
 133
 134 int client_import_add_conn(struct obd_import *imp, struct obd_uuid *uuid,
 135                            int priority)
 136 {
 137         return import_set_conn(imp, uuid, priority, 1);
 138 }
 139 EXPORT_SYMBOL(client_import_add_conn);
 140
 141 int client_import_del_conn(struct obd_import *imp, struct obd_uuid *uuid)
 142 {
 143         struct obd_import_conn *imp_conn;
 144         struct obd_export *dlmexp;
 145         int rc = -ENOENT;
 146         ENTRY;
 147
 148         spin_lock(&imp->imp_lock);
 149         if (list_empty(&imp->imp_conn_list)) {
 150                 LASSERT(!imp->imp_connection);
 151                 GOTO(out, rc);
 152         }
 153
 154         list_for_each_entry(imp_conn, &imp->imp_conn_list, oic_item) {
 155                 if (!obd_uuid_equals(uuid, &imp_conn->oic_uuid))
 156                         continue;
 157                 LASSERT(imp_conn->oic_conn);
 158
 159                 if (imp_conn == imp->imp_conn_current) {
 160                         LASSERT(imp_conn->oic_conn == imp->imp_connection);
 161
 162                         if (imp->imp_state != LUSTRE_IMP_CLOSED &&
 163                             imp->imp_state != LUSTRE_IMP_DISCON) {
 164                                 CERROR("can't remove current connection\n");
 165                                 GOTO(out, rc = -EBUSY);
 166                         }
 167
 168                         ptlrpc_connection_put(imp->imp_connection);
 169                         imp->imp_connection = NULL;
 170
 171                         dlmexp = class_conn2export(&imp->imp_dlm_handle);
 172                         if (dlmexp && dlmexp->exp_connection) {
 173                                 LASSERT(dlmexp->exp_connection ==
 174                                         imp_conn->oic_conn);
 175                                 ptlrpc_connection_put(dlmexp->exp_connection);
 176                                 dlmexp->exp_connection = NULL;
 177                         }
 178
 179                         if (dlmexp != NULL)
 180                                 class_export_put(dlmexp);
 181                 }
 182
 183                 list_del(&imp_conn->oic_item);
 184                 ptlrpc_connection_put(imp_conn->oic_conn);
 185                 OBD_FREE(imp_conn, sizeof(*imp_conn));
 186                 CDEBUG(D_HA, "imp %p@%s: remove connection %s\n",
 187                        imp, imp->imp_obd->obd_name, uuid->uuid);
 188                 rc = 0;
 189                 break;
 190         }
 191 out:
 192         spin_unlock(&imp->imp_lock);
 193         if (rc == -ENOENT)
 194                 CERROR("connection %s not found\n", uuid->uuid);
 195         RETURN(rc);
 196 }
 197 EXPORT_SYMBOL(client_import_del_conn);
 198
 199 /**
 200  * Find conn UUID by peer NID. \a peer is a server NID. This function is used
 201  * to find a conn uuid of \a imp which can reach \a peer.
 202  */
 203 int client_import_find_conn(struct obd_import *imp, lnet_nid_t peer,
 204                             struct obd_uuid *uuid)
 205 {
 206         struct obd_import_conn *conn;
 207         int rc = -ENOENT;
 208         ENTRY;
 209
 210         spin_lock(&imp->imp_lock);
 211         list_for_each_entry(conn, &imp->imp_conn_list, oic_item) {
 212                 /* Check if conn UUID does have this peer NID. */
 213                 if (class_check_uuid(&conn->oic_uuid, peer)) {
 214                         *uuid = conn->oic_uuid;
 215                         rc = 0;
 216                         break;
 217                 }
 218         }
 219         spin_unlock(&imp->imp_lock);
 220         RETURN(rc);
 221 }
 222 EXPORT_SYMBOL(client_import_find_conn);
 223
 224 void client_destroy_import(struct obd_import *imp)
 225 {
 226         /* Drop security policy instance after all RPCs have finished/aborted
 227          * to let all busy contexts be released. */
 228         class_import_get(imp);
 229         class_destroy_import(imp);
 230         sptlrpc_import_sec_put(imp);
 231         class_import_put(imp);
 232 }
 233 EXPORT_SYMBOL(client_destroy_import);
 234
 235 /**
 236  * Check whether or not the OSC is on MDT.
 237  * In the config log,
 238  * osc on MDT
 239  *      setup 0:{fsname}-OSTxxxx-osc[-MDTxxxx] 1:lustre-OST0000_UUID 2:NID
 240  * osc on client
 241  *      setup 0:{fsname}-OSTxxxx-osc 1:lustre-OST0000_UUID 2:NID
 242  *
 243  **/
 244 static int osc_on_mdt(char *obdname)
 245 {
 246         char *ptr;
 247
 248         ptr = strrchr(obdname, '-');
 249         if (ptr == NULL)
 250                 return 0;
 251
 252         if (strncmp(ptr + 1, "MDT", 3) == 0)
 253                 return 1;
 254
 255         return 0;
 256 }
 257
 258 /* Configure an RPC client OBD device.
 259  *
 260  * lcfg parameters:
 261  * 1 - client UUID
 262  * 2 - server UUID
 263  * 3 - inactive-on-startup
 264  * 4 - restrictive net
 265  */
 266 int client_obd_setup(struct obd_device *obddev, struct lustre_cfg *lcfg)
 267 {
 268         struct client_obd *cli = &obddev->u.cli;
 269         struct obd_import *imp;
 270         struct obd_uuid server_uuid;
 271         int rq_portal, rp_portal, connect_op;
 272         char *name = obddev->obd_type->typ_name;
 273         enum ldlm_ns_type ns_type = LDLM_NS_TYPE_UNKNOWN;
 274         char *cli_name = lustre_cfg_buf(lcfg, 0);
 275         struct ptlrpc_connection fake_conn = { .c_self = 0,
 276                                                .c_remote_uuid.uuid[0] = 0 };
 277         int rc;
 278         ENTRY;
 279
 280         /* In a more perfect world, we would hang a ptlrpc_client off of
 281          * obd_type and just use the values from there. */
 282         if (!strcmp(name, LUSTRE_OSC_NAME)) {
 283                 rq_portal = OST_REQUEST_PORTAL;
 284                 rp_portal = OSC_REPLY_PORTAL;
 285                 connect_op = OST_CONNECT;
 286                 cli->cl_sp_me = LUSTRE_SP_CLI;
 287                 cli->cl_sp_to = LUSTRE_SP_OST;
 288                 ns_type = LDLM_NS_TYPE_OSC;
 289         } else if (!strcmp(name, LUSTRE_MDC_NAME) ||
 290                    !strcmp(name, LUSTRE_LWP_NAME)) {
 291                 rq_portal = MDS_REQUEST_PORTAL;
 292                 rp_portal = MDC_REPLY_PORTAL;
 293                 connect_op = MDS_CONNECT;
 294                 if (is_lwp_on_ost(cli_name))
 295                         cli->cl_sp_me = LUSTRE_SP_OST;
 296                 else if (is_lwp_on_mdt(cli_name))
 297                         cli->cl_sp_me = LUSTRE_SP_MDT;
 298                 else
 299                         cli->cl_sp_me = LUSTRE_SP_CLI;
 300                 cli->cl_sp_to = LUSTRE_SP_MDT;
 301                 ns_type = LDLM_NS_TYPE_MDC;
 302         } else if (!strcmp(name, LUSTRE_OSP_NAME)) {
 303                 if (strstr(lustre_cfg_buf(lcfg, 1), "OST") == NULL) {
 304                         /* OSP_on_MDT for other MDTs */
 305                         connect_op = MDS_CONNECT;
 306                         cli->cl_sp_to = LUSTRE_SP_MDT;
 307                         ns_type = LDLM_NS_TYPE_MDC;
 308                         rq_portal = OUT_PORTAL;
 309                 } else {
 310                         /* OSP on MDT for OST */
 311                         connect_op = OST_CONNECT;
 312                         cli->cl_sp_to = LUSTRE_SP_OST;
 313                         ns_type = LDLM_NS_TYPE_OSC;
 314                         rq_portal = OST_REQUEST_PORTAL;
 315                 }
 316                 rp_portal = OSC_REPLY_PORTAL;
 317                 cli->cl_sp_me = LUSTRE_SP_MDT;
 318         } else if (!strcmp(name, LUSTRE_MGC_NAME)) {
 319                 rq_portal = MGS_REQUEST_PORTAL;
 320                 rp_portal = MGC_REPLY_PORTAL;
 321                 connect_op = MGS_CONNECT;
 322                 cli->cl_sp_me = LUSTRE_SP_MGC;
 323                 cli->cl_sp_to = LUSTRE_SP_MGS;
 324                 cli->cl_flvr_mgc.sf_rpc = SPTLRPC_FLVR_INVALID;
 325                 ns_type = LDLM_NS_TYPE_MGC;
 326         } else {
 327                 CERROR("unknown client OBD type \"%s\", can't setup\n",
 328                        name);
 329                 RETURN(-EINVAL);
 330         }
 331
 332         if (LUSTRE_CFG_BUFLEN(lcfg, 1) < 1) {
 333                 CERROR("requires a TARGET UUID\n");
 334                 RETURN(-EINVAL);
 335         }
 336
 337         if (LUSTRE_CFG_BUFLEN(lcfg, 1) > 37) {
 338                 CERROR("client UUID must be less than 38 characters\n");
 339                 RETURN(-EINVAL);
 340         }
 341
 342         if (LUSTRE_CFG_BUFLEN(lcfg, 2) < 1) {
 343                 CERROR("setup requires a SERVER UUID\n");
 344                 RETURN(-EINVAL);
 345         }
 346
 347         if (LUSTRE_CFG_BUFLEN(lcfg, 2) > 37) {
 348                 CERROR("target UUID must be less than 38 characters\n");
 349                 RETURN(-EINVAL);
 350         }
 351
 352         init_rwsem(&cli->cl_sem);
 353         mutex_init(&cli->cl_mgc_mutex);
 354         cli->cl_seq = NULL;
 355         init_rwsem(&cli->cl_seq_rwsem);
 356         cli->cl_conn_count = 0;
 357         memcpy(server_uuid.uuid, lustre_cfg_buf(lcfg, 2),
 358                min_t(unsigned int, LUSTRE_CFG_BUFLEN(lcfg, 2),
 359                      sizeof(server_uuid)));
 360
 361         cli->cl_dirty_pages = 0;
 362         cli->cl_avail_grant = 0;
 363         /* FIXME: Should limit this for the sum of all cl_dirty_max_pages. */
 364         /* cl_dirty_max_pages may be changed at connect time in
 365          * ptlrpc_connect_interpret(). */
 366         client_adjust_max_dirty(cli);
 367         INIT_LIST_HEAD(&cli->cl_cache_waiters);
 368         INIT_LIST_HEAD(&cli->cl_loi_ready_list);
 369         INIT_LIST_HEAD(&cli->cl_loi_hp_ready_list);
 370         INIT_LIST_HEAD(&cli->cl_loi_write_list);
 371         INIT_LIST_HEAD(&cli->cl_loi_read_list);
 372         spin_lock_init(&cli->cl_loi_list_lock);
 373         atomic_set(&cli->cl_pending_w_pages, 0);
 374         atomic_set(&cli->cl_pending_r_pages, 0);
 375         cli->cl_r_in_flight = 0;
 376         cli->cl_w_in_flight = 0;
 377
 378         spin_lock_init(&cli->cl_read_rpc_hist.oh_lock);
 379         spin_lock_init(&cli->cl_write_rpc_hist.oh_lock);
 380         spin_lock_init(&cli->cl_read_page_hist.oh_lock);
 381         spin_lock_init(&cli->cl_write_page_hist.oh_lock);
 382         spin_lock_init(&cli->cl_read_offset_hist.oh_lock);
 383         spin_lock_init(&cli->cl_write_offset_hist.oh_lock);
 384
 385         /* lru for osc. */
 386         INIT_LIST_HEAD(&cli->cl_lru_osc);
 387         atomic_set(&cli->cl_lru_shrinkers, 0);
 388         atomic_long_set(&cli->cl_lru_busy, 0);
 389         atomic_long_set(&cli->cl_lru_in_list, 0);
 390         INIT_LIST_HEAD(&cli->cl_lru_list);
 391         spin_lock_init(&cli->cl_lru_list_lock);
 392         atomic_long_set(&cli->cl_unstable_count, 0);
 393         INIT_LIST_HEAD(&cli->cl_shrink_list);
 394         INIT_LIST_HEAD(&cli->cl_grant_chain);
 395
 396         INIT_LIST_HEAD(&cli->cl_flight_waiters);
 397         cli->cl_rpcs_in_flight = 0;
 398
 399         init_waitqueue_head(&cli->cl_destroy_waitq);
 400         atomic_set(&cli->cl_destroy_in_flight, 0);
 401
 402         cli->cl_supp_cksum_types = OBD_CKSUM_CRC32;
 403 #ifdef ENABLE_CHECKSUM
 404         /* Turn on checksumming by default. */
 405         cli->cl_checksum = 1;
 406         /*
 407          * The supported checksum types will be worked out at connect time
 408          * Set cl_chksum* to CRC32 for now to avoid returning screwed info
 409          * through procfs.
 410          */
 411         cli->cl_cksum_type = cli->cl_supp_cksum_types;
 412 #endif
 413         atomic_set(&cli->cl_resends, OSC_DEFAULT_RESENDS);
 414
 415         /* Set it to possible maximum size. It may be reduced by ocd_brw_size
 416          * from OFD after connecting. */
 417         cli->cl_max_pages_per_rpc = PTLRPC_MAX_BRW_PAGES;
 418
 419         cli->cl_max_short_io_bytes = OBD_MAX_SHORT_IO_BYTES;
 420
 421         /* set cl_chunkbits default value to PAGE_SHIFT,
 422          * it will be updated at OSC connection time. */
 423         cli->cl_chunkbits = PAGE_SHIFT;
 424
 425         if (!strcmp(name, LUSTRE_MDC_NAME)) {
 426                 cli->cl_max_rpcs_in_flight = OBD_MAX_RIF_DEFAULT;
 427         } else if (totalram_pages >> (20 - PAGE_SHIFT) <= 128 /* MB */) {
 428                 cli->cl_max_rpcs_in_flight = 2;
 429         } else if (totalram_pages >> (20 - PAGE_SHIFT) <= 256 /* MB */) {
 430                 cli->cl_max_rpcs_in_flight = 3;
 431         } else if (totalram_pages >> (20 - PAGE_SHIFT) <= 512 /* MB */) {
 432                 cli->cl_max_rpcs_in_flight = 4;
 433         } else {
 434                 if (osc_on_mdt(obddev->obd_name))
 435                         cli->cl_max_rpcs_in_flight = OBD_MAX_RIF_MAX;
 436                 else
 437                         cli->cl_max_rpcs_in_flight = OBD_MAX_RIF_DEFAULT;
 438         }
 439
 440         spin_lock_init(&cli->cl_mod_rpcs_lock);
 441         spin_lock_init(&cli->cl_mod_rpcs_hist.oh_lock);
 442         cli->cl_max_mod_rpcs_in_flight = 0;
 443         cli->cl_mod_rpcs_in_flight = 0;
 444         cli->cl_close_rpcs_in_flight = 0;
 445         init_waitqueue_head(&cli->cl_mod_rpcs_waitq);
 446         cli->cl_mod_tag_bitmap = NULL;
 447
 448         INIT_LIST_HEAD(&cli->cl_chg_dev_linkage);
 449
 450         if (connect_op == MDS_CONNECT) {
 451                 cli->cl_max_mod_rpcs_in_flight = cli->cl_max_rpcs_in_flight - 1;
 452                 OBD_ALLOC(cli->cl_mod_tag_bitmap,
 453                           BITS_TO_LONGS(OBD_MAX_RIF_MAX) * sizeof(long));
 454                 if (cli->cl_mod_tag_bitmap == NULL)
 455                         GOTO(err, rc = -ENOMEM);
 456         }
 457
 458         rc = ldlm_get_ref();
 459         if (rc) {
 460                 CERROR("ldlm_get_ref failed: %d\n", rc);
 461                 GOTO(err, rc);
 462         }
 463
 464         ptlrpc_init_client(rq_portal, rp_portal, name,
 465                            &obddev->obd_ldlm_client);
 466
 467         imp = class_new_import(obddev);
 468         if (imp == NULL)
 469                 GOTO(err_ldlm, rc = -ENOENT);
 470         imp->imp_client = &obddev->obd_ldlm_client;
 471         imp->imp_connect_op = connect_op;
 472         memcpy(cli->cl_target_uuid.uuid, lustre_cfg_buf(lcfg, 1),
 473                LUSTRE_CFG_BUFLEN(lcfg, 1));
 474         class_import_put(imp);
 475
 476         if (lustre_cfg_buf(lcfg, 4)) {
 477                 __u32 refnet = libcfs_str2net(lustre_cfg_string(lcfg, 4));
 478
 479                 if (refnet == LNET_NIDNET(LNET_NID_ANY)) {
 480                         rc = -EINVAL;
 481                         CERROR("%s: bad mount option 'network=%s': rc = %d\n",
 482                                obddev->obd_name, lustre_cfg_string(lcfg, 4),
 483                                rc);
 484                         GOTO(err_import, rc);
 485                 }
 486                 fake_conn.c_self = LNET_MKNID(refnet, 0);
 487                 imp->imp_connection = &fake_conn;
 488         }
 489
 490         rc = client_import_add_conn(imp, &server_uuid, 1);
 491         if (rc) {
 492                 CERROR("can't add initial connection\n");
 493                 GOTO(err_import, rc);
 494         }
 495         imp->imp_connection = NULL;
 496
 497         cli->cl_import = imp;
 498         /* cli->cl_max_mds_easize updated by mdc_init_ea_size() */
 499         cli->cl_max_mds_easize = sizeof(struct lov_mds_md_v3);
 500
 501         if (LUSTRE_CFG_BUFLEN(lcfg, 3) > 0) {
 502                 if (!strcmp(lustre_cfg_string(lcfg, 3), "inactive")) {
 503                         CDEBUG(D_HA, "marking %s %s->%s as inactive\n",
 504                                name, obddev->obd_name,
 505                                cli->cl_target_uuid.uuid);
 506                         spin_lock(&imp->imp_lock);
 507                         imp->imp_deactive = 1;
 508                         spin_unlock(&imp->imp_lock);
 509                 }
 510         }
 511
 512         obddev->obd_namespace = ldlm_namespace_new(obddev, obddev->obd_name,
 513                                                    LDLM_NAMESPACE_CLIENT,
 514                                                    LDLM_NAMESPACE_GREEDY,
 515                                                    ns_type);
 516         if (obddev->obd_namespace == NULL) {
 517                 CERROR("Unable to create client namespace - %s\n",
 518                        obddev->obd_name);
 519                 GOTO(err_import, rc = -ENOMEM);
 520         }
 521
 522         RETURN(rc);
 523
 524 err_import:
 525         class_destroy_import(imp);
 526 err_ldlm:
 527         ldlm_put_ref();
 528 err:
 529         if (cli->cl_mod_tag_bitmap != NULL)
 530                 OBD_FREE(cli->cl_mod_tag_bitmap,
 531                          BITS_TO_LONGS(OBD_MAX_RIF_MAX) * sizeof(long));
 532         cli->cl_mod_tag_bitmap = NULL;
 533         RETURN(rc);
 534
 535 }
 536 EXPORT_SYMBOL(client_obd_setup);
 537
 538 int client_obd_cleanup(struct obd_device *obddev)
 539 {
 540         struct client_obd *cli = &obddev->u.cli;
 541         ENTRY;
 542
 543         ldlm_namespace_free_post(obddev->obd_namespace);
 544         obddev->obd_namespace = NULL;
 545
 546         obd_cleanup_client_import(obddev);
 547         LASSERT(obddev->u.cli.cl_import == NULL);
 548
 549         ldlm_put_ref();
 550
 551         if (cli->cl_mod_tag_bitmap != NULL)
 552                 OBD_FREE(cli->cl_mod_tag_bitmap,
 553                          BITS_TO_LONGS(OBD_MAX_RIF_MAX) * sizeof(long));
 554         cli->cl_mod_tag_bitmap = NULL;
 555
 556         RETURN(0);
 557 }
 558 EXPORT_SYMBOL(client_obd_cleanup);
 559
 560 /* ->o_connect() method for client side (OSC and MDC and MGC) */
 561 int client_connect_import(const struct lu_env *env,
 562                           struct obd_export **exp,
 563                           struct obd_device *obd, struct obd_uuid *cluuid,
 564                           struct obd_connect_data *data, void *localdata)
 565 {
 566         struct client_obd       *cli    = &obd->u.cli;
 567         struct obd_import       *imp    = cli->cl_import;
 568         struct obd_connect_data *ocd;
 569         struct lustre_handle    conn    = { 0 };
 570         int                     rc;
 571         ENTRY;
 572
 573         *exp = NULL;
 574         down_write(&cli->cl_sem);
 575         if (cli->cl_conn_count > 0)
 576                 GOTO(out_sem, rc = -EALREADY);
 577
 578         rc = class_connect(&conn, obd, cluuid);
 579         if (rc)
 580                 GOTO(out_sem, rc);
 581
 582         cli->cl_conn_count++;
 583         *exp = class_conn2export(&conn);
 584
 585         LASSERT(obd->obd_namespace);
 586
 587         imp->imp_dlm_handle = conn;
 588         rc = ptlrpc_init_import(imp);
 589         if (rc != 0)
 590                 GOTO(out_ldlm, rc);
 591
 592         ocd = &imp->imp_connect_data;
 593         if (data) {
 594                 *ocd = *data;
 595                 imp->imp_connect_flags_orig = data->ocd_connect_flags;
 596                 imp->imp_connect_flags2_orig = data->ocd_connect_flags2;
 597         }
 598
 599         rc = ptlrpc_connect_import(imp);
 600         if (rc != 0) {
 601                 LASSERT(imp->imp_state == LUSTRE_IMP_DISCON);
 602                 GOTO(out_ldlm, rc);
 603         }
 604         LASSERT(*exp != NULL && (*exp)->exp_connection);
 605
 606         if (data) {
 607                 LASSERTF((ocd->ocd_connect_flags & data->ocd_connect_flags) ==
 608                          ocd->ocd_connect_flags, "old %#llx, new %#llx\n",
 609                          data->ocd_connect_flags, ocd->ocd_connect_flags);
 610                 data->ocd_connect_flags = ocd->ocd_connect_flags;
 611                 data->ocd_connect_flags2 = ocd->ocd_connect_flags2;
 612         }
 613
 614         ptlrpc_pinger_add_import(imp);
 615
 616         EXIT;
 617
 618         if (rc) {
 619 out_ldlm:
 620                 cli->cl_conn_count--;
 621                 class_disconnect(*exp);
 622                 *exp = NULL;
 623         }
 624 out_sem:
 625         up_write(&cli->cl_sem);
 626
 627         return rc;
 628 }
 629 EXPORT_SYMBOL(client_connect_import);
 630
 631 int client_disconnect_export(struct obd_export *exp)
 632 {
 633         struct obd_device *obd = class_exp2obd(exp);
 634         struct client_obd *cli;
 635         struct obd_import *imp;
 636         int rc = 0, err;
 637         ENTRY;
 638
 639         if (!obd) {
 640                 CERROR("invalid export for disconnect: exp %p cookie %#llx\n",
 641                        exp, exp ? exp->exp_handle.h_cookie : -1);
 642                 RETURN(-EINVAL);
 643         }
 644
 645         cli = &obd->u.cli;
 646         imp = cli->cl_import;
 647
 648         down_write(&cli->cl_sem);
 649         CDEBUG(D_INFO, "disconnect %s - %zu\n", obd->obd_name,
 650                 cli->cl_conn_count);
 651
 652         if (cli->cl_conn_count == 0) {
 653                 CERROR("disconnecting disconnected device (%s)\n",
 654                        obd->obd_name);
 655                 GOTO(out_disconnect, rc = -EINVAL);
 656         }
 657
 658         cli->cl_conn_count--;
 659         if (cli->cl_conn_count != 0)
 660                 GOTO(out_disconnect, rc = 0);
 661
 662         /* Mark import deactivated now, so we don't try to reconnect if any
 663          * of the cleanup RPCs fails (e.g. LDLM cancel, etc).  We don't
 664          * fully deactivate the import, or that would drop all requests. */
 665         spin_lock(&imp->imp_lock);
 666         imp->imp_deactive = 1;
 667         spin_unlock(&imp->imp_lock);
 668
 669         /* Some non-replayable imports (MDS's OSCs) are pinged, so just
 670          * delete it regardless.  (It's safe to delete an import that was
 671          * never added.) */
 672         (void)ptlrpc_pinger_del_import(imp);
 673
 674         if (obd->obd_namespace != NULL) {
 675                 /* obd_force == local only */
 676                 ldlm_cli_cancel_unused(obd->obd_namespace, NULL,
 677                                        obd->obd_force ? LCF_LOCAL : 0, NULL);
 678                 ldlm_namespace_free_prior(obd->obd_namespace, imp, obd->obd_force);
 679         }
 680
 681         /* There's no need to hold sem while disconnecting an import,
 682          * and it may actually cause deadlock in GSS. */
 683         up_write(&cli->cl_sem);
 684         rc = ptlrpc_disconnect_import(imp, 0);
 685         down_write(&cli->cl_sem);
 686
 687         ptlrpc_invalidate_import(imp);
 688
 689         EXIT;
 690
 691 out_disconnect:
 692         /* Use server style - class_disconnect should be always called for
 693          * o_disconnect. */
 694         err = class_disconnect(exp);
 695         if (!rc && err)
 696                 rc = err;
 697
 698         up_write(&cli->cl_sem);
 699
 700         RETURN(rc);
 701 }
 702 EXPORT_SYMBOL(client_disconnect_export);
 703
 704 #ifdef HAVE_SERVER_SUPPORT
 705 int server_disconnect_export(struct obd_export *exp)
 706 {
 707         int rc;
 708         ENTRY;
 709
 710         /* Disconnect early so that clients can't keep using export. */
 711         rc = class_disconnect(exp);
 712         /* Close import to avoid sending any requests. */
 713         if (exp->exp_imp_reverse)
 714                 ptlrpc_cleanup_imp(exp->exp_imp_reverse);
 715
 716         ldlm_bl_thread_wakeup();
 717
 718         /* complete all outstanding replies */
 719         spin_lock(&exp->exp_lock);
 720         while (!list_empty(&exp->exp_outstanding_replies)) {
 721                 struct ptlrpc_reply_state *rs =
 722                         list_entry(exp->exp_outstanding_replies.next,
 723                                        struct ptlrpc_reply_state, rs_exp_list);
 724                 struct ptlrpc_service_part *svcpt = rs->rs_svcpt;
 725
 726                 spin_lock(&svcpt->scp_rep_lock);
 727
 728                 list_del_init(&rs->rs_exp_list);
 729
 730                 spin_lock(&rs->rs_lock);
 731                 /* clear rs_convert_lock to make sure rs is handled and put */
 732                 rs->rs_convert_lock = 0;
 733                 ptlrpc_schedule_difficult_reply(rs);
 734                 spin_unlock(&rs->rs_lock);
 735
 736                 spin_unlock(&svcpt->scp_rep_lock);
 737         }
 738         spin_unlock(&exp->exp_lock);
 739
 740         RETURN(rc);
 741 }
 742 EXPORT_SYMBOL(server_disconnect_export);
 743
 744 /* --------------------------------------------------------------------------
 745  * from old lib/target.c
 746  * -------------------------------------------------------------------------- */
 747
 748 static int target_handle_reconnect(struct lustre_handle *conn,
 749                                    struct obd_export *exp,
 750                                    struct obd_uuid *cluuid)
 751 {
 752         struct obd_device *target;
 753         struct lustre_handle *hdl;
 754         time64_t deadline;
 755         time64_t timeout;
 756         time64_t now;
 757         int rc = 0;
 758
 759         ENTRY;
 760         hdl = &exp->exp_imp_reverse->imp_remote_handle;
 761         if (!exp->exp_connection || !lustre_handle_is_used(hdl)) {
 762                 conn->cookie = exp->exp_handle.h_cookie;
 763                 CDEBUG(D_HA, "connect export for UUID '%s' at %p,"
 764                        " cookie %#llx\n", cluuid->uuid, exp, conn->cookie);
 765                 RETURN(0);
 766         }
 767
 768         target = exp->exp_obd;
 769
 770         /* Might be a re-connect after a partition. */
 771         if (memcmp(&conn->cookie, &hdl->cookie, sizeof conn->cookie)) {
 772                 LCONSOLE_WARN("%s: already connected client %s (at %s) "
 773                               "with handle %#llx. Rejecting client "
 774                               "with the same UUID trying to reconnect "
 775                               "with handle %#llx\n", target->obd_name,
 776                               obd_uuid2str(&exp->exp_client_uuid),
 777                               obd_export_nid2str(exp),
 778                               hdl->cookie, conn->cookie);
 779                 memset(conn, 0, sizeof *conn);
 780                 /* target_handle_connect() treats EALREADY and
 781                  * -EALREADY differently.  -EALREADY is an error
 782                  * (same UUID, different handle). */
 783                 RETURN(-EALREADY);
 784         }
 785
 786         if (!target->obd_recovering) {
 787                 LCONSOLE_WARN("%s: Client %s (at %s) reconnecting\n",
 788                         target->obd_name, obd_uuid2str(&exp->exp_client_uuid),
 789                         obd_export_nid2str(exp));
 790                 GOTO(out_already, rc);
 791         }
 792
 793         now = ktime_get_seconds();
 794         deadline = jiffies_to_msecs(target->obd_recovery_timer.expires) /
 795                    MSEC_PER_SEC;
 796         if (now < deadline) {
 797                 struct target_distribute_txn_data *tdtd;
 798                 int size = 0;
 799                 int count = 0;
 800                 char *buf = NULL;
 801
 802                 timeout = deadline - now;
 803                 tdtd = class_exp2tgt(exp)->lut_tdtd;
 804                 if (tdtd && tdtd->tdtd_show_update_logs_retrievers)
 805                         buf = tdtd->tdtd_show_update_logs_retrievers(
 806                                 tdtd->tdtd_show_retrievers_cbdata,
 807                                 &size, &count);
 808
 809                 if (count > 0)
 810                         LCONSOLE_WARN("%s: Recovery already passed deadline "
 811                                       "%lld:%.02lld. It is due to DNE recovery "
 812                                       "failed/stuck on the %d MDT(s):%s. "
 813                                       "Please wait until all MDTs recovered "
 814                                       "or abort the recovery by force.\n",
 815                                       target->obd_name, timeout / 60,
 816                                       timeout % 60, count,
 817                                       buf ? buf : "unknown (not enough RAM)");
 818                 else
 819                         LCONSOLE_WARN("%s: Recovery already passed deadline "
 820                                       "%lld:%.02lld. If you do not want to wait "
 821                                       "more, please abort the recovery by "
 822                                       "force.\n", target->obd_name,
 823                                       timeout / 60, timeout % 60);
 824
 825                 if (buf != NULL)
 826                         OBD_FREE(buf, size);
 827         } else {
 828                 timeout = now - deadline;
 829                 LCONSOLE_WARN("%s: Recovery already passed deadline"
 830                         " %lld:%.02lld, It is most likely due to DNE"
 831                         " recovery is failed or stuck, please wait a"
 832                         " few more minutes or abort the recovery.\n",
 833                         target->obd_name, timeout / 60, timeout % 60);
 834         }
 835
 836 out_already:
 837         conn->cookie = exp->exp_handle.h_cookie;
 838         /* target_handle_connect() treats EALREADY and
 839          * -EALREADY differently.  EALREADY means we are
 840          * doing a valid reconnect from the same client. */
 841         RETURN(EALREADY);
 842 }
 843
 844 static void
 845 check_and_start_recovery_timer(struct obd_device *obd,
 846                                struct ptlrpc_request *req, int new_client);
 847
 848 /**
 849  * update flags for import during reconnect process
 850  */
 851 static int rev_import_flags_update(struct obd_import *revimp,
 852                                    struct ptlrpc_request *req)
 853 {
 854         int rc;
 855         struct obd_connect_data *data;
 856
 857         data = req_capsule_client_get(&req->rq_pill, &RMF_CONNECT_DATA);
 858
 859         if (data->ocd_connect_flags & OBD_CONNECT_AT)
 860                 revimp->imp_msghdr_flags |= MSGHDR_AT_SUPPORT;
 861         else
 862                 revimp->imp_msghdr_flags &= ~MSGHDR_AT_SUPPORT;
 863
 864         revimp->imp_msghdr_flags |= MSGHDR_CKSUM_INCOMPAT18;
 865
 866         rc = sptlrpc_import_sec_adapt(revimp, req->rq_svc_ctx, &req->rq_flvr);
 867         if (rc) {
 868                 CERROR("%s: cannot get reverse import %s security: rc = %d\n",
 869                         revimp->imp_client->cli_name,
 870                         libcfs_id2str(req->rq_peer), rc);
 871                 return rc;
 872         }
 873
 874         return 0;
 875 }
 876
 877 /**
 878  * Allocate a new reverse import for an export.
 879  *
 880  * \retval -errno in case error hit
 881  * \retval 0 if reverse import correctly init
 882  **/
 883 int rev_import_init(struct obd_export *export)
 884 {
 885         struct obd_device *obd = export->exp_obd;
 886         struct obd_import *revimp;
 887
 888         LASSERT(export->exp_imp_reverse == NULL);
 889
 890         revimp = class_new_import(obd);
 891         if (revimp == NULL)
 892                 return -ENOMEM;
 893
 894         revimp->imp_remote_handle.cookie = 0ULL;
 895         revimp->imp_client = &obd->obd_ldlm_client;
 896         revimp->imp_dlm_fake = 1;
 897
 898         /* it is safe to connect import in new state as no sends possible */
 899         spin_lock(&export->exp_lock);
 900         export->exp_imp_reverse = revimp;
 901         spin_unlock(&export->exp_lock);
 902         class_import_put(revimp);
 903
 904         return 0;
 905 }
 906 EXPORT_SYMBOL(rev_import_init);
 907
 908 /**
 909  * Handle reconnect for an export.
 910  *
 911  * \param exp export to handle reconnect process
 912  * \param req client reconnect request
 913  *
 914  * \retval -rc in case securitfy flavor can't be changed
 915  * \retval 0 in case none problems
 916  */
 917 static int rev_import_reconnect(struct obd_export *exp,
 918                                 struct ptlrpc_request *req)
 919 {
 920         struct obd_import *revimp = exp->exp_imp_reverse;
 921         struct lustre_handle *lh;
 922         int rc;
 923
 924         /* avoid sending a request until import flags are changed */
 925         ptlrpc_import_enter_resend(revimp);
 926
 927         if (revimp->imp_connection != NULL)
 928                 ptlrpc_connection_put(revimp->imp_connection);
 929
 930         /*
 931          * client from recovery don't have a handle so we need to take from
 932          * request. it may produce situation when wrong client connected
 933          * to recovery as we trust a client uuid
 934          */
 935         lh = req_capsule_client_get(&req->rq_pill, &RMF_CONN);
 936         revimp->imp_remote_handle = *lh;
 937
 938         /* unknown versions will be caught in
 939          * ptlrpc_handle_server_req_in->lustre_unpack_msg() */
 940         revimp->imp_msg_magic = req->rq_reqmsg->lm_magic;
 941
 942         revimp->imp_connection = ptlrpc_connection_addref(exp->exp_connection);
 943
 944         rc = rev_import_flags_update(revimp, req);
 945         if (rc != 0) {
 946                 /* it is safe to still be in RECOVERY phase as we are not able
 947                  * to setup correct security flavor so requests are not able to
 948                  * be delivered correctly */
 949                 return rc;
 950         }
 951
 952         /* resend all rpc's via new connection */
 953         return ptlrpc_import_recovery_state_machine(revimp);
 954 }
 955
 956 int target_handle_connect(struct ptlrpc_request *req)
 957 {
 958         struct obd_device *target = NULL;
 959         struct obd_export *export = NULL;
 960         /* connect handle - filled from target_handle_reconnect in
 961          * reconnect case */
 962         struct lustre_handle conn;
 963         struct lustre_handle *tmp;
 964         struct obd_uuid cluuid;
 965         char *str;
 966         int rc = 0;
 967         char *target_start;
 968         int target_len;
 969         bool     mds_conn = false, lw_client = false, initial_conn = false;
 970         bool     mds_mds_conn = false;
 971         bool     new_mds_mds_conn = false;
 972         struct obd_connect_data *data, *tmpdata;
 973         int size, tmpsize;
 974         lnet_nid_t *client_nid = NULL;
 975         ENTRY;
 976
 977         OBD_RACE(OBD_FAIL_TGT_CONN_RACE);
 978
 979         str = req_capsule_client_get(&req->rq_pill, &RMF_TGTUUID);
 980         if (str == NULL) {
 981                 DEBUG_REQ(D_ERROR, req, "bad target UUID for connect");
 982                 GOTO(out, rc = -EINVAL);
 983         }
 984
 985         target = class_dev_by_str(str);
 986         if (!target) {
 987                 deuuidify(str, NULL, &target_start, &target_len);
 988                 LCONSOLE_ERROR_MSG(0x137, "%s: not available for connect "
 989                                    "from %s (no target). If you are running "
 990                                    "an HA pair check that the target is "
 991                                    "mounted on the other server.\n", str,
 992                                    libcfs_nid2str(req->rq_peer.nid));
 993                 GOTO(out, rc = -ENODEV);
 994         }
 995
 996         spin_lock(&target->obd_dev_lock);
 997
 998         target->obd_conn_inprogress++;
 999
1000         if (target->obd_stopping || !target->obd_set_up) {
1001                 spin_unlock(&target->obd_dev_lock);
1002
1003                 deuuidify(str, NULL, &target_start, &target_len);
1004                 LCONSOLE_INFO("%.*s: Not available for connect from %s (%s)\n",
1005                               target_len, target_start,
1006                               libcfs_nid2str(req->rq_peer.nid),
1007                               (target->obd_stopping ?
1008                                "stopping" : "not set up"));
1009                 GOTO(out, rc = -ENODEV);
1010         }
1011
1012         if (target->obd_no_conn) {
1013                 spin_unlock(&target->obd_dev_lock);
1014
1015                 CDEBUG(D_INFO, "%s: Temporarily refusing client connection "
1016                                "from %s\n", target->obd_name,
1017                                libcfs_nid2str(req->rq_peer.nid));
1018                 GOTO(out, rc = -EAGAIN);
1019         }
1020
1021         spin_unlock(&target->obd_dev_lock);
1022
1023         str = req_capsule_client_get(&req->rq_pill, &RMF_CLUUID);
1024         if (str == NULL) {
1025                 DEBUG_REQ(D_ERROR, req, "bad client UUID for connect");
1026                 GOTO(out, rc = -EINVAL);
1027         }
1028
1029         obd_str2uuid(&cluuid, str);
1030
1031         tmp = req_capsule_client_get(&req->rq_pill, &RMF_CONN);
1032         if (tmp == NULL)
1033                 GOTO(out, rc = -EPROTO);
1034
1035         conn = *tmp;
1036
1037         size = req_capsule_get_size(&req->rq_pill, &RMF_CONNECT_DATA,
1038                                     RCL_CLIENT);
1039         data = req_capsule_client_get(&req->rq_pill, &RMF_CONNECT_DATA);
1040         if (!data)
1041                 GOTO(out, rc = -EPROTO);
1042
1043         rc = req_capsule_server_pack(&req->rq_pill);
1044         if (rc)
1045                 GOTO(out, rc);
1046
1047 #if LUSTRE_VERSION_CODE < OBD_OCD_VERSION(3, 0, 53, 0)
1048         /* Don't allow clients to connect that are using old 1.8 format
1049          * protocol conventions (LUSTRE_MSG_MAGIC_v1, !MSGHDR_CKSUM_INCOMPAT18,
1050          * ldlm_flock_policy_wire format, MDT_ATTR_xTIME_SET, etc).  The
1051          * FULL20 flag should be set on all connections since 2.0, but no
1052          * longer affects behaviour.
1053          *
1054          * Later this check will be disabled and the flag can be retired
1055          * completely once interop with 3.0 is no longer needed.
1056          */
1057         if (!(data->ocd_connect_flags & OBD_CONNECT_FULL20))
1058                 GOTO(out, rc = -EPROTO);
1059
1060         /* Don't allow liblustre clients to connect.
1061          * - testing was disabled in v2_2_50_0-61-g6a75d65
1062          * - building was disabled in v2_5_58_0-28-g7277179
1063          * - client code was deleted in v2_6_50_0-101-gcdfbc72,
1064          * - clients were refused connect for version difference > 0.0.1.32  */
1065         if (lustre_msg_get_op_flags(req->rq_reqmsg) & MSG_CONNECT_LIBCLIENT) {
1066                 DEBUG_REQ(D_WARNING, req, "Refusing libclient connection");
1067                 GOTO(out, rc = -EPROTO);
1068         }
1069 #endif
1070
1071         /* Note: lw_client is needed in MDS-MDS failover during update log
1072          * processing, so we needs to allow lw_client to be connected at
1073          * anytime, instead of only the initial connection */
1074         lw_client = (data->ocd_connect_flags & OBD_CONNECT_LIGHTWEIGHT) != 0;
1075
1076         if (lustre_msg_get_op_flags(req->rq_reqmsg) & MSG_CONNECT_INITIAL) {
1077                 initial_conn = true;
1078                 mds_conn = (data->ocd_connect_flags & OBD_CONNECT_MDS) != 0;
1079                 mds_mds_conn = (data->ocd_connect_flags &
1080                                 OBD_CONNECT_MDS_MDS) != 0;
1081
1082                 /* OBD_CONNECT_MNE_SWAB is defined as OBD_CONNECT_MDS_MDS
1083                  * for Imperative Recovery connection from MGC to MGS.
1084                  *
1085                  * Via check OBD_CONNECT_FID, we can distinguish whether
1086                  * the OBD_CONNECT_MDS_MDS/OBD_CONNECT_MNE_SWAB is from
1087                  * MGC or MDT, since MGC does not use OBD_CONNECT_FID.
1088                  */
1089                 if (!lw_client &&
1090                     (data->ocd_connect_flags & OBD_CONNECT_MDS_MDS) &&
1091                     (data->ocd_connect_flags & OBD_CONNECT_FID) &&
1092                     (data->ocd_connect_flags & OBD_CONNECT_VERSION)) {
1093                         __u32 major = OBD_OCD_VERSION_MAJOR(data->ocd_version);
1094                         __u32 minor = OBD_OCD_VERSION_MINOR(data->ocd_version);
1095                         __u32 patch = OBD_OCD_VERSION_PATCH(data->ocd_version);
1096
1097                         /* We do not support the MDT-MDT interoperations with
1098                          * different version MDT because of protocol changes. */
1099                         if (unlikely(major != LUSTRE_MAJOR ||
1100                                      minor != LUSTRE_MINOR ||
1101                                      abs(patch - LUSTRE_PATCH) > 3)) {
1102                                 LCONSOLE_WARN("%s (%u.%u.%u.%u) refused the "
1103                                         "connection from different version MDT "
1104                                         "(%d.%d.%d.%d) %s %s\n",
1105                                         target->obd_name, LUSTRE_MAJOR,
1106                                         LUSTRE_MINOR, LUSTRE_PATCH, LUSTRE_FIX,
1107                                         major, minor, patch,
1108                                         OBD_OCD_VERSION_FIX(data->ocd_version),
1109                                         libcfs_nid2str(req->rq_peer.nid), str);
1110
1111                                 GOTO(out, rc = -EPROTO);
1112                         }
1113                 }
1114         }
1115
1116         /* lctl gets a backstage, all-access pass. */
1117         if (obd_uuid_equals(&cluuid, &target->obd_uuid))
1118                 goto dont_check_exports;
1119
1120         export = cfs_hash_lookup(target->obd_uuid_hash, &cluuid);
1121         if (!export)
1122                 goto no_export;
1123
1124         /* We've found an export in the hash. */
1125
1126         spin_lock(&export->exp_lock);
1127
1128         if (export->exp_connecting) { /* bug 9635, et. al. */
1129                 spin_unlock(&export->exp_lock);
1130                 LCONSOLE_WARN("%s: Export %p already connecting from %s\n",
1131                               export->exp_obd->obd_name, export,
1132                               libcfs_nid2str(req->rq_peer.nid));
1133                 class_export_put(export);
1134                 export = NULL;
1135                 rc = -EALREADY;
1136         } else if ((mds_conn || (lw_client && initial_conn) ||
1137                    data->ocd_connect_flags & OBD_CONNECT_MDS_MDS) &&
1138                    export->exp_connection != NULL) {
1139                 spin_unlock(&export->exp_lock);
1140                 if (req->rq_peer.nid != export->exp_connection->c_peer.nid) {
1141                         /* MDS or LWP reconnected after failover. */
1142                         LCONSOLE_WARN("%s: Received %s connection from "
1143                             "%s, removing former export from %s\n",
1144                             target->obd_name, mds_conn ? "MDS" : "LWP",
1145                             libcfs_nid2str(req->rq_peer.nid),
1146                             libcfs_nid2str(export->exp_connection->c_peer.nid));
1147                 } else {
1148                         /* New MDS connection from the same NID. */
1149                         LCONSOLE_WARN("%s: Received new %s connection from "
1150                                 "%s, removing former export from same NID\n",
1151                                 target->obd_name, mds_conn ? "MDS" : "LWP",
1152                                 libcfs_nid2str(req->rq_peer.nid));
1153                 }
1154
1155                 if (req->rq_peer.nid == export->exp_connection->c_peer.nid &&
1156                     data->ocd_connect_flags & OBD_CONNECT_MDS_MDS) {
1157                         /* Because exports between MDTs will always be
1158                          * kept, let's do not fail such export if they
1159                          * come from the same NID, otherwise it might
1160                          * cause eviction between MDTs, which might
1161                          * cause namespace inconsistency */
1162                         spin_lock(&export->exp_lock);
1163                         export->exp_connecting = 1;
1164                         export->exp_conn_cnt = 0;
1165                         spin_unlock(&export->exp_lock);
1166                         conn.cookie = export->exp_handle.h_cookie;
1167                         rc = EALREADY;
1168                 } else {
1169                         class_fail_export(export);
1170                         class_export_put(export);
1171                         export = NULL;
1172                         rc = 0;
1173                 }
1174         } else if (export->exp_connection != NULL && initial_conn &&
1175                    req->rq_peer.nid != export->exp_connection->c_peer.nid) {
1176                 spin_unlock(&export->exp_lock);
1177                 /* In MDS failover we have static UUID but NID can change. */
1178                 LCONSOLE_WARN("%s: Client %s seen on new nid %s when "
1179                               "existing nid %s is already connected\n",
1180                               target->obd_name, cluuid.uuid,
1181                               libcfs_nid2str(req->rq_peer.nid),
1182                               libcfs_nid2str(
1183                                       export->exp_connection->c_peer.nid));
1184                 rc = -EALREADY;
1185                 class_export_put(export);
1186                 export = NULL;
1187         } else {
1188                 export->exp_connecting = 1;
1189                 spin_unlock(&export->exp_lock);
1190                 LASSERT(export->exp_obd == target);
1191
1192                 rc = target_handle_reconnect(&conn, export, &cluuid);
1193         }
1194
1195         /* If we found an export, we already unlocked. */
1196         if (!export) {
1197 no_export:
1198                 OBD_FAIL_TIMEOUT(OBD_FAIL_TGT_DELAY_CONNECT, 2 * obd_timeout);
1199         } else if (req->rq_export == NULL &&
1200                    atomic_read(&export->exp_rpc_count) > 0) {
1201                 LCONSOLE_WARN("%s: Client %s (at %s) refused connection, "
1202                               "still busy with %d references\n",
1203                               target->obd_name, cluuid.uuid,
1204                               libcfs_nid2str(req->rq_peer.nid),
1205                               atomic_read(&export->exp_refcount));
1206                         GOTO(out, rc = -EBUSY);
1207         } else if (lustre_msg_get_conn_cnt(req->rq_reqmsg) == 1 &&
1208                    rc != EALREADY) {
1209                 if (!strstr(cluuid.uuid, "mdt"))
1210                         LCONSOLE_WARN("%s: Rejecting reconnect from the "
1211                                       "known client %s (at %s) because it "
1212                                       "is indicating it is a new client",
1213                                       target->obd_name, cluuid.uuid,
1214                                       libcfs_nid2str(req->rq_peer.nid));
1215                 GOTO(out, rc = -EALREADY);
1216         } else {
1217                 OBD_FAIL_TIMEOUT(OBD_FAIL_TGT_DELAY_RECONNECT, 2 * obd_timeout);
1218         }
1219
1220         if (rc < 0) {
1221                 GOTO(out, rc);
1222         }
1223
1224         CDEBUG(D_HA, "%s: connection from %s@%s %st%llu exp %p cur %lld last %lld\n",
1225                target->obd_name, cluuid.uuid, libcfs_nid2str(req->rq_peer.nid),
1226                target->obd_recovering ? "recovering/" : "", data->ocd_transno,
1227                export, ktime_get_real_seconds(),
1228                export ? export->exp_last_request_time : 0);
1229
1230         /* If this is the first time a client connects, reset the recovery
1231          * timer. Discard lightweight connections which might be local. */
1232         if (!lw_client && rc == 0 && target->obd_recovering)
1233                 check_and_start_recovery_timer(target, req, export == NULL);
1234
1235         /* We want to handle EALREADY but *not* -EALREADY from
1236          * target_handle_reconnect(), return reconnection state in a flag. */
1237         if (rc == EALREADY) {
1238                 lustre_msg_add_op_flags(req->rq_repmsg, MSG_CONNECT_RECONNECT);
1239                 rc = 0;
1240         } else {
1241                 LASSERT(rc == 0);
1242         }
1243
1244         /* Tell the client if we support replayable requests. */
1245         if (target->obd_replayable)
1246                 lustre_msg_add_op_flags(req->rq_repmsg, MSG_CONNECT_REPLAYABLE);
1247         client_nid = &req->rq_peer.nid;
1248
1249         if (export == NULL) {
1250                 /* allow lightweight connections during recovery */
1251                 /* allow "new" MDT to be connected during recovery, since we
1252                  * need retrieve recovery update records from it */
1253                 if (target->obd_recovering && !lw_client && !mds_mds_conn) {
1254                         time64_t now;
1255                         time64_t t;
1256                         char *msg;
1257                         int c; /* connected */
1258                         int i; /* in progress */
1259                         int k; /* known */
1260                         int s; /* stale/evicted */
1261
1262                         c = atomic_read(&target->obd_connected_clients);
1263                         i = atomic_read(&target->obd_lock_replay_clients);
1264                         k = target->obd_max_recoverable_clients;
1265                         s = target->obd_stale_clients;
1266                         t = jiffies_to_msecs(target->obd_recovery_timer.expires);
1267                         t /= MSEC_PER_SEC;
1268                         now = ktime_get_seconds();
1269                         if (now > t) {
1270                                 t = now - t;
1271                                 msg = "already passed deadline";
1272                         } else {
1273                                 t -= now;
1274                                 msg = "to recover in";
1275                         }
1276
1277                         LCONSOLE_WARN("%s: Denying connection for new client %s"
1278                                       "(at %s), waiting for %d known clients "
1279                                       "(%d recovered, %d in progress, and %d "
1280                                       "evicted) %s %lld:%.02lld\n",
1281                                       target->obd_name, cluuid.uuid,
1282                                       libcfs_nid2str(req->rq_peer.nid), k,
1283                                       c - i, i, s, msg, t / 60, t % 60);
1284                         rc = -EBUSY;
1285                 } else {
1286 dont_check_exports:
1287                         rc = obd_connect(req->rq_svc_thread->t_env,
1288                                          &export, target, &cluuid, data,
1289                                          client_nid);
1290                         if (mds_conn && OBD_FAIL_CHECK(OBD_FAIL_TGT_RCVG_FLAG))
1291                                 lustre_msg_add_op_flags(req->rq_repmsg,
1292                                                         MSG_CONNECT_RECOVERING);
1293                         if (rc == 0) {
1294                                 conn.cookie = export->exp_handle.h_cookie;
1295                                 rc = rev_import_init(export);
1296                         }
1297
1298                         if (mds_mds_conn)
1299                                 new_mds_mds_conn = true;
1300                 }
1301         } else {
1302                 rc = obd_reconnect(req->rq_svc_thread->t_env,
1303                                    export, target, &cluuid, data, client_nid);
1304         }
1305         if (rc)
1306                 GOTO(out, rc);
1307
1308         LASSERT(target->u.obt.obt_magic == OBT_MAGIC);
1309         data->ocd_instance = target->u.obt.obt_instance;
1310
1311         /* Return only the parts of obd_connect_data that we understand, so the
1312          * client knows that we don't understand the rest. */
1313         if (data) {
1314                 tmpsize = req_capsule_get_size(&req->rq_pill, &RMF_CONNECT_DATA,
1315                                                RCL_SERVER);
1316                 tmpdata = req_capsule_server_get(&req->rq_pill,
1317                                                  &RMF_CONNECT_DATA);
1318                 /* Don't use struct assignment here, because the client reply
1319                  * buffer may be smaller/larger than the local struct
1320                  * obd_connect_data. */
1321                 memcpy(tmpdata, data, min(tmpsize, size));
1322         }
1323
1324         /* If the client and the server are the same node, we will already
1325          * have an export that really points to the client's DLM export,
1326          * because we have a shared handles table.
1327          *
1328          * XXX this will go away when shaver stops sending the "connect" handle
1329          * in the real "remote handle" field of the request --phik 24 Apr 2003
1330          */
1331         ptlrpc_request_change_export(req, export);
1332
1333         spin_lock(&export->exp_lock);
1334         if (export->exp_conn_cnt >= lustre_msg_get_conn_cnt(req->rq_reqmsg)) {
1335                 spin_unlock(&export->exp_lock);
1336                 CDEBUG(D_RPCTRACE, "%s: %s already connected at greater "
1337                        "or equal conn_cnt: %d >= %d\n",
1338                        cluuid.uuid, libcfs_nid2str(req->rq_peer.nid),
1339                        export->exp_conn_cnt,
1340                        lustre_msg_get_conn_cnt(req->rq_reqmsg));
1341
1342                 GOTO(out, rc = -EALREADY);
1343         }
1344         LASSERT(lustre_msg_get_conn_cnt(req->rq_reqmsg) > 0);
1345         export->exp_conn_cnt = lustre_msg_get_conn_cnt(req->rq_reqmsg);
1346         spin_unlock(&export->exp_lock);
1347
1348         if (export->exp_connection != NULL) {
1349                 /* Check to see if connection came from another NID. */
1350                 if ((export->exp_connection->c_peer.nid != req->rq_peer.nid) &&
1351                     !hlist_unhashed(&export->exp_nid_hash))
1352                         cfs_hash_del(export->exp_obd->obd_nid_hash,
1353                                      &export->exp_connection->c_peer.nid,
1354                                      &export->exp_nid_hash);
1355
1356                 ptlrpc_connection_put(export->exp_connection);
1357         }
1358
1359         export->exp_connection = ptlrpc_connection_get(req->rq_peer,
1360                                                        req->rq_self,
1361                                                        &cluuid);
1362         if (hlist_unhashed(&export->exp_nid_hash))
1363                 cfs_hash_add(export->exp_obd->obd_nid_hash,
1364                              &export->exp_connection->c_peer.nid,
1365                              &export->exp_nid_hash);
1366
1367         lustre_msg_set_handle(req->rq_repmsg, &conn);
1368
1369         rc = rev_import_reconnect(export, req);
1370         if (rc != 0)
1371                 GOTO(out, rc);
1372
1373         if (target->obd_recovering && !export->exp_in_recovery && !lw_client) {
1374                 int has_transno;
1375                 __u64 transno = data->ocd_transno;
1376
1377                 spin_lock(&export->exp_lock);
1378                 /* possible race with class_disconnect_stale_exports,
1379                  * export may be already in the eviction process */
1380                 if (export->exp_failed) {
1381                         spin_unlock(&export->exp_lock);
1382                         GOTO(out, rc = -ENODEV);
1383                 }
1384                 export->exp_in_recovery = 1;
1385                 export->exp_req_replay_needed = 1;
1386                 export->exp_lock_replay_needed = 1;
1387                 spin_unlock(&export->exp_lock);
1388
1389                 has_transno = !!(lustre_msg_get_op_flags(req->rq_reqmsg) &
1390                                  MSG_CONNECT_TRANSNO);
1391                 if (has_transno && transno == 0)
1392                         CWARN("Connect with zero transno!\n");
1393
1394                 if (has_transno && transno > 0 &&
1395                     transno < target->obd_next_recovery_transno &&
1396                     transno > target->obd_last_committed) {
1397                         /* Another way is to use cmpxchg() to be lock-free. */
1398                         spin_lock(&target->obd_recovery_task_lock);
1399                         if (transno < target->obd_next_recovery_transno)
1400                                 target->obd_next_recovery_transno = transno;
1401                         spin_unlock(&target->obd_recovery_task_lock);
1402                 }
1403
1404                 atomic_inc(&target->obd_req_replay_clients);
1405                 atomic_inc(&target->obd_lock_replay_clients);
1406                 /* Note: MDS-MDS connection is allowed to be connected during
1407                  * recovery, no matter if the exports needs to be recoveried.
1408                  * Because we need retrieve updates logs from all other MDTs.
1409                  * So if the MDS-MDS export is new, obd_max_recoverable_clients
1410                  * also needs to be increased to match other recovery checking
1411                  * condition. */
1412                 if (new_mds_mds_conn)
1413                         target->obd_max_recoverable_clients++;
1414                 if (atomic_inc_return(&target->obd_connected_clients) ==
1415                     target->obd_max_recoverable_clients)
1416                         wake_up(&target->obd_next_transno_waitq);
1417         }
1418
1419         /* Tell the client we're in recovery, when client is involved in it. */
1420         if (target->obd_recovering && !lw_client)
1421                 lustre_msg_add_op_flags(req->rq_repmsg, MSG_CONNECT_RECOVERING);
1422
1423 out:
1424         if (export) {
1425                 spin_lock(&export->exp_lock);
1426                 export->exp_connecting = 0;
1427                 spin_unlock(&export->exp_lock);
1428
1429                 class_export_put(export);
1430         }
1431         if (target != NULL) {
1432                 spin_lock(&target->obd_dev_lock);
1433                 target->obd_conn_inprogress--;
1434                 spin_unlock(&target->obd_dev_lock);
1435                 class_decref(target, "find", current);
1436         }
1437         req->rq_status = rc;
1438         RETURN(rc);
1439 }
1440
1441 int target_handle_disconnect(struct ptlrpc_request *req)
1442 {
1443         int rc;
1444         ENTRY;
1445
1446         rc = req_capsule_server_pack(&req->rq_pill);
1447         if (rc)
1448                 RETURN(rc);
1449
1450         /* Keep the rq_export around so we can send the reply. */
1451         req->rq_status = obd_disconnect(class_export_get(req->rq_export));
1452
1453         RETURN(0);
1454 }
1455
1456 void target_destroy_export(struct obd_export *exp)
1457 {
1458         struct obd_import       *imp = NULL;
1459         /* exports created from last_rcvd data, and "fake"
1460            exports created by lctl don't have an import */
1461         spin_lock(&exp->exp_lock);
1462         if (exp->exp_imp_reverse != NULL) {
1463                 imp = exp->exp_imp_reverse;
1464                 exp->exp_imp_reverse = NULL;
1465         }
1466         spin_unlock(&exp->exp_lock);
1467         if (imp != NULL)
1468                 client_destroy_import(imp);
1469
1470         LASSERT_ATOMIC_ZERO(&exp->exp_locks_count);
1471         LASSERT_ATOMIC_ZERO(&exp->exp_rpc_count);
1472         LASSERT_ATOMIC_ZERO(&exp->exp_cb_count);
1473         LASSERT_ATOMIC_ZERO(&exp->exp_replay_count);
1474 }
1475 EXPORT_SYMBOL(target_destroy_export);
1476
1477 /*
1478  * Recovery functions
1479  */
1480 static void target_request_copy_get(struct ptlrpc_request *req)
1481 {
1482         class_export_rpc_inc(req->rq_export);
1483         LASSERT(list_empty(&req->rq_list));
1484         INIT_LIST_HEAD(&req->rq_replay_list);
1485
1486         /* Increase refcount to keep request in queue. */
1487         atomic_inc(&req->rq_refcount);
1488         /* Let export know it has replays to be handled. */
1489         atomic_inc(&req->rq_export->exp_replay_count);
1490 }
1491
1492 static void target_request_copy_put(struct ptlrpc_request *req)
1493 {
1494         LASSERT(list_empty(&req->rq_replay_list));
1495         LASSERT_ATOMIC_POS(&req->rq_export->exp_replay_count);
1496
1497         atomic_dec(&req->rq_export->exp_replay_count);
1498         class_export_rpc_dec(req->rq_export);
1499         ptlrpc_server_drop_request(req);
1500 }
1501
1502 static int target_exp_enqueue_req_replay(struct ptlrpc_request *req)
1503 {
1504         __u64                  transno = lustre_msg_get_transno(req->rq_reqmsg);
1505         struct obd_export     *exp = req->rq_export;
1506         struct ptlrpc_request *reqiter;
1507         struct ptlrpc_request *dup_req = NULL;
1508         int                    dup = 0;
1509
1510         LASSERT(exp);
1511
1512         spin_lock(&exp->exp_lock);
1513         list_for_each_entry(reqiter, &exp->exp_req_replay_queue,
1514                                 rq_replay_list) {
1515                 if (lustre_msg_get_transno(reqiter->rq_reqmsg) == transno) {
1516                         dup_req = reqiter;
1517                         dup = 1;
1518                         break;
1519                 }
1520         }
1521
1522         if (dup) {
1523                 /* We expect it with RESENT and REPLAY flags. */
1524                 if ((lustre_msg_get_flags(req->rq_reqmsg) &
1525                      (MSG_RESENT | MSG_REPLAY)) != (MSG_RESENT | MSG_REPLAY))
1526                         CERROR("invalid flags %x of resent replay\n",
1527                                lustre_msg_get_flags(req->rq_reqmsg));
1528
1529                 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY) {
1530                         __u32 new_conn;
1531
1532                         new_conn = lustre_msg_get_conn_cnt(req->rq_reqmsg);
1533                         if (new_conn >
1534                             lustre_msg_get_conn_cnt(dup_req->rq_reqmsg))
1535                                 lustre_msg_set_conn_cnt(dup_req->rq_reqmsg,
1536                                                         new_conn);
1537                 }
1538         } else {
1539                 list_add_tail(&req->rq_replay_list,
1540                                   &exp->exp_req_replay_queue);
1541         }
1542
1543         spin_unlock(&exp->exp_lock);
1544         return dup;
1545 }
1546
1547 static void target_exp_dequeue_req_replay(struct ptlrpc_request *req)
1548 {
1549         LASSERT(!list_empty(&req->rq_replay_list));
1550         LASSERT(req->rq_export);
1551
1552         spin_lock(&req->rq_export->exp_lock);
1553         list_del_init(&req->rq_replay_list);
1554         spin_unlock(&req->rq_export->exp_lock);
1555 }
1556
1557 static void target_finish_recovery(struct lu_target *lut)
1558 {
1559         struct obd_device *obd = lut->lut_obd;
1560         ENTRY;
1561
1562         /* Only log a recovery message when recovery has occurred. */
1563         if (obd->obd_recovery_start) {
1564                 time64_t now = ktime_get_real_seconds();
1565                 time64_t elapsed_time;
1566
1567                 elapsed_time = max_t(time64_t, now - obd->obd_recovery_start, 1);
1568                 LCONSOLE_INFO("%s: Recovery over after %lld:%.02lld, of %d clients "
1569                         "%d recovered and %d %s evicted.\n", obd->obd_name,
1570                         (s64)elapsed_time / 60, (s64)elapsed_time % 60,
1571                         obd->obd_max_recoverable_clients,
1572                         atomic_read(&obd->obd_connected_clients),
1573                         obd->obd_stale_clients,
1574                         obd->obd_stale_clients == 1 ? "was" : "were");
1575         }
1576
1577         ldlm_reprocess_recovery_done(obd->obd_namespace);
1578         spin_lock(&obd->obd_recovery_task_lock);
1579         if (!list_empty(&obd->obd_req_replay_queue) ||
1580             !list_empty(&obd->obd_lock_replay_queue) ||
1581             !list_empty(&obd->obd_final_req_queue)) {
1582                 CERROR("%s: Recovery queues ( %s%s%s) are not empty\n",
1583                        obd->obd_name,
1584                        list_empty(&obd->obd_req_replay_queue) ? "" : "req ",
1585                        list_empty(&obd->obd_lock_replay_queue) ? \
1586                                "" : "lock ",
1587                        list_empty(&obd->obd_final_req_queue) ? \
1588                                "" : "final ");
1589                 spin_unlock(&obd->obd_recovery_task_lock);
1590                 LBUG();
1591         }
1592         spin_unlock(&obd->obd_recovery_task_lock);
1593
1594         obd->obd_recovery_end = ktime_get_real_seconds();
1595
1596         /* When recovery finished, cleanup orphans on MDS and OST. */
1597         if (obd->obd_type && OBP(obd, postrecov)) {
1598                 int rc = OBP(obd, postrecov)(obd);
1599
1600                 if (rc < 0)
1601                         LCONSOLE_WARN("%s: Post recovery failed, rc %d\n",
1602                                       obd->obd_name, rc);
1603         }
1604         EXIT;
1605 }
1606
1607 static void abort_req_replay_queue(struct obd_device *obd)
1608 {
1609         struct ptlrpc_request *req, *n;
1610         struct list_head abort_list;
1611
1612         INIT_LIST_HEAD(&abort_list);
1613         spin_lock(&obd->obd_recovery_task_lock);
1614         list_splice_init(&obd->obd_req_replay_queue, &abort_list);
1615         spin_unlock(&obd->obd_recovery_task_lock);
1616         list_for_each_entry_safe(req, n, &abort_list, rq_list) {
1617                 DEBUG_REQ(D_WARNING, req, "aborted:");
1618                 req->rq_status = -ENOTCONN;
1619                 if (ptlrpc_error(req)) {
1620                         DEBUG_REQ(D_ERROR, req,
1621                                   "failed abort_req_reply; skipping");
1622                 }
1623                 target_exp_dequeue_req_replay(req);
1624                 target_request_copy_put(req);
1625         }
1626 }
1627
1628 static void abort_lock_replay_queue(struct obd_device *obd)
1629 {
1630         struct ptlrpc_request *req, *n;
1631         struct list_head abort_list;
1632
1633         INIT_LIST_HEAD(&abort_list);
1634         spin_lock(&obd->obd_recovery_task_lock);
1635         list_splice_init(&obd->obd_lock_replay_queue, &abort_list);
1636         spin_unlock(&obd->obd_recovery_task_lock);
1637         list_for_each_entry_safe(req, n, &abort_list, rq_list) {
1638                 DEBUG_REQ(D_ERROR, req, "aborted:");
1639                 req->rq_status = -ENOTCONN;
1640                 if (ptlrpc_error(req)) {
1641                         DEBUG_REQ(D_ERROR, req,
1642                                   "failed abort_lock_reply; skipping");
1643                 }
1644                 target_request_copy_put(req);
1645         }
1646 }
1647
1648 /* Called from a cleanup function if the device is being cleaned up
1649    forcefully.  The exports should all have been disconnected already,
1650    the only thing left to do is
1651      - clear the recovery flags
1652      - cancel the timer
1653      - free queued requests and replies, but don't send replies
1654    Because the obd_stopping flag is set, no new requests should be received.
1655
1656 */
1657 void target_cleanup_recovery(struct obd_device *obd)
1658 {
1659         struct ptlrpc_request *req, *n;
1660         struct list_head clean_list;
1661
1662         INIT_LIST_HEAD(&clean_list);
1663         spin_lock(&obd->obd_dev_lock);
1664         if (!obd->obd_recovering) {
1665                 spin_unlock(&obd->obd_dev_lock);
1666                 EXIT;
1667                 return;
1668         }
1669         obd->obd_recovering = obd->obd_abort_recovery = 0;
1670         spin_unlock(&obd->obd_dev_lock);
1671
1672         spin_lock(&obd->obd_recovery_task_lock);
1673         target_cancel_recovery_timer(obd);
1674         list_splice_init(&obd->obd_req_replay_queue, &clean_list);
1675         spin_unlock(&obd->obd_recovery_task_lock);
1676
1677         list_for_each_entry_safe(req, n, &clean_list, rq_list) {
1678                 LASSERT(req->rq_reply_state == NULL);
1679                 target_exp_dequeue_req_replay(req);
1680                 target_request_copy_put(req);
1681         }
1682
1683         spin_lock(&obd->obd_recovery_task_lock);
1684         list_splice_init(&obd->obd_lock_replay_queue, &clean_list);
1685         list_splice_init(&obd->obd_final_req_queue, &clean_list);
1686         spin_unlock(&obd->obd_recovery_task_lock);
1687
1688         list_for_each_entry_safe(req, n, &clean_list, rq_list) {
1689                 LASSERT(req->rq_reply_state == NULL);
1690                 target_request_copy_put(req);
1691         }
1692
1693         EXIT;
1694 }
1695 EXPORT_SYMBOL(target_cleanup_recovery);
1696
1697 /* obd_recovery_task_lock should be held */
1698 void target_cancel_recovery_timer(struct obd_device *obd)
1699 {
1700         CDEBUG(D_HA, "%s: cancel recovery timer\n", obd->obd_name);
1701         del_timer(&obd->obd_recovery_timer);
1702 }
1703
1704 static void target_start_recovery_timer(struct obd_device *obd)
1705 {
1706         if (obd->obd_recovery_start != 0)
1707                 return;
1708
1709         spin_lock(&obd->obd_dev_lock);
1710         if (!obd->obd_recovering || obd->obd_abort_recovery) {
1711                 spin_unlock(&obd->obd_dev_lock);
1712                 return;
1713         }
1714
1715         LASSERT(obd->obd_recovery_timeout != 0);
1716
1717         if (obd->obd_recovery_start != 0) {
1718                 spin_unlock(&obd->obd_dev_lock);
1719                 return;
1720         }
1721
1722         mod_timer(&obd->obd_recovery_timer,
1723                   jiffies + cfs_time_seconds(obd->obd_recovery_timeout));
1724         obd->obd_recovery_start = ktime_get_real_seconds();
1725         spin_unlock(&obd->obd_dev_lock);
1726
1727         LCONSOLE_WARN("%s: Will be in recovery for at least %llu:%02llu, or until %d client%s reconnect%s\n",
1728                       obd->obd_name,
1729                       obd->obd_recovery_timeout / 60,
1730                       obd->obd_recovery_timeout % 60,
1731                       obd->obd_max_recoverable_clients,
1732                       (obd->obd_max_recoverable_clients == 1) ? "" : "s",
1733                       (obd->obd_max_recoverable_clients == 1) ? "s": "");
1734 }
1735
1736 /**
1737  * extend recovery window.
1738  *
1739  * if @extend is true, extend recovery window to have @drt remaining at least;
1740  * otherwise, make sure the recovery timeout value is not less than @drt.
1741  */
1742 static void extend_recovery_timer(struct obd_device *obd, time64_t drt,
1743                                   bool extend)
1744 {
1745         time64_t now;
1746         time64_t end;
1747         time64_t left;
1748         time64_t to;
1749
1750         spin_lock(&obd->obd_dev_lock);
1751         if (!obd->obd_recovering || obd->obd_abort_recovery) {
1752                 spin_unlock(&obd->obd_dev_lock);
1753                 return;
1754         }
1755         LASSERT(obd->obd_recovery_start != 0);
1756
1757         now = ktime_get_real_seconds();
1758         to = obd->obd_recovery_timeout;
1759         end = obd->obd_recovery_start + to;
1760         left = end - now;
1761
1762         if (extend && (drt > left)) {
1763                 to += drt - left;
1764         } else if (!extend && (drt > to)) {
1765                 to = drt;
1766         }
1767
1768         if (to > obd->obd_recovery_time_hard) {
1769                 to = obd->obd_recovery_time_hard;
1770                 CWARN("%s: extended recovery timer reaching hard limit: %lld, extend: %d\n",
1771                       obd->obd_name, to, extend);
1772         }
1773
1774         if (obd->obd_recovery_timeout < to) {
1775                 obd->obd_recovery_timeout = to;
1776                 end = obd->obd_recovery_start + to;
1777                 mod_timer(&obd->obd_recovery_timer,
1778                           jiffies + cfs_time_seconds(end - now));
1779         }
1780         spin_unlock(&obd->obd_dev_lock);
1781
1782         CDEBUG(D_HA, "%s: recovery timer will expire in %lld seconds\n",
1783                 obd->obd_name, (s64)(end - now));
1784 }
1785
1786 /* Reset the timer with each new client connection */
1787 /*
1788  * This timer is actually reconnect_timer, which is for making sure
1789  * the total recovery window is at least as big as my reconnect
1790  * attempt timing. So the initial recovery time_out will be set to
1791  * OBD_RECOVERY_FACTOR * obd_timeout. If the timeout coming
1792  * from client is bigger than this, then the recovery time_out will
1793  * be extended to make sure the client could be reconnected, in the
1794  * process, the timeout from the new client should be ignored.
1795  */
1796
1797 static void
1798 check_and_start_recovery_timer(struct obd_device *obd,
1799                                struct ptlrpc_request *req,
1800                                int new_client)
1801 {
1802         time64_t service_time = lustre_msg_get_service_time(req->rq_reqmsg);
1803         struct obd_device_target *obt = &obd->u.obt;
1804
1805         if (!new_client && service_time)
1806                 /* Teach server about old server's estimates, as first guess
1807                  * at how long new requests will take. */
1808                 at_measured(&req->rq_rqbd->rqbd_svcpt->scp_at_estimate,
1809                             service_time);
1810
1811         target_start_recovery_timer(obd);
1812
1813         /* Convert the service time to RPC timeout,
1814          * and reuse service_time to limit stack usage.
1815          */
1816         service_time = at_est2timeout(service_time);
1817
1818         if (OBD_FAIL_CHECK(OBD_FAIL_TGT_SLUGGISH_NET) &&
1819             service_time < at_extra)
1820                 service_time = at_extra;
1821
1822         /* We expect other clients to timeout within service_time, then try
1823          * to reconnect, then try the failover server.  The max delay between
1824          * connect attempts is SWITCH_MAX + SWITCH_INC + INITIAL. */
1825         service_time += 2 * INITIAL_CONNECT_TIMEOUT;
1826
1827         LASSERT(obt->obt_magic == OBT_MAGIC);
1828         service_time += 2 * (CONNECTION_SWITCH_MAX + CONNECTION_SWITCH_INC);
1829         if (service_time > obd->obd_recovery_timeout && !new_client)
1830                 extend_recovery_timer(obd, service_time, false);
1831 }
1832
1833 /** Health checking routines */
1834 static inline int exp_connect_healthy(struct obd_export *exp)
1835 {
1836         return (exp->exp_in_recovery);
1837 }
1838
1839 /** if export done req_replay or has replay in queue */
1840 static inline int exp_req_replay_healthy(struct obd_export *exp)
1841 {
1842         return (!exp->exp_req_replay_needed ||
1843                 atomic_read(&exp->exp_replay_count) > 0);
1844 }
1845
1846
1847 static inline int exp_req_replay_healthy_or_from_mdt(struct obd_export *exp)
1848 {
1849         return (exp_connect_flags(exp) & OBD_CONNECT_MDS_MDS) ||
1850                exp_req_replay_healthy(exp);
1851 }
1852
1853 /** if export done lock_replay or has replay in queue */
1854 static inline int exp_lock_replay_healthy(struct obd_export *exp)
1855 {
1856         return (!exp->exp_lock_replay_needed ||
1857                 atomic_read(&exp->exp_replay_count) > 0);
1858 }
1859
1860 static inline int exp_vbr_healthy(struct obd_export *exp)
1861 {
1862         return (!exp->exp_vbr_failed);
1863 }
1864
1865 static inline int exp_finished(struct obd_export *exp)
1866 {
1867         return (exp->exp_in_recovery && !exp->exp_lock_replay_needed);
1868 }
1869
1870 static inline int exp_finished_or_from_mdt(struct obd_export *exp)
1871 {
1872         return (exp_connect_flags(exp) & OBD_CONNECT_MDS_MDS) ||
1873                 exp_finished(exp);
1874 }
1875
1876 static int check_for_next_transno(struct lu_target *lut)
1877 {
1878         struct ptlrpc_request *req = NULL;
1879         struct obd_device *obd = lut->lut_obd;
1880         struct target_distribute_txn_data *tdtd = lut->lut_tdtd;
1881         int wake_up = 0, connected, completed, queue_len;
1882         __u64 req_transno = 0;
1883         __u64 update_transno = 0;
1884         __u64 next_transno = 0;
1885         ENTRY;
1886
1887         spin_lock(&obd->obd_recovery_task_lock);
1888         if (!list_empty(&obd->obd_req_replay_queue)) {
1889                 req = list_entry(obd->obd_req_replay_queue.next,
1890                                      struct ptlrpc_request, rq_list);
1891                 req_transno = lustre_msg_get_transno(req->rq_reqmsg);
1892         }
1893
1894         if (tdtd != NULL)
1895                 update_transno = distribute_txn_get_next_transno(tdtd);
1896
1897         connected = atomic_read(&obd->obd_connected_clients);
1898         completed = connected - atomic_read(&obd->obd_req_replay_clients);
1899         queue_len = obd->obd_requests_queued_for_recovery;
1900         next_transno = obd->obd_next_recovery_transno;
1901
1902         CDEBUG(D_HA, "max: %d, connected: %d, completed: %d, queue_len: %d, "
1903                "req_transno: %llu, next_transno: %llu\n",
1904                obd->obd_max_recoverable_clients, connected, completed,
1905                queue_len, req_transno, next_transno);
1906
1907         if (obd->obd_abort_recovery) {
1908                 CDEBUG(D_HA, "waking for aborted recovery\n");
1909                 wake_up = 1;
1910         } else if (obd->obd_recovery_expired) {
1911                 CDEBUG(D_HA, "waking for expired recovery\n");
1912                 wake_up = 1;
1913         } else if (tdtd != NULL && req != NULL &&
1914                    is_req_replayed_by_update(req)) {
1915                 LASSERTF(req_transno < next_transno, "req_transno %llu"
1916                          "next_transno%llu\n", req_transno, next_transno);
1917                 CDEBUG(D_HA, "waking for duplicate req (%llu)\n",
1918                        req_transno);
1919                 wake_up = 1;
1920         } else if (req_transno == next_transno ||
1921                    (update_transno != 0 && update_transno <= next_transno)) {
1922                 CDEBUG(D_HA, "waking for next (%lld)\n", next_transno);
1923                 wake_up = 1;
1924         } else if (queue_len > 0 &&
1925                    queue_len == atomic_read(&obd->obd_req_replay_clients)) {
1926                 /** handle gaps occured due to lost reply or VBR */
1927                 LASSERTF(req_transno >= next_transno,
1928                          "req_transno: %llu, next_transno: %llu\n",
1929                          req_transno, next_transno);
1930                 CDEBUG(D_HA,
1931                        "%s: waking for gap in transno, VBR is %s (skip: "
1932                        "%lld, ql: %d, comp: %d, conn: %d, next: %lld"
1933                        ", next_update %lld last_committed: %lld)\n",
1934                        obd->obd_name, obd->obd_version_recov ? "ON" : "OFF",
1935                        next_transno, queue_len, completed, connected,
1936                        req_transno, update_transno, obd->obd_last_committed);
1937                 obd->obd_next_recovery_transno = req_transno;
1938                 wake_up = 1;
1939         } else if (atomic_read(&obd->obd_req_replay_clients) == 0) {
1940                 CDEBUG(D_HA, "waking for completed recovery\n");
1941                 wake_up = 1;
1942         } else if (OBD_FAIL_CHECK(OBD_FAIL_MDS_RECOVERY_ACCEPTS_GAPS)) {
1943                 CDEBUG(D_HA, "accepting transno gaps is explicitly allowed"
1944                        " by fail_lock, waking up (%lld)\n", next_transno);
1945                 obd->obd_next_recovery_transno = req_transno;
1946                 wake_up = 1;
1947         }
1948         spin_unlock(&obd->obd_recovery_task_lock);
1949         return wake_up;
1950 }
1951
1952 static int check_for_next_lock(struct lu_target *lut)
1953 {
1954         struct obd_device *obd = lut->lut_obd;
1955         int wake_up = 0;
1956
1957         spin_lock(&obd->obd_recovery_task_lock);
1958         if (!list_empty(&obd->obd_lock_replay_queue)) {
1959                 CDEBUG(D_HA, "waking for next lock\n");
1960                 wake_up = 1;
1961         } else if (atomic_read(&obd->obd_lock_replay_clients) == 0) {
1962                 CDEBUG(D_HA, "waking for completed lock replay\n");
1963                 wake_up = 1;
1964         } else if (obd->obd_abort_recovery) {
1965                 CDEBUG(D_HA, "waking for aborted recovery\n");
1966                 wake_up = 1;
1967         } else if (obd->obd_recovery_expired) {
1968                 CDEBUG(D_HA, "waking for expired recovery\n");
1969                 wake_up = 1;
1970         }
1971         spin_unlock(&obd->obd_recovery_task_lock);
1972
1973         return wake_up;
1974 }
1975
1976 /**
1977  * wait for recovery events,
1978  * check its status with help of check_routine
1979  * evict dead clients via health_check
1980  */
1981 static int target_recovery_overseer(struct lu_target *lut,
1982                                     int (*check_routine)(struct lu_target *),
1983                                     int (*health_check)(struct obd_export *))
1984 {
1985         struct obd_device       *obd = lut->lut_obd;
1986         struct target_distribute_txn_data *tdtd;
1987         time64_t last = 0;
1988         time64_t now;
1989 repeat:
1990         if (obd->obd_recovering && obd->obd_recovery_start == 0) {
1991                 now = ktime_get_seconds();
1992                 if (now - last > 600) {
1993                         LCONSOLE_INFO("%s: in recovery but waiting for "
1994                                       "the first client to connect\n",
1995                                       obd->obd_name);
1996                         last = now;
1997                 }
1998         }
1999         if (obd->obd_recovery_start != 0 && ktime_get_real_seconds() >=
2000               (obd->obd_recovery_start + obd->obd_recovery_time_hard)) {
2001                 __u64 next_update_transno = 0;
2002
2003                 /* Only abort the recovery if there are no update recovery
2004                  * left in the queue */
2005                 spin_lock(&obd->obd_recovery_task_lock);
2006                 if (lut->lut_tdtd != NULL) {
2007                         next_update_transno =
2008                                 distribute_txn_get_next_transno(lut->lut_tdtd);
2009
2010                         tdtd = lut->lut_tdtd;
2011                         /* If next_update_transno == 0, it probably because
2012                          * updatelog retrieve threads did not get any records
2013                          * yet, let's wait those threads stopped */
2014                         if (next_update_transno == 0) {
2015                                 struct l_wait_info lwi = { 0 };
2016
2017                                 l_wait_event(tdtd->tdtd_recovery_threads_waitq,
2018                                        atomic_read(
2019                                        &tdtd->tdtd_recovery_threads_count) == 0,
2020                                        &lwi);
2021
2022                                 next_update_transno =
2023                                         distribute_txn_get_next_transno(
2024                                                                 lut->lut_tdtd);
2025                         }
2026                 }
2027
2028                 if (next_update_transno != 0 && !obd->obd_abort_recovery) {
2029                         obd->obd_next_recovery_transno = next_update_transno;
2030                         spin_unlock(&obd->obd_recovery_task_lock);
2031                         /* Disconnect unfinished exports from clients, and
2032                          * keep connection from MDT to make sure the update
2033                          * recovery will still keep trying until some one
2034                          * manually abort the recovery */
2035                         class_disconnect_stale_exports(obd,
2036                                                 exp_finished_or_from_mdt);
2037                         /* Abort all of replay and replay lock req from
2038                          * clients */
2039                         abort_req_replay_queue(obd);
2040                         abort_lock_replay_queue(obd);
2041                         CDEBUG(D_HA, "%s: there are still update replay (%#llx"
2042                                ")in the queue.\n", obd->obd_name,
2043                                next_update_transno);
2044                 } else {
2045                         obd->obd_abort_recovery = 1;
2046                         spin_unlock(&obd->obd_recovery_task_lock);
2047                         CWARN("%s recovery is aborted by hard timeout\n",
2048                               obd->obd_name);
2049                 }
2050         }
2051
2052         while (wait_event_timeout(obd->obd_next_transno_waitq,
2053                                   check_routine(lut),
2054                                   msecs_to_jiffies(60 * MSEC_PER_SEC)) == 0)
2055                 /* wait indefinitely for event, but don't trigger watchdog */;
2056
2057         if (obd->obd_abort_recovery) {
2058                 CWARN("recovery is aborted, evict exports in recovery\n");
2059                 if (lut->lut_tdtd != NULL) {
2060                         struct l_wait_info lwi = { 0 };
2061
2062                         tdtd = lut->lut_tdtd;
2063                         /* Let's wait all of the update log recovery thread
2064                          * finished */
2065                         l_wait_event(tdtd->tdtd_recovery_threads_waitq,
2066                          atomic_read(&tdtd->tdtd_recovery_threads_count) == 0,
2067                              &lwi);
2068                         /* Then abort the update recovery list */
2069                         dtrq_list_destroy(lut->lut_tdtd);
2070                 }
2071
2072                 /** evict exports which didn't finish recovery yet */
2073                 class_disconnect_stale_exports(obd, exp_finished);
2074                 return 1;
2075         } else if (obd->obd_recovery_expired) {
2076                 obd->obd_recovery_expired = 0;
2077                 /** If some clients died being recovered, evict them */
2078                 LCONSOLE_WARN("%s: recovery is timed out, "
2079                               "evict stale exports\n", obd->obd_name);
2080                 /** evict cexports with no replay in queue, they are stalled */
2081                 class_disconnect_stale_exports(obd, health_check);
2082
2083                 /** continue with VBR */
2084                 spin_lock(&obd->obd_dev_lock);
2085                 obd->obd_version_recov = 1;
2086                 spin_unlock(&obd->obd_dev_lock);
2087                 /**
2088                  * reset timer, recovery will proceed with versions now,
2089                  * timeout is set just to handle reconnection delays
2090                  */
2091                 extend_recovery_timer(obd, RECONNECT_DELAY_MAX, true);
2092                 /** Wait for recovery events again, after evicting bad clients */
2093                 goto repeat;
2094         }
2095         return 0;
2096 }
2097
2098 static struct ptlrpc_request *target_next_replay_lock(struct lu_target *lut)
2099 {
2100         struct obd_device       *obd = lut->lut_obd;
2101         struct ptlrpc_request *req = NULL;
2102
2103         CDEBUG(D_HA, "Waiting for lock\n");
2104         if (target_recovery_overseer(lut, check_for_next_lock,
2105                                      exp_lock_replay_healthy))
2106                 abort_lock_replay_queue(obd);
2107
2108         spin_lock(&obd->obd_recovery_task_lock);
2109         if (!list_empty(&obd->obd_lock_replay_queue)) {
2110                 req = list_entry(obd->obd_lock_replay_queue.next,
2111                                      struct ptlrpc_request, rq_list);
2112                 list_del_init(&req->rq_list);
2113                 spin_unlock(&obd->obd_recovery_task_lock);
2114         } else {
2115                 spin_unlock(&obd->obd_recovery_task_lock);
2116                 LASSERT(list_empty(&obd->obd_lock_replay_queue));
2117                 LASSERT(atomic_read(&obd->obd_lock_replay_clients) == 0);
2118                 /** evict exports failed VBR */
2119                 class_disconnect_stale_exports(obd, exp_vbr_healthy);
2120         }
2121         return req;
2122 }
2123
2124 static struct ptlrpc_request *target_next_final_ping(struct obd_device *obd)
2125 {
2126         struct ptlrpc_request *req = NULL;
2127
2128         spin_lock(&obd->obd_recovery_task_lock);
2129         if (!list_empty(&obd->obd_final_req_queue)) {
2130                 req = list_entry(obd->obd_final_req_queue.next,
2131                                      struct ptlrpc_request, rq_list);
2132                 list_del_init(&req->rq_list);
2133                 spin_unlock(&obd->obd_recovery_task_lock);
2134                 if (req->rq_export->exp_in_recovery) {
2135                         spin_lock(&req->rq_export->exp_lock);
2136                         req->rq_export->exp_in_recovery = 0;
2137                         spin_unlock(&req->rq_export->exp_lock);
2138                 }
2139         } else {
2140                 spin_unlock(&obd->obd_recovery_task_lock);
2141         }
2142         return req;
2143 }
2144
2145 static void handle_recovery_req(struct ptlrpc_thread *thread,
2146                                 struct ptlrpc_request *req,
2147                                 svc_handler_t handler)
2148 {
2149         ENTRY;
2150
2151         /**
2152          * export can be evicted during recovery, no need to handle replays for
2153          * it after that, discard such request silently
2154          */
2155         if (req->rq_export->exp_disconnected)
2156                 RETURN_EXIT;
2157
2158         req->rq_session.lc_thread = thread;
2159         req->rq_svc_thread = thread;
2160         req->rq_svc_thread->t_env->le_ses = &req->rq_session;
2161
2162         /* thread context */
2163         lu_context_enter(&thread->t_env->le_ctx);
2164         (void)handler(req);
2165         lu_context_exit(&thread->t_env->le_ctx);
2166
2167         /* don't reset timer for final stage */
2168         if (!exp_finished(req->rq_export)) {
2169                 time64_t to = obd_timeout;
2170
2171                 /**
2172                  * Add request timeout to the recovery time so next request from
2173                  * this client may come in recovery time
2174                  */
2175                 if (!AT_OFF) {
2176                         struct ptlrpc_service_part *svcpt;
2177
2178                         svcpt = req->rq_rqbd->rqbd_svcpt;
2179                         /* If the server sent early reply for this request,
2180                          * the client will recalculate the timeout according to
2181                          * current server estimate service time, so we will
2182                          * use the maxium timeout here for waiting the client
2183                          * sending the next req */
2184                         to = max((int)at_est2timeout(
2185                                  at_get(&svcpt->scp_at_estimate)),
2186                                  (int)lustre_msg_get_timeout(req->rq_reqmsg));
2187                         /* Add 2 net_latency, one for balance rq_deadline
2188                          * (see ptl_send_rpc), one for resend the req to server,
2189                          * Note: client will pack net_latency in replay req
2190                          * (see ptlrpc_replay_req) */
2191                         to += 2 * lustre_msg_get_service_time(req->rq_reqmsg);
2192                 }
2193                 extend_recovery_timer(class_exp2obd(req->rq_export), to, true);
2194         }
2195         EXIT;
2196 }
2197
2198 /** Checking routines for recovery */
2199 static int check_for_recovery_ready(struct lu_target *lut)
2200 {
2201         struct obd_device *obd = lut->lut_obd;
2202         unsigned int clnts = atomic_read(&obd->obd_connected_clients);
2203
2204         CDEBUG(D_HA, "connected %d stale %d max_recoverable_clients %d"
2205                " abort %d expired %d\n", clnts, obd->obd_stale_clients,
2206                obd->obd_max_recoverable_clients, obd->obd_abort_recovery,
2207                obd->obd_recovery_expired);
2208
2209         if (!obd->obd_abort_recovery && !obd->obd_recovery_expired) {
2210                 LASSERT(clnts <= obd->obd_max_recoverable_clients);
2211                 if (clnts + obd->obd_stale_clients <
2212                     obd->obd_max_recoverable_clients)
2213                         return 0;
2214         }
2215
2216         if (lut->lut_tdtd != NULL) {
2217                 if (!lut->lut_tdtd->tdtd_replay_ready &&
2218                     !obd->obd_abort_recovery) {
2219                         /* Let's extend recovery timer, in case the recovery
2220                          * timer expired, and some clients got evicted */
2221                         extend_recovery_timer(obd, obd->obd_recovery_timeout,
2222                                               true);
2223                         CDEBUG(D_HA, "%s update recovery is not ready, extend recovery %llu\n",
2224                                obd->obd_name, obd->obd_recovery_timeout);
2225                         return 0;
2226                 }
2227         }
2228
2229         return 1;
2230 }
2231
2232 enum {
2233         REQUEST_RECOVERY = 1,
2234         UPDATE_RECOVERY = 2,
2235 };
2236
2237 static __u64 get_next_replay_req_transno(struct obd_device *obd)
2238 {
2239         __u64 transno = 0;
2240
2241         if (!list_empty(&obd->obd_req_replay_queue)) {
2242                 struct ptlrpc_request *req;
2243
2244                 req = list_entry(obd->obd_req_replay_queue.next,
2245                                  struct ptlrpc_request, rq_list);
2246                 transno = lustre_msg_get_transno(req->rq_reqmsg);
2247         }
2248
2249         return transno;
2250 }
2251
2252 static __u64 get_next_transno(struct lu_target *lut, int *type)
2253 {
2254         struct obd_device *obd = lut->lut_obd;
2255         struct target_distribute_txn_data *tdtd = lut->lut_tdtd;
2256         __u64 transno = 0;
2257         __u64 update_transno;
2258         ENTRY;
2259
2260         transno = get_next_replay_req_transno(obd);
2261         if (type != NULL)
2262                 *type = REQUEST_RECOVERY;
2263
2264         if (tdtd == NULL)
2265                 RETURN(transno);
2266
2267         update_transno = distribute_txn_get_next_transno(tdtd);
2268         if (transno == 0 || (transno >= update_transno &&
2269                              update_transno != 0)) {
2270                 transno = update_transno;
2271                 if (type != NULL)
2272                         *type = UPDATE_RECOVERY;
2273         }
2274
2275         RETURN(transno);
2276 }
2277
2278 /**
2279  * drop duplicate replay request
2280  *
2281  * Because the operation has been replayed by update recovery, the request
2282  * with the same transno will be dropped and also notify the client to send
2283  * next replay request.
2284  *
2285  * \param[in] env       execution environment
2286  * \param[in] obd       failover obd device
2287  * \param[in] req       request to be dropped
2288  */
2289 static void drop_duplicate_replay_req(struct lu_env *env,
2290                                       struct obd_device *obd,
2291                                       struct ptlrpc_request *req)
2292 {
2293         DEBUG_REQ(D_HA, req, "remove t%lld from %s because of duplicate"
2294                   " update records are found.\n",
2295                   lustre_msg_get_transno(req->rq_reqmsg),
2296                   libcfs_nid2str(req->rq_peer.nid));
2297
2298         /* Right now, only for MDS reint operation update replay and
2299          * normal request replay can have the same transno */
2300         if (lustre_msg_get_opc(req->rq_reqmsg) == MDS_REINT) {
2301                 req_capsule_set(&req->rq_pill, &RQF_MDS_REINT);
2302                 req->rq_status = req_capsule_server_pack(&req->rq_pill);
2303                 if (likely(req->rq_export))
2304                         target_committed_to_req(req);
2305                 lustre_msg_set_transno(req->rq_repmsg, req->rq_transno);
2306                 target_send_reply(req, req->rq_status, 0);
2307         } else {
2308                 DEBUG_REQ(D_ERROR, req, "wrong opc" "from %s\n",
2309                 libcfs_nid2str(req->rq_peer.nid));
2310         }
2311         target_exp_dequeue_req_replay(req);
2312         target_request_copy_put(req);
2313         obd->obd_replayed_requests++;
2314 }
2315
2316 static void replay_request_or_update(struct lu_env *env,
2317                                      struct lu_target *lut,
2318                                      struct target_recovery_data *trd,
2319                                      struct ptlrpc_thread *thread)
2320 {
2321         struct obd_device *obd = lut->lut_obd;
2322         struct ptlrpc_request *req = NULL;
2323         int                     type;
2324         __u64                   transno;
2325         ENTRY;
2326
2327         CDEBUG(D_HA, "Waiting for transno %lld\n",
2328                obd->obd_next_recovery_transno);
2329
2330         /* Replay all of request and update by transno */
2331         do {
2332                 struct target_distribute_txn_data *tdtd = lut->lut_tdtd;
2333
2334                 CFS_FAIL_TIMEOUT(OBD_FAIL_TGT_REPLAY_DELAY2, cfs_fail_val);
2335
2336                 /** It is needed to extend recovery window above
2337                  *  recovery_time_soft. Extending is possible only in the
2338                  *  end of recovery window (see more details in
2339                  *  handle_recovery_req()).
2340                  */
2341                 CFS_FAIL_TIMEOUT_MS(OBD_FAIL_TGT_REPLAY_DELAY, 300);
2342
2343                 if (target_recovery_overseer(lut, check_for_next_transno,
2344                                         exp_req_replay_healthy_or_from_mdt)) {
2345                         abort_req_replay_queue(obd);
2346                         abort_lock_replay_queue(obd);
2347                         goto abort;
2348                 }
2349
2350                 spin_lock(&obd->obd_recovery_task_lock);
2351                 transno = get_next_transno(lut, &type);
2352                 if (type == REQUEST_RECOVERY && transno != 0) {
2353                         /* Drop replay request from client side, if the
2354                          * replay has been executed by update with the
2355                          * same transno */
2356                         req = list_entry(obd->obd_req_replay_queue.next,
2357                                         struct ptlrpc_request, rq_list);
2358
2359                         list_del_init(&req->rq_list);
2360                         obd->obd_requests_queued_for_recovery--;
2361                         spin_unlock(&obd->obd_recovery_task_lock);
2362
2363                         /* Let's check if the request has been redone by
2364                          * update replay */
2365                         if (is_req_replayed_by_update(req)) {
2366                                 struct distribute_txn_replay_req *dtrq;
2367
2368                                 dtrq = distribute_txn_lookup_finish_list(tdtd,
2369                                                                    req->rq_xid);
2370                                 LASSERT(dtrq != NULL);
2371                                 spin_lock(&tdtd->tdtd_replay_list_lock);
2372                                 list_del_init(&dtrq->dtrq_list);
2373                                 spin_unlock(&tdtd->tdtd_replay_list_lock);
2374                                 dtrq_destroy(dtrq);
2375
2376                                 drop_duplicate_replay_req(env, obd, req);
2377
2378                                 continue;
2379                         }
2380
2381                         LASSERT(trd->trd_processing_task == current_pid());
2382                         DEBUG_REQ(D_HA, req, "processing t%lld from %s",
2383                                   lustre_msg_get_transno(req->rq_reqmsg),
2384                                   libcfs_nid2str(req->rq_peer.nid));
2385
2386                         handle_recovery_req(thread, req,
2387                                             trd->trd_recovery_handler);
2388                         /**
2389                          * bz18031: increase next_recovery_transno before
2390                          * target_request_copy_put() will drop exp_rpc reference
2391                          */
2392                         spin_lock(&obd->obd_recovery_task_lock);
2393                         obd->obd_next_recovery_transno++;
2394                         spin_unlock(&obd->obd_recovery_task_lock);
2395                         target_exp_dequeue_req_replay(req);
2396                         target_request_copy_put(req);
2397                         obd->obd_replayed_requests++;
2398                 } else if (type == UPDATE_RECOVERY && transno != 0) {
2399                         struct distribute_txn_replay_req *dtrq;
2400                         int rc;
2401
2402                         spin_unlock(&obd->obd_recovery_task_lock);
2403
2404                         LASSERT(tdtd != NULL);
2405                         dtrq = distribute_txn_get_next_req(tdtd);
2406                         lu_context_enter(&thread->t_env->le_ctx);
2407                         rc = tdtd->tdtd_replay_handler(env, tdtd, dtrq);
2408                         lu_context_exit(&thread->t_env->le_ctx);
2409                         extend_recovery_timer(obd, obd_timeout, true);
2410
2411                         if (rc == 0 && dtrq->dtrq_xid != 0) {
2412                                 CDEBUG(D_HA, "Move x%llu t%llu"
2413                                        " to finish list\n", dtrq->dtrq_xid,
2414                                        dtrq->dtrq_master_transno);
2415
2416                                 /* Add it to the replay finish list */
2417                                 spin_lock(&tdtd->tdtd_replay_list_lock);
2418                                 list_add(&dtrq->dtrq_list,
2419                                          &tdtd->tdtd_replay_finish_list);
2420                                 spin_unlock(&tdtd->tdtd_replay_list_lock);
2421
2422                                 spin_lock(&obd->obd_recovery_task_lock);
2423                                 if (transno == obd->obd_next_recovery_transno)
2424                                         obd->obd_next_recovery_transno++;
2425                                 else if (transno >
2426                                          obd->obd_next_recovery_transno)
2427                                         obd->obd_next_recovery_transno =
2428                                                                 transno + 1;
2429                                 spin_unlock(&obd->obd_recovery_task_lock);
2430                         } else {
2431                                 dtrq_destroy(dtrq);
2432                         }
2433                 } else {
2434                         spin_unlock(&obd->obd_recovery_task_lock);
2435 abort:
2436                         LASSERT(list_empty(&obd->obd_req_replay_queue));
2437                         LASSERT(atomic_read(&obd->obd_req_replay_clients) == 0);
2438                         /** evict exports failed VBR */
2439                         class_disconnect_stale_exports(obd, exp_vbr_healthy);
2440                         break;
2441                 }
2442         } while (1);
2443 }
2444
2445 static int target_recovery_thread(void *arg)
2446 {
2447         struct lu_target *lut = arg;
2448         struct obd_device *obd = lut->lut_obd;
2449         struct ptlrpc_request *req;
2450         struct target_recovery_data *trd = &obd->obd_recovery_data;
2451         unsigned long delta;
2452         struct lu_env *env;
2453         struct ptlrpc_thread *thread = NULL;
2454         int rc = 0;
2455         ENTRY;
2456
2457         unshare_fs_struct();
2458         OBD_ALLOC_PTR(thread);
2459         if (thread == NULL)
2460                 RETURN(-ENOMEM);
2461
2462         OBD_ALLOC_PTR(env);
2463         if (env == NULL) {
2464                 OBD_FREE_PTR(thread);
2465                 RETURN(-ENOMEM);
2466         }
2467
2468         rc = lu_context_init(&env->le_ctx, LCT_MD_THREAD | LCT_DT_THREAD);
2469         if (rc) {
2470                 OBD_FREE_PTR(thread);
2471                 OBD_FREE_PTR(env);
2472                 RETURN(rc);
2473         }
2474
2475         thread->t_env = env;
2476         thread->t_id = -1; /* force filter_iobuf_get/put to use local buffers */
2477         env->le_ctx.lc_thread = thread;
2478         tgt_io_thread_init(thread); /* init thread_big_cache for IO requests */
2479
2480         CDEBUG(D_HA, "%s: started recovery thread pid %d\n", obd->obd_name,
2481                current_pid());
2482         trd->trd_processing_task = current_pid();
2483
2484         spin_lock(&obd->obd_dev_lock);
2485         obd->obd_recovering = 1;
2486         spin_unlock(&obd->obd_dev_lock);
2487         complete(&trd->trd_starting);
2488
2489         /* first of all, we have to know the first transno to replay */
2490         if (target_recovery_overseer(lut, check_for_recovery_ready,
2491                                      exp_connect_healthy)) {
2492                 abort_req_replay_queue(obd);
2493                 abort_lock_replay_queue(obd);
2494                 if (lut->lut_tdtd != NULL)
2495                         dtrq_list_destroy(lut->lut_tdtd);
2496         }
2497
2498         /* next stage: replay requests or update */
2499         delta = jiffies;
2500         CDEBUG(D_INFO, "1: request replay stage - %d clients from t%llu\n",
2501                atomic_read(&obd->obd_req_replay_clients),
2502                obd->obd_next_recovery_transno);
2503         replay_request_or_update(env, lut, trd, thread);
2504
2505         /**
2506          * The second stage: replay locks
2507          */
2508         CDEBUG(D_INFO, "2: lock replay stage - %d clients\n",
2509                atomic_read(&obd->obd_lock_replay_clients));
2510         while ((req = target_next_replay_lock(lut))) {
2511                 LASSERT(trd->trd_processing_task == current_pid());
2512                 DEBUG_REQ(D_HA, req, "processing lock from %s: ",
2513                           libcfs_nid2str(req->rq_peer.nid));
2514                 handle_recovery_req(thread, req,
2515                                     trd->trd_recovery_handler);
2516                 target_request_copy_put(req);
2517                 obd->obd_replayed_locks++;
2518         }
2519
2520         /**
2521          * The third stage: reply on final pings, at this moment all clients
2522          * must have request in final queue
2523          */
2524         CFS_FAIL_TIMEOUT(OBD_FAIL_TGT_REPLAY_RECONNECT, cfs_fail_val);
2525         CDEBUG(D_INFO, "3: final stage - process recovery completion pings\n");
2526         /** Update server last boot epoch */
2527         tgt_boot_epoch_update(lut);
2528         /* We drop recoverying flag to forward all new requests
2529          * to regular mds_handle() since now */
2530         spin_lock(&obd->obd_dev_lock);
2531         obd->obd_recovering = obd->obd_abort_recovery = 0;
2532         spin_unlock(&obd->obd_dev_lock);
2533         spin_lock(&obd->obd_recovery_task_lock);
2534         target_cancel_recovery_timer(obd);
2535         spin_unlock(&obd->obd_recovery_task_lock);
2536         while ((req = target_next_final_ping(obd))) {
2537                 LASSERT(trd->trd_processing_task == current_pid());
2538                 DEBUG_REQ(D_HA, req, "processing final ping from %s: ",
2539                           libcfs_nid2str(req->rq_peer.nid));
2540                 handle_recovery_req(thread, req,
2541                                     trd->trd_recovery_handler);
2542                 /* Because the waiting client can not send ping to server,
2543                  * so we need refresh the last_request_time, to avoid the
2544                  * export is being evicted */
2545                 ptlrpc_update_export_timer(req->rq_export, 0);
2546                 target_request_copy_put(req);
2547         }
2548
2549         delta = jiffies_to_msecs(jiffies - delta) / MSEC_PER_SEC;
2550         CDEBUG(D_INFO,"4: recovery completed in %lus - %d/%d reqs/locks\n",
2551               delta, obd->obd_replayed_requests, obd->obd_replayed_locks);
2552         if (delta > OBD_RECOVERY_TIME_SOFT) {
2553                 CWARN("too long recovery - read logs\n");
2554                 libcfs_debug_dumplog();
2555         }
2556
2557         target_finish_recovery(lut);
2558
2559         lu_context_fini(&env->le_ctx);
2560         trd->trd_processing_task = 0;
2561         complete(&trd->trd_finishing);
2562
2563         tgt_io_thread_done(thread);
2564         OBD_FREE_PTR(thread);
2565         OBD_FREE_PTR(env);
2566         RETURN(rc);
2567 }
2568
2569 static int target_start_recovery_thread(struct lu_target *lut,
2570                                         svc_handler_t handler)
2571 {
2572         struct obd_device *obd = lut->lut_obd;
2573         int rc = 0;
2574         struct target_recovery_data *trd = &obd->obd_recovery_data;
2575         int index;
2576
2577         memset(trd, 0, sizeof(*trd));
2578         init_completion(&trd->trd_starting);
2579         init_completion(&trd->trd_finishing);
2580         trd->trd_recovery_handler = handler;
2581
2582         rc = server_name2index(obd->obd_name, &index, NULL);
2583         if (rc < 0)
2584                 return rc;
2585
2586         if (!IS_ERR(kthread_run(target_recovery_thread,
2587                                 lut, "tgt_recover_%d", index))) {
2588                 wait_for_completion(&trd->trd_starting);
2589                 LASSERT(obd->obd_recovering != 0);
2590         } else {
2591                 rc = -ECHILD;
2592         }
2593
2594         return rc;
2595 }
2596
2597 void target_stop_recovery_thread(struct obd_device *obd)
2598 {
2599         if (obd->obd_recovery_data.trd_processing_task > 0) {
2600                 struct target_recovery_data *trd = &obd->obd_recovery_data;
2601                 /** recovery can be done but postrecovery is not yet */
2602                 spin_lock(&obd->obd_dev_lock);
2603                 if (obd->obd_recovering) {
2604                         CERROR("%s: Aborting recovery\n", obd->obd_name);
2605                         obd->obd_abort_recovery = 1;
2606                         wake_up(&obd->obd_next_transno_waitq);
2607                 }
2608                 spin_unlock(&obd->obd_dev_lock);
2609                 wait_for_completion(&trd->trd_finishing);
2610         }
2611 }
2612 EXPORT_SYMBOL(target_stop_recovery_thread);
2613
2614 void target_recovery_fini(struct obd_device *obd)
2615 {
2616         class_disconnect_exports(obd);
2617         target_stop_recovery_thread(obd);
2618         target_cleanup_recovery(obd);
2619 }
2620 EXPORT_SYMBOL(target_recovery_fini);
2621
2622 static void target_recovery_expired(cfs_timer_cb_arg_t data)
2623 {
2624         struct obd_device *obd = cfs_from_timer(obd, data, obd_recovery_timer);
2625         CDEBUG(D_HA, "%s: recovery timed out; %d clients are still in recovery"
2626                " after %llus (%d clients connected)\n",
2627                obd->obd_name, atomic_read(&obd->obd_lock_replay_clients),
2628                (s64)(ktime_get_real_seconds() - obd->obd_recovery_start),
2629                atomic_read(&obd->obd_connected_clients));
2630
2631         obd->obd_recovery_expired = 1;
2632         wake_up(&obd->obd_next_transno_waitq);
2633 }
2634
2635 void target_recovery_init(struct lu_target *lut, svc_handler_t handler)
2636 {
2637         struct obd_device *obd = lut->lut_obd;
2638
2639         if (lut->lut_bottom->dd_rdonly)
2640                 return;
2641
2642         if (obd->obd_max_recoverable_clients == 0) {
2643                 /** Update server last boot epoch */
2644                 tgt_boot_epoch_update(lut);
2645                 return;
2646         }
2647
2648         CDEBUG(D_HA, "RECOVERY: service %s, %d recoverable clients, "
2649                "last_transno %llu\n", obd->obd_name,
2650                obd->obd_max_recoverable_clients, obd->obd_last_committed);
2651         LASSERT(obd->obd_stopping == 0);
2652         obd->obd_next_recovery_transno = obd->obd_last_committed + 1;
2653         obd->obd_recovery_start = 0;
2654         obd->obd_recovery_end = 0;
2655
2656         cfs_timer_setup(&obd->obd_recovery_timer, target_recovery_expired,
2657                         (unsigned long)obd, 0);
2658         target_start_recovery_thread(lut, handler);
2659 }
2660 EXPORT_SYMBOL(target_recovery_init);
2661
2662 static int target_process_req_flags(struct obd_device *obd,
2663                                     struct ptlrpc_request *req)
2664 {
2665         struct obd_export *exp = req->rq_export;
2666         LASSERT(exp != NULL);
2667         if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REQ_REPLAY_DONE) {
2668                 /* client declares he's ready to replay locks */
2669                 spin_lock(&exp->exp_lock);
2670                 if (exp->exp_req_replay_needed) {
2671                         exp->exp_req_replay_needed = 0;
2672                         spin_unlock(&exp->exp_lock);
2673
2674                         LASSERT_ATOMIC_POS(&obd->obd_req_replay_clients);
2675                         atomic_dec(&obd->obd_req_replay_clients);
2676                 } else {
2677                         spin_unlock(&exp->exp_lock);
2678                 }
2679         }
2680         if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_LOCK_REPLAY_DONE) {
2681                 /* client declares he's ready to complete recovery
2682                  * so, we put the request on th final queue */
2683                 spin_lock(&exp->exp_lock);
2684                 if (exp->exp_lock_replay_needed) {
2685                         exp->exp_lock_replay_needed = 0;
2686                         spin_unlock(&exp->exp_lock);
2687
2688                         LASSERT_ATOMIC_POS(&obd->obd_lock_replay_clients);
2689                         atomic_dec(&obd->obd_lock_replay_clients);
2690                 } else {
2691                         spin_unlock(&exp->exp_lock);
2692                 }
2693         }
2694         return 0;
2695 }
2696
2697 int target_queue_recovery_request(struct ptlrpc_request *req,
2698                                   struct obd_device *obd)
2699 {
2700         __u64 transno = lustre_msg_get_transno(req->rq_reqmsg);
2701         struct ptlrpc_request *reqiter;
2702         int inserted = 0;
2703         ENTRY;
2704
2705         if (obd->obd_recovery_data.trd_processing_task == current_pid()) {
2706                 /* Processing the queue right now, don't re-add. */
2707                 RETURN(1);
2708         }
2709
2710         target_process_req_flags(obd, req);
2711
2712         if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_LOCK_REPLAY_DONE) {
2713                 if (unlikely(OBD_FAIL_CHECK(OBD_FAIL_TGT_RECOVERY_REQ_RACE))) {
2714                         if (cfs_fail_val == 1) {
2715                                 cfs_race_state = 1;
2716                                 cfs_fail_val = 0;
2717                                 wake_up(&cfs_race_waitq);
2718
2719                                 set_current_state(TASK_INTERRUPTIBLE);
2720                                 schedule_timeout(cfs_time_seconds(1));
2721                         }
2722                 }
2723
2724                 /* client declares he's ready to complete recovery
2725                  * so, we put the request on th final queue */
2726                 target_request_copy_get(req);
2727                 DEBUG_REQ(D_HA, req, "queue final req");
2728                 wake_up(&obd->obd_next_transno_waitq);
2729                 spin_lock(&obd->obd_recovery_task_lock);
2730                 if (obd->obd_recovering) {
2731                         struct ptlrpc_request *tmp;
2732                         struct ptlrpc_request *duplicate = NULL;
2733
2734                         if (likely(!req->rq_export->exp_replay_done)) {
2735                                 req->rq_export->exp_replay_done = 1;
2736                                 list_add_tail(&req->rq_list,
2737                                               &obd->obd_final_req_queue);
2738                                 spin_unlock(&obd->obd_recovery_task_lock);
2739                                 RETURN(0);
2740                         }
2741
2742                         /* XXX O(n), but only happens if final ping is
2743                          * timed out, probably reorganize the list as
2744                          * a hash list later */
2745                         list_for_each_entry_safe(reqiter, tmp,
2746                                                  &obd->obd_final_req_queue,
2747                                                  rq_list) {
2748                                 if (reqiter->rq_export == req->rq_export) {
2749                                         list_del_init(&reqiter->rq_list);
2750                                         duplicate = reqiter;
2751                                         break;
2752                                 }
2753                         }
2754
2755                         list_add_tail(&req->rq_list,
2756                                       &obd->obd_final_req_queue);
2757                         req->rq_export->exp_replay_done = 1;
2758                         spin_unlock(&obd->obd_recovery_task_lock);
2759
2760                         if (duplicate != NULL) {
2761                                 DEBUG_REQ(D_HA, duplicate,
2762                                           "put prev final req\n");
2763                                 target_request_copy_put(duplicate);
2764                         }
2765                         RETURN(0);
2766                 } else {
2767                         spin_unlock(&obd->obd_recovery_task_lock);
2768                         target_request_copy_put(req);
2769                         RETURN(obd->obd_stopping ? -ENOTCONN : 1);
2770                 }
2771         }
2772         if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REQ_REPLAY_DONE) {
2773                 /* client declares he's ready to replay locks */
2774                 target_request_copy_get(req);
2775                 DEBUG_REQ(D_HA, req, "queue lock replay req");
2776                 wake_up(&obd->obd_next_transno_waitq);
2777                 spin_lock(&obd->obd_recovery_task_lock);
2778                 LASSERT(obd->obd_recovering);
2779                 /* usually due to recovery abort */
2780                 if (!req->rq_export->exp_in_recovery) {
2781                         spin_unlock(&obd->obd_recovery_task_lock);
2782                         target_request_copy_put(req);
2783                         RETURN(-ENOTCONN);
2784                 }
2785                 LASSERT(req->rq_export->exp_lock_replay_needed);
2786                 list_add_tail(&req->rq_list, &obd->obd_lock_replay_queue);
2787                 spin_unlock(&obd->obd_recovery_task_lock);
2788                 RETURN(0);
2789         }
2790
2791         /* CAVEAT EMPTOR: The incoming request message has been swabbed
2792          * (i.e. buflens etc are in my own byte order), but type-dependent
2793          * buffers (eg mdt_body, ost_body etc) have NOT been swabbed. */
2794
2795         if (!transno) {
2796                 INIT_LIST_HEAD(&req->rq_list);
2797                 DEBUG_REQ(D_HA, req, "not queueing");
2798                 RETURN(1);
2799         }
2800
2801         /* If we're processing the queue, we want don't want to queue this
2802          * message.
2803          *
2804          * Also, if this request has a transno less than the one we're waiting
2805          * for, we should process it now.  It could (and currently always will)
2806          * be an open request for a descriptor that was opened some time ago.
2807          *
2808          * Also, a resent, replayed request that has already been
2809          * handled will pass through here and be processed immediately.
2810          */
2811         CDEBUG(D_HA, "Next recovery transno: %llu"
2812                ", current: %llu, replaying\n",
2813                obd->obd_next_recovery_transno, transno);
2814
2815         /* If the request has been replayed by update replay, then sends this
2816          * request to the recovery thread (replay_request_or_update()), where
2817          * it will be handled */
2818         spin_lock(&obd->obd_recovery_task_lock);
2819         if (transno < obd->obd_next_recovery_transno &&
2820             !is_req_replayed_by_update(req)) {
2821                 /* Processing the queue right now, don't re-add. */
2822                 LASSERT(list_empty(&req->rq_list));
2823                 spin_unlock(&obd->obd_recovery_task_lock);
2824                 RETURN(1);
2825         }
2826         spin_unlock(&obd->obd_recovery_task_lock);
2827
2828         if (OBD_FAIL_CHECK(OBD_FAIL_TGT_REPLAY_DROP))
2829                 RETURN(0);
2830
2831         target_request_copy_get(req);
2832         if (!req->rq_export->exp_in_recovery) {
2833                 target_request_copy_put(req);
2834                 RETURN(-ENOTCONN);
2835         }
2836         LASSERT(req->rq_export->exp_req_replay_needed);
2837
2838         if (target_exp_enqueue_req_replay(req)) {
2839                 DEBUG_REQ(D_ERROR, req, "dropping resent queued req");
2840                 target_request_copy_put(req);
2841                 RETURN(0);
2842         }
2843
2844         /* XXX O(n^2) */
2845         spin_lock(&obd->obd_recovery_task_lock);
2846         LASSERT(obd->obd_recovering);
2847         list_for_each_entry(reqiter, &obd->obd_req_replay_queue, rq_list) {
2848                 if (lustre_msg_get_transno(reqiter->rq_reqmsg) > transno) {
2849                         list_add_tail(&req->rq_list, &reqiter->rq_list);
2850                         inserted = 1;
2851                         goto added;
2852                 }
2853
2854                 if (unlikely(lustre_msg_get_transno(reqiter->rq_reqmsg) ==
2855                              transno)) {
2856                         DEBUG_REQ(D_ERROR, req, "dropping replay: transno "
2857                                   "has been claimed by another client");
2858                         spin_unlock(&obd->obd_recovery_task_lock);
2859                         target_exp_dequeue_req_replay(req);
2860                         target_request_copy_put(req);
2861                         RETURN(0);
2862                 }
2863         }
2864 added:
2865         if (!inserted)
2866                 list_add_tail(&req->rq_list, &obd->obd_req_replay_queue);
2867
2868         obd->obd_requests_queued_for_recovery++;
2869         spin_unlock(&obd->obd_recovery_task_lock);
2870         wake_up(&obd->obd_next_transno_waitq);
2871         RETURN(0);
2872 }
2873
2874 int target_handle_ping(struct ptlrpc_request *req)
2875 {
2876         obd_ping(req->rq_svc_thread->t_env, req->rq_export);
2877         return req_capsule_server_pack(&req->rq_pill);
2878 }
2879
2880 void target_committed_to_req(struct ptlrpc_request *req)
2881 {
2882         struct obd_export *exp = req->rq_export;
2883
2884         if (!exp->exp_obd->obd_no_transno && req->rq_repmsg != NULL)
2885                 lustre_msg_set_last_committed(req->rq_repmsg,
2886                                               exp->exp_last_committed);
2887         else
2888                 DEBUG_REQ(D_IOCTL, req, "not sending last_committed update (%d/"
2889                           "%d)", exp->exp_obd->obd_no_transno,
2890                           req->rq_repmsg == NULL);
2891
2892         CDEBUG(D_INFO, "last_committed %llu, transno %llu, xid %llu\n",
2893                exp->exp_last_committed, req->rq_transno, req->rq_xid);
2894 }
2895
2896 #endif /* HAVE_SERVER_SUPPORT */
2897
2898 /**
2899  * Packs current SLV and Limit into \a req.
2900  */
2901 int target_pack_pool_reply(struct ptlrpc_request *req)
2902 {
2903         struct obd_device *obd;
2904         ENTRY;
2905
2906         /* Check that we still have all structures alive as this may
2907          * be some late RPC at shutdown time. */
2908         if (unlikely(!req->rq_export || !req->rq_export->exp_obd ||
2909                      !exp_connect_lru_resize(req->rq_export))) {
2910                 lustre_msg_set_slv(req->rq_repmsg, 0);
2911                 lustre_msg_set_limit(req->rq_repmsg, 0);
2912                 RETURN(0);
2913         }
2914
2915         /* OBD is alive here as export is alive, which we checked above. */
2916         obd = req->rq_export->exp_obd;
2917
2918         read_lock(&obd->obd_pool_lock);
2919         lustre_msg_set_slv(req->rq_repmsg, obd->obd_pool_slv);
2920         lustre_msg_set_limit(req->rq_repmsg, obd->obd_pool_limit);
2921         read_unlock(&obd->obd_pool_lock);
2922
2923         RETURN(0);
2924 }
2925
2926 static int target_send_reply_msg(struct ptlrpc_request *req,
2927                                  int rc, int fail_id)
2928 {
2929         if (OBD_FAIL_CHECK_ORSET(fail_id & ~OBD_FAIL_ONCE, OBD_FAIL_ONCE)) {
2930                 DEBUG_REQ(D_ERROR, req, "dropping reply");
2931                 return -ECOMM;
2932         }
2933         /* We can have a null rq_reqmsg in the event of bad signature or
2934          * no context when unwrapping */
2935         if (req->rq_reqmsg &&
2936             unlikely(lustre_msg_get_opc(req->rq_reqmsg) == MDS_REINT &&
2937             OBD_FAIL_CHECK(OBD_FAIL_MDS_REINT_MULTI_NET_REP)))
2938                 return -ECOMM;
2939
2940         if (unlikely(rc)) {
2941                 DEBUG_REQ(D_NET, req, "processing error (%d)", rc);
2942                 req->rq_status = rc;
2943                 return ptlrpc_send_error(req, 1);
2944         } else {
2945                 DEBUG_REQ(D_NET, req, "sending reply");
2946         }
2947
2948         return ptlrpc_send_reply(req, PTLRPC_REPLY_MAYBE_DIFFICULT);
2949 }
2950
2951 void target_send_reply(struct ptlrpc_request *req, int rc, int fail_id)
2952 {
2953         struct ptlrpc_service_part *svcpt;
2954         int                        netrc;
2955         struct ptlrpc_reply_state *rs;
2956         struct obd_export         *exp;
2957         ENTRY;
2958
2959         if (req->rq_no_reply) {
2960                 EXIT;
2961                 return;
2962         }
2963
2964         svcpt = req->rq_rqbd->rqbd_svcpt;
2965         rs = req->rq_reply_state;
2966         if (rs == NULL || !rs->rs_difficult) {
2967                 /* no notifiers */
2968                 target_send_reply_msg (req, rc, fail_id);
2969                 EXIT;
2970                 return;
2971         }
2972
2973         /* must be an export if locks saved */
2974         LASSERT(req->rq_export != NULL);
2975         /* req/reply consistent */
2976         LASSERT(rs->rs_svcpt == svcpt);
2977
2978         /* "fresh" reply */
2979         LASSERT(!rs->rs_scheduled);
2980         LASSERT(!rs->rs_scheduled_ever);
2981         LASSERT(!rs->rs_handled);
2982         LASSERT(!rs->rs_on_net);
2983         LASSERT(rs->rs_export == NULL);
2984         LASSERT(list_empty(&rs->rs_obd_list));
2985         LASSERT(list_empty(&rs->rs_exp_list));
2986
2987         exp = class_export_get(req->rq_export);
2988
2989         /* disable reply scheduling while I'm setting up */
2990         rs->rs_scheduled = 1;
2991         rs->rs_on_net    = 1;
2992         rs->rs_xid       = req->rq_xid;
2993         rs->rs_transno   = req->rq_transno;
2994         rs->rs_export    = exp;
2995         rs->rs_opc       = lustre_msg_get_opc(req->rq_reqmsg);
2996
2997         spin_lock(&exp->exp_uncommitted_replies_lock);
2998         CDEBUG(D_NET, "rs transno = %llu, last committed = %llu\n",
2999                rs->rs_transno, exp->exp_last_committed);
3000         if (rs->rs_transno > exp->exp_last_committed) {
3001                 /* not committed already */
3002                 list_add_tail(&rs->rs_obd_list,
3003                                   &exp->exp_uncommitted_replies);
3004         }
3005         spin_unlock(&exp->exp_uncommitted_replies_lock);
3006
3007         spin_lock(&exp->exp_lock);
3008         list_add_tail(&rs->rs_exp_list, &exp->exp_outstanding_replies);
3009         spin_unlock(&exp->exp_lock);
3010
3011         netrc = target_send_reply_msg(req, rc, fail_id);
3012
3013         spin_lock(&svcpt->scp_rep_lock);
3014
3015         atomic_inc(&svcpt->scp_nreps_difficult);
3016
3017         if (netrc != 0) {
3018                 /* error sending: reply is off the net.  Also we need +1
3019                  * reply ref until ptlrpc_handle_rs() is done
3020                  * with the reply state (if the send was successful, there
3021                  * would have been +1 ref for the net, which
3022                  * reply_out_callback leaves alone) */
3023                 rs->rs_on_net = 0;
3024                 ptlrpc_rs_addref(rs);
3025         }
3026
3027         spin_lock(&rs->rs_lock);
3028         if (rs->rs_transno <= exp->exp_last_committed ||
3029             (!rs->rs_on_net && !rs->rs_no_ack) ||
3030             list_empty(&rs->rs_exp_list) ||     /* completed already */
3031             list_empty(&rs->rs_obd_list)) {
3032                 CDEBUG(D_HA, "Schedule reply immediately\n");
3033                 ptlrpc_dispatch_difficult_reply(rs);
3034         } else {
3035                 list_add(&rs->rs_list, &svcpt->scp_rep_active);
3036                 rs->rs_scheduled = 0;   /* allow notifier to schedule */
3037         }
3038         spin_unlock(&rs->rs_lock);
3039         spin_unlock(&svcpt->scp_rep_lock);
3040         EXIT;
3041 }
3042
3043 enum ldlm_mode lck_compat_array[] = {
3044         [LCK_EX]    = LCK_COMPAT_EX,
3045         [LCK_PW]    = LCK_COMPAT_PW,
3046         [LCK_PR]    = LCK_COMPAT_PR,
3047         [LCK_CW]    = LCK_COMPAT_CW,
3048         [LCK_CR]    = LCK_COMPAT_CR,
3049         [LCK_NL]    = LCK_COMPAT_NL,
3050         [LCK_GROUP] = LCK_COMPAT_GROUP,
3051         [LCK_COS]   = LCK_COMPAT_COS,
3052 };
3053
3054 /**
3055  * Rather arbitrary mapping from LDLM error codes to errno values. This should
3056  * not escape to the user level.
3057  */
3058 int ldlm_error2errno(enum ldlm_error error)
3059 {
3060         int result;
3061
3062         switch (error) {
3063         case ELDLM_OK:
3064         case ELDLM_LOCK_MATCHED:
3065                 result = 0;
3066                 break;
3067         case ELDLM_LOCK_CHANGED:
3068                 result = -ESTALE;
3069                 break;
3070         case ELDLM_LOCK_ABORTED:
3071                 result = -ENAVAIL;
3072                 break;
3073         case ELDLM_LOCK_REPLACED:
3074                 result = -ESRCH;
3075                 break;
3076         case ELDLM_NO_LOCK_DATA:
3077                 result = -ENOENT;
3078                 break;
3079         case ELDLM_NAMESPACE_EXISTS:
3080                 result = -EEXIST;
3081                 break;
3082         case ELDLM_BAD_NAMESPACE:
3083                 result = -EBADF;
3084                 break;
3085         default:
3086                 if (((int)error) < 0) { /* cast to signed type */
3087                         result = error; /* as ldlm_error can be unsigned */
3088                 } else {
3089                         CERROR("Invalid DLM result code: %d\n", error);
3090                         result = -EPROTO;
3091                 }
3092         }
3093         return result;
3094 }
3095 EXPORT_SYMBOL(ldlm_error2errno);
3096
3097 /**
3098  * Dual to ldlm_error2errno(): maps errno values back to enum ldlm_error.
3099  */
3100 enum ldlm_error ldlm_errno2error(int err_no)
3101 {
3102         int error;
3103
3104         switch (err_no) {
3105         case 0:
3106                 error = ELDLM_OK;
3107                 break;
3108         case -ESTALE:
3109                 error = ELDLM_LOCK_CHANGED;
3110                 break;
3111         case -ENAVAIL:
3112                 error = ELDLM_LOCK_ABORTED;
3113                 break;
3114         case -ESRCH:
3115                 error = ELDLM_LOCK_REPLACED;
3116                 break;
3117         case -ENOENT:
3118                 error = ELDLM_NO_LOCK_DATA;
3119                 break;
3120         case -EEXIST:
3121                 error = ELDLM_NAMESPACE_EXISTS;
3122                 break;
3123         case -EBADF:
3124                 error = ELDLM_BAD_NAMESPACE;
3125                 break;
3126         default:
3127                 error = err_no;
3128         }
3129         return error;
3130 }
3131
3132 #if LUSTRE_TRACKS_LOCK_EXP_REFS
3133 void ldlm_dump_export_locks(struct obd_export *exp)
3134 {
3135         spin_lock(&exp->exp_locks_list_guard);
3136         if (!list_empty(&exp->exp_locks_list)) {
3137                 struct ldlm_lock *lock;
3138
3139                 CERROR("dumping locks for export %p,"
3140                        "ignore if the unmount doesn't hang\n", exp);
3141                 list_for_each_entry(lock, &exp->exp_locks_list,
3142                                         l_exp_refs_link)
3143                         LDLM_ERROR(lock, "lock:");
3144         }
3145         spin_unlock(&exp->exp_locks_list_guard);
3146 }
3147 #endif
3148
3149 #ifdef HAVE_SERVER_SUPPORT
3150 static int target_bulk_timeout(void *data)
3151 {
3152         ENTRY;
3153         /* We don't fail the connection here, because having the export
3154          * killed makes the (vital) call to commitrw very sad.
3155          */
3156         RETURN(1);
3157 }
3158
3159 static inline const char *bulk2type(struct ptlrpc_request *req)
3160 {
3161         if (req->rq_bulk_read)
3162                 return "READ";
3163         if (req->rq_bulk_write)
3164                 return "WRITE";
3165         return "UNKNOWN";
3166 }
3167
3168 int target_bulk_io(struct obd_export *exp, struct ptlrpc_bulk_desc *desc,
3169                    struct l_wait_info *lwi)
3170 {
3171         struct ptlrpc_request *req = desc->bd_req;
3172         time64_t start = ktime_get_real_seconds();
3173         time64_t deadline;
3174         int rc = 0;
3175
3176         ENTRY;
3177
3178         /* If there is eviction in progress, wait for it to finish. */
3179         if (unlikely(atomic_read(&exp->exp_obd->obd_evict_inprogress))) {
3180                 *lwi = LWI_INTR(NULL, NULL);
3181                 rc = l_wait_event(exp->exp_obd->obd_evict_inprogress_waitq,
3182                                   !atomic_read(&exp->exp_obd->
3183                                                    obd_evict_inprogress),
3184                                   lwi);
3185         }
3186
3187         /* Check if client was evicted or reconnected already. */
3188         if (exp->exp_failed ||
3189             exp->exp_conn_cnt > lustre_msg_get_conn_cnt(req->rq_reqmsg)) {
3190                 rc = -ENOTCONN;
3191         } else {
3192                 if (req->rq_bulk_read)
3193                         rc = sptlrpc_svc_wrap_bulk(req, desc);
3194
3195                 if (OCD_HAS_FLAG(&exp->exp_connect_data, BULK_MBITS))
3196                         req->rq_mbits = lustre_msg_get_mbits(req->rq_reqmsg);
3197                 else /* old version, bulk matchbits is rq_xid */
3198                         req->rq_mbits = req->rq_xid;
3199
3200                 if (rc == 0)
3201                         rc = ptlrpc_start_bulk_transfer(desc);
3202         }
3203
3204         if (rc < 0) {
3205                 DEBUG_REQ(D_ERROR, req, "bulk %s failed: rc %d",
3206                           bulk2type(req), rc);
3207                 RETURN(rc);
3208         }
3209
3210         if (OBD_FAIL_CHECK(OBD_FAIL_MDS_SENDPAGE)) {
3211                 ptlrpc_abort_bulk(desc);
3212                 RETURN(0);
3213         }
3214
3215         /* limit actual bulk transfer to bulk_timeout seconds */
3216         deadline = start + bulk_timeout;
3217         if (deadline > req->rq_deadline)
3218                 deadline = req->rq_deadline;
3219
3220         do {
3221                 time64_t timeoutl = deadline - ktime_get_real_seconds();
3222                 long timeout_jiffies = timeoutl <= 0 ?
3223                                        1 : cfs_time_seconds(timeoutl);
3224                 time64_t rq_deadline;
3225
3226                 *lwi = LWI_TIMEOUT_INTERVAL(timeout_jiffies,
3227                                             cfs_time_seconds(1),
3228                                             target_bulk_timeout, desc);
3229                 rc = l_wait_event(desc->bd_waitq,
3230                                   !ptlrpc_server_bulk_active(desc) ||
3231                                   exp->exp_failed ||
3232                                   exp->exp_conn_cnt >
3233                                   lustre_msg_get_conn_cnt(req->rq_reqmsg),
3234                                   lwi);
3235                 LASSERT(rc == 0 || rc == -ETIMEDOUT);
3236                 /* Wait again if we changed rq_deadline. */
3237                 rq_deadline = READ_ONCE(req->rq_deadline);
3238                 deadline = start + bulk_timeout;
3239                 if (deadline > rq_deadline)
3240                         deadline = rq_deadline;
3241         } while (rc == -ETIMEDOUT &&
3242                  deadline > ktime_get_real_seconds());
3243
3244         if (rc == -ETIMEDOUT) {
3245                 DEBUG_REQ(D_ERROR, req, "timeout on bulk %s after %lld%+llds",
3246                           bulk2type(req), deadline - start,
3247                           ktime_get_real_seconds() - deadline);
3248                 ptlrpc_abort_bulk(desc);
3249         } else if (exp->exp_failed) {
3250                 DEBUG_REQ(D_ERROR, req, "Eviction on bulk %s",
3251                           bulk2type(req));
3252                 rc = -ENOTCONN;
3253                 ptlrpc_abort_bulk(desc);
3254         } else if (exp->exp_conn_cnt >
3255                    lustre_msg_get_conn_cnt(req->rq_reqmsg)) {
3256                 DEBUG_REQ(D_ERROR, req, "Reconnect on bulk %s",
3257                           bulk2type(req));
3258                 /* We don't reply anyway. */
3259                 rc = -ETIMEDOUT;
3260                 ptlrpc_abort_bulk(desc);
3261         } else if (desc->bd_failure) {
3262                 DEBUG_REQ(D_ERROR, req, "network error on bulk %s",
3263                           bulk2type(req));
3264                 /* XXX should this be a different errno? */
3265                 rc = -ETIMEDOUT;
3266         } else {
3267                 if (req->rq_bulk_write)
3268                         rc = sptlrpc_svc_unwrap_bulk(req, desc);
3269                 if (rc == 0 && desc->bd_nob_transferred != desc->bd_nob) {
3270                         DEBUG_REQ(D_ERROR, req, "truncated bulk %s %d(%d)",
3271                                   bulk2type(req), desc->bd_nob_transferred,
3272                                   desc->bd_nob);
3273                         /* XXX should this be a different errno? */
3274                         rc = -ETIMEDOUT;
3275                 }
3276         }
3277
3278         RETURN(rc);
3279 }
3280 EXPORT_SYMBOL(target_bulk_io);
3281
3282 #endif /* HAVE_SERVER_SUPPORT */