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