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 <cl_object.h>
  43 #include <linux/jiffies.h>
  44 #include <linux/kernel.h>
  45 #include <linux/kthread.h>
  46 #include <libcfs/libcfs.h>
  47 #include <obd.h>
  48 #include <obd_class.h>
  49 #include <lustre_dlm.h>
  50 #include <lustre_net.h>
  51 #include <lustre_sec.h>
  52 #include "ldlm_internal.h"
  53
  54 /*
  55  * @priority: If non-zero, move the selected connection to the list head.
  56  * @create: If zero, only search in existing connections.
  57  */
  58 static int import_set_conn(struct obd_import *imp, struct obd_uuid *uuid,
  59                            int priority, int create)
  60 {
  61         struct ptlrpc_connection *ptlrpc_conn;
  62         struct obd_import_conn *imp_conn = NULL, *item;
  63         lnet_nid_t nid4refnet = LNET_NID_ANY;
  64         int rc = 0;
  65
  66         ENTRY;
  67
  68         if (!create && !priority) {
  69                 CDEBUG(D_HA, "Nothing to do\n");
  70                 RETURN(-EINVAL);
  71         }
  72
  73         if (imp->imp_connection &&
  74             imp->imp_connection->c_remote_uuid.uuid[0] == 0)
  75                 /* nid4refnet is used to restrict network connections */
  76                 nid4refnet = imp->imp_connection->c_self;
  77         ptlrpc_conn = ptlrpc_uuid_to_connection(uuid, nid4refnet);
  78         if (!ptlrpc_conn) {
  79                 CDEBUG(D_HA, "can't find connection %s\n", uuid->uuid);
  80                 RETURN(-ENOENT);
  81         }
  82
  83         if (create) {
  84                 OBD_ALLOC(imp_conn, sizeof(*imp_conn));
  85                 if (!imp_conn)
  86                         GOTO(out_put, rc = -ENOMEM);
  87         }
  88
  89         spin_lock(&imp->imp_lock);
  90         list_for_each_entry(item, &imp->imp_conn_list, oic_item) {
  91                 if (obd_uuid_equals(uuid, &item->oic_uuid)) {
  92                         if (priority) {
  93                                 list_move(&item->oic_item,
  94                                           &imp->imp_conn_list);
  95                                 item->oic_last_attempt = 0;
  96                         }
  97                         CDEBUG(D_HA, "imp %p@%s: found existing conn %s%s\n",
  98                                imp, imp->imp_obd->obd_name, uuid->uuid,
  99                                (priority ? ", moved to head" : ""));
 100                         spin_unlock(&imp->imp_lock);
 101                         GOTO(out_free, rc = 0);
 102                 }
 103         }
 104         /* No existing import connection found for \a uuid. */
 105         if (create) {
 106                 imp_conn->oic_conn = ptlrpc_conn;
 107                 imp_conn->oic_uuid = *uuid;
 108                 imp_conn->oic_last_attempt = 0;
 109                 if (priority)
 110                         list_add(&imp_conn->oic_item, &imp->imp_conn_list);
 111                 else
 112                         list_add_tail(&imp_conn->oic_item,
 113                                       &imp->imp_conn_list);
 114                 CDEBUG(D_HA, "imp %p@%s: add connection %s at %s\n",
 115                        imp, imp->imp_obd->obd_name, uuid->uuid,
 116                        (priority ? "head" : "tail"));
 117         } else {
 118                 spin_unlock(&imp->imp_lock);
 119                 GOTO(out_free, rc = -ENOENT);
 120         }
 121
 122         spin_unlock(&imp->imp_lock);
 123         RETURN(0);
 124 out_free:
 125         if (imp_conn)
 126                 OBD_FREE(imp_conn, sizeof(*imp_conn));
 127 out_put:
 128         ptlrpc_connection_put(ptlrpc_conn);
 129         RETURN(rc);
 130 }
 131
 132 int import_set_conn_priority(struct obd_import *imp, struct obd_uuid *uuid)
 133 {
 134         return import_set_conn(imp, uuid, 1, 0);
 135 }
 136
 137 int client_import_add_conn(struct obd_import *imp, struct obd_uuid *uuid,
 138                            int priority)
 139 {
 140         return import_set_conn(imp, uuid, priority, 1);
 141 }
 142 EXPORT_SYMBOL(client_import_add_conn);
 143
 144 int client_import_del_conn(struct obd_import *imp, struct obd_uuid *uuid)
 145 {
 146         struct obd_import_conn *imp_conn;
 147         struct obd_export *dlmexp;
 148         int rc = -ENOENT;
 149
 150         ENTRY;
 151
 152         spin_lock(&imp->imp_lock);
 153         if (list_empty(&imp->imp_conn_list)) {
 154                 LASSERT(!imp->imp_connection);
 155                 GOTO(out, rc);
 156         }
 157
 158         list_for_each_entry(imp_conn, &imp->imp_conn_list, oic_item) {
 159                 if (!obd_uuid_equals(uuid, &imp_conn->oic_uuid))
 160                         continue;
 161                 LASSERT(imp_conn->oic_conn);
 162
 163                 if (imp_conn == imp->imp_conn_current) {
 164                         LASSERT(imp_conn->oic_conn == imp->imp_connection);
 165
 166                         if (imp->imp_state != LUSTRE_IMP_CLOSED &&
 167                             imp->imp_state != LUSTRE_IMP_DISCON) {
 168                                 CERROR("can't remove current connection\n");
 169                                 GOTO(out, rc = -EBUSY);
 170                         }
 171
 172                         ptlrpc_connection_put(imp->imp_connection);
 173                         imp->imp_connection = NULL;
 174
 175                         dlmexp = class_conn2export(&imp->imp_dlm_handle);
 176                         if (dlmexp && dlmexp->exp_connection) {
 177                                 LASSERT(dlmexp->exp_connection ==
 178                                         imp_conn->oic_conn);
 179                                 ptlrpc_connection_put(dlmexp->exp_connection);
 180                                 dlmexp->exp_connection = NULL;
 181                         }
 182
 183                         if (dlmexp != NULL)
 184                                 class_export_put(dlmexp);
 185                 }
 186
 187                 list_del(&imp_conn->oic_item);
 188                 ptlrpc_connection_put(imp_conn->oic_conn);
 189                 OBD_FREE(imp_conn, sizeof(*imp_conn));
 190                 CDEBUG(D_HA, "imp %p@%s: remove connection %s\n",
 191                        imp, imp->imp_obd->obd_name, uuid->uuid);
 192                 rc = 0;
 193                 break;
 194         }
 195 out:
 196         spin_unlock(&imp->imp_lock);
 197         if (rc == -ENOENT)
 198                 CERROR("connection %s not found\n", uuid->uuid);
 199         RETURN(rc);
 200 }
 201 EXPORT_SYMBOL(client_import_del_conn);
 202
 203 /**
 204  * Find conn UUID by peer NID. \a peer is a server NID. This function is used
 205  * to find a conn uuid of \a imp which can reach \a peer.
 206  */
 207 int client_import_find_conn(struct obd_import *imp, lnet_nid_t peer,
 208                             struct obd_uuid *uuid)
 209 {
 210         struct obd_import_conn *conn;
 211         int rc = -ENOENT;
 212
 213         ENTRY;
 214
 215         spin_lock(&imp->imp_lock);
 216         list_for_each_entry(conn, &imp->imp_conn_list, oic_item) {
 217                 /* Check if conn UUID does have this peer NID. */
 218                 if (class_check_uuid(&conn->oic_uuid, peer)) {
 219                         *uuid = conn->oic_uuid;
 220                         rc = 0;
 221                         break;
 222                 }
 223         }
 224         spin_unlock(&imp->imp_lock);
 225         RETURN(rc);
 226 }
 227 EXPORT_SYMBOL(client_import_find_conn);
 228
 229 void client_destroy_import(struct obd_import *imp)
 230 {
 231         /*
 232          * Drop security policy instance after all RPCs have finished/aborted
 233          * to let all busy contexts be released.
 234          */
 235         class_import_get(imp);
 236         class_destroy_import(imp);
 237         sptlrpc_import_sec_put(imp);
 238         class_import_put(imp);
 239 }
 240 EXPORT_SYMBOL(client_destroy_import);
 241
 242 /**
 243  * Check whether or not the OSC is on MDT.
 244  * In the config log,
 245  * osc on MDT
 246  *      setup 0:{fsname}-OSTxxxx-osc[-MDTxxxx] 1:lustre-OST0000_UUID 2:NID
 247  * osc on client
 248  *      setup 0:{fsname}-OSTxxxx-osc 1:lustre-OST0000_UUID 2:NID
 249  *
 250  **/
 251 static int osc_on_mdt(char *obdname)
 252 {
 253         char *ptr;
 254
 255         ptr = strrchr(obdname, '-');
 256         if (ptr == NULL)
 257                 return 0;
 258
 259         if (strncmp(ptr + 1, "MDT", 3) == 0)
 260                 return 1;
 261
 262         return 0;
 263 }
 264
 265 /*
 266  * Configure an RPC client OBD device.
 267  *
 268  * lcfg parameters:
 269  * 1 - client UUID
 270  * 2 - server UUID
 271  * 3 - inactive-on-startup
 272  * 4 - restrictive net
 273  */
 274 int client_obd_setup(struct obd_device *obd, struct lustre_cfg *lcfg)
 275 {
 276         struct client_obd *cli = &obd->u.cli;
 277         struct obd_import *imp;
 278         struct obd_uuid server_uuid;
 279         int rq_portal, rp_portal, connect_op;
 280         const char *name = obd->obd_type->typ_name;
 281         enum ldlm_ns_type ns_type = LDLM_NS_TYPE_UNKNOWN;
 282         char *cli_name = lustre_cfg_buf(lcfg, 0);
 283         struct ptlrpc_connection fake_conn = { .c_self = 0,
 284                                                .c_remote_uuid.uuid[0] = 0 };
 285         int rc;
 286
 287         ENTRY;
 288
 289         /*
 290          * In a more perfect world, we would hang a ptlrpc_client off of
 291          * obd_type and just use the values from there.
 292          */
 293         if (!strcmp(name, LUSTRE_OSC_NAME)) {
 294                 rq_portal = OST_REQUEST_PORTAL;
 295                 rp_portal = OSC_REPLY_PORTAL;
 296                 connect_op = OST_CONNECT;
 297                 cli->cl_sp_me = LUSTRE_SP_CLI;
 298                 cli->cl_sp_to = LUSTRE_SP_OST;
 299                 ns_type = LDLM_NS_TYPE_OSC;
 300         } else if (!strcmp(name, LUSTRE_MDC_NAME) ||
 301                    !strcmp(name, LUSTRE_LWP_NAME)) {
 302                 rq_portal = MDS_REQUEST_PORTAL;
 303                 rp_portal = MDC_REPLY_PORTAL;
 304                 connect_op = MDS_CONNECT;
 305                 if (is_lwp_on_ost(cli_name))
 306                         cli->cl_sp_me = LUSTRE_SP_OST;
 307                 else if (is_lwp_on_mdt(cli_name))
 308                         cli->cl_sp_me = LUSTRE_SP_MDT;
 309                 else
 310                         cli->cl_sp_me = LUSTRE_SP_CLI;
 311                 cli->cl_sp_to = LUSTRE_SP_MDT;
 312                 ns_type = LDLM_NS_TYPE_MDC;
 313         } else if (!strcmp(name, LUSTRE_OSP_NAME)) {
 314                 if (strstr(lustre_cfg_buf(lcfg, 1), "OST") == NULL) {
 315                         /* OSP_on_MDT for other MDTs */
 316                         connect_op = MDS_CONNECT;
 317                         cli->cl_sp_to = LUSTRE_SP_MDT;
 318                         ns_type = LDLM_NS_TYPE_MDC;
 319                         rq_portal = OUT_PORTAL;
 320                 } else {
 321                         /* OSP on MDT for OST */
 322                         connect_op = OST_CONNECT;
 323                         cli->cl_sp_to = LUSTRE_SP_OST;
 324                         ns_type = LDLM_NS_TYPE_OSC;
 325                         rq_portal = OST_REQUEST_PORTAL;
 326                 }
 327                 rp_portal = OSC_REPLY_PORTAL;
 328                 cli->cl_sp_me = LUSTRE_SP_MDT;
 329         } else if (!strcmp(name, LUSTRE_MGC_NAME)) {
 330                 rq_portal = MGS_REQUEST_PORTAL;
 331                 rp_portal = MGC_REPLY_PORTAL;
 332                 connect_op = MGS_CONNECT;
 333                 cli->cl_sp_me = LUSTRE_SP_MGC;
 334                 cli->cl_sp_to = LUSTRE_SP_MGS;
 335                 cli->cl_flvr_mgc.sf_rpc = SPTLRPC_FLVR_INVALID;
 336                 ns_type = LDLM_NS_TYPE_MGC;
 337         } else {
 338                 CERROR("unknown client OBD type \"%s\", can't setup\n",
 339                        name);
 340                 RETURN(-EINVAL);
 341         }
 342
 343         if (LUSTRE_CFG_BUFLEN(lcfg, 1) < 1) {
 344                 CERROR("requires a TARGET UUID\n");
 345                 RETURN(-EINVAL);
 346         }
 347
 348         if (LUSTRE_CFG_BUFLEN(lcfg, 1) > 37) {
 349                 CERROR("client UUID must be less than 38 characters\n");
 350                 RETURN(-EINVAL);
 351         }
 352
 353         if (LUSTRE_CFG_BUFLEN(lcfg, 2) < 1) {
 354                 CERROR("setup requires a SERVER UUID\n");
 355                 RETURN(-EINVAL);
 356         }
 357
 358         if (LUSTRE_CFG_BUFLEN(lcfg, 2) > 37) {
 359                 CERROR("target UUID must be less than 38 characters\n");
 360                 RETURN(-EINVAL);
 361         }
 362
 363         init_rwsem(&cli->cl_sem);
 364         mutex_init(&cli->cl_mgc_mutex);
 365         cli->cl_seq = NULL;
 366         init_rwsem(&cli->cl_seq_rwsem);
 367         cli->cl_conn_count = 0;
 368         memcpy(server_uuid.uuid, lustre_cfg_buf(lcfg, 2),
 369                min_t(unsigned int, LUSTRE_CFG_BUFLEN(lcfg, 2),
 370                      sizeof(server_uuid)));
 371
 372         cli->cl_dirty_pages = 0;
 373         cli->cl_dirty_max_pages = 0;
 374         cli->cl_avail_grant = 0;
 375         /* FIXME: Should limit this for the sum of all cl_dirty_max_pages. */
 376         /*
 377          * cl_dirty_max_pages may be changed at connect time in
 378          * ptlrpc_connect_interpret().
 379          */
 380         client_adjust_max_dirty(cli);
 381         init_waitqueue_head(&cli->cl_cache_waiters);
 382         INIT_LIST_HEAD(&cli->cl_loi_ready_list);
 383         INIT_LIST_HEAD(&cli->cl_loi_hp_ready_list);
 384         INIT_LIST_HEAD(&cli->cl_loi_write_list);
 385         INIT_LIST_HEAD(&cli->cl_loi_read_list);
 386         spin_lock_init(&cli->cl_loi_list_lock);
 387         atomic_set(&cli->cl_pending_w_pages, 0);
 388         atomic_set(&cli->cl_pending_r_pages, 0);
 389         cli->cl_r_in_flight = 0;
 390         cli->cl_w_in_flight = 0;
 391
 392         spin_lock_init(&cli->cl_read_rpc_hist.oh_lock);
 393         spin_lock_init(&cli->cl_write_rpc_hist.oh_lock);
 394         spin_lock_init(&cli->cl_read_page_hist.oh_lock);
 395         spin_lock_init(&cli->cl_write_page_hist.oh_lock);
 396         spin_lock_init(&cli->cl_read_offset_hist.oh_lock);
 397         spin_lock_init(&cli->cl_write_offset_hist.oh_lock);
 398
 399         /* lru for osc. */
 400         INIT_LIST_HEAD(&cli->cl_lru_osc);
 401         atomic_set(&cli->cl_lru_shrinkers, 0);
 402         atomic_long_set(&cli->cl_lru_busy, 0);
 403         atomic_long_set(&cli->cl_lru_in_list, 0);
 404         INIT_LIST_HEAD(&cli->cl_lru_list);
 405         spin_lock_init(&cli->cl_lru_list_lock);
 406         atomic_long_set(&cli->cl_unstable_count, 0);
 407         INIT_LIST_HEAD(&cli->cl_shrink_list);
 408         INIT_LIST_HEAD(&cli->cl_grant_chain);
 409
 410         INIT_LIST_HEAD(&cli->cl_flight_waiters);
 411         cli->cl_rpcs_in_flight = 0;
 412
 413         init_waitqueue_head(&cli->cl_destroy_waitq);
 414         atomic_set(&cli->cl_destroy_in_flight, 0);
 415
 416
 417         cli->cl_supp_cksum_types = OBD_CKSUM_CRC32;
 418         cli->cl_preferred_cksum_type = 0;
 419 #ifdef ENABLE_CHECKSUM
 420         /* Turn on checksumming by default. */
 421         cli->cl_checksum = 1;
 422         /*
 423          * The supported checksum types will be worked out at connect time
 424          * Set cl_chksum* to CRC32 for now to avoid returning screwed info
 425          * through procfs.
 426          */
 427         cli->cl_cksum_type = cli->cl_supp_cksum_types;
 428 #endif
 429         atomic_set(&cli->cl_resends, OSC_DEFAULT_RESENDS);
 430
 431         /*
 432          * Set it to possible maximum size. It may be reduced by ocd_brw_size
 433          * from OFD after connecting.
 434          */
 435         cli->cl_max_pages_per_rpc = PTLRPC_MAX_BRW_PAGES;
 436
 437         cli->cl_max_short_io_bytes = OBD_DEF_SHORT_IO_BYTES;
 438
 439         /*
 440          * set cl_chunkbits default value to PAGE_SHIFT,
 441          * it will be updated at OSC connection time.
 442          */
 443         cli->cl_chunkbits = PAGE_SHIFT;
 444
 445         if (!strcmp(name, LUSTRE_MDC_NAME)) {
 446                 cli->cl_max_rpcs_in_flight = OBD_MAX_RIF_DEFAULT;
 447         } else if (cfs_totalram_pages() >> (20 - PAGE_SHIFT) <= 128 /* MB */) {
 448                 cli->cl_max_rpcs_in_flight = 2;
 449         } else if (cfs_totalram_pages() >> (20 - PAGE_SHIFT) <= 256 /* MB */) {
 450                 cli->cl_max_rpcs_in_flight = 3;
 451         } else if (cfs_totalram_pages() >> (20 - PAGE_SHIFT) <= 512 /* MB */) {
 452                 cli->cl_max_rpcs_in_flight = 4;
 453         } else {
 454                 if (osc_on_mdt(obd->obd_name))
 455                         cli->cl_max_rpcs_in_flight = OBD_MAX_RIF_MAX;
 456                 else
 457                         cli->cl_max_rpcs_in_flight = OBD_MAX_RIF_DEFAULT;
 458         }
 459
 460         spin_lock_init(&cli->cl_mod_rpcs_lock);
 461         spin_lock_init(&cli->cl_mod_rpcs_hist.oh_lock);
 462         cli->cl_max_mod_rpcs_in_flight = 0;
 463         cli->cl_mod_rpcs_in_flight = 0;
 464         cli->cl_close_rpcs_in_flight = 0;
 465         init_waitqueue_head(&cli->cl_mod_rpcs_waitq);
 466         cli->cl_mod_tag_bitmap = NULL;
 467
 468         INIT_LIST_HEAD(&cli->cl_chg_dev_linkage);
 469
 470         if (connect_op == MDS_CONNECT) {
 471                 cli->cl_max_mod_rpcs_in_flight = cli->cl_max_rpcs_in_flight - 1;
 472                 OBD_ALLOC(cli->cl_mod_tag_bitmap,
 473                           BITS_TO_LONGS(OBD_MAX_RIF_MAX) * sizeof(long));
 474                 if (cli->cl_mod_tag_bitmap == NULL)
 475                         GOTO(err, rc = -ENOMEM);
 476         }
 477
 478         rc = ldlm_get_ref();
 479         if (rc) {
 480                 CERROR("ldlm_get_ref failed: %d\n", rc);
 481                 GOTO(err, rc);
 482         }
 483
 484         ptlrpc_init_client(rq_portal, rp_portal, name,
 485                            &obd->obd_ldlm_client);
 486
 487         imp = class_new_import(obd);
 488         if (imp == NULL)
 489                 GOTO(err_ldlm, rc = -ENOENT);
 490         imp->imp_client = &obd->obd_ldlm_client;
 491         imp->imp_connect_op = connect_op;
 492         memcpy(cli->cl_target_uuid.uuid, lustre_cfg_buf(lcfg, 1),
 493                LUSTRE_CFG_BUFLEN(lcfg, 1));
 494         class_import_put(imp);
 495
 496         if (lustre_cfg_buf(lcfg, 4)) {
 497                 __u32 refnet = libcfs_str2net(lustre_cfg_string(lcfg, 4));
 498
 499                 if (refnet == LNET_NIDNET(LNET_NID_ANY)) {
 500                         rc = -EINVAL;
 501                         CERROR("%s: bad mount option 'network=%s': rc = %d\n",
 502                                obd->obd_name, lustre_cfg_string(lcfg, 4),
 503                                rc);
 504                         GOTO(err_import, rc);
 505                 }
 506                 fake_conn.c_self = LNET_MKNID(refnet, 0);
 507                 imp->imp_connection = &fake_conn;
 508         }
 509
 510         rc = client_import_add_conn(imp, &server_uuid, 1);
 511         if (rc) {
 512                 CERROR("can't add initial connection\n");
 513                 GOTO(err_import, rc);
 514         }
 515         imp->imp_connection = NULL;
 516
 517         cli->cl_import = imp;
 518         /* cli->cl_max_mds_easize updated by mdc_init_ea_size() */
 519         cli->cl_max_mds_easize = sizeof(struct lov_mds_md_v3);
 520
 521         if (LUSTRE_CFG_BUFLEN(lcfg, 3) > 0) {
 522                 if (!strcmp(lustre_cfg_string(lcfg, 3), "inactive")) {
 523                         CDEBUG(D_HA, "marking %s %s->%s as inactive\n",
 524                                name, obd->obd_name,
 525                                cli->cl_target_uuid.uuid);
 526                         spin_lock(&imp->imp_lock);
 527                         imp->imp_deactive = 1;
 528                         spin_unlock(&imp->imp_lock);
 529                 }
 530         }
 531
 532         obd->obd_namespace = ldlm_namespace_new(obd, obd->obd_name,
 533                                                 LDLM_NAMESPACE_CLIENT,
 534                                                 LDLM_NAMESPACE_GREEDY,
 535                                                 ns_type);
 536         if (obd->obd_namespace == NULL) {
 537                 CERROR("Unable to create client namespace - %s\n",
 538                        obd->obd_name);
 539                 GOTO(err_import, rc = -ENOMEM);
 540         }
 541
 542         RETURN(rc);
 543
 544 err_import:
 545         class_destroy_import(imp);
 546 err_ldlm:
 547         ldlm_put_ref();
 548 err:
 549         if (cli->cl_mod_tag_bitmap != NULL)
 550                 OBD_FREE(cli->cl_mod_tag_bitmap,
 551                          BITS_TO_LONGS(OBD_MAX_RIF_MAX) * sizeof(long));
 552         cli->cl_mod_tag_bitmap = NULL;
 553         RETURN(rc);
 554
 555 }
 556 EXPORT_SYMBOL(client_obd_setup);
 557
 558 int client_obd_cleanup(struct obd_device *obd)
 559 {
 560         struct client_obd *cli = &obd->u.cli;
 561
 562         ENTRY;
 563
 564         ldlm_namespace_free_post(obd->obd_namespace);
 565         obd->obd_namespace = NULL;
 566
 567         obd_cleanup_client_import(obd);
 568         LASSERT(obd->u.cli.cl_import == NULL);
 569
 570         ldlm_put_ref();
 571
 572         if (cli->cl_mod_tag_bitmap != NULL)
 573                 OBD_FREE(cli->cl_mod_tag_bitmap,
 574                          BITS_TO_LONGS(OBD_MAX_RIF_MAX) * sizeof(long));
 575         cli->cl_mod_tag_bitmap = NULL;
 576
 577         RETURN(0);
 578 }
 579 EXPORT_SYMBOL(client_obd_cleanup);
 580
 581 /* ->o_connect() method for client side (OSC and MDC and MGC) */
 582 int client_connect_import(const struct lu_env *env,
 583                           struct obd_export **exp,
 584                           struct obd_device *obd, struct obd_uuid *cluuid,
 585                           struct obd_connect_data *data, void *localdata)
 586 {
 587         struct client_obd *cli = &obd->u.cli;
 588         struct obd_import *imp = cli->cl_import;
 589         struct obd_connect_data *ocd;
 590         struct lustre_handle conn = { 0 };
 591         int rc;
 592
 593         ENTRY;
 594
 595         *exp = NULL;
 596         down_write(&cli->cl_sem);
 597         if (cli->cl_conn_count > 0)
 598                 GOTO(out_sem, rc = -EALREADY);
 599
 600         rc = class_connect(&conn, obd, cluuid);
 601         if (rc)
 602                 GOTO(out_sem, rc);
 603
 604         cli->cl_conn_count++;
 605         *exp = class_conn2export(&conn);
 606
 607         LASSERT(obd->obd_namespace);
 608
 609         imp->imp_dlm_handle = conn;
 610         rc = ptlrpc_init_import(imp);
 611         if (rc != 0)
 612                 GOTO(out_ldlm, rc);
 613
 614         ocd = &imp->imp_connect_data;
 615         if (data) {
 616                 *ocd = *data;
 617                 imp->imp_connect_flags_orig = data->ocd_connect_flags;
 618                 imp->imp_connect_flags2_orig = data->ocd_connect_flags2;
 619         }
 620
 621         rc = ptlrpc_connect_import(imp);
 622         if (rc != 0) {
 623                 LASSERT(imp->imp_state == LUSTRE_IMP_DISCON);
 624                 GOTO(out_ldlm, rc);
 625         }
 626         LASSERT(*exp != NULL && (*exp)->exp_connection);
 627
 628         if (data) {
 629                 LASSERTF((ocd->ocd_connect_flags & data->ocd_connect_flags) ==
 630                          ocd->ocd_connect_flags, "old %#llx, new %#llx\n",
 631                          data->ocd_connect_flags, ocd->ocd_connect_flags);
 632                 data->ocd_connect_flags = ocd->ocd_connect_flags;
 633                 data->ocd_connect_flags2 = ocd->ocd_connect_flags2;
 634         }
 635
 636         ptlrpc_pinger_add_import(imp);
 637
 638         EXIT;
 639
 640         if (rc) {
 641 out_ldlm:
 642                 cli->cl_conn_count--;
 643                 class_disconnect(*exp);
 644                 *exp = NULL;
 645         }
 646 out_sem:
 647         up_write(&cli->cl_sem);
 648
 649         if (!rc && localdata) {
 650                 LASSERT(cli->cl_cache == NULL); /* only once */
 651                 cli->cl_cache = (struct cl_client_cache *)localdata;
 652                 cl_cache_incref(cli->cl_cache);
 653                 cli->cl_lru_left = &cli->cl_cache->ccc_lru_left;
 654
 655                 /* add this osc into entity list */
 656                 LASSERT(list_empty(&cli->cl_lru_osc));
 657                 spin_lock(&cli->cl_cache->ccc_lru_lock);
 658                 list_add(&cli->cl_lru_osc, &cli->cl_cache->ccc_lru);
 659                 spin_unlock(&cli->cl_cache->ccc_lru_lock);
 660         }
 661
 662         return rc;
 663 }
 664 EXPORT_SYMBOL(client_connect_import);
 665
 666 int client_disconnect_export(struct obd_export *exp)
 667 {
 668         struct obd_device *obd = class_exp2obd(exp);
 669         struct client_obd *cli;
 670         struct obd_import *imp;
 671         int rc = 0, err;
 672
 673         ENTRY;
 674
 675         if (!obd) {
 676                 CERROR("invalid export for disconnect: exp %p cookie %#llx\n",
 677                        exp, exp ? exp->exp_handle.h_cookie : -1);
 678                 RETURN(-EINVAL);
 679         }
 680
 681         cli = &obd->u.cli;
 682         imp = cli->cl_import;
 683
 684         down_write(&cli->cl_sem);
 685         CDEBUG(D_INFO, "disconnect %s - %zu\n", obd->obd_name,
 686                 cli->cl_conn_count);
 687
 688         if (cli->cl_conn_count == 0) {
 689                 CERROR("disconnecting disconnected device (%s)\n",
 690                        obd->obd_name);
 691                 GOTO(out_disconnect, rc = -EINVAL);
 692         }
 693
 694         cli->cl_conn_count--;
 695         if (cli->cl_conn_count != 0)
 696                 GOTO(out_disconnect, rc = 0);
 697
 698         /*
 699          * Mark import deactivated now, so we don't try to reconnect if any
 700          * of the cleanup RPCs fails (e.g. LDLM cancel, etc).  We don't
 701          * fully deactivate the import, or that would drop all requests.
 702          */
 703         spin_lock(&imp->imp_lock);
 704         imp->imp_deactive = 1;
 705         spin_unlock(&imp->imp_lock);
 706
 707         /*
 708          * Some non-replayable imports (MDS's OSCs) are pinged, so just
 709          * delete it regardless.  (It's safe to delete an import that was
 710          * never added.)
 711          */
 712         (void)ptlrpc_pinger_del_import(imp);
 713
 714         if (obd->obd_namespace != NULL) {
 715                 /* obd_force == local only */
 716                 ldlm_cli_cancel_unused(obd->obd_namespace, NULL,
 717                                        obd->obd_force ? LCF_LOCAL : 0, NULL);
 718                 ldlm_namespace_free_prior(obd->obd_namespace, imp,
 719                                           obd->obd_force);
 720         }
 721
 722         /*
 723          * There's no need to hold sem while disconnecting an import,
 724          * and it may actually cause deadlock in GSS.
 725          */
 726         up_write(&cli->cl_sem);
 727         rc = ptlrpc_disconnect_import(imp, 0);
 728         down_write(&cli->cl_sem);
 729
 730         ptlrpc_invalidate_import(imp);
 731
 732         EXIT;
 733
 734 out_disconnect:
 735         /*
 736          * Use server style - class_disconnect should be always called for
 737          * o_disconnect.
 738          */
 739         err = class_disconnect(exp);
 740         if (!rc && err)
 741                 rc = err;
 742
 743         up_write(&cli->cl_sem);
 744
 745         RETURN(rc);
 746 }
 747 EXPORT_SYMBOL(client_disconnect_export);
 748
 749 #ifdef HAVE_SERVER_SUPPORT
 750 int server_disconnect_export(struct obd_export *exp)
 751 {
 752         int rc;
 753
 754         ENTRY;
 755
 756         /* Disconnect early so that clients can't keep using export. */
 757         rc = class_disconnect(exp);
 758         /* Close import to avoid sending any requests. */
 759         if (exp->exp_imp_reverse)
 760                 ptlrpc_cleanup_imp(exp->exp_imp_reverse);
 761
 762         ldlm_bl_thread_wakeup();
 763
 764         /* complete all outstanding replies */
 765         spin_lock(&exp->exp_lock);
 766         while (!list_empty(&exp->exp_outstanding_replies)) {
 767                 struct ptlrpc_reply_state *rs =
 768                         list_entry(exp->exp_outstanding_replies.next,
 769                                        struct ptlrpc_reply_state, rs_exp_list);
 770                 struct ptlrpc_service_part *svcpt = rs->rs_svcpt;
 771
 772                 spin_lock(&svcpt->scp_rep_lock);
 773
 774                 list_del_init(&rs->rs_exp_list);
 775
 776                 spin_lock(&rs->rs_lock);
 777                 /* clear rs_convert_lock to make sure rs is handled and put */
 778                 rs->rs_convert_lock = 0;
 779                 ptlrpc_schedule_difficult_reply(rs);
 780                 spin_unlock(&rs->rs_lock);
 781
 782                 spin_unlock(&svcpt->scp_rep_lock);
 783         }
 784         spin_unlock(&exp->exp_lock);
 785
 786         RETURN(rc);
 787 }
 788 EXPORT_SYMBOL(server_disconnect_export);
 789
 790 static inline int target_check_recovery_timer(struct obd_device *target)
 791 {
 792         ktime_t remaining;
 793         s64 timeout;
 794
 795         if (!target->obd_recovering || target->obd_recovery_start == 0)
 796                 return 0;
 797
 798         remaining = hrtimer_expires_remaining(&target->obd_recovery_timer);
 799         timeout = ktime_divns(remaining, NSEC_PER_SEC);
 800         if (timeout > -30)
 801                 return 0;
 802
 803         /* the recovery timer should expire, but it isn't triggered,
 804          * it's better to abort the recovery of this target to speed up
 805          * the recovery of the whole cluster. */
 806         spin_lock(&target->obd_dev_lock);
 807         if (target->obd_recovering) {
 808                 CERROR("%s: Aborting recovery\n", target->obd_name);
 809                 target->obd_abort_recovery = 1;
 810                 wake_up(&target->obd_next_transno_waitq);
 811         }
 812         spin_unlock(&target->obd_dev_lock);
 813         return 0;
 814 }
 815
 816 /*
 817  * --------------------------------------------------------------------------
 818  * from old lib/target.c
 819  * --------------------------------------------------------------------------
 820  */
 821 static int target_handle_reconnect(struct lustre_handle *conn,
 822                                    struct obd_export *exp,
 823                                    struct obd_uuid *cluuid)
 824 {
 825         struct obd_device *target;
 826         struct lustre_handle *hdl;
 827         ktime_t remaining;
 828         s64 timeout;
 829         int rc = 0;
 830
 831         ENTRY;
 832         hdl = &exp->exp_imp_reverse->imp_remote_handle;
 833         if (!exp->exp_connection || !lustre_handle_is_used(hdl)) {
 834                 conn->cookie = exp->exp_handle.h_cookie;
 835                 CDEBUG(D_HA,
 836                        "connect export for UUID '%s' at %p, cookie %#llx\n",
 837                        cluuid->uuid, exp, conn->cookie);
 838                 RETURN(0);
 839         }
 840
 841         target = exp->exp_obd;
 842
 843         /* Might be a re-connect after a partition. */
 844         if (memcmp(&conn->cookie, &hdl->cookie, sizeof(conn->cookie))) {
 845                 LCONSOLE_WARN("%s: already connected client %s (at %s) with handle %#llx. Rejecting client with the same UUID trying to reconnect with handle %#llx\n",
 846                               target->obd_name,
 847                               obd_uuid2str(&exp->exp_client_uuid),
 848                               obd_export_nid2str(exp),
 849                               hdl->cookie, conn->cookie);
 850                 memset(conn, 0, sizeof(*conn));
 851                 /*
 852                  * target_handle_connect() treats EALREADY and
 853                  * -EALREADY differently.  -EALREADY is an error
 854                  * (same UUID, different handle).
 855                  */
 856                 RETURN(-EALREADY);
 857         }
 858
 859         if (!target->obd_recovering) {
 860                 LCONSOLE_WARN("%s: Client %s (at %s) reconnecting\n",
 861                         target->obd_name, obd_uuid2str(&exp->exp_client_uuid),
 862                         obd_export_nid2str(exp));
 863                 GOTO(out_already, rc);
 864         }
 865
 866         remaining = hrtimer_expires_remaining(&target->obd_recovery_timer);
 867         timeout = ktime_divns(remaining, NSEC_PER_SEC);
 868         if (timeout > 0) {
 869                 LCONSOLE_WARN("%s: Client %s (at %s) reconnected, waiting for %d clients in recovery for %lld:%.02lld\n",
 870                               target->obd_name,
 871                               obd_uuid2str(&exp->exp_client_uuid),
 872                               obd_export_nid2str(exp),
 873                               atomic_read(&target->obd_max_recoverable_clients),
 874                               timeout / 60, timeout % 60);
 875         } else {
 876                 struct target_distribute_txn_data *tdtd;
 877                 int size = 0;
 878                 int count = 0;
 879                 char *buf = NULL;
 880
 881                 target_check_recovery_timer(target);
 882
 883                 tdtd = class_exp2tgt(exp)->lut_tdtd;
 884                 if (tdtd && tdtd->tdtd_show_update_logs_retrievers)
 885                         buf = tdtd->tdtd_show_update_logs_retrievers(
 886                                 tdtd->tdtd_show_retrievers_cbdata,
 887                                 &size, &count);
 888
 889                 if (count > 0)
 890                         LCONSOLE_WARN("%s: Client %s (at %s) reconnecting, waiting for %d MDTs (%s) in recovery for %lld:%.02lld. Please wait until all MDTs recovered or you may force MDT evicition via 'lctl --device %s abort_recovery.\n",
 891                                       target->obd_name,
 892                                       obd_uuid2str(&exp->exp_client_uuid),
 893                                       obd_export_nid2str(exp), count,
 894                                       buf ? buf : "unknown (not enough RAM)",
 895                                       (abs(timeout) + target->obd_recovery_timeout) / 60,
 896                                       (abs(timeout) + target->obd_recovery_timeout) % 60,
 897                                       target->obd_name);
 898                 else
 899                         LCONSOLE_WARN("%s: Recovery already passed deadline %lld:%.02lld. If you do not want to wait more, you may force taget eviction via 'lctl --device %s abort_recovery.\n",
 900                                       target->obd_name, abs(timeout) / 60,
 901                                       abs(timeout) % 60, target->obd_name);
 902
 903                 if (buf != NULL)
 904                         OBD_FREE(buf, size);
 905         }
 906
 907 out_already:
 908         conn->cookie = exp->exp_handle.h_cookie;
 909         /*
 910          * target_handle_connect() treats EALREADY and
 911          * -EALREADY differently.  EALREADY means we are
 912          * doing a valid reconnect from the same client.
 913          */
 914         RETURN(EALREADY);
 915 }
 916
 917 static void
 918 check_and_start_recovery_timer(struct obd_device *obd,
 919                                struct ptlrpc_request *req, int new_client);
 920
 921 /**
 922  * update flags for import during reconnect process
 923  */
 924 static int rev_import_flags_update(struct obd_import *revimp,
 925                                    struct ptlrpc_request *req)
 926 {
 927         int rc;
 928         struct obd_connect_data *data;
 929
 930         data = req_capsule_client_get(&req->rq_pill, &RMF_CONNECT_DATA);
 931
 932         if (data->ocd_connect_flags & OBD_CONNECT_AT)
 933                 revimp->imp_msghdr_flags |= MSGHDR_AT_SUPPORT;
 934         else
 935                 revimp->imp_msghdr_flags &= ~MSGHDR_AT_SUPPORT;
 936
 937         revimp->imp_msghdr_flags |= MSGHDR_CKSUM_INCOMPAT18;
 938
 939         rc = sptlrpc_import_sec_adapt(revimp, req->rq_svc_ctx, &req->rq_flvr);
 940         if (rc) {
 941                 CERROR("%s: cannot get reverse import %s security: rc = %d\n",
 942                         revimp->imp_client->cli_name,
 943                         libcfs_id2str(req->rq_peer), rc);
 944                 return rc;
 945         }
 946
 947         return 0;
 948 }
 949
 950 /**
 951  * Allocate a new reverse import for an export.
 952  *
 953  * \retval -errno in case error hit
 954  * \retval 0 if reverse import correctly init
 955  **/
 956 int rev_import_init(struct obd_export *export)
 957 {
 958         struct obd_device *obd = export->exp_obd;
 959         struct obd_import *revimp;
 960
 961         LASSERT(export->exp_imp_reverse == NULL);
 962
 963         revimp = class_new_import(obd);
 964         if (revimp == NULL)
 965                 return -ENOMEM;
 966
 967         revimp->imp_remote_handle.cookie = 0ULL;
 968         revimp->imp_client = &obd->obd_ldlm_client;
 969         revimp->imp_dlm_fake = 1;
 970
 971         /* it is safe to connect import in new state as no sends possible */
 972         spin_lock(&export->exp_lock);
 973         export->exp_imp_reverse = revimp;
 974         spin_unlock(&export->exp_lock);
 975         class_import_put(revimp);
 976
 977         return 0;
 978 }
 979 EXPORT_SYMBOL(rev_import_init);
 980
 981 /**
 982  * Handle reconnect for an export.
 983  *
 984  * \param exp export to handle reconnect process
 985  * \param req client reconnect request
 986  *
 987  * \retval -rc in case securitfy flavor can't be changed
 988  * \retval 0 in case none problems
 989  */
 990 static int rev_import_reconnect(struct obd_export *exp,
 991                                 struct ptlrpc_request *req)
 992 {
 993         struct obd_import *revimp = exp->exp_imp_reverse;
 994         struct lustre_handle *lh;
 995         int rc;
 996
 997         /* avoid sending a request until import flags are changed */
 998         ptlrpc_import_enter_resend(revimp);
 999
1000         if (revimp->imp_connection != NULL)
1001                 ptlrpc_connection_put(revimp->imp_connection);
1002
1003         /*
1004          * client from recovery don't have a handle so we need to take from
1005          * request. it may produce situation when wrong client connected
1006          * to recovery as we trust a client uuid
1007          */
1008         lh = req_capsule_client_get(&req->rq_pill, &RMF_CONN);
1009         revimp->imp_remote_handle = *lh;
1010
1011         /*
1012          * unknown versions will be caught in
1013          * ptlrpc_handle_server_req_in->lustre_unpack_msg()
1014          */
1015         revimp->imp_msg_magic = req->rq_reqmsg->lm_magic;
1016
1017         revimp->imp_connection = ptlrpc_connection_addref(exp->exp_connection);
1018
1019         rc = rev_import_flags_update(revimp, req);
1020         if (rc != 0) {
1021                 /*
1022                  * it is safe to still be in RECOVERY phase as we are not able
1023                  * to setup correct security flavor so requests are not able to
1024                  * be delivered correctly
1025                  */
1026                 return rc;
1027         }
1028
1029         /* resend all rpc's via new connection */
1030         return ptlrpc_import_recovery_state_machine(revimp);
1031 }
1032
1033 int target_handle_connect(struct ptlrpc_request *req)
1034 {
1035         struct obd_device *target = NULL;
1036         struct obd_export *export = NULL;
1037         /*
1038          * connect handle - filled from target_handle_reconnect in
1039          * reconnect case
1040          */
1041         struct lustre_handle conn;
1042         struct lustre_handle *tmp;
1043         struct obd_uuid cluuid;
1044         char *str;
1045         int rc = 0;
1046         char *target_start;
1047         int target_len;
1048         bool mds_conn = false, lw_client = false, initial_conn = false;
1049         bool mds_mds_conn = false;
1050         bool new_mds_mds_conn = false;
1051         struct obd_connect_data *data, *tmpdata;
1052         int size, tmpsize;
1053         lnet_nid_t *client_nid = NULL;
1054
1055         ENTRY;
1056
1057         OBD_RACE(OBD_FAIL_TGT_CONN_RACE);
1058
1059         str = req_capsule_client_get(&req->rq_pill, &RMF_TGTUUID);
1060         if (str == NULL) {
1061                 DEBUG_REQ(D_ERROR, req, "bad target UUID for connect");
1062                 GOTO(out, rc = -EINVAL);
1063         }
1064
1065         target = class_dev_by_str(str);
1066         if (!target) {
1067                 deuuidify(str, NULL, &target_start, &target_len);
1068                 LCONSOLE_ERROR_MSG(0x137,
1069                                    "%s: not available for connect from %s (no target). If you are running an HA pair check that the target is mounted on the other server.\n",
1070                                    str, libcfs_nid2str(req->rq_peer.nid));
1071                 GOTO(out, rc = -ENODEV);
1072         }
1073
1074         spin_lock(&target->obd_dev_lock);
1075
1076         target->obd_conn_inprogress++;
1077
1078         if (target->obd_stopping || !target->obd_set_up) {
1079                 spin_unlock(&target->obd_dev_lock);
1080
1081                 deuuidify(str, NULL, &target_start, &target_len);
1082                 LCONSOLE_INFO("%.*s: Not available for connect from %s (%s)\n",
1083                               target_len, target_start,
1084                               libcfs_nid2str(req->rq_peer.nid),
1085                               (target->obd_stopping ?
1086                                "stopping" : "not set up"));
1087                 GOTO(out, rc = -ENODEV);
1088         }
1089
1090         if (target->obd_no_conn) {
1091                 spin_unlock(&target->obd_dev_lock);
1092
1093                 CDEBUG(D_INFO,
1094                        "%s: Temporarily refusing client connection from %s\n",
1095                        target->obd_name, libcfs_nid2str(req->rq_peer.nid));
1096                 GOTO(out, rc = -EAGAIN);
1097         }
1098
1099         spin_unlock(&target->obd_dev_lock);
1100
1101         str = req_capsule_client_get(&req->rq_pill, &RMF_CLUUID);
1102         if (str == NULL) {
1103                 DEBUG_REQ(D_ERROR, req, "bad client UUID for connect");
1104                 GOTO(out, rc = -EINVAL);
1105         }
1106
1107         obd_str2uuid(&cluuid, str);
1108
1109         tmp = req_capsule_client_get(&req->rq_pill, &RMF_CONN);
1110         if (tmp == NULL)
1111                 GOTO(out, rc = -EPROTO);
1112
1113         conn = *tmp;
1114
1115         size = req_capsule_get_size(&req->rq_pill, &RMF_CONNECT_DATA,
1116                                     RCL_CLIENT);
1117         if (size < 0 || size > 8 * sizeof(struct obd_connect_data))
1118                 GOTO(out, rc = -EPROTO);
1119         data = req_capsule_client_get(&req->rq_pill, &RMF_CONNECT_DATA);
1120         if (!data)
1121                 GOTO(out, rc = -EPROTO);
1122
1123         rc = req_capsule_server_pack(&req->rq_pill);
1124         if (rc)
1125                 GOTO(out, rc);
1126
1127 #if LUSTRE_VERSION_CODE < OBD_OCD_VERSION(3, 0, 53, 0)
1128         /*
1129          * Don't allow clients to connect that are using old 1.8 format
1130          * protocol conventions (LUSTRE_MSG_MAGIC_v1, !MSGHDR_CKSUM_INCOMPAT18,
1131          * ldlm_flock_policy_wire format, MDT_ATTR_xTIME_SET, etc).  The
1132          * FULL20 flag should be set on all connections since 2.0, but no
1133          * longer affects behaviour.
1134          *
1135          * Later this check will be disabled and the flag can be retired
1136          * completely once interop with 3.0 is no longer needed.
1137          */
1138         if (!(data->ocd_connect_flags & OBD_CONNECT_FULL20))
1139                 GOTO(out, rc = -EPROTO);
1140
1141         /*
1142          * Don't allow liblustre clients to connect.
1143          * - testing was disabled in v2_2_50_0-61-g6a75d65
1144          * - building was disabled in v2_5_58_0-28-g7277179
1145          * - client code was deleted in v2_6_50_0-101-gcdfbc72,
1146          * - clients were refused connect for version difference > 0.0.1.32
1147          */
1148         if (lustre_msg_get_op_flags(req->rq_reqmsg) & MSG_CONNECT_LIBCLIENT) {
1149                 DEBUG_REQ(D_WARNING, req, "Refusing libclient connection");
1150                 GOTO(out, rc = -EPROTO);
1151         }
1152 #endif
1153
1154         /*
1155          * Note: lw_client is needed in MDS-MDS failover during update log
1156          * processing, so we needs to allow lw_client to be connected at
1157          * anytime, instead of only the initial connection
1158          */
1159         lw_client = (data->ocd_connect_flags & OBD_CONNECT_LIGHTWEIGHT) != 0;
1160
1161         if (lustre_msg_get_op_flags(req->rq_reqmsg) & MSG_CONNECT_INITIAL) {
1162                 initial_conn = true;
1163                 mds_conn = (data->ocd_connect_flags & OBD_CONNECT_MDS) != 0;
1164                 mds_mds_conn = (data->ocd_connect_flags &
1165                                 OBD_CONNECT_MDS_MDS) != 0;
1166
1167                 /*
1168                  * OBD_CONNECT_MNE_SWAB is removed at 2.14
1169                  * Checking OBD_CONNECT_FID can be removed in the future.
1170                  *
1171                  * Via check OBD_CONNECT_FID, we can distinguish whether
1172                  * the OBD_CONNECT_MDS_MDS/OBD_CONNECT_MNE_SWAB is from
1173                  * MGC or MDT, since MGC does not use OBD_CONNECT_FID.
1174                  */
1175                 if (!lw_client &&
1176                     (data->ocd_connect_flags & OBD_CONNECT_MDS_MDS) &&
1177                     (data->ocd_connect_flags & OBD_CONNECT_FID) &&
1178                     (data->ocd_connect_flags & OBD_CONNECT_VERSION)) {
1179                         __u32 major = OBD_OCD_VERSION_MAJOR(data->ocd_version);
1180                         __u32 minor = OBD_OCD_VERSION_MINOR(data->ocd_version);
1181                         __u32 patch = OBD_OCD_VERSION_PATCH(data->ocd_version);
1182
1183                         /*
1184                          * We do not support the MDT-MDT interoperations with
1185                          * different version MDT because of protocol changes.
1186                          */
1187                         if (unlikely(major != LUSTRE_MAJOR ||
1188                                      minor != LUSTRE_MINOR ||
1189                                      abs(patch - LUSTRE_PATCH) > 3)) {
1190                                 LCONSOLE_WARN("%s (%u.%u.%u.%u) refused the connection from different version MDT (%d.%d.%d.%d) %s %s\n",
1191                                               target->obd_name, LUSTRE_MAJOR,
1192                                               LUSTRE_MINOR, LUSTRE_PATCH,
1193                                               LUSTRE_FIX, major, minor, patch,
1194                                               OBD_OCD_VERSION_FIX(data->ocd_version),
1195                                               libcfs_nid2str(req->rq_peer.nid),
1196                                               str);
1197                                 GOTO(out, rc = -EPROTO);
1198                         }
1199                 }
1200         }
1201
1202         /* lctl gets a backstage, all-access pass. */
1203         if (obd_uuid_equals(&cluuid, &target->obd_uuid))
1204                 goto dont_check_exports;
1205
1206         export = obd_uuid_lookup(target, &cluuid);
1207         if (!export)
1208                 goto no_export;
1209
1210         /* We've found an export in the hash. */
1211
1212         spin_lock(&export->exp_lock);
1213
1214         if (export->exp_connecting) { /* b=9635, et. al. */
1215                 spin_unlock(&export->exp_lock);
1216                 LCONSOLE_WARN("%s: Export %p already connecting from %s\n",
1217                               export->exp_obd->obd_name, export,
1218                               libcfs_nid2str(req->rq_peer.nid));
1219                 class_export_put(export);
1220                 export = NULL;
1221                 rc = -EALREADY;
1222         } else if ((mds_conn || (lw_client && initial_conn) ||
1223                    data->ocd_connect_flags & OBD_CONNECT_MDS_MDS) &&
1224                    export->exp_connection != NULL) {
1225                 spin_unlock(&export->exp_lock);
1226                 if (req->rq_peer.nid != export->exp_connection->c_peer.nid) {
1227                         /* MDS or LWP reconnected after failover. */
1228                         LCONSOLE_WARN("%s: Received %s connection from %s, removing former export from %s\n",
1229                                       target->obd_name,
1230                                       mds_conn ? "MDS" : "LWP",
1231                                       libcfs_nid2str(req->rq_peer.nid),
1232                                       libcfs_nid2str(export->exp_connection->c_peer.nid));
1233                 } else {
1234                         /* New MDS connection from the same NID. */
1235                         LCONSOLE_WARN("%s: Received new %s connection from %s, removing former export from same NID\n",
1236                                       target->obd_name,
1237                                       mds_conn ? "MDS" : "LWP",
1238                                       libcfs_nid2str(req->rq_peer.nid));
1239                 }
1240
1241                 if (req->rq_peer.nid == export->exp_connection->c_peer.nid &&
1242                     data->ocd_connect_flags & OBD_CONNECT_MDS_MDS) {
1243                         /*
1244                          * Because exports between MDTs will always be
1245                          * kept, let's do not fail such export if they
1246                          * come from the same NID, otherwise it might
1247                          * cause eviction between MDTs, which might
1248                          * cause namespace inconsistency
1249                          */
1250                         spin_lock(&export->exp_lock);
1251                         export->exp_connecting = 1;
1252                         export->exp_conn_cnt = 0;
1253                         spin_unlock(&export->exp_lock);
1254                         conn.cookie = export->exp_handle.h_cookie;
1255                         rc = EALREADY;
1256                 } else {
1257                         class_fail_export(export);
1258                         class_export_put(export);
1259                         export = NULL;
1260                         rc = 0;
1261                 }
1262         } else if (export->exp_connection != NULL && initial_conn &&
1263                    req->rq_peer.nid != export->exp_connection->c_peer.nid) {
1264                 spin_unlock(&export->exp_lock);
1265                 /* In MDS failover we have static UUID but NID can change. */
1266                 LCONSOLE_WARN("%s: Client %s seen on new nid %s when existing nid %s is already connected\n",
1267                               target->obd_name, cluuid.uuid,
1268                               libcfs_nid2str(req->rq_peer.nid),
1269                               libcfs_nid2str(
1270                                         export->exp_connection->c_peer.nid));
1271                 rc = -EALREADY;
1272                 class_export_put(export);
1273                 export = NULL;
1274         } else if (OBD_FAIL_PRECHECK(OBD_FAIL_TGT_RECOVERY_CONNECT) &&
1275                    !lw_client) {
1276                 spin_unlock(&export->exp_lock);
1277                 rc = -EAGAIN;
1278         } else {
1279                 export->exp_connecting = 1;
1280                 spin_unlock(&export->exp_lock);
1281                 LASSERT(export->exp_obd == target);
1282
1283                 rc = target_handle_reconnect(&conn, export, &cluuid);
1284         }
1285
1286         /* If we found an export, we already unlocked. */
1287         if (!export) {
1288 no_export:
1289                 OBD_FAIL_TIMEOUT(OBD_FAIL_TGT_DELAY_CONNECT, 2 * obd_timeout);
1290         } else if (req->rq_export == NULL &&
1291                    atomic_read(&export->exp_rpc_count) > 0) {
1292                 LCONSOLE_WARN("%s: Client %s (at %s) refused connection, still busy with %d references\n",
1293                               target->obd_name, cluuid.uuid,
1294                               libcfs_nid2str(req->rq_peer.nid),
1295                               refcount_read(&export->exp_handle.h_ref));
1296                         GOTO(out, rc = -EBUSY);
1297         } else if (lustre_msg_get_conn_cnt(req->rq_reqmsg) == 1 &&
1298                    rc != EALREADY) {
1299                 if (!strstr(cluuid.uuid, "mdt"))
1300                         LCONSOLE_WARN("%s: Rejecting reconnect from the known client %s (at %s) because it is indicating it is a new client",
1301                                       target->obd_name, cluuid.uuid,
1302                                       libcfs_nid2str(req->rq_peer.nid));
1303                 GOTO(out, rc = -EALREADY);
1304         } else {
1305                 OBD_FAIL_TIMEOUT(OBD_FAIL_TGT_DELAY_RECONNECT, 2 * obd_timeout);
1306         }
1307
1308         if (rc < 0)
1309                 GOTO(out, rc);
1310
1311         CDEBUG(D_HA, "%s: connection from %s@%s %st%llu exp %p cur %lld last %lld\n",
1312                target->obd_name, cluuid.uuid, libcfs_nid2str(req->rq_peer.nid),
1313                target->obd_recovering ? "recovering/" : "", data->ocd_transno,
1314                export, ktime_get_seconds(),
1315                export ? export->exp_last_request_time : 0);
1316
1317         /*
1318          * If this is the first time a client connects, reset the recovery
1319          * timer. Discard lightweight connections which might be local.
1320          */
1321         if (!lw_client && rc == 0 && target->obd_recovering)
1322                 check_and_start_recovery_timer(target, req, export == NULL);
1323
1324         /*
1325          * We want to handle EALREADY but *not* -EALREADY from
1326          * target_handle_reconnect(), return reconnection state in a flag.
1327          */
1328         if (rc == EALREADY) {
1329                 lustre_msg_add_op_flags(req->rq_repmsg, MSG_CONNECT_RECONNECT);
1330                 rc = 0;
1331         } else {
1332                 LASSERT(rc == 0);
1333         }
1334
1335         /* Tell the client if we support replayable requests. */
1336         if (target->obd_replayable)
1337                 lustre_msg_add_op_flags(req->rq_repmsg, MSG_CONNECT_REPLAYABLE);
1338         client_nid = &req->rq_peer.nid;
1339
1340         if (export == NULL) {
1341                 /* allow lightweight connections during recovery */
1342                 /*
1343                  * allow "new" MDT to be connected during recovery, since we
1344                  * need retrieve recovery update records from it
1345                  */
1346                 if (target->obd_recovering && !lw_client && !mds_mds_conn) {
1347                         struct hrtimer *timer = &target->obd_recovery_timer;
1348                         ktime_t remaining;
1349                         s64 timeout, left;
1350                         int in_progress;
1351                         int connected;
1352                         int known;
1353                         int stale;
1354                         char *msg;
1355
1356                         connected = atomic_read(&target->obd_connected_clients);
1357                         in_progress = atomic_read(&target->obd_lock_replay_clients);
1358                         known =
1359                            atomic_read(&target->obd_max_recoverable_clients);
1360                         stale = target->obd_stale_clients;
1361                         remaining = hrtimer_expires_remaining(timer);
1362                         left = ktime_divns(remaining, NSEC_PER_SEC);
1363
1364                         if (ktime_to_ns(remaining) > 0) {
1365                                 msg = "to recover in";
1366                                 timeout = left;
1367                         } else {
1368                                 msg = "already passed deadline";
1369                                 timeout = -left;
1370
1371                                 target_check_recovery_timer(target);
1372                         }
1373
1374                         LCONSOLE_WARN("%s: Denying connection for new client %s (at %s), waiting for %d known clients (%d recovered, %d in progress, and %d evicted) %s %lld:%.02lld\n",
1375                                       target->obd_name, cluuid.uuid,
1376                                       libcfs_nid2str(req->rq_peer.nid), known,
1377                                       connected - in_progress, in_progress,
1378                                       stale, msg, timeout / 60, timeout % 60);
1379                         rc = -EBUSY;
1380                 } else {
1381 dont_check_exports:
1382                         rc = obd_connect(req->rq_svc_thread->t_env,
1383                                          &export, target, &cluuid, data,
1384                                          client_nid);
1385                         if (mds_conn && OBD_FAIL_CHECK(OBD_FAIL_TGT_RCVG_FLAG))
1386                                 lustre_msg_add_op_flags(req->rq_repmsg,
1387                                                         MSG_CONNECT_RECOVERING);
1388                         if (rc == 0) {
1389                                 conn.cookie = export->exp_handle.h_cookie;
1390                                 rc = rev_import_init(export);
1391                         }
1392
1393                         if (mds_mds_conn)
1394                                 new_mds_mds_conn = true;
1395                 }
1396         } else {
1397                 rc = obd_reconnect(req->rq_svc_thread->t_env,
1398                                    export, target, &cluuid, data, client_nid);
1399         }
1400         if (rc)
1401                 GOTO(out, rc);
1402
1403         LASSERT(target->u.obt.obt_magic == OBT_MAGIC);
1404         data->ocd_instance = target->u.obt.obt_instance;
1405
1406         /*
1407          * Return only the parts of obd_connect_data that we understand, so the
1408          * client knows that we don't understand the rest.
1409          */
1410         if (data) {
1411                 tmpsize = req_capsule_get_size(&req->rq_pill, &RMF_CONNECT_DATA,
1412                                                RCL_SERVER);
1413                 tmpdata = req_capsule_server_get(&req->rq_pill,
1414                                                  &RMF_CONNECT_DATA);
1415                 /*
1416                  * Don't use struct assignment here, because the client reply
1417                  * buffer may be smaller/larger than the local struct
1418                  * obd_connect_data.
1419                  */
1420                 memcpy(tmpdata, data, min(tmpsize, size));
1421         }
1422
1423         /*
1424          * If the client and the server are the same node, we will already
1425          * have an export that really points to the client's DLM export,
1426          * because we have a shared handles table.
1427          *
1428          * XXX this will go away when shaver stops sending the "connect" handle
1429          * in the real "remote handle" field of the request --phik 24 Apr 2003
1430          */
1431         ptlrpc_request_change_export(req, export);
1432
1433         spin_lock(&export->exp_lock);
1434         if (export->exp_conn_cnt >= lustre_msg_get_conn_cnt(req->rq_reqmsg)) {
1435                 spin_unlock(&export->exp_lock);
1436                 CDEBUG(D_RPCTRACE,
1437                        "%s: %s already connected at greater or equal conn_cnt: %d >= %d\n",
1438                        cluuid.uuid, libcfs_nid2str(req->rq_peer.nid),
1439                        export->exp_conn_cnt,
1440                        lustre_msg_get_conn_cnt(req->rq_reqmsg));
1441
1442                 GOTO(out, rc = -EALREADY);
1443         }
1444         LASSERT(lustre_msg_get_conn_cnt(req->rq_reqmsg) > 0);
1445         export->exp_conn_cnt = lustre_msg_get_conn_cnt(req->rq_reqmsg);
1446         spin_unlock(&export->exp_lock);
1447
1448         if (export->exp_connection != NULL) {
1449                 /* Check to see if connection came from another NID. */
1450                 if (export->exp_connection->c_peer.nid != req->rq_peer.nid)
1451                         obd_nid_del(export->exp_obd, export);
1452
1453                 ptlrpc_connection_put(export->exp_connection);
1454         }
1455
1456         export->exp_connection = ptlrpc_connection_get(req->rq_peer,
1457                                                        req->rq_self,
1458                                                        &cluuid);
1459         if (!export->exp_connection)
1460                 GOTO(out, rc = -ENOTCONN);
1461
1462         obd_nid_add(export->exp_obd, export);
1463
1464         lustre_msg_set_handle(req->rq_repmsg, &conn);
1465
1466         rc = rev_import_reconnect(export, req);
1467         if (rc != 0)
1468                 GOTO(out, rc);
1469
1470         if (target->obd_recovering && !export->exp_in_recovery && !lw_client) {
1471                 int has_transno;
1472                 __u64 transno = data->ocd_transno;
1473
1474                 spin_lock(&export->exp_lock);
1475                 /*
1476                  * possible race with class_disconnect_stale_exports,
1477                  * export may be already in the eviction process
1478                  */
1479                 if (export->exp_failed) {
1480                         spin_unlock(&export->exp_lock);
1481                         GOTO(out, rc = -ENODEV);
1482                 }
1483                 export->exp_in_recovery = 1;
1484                 export->exp_req_replay_needed = 1;
1485                 export->exp_lock_replay_needed = 1;
1486                 spin_unlock(&export->exp_lock);
1487
1488                 has_transno = !!(lustre_msg_get_op_flags(req->rq_reqmsg) &
1489                                  MSG_CONNECT_TRANSNO);
1490                 if (has_transno && transno == 0)
1491                         CWARN("Connect with zero transno!\n");
1492
1493                 if (has_transno && transno > 0 &&
1494                     transno < target->obd_next_recovery_transno &&
1495                     transno > target->obd_last_committed) {
1496                         /* Another way is to use cmpxchg() to be lock-free. */
1497                         spin_lock(&target->obd_recovery_task_lock);
1498                         if (transno < target->obd_next_recovery_transno)
1499                                 target->obd_next_recovery_transno = transno;
1500                         spin_unlock(&target->obd_recovery_task_lock);
1501                 }
1502
1503                 atomic_inc(&target->obd_req_replay_clients);
1504                 atomic_inc(&target->obd_lock_replay_clients);
1505                 /*
1506                  * Note: MDS-MDS connection is allowed to be connected during
1507                  * recovery, no matter if the exports needs to be recoveried.
1508                  * Because we need retrieve updates logs from all other MDTs.
1509                  * So if the MDS-MDS export is new, obd_max_recoverable_clients
1510                  * also needs to be increased to match other recovery checking
1511                  * condition.
1512                  */
1513                 if (new_mds_mds_conn)
1514                         atomic_inc(&target->obd_max_recoverable_clients);
1515
1516                 if (atomic_inc_return(&target->obd_connected_clients) ==
1517                     atomic_read(&target->obd_max_recoverable_clients))
1518                         wake_up(&target->obd_next_transno_waitq);
1519         }
1520
1521         /* Tell the client we're in recovery, when client is involved in it. */
1522         if (target->obd_recovering && !lw_client)
1523                 lustre_msg_add_op_flags(req->rq_repmsg, MSG_CONNECT_RECOVERING);
1524
1525 out:
1526         if (export) {
1527                 spin_lock(&export->exp_lock);
1528                 export->exp_connecting = 0;
1529                 spin_unlock(&export->exp_lock);
1530
1531                 class_export_put(export);
1532         }
1533         if (target != NULL) {
1534                 spin_lock(&target->obd_dev_lock);
1535                 target->obd_conn_inprogress--;
1536                 spin_unlock(&target->obd_dev_lock);
1537                 class_decref(target, "find", current);
1538         }
1539         req->rq_status = rc;
1540         RETURN(rc);
1541 }
1542
1543 int target_handle_disconnect(struct ptlrpc_request *req)
1544 {
1545         int rc;
1546
1547         ENTRY;
1548
1549         rc = req_capsule_server_pack(&req->rq_pill);
1550         if (rc)
1551                 RETURN(rc);
1552
1553         /* Keep the rq_export around so we can send the reply. */
1554         req->rq_status = obd_disconnect(class_export_get(req->rq_export));
1555
1556         RETURN(0);
1557 }
1558
1559 void target_destroy_export(struct obd_export *exp)
1560 {
1561         struct obd_import *imp = NULL;
1562         /*
1563          * exports created from last_rcvd data, and "fake"
1564          * exports created by lctl don't have an import
1565          */
1566         spin_lock(&exp->exp_lock);
1567         if (exp->exp_imp_reverse != NULL) {
1568                 imp = exp->exp_imp_reverse;
1569                 exp->exp_imp_reverse = NULL;
1570         }
1571         spin_unlock(&exp->exp_lock);
1572         if (imp != NULL)
1573                 client_destroy_import(imp);
1574
1575         LASSERT_ATOMIC_ZERO(&exp->exp_locks_count);
1576         LASSERT_ATOMIC_ZERO(&exp->exp_rpc_count);
1577         LASSERT_ATOMIC_ZERO(&exp->exp_cb_count);
1578         LASSERT_ATOMIC_ZERO(&exp->exp_replay_count);
1579 }
1580 EXPORT_SYMBOL(target_destroy_export);
1581
1582 /*
1583  * Recovery functions
1584  */
1585 static void target_request_copy_get(struct ptlrpc_request *req)
1586 {
1587         class_export_rpc_inc(req->rq_export);
1588         LASSERT(list_empty(&req->rq_list));
1589         INIT_LIST_HEAD(&req->rq_replay_list);
1590
1591         /* Increase refcount to keep request in queue. */
1592         atomic_inc(&req->rq_refcount);
1593         /* Let export know it has replays to be handled. */
1594         atomic_inc(&req->rq_export->exp_replay_count);
1595 }
1596
1597 static void target_request_copy_put(struct ptlrpc_request *req)
1598 {
1599         LASSERT(list_empty(&req->rq_replay_list));
1600         LASSERT_ATOMIC_POS(&req->rq_export->exp_replay_count);
1601
1602         atomic_dec(&req->rq_export->exp_replay_count);
1603         class_export_rpc_dec(req->rq_export);
1604         ptlrpc_server_drop_request(req);
1605 }
1606
1607 static int target_exp_enqueue_req_replay(struct ptlrpc_request *req)
1608 {
1609         __u64 transno = lustre_msg_get_transno(req->rq_reqmsg);
1610         struct obd_export *exp = req->rq_export;
1611         struct ptlrpc_request *reqiter;
1612         struct ptlrpc_request *dup_req = NULL;
1613         int dup = 0;
1614
1615         LASSERT(exp);
1616
1617         spin_lock(&exp->exp_lock);
1618         list_for_each_entry(reqiter, &exp->exp_req_replay_queue,
1619                             rq_replay_list) {
1620                 if (lustre_msg_get_transno(reqiter->rq_reqmsg) == transno) {
1621                         dup_req = reqiter;
1622                         dup = 1;
1623                         break;
1624                 }
1625         }
1626
1627         if (dup) {
1628                 /* We expect it with RESENT and REPLAY flags. */
1629                 if ((lustre_msg_get_flags(req->rq_reqmsg) &
1630                     (MSG_RESENT | MSG_REPLAY)) != (MSG_RESENT | MSG_REPLAY))
1631                         CERROR("invalid flags %x of resent replay\n",
1632                                lustre_msg_get_flags(req->rq_reqmsg));
1633
1634                 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY) {
1635                         __u32 new_conn;
1636
1637                         new_conn = lustre_msg_get_conn_cnt(req->rq_reqmsg);
1638                         if (new_conn >
1639                             lustre_msg_get_conn_cnt(dup_req->rq_reqmsg))
1640                                 lustre_msg_set_conn_cnt(dup_req->rq_reqmsg,
1641                                                         new_conn);
1642                 }
1643         } else {
1644                 list_add_tail(&req->rq_replay_list,
1645                               &exp->exp_req_replay_queue);
1646         }
1647
1648         spin_unlock(&exp->exp_lock);
1649         return dup;
1650 }
1651
1652 static void target_exp_dequeue_req_replay(struct ptlrpc_request *req)
1653 {
1654         LASSERT(!list_empty(&req->rq_replay_list));
1655         LASSERT(req->rq_export);
1656
1657         spin_lock(&req->rq_export->exp_lock);
1658         list_del_init(&req->rq_replay_list);
1659         spin_unlock(&req->rq_export->exp_lock);
1660 }
1661
1662 static void target_finish_recovery(struct lu_target *lut)
1663 {
1664         struct obd_device *obd = lut->lut_obd;
1665
1666         ENTRY;
1667
1668         /* Only log a recovery message when recovery has occurred. */
1669         if (obd->obd_recovery_start) {
1670                 time64_t now = ktime_get_seconds();
1671                 time64_t elapsed_time;
1672
1673                 elapsed_time = max_t(time64_t, now - obd->obd_recovery_start,
1674                                      1);
1675                 LCONSOLE_INFO("%s: Recovery over after %lld:%.02lld, of %d clients %d recovered and %d %s evicted.\n",
1676                               obd->obd_name, elapsed_time / 60,
1677                               elapsed_time % 60,
1678                               atomic_read(&obd->obd_max_recoverable_clients),
1679                               atomic_read(&obd->obd_connected_clients),
1680                               obd->obd_stale_clients,
1681                               obd->obd_stale_clients == 1 ? "was" : "were");
1682         }
1683
1684         ldlm_reprocess_recovery_done(obd->obd_namespace);
1685         spin_lock(&obd->obd_recovery_task_lock);
1686         if (!list_empty(&obd->obd_req_replay_queue) ||
1687             !list_empty(&obd->obd_lock_replay_queue) ||
1688             !list_empty(&obd->obd_final_req_queue)) {
1689                 CERROR("%s: Recovery queues ( %s%s%s) are not empty\n",
1690                        obd->obd_name,
1691                        list_empty(&obd->obd_req_replay_queue) ? "" : "req ",
1692                        list_empty(&obd->obd_lock_replay_queue) ? \
1693                                   "" : "lock ",
1694                        list_empty(&obd->obd_final_req_queue) ? \
1695                                   "" : "final ");
1696                 spin_unlock(&obd->obd_recovery_task_lock);
1697                 LBUG();
1698         }
1699         spin_unlock(&obd->obd_recovery_task_lock);
1700
1701         obd->obd_recovery_end = ktime_get_seconds();
1702
1703         /* When recovery finished, cleanup orphans on MDS and OST. */
1704         if (obd->obd_type && OBP(obd, postrecov)) {
1705                 int rc = OBP(obd, postrecov)(obd);
1706
1707                 if (rc < 0)
1708                         LCONSOLE_WARN("%s: Post recovery failed, rc %d\n",
1709                                       obd->obd_name, rc);
1710         }
1711         EXIT;
1712 }
1713
1714 static void abort_req_replay_queue(struct obd_device *obd)
1715 {
1716         struct ptlrpc_request *req, *n;
1717         LIST_HEAD(abort_list);
1718
1719         spin_lock(&obd->obd_recovery_task_lock);
1720         list_splice_init(&obd->obd_req_replay_queue, &abort_list);
1721         spin_unlock(&obd->obd_recovery_task_lock);
1722         list_for_each_entry_safe(req, n, &abort_list, rq_list) {
1723                 DEBUG_REQ(D_WARNING, req, "aborted:");
1724                 req->rq_status = -ENOTCONN;
1725                 if (ptlrpc_error(req)) {
1726                         DEBUG_REQ(D_ERROR, req,
1727                                   "failed abort_req_reply; skipping");
1728                 }
1729                 target_exp_dequeue_req_replay(req);
1730                 target_request_copy_put(req);
1731         }
1732 }
1733
1734 static void abort_lock_replay_queue(struct obd_device *obd)
1735 {
1736         struct ptlrpc_request *req, *n;
1737         LIST_HEAD(abort_list);
1738
1739         spin_lock(&obd->obd_recovery_task_lock);
1740         list_splice_init(&obd->obd_lock_replay_queue, &abort_list);
1741         spin_unlock(&obd->obd_recovery_task_lock);
1742         list_for_each_entry_safe(req, n, &abort_list, rq_list) {
1743                 DEBUG_REQ(D_ERROR, req, "aborted:");
1744                 req->rq_status = -ENOTCONN;
1745                 if (ptlrpc_error(req)) {
1746                         DEBUG_REQ(D_ERROR, req,
1747                                   "failed abort_lock_reply; skipping");
1748                 }
1749                 target_request_copy_put(req);
1750         }
1751 }
1752
1753 /*
1754  * Called from a cleanup function if the device is being cleaned up
1755  * forcefully.  The exports should all have been disconnected already,
1756  * the only thing left to do is
1757  * - clear the recovery flags
1758  * - cancel the timer
1759  * - free queued requests and replies, but don't send replies
1760  * Because the obd_stopping flag is set, no new requests should be received.
1761  */
1762 void target_cleanup_recovery(struct obd_device *obd)
1763 {
1764         struct ptlrpc_request *req, *n;
1765         LIST_HEAD(clean_list);
1766
1767         spin_lock(&obd->obd_dev_lock);
1768         if (!obd->obd_recovering) {
1769                 spin_unlock(&obd->obd_dev_lock);
1770                 EXIT;
1771                 return;
1772         }
1773         obd->obd_recovering = obd->obd_abort_recovery = 0;
1774         spin_unlock(&obd->obd_dev_lock);
1775
1776         spin_lock(&obd->obd_recovery_task_lock);
1777         target_cancel_recovery_timer(obd);
1778         list_splice_init(&obd->obd_req_replay_queue, &clean_list);
1779         spin_unlock(&obd->obd_recovery_task_lock);
1780
1781         list_for_each_entry_safe(req, n, &clean_list, rq_list) {
1782                 LASSERT(req->rq_reply_state == NULL);
1783                 target_exp_dequeue_req_replay(req);
1784                 target_request_copy_put(req);
1785         }
1786
1787         spin_lock(&obd->obd_recovery_task_lock);
1788         list_splice_init(&obd->obd_lock_replay_queue, &clean_list);
1789         list_splice_init(&obd->obd_final_req_queue, &clean_list);
1790         spin_unlock(&obd->obd_recovery_task_lock);
1791
1792         list_for_each_entry_safe(req, n, &clean_list, rq_list) {
1793                 LASSERT(req->rq_reply_state == NULL);
1794                 target_request_copy_put(req);
1795         }
1796
1797         EXIT;
1798 }
1799 EXPORT_SYMBOL(target_cleanup_recovery);
1800
1801 /* obd_recovery_task_lock should be held */
1802 void target_cancel_recovery_timer(struct obd_device *obd)
1803 {
1804         CDEBUG(D_HA, "%s: cancel recovery timer\n", obd->obd_name);
1805         hrtimer_cancel(&obd->obd_recovery_timer);
1806 }
1807
1808 static void target_start_recovery_timer(struct obd_device *obd)
1809 {
1810         ktime_t delay;
1811
1812         if (obd->obd_recovery_start != 0)
1813                 return;
1814
1815         spin_lock(&obd->obd_dev_lock);
1816         if (!obd->obd_recovering || obd->obd_abort_recovery) {
1817                 spin_unlock(&obd->obd_dev_lock);
1818                 return;
1819         }
1820
1821         LASSERT(obd->obd_recovery_timeout != 0);
1822
1823         if (obd->obd_recovery_start != 0) {
1824                 spin_unlock(&obd->obd_dev_lock);
1825                 return;
1826         }
1827
1828         obd->obd_recovery_start = ktime_get_seconds();
1829         delay = ktime_set(obd->obd_recovery_start +
1830                           obd->obd_recovery_timeout, 0);
1831         hrtimer_start(&obd->obd_recovery_timer, delay, HRTIMER_MODE_ABS);
1832         spin_unlock(&obd->obd_dev_lock);
1833
1834         LCONSOLE_WARN("%s: Will be in recovery for at least %u:%02u, or until %d client%s reconnect%s\n",
1835                       obd->obd_name,
1836                       obd->obd_recovery_timeout / 60,
1837                       obd->obd_recovery_timeout % 60,
1838                       atomic_read(&obd->obd_max_recoverable_clients),
1839                       (atomic_read(&obd->obd_max_recoverable_clients) == 1) ?
1840                       "" : "s",
1841                       (atomic_read(&obd->obd_max_recoverable_clients) == 1) ?
1842                       "s" : "");
1843 }
1844
1845 /**
1846  * extend recovery window.
1847  *
1848  * if @extend is true, extend recovery window to have @dr_timeout remaining
1849  * at least; otherwise, make sure the recovery timeout value is not less
1850  * than @dr_timeout.
1851  */
1852 static void extend_recovery_timer(struct obd_device *obd, timeout_t dr_timeout,
1853                                   bool extend)
1854 {
1855         ktime_t left_ns;
1856         timeout_t timeout;
1857         timeout_t left;
1858
1859         spin_lock(&obd->obd_dev_lock);
1860         if (!obd->obd_recovering || obd->obd_abort_recovery ||
1861             obd->obd_stopping) {
1862                 spin_unlock(&obd->obd_dev_lock);
1863                 return;
1864         }
1865         LASSERT(obd->obd_recovery_start != 0);
1866
1867         left_ns = hrtimer_expires_remaining(&obd->obd_recovery_timer);
1868         left = ktime_divns(left_ns, NSEC_PER_SEC);
1869
1870         if (extend) {
1871                 timeout = obd->obd_recovery_timeout;
1872                 /* dr_timeout will happen after the hrtimer has expired.
1873                  * Add the excess time to the soft recovery timeout without
1874                  * exceeding the hard recovery timeout.
1875                  */
1876                 if (dr_timeout > left) {
1877                         timeout += dr_timeout - left;
1878                         timeout = min_t(timeout_t, obd->obd_recovery_time_hard,
1879                                         timeout);
1880                 }
1881         } else {
1882                 timeout = clamp_t(timeout_t, dr_timeout,
1883                                   obd->obd_recovery_timeout,
1884                                   obd->obd_recovery_time_hard);
1885         }
1886
1887         if (timeout == obd->obd_recovery_time_hard)
1888                 CWARN("%s: extended recovery timer reached hard limit: %d, extend: %d\n",
1889                       obd->obd_name, timeout, extend);
1890
1891         if (obd->obd_recovery_timeout < timeout) {
1892                 ktime_t end, now;
1893
1894                 obd->obd_recovery_timeout = timeout;
1895                 end = ktime_set(obd->obd_recovery_start + timeout, 0);
1896                 now = ktime_set(ktime_get_seconds(), 0);
1897                 left_ns = ktime_sub(end, now);
1898                 hrtimer_start(&obd->obd_recovery_timer, end, HRTIMER_MODE_ABS);
1899                 left = ktime_divns(left_ns, NSEC_PER_SEC);
1900         }
1901         spin_unlock(&obd->obd_dev_lock);
1902
1903         CDEBUG(D_HA, "%s: recovery timer will expire in %d seconds\n",
1904                 obd->obd_name, left);
1905 }
1906
1907 /* Reset the timer with each new client connection */
1908 /*
1909  * This timer is actually reconnect_timer, which is for making sure
1910  * the total recovery window is at least as big as my reconnect
1911  * attempt timing. So the initial recovery time_out will be set to
1912  * OBD_RECOVERY_FACTOR * obd_timeout. If the timeout coming
1913  * from client is bigger than this, then the recovery time_out will
1914  * be extended to make sure the client could be reconnected, in the
1915  * process, the timeout from the new client should be ignored.
1916  */
1917 static void
1918 check_and_start_recovery_timer(struct obd_device *obd,
1919                                struct ptlrpc_request *req,
1920                                int new_client)
1921 {
1922         timeout_t service_timeout = lustre_msg_get_service_timeout(req->rq_reqmsg);
1923         struct obd_device_target *obt = &obd->u.obt;
1924
1925         if (!new_client && service_timeout)
1926                 /*
1927                  * Teach server about old server's estimates, as first guess
1928                  * at how long new requests will take.
1929                  */
1930                 at_measured(&req->rq_rqbd->rqbd_svcpt->scp_at_estimate,
1931                             service_timeout);
1932
1933         target_start_recovery_timer(obd);
1934
1935         /*
1936          * Convert the service time to RPC timeout,
1937          * and reuse service_timeout to limit stack usage.
1938          */
1939         service_timeout = at_est2timeout(service_timeout);
1940
1941         if (OBD_FAIL_CHECK(OBD_FAIL_TGT_SLUGGISH_NET) &&
1942             service_timeout < at_extra)
1943                 service_timeout = at_extra;
1944
1945         /*
1946          * We expect other clients to timeout within service_timeout, then try
1947          * to reconnect, then try the failover server.  The max delay between
1948          * connect attempts is SWITCH_MAX + SWITCH_INC + INITIAL.
1949          */
1950         service_timeout += 2 * INITIAL_CONNECT_TIMEOUT;
1951
1952         LASSERT(obt->obt_magic == OBT_MAGIC);
1953         service_timeout += 2 * (CONNECTION_SWITCH_MAX + CONNECTION_SWITCH_INC);
1954         if (service_timeout > obd->obd_recovery_timeout && !new_client)
1955                 extend_recovery_timer(obd, service_timeout, false);
1956 }
1957
1958 /** Health checking routines */
1959 static inline int exp_connect_healthy(struct obd_export *exp)
1960 {
1961         return exp->exp_in_recovery;
1962 }
1963
1964 /** if export done req_replay or has replay in queue */
1965 static inline int exp_req_replay_healthy(struct obd_export *exp)
1966 {
1967         return (!exp->exp_req_replay_needed ||
1968                 atomic_read(&exp->exp_replay_count) > 0);
1969 }
1970
1971
1972 static inline int exp_req_replay_healthy_or_from_mdt(struct obd_export *exp)
1973 {
1974         return (exp_connect_flags(exp) & OBD_CONNECT_MDS_MDS) ||
1975                exp_req_replay_healthy(exp);
1976 }
1977
1978 /** if export done lock_replay or has replay in queue */
1979 static inline int exp_lock_replay_healthy(struct obd_export *exp)
1980 {
1981         return (!exp->exp_lock_replay_needed ||
1982                 atomic_read(&exp->exp_replay_count) > 0);
1983 }
1984
1985 static inline int exp_vbr_healthy(struct obd_export *exp)
1986 {
1987         return !exp->exp_vbr_failed;
1988 }
1989
1990 static inline int exp_finished(struct obd_export *exp)
1991 {
1992         return exp->exp_in_recovery && !exp->exp_lock_replay_needed;
1993 }
1994
1995 static inline int exp_finished_or_from_mdt(struct obd_export *exp)
1996 {
1997         return (exp_connect_flags(exp) & OBD_CONNECT_MDS_MDS) ||
1998                 exp_finished(exp);
1999 }
2000
2001 static int check_for_next_transno(struct lu_target *lut)
2002 {
2003         struct ptlrpc_request *req = NULL;
2004         struct obd_device *obd = lut->lut_obd;
2005         struct target_distribute_txn_data *tdtd = lut->lut_tdtd;
2006         int wake_up = 0, connected, completed, queue_len;
2007         __u64 req_transno = 0;
2008         __u64 update_transno = 0;
2009         __u64 next_transno = 0;
2010
2011         ENTRY;
2012
2013         spin_lock(&obd->obd_recovery_task_lock);
2014         if (!list_empty(&obd->obd_req_replay_queue)) {
2015                 req = list_entry(obd->obd_req_replay_queue.next,
2016                                      struct ptlrpc_request, rq_list);
2017                 req_transno = lustre_msg_get_transno(req->rq_reqmsg);
2018         }
2019
2020         if (tdtd != NULL)
2021                 update_transno = distribute_txn_get_next_transno(tdtd);
2022
2023         connected = atomic_read(&obd->obd_connected_clients);
2024         completed = connected - atomic_read(&obd->obd_req_replay_clients);
2025         queue_len = obd->obd_requests_queued_for_recovery;
2026         next_transno = obd->obd_next_recovery_transno;
2027
2028         CDEBUG(D_HA,
2029                "max: %d, connected: %d, completed: %d, queue_len: %d, req_transno: %llu, next_transno: %llu\n",
2030                atomic_read(&obd->obd_max_recoverable_clients),
2031                connected, completed,
2032                queue_len, req_transno, next_transno);
2033
2034         if (obd->obd_abort_recovery) {
2035                 CDEBUG(D_HA, "waking for aborted recovery\n");
2036                 wake_up = 1;
2037         } else if (obd->obd_recovery_expired) {
2038                 CDEBUG(D_HA, "waking for expired recovery\n");
2039                 wake_up = 1;
2040         } else if (tdtd != NULL && req != NULL &&
2041                    is_req_replayed_by_update(req)) {
2042                 LASSERTF(req_transno < next_transno,
2043                          "req_transno %llu next_transno%llu\n", req_transno,
2044                          next_transno);
2045                 CDEBUG(D_HA, "waking for duplicate req (%llu)\n",
2046                        req_transno);
2047                 wake_up = 1;
2048         } else if (req_transno == next_transno ||
2049                    (update_transno != 0 && update_transno <= next_transno)) {
2050                 CDEBUG(D_HA, "waking for next (%lld)\n", next_transno);
2051                 wake_up = 1;
2052         } else if (queue_len > 0 &&
2053                    queue_len == atomic_read(&obd->obd_req_replay_clients)) {
2054                 /** handle gaps occured due to lost reply or VBR */
2055                 LASSERTF(req_transno >= next_transno,
2056                          "req_transno: %llu, next_transno: %llu\n",
2057                          req_transno, next_transno);
2058                 CDEBUG(D_HA,
2059                        "%s: waking for gap in transno, VBR is %s (skip: %lld, ql: %d, comp: %d, conn: %d, next: %lld, next_update %lld last_committed: %lld)\n",
2060                        obd->obd_name, obd->obd_version_recov ? "ON" : "OFF",
2061                        next_transno, queue_len, completed, connected,
2062                        req_transno, update_transno, obd->obd_last_committed);
2063                 obd->obd_next_recovery_transno = req_transno;
2064                 wake_up = 1;
2065         } else if (atomic_read(&obd->obd_req_replay_clients) == 0) {
2066                 CDEBUG(D_HA, "waking for completed recovery\n");
2067                 wake_up = 1;
2068         } else if (OBD_FAIL_CHECK(OBD_FAIL_MDS_RECOVERY_ACCEPTS_GAPS)) {
2069                 CDEBUG(D_HA,
2070                        "accepting transno gaps is explicitly allowed by fail_lock, waking up (%lld)\n",
2071                        next_transno);
2072                 obd->obd_next_recovery_transno = req_transno;
2073                 wake_up = 1;
2074         }
2075         spin_unlock(&obd->obd_recovery_task_lock);
2076         return wake_up;
2077 }
2078
2079 static int check_for_next_lock(struct lu_target *lut)
2080 {
2081         struct obd_device *obd = lut->lut_obd;
2082         int wake_up = 0;
2083
2084         spin_lock(&obd->obd_recovery_task_lock);
2085         if (!list_empty(&obd->obd_lock_replay_queue)) {
2086                 CDEBUG(D_HA, "waking for next lock\n");
2087                 wake_up = 1;
2088         } else if (atomic_read(&obd->obd_lock_replay_clients) == 0) {
2089                 CDEBUG(D_HA, "waking for completed lock replay\n");
2090                 wake_up = 1;
2091         } else if (obd->obd_abort_recovery) {
2092                 CDEBUG(D_HA, "waking for aborted recovery\n");
2093                 wake_up = 1;
2094         } else if (obd->obd_recovery_expired) {
2095                 CDEBUG(D_HA, "waking for expired recovery\n");
2096                 wake_up = 1;
2097         }
2098         spin_unlock(&obd->obd_recovery_task_lock);
2099
2100         return wake_up;
2101 }
2102
2103 static int check_update_llog(struct lu_target *lut)
2104 {
2105         struct obd_device *obd = lut->lut_obd;
2106         struct target_distribute_txn_data *tdtd = lut->lut_tdtd;
2107
2108         if (obd->obd_abort_recovery) {
2109                 CDEBUG(D_HA, "waking for aborted recovery\n");
2110                 return 1;
2111         }
2112
2113         if (atomic_read(&tdtd->tdtd_recovery_threads_count) == 0) {
2114                 CDEBUG(D_HA, "waking for completion of reading update log\n");
2115                 return 1;
2116         }
2117
2118         return 0;
2119 }
2120
2121 /**
2122  * wait for recovery events,
2123  * check its status with help of check_routine
2124  * evict dead clients via health_check
2125  */
2126 static int target_recovery_overseer(struct lu_target *lut,
2127                                     int (*check_routine)(struct lu_target *),
2128                                     int (*health_check)(struct obd_export *))
2129 {
2130         struct obd_device *obd = lut->lut_obd;
2131         struct target_distribute_txn_data *tdtd;
2132         time64_t last = 0;
2133         time64_t now;
2134 repeat:
2135         if (obd->obd_recovering && obd->obd_recovery_start == 0) {
2136                 now = ktime_get_seconds();
2137                 if (now - last > 600) {
2138                         LCONSOLE_INFO("%s: in recovery but waiting for the first client to connect\n",
2139                                       obd->obd_name);
2140                         last = now;
2141                 }
2142         }
2143         if (obd->obd_recovery_start != 0 && ktime_get_seconds() >=
2144               (obd->obd_recovery_start + obd->obd_recovery_time_hard)) {
2145                 __u64 next_update_transno = 0;
2146
2147                 /*
2148                  * Only abort the recovery if there are no update recovery
2149                  * left in the queue
2150                  */
2151                 spin_lock(&obd->obd_recovery_task_lock);
2152                 if (lut->lut_tdtd != NULL) {
2153                         next_update_transno =
2154                                 distribute_txn_get_next_transno(lut->lut_tdtd);
2155
2156                         tdtd = lut->lut_tdtd;
2157                         /*
2158                          * If next_update_transno == 0, it probably because
2159                          * updatelog retrieve threads did not get any records
2160                          * yet, let's wait those threads stopped
2161                          */
2162                         if (next_update_transno == 0) {
2163                                 spin_unlock(&obd->obd_recovery_task_lock);
2164
2165                                 while (wait_event_timeout(
2166                                         tdtd->tdtd_recovery_threads_waitq,
2167                                         check_update_llog(lut),
2168                                         cfs_time_seconds(60)) == 0);
2169
2170                                 spin_lock(&obd->obd_recovery_task_lock);
2171                                 next_update_transno =
2172                                         distribute_txn_get_next_transno(tdtd);
2173                         }
2174                 }
2175
2176                 if (next_update_transno != 0 && !obd->obd_abort_recovery) {
2177                         obd->obd_next_recovery_transno = next_update_transno;
2178                         spin_unlock(&obd->obd_recovery_task_lock);
2179                         /*
2180                          * Disconnect unfinished exports from clients, and
2181                          * keep connection from MDT to make sure the update
2182                          * recovery will still keep trying until some one
2183                          * manually abort the recovery
2184                          */
2185                         class_disconnect_stale_exports(obd,
2186                                                 exp_finished_or_from_mdt);
2187                         /* Abort all of replay & replay lock req from clients */
2188                         abort_req_replay_queue(obd);
2189                         abort_lock_replay_queue(obd);
2190                         CDEBUG(D_HA,
2191                                "%s: there are still update replay (%#llx)in the queue.\n",
2192                                obd->obd_name, next_update_transno);
2193                 } else {
2194                         obd->obd_abort_recovery = 1;
2195                         spin_unlock(&obd->obd_recovery_task_lock);
2196                         CWARN("%s recovery is aborted by hard timeout\n",
2197                               obd->obd_name);
2198                 }
2199         }
2200
2201         while (wait_event_timeout(obd->obd_next_transno_waitq,
2202                                   check_routine(lut),
2203                                   cfs_time_seconds(60)) == 0)
2204                 ; /* wait indefinitely for event, but don't trigger watchdog */
2205
2206         if (obd->obd_abort_recovery) {
2207                 CWARN("recovery is aborted, evict exports in recovery\n");
2208                 if (lut->lut_tdtd != NULL) {
2209                         tdtd = lut->lut_tdtd;
2210                         /*
2211                          * Let's wait all of the update log recovery thread
2212                          * finished
2213                          */
2214                         wait_event_idle(
2215                                 tdtd->tdtd_recovery_threads_waitq,
2216                                 atomic_read(&tdtd->tdtd_recovery_threads_count)
2217                                 == 0);
2218                         /* Then abort the update recovery list */
2219                         dtrq_list_destroy(lut->lut_tdtd);
2220                 }
2221
2222                 /** evict exports which didn't finish recovery yet */
2223                 class_disconnect_stale_exports(obd, exp_finished);
2224                 return 1;
2225         } else if (obd->obd_recovery_expired &&
2226                    obd->obd_recovery_timeout < obd->obd_recovery_time_hard) {
2227                 obd->obd_recovery_expired = 0;
2228
2229                 /** If some clients died being recovered, evict them */
2230                 LCONSOLE_WARN("%s: recovery is timed out, evict stale exports\n",
2231                               obd->obd_name);
2232                 /** evict cexports with no replay in queue, they are stalled */
2233                 class_disconnect_stale_exports(obd, health_check);
2234
2235                 /** continue with VBR */
2236                 spin_lock(&obd->obd_dev_lock);
2237                 obd->obd_version_recov = 1;
2238                 spin_unlock(&obd->obd_dev_lock);
2239                 /**
2240                  * reset timer, recovery will proceed with versions now,
2241                  * timeout is set just to handle reconnection delays
2242                  */
2243                 extend_recovery_timer(obd, RECONNECT_DELAY_MAX, true);
2244                 /**
2245                  * Wait for recovery events again, after evicting bad clients
2246                  */
2247                 goto repeat;
2248         }
2249         return 0;
2250 }
2251
2252 static struct ptlrpc_request *target_next_replay_lock(struct lu_target *lut)
2253 {
2254         struct obd_device *obd = lut->lut_obd;
2255         struct ptlrpc_request *req = NULL;
2256
2257         CDEBUG(D_HA, "Waiting for lock\n");
2258         if (target_recovery_overseer(lut, check_for_next_lock,
2259                                      exp_lock_replay_healthy))
2260                 abort_lock_replay_queue(obd);
2261
2262         spin_lock(&obd->obd_recovery_task_lock);
2263         if (!list_empty(&obd->obd_lock_replay_queue)) {
2264                 req = list_entry(obd->obd_lock_replay_queue.next,
2265                                      struct ptlrpc_request, rq_list);
2266                 list_del_init(&req->rq_list);
2267                 spin_unlock(&obd->obd_recovery_task_lock);
2268         } else {
2269                 spin_unlock(&obd->obd_recovery_task_lock);
2270                 LASSERT(list_empty(&obd->obd_lock_replay_queue));
2271                 LASSERT(atomic_read(&obd->obd_lock_replay_clients) == 0);
2272                 /** evict exports failed VBR */
2273                 class_disconnect_stale_exports(obd, exp_vbr_healthy);
2274         }
2275         return req;
2276 }
2277
2278 static struct ptlrpc_request *target_next_final_ping(struct obd_device *obd)
2279 {
2280         struct ptlrpc_request *req = NULL;
2281
2282         spin_lock(&obd->obd_recovery_task_lock);
2283         if (!list_empty(&obd->obd_final_req_queue)) {
2284                 req = list_entry(obd->obd_final_req_queue.next,
2285                                      struct ptlrpc_request, rq_list);
2286                 list_del_init(&req->rq_list);
2287                 spin_unlock(&obd->obd_recovery_task_lock);
2288                 if (req->rq_export->exp_in_recovery) {
2289                         spin_lock(&req->rq_export->exp_lock);
2290                         req->rq_export->exp_in_recovery = 0;
2291                         spin_unlock(&req->rq_export->exp_lock);
2292                 }
2293         } else {
2294                 spin_unlock(&obd->obd_recovery_task_lock);
2295         }
2296         return req;
2297 }
2298
2299 static void handle_recovery_req(struct ptlrpc_thread *thread,
2300                                 struct ptlrpc_request *req,
2301                                 svc_handler_t handler)
2302 {
2303         ENTRY;
2304
2305         /**
2306          * export can be evicted during recovery, no need to handle replays for
2307          * it after that, discard such request silently
2308          */
2309         if (req->rq_export->exp_disconnected)
2310                 RETURN_EXIT;
2311
2312         req->rq_session.lc_thread = thread;
2313         req->rq_svc_thread = thread;
2314         req->rq_svc_thread->t_env->le_ses = &req->rq_session;
2315
2316         /* thread context */
2317         lu_context_enter(&thread->t_env->le_ctx);
2318         (void)handler(req);
2319         lu_context_exit(&thread->t_env->le_ctx);
2320
2321         req->rq_svc_thread->t_env->le_ses = NULL;
2322
2323         /* don't reset timer for final stage */
2324         if (!exp_finished(req->rq_export)) {
2325                 timeout_t timeout = obd_timeout;
2326
2327                 /**
2328                  * Add request @timeout to the recovery time so next request from
2329                  * this client may come in recovery time
2330                  */
2331                 if (!AT_OFF) {
2332                         struct ptlrpc_service_part *svcpt;
2333                         timeout_t est_timeout;
2334
2335                         svcpt = req->rq_rqbd->rqbd_svcpt;
2336                         /*
2337                          * If the server sent early reply for this request,
2338                          * the client will recalculate the timeout according to
2339                          * current server estimate service time, so we will
2340                          * use the maxium timeout here for waiting the client
2341                          * sending the next req
2342                          */
2343                         est_timeout = at_get(&svcpt->scp_at_estimate);
2344                         timeout = max_t(timeout_t, at_est2timeout(est_timeout),
2345                                         lustre_msg_get_timeout(req->rq_reqmsg));
2346                         /*
2347                          * Add 2 net_latency, one for balance rq_deadline
2348                          * (see ptl_send_rpc), one for resend the req to server,
2349                          * Note: client will pack net_latency in replay req
2350                          * (see ptlrpc_replay_req)
2351                          */
2352                         timeout += 2 * lustre_msg_get_service_timeout(req->rq_reqmsg);
2353                 }
2354                 extend_recovery_timer(class_exp2obd(req->rq_export), timeout,
2355                                       true);
2356         }
2357         EXIT;
2358 }
2359
2360 /** Checking routines for recovery */
2361 static int check_for_recovery_ready(struct lu_target *lut)
2362 {
2363         struct obd_device *obd = lut->lut_obd;
2364         unsigned int clnts = atomic_read(&obd->obd_connected_clients);
2365
2366         CDEBUG(D_HA,
2367                "connected %d stale %d max_recoverable_clients %d abort %d expired %d\n",
2368                clnts, obd->obd_stale_clients,
2369                atomic_read(&obd->obd_max_recoverable_clients),
2370                obd->obd_abort_recovery, obd->obd_recovery_expired);
2371
2372         if (!obd->obd_abort_recovery && !obd->obd_recovery_expired) {
2373                 LASSERT(clnts <=
2374                         atomic_read(&obd->obd_max_recoverable_clients));
2375                 if (clnts + obd->obd_stale_clients <
2376                     atomic_read(&obd->obd_max_recoverable_clients))
2377                         return 0;
2378         }
2379
2380         if (lut->lut_tdtd != NULL) {
2381                 if (!lut->lut_tdtd->tdtd_replay_ready &&
2382                     !obd->obd_abort_recovery && !obd->obd_stopping) {
2383                         /*
2384                          * Let's extend recovery timer, in case the recovery
2385                          * timer expired, and some clients got evicted
2386                          */
2387                         extend_recovery_timer(obd, obd->obd_recovery_timeout,
2388                                               true);
2389                         CDEBUG(D_HA,
2390                                "%s update recovery is not ready, extend recovery %d\n",
2391                                obd->obd_name, obd->obd_recovery_timeout);
2392                         return 0;
2393                 }
2394         }
2395
2396         return 1;
2397 }
2398
2399 enum {
2400         REQUEST_RECOVERY = 1,
2401         UPDATE_RECOVERY = 2,
2402 };
2403
2404 static __u64 get_next_replay_req_transno(struct obd_device *obd)
2405 {
2406         __u64 transno = 0;
2407
2408         if (!list_empty(&obd->obd_req_replay_queue)) {
2409                 struct ptlrpc_request *req;
2410
2411                 req = list_entry(obd->obd_req_replay_queue.next,
2412                                  struct ptlrpc_request, rq_list);
2413                 transno = lustre_msg_get_transno(req->rq_reqmsg);
2414         }
2415
2416         return transno;
2417 }
2418
2419 static __u64 get_next_transno(struct lu_target *lut, int *type)
2420 {
2421         struct obd_device *obd = lut->lut_obd;
2422         struct target_distribute_txn_data *tdtd = lut->lut_tdtd;
2423         __u64 transno = 0;
2424         __u64 update_transno;
2425
2426         ENTRY;
2427
2428         transno = get_next_replay_req_transno(obd);
2429         if (type != NULL)
2430                 *type = REQUEST_RECOVERY;
2431
2432         if (tdtd == NULL)
2433                 RETURN(transno);
2434
2435         update_transno = distribute_txn_get_next_transno(tdtd);
2436         if (transno == 0 || (transno >= update_transno &&
2437                              update_transno != 0)) {
2438                 transno = update_transno;
2439                 if (type != NULL)
2440                         *type = UPDATE_RECOVERY;
2441         }
2442
2443         RETURN(transno);
2444 }
2445
2446 /**
2447  * drop duplicate replay request
2448  *
2449  * Because the operation has been replayed by update recovery, the request
2450  * with the same transno will be dropped and also notify the client to send
2451  * next replay request.
2452  *
2453  * \param[in] env       execution environment
2454  * \param[in] obd       failover obd device
2455  * \param[in] req       request to be dropped
2456  */
2457 static void drop_duplicate_replay_req(struct lu_env *env,
2458                                       struct obd_device *obd,
2459                                       struct ptlrpc_request *req)
2460 {
2461         DEBUG_REQ(D_HA, req,
2462                   "remove t%lld from %s because duplicate update records found",
2463                   lustre_msg_get_transno(req->rq_reqmsg),
2464                   libcfs_nid2str(req->rq_peer.nid));
2465
2466         /*
2467          * Right now, only for MDS reint operation update replay and
2468          * normal request replay can have the same transno
2469          */
2470         if (lustre_msg_get_opc(req->rq_reqmsg) == MDS_REINT) {
2471                 req_capsule_set(&req->rq_pill, &RQF_MDS_REINT);
2472                 req->rq_status = req_capsule_server_pack(&req->rq_pill);
2473                 if (likely(req->rq_export))
2474                         target_committed_to_req(req);
2475                 lustre_msg_set_transno(req->rq_repmsg, req->rq_transno);
2476                 target_send_reply(req, req->rq_status, 0);
2477         } else {
2478                 DEBUG_REQ(D_ERROR, req, "wrong opc from %s",
2479                 libcfs_nid2str(req->rq_peer.nid));
2480         }
2481         target_exp_dequeue_req_replay(req);
2482         target_request_copy_put(req);
2483         obd->obd_replayed_requests++;
2484 }
2485
2486 #define WATCHDOG_TIMEOUT (obd_timeout * 10)
2487
2488 static void replay_request_or_update(struct lu_env *env,
2489                                      struct lu_target *lut,
2490                                      struct target_recovery_data *trd,
2491                                      struct ptlrpc_thread *thread)
2492 {
2493         struct obd_device *obd = lut->lut_obd;
2494         struct ptlrpc_request *req = NULL;
2495         int type;
2496         __u64 transno;
2497
2498         ENTRY;
2499
2500         CDEBUG(D_HA, "Waiting for transno %lld\n",
2501                obd->obd_next_recovery_transno);
2502
2503         /* Replay all of request and update by transno */
2504         do {
2505                 struct target_distribute_txn_data *tdtd = lut->lut_tdtd;
2506
2507                 CFS_FAIL_TIMEOUT(OBD_FAIL_TGT_REPLAY_DELAY2, cfs_fail_val);
2508
2509                 /**
2510                  * It is needed to extend recovery window above
2511                  *  recovery_time_soft. Extending is possible only in the
2512                  *  end of recovery window (see more details in
2513                  *  handle_recovery_req()).
2514                  */
2515                 CFS_FAIL_TIMEOUT_MS(OBD_FAIL_TGT_REPLAY_DELAY, 300);
2516
2517                 if (target_recovery_overseer(lut, check_for_next_transno,
2518                                         exp_req_replay_healthy_or_from_mdt)) {
2519                         abort_req_replay_queue(obd);
2520                         abort_lock_replay_queue(obd);
2521                         goto abort;
2522                 }
2523
2524                 spin_lock(&obd->obd_recovery_task_lock);
2525                 transno = get_next_transno(lut, &type);
2526                 if (type == REQUEST_RECOVERY && transno != 0) {
2527                         /*
2528                          * Drop replay request from client side, if the
2529                          * replay has been executed by update with the
2530                          * same transno
2531                          */
2532                         req = list_entry(obd->obd_req_replay_queue.next,
2533                                         struct ptlrpc_request, rq_list);
2534
2535                         list_del_init(&req->rq_list);
2536                         obd->obd_requests_queued_for_recovery--;
2537                         spin_unlock(&obd->obd_recovery_task_lock);
2538
2539                         /*
2540                          * Let's check if the request has been redone by
2541                          * update replay
2542                          */
2543                         if (is_req_replayed_by_update(req)) {
2544                                 struct distribute_txn_replay_req *dtrq;
2545
2546                                 dtrq = distribute_txn_lookup_finish_list(tdtd,
2547                                                                       transno);
2548                                 LASSERT(dtrq != NULL);
2549                                 spin_lock(&tdtd->tdtd_replay_list_lock);
2550                                 list_del_init(&dtrq->dtrq_list);
2551                                 spin_unlock(&tdtd->tdtd_replay_list_lock);
2552                                 dtrq_destroy(dtrq);
2553
2554                                 drop_duplicate_replay_req(env, obd, req);
2555
2556                                 continue;
2557                         }
2558
2559                         LASSERT(trd->trd_processing_task == current->pid);
2560                         DEBUG_REQ(D_HA, req, "processing x%llu t%lld from %s",
2561                                   req->rq_xid,
2562                                   lustre_msg_get_transno(req->rq_reqmsg),
2563                                   libcfs_nid2str(req->rq_peer.nid));
2564
2565                         ptlrpc_watchdog_init(&thread->t_watchdog,
2566                                              WATCHDOG_TIMEOUT);
2567                         handle_recovery_req(thread, req,
2568                                             trd->trd_recovery_handler);
2569                         ptlrpc_watchdog_disable(&thread->t_watchdog);
2570
2571                         /**
2572                          * bz18031: increase next_recovery_transno before
2573                          * target_request_copy_put() will drop exp_rpc reference
2574                          */
2575                         spin_lock(&obd->obd_recovery_task_lock);
2576                         obd->obd_next_recovery_transno++;
2577                         spin_unlock(&obd->obd_recovery_task_lock);
2578                         target_exp_dequeue_req_replay(req);
2579                         target_request_copy_put(req);
2580                         obd->obd_replayed_requests++;
2581                 } else if (type == UPDATE_RECOVERY && transno != 0) {
2582                         struct distribute_txn_replay_req *dtrq;
2583                         int rc;
2584
2585                         spin_unlock(&obd->obd_recovery_task_lock);
2586
2587                         LASSERT(tdtd != NULL);
2588                         dtrq = distribute_txn_get_next_req(tdtd);
2589                         lu_context_enter(&thread->t_env->le_ctx);
2590                         ptlrpc_watchdog_init(&thread->t_watchdog,
2591                                              WATCHDOG_TIMEOUT);
2592                         rc = tdtd->tdtd_replay_handler(env, tdtd, dtrq);
2593                         ptlrpc_watchdog_disable(&thread->t_watchdog);
2594                         lu_context_exit(&thread->t_env->le_ctx);
2595                         extend_recovery_timer(obd, obd_timeout, true);
2596
2597                         if (rc == 0 && dtrq->dtrq_xid != 0) {
2598                                 CDEBUG(D_HA,
2599                                        "Move x%llu t%llu to finish list\n",
2600                                        dtrq->dtrq_xid,
2601                                        dtrq->dtrq_master_transno);
2602
2603                                 /* Add it to the replay finish list */
2604                                 spin_lock(&tdtd->tdtd_replay_list_lock);
2605                                 list_add(&dtrq->dtrq_list,
2606                                          &tdtd->tdtd_replay_finish_list);
2607                                 spin_unlock(&tdtd->tdtd_replay_list_lock);
2608
2609                                 spin_lock(&obd->obd_recovery_task_lock);
2610                                 if (transno == obd->obd_next_recovery_transno)
2611                                         obd->obd_next_recovery_transno++;
2612                                 else if (transno >
2613                                          obd->obd_next_recovery_transno)
2614                                         obd->obd_next_recovery_transno =
2615                                                                 transno + 1;
2616                                 spin_unlock(&obd->obd_recovery_task_lock);
2617                         } else {
2618                                 dtrq_destroy(dtrq);
2619                         }
2620                 } else {
2621                         spin_unlock(&obd->obd_recovery_task_lock);
2622 abort:
2623                         LASSERT(list_empty(&obd->obd_req_replay_queue));
2624                         LASSERT(atomic_read(&obd->obd_req_replay_clients) == 0);
2625                         /** evict exports failed VBR */
2626                         class_disconnect_stale_exports(obd, exp_vbr_healthy);
2627                         break;
2628                 }
2629         } while (1);
2630 }
2631
2632 static int target_recovery_thread(void *arg)
2633 {
2634         struct lu_target *lut = arg;
2635         struct obd_device *obd = lut->lut_obd;
2636         struct ptlrpc_request *req;
2637         struct target_recovery_data *trd = &obd->obd_recovery_data;
2638         unsigned long delta;
2639         struct lu_env *env;
2640         struct ptlrpc_thread *thread = NULL;
2641         int rc = 0;
2642
2643         ENTRY;
2644
2645         unshare_fs_struct();
2646         OBD_ALLOC_PTR(thread);
2647         if (thread == NULL)
2648                 RETURN(-ENOMEM);
2649
2650         OBD_ALLOC_PTR(env);
2651         if (env == NULL)
2652                 GOTO(out_thread, rc = -ENOMEM);
2653         rc = lu_env_add(env);
2654         if (rc)
2655                 GOTO(out_env, rc);
2656
2657         rc = lu_context_init(&env->le_ctx, LCT_MD_THREAD | LCT_DT_THREAD);
2658         if (rc)
2659                 GOTO(out_env_remove, rc);
2660
2661         thread->t_env = env;
2662         thread->t_id = -1; /* force filter_iobuf_get/put to use local buffers */
2663         env->le_ctx.lc_thread = thread;
2664         tgt_io_thread_init(thread); /* init thread_big_cache for IO requests */
2665
2666         CDEBUG(D_HA, "%s: started recovery thread pid %d\n", obd->obd_name,
2667                current->pid);
2668         trd->trd_processing_task = current->pid;
2669
2670         spin_lock(&obd->obd_dev_lock);
2671         obd->obd_recovering = 1;
2672         spin_unlock(&obd->obd_dev_lock);
2673         complete(&trd->trd_starting);
2674
2675         /* first of all, we have to know the first transno to replay */
2676         if (target_recovery_overseer(lut, check_for_recovery_ready,
2677                                      exp_connect_healthy)) {
2678                 abort_req_replay_queue(obd);
2679                 abort_lock_replay_queue(obd);
2680                 if (lut->lut_tdtd != NULL)
2681                         dtrq_list_destroy(lut->lut_tdtd);
2682         }
2683
2684         /* next stage: replay requests or update */
2685         delta = jiffies;
2686         CDEBUG(D_INFO, "1: request replay stage - %d clients from t%llu\n",
2687                atomic_read(&obd->obd_req_replay_clients),
2688                obd->obd_next_recovery_transno);
2689         replay_request_or_update(env, lut, trd, thread);
2690
2691         /**
2692          * The second stage: replay locks
2693          */
2694         CDEBUG(D_INFO, "2: lock replay stage - %d clients\n",
2695                atomic_read(&obd->obd_lock_replay_clients));
2696         while ((req = target_next_replay_lock(lut))) {
2697                 LASSERT(trd->trd_processing_task == current->pid);
2698                 DEBUG_REQ(D_HA, req, "processing lock from %s:",
2699                           libcfs_nid2str(req->rq_peer.nid));
2700                 handle_recovery_req(thread, req,
2701                                     trd->trd_recovery_handler);
2702                 target_request_copy_put(req);
2703                 obd->obd_replayed_locks++;
2704         }
2705
2706         /**
2707          * The third stage: reply on final pings, at this moment all clients
2708          * must have request in final queue
2709          */
2710         CFS_FAIL_TIMEOUT(OBD_FAIL_TGT_REPLAY_RECONNECT, cfs_fail_val);
2711         CDEBUG(D_INFO, "3: final stage - process recovery completion pings\n");
2712         /** Update server last boot epoch */
2713         tgt_boot_epoch_update(lut);
2714         /*
2715          * We drop recoverying flag to forward all new requests
2716          * to regular mds_handle() since now
2717          */
2718         spin_lock(&obd->obd_dev_lock);
2719         obd->obd_recovering = obd->obd_abort_recovery = 0;
2720         spin_unlock(&obd->obd_dev_lock);
2721         spin_lock(&obd->obd_recovery_task_lock);
2722         target_cancel_recovery_timer(obd);
2723         spin_unlock(&obd->obd_recovery_task_lock);
2724         while ((req = target_next_final_ping(obd))) {
2725                 LASSERT(trd->trd_processing_task == current->pid);
2726                 DEBUG_REQ(D_HA, req, "processing final ping from %s:",
2727                           libcfs_nid2str(req->rq_peer.nid));
2728                 handle_recovery_req(thread, req,
2729                                     trd->trd_recovery_handler);
2730                 /*
2731                  * Because the waiting client can not send ping to server,
2732                  * so we need refresh the last_request_time, to avoid the
2733                  * export is being evicted
2734                  */
2735                 ptlrpc_update_export_timer(req->rq_export, 0);
2736                 target_request_copy_put(req);
2737         }
2738
2739         delta = jiffies_to_msecs(jiffies - delta) / MSEC_PER_SEC;
2740         CDEBUG(D_INFO, "4: recovery completed in %lus - %d/%d reqs/locks\n",
2741                delta, obd->obd_replayed_requests, obd->obd_replayed_locks);
2742         if (delta > OBD_RECOVERY_TIME_SOFT) {
2743                 CWARN("too long recovery - read logs\n");
2744                 libcfs_debug_dumplog();
2745         }
2746
2747         target_finish_recovery(lut);
2748         lu_context_fini(&env->le_ctx);
2749         trd->trd_processing_task = 0;
2750         complete_all(&trd->trd_finishing);
2751         tgt_io_thread_done(thread);
2752 out_env_remove:
2753         lu_env_remove(env);
2754 out_env:
2755         OBD_FREE_PTR(env);
2756 out_thread:
2757         OBD_FREE_PTR(thread);
2758         RETURN(rc);
2759 }
2760
2761 static int target_start_recovery_thread(struct lu_target *lut,
2762                                         svc_handler_t handler)
2763 {
2764         struct obd_device *obd = lut->lut_obd;
2765         int rc = 0;
2766         struct target_recovery_data *trd = &obd->obd_recovery_data;
2767         int index;
2768
2769         memset(trd, 0, sizeof(*trd));
2770         init_completion(&trd->trd_starting);
2771         init_completion(&trd->trd_finishing);
2772         trd->trd_recovery_handler = handler;
2773
2774         rc = server_name2index(obd->obd_name, &index, NULL);
2775         if (rc < 0)
2776                 return rc;
2777
2778         if (!IS_ERR(kthread_run(target_recovery_thread,
2779                                 lut, "tgt_recover_%d", index))) {
2780                 wait_for_completion(&trd->trd_starting);
2781                 LASSERT(obd->obd_recovering != 0);
2782         } else {
2783                 rc = -ECHILD;
2784         }
2785
2786         return rc;
2787 }
2788
2789 void target_stop_recovery_thread(struct obd_device *obd)
2790 {
2791         if (obd->obd_recovery_data.trd_processing_task > 0) {
2792                 struct target_recovery_data *trd = &obd->obd_recovery_data;
2793                 /** recovery can be done but postrecovery is not yet */
2794                 spin_lock(&obd->obd_dev_lock);
2795                 if (obd->obd_recovering) {
2796                         CERROR("%s: Aborting recovery\n", obd->obd_name);
2797                         obd->obd_abort_recovery = 1;
2798                         wake_up(&obd->obd_next_transno_waitq);
2799                 }
2800                 spin_unlock(&obd->obd_dev_lock);
2801                 wait_for_completion(&trd->trd_finishing);
2802         }
2803 }
2804 EXPORT_SYMBOL(target_stop_recovery_thread);
2805
2806 void target_recovery_fini(struct obd_device *obd)
2807 {
2808         class_disconnect_exports(obd);
2809         target_stop_recovery_thread(obd);
2810         target_cleanup_recovery(obd);
2811 }
2812 EXPORT_SYMBOL(target_recovery_fini);
2813
2814 static enum hrtimer_restart target_recovery_expired(struct hrtimer *timer)
2815 {
2816         struct obd_device *obd = container_of(timer, struct obd_device,
2817                                               obd_recovery_timer);
2818
2819         CDEBUG(D_HA,
2820                "%s: recovery timed out; %d clients are still in recovery after %llu seconds (%d clients connected)\n",
2821                obd->obd_name, atomic_read(&obd->obd_lock_replay_clients),
2822                ktime_get_seconds() - obd->obd_recovery_start,
2823                atomic_read(&obd->obd_connected_clients));
2824
2825         obd->obd_recovery_expired = 1;
2826         wake_up(&obd->obd_next_transno_waitq);
2827         return HRTIMER_NORESTART;
2828 }
2829
2830 void target_recovery_init(struct lu_target *lut, svc_handler_t handler)
2831 {
2832         struct obd_device *obd = lut->lut_obd;
2833
2834         if (lut->lut_bottom->dd_rdonly)
2835                 return;
2836
2837         if (atomic_read(&obd->obd_max_recoverable_clients) == 0) {
2838                 /** Update server last boot epoch */
2839                 tgt_boot_epoch_update(lut);
2840                 return;
2841         }
2842
2843         CDEBUG(D_HA, "RECOVERY: service %s, %d recoverable clients, "
2844                "last_transno %llu\n", obd->obd_name,
2845                atomic_read(&obd->obd_max_recoverable_clients),
2846                obd->obd_last_committed);
2847         LASSERT(obd->obd_stopping == 0);
2848         obd->obd_next_recovery_transno = obd->obd_last_committed + 1;
2849         obd->obd_recovery_start = 0;
2850         obd->obd_recovery_end = 0;
2851
2852         hrtimer_init(&obd->obd_recovery_timer, CLOCK_MONOTONIC,
2853                      HRTIMER_MODE_ABS);
2854         obd->obd_recovery_timer.function = &target_recovery_expired;
2855         target_start_recovery_thread(lut, handler);
2856 }
2857 EXPORT_SYMBOL(target_recovery_init);
2858
2859 static int target_process_req_flags(struct obd_device *obd,
2860                                     struct ptlrpc_request *req)
2861 {
2862         struct obd_export *exp = req->rq_export;
2863
2864         LASSERT(exp != NULL);
2865         if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REQ_REPLAY_DONE) {
2866                 /* client declares he's ready to replay locks */
2867                 spin_lock(&exp->exp_lock);
2868                 if (exp->exp_req_replay_needed) {
2869                         exp->exp_req_replay_needed = 0;
2870                         spin_unlock(&exp->exp_lock);
2871
2872                         LASSERT_ATOMIC_POS(&obd->obd_req_replay_clients);
2873                         atomic_dec(&obd->obd_req_replay_clients);
2874                 } else {
2875                         spin_unlock(&exp->exp_lock);
2876                 }
2877         }
2878         if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_LOCK_REPLAY_DONE) {
2879                 /*
2880                  * client declares he's ready to complete recovery
2881                  * so, we put the request on th final queue
2882                  */
2883                 spin_lock(&exp->exp_lock);
2884                 if (exp->exp_lock_replay_needed) {
2885                         exp->exp_lock_replay_needed = 0;
2886                         spin_unlock(&exp->exp_lock);
2887
2888                         LASSERT_ATOMIC_POS(&obd->obd_lock_replay_clients);
2889                         atomic_dec(&obd->obd_lock_replay_clients);
2890                 } else {
2891                         spin_unlock(&exp->exp_lock);
2892                 }
2893         }
2894         return 0;
2895 }
2896
2897 int target_queue_recovery_request(struct ptlrpc_request *req,
2898                                   struct obd_device *obd)
2899 {
2900         __u64 transno = lustre_msg_get_transno(req->rq_reqmsg);
2901         struct ptlrpc_request *reqiter;
2902         int inserted = 0;
2903
2904         ENTRY;
2905
2906         if (obd->obd_recovery_data.trd_processing_task == current->pid) {
2907                 /* Processing the queue right now, don't re-add. */
2908                 RETURN(1);
2909         }
2910
2911         target_process_req_flags(obd, req);
2912
2913         if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_LOCK_REPLAY_DONE) {
2914                 if (unlikely(OBD_FAIL_CHECK(OBD_FAIL_TGT_RECOVERY_REQ_RACE))) {
2915                         if (cfs_fail_val == 1) {
2916                                 cfs_race_state = 1;
2917                                 cfs_fail_val = 0;
2918                                 wake_up(&cfs_race_waitq);
2919
2920                                 schedule_timeout_interruptible(
2921                                         cfs_time_seconds(1));
2922                         }
2923                 }
2924
2925                 /*
2926                  * client declares he's ready to complete recovery
2927                  * so, we put the request on th final queue
2928                  */
2929                 target_request_copy_get(req);
2930                 DEBUG_REQ(D_HA, req, "queue final req");
2931                 wake_up(&obd->obd_next_transno_waitq);
2932                 spin_lock(&obd->obd_recovery_task_lock);
2933                 if (obd->obd_recovering) {
2934                         struct ptlrpc_request *tmp;
2935                         struct ptlrpc_request *duplicate = NULL;
2936
2937                         if (likely(!req->rq_export->exp_replay_done)) {
2938                                 req->rq_export->exp_replay_done = 1;
2939                                 list_add_tail(&req->rq_list,
2940                                               &obd->obd_final_req_queue);
2941                                 spin_unlock(&obd->obd_recovery_task_lock);
2942                                 RETURN(0);
2943                         }
2944
2945                         /*
2946                          * XXX O(n), but only happens if final ping is
2947                          * timed out, probably reorganize the list as
2948                          * a hash list later
2949                          */
2950                         list_for_each_entry_safe(reqiter, tmp,
2951                                                  &obd->obd_final_req_queue,
2952                                                  rq_list) {
2953                                 if (reqiter->rq_export == req->rq_export) {
2954                                         list_del_init(&reqiter->rq_list);
2955                                         duplicate = reqiter;
2956                                         break;
2957                                 }
2958                         }
2959
2960                         list_add_tail(&req->rq_list,
2961                                       &obd->obd_final_req_queue);
2962                         req->rq_export->exp_replay_done = 1;
2963                         spin_unlock(&obd->obd_recovery_task_lock);
2964
2965                         if (duplicate != NULL) {
2966                                 DEBUG_REQ(D_HA, duplicate,
2967                                           "put prev final req");
2968                                 target_request_copy_put(duplicate);
2969                         }
2970                         RETURN(0);
2971                 } else {
2972                         spin_unlock(&obd->obd_recovery_task_lock);
2973                         target_request_copy_put(req);
2974                         RETURN(obd->obd_stopping ? -ENOTCONN : 1);
2975                 }
2976         }
2977         if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REQ_REPLAY_DONE) {
2978                 /* client declares he's ready to replay locks */
2979                 target_request_copy_get(req);
2980                 DEBUG_REQ(D_HA, req, "queue lock replay req");
2981                 wake_up(&obd->obd_next_transno_waitq);
2982                 spin_lock(&obd->obd_recovery_task_lock);
2983                 LASSERT(obd->obd_recovering);
2984                 /* usually due to recovery abort */
2985                 if (!req->rq_export->exp_in_recovery) {
2986                         spin_unlock(&obd->obd_recovery_task_lock);
2987                         target_request_copy_put(req);
2988                         RETURN(-ENOTCONN);
2989                 }
2990                 LASSERT(req->rq_export->exp_lock_replay_needed);
2991                 list_add_tail(&req->rq_list, &obd->obd_lock_replay_queue);
2992                 spin_unlock(&obd->obd_recovery_task_lock);
2993                 RETURN(0);
2994         }
2995
2996         /*
2997          * CAVEAT EMPTOR: The incoming request message has been swabbed
2998          * (i.e. buflens etc are in my own byte order), but type-dependent
2999          * buffers (eg mdt_body, ost_body etc) have NOT been swabbed.
3000          */
3001
3002         if (!transno) {
3003                 INIT_LIST_HEAD(&req->rq_list);
3004                 DEBUG_REQ(D_HA, req, "not queueing");
3005                 RETURN(1);
3006         }
3007
3008         /*
3009          * If we're processing the queue, we want don't want to queue this
3010          * message.
3011          *
3012          * Also, if this request has a transno less than the one we're waiting
3013          * for, we should process it now.  It could (and currently always will)
3014          * be an open request for a descriptor that was opened some time ago.
3015          *
3016          * Also, a resent, replayed request that has already been
3017          * handled will pass through here and be processed immediately.
3018          */
3019         CDEBUG(D_HA,
3020                "Next recovery transno: %llu, current: %llu, replaying\n",
3021                obd->obd_next_recovery_transno, transno);
3022
3023         /*
3024          * If the request has been replayed by update replay, then sends this
3025          * request to the recovery thread (replay_request_or_update()), where
3026          * it will be handled
3027          */
3028         spin_lock(&obd->obd_recovery_task_lock);
3029         if (transno < obd->obd_next_recovery_transno &&
3030             !is_req_replayed_by_update(req)) {
3031                 /* Processing the queue right now, don't re-add. */
3032                 LASSERT(list_empty(&req->rq_list));
3033                 spin_unlock(&obd->obd_recovery_task_lock);
3034                 RETURN(1);
3035         }
3036         spin_unlock(&obd->obd_recovery_task_lock);
3037
3038         if (OBD_FAIL_CHECK(OBD_FAIL_TGT_REPLAY_DROP))
3039                 RETURN(0);
3040
3041         target_request_copy_get(req);
3042         if (!req->rq_export->exp_in_recovery) {
3043                 target_request_copy_put(req);
3044                 RETURN(-ENOTCONN);
3045         }
3046         LASSERT(req->rq_export->exp_req_replay_needed);
3047
3048         if (target_exp_enqueue_req_replay(req)) {
3049                 DEBUG_REQ(D_ERROR, req, "dropping resent queued req");
3050                 target_request_copy_put(req);
3051                 RETURN(0);
3052         }
3053
3054         /* XXX O(n^2) */
3055         spin_lock(&obd->obd_recovery_task_lock);
3056         LASSERT(obd->obd_recovering);
3057         list_for_each_entry(reqiter, &obd->obd_req_replay_queue, rq_list) {
3058                 if (lustre_msg_get_transno(reqiter->rq_reqmsg) > transno) {
3059                         list_add_tail(&req->rq_list, &reqiter->rq_list);
3060                         inserted = 1;
3061                         goto added;
3062                 }
3063
3064                 if (unlikely(lustre_msg_get_transno(reqiter->rq_reqmsg) ==
3065                              transno)) {
3066                         DEBUG_REQ(D_ERROR, req,
3067                                   "dropping replay: transno has been claimed by another client");
3068                         spin_unlock(&obd->obd_recovery_task_lock);
3069                         target_exp_dequeue_req_replay(req);
3070                         target_request_copy_put(req);
3071                         RETURN(0);
3072                 }
3073         }
3074 added:
3075         if (!inserted)
3076                 list_add_tail(&req->rq_list, &obd->obd_req_replay_queue);
3077
3078         obd->obd_requests_queued_for_recovery++;
3079         spin_unlock(&obd->obd_recovery_task_lock);
3080         wake_up(&obd->obd_next_transno_waitq);
3081         RETURN(0);
3082 }
3083
3084 void target_committed_to_req(struct ptlrpc_request *req)
3085 {
3086         struct obd_export *exp = req->rq_export;
3087
3088         if (!exp->exp_obd->obd_no_transno && req->rq_repmsg != NULL)
3089                 lustre_msg_set_last_committed(req->rq_repmsg,
3090                                               exp->exp_last_committed);
3091         else
3092                 DEBUG_REQ(D_IOCTL, req,
3093                           "not sending last_committed update (%d/%d)",
3094                           exp->exp_obd->obd_no_transno,
3095                           req->rq_repmsg == NULL);
3096
3097         CDEBUG(D_INFO, "last_committed %llu, transno %llu, xid %llu\n",
3098                exp->exp_last_committed, req->rq_transno, req->rq_xid);
3099 }
3100
3101 #endif /* HAVE_SERVER_SUPPORT */
3102
3103 /**
3104  * Packs current SLV and Limit into \a req.
3105  */
3106 int target_pack_pool_reply(struct ptlrpc_request *req)
3107 {
3108         struct obd_device *obd;
3109
3110         ENTRY;
3111
3112         /*
3113          * Check that we still have all structures alive as this may
3114          * be some late RPC at shutdown time.
3115          */
3116         if (unlikely(!req->rq_export || !req->rq_export->exp_obd ||
3117                      !exp_connect_lru_resize(req->rq_export))) {
3118                 lustre_msg_set_slv(req->rq_repmsg, 0);
3119                 lustre_msg_set_limit(req->rq_repmsg, 0);
3120                 RETURN(0);
3121         }
3122
3123         /* OBD is alive here as export is alive, which we checked above. */
3124         obd = req->rq_export->exp_obd;
3125
3126         read_lock(&obd->obd_pool_lock);
3127         lustre_msg_set_slv(req->rq_repmsg, obd->obd_pool_slv);
3128         lustre_msg_set_limit(req->rq_repmsg, obd->obd_pool_limit);
3129         read_unlock(&obd->obd_pool_lock);
3130
3131         RETURN(0);
3132 }
3133
3134 static int target_send_reply_msg(struct ptlrpc_request *req,
3135                                  int rc, int fail_id)
3136 {
3137         if (OBD_FAIL_CHECK_ORSET(fail_id & ~OBD_FAIL_ONCE, OBD_FAIL_ONCE)) {
3138                 DEBUG_REQ(D_ERROR, req, "dropping reply");
3139                 return -ECOMM;
3140         }
3141         /*
3142          * We can have a null rq_reqmsg in the event of bad signature or
3143          * no context when unwrapping
3144          */
3145         if (req->rq_reqmsg &&
3146             unlikely(lustre_msg_get_opc(req->rq_reqmsg) == MDS_REINT &&
3147             OBD_FAIL_CHECK(OBD_FAIL_MDS_REINT_MULTI_NET_REP)))
3148                 return -ECOMM;
3149
3150         if (unlikely(rc)) {
3151                 DEBUG_REQ(D_NET, req, "processing error (%d)", rc);
3152                 req->rq_status = rc;
3153                 return ptlrpc_send_error(req, 1);
3154         }
3155         DEBUG_REQ(D_NET, req, "sending reply");
3156
3157         return ptlrpc_send_reply(req, PTLRPC_REPLY_MAYBE_DIFFICULT);
3158 }
3159
3160 void target_send_reply(struct ptlrpc_request *req, int rc, int fail_id)
3161 {
3162         struct ptlrpc_service_part *svcpt;
3163         int netrc;
3164         struct ptlrpc_reply_state *rs;
3165         struct obd_export *exp;
3166
3167         ENTRY;
3168
3169         if (req->rq_no_reply) {
3170                 EXIT;
3171                 return;
3172         }
3173
3174         svcpt = req->rq_rqbd->rqbd_svcpt;
3175         rs = req->rq_reply_state;
3176         if (rs == NULL || !rs->rs_difficult) {
3177                 /* no notifiers */
3178                 target_send_reply_msg(req, rc, fail_id);
3179                 EXIT;
3180                 return;
3181         }
3182
3183         /* must be an export if locks saved */
3184         LASSERT(req->rq_export != NULL);
3185         /* req/reply consistent */
3186         LASSERT(rs->rs_svcpt == svcpt);
3187
3188         /* "fresh" reply */
3189         LASSERT(!rs->rs_scheduled);
3190         LASSERT(!rs->rs_scheduled_ever);
3191         LASSERT(!rs->rs_handled);
3192         LASSERT(!rs->rs_on_net);
3193         LASSERT(rs->rs_export == NULL);
3194         LASSERT(list_empty(&rs->rs_obd_list));
3195         LASSERT(list_empty(&rs->rs_exp_list));
3196
3197         exp = class_export_get(req->rq_export);
3198
3199         /* disable reply scheduling while I'm setting up */
3200         rs->rs_scheduled = 1;
3201         rs->rs_on_net    = 1;
3202         rs->rs_xid       = req->rq_xid;
3203         rs->rs_transno   = req->rq_transno;
3204         rs->rs_export    = exp;
3205         rs->rs_opc       = lustre_msg_get_opc(req->rq_reqmsg);
3206
3207         spin_lock(&exp->exp_uncommitted_replies_lock);
3208         CDEBUG(D_NET, "rs transno = %llu, last committed = %llu\n",
3209                rs->rs_transno, exp->exp_last_committed);
3210         if (rs->rs_transno > exp->exp_last_committed) {
3211                 /* not committed already */
3212                 list_add_tail(&rs->rs_obd_list,
3213                                   &exp->exp_uncommitted_replies);
3214         }
3215         spin_unlock(&exp->exp_uncommitted_replies_lock);
3216
3217         spin_lock(&exp->exp_lock);
3218         list_add_tail(&rs->rs_exp_list, &exp->exp_outstanding_replies);
3219         spin_unlock(&exp->exp_lock);
3220
3221         netrc = target_send_reply_msg(req, rc, fail_id);
3222
3223         spin_lock(&svcpt->scp_rep_lock);
3224
3225         atomic_inc(&svcpt->scp_nreps_difficult);
3226
3227         if (netrc != 0) {
3228                 /*
3229                  * error sending: reply is off the net.  Also we need +1
3230                  * reply ref until ptlrpc_handle_rs() is done
3231                  * with the reply state (if the send was successful, there
3232                  * would have been +1 ref for the net, which
3233                  * reply_out_callback leaves alone)
3234                  */
3235                 rs->rs_on_net = 0;
3236                 ptlrpc_rs_addref(rs);
3237         }
3238
3239         spin_lock(&rs->rs_lock);
3240         if (rs->rs_transno <= exp->exp_last_committed ||
3241             (!rs->rs_on_net && !rs->rs_no_ack) ||
3242             list_empty(&rs->rs_exp_list) ||     /* completed already */
3243             list_empty(&rs->rs_obd_list)) {
3244                 CDEBUG(D_HA, "Schedule reply immediately\n");
3245                 ptlrpc_dispatch_difficult_reply(rs);
3246         } else {
3247                 list_add(&rs->rs_list, &svcpt->scp_rep_active);
3248                 rs->rs_scheduled = 0;   /* allow notifier to schedule */
3249         }
3250         spin_unlock(&rs->rs_lock);
3251         spin_unlock(&svcpt->scp_rep_lock);
3252         EXIT;
3253 }
3254
3255 enum ldlm_mode lck_compat_array[] = {
3256         [LCK_EX]    = LCK_COMPAT_EX,
3257         [LCK_PW]    = LCK_COMPAT_PW,
3258         [LCK_PR]    = LCK_COMPAT_PR,
3259         [LCK_CW]    = LCK_COMPAT_CW,
3260         [LCK_CR]    = LCK_COMPAT_CR,
3261         [LCK_NL]    = LCK_COMPAT_NL,
3262         [LCK_GROUP] = LCK_COMPAT_GROUP,
3263         [LCK_COS]   = LCK_COMPAT_COS,
3264 };
3265
3266 /**
3267  * Rather arbitrary mapping from LDLM error codes to errno values. This should
3268  * not escape to the user level.
3269  */
3270 int ldlm_error2errno(enum ldlm_error error)
3271 {
3272         int result;
3273
3274         switch (error) {
3275         case ELDLM_OK:
3276         case ELDLM_LOCK_MATCHED:
3277                 result = 0;
3278                 break;
3279         case ELDLM_LOCK_CHANGED:
3280                 result = -ESTALE;
3281                 break;
3282         case ELDLM_LOCK_ABORTED:
3283                 result = -ENAVAIL;
3284                 break;
3285         case ELDLM_LOCK_REPLACED:
3286                 result = -ESRCH;
3287                 break;
3288         case ELDLM_NO_LOCK_DATA:
3289                 result = -ENOENT;
3290                 break;
3291         case ELDLM_NAMESPACE_EXISTS:
3292                 result = -EEXIST;
3293                 break;
3294         case ELDLM_BAD_NAMESPACE:
3295                 result = -EBADF;
3296                 break;
3297         default:
3298                 if (((int)error) < 0) { /* cast to signed type */
3299                         result = error; /* as ldlm_error can be unsigned */
3300                 } else {
3301                         CERROR("Invalid DLM result code: %d\n", error);
3302                         result = -EPROTO;
3303                 }
3304         }
3305         return result;
3306 }
3307 EXPORT_SYMBOL(ldlm_error2errno);
3308
3309 /**
3310  * Dual to ldlm_error2errno(): maps errno values back to enum ldlm_error.
3311  */
3312 enum ldlm_error ldlm_errno2error(int err_no)
3313 {
3314         int error;
3315
3316         switch (err_no) {
3317         case 0:
3318                 error = ELDLM_OK;
3319                 break;
3320         case -ESTALE:
3321                 error = ELDLM_LOCK_CHANGED;
3322                 break;
3323         case -ENAVAIL:
3324                 error = ELDLM_LOCK_ABORTED;
3325                 break;
3326         case -ESRCH:
3327                 error = ELDLM_LOCK_REPLACED;
3328                 break;
3329         case -ENOENT:
3330                 error = ELDLM_NO_LOCK_DATA;
3331                 break;
3332         case -EEXIST:
3333                 error = ELDLM_NAMESPACE_EXISTS;
3334                 break;
3335         case -EBADF:
3336                 error = ELDLM_BAD_NAMESPACE;
3337                 break;
3338         default:
3339                 error = err_no;
3340         }
3341         return error;
3342 }
3343
3344 #if LUSTRE_TRACKS_LOCK_EXP_REFS
3345 void ldlm_dump_export_locks(struct obd_export *exp)
3346 {
3347         spin_lock(&exp->exp_locks_list_guard);
3348         if (!list_empty(&exp->exp_locks_list)) {
3349                 struct ldlm_lock *lock;
3350
3351                 CERROR("dumping locks for export %p, ignore if the unmount doesn't hang\n",
3352                        exp);
3353                 list_for_each_entry(lock, &exp->exp_locks_list,
3354                                         l_exp_refs_link)
3355                         LDLM_ERROR(lock, "lock:");
3356         }
3357         spin_unlock(&exp->exp_locks_list_guard);
3358 }
3359 #endif
3360
3361 #ifdef HAVE_SERVER_SUPPORT
3362 static inline const char *bulk2type(struct ptlrpc_request *req)
3363 {
3364         if (req->rq_bulk_read)
3365                 return "READ";
3366         if (req->rq_bulk_write)
3367                 return "WRITE";
3368         return "UNKNOWN";
3369 }
3370
3371 int target_bulk_io(struct obd_export *exp, struct ptlrpc_bulk_desc *desc)
3372 {
3373         struct ptlrpc_request *req = desc->bd_req;
3374         time64_t start = ktime_get_seconds();
3375         time64_t deadline;
3376         int rc = 0;
3377
3378         ENTRY;
3379
3380         /* If there is eviction in progress, wait for it to finish. */
3381         wait_event_idle(
3382                 exp->exp_obd->obd_evict_inprogress_waitq,
3383                 !atomic_read(&exp->exp_obd->obd_evict_inprogress));
3384
3385         /* Check if client was evicted or reconnected already. */
3386         if (exp->exp_failed ||
3387             exp->exp_conn_cnt > lustre_msg_get_conn_cnt(req->rq_reqmsg)) {
3388                 rc = -ENOTCONN;
3389         } else {
3390                 if (req->rq_bulk_read)
3391                         rc = sptlrpc_svc_wrap_bulk(req, desc);
3392
3393                 if (OCD_HAS_FLAG(&exp->exp_connect_data, BULK_MBITS))
3394                         req->rq_mbits = lustre_msg_get_mbits(req->rq_reqmsg);
3395                 else /* old version, bulk matchbits is rq_xid */
3396                         req->rq_mbits = req->rq_xid;
3397
3398                 if (rc == 0)
3399                         rc = ptlrpc_start_bulk_transfer(desc);
3400         }
3401
3402         if (rc < 0) {
3403                 DEBUG_REQ(D_ERROR, req, "bulk %s failed: rc = %d",
3404                           bulk2type(req), rc);
3405                 RETURN(rc);
3406         }
3407
3408         if (OBD_FAIL_CHECK(OBD_FAIL_MDS_SENDPAGE)) {
3409                 ptlrpc_abort_bulk(desc);
3410                 RETURN(0);
3411         }
3412
3413         /* limit actual bulk transfer to bulk_timeout seconds */
3414         deadline = start + bulk_timeout;
3415         if (deadline > req->rq_deadline)
3416                 deadline = req->rq_deadline;
3417
3418         do {
3419                 time64_t timeoutl = deadline - ktime_get_seconds();
3420                 time64_t rq_deadline;
3421
3422                 while (timeoutl >= 0 &&
3423                        wait_event_idle_timeout(
3424                                desc->bd_waitq,
3425                                !ptlrpc_server_bulk_active(desc) ||
3426                                exp->exp_failed ||
3427                                exp->exp_conn_cnt >
3428                                lustre_msg_get_conn_cnt(req->rq_reqmsg),
3429                                timeoutl ? cfs_time_seconds(1) : 1) == 0)
3430                         timeoutl -= 1;
3431                 rc = timeoutl < 0 ? -ETIMEDOUT : 0;
3432
3433                 /* Wait again if we changed rq_deadline. */
3434                 rq_deadline = READ_ONCE(req->rq_deadline);
3435                 deadline = start + bulk_timeout;
3436                 if (deadline > rq_deadline)
3437                         deadline = rq_deadline;
3438         } while (rc == -ETIMEDOUT &&
3439                  deadline > ktime_get_seconds());
3440
3441         if (rc == -ETIMEDOUT) {
3442                 DEBUG_REQ(D_ERROR, req, "timeout on bulk %s after %lld%+llds",
3443                           bulk2type(req), deadline - start,
3444                           ktime_get_real_seconds() - deadline);
3445                 ptlrpc_abort_bulk(desc);
3446         } else if (exp->exp_failed) {
3447                 DEBUG_REQ(D_ERROR, req, "Eviction on bulk %s",
3448                           bulk2type(req));
3449                 rc = -ENOTCONN;
3450                 ptlrpc_abort_bulk(desc);
3451         } else if (exp->exp_conn_cnt >
3452                    lustre_msg_get_conn_cnt(req->rq_reqmsg)) {
3453                 DEBUG_REQ(D_ERROR, req, "Reconnect on bulk %s",
3454                           bulk2type(req));
3455                 /* We don't reply anyway. */
3456                 rc = -ETIMEDOUT;
3457                 ptlrpc_abort_bulk(desc);
3458         } else if (desc->bd_failure) {
3459                 DEBUG_REQ(D_ERROR, req, "network error on bulk %s",
3460                           bulk2type(req));
3461                 /* XXX should this be a different errno? */
3462                 rc = -ETIMEDOUT;
3463         } else {
3464                 if (req->rq_bulk_write)
3465                         rc = sptlrpc_svc_unwrap_bulk(req, desc);
3466                 if (rc == 0 && desc->bd_nob_transferred != desc->bd_nob) {
3467                         DEBUG_REQ(D_ERROR, req, "truncated bulk %s %d(%d)",
3468                                   bulk2type(req), desc->bd_nob_transferred,
3469                                   desc->bd_nob);
3470                         /* XXX should this be a different errno? */
3471                         rc = -ETIMEDOUT;
3472                 }
3473         }
3474
3475         RETURN(rc);
3476 }
3477 EXPORT_SYMBOL(target_bulk_io);
3478
3479 #endif /* HAVE_SERVER_SUPPORT */