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