lustre/ldlm/ldlm_lib.c

   1 /*
   2  * GPL HEADER START
   3  *
   4  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
   5  *
   6  * This program is free software; you can redistribute it and/or modify
   7  * it under the terms of the GNU General Public License version 2 only,
   8  * as published by the Free Software Foundation.
   9  *
  10  * This program is distributed in the hope that it will be useful, but
  11  * WITHOUT ANY WARRANTY; without even the implied warranty of
  12  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  13  * General Public License version 2 for more details (a copy is included
  14  * in the LICENSE file that accompanied this code).
  15  *
  16  * You should have received a copy of the GNU General Public License
  17  * version 2 along with this program; If not, see
  18  * http://www.gnu.org/licenses/gpl-2.0.html
  19  *
  20  * GPL HEADER END
  21  */
  22 /*
  23  * Copyright (c) 2003, 2010, Oracle and/or its affiliates. All rights reserved.
  24  * Use is subject to license terms.
  25  *
  26  * Copyright (c) 2010, 2017, Intel Corporation.
  27  */
  28 /*
  29  * This file is part of Lustre, http://www.lustre.org/
  30  * Lustre is a trademark of Sun Microsystems, Inc.
  31  */
  32
  33 /**
  34  * This file deals with various client/target related logic including recovery.
  35  *
  36  * TODO: This code more logically belongs in the ptlrpc module than in ldlm and
  37  * should be moved.
  38  */
  39
  40 #define DEBUG_SUBSYSTEM S_LDLM
  41
  42 #include <linux/jiffies.h>
  43 #include <linux/kernel.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_waitqueue_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         return rc;
 630 }
 631 EXPORT_SYMBOL(client_connect_import);
 632
 633 int client_disconnect_export(struct obd_export *exp)
 634 {
 635         struct obd_device *obd = class_exp2obd(exp);
 636         struct client_obd *cli;
 637         struct obd_import *imp;
 638         int rc = 0, err;
 639         ENTRY;
 640
 641         if (!obd) {
 642                 CERROR("invalid export for disconnect: exp %p cookie %#llx\n",
 643                        exp, exp ? exp->exp_handle.h_cookie : -1);
 644                 RETURN(-EINVAL);
 645         }
 646
 647         cli = &obd->u.cli;
 648         imp = cli->cl_import;
 649
 650         down_write(&cli->cl_sem);
 651         CDEBUG(D_INFO, "disconnect %s - %zu\n", obd->obd_name,
 652                 cli->cl_conn_count);
 653
 654         if (cli->cl_conn_count == 0) {
 655                 CERROR("disconnecting disconnected device (%s)\n",
 656                        obd->obd_name);
 657                 GOTO(out_disconnect, rc = -EINVAL);
 658         }
 659
 660         cli->cl_conn_count--;
 661         if (cli->cl_conn_count != 0)
 662                 GOTO(out_disconnect, rc = 0);
 663
 664         /* Mark import deactivated now, so we don't try to reconnect if any
 665          * of the cleanup RPCs fails (e.g. LDLM cancel, etc).  We don't
 666          * fully deactivate the import, or that would drop all requests. */
 667         spin_lock(&imp->imp_lock);
 668         imp->imp_deactive = 1;
 669         spin_unlock(&imp->imp_lock);
 670
 671         /* Some non-replayable imports (MDS's OSCs) are pinged, so just
 672          * delete it regardless.  (It's safe to delete an import that was
 673          * never added.) */
 674         (void)ptlrpc_pinger_del_import(imp);
 675
 676         if (obd->obd_namespace != NULL) {
 677                 /* obd_force == local only */
 678                 ldlm_cli_cancel_unused(obd->obd_namespace, NULL,
 679                                        obd->obd_force ? LCF_LOCAL : 0, NULL);
 680                 ldlm_namespace_free_prior(obd->obd_namespace, imp, obd->obd_force);
 681         }
 682
 683         /* There's no need to hold sem while disconnecting an import,
 684          * and it may actually cause deadlock in GSS. */
 685         up_write(&cli->cl_sem);
 686         rc = ptlrpc_disconnect_import(imp, 0);
 687         down_write(&cli->cl_sem);
 688
 689         ptlrpc_invalidate_import(imp);
 690
 691         EXIT;
 692
 693 out_disconnect:
 694         /* Use server style - class_disconnect should be always called for
 695          * o_disconnect. */
 696         err = class_disconnect(exp);
 697         if (!rc && err)
 698                 rc = err;
 699
 700         up_write(&cli->cl_sem);
 701
 702         RETURN(rc);
 703 }
 704 EXPORT_SYMBOL(client_disconnect_export);
 705
 706 #ifdef HAVE_SERVER_SUPPORT
 707 int server_disconnect_export(struct obd_export *exp)
 708 {
 709         int rc;
 710         ENTRY;
 711
 712         /* Disconnect early so that clients can't keep using export. */
 713         rc = class_disconnect(exp);
 714         /* Close import to avoid sending any requests. */
 715         if (exp->exp_imp_reverse)
 716                 ptlrpc_cleanup_imp(exp->exp_imp_reverse);
 717
 718         ldlm_bl_thread_wakeup();
 719
 720         /* complete all outstanding replies */
 721         spin_lock(&exp->exp_lock);
 722         while (!list_empty(&exp->exp_outstanding_replies)) {
 723                 struct ptlrpc_reply_state *rs =
 724                         list_entry(exp->exp_outstanding_replies.next,
 725                                        struct ptlrpc_reply_state, rs_exp_list);
 726                 struct ptlrpc_service_part *svcpt = rs->rs_svcpt;
 727
 728                 spin_lock(&svcpt->scp_rep_lock);
 729
 730                 list_del_init(&rs->rs_exp_list);
 731
 732                 spin_lock(&rs->rs_lock);
 733                 /* clear rs_convert_lock to make sure rs is handled and put */
 734                 rs->rs_convert_lock = 0;
 735                 ptlrpc_schedule_difficult_reply(rs);
 736                 spin_unlock(&rs->rs_lock);
 737
 738                 spin_unlock(&svcpt->scp_rep_lock);
 739         }
 740         spin_unlock(&exp->exp_lock);
 741
 742         RETURN(rc);
 743 }
 744 EXPORT_SYMBOL(server_disconnect_export);
 745
 746 static inline int target_check_recovery_timer(struct obd_device *target)
 747 {
 748         ktime_t remaining;
 749         s64 timeout;
 750
 751         if (!target->obd_recovering || target->obd_recovery_start == 0)
 752                 return 0;
 753
 754         remaining = hrtimer_expires_remaining(&target->obd_recovery_timer);
 755         timeout = ktime_divns(remaining, NSEC_PER_SEC);
 756         if (timeout > -30)
 757                 return 0;
 758
 759         /* the recovery timer should expire, but it isn't triggered,
 760          * it's better to abort the recovery of this target to speed up
 761          * the recovery of the whole cluster. */
 762         spin_lock(&target->obd_dev_lock);
 763         if (target->obd_recovering) {
 764                 CERROR("%s: Aborting recovery\n", target->obd_name);
 765                 target->obd_abort_recovery = 1;
 766                 wake_up(&target->obd_next_transno_waitq);
 767         }
 768         spin_unlock(&target->obd_dev_lock);
 769         return 0;
 770 }
 771
 772 /* --------------------------------------------------------------------------
 773  * from old lib/target.c
 774  * -------------------------------------------------------------------------- */
 775
 776 static int target_handle_reconnect(struct lustre_handle *conn,
 777                                    struct obd_export *exp,
 778                                    struct obd_uuid *cluuid)
 779 {
 780         struct obd_device *target;
 781         struct lustre_handle *hdl;
 782         ktime_t remaining;
 783         s64 timeout;
 784         int rc = 0;
 785
 786         ENTRY;
 787         hdl = &exp->exp_imp_reverse->imp_remote_handle;
 788         if (!exp->exp_connection || !lustre_handle_is_used(hdl)) {
 789                 conn->cookie = exp->exp_handle.h_cookie;
 790                 CDEBUG(D_HA, "connect export for UUID '%s' at %p,"
 791                        " cookie %#llx\n", cluuid->uuid, exp, conn->cookie);
 792                 RETURN(0);
 793         }
 794
 795         target = exp->exp_obd;
 796
 797         /* Might be a re-connect after a partition. */
 798         if (memcmp(&conn->cookie, &hdl->cookie, sizeof conn->cookie)) {
 799                 LCONSOLE_WARN("%s: already connected client %s (at %s) "
 800                               "with handle %#llx. Rejecting client "
 801                               "with the same UUID trying to reconnect "
 802                               "with handle %#llx\n", target->obd_name,
 803                               obd_uuid2str(&exp->exp_client_uuid),
 804                               obd_export_nid2str(exp),
 805                               hdl->cookie, conn->cookie);
 806                 memset(conn, 0, sizeof *conn);
 807                 /* target_handle_connect() treats EALREADY and
 808                  * -EALREADY differently.  -EALREADY is an error
 809                  * (same UUID, different handle). */
 810                 RETURN(-EALREADY);
 811         }
 812
 813         if (!target->obd_recovering) {
 814                 LCONSOLE_WARN("%s: Client %s (at %s) reconnecting\n",
 815                         target->obd_name, obd_uuid2str(&exp->exp_client_uuid),
 816                         obd_export_nid2str(exp));
 817                 GOTO(out_already, rc);
 818         }
 819
 820         remaining = hrtimer_expires_remaining(&target->obd_recovery_timer);
 821         timeout = ktime_divns(remaining, NSEC_PER_SEC);
 822         if (timeout > 0) {
 823                 LCONSOLE_WARN("%s: Client %s (at %s) reconnected, waiting for %d clients in recovery for %lld:%.02lld\n",
 824                               target->obd_name,
 825                               obd_uuid2str(&exp->exp_client_uuid),
 826                               obd_export_nid2str(exp),
 827                               atomic_read(&target->obd_max_recoverable_clients),
 828                               timeout / 60, timeout % 60);
 829         } else {
 830                 struct target_distribute_txn_data *tdtd;
 831                 int size = 0;
 832                 int count = 0;
 833                 char *buf = NULL;
 834
 835                 target_check_recovery_timer(target);
 836
 837                 tdtd = class_exp2tgt(exp)->lut_tdtd;
 838                 if (tdtd && tdtd->tdtd_show_update_logs_retrievers)
 839                         buf = tdtd->tdtd_show_update_logs_retrievers(
 840                                 tdtd->tdtd_show_retrievers_cbdata,
 841                                 &size, &count);
 842
 843                 if (count > 0)
 844                         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",
 845                                       target->obd_name,
 846                                       obd_uuid2str(&exp->exp_client_uuid),
 847                                       obd_export_nid2str(exp), count,
 848                                       buf ? buf : "unknown (not enough RAM)",
 849                                       (abs(timeout) + target->obd_recovery_timeout) / 60,
 850                                       (abs(timeout) + target->obd_recovery_timeout) % 60,
 851                                       target->obd_name);
 852                 else
 853                         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",
 854                                       target->obd_name, abs(timeout) / 60,
 855                                       abs(timeout) % 60, target->obd_name);
 856
 857                 if (buf != NULL)
 858                         OBD_FREE(buf, size);
 859         }
 860
 861 out_already:
 862         conn->cookie = exp->exp_handle.h_cookie;
 863         /* target_handle_connect() treats EALREADY and
 864          * -EALREADY differently.  EALREADY means we are
 865          * doing a valid reconnect from the same client. */
 866         RETURN(EALREADY);
 867 }
 868
 869 static void
 870 check_and_start_recovery_timer(struct obd_device *obd,
 871                                struct ptlrpc_request *req, int new_client);
 872
 873 /**
 874  * update flags for import during reconnect process
 875  */
 876 static int rev_import_flags_update(struct obd_import *revimp,
 877                                    struct ptlrpc_request *req)
 878 {
 879         int rc;
 880         struct obd_connect_data *data;
 881
 882         data = req_capsule_client_get(&req->rq_pill, &RMF_CONNECT_DATA);
 883
 884         if (data->ocd_connect_flags & OBD_CONNECT_AT)
 885                 revimp->imp_msghdr_flags |= MSGHDR_AT_SUPPORT;
 886         else
 887                 revimp->imp_msghdr_flags &= ~MSGHDR_AT_SUPPORT;
 888
 889         revimp->imp_msghdr_flags |= MSGHDR_CKSUM_INCOMPAT18;
 890
 891         rc = sptlrpc_import_sec_adapt(revimp, req->rq_svc_ctx, &req->rq_flvr);
 892         if (rc) {
 893                 CERROR("%s: cannot get reverse import %s security: rc = %d\n",
 894                         revimp->imp_client->cli_name,
 895                         libcfs_id2str(req->rq_peer), rc);
 896                 return rc;
 897         }
 898
 899         return 0;
 900 }
 901
 902 /**
 903  * Allocate a new reverse import for an export.
 904  *
 905  * \retval -errno in case error hit
 906  * \retval 0 if reverse import correctly init
 907  **/
 908 int rev_import_init(struct obd_export *export)
 909 {
 910         struct obd_device *obd = export->exp_obd;
 911         struct obd_import *revimp;
 912
 913         LASSERT(export->exp_imp_reverse == NULL);
 914
 915         revimp = class_new_import(obd);
 916         if (revimp == NULL)
 917                 return -ENOMEM;
 918
 919         revimp->imp_remote_handle.cookie = 0ULL;
 920         revimp->imp_client = &obd->obd_ldlm_client;
 921         revimp->imp_dlm_fake = 1;
 922
 923         /* it is safe to connect import in new state as no sends possible */
 924         spin_lock(&export->exp_lock);
 925         export->exp_imp_reverse = revimp;
 926         spin_unlock(&export->exp_lock);
 927         class_import_put(revimp);
 928
 929         return 0;
 930 }
 931 EXPORT_SYMBOL(rev_import_init);
 932
 933 /**
 934  * Handle reconnect for an export.
 935  *
 936  * \param exp export to handle reconnect process
 937  * \param req client reconnect request
 938  *
 939  * \retval -rc in case securitfy flavor can't be changed
 940  * \retval 0 in case none problems
 941  */
 942 static int rev_import_reconnect(struct obd_export *exp,
 943                                 struct ptlrpc_request *req)
 944 {
 945         struct obd_import *revimp = exp->exp_imp_reverse;
 946         struct lustre_handle *lh;
 947         int rc;
 948
 949         /* avoid sending a request until import flags are changed */
 950         ptlrpc_import_enter_resend(revimp);
 951
 952         if (revimp->imp_connection != NULL)
 953                 ptlrpc_connection_put(revimp->imp_connection);
 954
 955         /*
 956          * client from recovery don't have a handle so we need to take from
 957          * request. it may produce situation when wrong client connected
 958          * to recovery as we trust a client uuid
 959          */
 960         lh = req_capsule_client_get(&req->rq_pill, &RMF_CONN);
 961         revimp->imp_remote_handle = *lh;
 962
 963         /* unknown versions will be caught in
 964          * ptlrpc_handle_server_req_in->lustre_unpack_msg() */
 965         revimp->imp_msg_magic = req->rq_reqmsg->lm_magic;
 966
 967         revimp->imp_connection = ptlrpc_connection_addref(exp->exp_connection);
 968
 969         rc = rev_import_flags_update(revimp, req);
 970         if (rc != 0) {
 971                 /* it is safe to still be in RECOVERY phase as we are not able
 972                  * to setup correct security flavor so requests are not able to
 973                  * be delivered correctly */
 974                 return rc;
 975         }
 976
 977         /* resend all rpc's via new connection */
 978         return ptlrpc_import_recovery_state_machine(revimp);
 979 }
 980
 981 int target_handle_connect(struct ptlrpc_request *req)
 982 {
 983         struct obd_device *target = NULL;
 984         struct obd_export *export = NULL;
 985         /* connect handle - filled from target_handle_reconnect in
 986          * reconnect case */
 987         struct lustre_handle conn;
 988         struct lustre_handle *tmp;
 989         struct obd_uuid cluuid;
 990         char *str;
 991         int rc = 0;
 992         char *target_start;
 993         int target_len;
 994         bool     mds_conn = false, lw_client = false, initial_conn = false;
 995         bool     mds_mds_conn = false;
 996         bool     new_mds_mds_conn = false;
 997         struct obd_connect_data *data, *tmpdata;
 998         int size, tmpsize;
 999         lnet_nid_t *client_nid = NULL;
1000         ENTRY;
1001
1002         OBD_RACE(OBD_FAIL_TGT_CONN_RACE);
1003
1004         str = req_capsule_client_get(&req->rq_pill, &RMF_TGTUUID);
1005         if (str == NULL) {
1006                 DEBUG_REQ(D_ERROR, req, "bad target UUID for connect");
1007                 GOTO(out, rc = -EINVAL);
1008         }
1009
1010         target = class_dev_by_str(str);
1011         if (!target) {
1012                 deuuidify(str, NULL, &target_start, &target_len);
1013                 LCONSOLE_ERROR_MSG(0x137, "%s: not available for connect "
1014                                    "from %s (no target). If you are running "
1015                                    "an HA pair check that the target is "
1016                                    "mounted on the other server.\n", str,
1017                                    libcfs_nid2str(req->rq_peer.nid));
1018                 GOTO(out, rc = -ENODEV);
1019         }
1020
1021         spin_lock(&target->obd_dev_lock);
1022
1023         target->obd_conn_inprogress++;
1024
1025         if (target->obd_stopping || !target->obd_set_up) {
1026                 spin_unlock(&target->obd_dev_lock);
1027
1028                 deuuidify(str, NULL, &target_start, &target_len);
1029                 LCONSOLE_INFO("%.*s: Not available for connect from %s (%s)\n",
1030                               target_len, target_start,
1031                               libcfs_nid2str(req->rq_peer.nid),
1032                               (target->obd_stopping ?
1033                                "stopping" : "not set up"));
1034                 GOTO(out, rc = -ENODEV);
1035         }
1036
1037         if (target->obd_no_conn) {
1038                 spin_unlock(&target->obd_dev_lock);
1039
1040                 CDEBUG(D_INFO, "%s: Temporarily refusing client connection "
1041                                "from %s\n", target->obd_name,
1042                                libcfs_nid2str(req->rq_peer.nid));
1043                 GOTO(out, rc = -EAGAIN);
1044         }
1045
1046         spin_unlock(&target->obd_dev_lock);
1047
1048         str = req_capsule_client_get(&req->rq_pill, &RMF_CLUUID);
1049         if (str == NULL) {
1050                 DEBUG_REQ(D_ERROR, req, "bad client UUID for connect");
1051                 GOTO(out, rc = -EINVAL);
1052         }
1053
1054         obd_str2uuid(&cluuid, str);
1055
1056         tmp = req_capsule_client_get(&req->rq_pill, &RMF_CONN);
1057         if (tmp == NULL)
1058                 GOTO(out, rc = -EPROTO);
1059
1060         conn = *tmp;
1061
1062         size = req_capsule_get_size(&req->rq_pill, &RMF_CONNECT_DATA,
1063                                     RCL_CLIENT);
1064         if (size < 0 || size > 8 * sizeof(struct obd_connect_data))
1065                 GOTO(out, rc = -EPROTO);
1066         data = req_capsule_client_get(&req->rq_pill, &RMF_CONNECT_DATA);
1067         if (!data)
1068                 GOTO(out, rc = -EPROTO);
1069
1070         rc = req_capsule_server_pack(&req->rq_pill);
1071         if (rc)
1072                 GOTO(out, rc);
1073
1074 #if LUSTRE_VERSION_CODE < OBD_OCD_VERSION(3, 0, 53, 0)
1075         /* Don't allow clients to connect that are using old 1.8 format
1076          * protocol conventions (LUSTRE_MSG_MAGIC_v1, !MSGHDR_CKSUM_INCOMPAT18,
1077          * ldlm_flock_policy_wire format, MDT_ATTR_xTIME_SET, etc).  The
1078          * FULL20 flag should be set on all connections since 2.0, but no
1079          * longer affects behaviour.
1080          *
1081          * Later this check will be disabled and the flag can be retired
1082          * completely once interop with 3.0 is no longer needed.
1083          */
1084         if (!(data->ocd_connect_flags & OBD_CONNECT_FULL20))
1085                 GOTO(out, rc = -EPROTO);
1086
1087         /* Don't allow liblustre clients to connect.
1088          * - testing was disabled in v2_2_50_0-61-g6a75d65
1089          * - building was disabled in v2_5_58_0-28-g7277179
1090          * - client code was deleted in v2_6_50_0-101-gcdfbc72,
1091          * - clients were refused connect for version difference > 0.0.1.32  */
1092         if (lustre_msg_get_op_flags(req->rq_reqmsg) & MSG_CONNECT_LIBCLIENT) {
1093                 DEBUG_REQ(D_WARNING, req, "Refusing libclient connection");
1094                 GOTO(out, rc = -EPROTO);
1095         }
1096 #endif
1097
1098         /* Note: lw_client is needed in MDS-MDS failover during update log
1099          * processing, so we needs to allow lw_client to be connected at
1100          * anytime, instead of only the initial connection */
1101         lw_client = (data->ocd_connect_flags & OBD_CONNECT_LIGHTWEIGHT) != 0;
1102
1103         if (lustre_msg_get_op_flags(req->rq_reqmsg) & MSG_CONNECT_INITIAL) {
1104                 initial_conn = true;
1105                 mds_conn = (data->ocd_connect_flags & OBD_CONNECT_MDS) != 0;
1106                 mds_mds_conn = (data->ocd_connect_flags &
1107                                 OBD_CONNECT_MDS_MDS) != 0;
1108
1109                 /* OBD_CONNECT_MNE_SWAB is defined as OBD_CONNECT_MDS_MDS
1110                  * for Imperative Recovery connection from MGC to MGS.
1111                  *
1112                  * Via check OBD_CONNECT_FID, we can distinguish whether
1113                  * the OBD_CONNECT_MDS_MDS/OBD_CONNECT_MNE_SWAB is from
1114                  * MGC or MDT, since MGC does not use OBD_CONNECT_FID.
1115                  */
1116                 if (!lw_client &&
1117                     (data->ocd_connect_flags & OBD_CONNECT_MDS_MDS) &&
1118                     (data->ocd_connect_flags & OBD_CONNECT_FID) &&
1119                     (data->ocd_connect_flags & OBD_CONNECT_VERSION)) {
1120                         __u32 major = OBD_OCD_VERSION_MAJOR(data->ocd_version);
1121                         __u32 minor = OBD_OCD_VERSION_MINOR(data->ocd_version);
1122                         __u32 patch = OBD_OCD_VERSION_PATCH(data->ocd_version);
1123
1124                         /* We do not support the MDT-MDT interoperations with
1125                          * different version MDT because of protocol changes. */
1126                         if (unlikely(major != LUSTRE_MAJOR ||
1127                                      minor != LUSTRE_MINOR ||
1128                                      abs(patch - LUSTRE_PATCH) > 3)) {
1129                                 LCONSOLE_WARN("%s (%u.%u.%u.%u) refused the "
1130                                         "connection from different version MDT "
1131                                         "(%d.%d.%d.%d) %s %s\n",
1132                                         target->obd_name, LUSTRE_MAJOR,
1133                                         LUSTRE_MINOR, LUSTRE_PATCH, LUSTRE_FIX,
1134                                         major, minor, patch,
1135                                         OBD_OCD_VERSION_FIX(data->ocd_version),
1136                                         libcfs_nid2str(req->rq_peer.nid), str);
1137
1138                                 GOTO(out, rc = -EPROTO);
1139                         }
1140                 }
1141         }
1142
1143         /* lctl gets a backstage, all-access pass. */
1144         if (obd_uuid_equals(&cluuid, &target->obd_uuid))
1145                 goto dont_check_exports;
1146
1147         export = cfs_hash_lookup(target->obd_uuid_hash, &cluuid);
1148         if (!export)
1149                 goto no_export;
1150
1151         /* We've found an export in the hash. */
1152
1153         spin_lock(&export->exp_lock);
1154
1155         if (export->exp_connecting) { /* bug 9635, et. al. */
1156                 spin_unlock(&export->exp_lock);
1157                 LCONSOLE_WARN("%s: Export %p already connecting from %s\n",
1158                               export->exp_obd->obd_name, export,
1159                               libcfs_nid2str(req->rq_peer.nid));
1160                 class_export_put(export);
1161                 export = NULL;
1162                 rc = -EALREADY;
1163         } else if ((mds_conn || (lw_client && initial_conn) ||
1164                    data->ocd_connect_flags & OBD_CONNECT_MDS_MDS) &&
1165                    export->exp_connection != NULL) {
1166                 spin_unlock(&export->exp_lock);
1167                 if (req->rq_peer.nid != export->exp_connection->c_peer.nid) {
1168                         /* MDS or LWP reconnected after failover. */
1169                         LCONSOLE_WARN("%s: Received %s connection from "
1170                             "%s, removing former export from %s\n",
1171                             target->obd_name, mds_conn ? "MDS" : "LWP",
1172                             libcfs_nid2str(req->rq_peer.nid),
1173                             libcfs_nid2str(export->exp_connection->c_peer.nid));
1174                 } else {
1175                         /* New MDS connection from the same NID. */
1176                         LCONSOLE_WARN("%s: Received new %s connection from "
1177                                 "%s, removing former export from same NID\n",
1178                                 target->obd_name, mds_conn ? "MDS" : "LWP",
1179                                 libcfs_nid2str(req->rq_peer.nid));
1180                 }
1181
1182                 if (req->rq_peer.nid == export->exp_connection->c_peer.nid &&
1183                     data->ocd_connect_flags & OBD_CONNECT_MDS_MDS) {
1184                         /* Because exports between MDTs will always be
1185                          * kept, let's do not fail such export if they
1186                          * come from the same NID, otherwise it might
1187                          * cause eviction between MDTs, which might
1188                          * cause namespace inconsistency */
1189                         spin_lock(&export->exp_lock);
1190                         export->exp_connecting = 1;
1191                         export->exp_conn_cnt = 0;
1192                         spin_unlock(&export->exp_lock);
1193                         conn.cookie = export->exp_handle.h_cookie;
1194                         rc = EALREADY;
1195                 } else {
1196                         class_fail_export(export);
1197                         class_export_put(export);
1198                         export = NULL;
1199                         rc = 0;
1200                 }
1201         } else if (export->exp_connection != NULL && initial_conn &&
1202                    req->rq_peer.nid != export->exp_connection->c_peer.nid) {
1203                 spin_unlock(&export->exp_lock);
1204                 /* In MDS failover we have static UUID but NID can change. */
1205                 LCONSOLE_WARN("%s: Client %s seen on new nid %s when "
1206                               "existing nid %s is already connected\n",
1207                               target->obd_name, cluuid.uuid,
1208                               libcfs_nid2str(req->rq_peer.nid),
1209                               libcfs_nid2str(
1210                                       export->exp_connection->c_peer.nid));
1211                 rc = -EALREADY;
1212                 class_export_put(export);
1213                 export = NULL;
1214         } else {
1215                 export->exp_connecting = 1;
1216                 spin_unlock(&export->exp_lock);
1217                 LASSERT(export->exp_obd == target);
1218
1219                 rc = target_handle_reconnect(&conn, export, &cluuid);
1220         }
1221
1222         /* If we found an export, we already unlocked. */
1223         if (!export) {
1224 no_export:
1225                 OBD_FAIL_TIMEOUT(OBD_FAIL_TGT_DELAY_CONNECT, 2 * obd_timeout);
1226         } else if (req->rq_export == NULL &&
1227                    atomic_read(&export->exp_rpc_count) > 0) {
1228                 LCONSOLE_WARN("%s: Client %s (at %s) refused connection, "
1229                               "still busy with %d references\n",
1230                               target->obd_name, cluuid.uuid,
1231                               libcfs_nid2str(req->rq_peer.nid),
1232                               atomic_read(&export->exp_refcount));
1233                         GOTO(out, rc = -EBUSY);
1234         } else if (lustre_msg_get_conn_cnt(req->rq_reqmsg) == 1 &&
1235                    rc != EALREADY) {
1236                 if (!strstr(cluuid.uuid, "mdt"))
1237                         LCONSOLE_WARN("%s: Rejecting reconnect from the "
1238                                       "known client %s (at %s) because it "
1239                                       "is indicating it is a new client",
1240                                       target->obd_name, cluuid.uuid,
1241                                       libcfs_nid2str(req->rq_peer.nid));
1242                 GOTO(out, rc = -EALREADY);
1243         } else {
1244                 OBD_FAIL_TIMEOUT(OBD_FAIL_TGT_DELAY_RECONNECT, 2 * obd_timeout);
1245         }
1246
1247         if (rc < 0) {
1248                 GOTO(out, rc);
1249         }
1250
1251         CDEBUG(D_HA, "%s: connection from %s@%s %st%llu exp %p cur %lld last %lld\n",
1252                target->obd_name, cluuid.uuid, libcfs_nid2str(req->rq_peer.nid),
1253                target->obd_recovering ? "recovering/" : "", data->ocd_transno,
1254                export, ktime_get_seconds(),
1255                export ? export->exp_last_request_time : 0);
1256
1257         /* If this is the first time a client connects, reset the recovery
1258          * timer. Discard lightweight connections which might be local. */
1259         if (!lw_client && rc == 0 && target->obd_recovering)
1260                 check_and_start_recovery_timer(target, req, export == NULL);
1261
1262         /* We want to handle EALREADY but *not* -EALREADY from
1263          * target_handle_reconnect(), return reconnection state in a flag. */
1264         if (rc == EALREADY) {
1265                 lustre_msg_add_op_flags(req->rq_repmsg, MSG_CONNECT_RECONNECT);
1266                 rc = 0;
1267         } else {
1268                 LASSERT(rc == 0);
1269         }
1270
1271         /* Tell the client if we support replayable requests. */
1272         if (target->obd_replayable)
1273                 lustre_msg_add_op_flags(req->rq_repmsg, MSG_CONNECT_REPLAYABLE);
1274         client_nid = &req->rq_peer.nid;
1275
1276         if (export == NULL) {
1277                 /* allow lightweight connections during recovery */
1278                 /* allow "new" MDT to be connected during recovery, since we
1279                  * need retrieve recovery update records from it */
1280                 if (target->obd_recovering && !lw_client && !mds_mds_conn) {
1281                         struct hrtimer *timer = &target->obd_recovery_timer;
1282                         ktime_t remaining;
1283                         s64 timeout, left;
1284                         int in_progress;
1285                         int connected;
1286                         int known;
1287                         int stale;
1288                         char *msg;
1289
1290                         connected = atomic_read(&target->obd_connected_clients);
1291                         in_progress = atomic_read(&target->obd_lock_replay_clients);
1292                         known =
1293                            atomic_read(&target->obd_max_recoverable_clients);
1294                         stale = target->obd_stale_clients;
1295                         remaining = hrtimer_expires_remaining(timer);
1296                         left = ktime_divns(remaining, NSEC_PER_SEC);
1297                         if (ktime_to_ns(remaining) > 0) {
1298                                 msg = "to recover in";
1299                                 timeout = left;
1300                         } else {
1301                                 msg = "already passed deadline";
1302                                 timeout = -left;
1303
1304                                 target_check_recovery_timer(target);
1305                         }
1306
1307                         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",
1308                                       target->obd_name, cluuid.uuid,
1309                                       libcfs_nid2str(req->rq_peer.nid), known,
1310                                       connected - in_progress, in_progress,
1311                                       stale, msg, timeout / 60, timeout % 60);
1312                         rc = -EBUSY;
1313                 } else {
1314 dont_check_exports:
1315                         rc = obd_connect(req->rq_svc_thread->t_env,
1316                                          &export, target, &cluuid, data,
1317                                          client_nid);
1318                         if (mds_conn && OBD_FAIL_CHECK(OBD_FAIL_TGT_RCVG_FLAG))
1319                                 lustre_msg_add_op_flags(req->rq_repmsg,
1320                                                         MSG_CONNECT_RECOVERING);
1321                         if (rc == 0) {
1322                                 conn.cookie = export->exp_handle.h_cookie;
1323                                 rc = rev_import_init(export);
1324                         }
1325
1326                         if (mds_mds_conn)
1327                                 new_mds_mds_conn = true;
1328                 }
1329         } else {
1330                 rc = obd_reconnect(req->rq_svc_thread->t_env,
1331                                    export, target, &cluuid, data, client_nid);
1332         }
1333         if (rc)
1334                 GOTO(out, rc);
1335
1336         LASSERT(target->u.obt.obt_magic == OBT_MAGIC);
1337         data->ocd_instance = target->u.obt.obt_instance;
1338
1339         /* Return only the parts of obd_connect_data that we understand, so the
1340          * client knows that we don't understand the rest. */
1341         if (data) {
1342                 tmpsize = req_capsule_get_size(&req->rq_pill, &RMF_CONNECT_DATA,
1343                                                RCL_SERVER);
1344                 tmpdata = req_capsule_server_get(&req->rq_pill,
1345                                                  &RMF_CONNECT_DATA);
1346                 /* Don't use struct assignment here, because the client reply
1347                  * buffer may be smaller/larger than the local struct
1348                  * obd_connect_data. */
1349                 memcpy(tmpdata, data, min(tmpsize, size));
1350         }
1351
1352         /* If the client and the server are the same node, we will already
1353          * have an export that really points to the client's DLM export,
1354          * because we have a shared handles table.
1355          *
1356          * XXX this will go away when shaver stops sending the "connect" handle
1357          * in the real "remote handle" field of the request --phik 24 Apr 2003
1358          */
1359         ptlrpc_request_change_export(req, export);
1360
1361         spin_lock(&export->exp_lock);
1362         if (export->exp_conn_cnt >= lustre_msg_get_conn_cnt(req->rq_reqmsg)) {
1363                 spin_unlock(&export->exp_lock);
1364                 CDEBUG(D_RPCTRACE, "%s: %s already connected at greater "
1365                        "or equal conn_cnt: %d >= %d\n",
1366                        cluuid.uuid, libcfs_nid2str(req->rq_peer.nid),
1367                        export->exp_conn_cnt,
1368                        lustre_msg_get_conn_cnt(req->rq_reqmsg));
1369
1370                 GOTO(out, rc = -EALREADY);
1371         }
1372         LASSERT(lustre_msg_get_conn_cnt(req->rq_reqmsg) > 0);
1373         export->exp_conn_cnt = lustre_msg_get_conn_cnt(req->rq_reqmsg);
1374         spin_unlock(&export->exp_lock);
1375
1376         if (export->exp_connection != NULL) {
1377                 /* Check to see if connection came from another NID. */
1378                 if ((export->exp_connection->c_peer.nid != req->rq_peer.nid) &&
1379                     !hlist_unhashed(&export->exp_nid_hash))
1380                         cfs_hash_del(export->exp_obd->obd_nid_hash,
1381                                      &export->exp_connection->c_peer.nid,
1382                                      &export->exp_nid_hash);
1383
1384                 ptlrpc_connection_put(export->exp_connection);
1385         }
1386
1387         export->exp_connection = ptlrpc_connection_get(req->rq_peer,
1388                                                        req->rq_self,
1389                                                        &cluuid);
1390         if (hlist_unhashed(&export->exp_nid_hash))
1391                 cfs_hash_add(export->exp_obd->obd_nid_hash,
1392                              &export->exp_connection->c_peer.nid,
1393                              &export->exp_nid_hash);
1394
1395         lustre_msg_set_handle(req->rq_repmsg, &conn);
1396
1397         rc = rev_import_reconnect(export, req);
1398         if (rc != 0)
1399                 GOTO(out, rc);
1400
1401         if (target->obd_recovering && !export->exp_in_recovery && !lw_client) {
1402                 int has_transno;
1403                 __u64 transno = data->ocd_transno;
1404
1405                 spin_lock(&export->exp_lock);
1406                 /* possible race with class_disconnect_stale_exports,
1407                  * export may be already in the eviction process */
1408                 if (export->exp_failed) {
1409                         spin_unlock(&export->exp_lock);
1410                         GOTO(out, rc = -ENODEV);
1411                 }
1412                 export->exp_in_recovery = 1;
1413                 export->exp_req_replay_needed = 1;
1414                 export->exp_lock_replay_needed = 1;
1415                 spin_unlock(&export->exp_lock);
1416
1417                 has_transno = !!(lustre_msg_get_op_flags(req->rq_reqmsg) &
1418                                  MSG_CONNECT_TRANSNO);
1419                 if (has_transno && transno == 0)
1420                         CWARN("Connect with zero transno!\n");
1421
1422                 if (has_transno && transno > 0 &&
1423                     transno < target->obd_next_recovery_transno &&
1424                     transno > target->obd_last_committed) {
1425                         /* Another way is to use cmpxchg() to be lock-free. */
1426                         spin_lock(&target->obd_recovery_task_lock);
1427                         if (transno < target->obd_next_recovery_transno)
1428                                 target->obd_next_recovery_transno = transno;
1429                         spin_unlock(&target->obd_recovery_task_lock);
1430                 }
1431
1432                 atomic_inc(&target->obd_req_replay_clients);
1433                 atomic_inc(&target->obd_lock_replay_clients);
1434                 /* Note: MDS-MDS connection is allowed to be connected during
1435                  * recovery, no matter if the exports needs to be recoveried.
1436                  * Because we need retrieve updates logs from all other MDTs.
1437                  * So if the MDS-MDS export is new, obd_max_recoverable_clients
1438                  * also needs to be increased to match other recovery checking
1439                  * condition. */
1440                 if (new_mds_mds_conn)
1441                         atomic_inc(&target->obd_max_recoverable_clients);
1442
1443                 if (atomic_inc_return(&target->obd_connected_clients) ==
1444                     atomic_read(&target->obd_max_recoverable_clients))
1445                         wake_up(&target->obd_next_transno_waitq);
1446         }
1447
1448         /* Tell the client we're in recovery, when client is involved in it. */
1449         if (target->obd_recovering && !lw_client)
1450                 lustre_msg_add_op_flags(req->rq_repmsg, MSG_CONNECT_RECOVERING);
1451
1452 out:
1453         if (export) {
1454                 spin_lock(&export->exp_lock);
1455                 export->exp_connecting = 0;
1456                 spin_unlock(&export->exp_lock);
1457
1458                 class_export_put(export);
1459         }
1460         if (target != NULL) {
1461                 spin_lock(&target->obd_dev_lock);
1462                 target->obd_conn_inprogress--;
1463                 spin_unlock(&target->obd_dev_lock);
1464                 class_decref(target, "find", current);
1465         }
1466         req->rq_status = rc;
1467         RETURN(rc);
1468 }
1469
1470 int target_handle_disconnect(struct ptlrpc_request *req)
1471 {
1472         int rc;
1473         ENTRY;
1474
1475         rc = req_capsule_server_pack(&req->rq_pill);
1476         if (rc)
1477                 RETURN(rc);
1478
1479         /* In case of target disconnect, updating sec ctx immediately is
1480          * required in order to record latest sequence number used.
1481          * Sequence is normally updated on export destroy, but this event
1482          * can occur too late, ie after a new target connect request has
1483          * been processed.
1484          * Maintaining correct sequence when client connection becomes idle
1485          * ensures that GSS does not erroneously consider requests as replays.
1486          */
1487         rc = sptlrpc_export_update_ctx(req->rq_export);
1488         if (rc)
1489                 RETURN(rc);
1490
1491         /* Keep the rq_export around so we can send the reply. */
1492         req->rq_status = obd_disconnect(class_export_get(req->rq_export));
1493
1494         RETURN(0);
1495 }
1496
1497 void target_destroy_export(struct obd_export *exp)
1498 {
1499         struct obd_import       *imp = NULL;
1500         /* exports created from last_rcvd data, and "fake"
1501            exports created by lctl don't have an import */
1502         spin_lock(&exp->exp_lock);
1503         if (exp->exp_imp_reverse != NULL) {
1504                 imp = exp->exp_imp_reverse;
1505                 exp->exp_imp_reverse = NULL;
1506         }
1507         spin_unlock(&exp->exp_lock);
1508         if (imp != NULL)
1509                 client_destroy_import(imp);
1510
1511         LASSERT_ATOMIC_ZERO(&exp->exp_locks_count);
1512         LASSERT_ATOMIC_ZERO(&exp->exp_rpc_count);
1513         LASSERT_ATOMIC_ZERO(&exp->exp_cb_count);
1514         LASSERT_ATOMIC_ZERO(&exp->exp_replay_count);
1515 }
1516 EXPORT_SYMBOL(target_destroy_export);
1517
1518 /*
1519  * Recovery functions
1520  */
1521 static void target_request_copy_get(struct ptlrpc_request *req)
1522 {
1523         class_export_rpc_inc(req->rq_export);
1524         LASSERT(list_empty(&req->rq_list));
1525         INIT_LIST_HEAD(&req->rq_replay_list);
1526
1527         /* Increase refcount to keep request in queue. */
1528         atomic_inc(&req->rq_refcount);
1529         /* Let export know it has replays to be handled. */
1530         atomic_inc(&req->rq_export->exp_replay_count);
1531 }
1532
1533 static void target_request_copy_put(struct ptlrpc_request *req)
1534 {
1535         LASSERT(list_empty(&req->rq_replay_list));
1536         LASSERT_ATOMIC_POS(&req->rq_export->exp_replay_count);
1537
1538         atomic_dec(&req->rq_export->exp_replay_count);
1539         class_export_rpc_dec(req->rq_export);
1540         ptlrpc_server_drop_request(req);
1541 }
1542
1543 static int target_exp_enqueue_req_replay(struct ptlrpc_request *req)
1544 {
1545         __u64                  transno = lustre_msg_get_transno(req->rq_reqmsg);
1546         struct obd_export     *exp = req->rq_export;
1547         struct ptlrpc_request *reqiter;
1548         struct ptlrpc_request *dup_req = NULL;
1549         int                    dup = 0;
1550
1551         LASSERT(exp);
1552
1553         spin_lock(&exp->exp_lock);
1554         list_for_each_entry(reqiter, &exp->exp_req_replay_queue,
1555                                 rq_replay_list) {
1556                 if (lustre_msg_get_transno(reqiter->rq_reqmsg) == transno) {
1557                         dup_req = reqiter;
1558                         dup = 1;
1559                         break;
1560                 }
1561         }
1562
1563         if (dup) {
1564                 /* We expect it with RESENT and REPLAY flags. */
1565                 if ((lustre_msg_get_flags(req->rq_reqmsg) &
1566                      (MSG_RESENT | MSG_REPLAY)) != (MSG_RESENT | MSG_REPLAY))
1567                         CERROR("invalid flags %x of resent replay\n",
1568                                lustre_msg_get_flags(req->rq_reqmsg));
1569
1570                 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY) {
1571                         __u32 new_conn;
1572
1573                         new_conn = lustre_msg_get_conn_cnt(req->rq_reqmsg);
1574                         if (new_conn >
1575                             lustre_msg_get_conn_cnt(dup_req->rq_reqmsg))
1576                                 lustre_msg_set_conn_cnt(dup_req->rq_reqmsg,
1577                                                         new_conn);
1578                 }
1579         } else {
1580                 list_add_tail(&req->rq_replay_list,
1581                                   &exp->exp_req_replay_queue);
1582         }
1583
1584         spin_unlock(&exp->exp_lock);
1585         return dup;
1586 }
1587
1588 static void target_exp_dequeue_req_replay(struct ptlrpc_request *req)
1589 {
1590         LASSERT(!list_empty(&req->rq_replay_list));
1591         LASSERT(req->rq_export);
1592
1593         spin_lock(&req->rq_export->exp_lock);
1594         list_del_init(&req->rq_replay_list);
1595         spin_unlock(&req->rq_export->exp_lock);
1596 }
1597
1598 static void target_finish_recovery(struct lu_target *lut)
1599 {
1600         struct obd_device *obd = lut->lut_obd;
1601         ENTRY;
1602
1603         /* Only log a recovery message when recovery has occurred. */
1604         if (obd->obd_recovery_start) {
1605                 time64_t now = ktime_get_seconds();
1606                 time64_t elapsed_time;
1607
1608                 elapsed_time = max_t(time64_t, now - obd->obd_recovery_start, 1);
1609                 LCONSOLE_INFO("%s: Recovery over after %lld:%.02lld, of %d clients "
1610                         "%d recovered and %d %s evicted.\n", obd->obd_name,
1611                         (s64)elapsed_time / 60, (s64)elapsed_time % 60,
1612                         atomic_read(&obd->obd_max_recoverable_clients),
1613                         atomic_read(&obd->obd_connected_clients),
1614                         obd->obd_stale_clients,
1615                         obd->obd_stale_clients == 1 ? "was" : "were");
1616         }
1617
1618         ldlm_reprocess_recovery_done(obd->obd_namespace);
1619         spin_lock(&obd->obd_recovery_task_lock);
1620         if (!list_empty(&obd->obd_req_replay_queue) ||
1621             !list_empty(&obd->obd_lock_replay_queue) ||
1622             !list_empty(&obd->obd_final_req_queue)) {
1623                 CERROR("%s: Recovery queues ( %s%s%s) are not empty\n",
1624                        obd->obd_name,
1625                        list_empty(&obd->obd_req_replay_queue) ? "" : "req ",
1626                        list_empty(&obd->obd_lock_replay_queue) ? \
1627                                "" : "lock ",
1628                        list_empty(&obd->obd_final_req_queue) ? \
1629                                "" : "final ");
1630                 spin_unlock(&obd->obd_recovery_task_lock);
1631                 LBUG();
1632         }
1633         spin_unlock(&obd->obd_recovery_task_lock);
1634
1635         obd->obd_recovery_end = ktime_get_seconds();
1636
1637         /* When recovery finished, cleanup orphans on MDS and OST. */
1638         if (obd->obd_type && OBP(obd, postrecov)) {
1639                 int rc = OBP(obd, postrecov)(obd);
1640
1641                 if (rc < 0)
1642                         LCONSOLE_WARN("%s: Post recovery failed, rc %d\n",
1643                                       obd->obd_name, rc);
1644         }
1645         EXIT;
1646 }
1647
1648 static void abort_req_replay_queue(struct obd_device *obd)
1649 {
1650         struct ptlrpc_request *req, *n;
1651         struct list_head abort_list;
1652
1653         INIT_LIST_HEAD(&abort_list);
1654         spin_lock(&obd->obd_recovery_task_lock);
1655         list_splice_init(&obd->obd_req_replay_queue, &abort_list);
1656         spin_unlock(&obd->obd_recovery_task_lock);
1657         list_for_each_entry_safe(req, n, &abort_list, rq_list) {
1658                 DEBUG_REQ(D_WARNING, req, "aborted:");
1659                 req->rq_status = -ENOTCONN;
1660                 if (ptlrpc_error(req)) {
1661                         DEBUG_REQ(D_ERROR, req,
1662                                   "failed abort_req_reply; skipping");
1663                 }
1664                 target_exp_dequeue_req_replay(req);
1665                 target_request_copy_put(req);
1666         }
1667 }
1668
1669 static void abort_lock_replay_queue(struct obd_device *obd)
1670 {
1671         struct ptlrpc_request *req, *n;
1672         struct list_head abort_list;
1673
1674         INIT_LIST_HEAD(&abort_list);
1675         spin_lock(&obd->obd_recovery_task_lock);
1676         list_splice_init(&obd->obd_lock_replay_queue, &abort_list);
1677         spin_unlock(&obd->obd_recovery_task_lock);
1678         list_for_each_entry_safe(req, n, &abort_list, rq_list) {
1679                 DEBUG_REQ(D_ERROR, req, "aborted:");
1680                 req->rq_status = -ENOTCONN;
1681                 if (ptlrpc_error(req)) {
1682                         DEBUG_REQ(D_ERROR, req,
1683                                   "failed abort_lock_reply; skipping");
1684                 }
1685                 target_request_copy_put(req);
1686         }
1687 }
1688
1689 /* Called from a cleanup function if the device is being cleaned up
1690    forcefully.  The exports should all have been disconnected already,
1691    the only thing left to do is
1692      - clear the recovery flags
1693      - cancel the timer
1694      - free queued requests and replies, but don't send replies
1695    Because the obd_stopping flag is set, no new requests should be received.
1696
1697 */
1698 void target_cleanup_recovery(struct obd_device *obd)
1699 {
1700         struct ptlrpc_request *req, *n;
1701         struct list_head clean_list;
1702
1703         INIT_LIST_HEAD(&clean_list);
1704         spin_lock(&obd->obd_dev_lock);
1705         if (!obd->obd_recovering) {
1706                 spin_unlock(&obd->obd_dev_lock);
1707                 EXIT;
1708                 return;
1709         }
1710         obd->obd_recovering = obd->obd_abort_recovery = 0;
1711         spin_unlock(&obd->obd_dev_lock);
1712
1713         spin_lock(&obd->obd_recovery_task_lock);
1714         target_cancel_recovery_timer(obd);
1715         list_splice_init(&obd->obd_req_replay_queue, &clean_list);
1716         spin_unlock(&obd->obd_recovery_task_lock);
1717
1718         list_for_each_entry_safe(req, n, &clean_list, rq_list) {
1719                 LASSERT(req->rq_reply_state == NULL);
1720                 target_exp_dequeue_req_replay(req);
1721                 target_request_copy_put(req);
1722         }
1723
1724         spin_lock(&obd->obd_recovery_task_lock);
1725         list_splice_init(&obd->obd_lock_replay_queue, &clean_list);
1726         list_splice_init(&obd->obd_final_req_queue, &clean_list);
1727         spin_unlock(&obd->obd_recovery_task_lock);
1728
1729         list_for_each_entry_safe(req, n, &clean_list, rq_list) {
1730                 LASSERT(req->rq_reply_state == NULL);
1731                 target_request_copy_put(req);
1732         }
1733
1734         EXIT;
1735 }
1736 EXPORT_SYMBOL(target_cleanup_recovery);
1737
1738 /* obd_recovery_task_lock should be held */
1739 void target_cancel_recovery_timer(struct obd_device *obd)
1740 {
1741         CDEBUG(D_HA, "%s: cancel recovery timer\n", obd->obd_name);
1742         hrtimer_cancel(&obd->obd_recovery_timer);
1743 }
1744
1745 static void target_start_recovery_timer(struct obd_device *obd)
1746 {
1747         ktime_t delay;
1748
1749         if (obd->obd_recovery_start != 0)
1750                 return;
1751
1752         spin_lock(&obd->obd_dev_lock);
1753         if (!obd->obd_recovering || obd->obd_abort_recovery) {
1754                 spin_unlock(&obd->obd_dev_lock);
1755                 return;
1756         }
1757
1758         LASSERT(obd->obd_recovery_timeout != 0);
1759
1760         if (obd->obd_recovery_start != 0) {
1761                 spin_unlock(&obd->obd_dev_lock);
1762                 return;
1763         }
1764
1765         obd->obd_recovery_start = ktime_get_seconds();
1766         delay = ktime_set(obd->obd_recovery_start +
1767                           obd->obd_recovery_timeout, 0);
1768         hrtimer_start(&obd->obd_recovery_timer, delay, HRTIMER_MODE_ABS);
1769         spin_unlock(&obd->obd_dev_lock);
1770
1771         LCONSOLE_WARN("%s: Will be in recovery for at least %llu:%02llu, or until %d client%s reconnect%s\n",
1772                       obd->obd_name,
1773                       obd->obd_recovery_timeout / 60,
1774                       obd->obd_recovery_timeout % 60,
1775                       atomic_read(&obd->obd_max_recoverable_clients),
1776                       (atomic_read(&obd->obd_max_recoverable_clients) == 1) ?
1777                       "" : "s",
1778                       (atomic_read(&obd->obd_max_recoverable_clients) == 1) ?
1779                       "s" : "");
1780 }
1781
1782 /**
1783  * extend recovery window.
1784  *
1785  * if @extend is true, extend recovery window to have @dr_timeout remaining
1786  * at least; otherwise, make sure the recovery timeout value is not less
1787  * than @dr_timeout.
1788  */
1789 static void extend_recovery_timer(struct obd_device *obd, time_t dr_timeout,
1790                                   bool extend)
1791 {
1792         ktime_t left_ns;
1793         time_t timeout;
1794         time_t left;
1795
1796         spin_lock(&obd->obd_dev_lock);
1797         if (!obd->obd_recovering || obd->obd_abort_recovery) {
1798                 spin_unlock(&obd->obd_dev_lock);
1799                 return;
1800         }
1801         LASSERT(obd->obd_recovery_start != 0);
1802
1803         left_ns = hrtimer_expires_remaining(&obd->obd_recovery_timer);
1804         left = ktime_divns(left_ns, NSEC_PER_SEC);
1805
1806         if (extend) {
1807                 timeout = obd->obd_recovery_timeout;
1808                 /* dr_timeout will happen after the hrtimer has expired.
1809                  * Add the excess time to the soft recovery timeout without
1810                  * exceeding the hard recovery timeout.
1811                  */
1812                 if (dr_timeout > left) {
1813                         timeout += dr_timeout - left;
1814                         timeout = min_t(time_t, obd->obd_recovery_time_hard,
1815                                         timeout);
1816                 }
1817         } else {
1818                 timeout = clamp_t(time_t, dr_timeout, obd->obd_recovery_timeout,
1819                                   obd->obd_recovery_time_hard);
1820         }
1821
1822         if (timeout == obd->obd_recovery_time_hard)
1823                 CWARN("%s: extended recovery timer reached hard limit: %ld, extend: %d\n",
1824                       obd->obd_name, timeout, extend);
1825
1826         if (obd->obd_recovery_timeout < timeout) {
1827                 ktime_t end, now;
1828
1829                 obd->obd_recovery_timeout = timeout;
1830                 end = ktime_set(obd->obd_recovery_start + timeout, 0);
1831                 now = ktime_set(ktime_get_seconds(), 0);
1832                 left_ns = ktime_sub(end, now);
1833                 hrtimer_start(&obd->obd_recovery_timer, end, HRTIMER_MODE_ABS);
1834                 left = ktime_divns(left_ns, NSEC_PER_SEC);
1835         }
1836         spin_unlock(&obd->obd_dev_lock);
1837
1838         CDEBUG(D_HA, "%s: recovery timer will expire in %ld seconds\n",
1839                 obd->obd_name, left);
1840 }
1841
1842 /* Reset the timer with each new client connection */
1843 /*
1844  * This timer is actually reconnect_timer, which is for making sure
1845  * the total recovery window is at least as big as my reconnect
1846  * attempt timing. So the initial recovery time_out will be set to
1847  * OBD_RECOVERY_FACTOR * obd_timeout. If the timeout coming
1848  * from client is bigger than this, then the recovery time_out will
1849  * be extended to make sure the client could be reconnected, in the
1850  * process, the timeout from the new client should be ignored.
1851  */
1852 static void
1853 check_and_start_recovery_timer(struct obd_device *obd,
1854                                struct ptlrpc_request *req,
1855                                int new_client)
1856 {
1857         time_t service_time = lustre_msg_get_service_time(req->rq_reqmsg);
1858         struct obd_device_target *obt = &obd->u.obt;
1859
1860         if (!new_client && service_time)
1861                 /* Teach server about old server's estimates, as first guess
1862                  * at how long new requests will take.
1863                  */
1864                 at_measured(&req->rq_rqbd->rqbd_svcpt->scp_at_estimate,
1865                             service_time);
1866
1867         target_start_recovery_timer(obd);
1868
1869         /* Convert the service time to RPC timeout,
1870          * and reuse service_time to limit stack usage.
1871          */
1872         service_time = at_est2timeout(service_time);
1873
1874         if (OBD_FAIL_CHECK(OBD_FAIL_TGT_SLUGGISH_NET) &&
1875             service_time < at_extra)
1876                 service_time = at_extra;
1877
1878         /* We expect other clients to timeout within service_time, then try
1879          * to reconnect, then try the failover server.  The max delay between
1880          * connect attempts is SWITCH_MAX + SWITCH_INC + INITIAL. */
1881         service_time += 2 * INITIAL_CONNECT_TIMEOUT;
1882
1883         LASSERT(obt->obt_magic == OBT_MAGIC);
1884         service_time += 2 * (CONNECTION_SWITCH_MAX + CONNECTION_SWITCH_INC);
1885         if (service_time > obd->obd_recovery_timeout && !new_client)
1886                 extend_recovery_timer(obd, service_time, false);
1887 }
1888
1889 /** Health checking routines */
1890 static inline int exp_connect_healthy(struct obd_export *exp)
1891 {
1892         return (exp->exp_in_recovery);
1893 }
1894
1895 /** if export done req_replay or has replay in queue */
1896 static inline int exp_req_replay_healthy(struct obd_export *exp)
1897 {
1898         return (!exp->exp_req_replay_needed ||
1899                 atomic_read(&exp->exp_replay_count) > 0);
1900 }
1901
1902
1903 static inline int exp_req_replay_healthy_or_from_mdt(struct obd_export *exp)
1904 {
1905         return (exp_connect_flags(exp) & OBD_CONNECT_MDS_MDS) ||
1906                exp_req_replay_healthy(exp);
1907 }
1908
1909 /** if export done lock_replay or has replay in queue */
1910 static inline int exp_lock_replay_healthy(struct obd_export *exp)
1911 {
1912         return (!exp->exp_lock_replay_needed ||
1913                 atomic_read(&exp->exp_replay_count) > 0);
1914 }
1915
1916 static inline int exp_vbr_healthy(struct obd_export *exp)
1917 {
1918         return (!exp->exp_vbr_failed);
1919 }
1920
1921 static inline int exp_finished(struct obd_export *exp)
1922 {
1923         return (exp->exp_in_recovery && !exp->exp_lock_replay_needed);
1924 }
1925
1926 static inline int exp_finished_or_from_mdt(struct obd_export *exp)
1927 {
1928         return (exp_connect_flags(exp) & OBD_CONNECT_MDS_MDS) ||
1929                 exp_finished(exp);
1930 }
1931
1932 static int check_for_next_transno(struct lu_target *lut)
1933 {
1934         struct ptlrpc_request *req = NULL;
1935         struct obd_device *obd = lut->lut_obd;
1936         struct target_distribute_txn_data *tdtd = lut->lut_tdtd;
1937         int wake_up = 0, connected, completed, queue_len;
1938         __u64 req_transno = 0;
1939         __u64 update_transno = 0;
1940         __u64 next_transno = 0;
1941         ENTRY;
1942
1943         spin_lock(&obd->obd_recovery_task_lock);
1944         if (!list_empty(&obd->obd_req_replay_queue)) {
1945                 req = list_entry(obd->obd_req_replay_queue.next,
1946                                      struct ptlrpc_request, rq_list);
1947                 req_transno = lustre_msg_get_transno(req->rq_reqmsg);
1948         }
1949
1950         if (tdtd != NULL)
1951                 update_transno = distribute_txn_get_next_transno(tdtd);
1952
1953         connected = atomic_read(&obd->obd_connected_clients);
1954         completed = connected - atomic_read(&obd->obd_req_replay_clients);
1955         queue_len = obd->obd_requests_queued_for_recovery;
1956         next_transno = obd->obd_next_recovery_transno;
1957
1958         CDEBUG(D_HA,
1959                "max: %d, connected: %d, completed: %d, queue_len: %d, req_transno: %llu, next_transno: %llu\n",
1960                atomic_read(&obd->obd_max_recoverable_clients),
1961                connected, completed,
1962                queue_len, req_transno, next_transno);
1963
1964         if (obd->obd_abort_recovery) {
1965                 CDEBUG(D_HA, "waking for aborted recovery\n");
1966                 wake_up = 1;
1967         } else if (obd->obd_recovery_expired) {
1968                 CDEBUG(D_HA, "waking for expired recovery\n");
1969                 wake_up = 1;
1970         } else if (tdtd != NULL && req != NULL &&
1971                    is_req_replayed_by_update(req)) {
1972                 LASSERTF(req_transno < next_transno, "req_transno %llu"
1973                          "next_transno%llu\n", req_transno, next_transno);
1974                 CDEBUG(D_HA, "waking for duplicate req (%llu)\n",
1975                        req_transno);
1976                 wake_up = 1;
1977         } else if (req_transno == next_transno ||
1978                    (update_transno != 0 && update_transno <= next_transno)) {
1979                 CDEBUG(D_HA, "waking for next (%lld)\n", next_transno);
1980                 wake_up = 1;
1981         } else if (queue_len > 0 &&
1982                    queue_len == atomic_read(&obd->obd_req_replay_clients)) {
1983                 /** handle gaps occured due to lost reply or VBR */
1984                 LASSERTF(req_transno >= next_transno,
1985                          "req_transno: %llu, next_transno: %llu\n",
1986                          req_transno, next_transno);
1987                 CDEBUG(D_HA,
1988                        "%s: waking for gap in transno, VBR is %s (skip: "
1989                        "%lld, ql: %d, comp: %d, conn: %d, next: %lld"
1990                        ", next_update %lld last_committed: %lld)\n",
1991                        obd->obd_name, obd->obd_version_recov ? "ON" : "OFF",
1992                        next_transno, queue_len, completed, connected,
1993                        req_transno, update_transno, obd->obd_last_committed);
1994                 obd->obd_next_recovery_transno = req_transno;
1995                 wake_up = 1;
1996         } else if (atomic_read(&obd->obd_req_replay_clients) == 0) {
1997                 CDEBUG(D_HA, "waking for completed recovery\n");
1998                 wake_up = 1;
1999         } else if (OBD_FAIL_CHECK(OBD_FAIL_MDS_RECOVERY_ACCEPTS_GAPS)) {
2000                 CDEBUG(D_HA, "accepting transno gaps is explicitly allowed"
2001                        " by fail_lock, waking up (%lld)\n", next_transno);
2002                 obd->obd_next_recovery_transno = req_transno;
2003                 wake_up = 1;
2004         }
2005         spin_unlock(&obd->obd_recovery_task_lock);
2006         return wake_up;
2007 }
2008
2009 static int check_for_next_lock(struct lu_target *lut)
2010 {
2011         struct obd_device *obd = lut->lut_obd;
2012         int wake_up = 0;
2013
2014         spin_lock(&obd->obd_recovery_task_lock);
2015         if (!list_empty(&obd->obd_lock_replay_queue)) {
2016                 CDEBUG(D_HA, "waking for next lock\n");
2017                 wake_up = 1;
2018         } else if (atomic_read(&obd->obd_lock_replay_clients) == 0) {
2019                 CDEBUG(D_HA, "waking for completed lock replay\n");
2020                 wake_up = 1;
2021         } else if (obd->obd_abort_recovery) {
2022                 CDEBUG(D_HA, "waking for aborted recovery\n");
2023                 wake_up = 1;
2024         } else if (obd->obd_recovery_expired) {
2025                 CDEBUG(D_HA, "waking for expired recovery\n");
2026                 wake_up = 1;
2027         }
2028         spin_unlock(&obd->obd_recovery_task_lock);
2029
2030         return wake_up;
2031 }
2032
2033 static int check_update_llog(struct lu_target *lut)
2034 {
2035         struct obd_device *obd = lut->lut_obd;
2036         struct target_distribute_txn_data *tdtd = lut->lut_tdtd;
2037
2038         if (obd->obd_abort_recovery) {
2039                 CDEBUG(D_HA, "waking for aborted recovery\n");
2040                 return 1;
2041         }
2042
2043         if (atomic_read(&tdtd->tdtd_recovery_threads_count) == 0) {
2044                 CDEBUG(D_HA, "waking for completion of reading update log\n");
2045                 return 1;
2046         }
2047
2048         return 0;
2049 }
2050
2051 /**
2052  * wait for recovery events,
2053  * check its status with help of check_routine
2054  * evict dead clients via health_check
2055  */
2056 static int target_recovery_overseer(struct lu_target *lut,
2057                                     int (*check_routine)(struct lu_target *),
2058                                     int (*health_check)(struct obd_export *))
2059 {
2060         struct obd_device       *obd = lut->lut_obd;
2061         struct target_distribute_txn_data *tdtd;
2062         time64_t last = 0;
2063         time64_t now;
2064 repeat:
2065         if (obd->obd_recovering && obd->obd_recovery_start == 0) {
2066                 now = ktime_get_seconds();
2067                 if (now - last > 600) {
2068                         LCONSOLE_INFO("%s: in recovery but waiting for "
2069                                       "the first client to connect\n",
2070                                       obd->obd_name);
2071                         last = now;
2072                 }
2073         }
2074         if (obd->obd_recovery_start != 0 && ktime_get_seconds() >=
2075               (obd->obd_recovery_start + obd->obd_recovery_time_hard)) {
2076                 __u64 next_update_transno = 0;
2077
2078                 /* Only abort the recovery if there are no update recovery
2079                  * left in the queue */
2080                 spin_lock(&obd->obd_recovery_task_lock);
2081                 if (lut->lut_tdtd != NULL) {
2082                         next_update_transno =
2083                                 distribute_txn_get_next_transno(lut->lut_tdtd);
2084
2085                         tdtd = lut->lut_tdtd;
2086                         /* If next_update_transno == 0, it probably because
2087                          * updatelog retrieve threads did not get any records
2088                          * yet, let's wait those threads stopped */
2089                         if (next_update_transno == 0) {
2090                                 spin_unlock(&obd->obd_recovery_task_lock);
2091
2092                                 while (wait_event_timeout(
2093                                         tdtd->tdtd_recovery_threads_waitq,
2094                                         check_update_llog(lut),
2095                                         cfs_time_seconds(60)) == 0);
2096
2097                                 spin_lock(&obd->obd_recovery_task_lock);
2098                                 next_update_transno =
2099                                         distribute_txn_get_next_transno(tdtd);
2100                         }
2101                 }
2102
2103                 if (next_update_transno != 0 && !obd->obd_abort_recovery) {
2104                         obd->obd_next_recovery_transno = next_update_transno;
2105                         spin_unlock(&obd->obd_recovery_task_lock);
2106                         /* Disconnect unfinished exports from clients, and
2107                          * keep connection from MDT to make sure the update
2108                          * recovery will still keep trying until some one
2109                          * manually abort the recovery */
2110                         class_disconnect_stale_exports(obd,
2111                                                 exp_finished_or_from_mdt);
2112                         /* Abort all of replay and replay lock req from
2113                          * clients */
2114                         abort_req_replay_queue(obd);
2115                         abort_lock_replay_queue(obd);
2116                         CDEBUG(D_HA, "%s: there are still update replay (%#llx"
2117                                ")in the queue.\n", obd->obd_name,
2118                                next_update_transno);
2119                 } else {
2120                         obd->obd_abort_recovery = 1;
2121                         spin_unlock(&obd->obd_recovery_task_lock);
2122                         CWARN("%s recovery is aborted by hard timeout\n",
2123                               obd->obd_name);
2124                 }
2125         }
2126
2127         while (wait_event_timeout(obd->obd_next_transno_waitq,
2128                                   check_routine(lut),
2129                                   msecs_to_jiffies(60 * MSEC_PER_SEC)) == 0)
2130                 /* wait indefinitely for event, but don't trigger watchdog */;
2131
2132         if (obd->obd_abort_recovery) {
2133                 CWARN("recovery is aborted, evict exports in recovery\n");
2134                 if (lut->lut_tdtd != NULL) {
2135                         struct l_wait_info lwi = { 0 };
2136
2137                         tdtd = lut->lut_tdtd;
2138                         /* Let's wait all of the update log recovery thread
2139                          * finished */
2140                         l_wait_event(tdtd->tdtd_recovery_threads_waitq,
2141                          atomic_read(&tdtd->tdtd_recovery_threads_count) == 0,
2142                              &lwi);
2143                         /* Then abort the update recovery list */
2144                         dtrq_list_destroy(lut->lut_tdtd);
2145                 }
2146
2147                 /** evict exports which didn't finish recovery yet */
2148                 class_disconnect_stale_exports(obd, exp_finished);
2149                 return 1;
2150         } else if (obd->obd_recovery_expired) {
2151                 obd->obd_recovery_expired = 0;
2152
2153                 /** If some clients died being recovered, evict them */
2154                 LCONSOLE_WARN("%s: recovery is timed out, "
2155                               "evict stale exports\n", obd->obd_name);
2156                 /** evict cexports with no replay in queue, they are stalled */
2157                 class_disconnect_stale_exports(obd, health_check);
2158
2159                 /** continue with VBR */
2160                 spin_lock(&obd->obd_dev_lock);
2161                 obd->obd_version_recov = 1;
2162                 spin_unlock(&obd->obd_dev_lock);
2163                 /**
2164                  * reset timer, recovery will proceed with versions now,
2165                  * timeout is set just to handle reconnection delays
2166                  */
2167                 extend_recovery_timer(obd, RECONNECT_DELAY_MAX, true);
2168                 /** Wait for recovery events again, after evicting bad clients */
2169                 goto repeat;
2170         }
2171         return 0;
2172 }
2173
2174 static struct ptlrpc_request *target_next_replay_lock(struct lu_target *lut)
2175 {
2176         struct obd_device       *obd = lut->lut_obd;
2177         struct ptlrpc_request *req = NULL;
2178
2179         CDEBUG(D_HA, "Waiting for lock\n");
2180         if (target_recovery_overseer(lut, check_for_next_lock,
2181                                      exp_lock_replay_healthy))
2182                 abort_lock_replay_queue(obd);
2183
2184         spin_lock(&obd->obd_recovery_task_lock);
2185         if (!list_empty(&obd->obd_lock_replay_queue)) {
2186                 req = list_entry(obd->obd_lock_replay_queue.next,
2187                                      struct ptlrpc_request, rq_list);
2188                 list_del_init(&req->rq_list);
2189                 spin_unlock(&obd->obd_recovery_task_lock);
2190         } else {
2191                 spin_unlock(&obd->obd_recovery_task_lock);
2192                 LASSERT(list_empty(&obd->obd_lock_replay_queue));
2193                 LASSERT(atomic_read(&obd->obd_lock_replay_clients) == 0);
2194                 /** evict exports failed VBR */
2195                 class_disconnect_stale_exports(obd, exp_vbr_healthy);
2196         }
2197         return req;
2198 }
2199
2200 static struct ptlrpc_request *target_next_final_ping(struct obd_device *obd)
2201 {
2202         struct ptlrpc_request *req = NULL;
2203
2204         spin_lock(&obd->obd_recovery_task_lock);
2205         if (!list_empty(&obd->obd_final_req_queue)) {
2206                 req = list_entry(obd->obd_final_req_queue.next,
2207                                      struct ptlrpc_request, rq_list);
2208                 list_del_init(&req->rq_list);
2209                 spin_unlock(&obd->obd_recovery_task_lock);
2210                 if (req->rq_export->exp_in_recovery) {
2211                         spin_lock(&req->rq_export->exp_lock);
2212                         req->rq_export->exp_in_recovery = 0;
2213                         spin_unlock(&req->rq_export->exp_lock);
2214                 }
2215         } else {
2216                 spin_unlock(&obd->obd_recovery_task_lock);
2217         }
2218         return req;
2219 }
2220
2221 static void handle_recovery_req(struct ptlrpc_thread *thread,
2222                                 struct ptlrpc_request *req,
2223                                 svc_handler_t handler)
2224 {
2225         ENTRY;
2226
2227         /**
2228          * export can be evicted during recovery, no need to handle replays for
2229          * it after that, discard such request silently
2230          */
2231         if (req->rq_export->exp_disconnected)
2232                 RETURN_EXIT;
2233
2234         req->rq_session.lc_thread = thread;
2235         req->rq_svc_thread = thread;
2236         req->rq_svc_thread->t_env->le_ses = &req->rq_session;
2237
2238         /* thread context */
2239         lu_context_enter(&thread->t_env->le_ctx);
2240         (void)handler(req);
2241         lu_context_exit(&thread->t_env->le_ctx);
2242
2243         req->rq_svc_thread->t_env->le_ses = NULL;
2244
2245         /* don't reset timer for final stage */
2246         if (!exp_finished(req->rq_export)) {
2247                 time_t to = obd_timeout;
2248
2249                 /**
2250                  * Add request timeout to the recovery time so next request from
2251                  * this client may come in recovery time
2252                  */
2253                 if (!AT_OFF) {
2254                         struct ptlrpc_service_part *svcpt;
2255
2256                         svcpt = req->rq_rqbd->rqbd_svcpt;
2257                         /* If the server sent early reply for this request,
2258                          * the client will recalculate the timeout according to
2259                          * current server estimate service time, so we will
2260                          * use the maxium timeout here for waiting the client
2261                          * sending the next req
2262                          */
2263                         to = max_t(time_t, at_est2timeout(at_get(&svcpt->scp_at_estimate)),
2264                                    lustre_msg_get_timeout(req->rq_reqmsg));
2265                         /* Add 2 net_latency, one for balance rq_deadline
2266                          * (see ptl_send_rpc), one for resend the req to server,
2267                          * Note: client will pack net_latency in replay req
2268                          * (see ptlrpc_replay_req) */
2269                         to += 2 * lustre_msg_get_service_time(req->rq_reqmsg);
2270                 }
2271                 extend_recovery_timer(class_exp2obd(req->rq_export), to, true);
2272         }
2273         EXIT;
2274 }
2275
2276 /** Checking routines for recovery */
2277 static int check_for_recovery_ready(struct lu_target *lut)
2278 {
2279         struct obd_device *obd = lut->lut_obd;
2280         unsigned int clnts = atomic_read(&obd->obd_connected_clients);
2281
2282         CDEBUG(D_HA,
2283                "connected %d stale %d max_recoverable_clients %d abort %d expired %d\n",
2284                clnts, obd->obd_stale_clients,
2285                atomic_read(&obd->obd_max_recoverable_clients),
2286                obd->obd_abort_recovery, obd->obd_recovery_expired);
2287
2288         if (!obd->obd_abort_recovery && !obd->obd_recovery_expired) {
2289                 LASSERT(clnts <=
2290                         atomic_read(&obd->obd_max_recoverable_clients));
2291                 if (clnts + obd->obd_stale_clients <
2292                     atomic_read(&obd->obd_max_recoverable_clients))
2293                         return 0;
2294         }
2295
2296         if (lut->lut_tdtd != NULL) {
2297                 if (!lut->lut_tdtd->tdtd_replay_ready &&
2298                     !obd->obd_abort_recovery) {
2299                         /* Let's extend recovery timer, in case the recovery
2300                          * timer expired, and some clients got evicted */
2301                         extend_recovery_timer(obd, obd->obd_recovery_timeout,
2302                                               true);
2303                         CDEBUG(D_HA, "%s update recovery is not ready, extend recovery %llu\n",
2304                                obd->obd_name, obd->obd_recovery_timeout);
2305                         return 0;
2306                 }
2307         }
2308
2309         return 1;
2310 }
2311
2312 enum {
2313         REQUEST_RECOVERY = 1,
2314         UPDATE_RECOVERY = 2,
2315 };
2316
2317 static __u64 get_next_replay_req_transno(struct obd_device *obd)
2318 {
2319         __u64 transno = 0;
2320
2321         if (!list_empty(&obd->obd_req_replay_queue)) {
2322                 struct ptlrpc_request *req;
2323
2324                 req = list_entry(obd->obd_req_replay_queue.next,
2325                                  struct ptlrpc_request, rq_list);
2326                 transno = lustre_msg_get_transno(req->rq_reqmsg);
2327         }
2328
2329         return transno;
2330 }
2331
2332 static __u64 get_next_transno(struct lu_target *lut, int *type)
2333 {
2334         struct obd_device *obd = lut->lut_obd;
2335         struct target_distribute_txn_data *tdtd = lut->lut_tdtd;
2336         __u64 transno = 0;
2337         __u64 update_transno;
2338         ENTRY;
2339
2340         transno = get_next_replay_req_transno(obd);
2341         if (type != NULL)
2342                 *type = REQUEST_RECOVERY;
2343
2344         if (tdtd == NULL)
2345                 RETURN(transno);
2346
2347         update_transno = distribute_txn_get_next_transno(tdtd);
2348         if (transno == 0 || (transno >= update_transno &&
2349                              update_transno != 0)) {
2350                 transno = update_transno;
2351                 if (type != NULL)
2352                         *type = UPDATE_RECOVERY;
2353         }
2354
2355         RETURN(transno);
2356 }
2357
2358 /**
2359  * drop duplicate replay request
2360  *
2361  * Because the operation has been replayed by update recovery, the request
2362  * with the same transno will be dropped and also notify the client to send
2363  * next replay request.
2364  *
2365  * \param[in] env       execution environment
2366  * \param[in] obd       failover obd device
2367  * \param[in] req       request to be dropped
2368  */
2369 static void drop_duplicate_replay_req(struct lu_env *env,
2370                                       struct obd_device *obd,
2371                                       struct ptlrpc_request *req)
2372 {
2373         DEBUG_REQ(D_HA, req, "remove t%lld from %s because of duplicate"
2374                   " update records are found.\n",
2375                   lustre_msg_get_transno(req->rq_reqmsg),
2376                   libcfs_nid2str(req->rq_peer.nid));
2377
2378         /* Right now, only for MDS reint operation update replay and
2379          * normal request replay can have the same transno */
2380         if (lustre_msg_get_opc(req->rq_reqmsg) == MDS_REINT) {
2381                 req_capsule_set(&req->rq_pill, &RQF_MDS_REINT);
2382                 req->rq_status = req_capsule_server_pack(&req->rq_pill);
2383                 if (likely(req->rq_export))
2384                         target_committed_to_req(req);
2385                 lustre_msg_set_transno(req->rq_repmsg, req->rq_transno);
2386                 target_send_reply(req, req->rq_status, 0);
2387         } else {
2388                 DEBUG_REQ(D_ERROR, req, "wrong opc" "from %s\n",
2389                 libcfs_nid2str(req->rq_peer.nid));
2390         }
2391         target_exp_dequeue_req_replay(req);
2392         target_request_copy_put(req);
2393         obd->obd_replayed_requests++;
2394 }
2395
2396 #define WATCHDOG_TIMEOUT (obd_timeout * 10)
2397
2398 static void replay_request_or_update(struct lu_env *env,
2399                                      struct lu_target *lut,
2400                                      struct target_recovery_data *trd,
2401                                      struct ptlrpc_thread *thread)
2402 {
2403         struct obd_device *obd = lut->lut_obd;
2404         struct ptlrpc_request *req = NULL;
2405         int                     type;
2406         __u64                   transno;
2407         ENTRY;
2408
2409         CDEBUG(D_HA, "Waiting for transno %lld\n",
2410                obd->obd_next_recovery_transno);
2411
2412         /* Replay all of request and update by transno */
2413         do {
2414                 struct target_distribute_txn_data *tdtd = lut->lut_tdtd;
2415
2416                 CFS_FAIL_TIMEOUT(OBD_FAIL_TGT_REPLAY_DELAY2, cfs_fail_val);
2417
2418                 /** It is needed to extend recovery window above
2419                  *  recovery_time_soft. Extending is possible only in the
2420                  *  end of recovery window (see more details in
2421                  *  handle_recovery_req()).
2422                  */
2423                 CFS_FAIL_TIMEOUT_MS(OBD_FAIL_TGT_REPLAY_DELAY, 300);
2424
2425                 if (target_recovery_overseer(lut, check_for_next_transno,
2426                                         exp_req_replay_healthy_or_from_mdt)) {
2427                         abort_req_replay_queue(obd);
2428                         abort_lock_replay_queue(obd);
2429                         goto abort;
2430                 }
2431
2432                 spin_lock(&obd->obd_recovery_task_lock);
2433                 transno = get_next_transno(lut, &type);
2434                 if (type == REQUEST_RECOVERY && transno != 0) {
2435                         /* Drop replay request from client side, if the
2436                          * replay has been executed by update with the
2437                          * same transno */
2438                         req = list_entry(obd->obd_req_replay_queue.next,
2439                                         struct ptlrpc_request, rq_list);
2440
2441                         list_del_init(&req->rq_list);
2442                         obd->obd_requests_queued_for_recovery--;
2443                         spin_unlock(&obd->obd_recovery_task_lock);
2444
2445                         /* Let's check if the request has been redone by
2446                          * update replay */
2447                         if (is_req_replayed_by_update(req)) {
2448                                 struct distribute_txn_replay_req *dtrq;
2449
2450                                 dtrq = distribute_txn_lookup_finish_list(tdtd,
2451                                                                    req->rq_xid);
2452                                 LASSERT(dtrq != NULL);
2453                                 spin_lock(&tdtd->tdtd_replay_list_lock);
2454                                 list_del_init(&dtrq->dtrq_list);
2455                                 spin_unlock(&tdtd->tdtd_replay_list_lock);
2456                                 dtrq_destroy(dtrq);
2457
2458                                 drop_duplicate_replay_req(env, obd, req);
2459
2460                                 continue;
2461                         }
2462
2463                         LASSERT(trd->trd_processing_task == current_pid());
2464                         DEBUG_REQ(D_HA, req, "processing t%lld from %s",
2465                                   lustre_msg_get_transno(req->rq_reqmsg),
2466                                   libcfs_nid2str(req->rq_peer.nid));
2467
2468                         thread->t_watchdog = lc_watchdog_add(WATCHDOG_TIMEOUT,
2469                                                              NULL, NULL);
2470                         handle_recovery_req(thread, req,
2471                                             trd->trd_recovery_handler);
2472                         lc_watchdog_delete(thread->t_watchdog);
2473                         thread->t_watchdog = NULL;
2474
2475                         /**
2476                          * bz18031: increase next_recovery_transno before
2477                          * target_request_copy_put() will drop exp_rpc reference
2478                          */
2479                         spin_lock(&obd->obd_recovery_task_lock);
2480                         obd->obd_next_recovery_transno++;
2481                         spin_unlock(&obd->obd_recovery_task_lock);
2482                         target_exp_dequeue_req_replay(req);
2483                         target_request_copy_put(req);
2484                         obd->obd_replayed_requests++;
2485                 } else if (type == UPDATE_RECOVERY && transno != 0) {
2486                         struct distribute_txn_replay_req *dtrq;
2487                         int rc;
2488
2489                         spin_unlock(&obd->obd_recovery_task_lock);
2490
2491                         LASSERT(tdtd != NULL);
2492                         dtrq = distribute_txn_get_next_req(tdtd);
2493                         lu_context_enter(&thread->t_env->le_ctx);
2494                         thread->t_watchdog = lc_watchdog_add(WATCHDOG_TIMEOUT,
2495                                                              NULL, NULL);
2496                         rc = tdtd->tdtd_replay_handler(env, tdtd, dtrq);
2497                         lc_watchdog_delete(thread->t_watchdog);
2498                         thread->t_watchdog = NULL;
2499                         lu_context_exit(&thread->t_env->le_ctx);
2500                         extend_recovery_timer(obd, obd_timeout, true);
2501
2502                         if (rc == 0 && dtrq->dtrq_xid != 0) {
2503                                 CDEBUG(D_HA, "Move x%llu t%llu"
2504                                        " to finish list\n", dtrq->dtrq_xid,
2505                                        dtrq->dtrq_master_transno);
2506
2507                                 /* Add it to the replay finish list */
2508                                 spin_lock(&tdtd->tdtd_replay_list_lock);
2509                                 list_add(&dtrq->dtrq_list,
2510                                          &tdtd->tdtd_replay_finish_list);
2511                                 spin_unlock(&tdtd->tdtd_replay_list_lock);
2512
2513                                 spin_lock(&obd->obd_recovery_task_lock);
2514                                 if (transno == obd->obd_next_recovery_transno)
2515                                         obd->obd_next_recovery_transno++;
2516                                 else if (transno >
2517                                          obd->obd_next_recovery_transno)
2518                                         obd->obd_next_recovery_transno =
2519                                                                 transno + 1;
2520                                 spin_unlock(&obd->obd_recovery_task_lock);
2521                         } else {
2522                                 dtrq_destroy(dtrq);
2523                         }
2524                 } else {
2525                         spin_unlock(&obd->obd_recovery_task_lock);
2526 abort:
2527                         LASSERT(list_empty(&obd->obd_req_replay_queue));
2528                         LASSERT(atomic_read(&obd->obd_req_replay_clients) == 0);
2529                         /** evict exports failed VBR */
2530                         class_disconnect_stale_exports(obd, exp_vbr_healthy);
2531                         break;
2532                 }
2533         } while (1);
2534 }
2535
2536 static int target_recovery_thread(void *arg)
2537 {
2538         struct lu_target *lut = arg;
2539         struct obd_device *obd = lut->lut_obd;
2540         struct ptlrpc_request *req;
2541         struct target_recovery_data *trd = &obd->obd_recovery_data;
2542         unsigned long delta;
2543         struct lu_env *env;
2544         struct ptlrpc_thread *thread = NULL;
2545         int rc = 0;
2546         ENTRY;
2547
2548         unshare_fs_struct();
2549         OBD_ALLOC_PTR(thread);
2550         if (thread == NULL)
2551                 RETURN(-ENOMEM);
2552
2553         OBD_ALLOC_PTR(env);
2554         if (env == NULL)
2555                 GOTO(out_thread, rc = -ENOMEM);
2556         rc = lu_env_add(env);
2557         if (rc)
2558                 GOTO(out_env, rc);
2559
2560         rc = lu_context_init(&env->le_ctx, LCT_MD_THREAD | LCT_DT_THREAD);
2561         if (rc)
2562                 GOTO(out_env_remove, rc);
2563
2564         thread->t_env = env;
2565         thread->t_id = -1; /* force filter_iobuf_get/put to use local buffers */
2566         env->le_ctx.lc_thread = thread;
2567         tgt_io_thread_init(thread); /* init thread_big_cache for IO requests */
2568         thread->t_watchdog = NULL;
2569
2570         CDEBUG(D_HA, "%s: started recovery thread pid %d\n", obd->obd_name,
2571                current_pid());
2572         trd->trd_processing_task = current_pid();
2573
2574         spin_lock(&obd->obd_dev_lock);
2575         obd->obd_recovering = 1;
2576         spin_unlock(&obd->obd_dev_lock);
2577         complete(&trd->trd_starting);
2578
2579         /* first of all, we have to know the first transno to replay */
2580         if (target_recovery_overseer(lut, check_for_recovery_ready,
2581                                      exp_connect_healthy)) {
2582                 abort_req_replay_queue(obd);
2583                 abort_lock_replay_queue(obd);
2584                 if (lut->lut_tdtd != NULL)
2585                         dtrq_list_destroy(lut->lut_tdtd);
2586         }
2587
2588         /* next stage: replay requests or update */
2589         delta = jiffies;
2590         CDEBUG(D_INFO, "1: request replay stage - %d clients from t%llu\n",
2591                atomic_read(&obd->obd_req_replay_clients),
2592                obd->obd_next_recovery_transno);
2593         replay_request_or_update(env, lut, trd, thread);
2594
2595         /**
2596          * The second stage: replay locks
2597          */
2598         CDEBUG(D_INFO, "2: lock replay stage - %d clients\n",
2599                atomic_read(&obd->obd_lock_replay_clients));
2600         while ((req = target_next_replay_lock(lut))) {
2601                 LASSERT(trd->trd_processing_task == current_pid());
2602                 DEBUG_REQ(D_HA, req, "processing lock from %s: ",
2603                           libcfs_nid2str(req->rq_peer.nid));
2604                 if (OBD_FAIL_CHECK(OBD_FAIL_LDLM_LOCK_REPLAY)) {
2605                         req->rq_status = -ENODEV;
2606                         target_request_copy_put(req);
2607                         continue;
2608                 }
2609                 handle_recovery_req(thread, req,
2610                                     trd->trd_recovery_handler);
2611                 target_request_copy_put(req);
2612                 obd->obd_replayed_locks++;
2613         }
2614
2615         /**
2616          * The third stage: reply on final pings, at this moment all clients
2617          * must have request in final queue
2618          */
2619         CFS_FAIL_TIMEOUT(OBD_FAIL_TGT_REPLAY_RECONNECT, cfs_fail_val);
2620         CDEBUG(D_INFO, "3: final stage - process recovery completion pings\n");
2621         /** Update server last boot epoch */
2622         tgt_boot_epoch_update(lut);
2623         /* We drop recoverying flag to forward all new requests
2624          * to regular mds_handle() since now */
2625         spin_lock(&obd->obd_dev_lock);
2626         obd->obd_recovering = obd->obd_abort_recovery = 0;
2627         spin_unlock(&obd->obd_dev_lock);
2628         spin_lock(&obd->obd_recovery_task_lock);
2629         target_cancel_recovery_timer(obd);
2630         spin_unlock(&obd->obd_recovery_task_lock);
2631         while ((req = target_next_final_ping(obd))) {
2632                 LASSERT(trd->trd_processing_task == current_pid());
2633                 DEBUG_REQ(D_HA, req, "processing final ping from %s: ",
2634                           libcfs_nid2str(req->rq_peer.nid));
2635                 handle_recovery_req(thread, req,
2636                                     trd->trd_recovery_handler);
2637                 /* Because the waiting client can not send ping to server,
2638                  * so we need refresh the last_request_time, to avoid the
2639                  * export is being evicted */
2640                 ptlrpc_update_export_timer(req->rq_export, 0);
2641                 target_request_copy_put(req);
2642         }
2643
2644         delta = jiffies_to_msecs(jiffies - delta) / MSEC_PER_SEC;
2645         CDEBUG(D_INFO,"4: recovery completed in %lus - %d/%d reqs/locks\n",
2646               delta, obd->obd_replayed_requests, obd->obd_replayed_locks);
2647         if (delta > OBD_RECOVERY_TIME_SOFT) {
2648                 CWARN("too long recovery - read logs\n");
2649                 libcfs_debug_dumplog();
2650         }
2651
2652         target_finish_recovery(lut);
2653
2654         lu_context_fini(&env->le_ctx);
2655         trd->trd_processing_task = 0;
2656         complete(&trd->trd_finishing);
2657
2658         tgt_io_thread_done(thread);
2659 out_env_remove:
2660         lu_env_remove(env);
2661 out_env:
2662         OBD_FREE_PTR(env);
2663 out_thread:
2664         OBD_FREE_PTR(thread);
2665         RETURN(rc);
2666 }
2667
2668 static int target_start_recovery_thread(struct lu_target *lut,
2669                                         svc_handler_t handler)
2670 {
2671         struct obd_device *obd = lut->lut_obd;
2672         int rc = 0;
2673         struct target_recovery_data *trd = &obd->obd_recovery_data;
2674         int index;
2675
2676         memset(trd, 0, sizeof(*trd));
2677         init_completion(&trd->trd_starting);
2678         init_completion(&trd->trd_finishing);
2679         trd->trd_recovery_handler = handler;
2680
2681         rc = server_name2index(obd->obd_name, &index, NULL);
2682         if (rc < 0)
2683                 return rc;
2684
2685         if (!IS_ERR(kthread_run(target_recovery_thread,
2686                                 lut, "tgt_recover_%d", index))) {
2687                 wait_for_completion(&trd->trd_starting);
2688                 LASSERT(obd->obd_recovering != 0);
2689         } else {
2690                 rc = -ECHILD;
2691         }
2692
2693         return rc;
2694 }
2695
2696 void target_stop_recovery_thread(struct obd_device *obd)
2697 {
2698         if (obd->obd_recovery_data.trd_processing_task > 0) {
2699                 struct target_recovery_data *trd = &obd->obd_recovery_data;
2700                 /** recovery can be done but postrecovery is not yet */
2701                 spin_lock(&obd->obd_dev_lock);
2702                 if (obd->obd_recovering) {
2703                         CERROR("%s: Aborting recovery\n", obd->obd_name);
2704                         obd->obd_abort_recovery = 1;
2705                         wake_up(&obd->obd_next_transno_waitq);
2706                 }
2707                 spin_unlock(&obd->obd_dev_lock);
2708                 wait_for_completion(&trd->trd_finishing);
2709         }
2710 }
2711 EXPORT_SYMBOL(target_stop_recovery_thread);
2712
2713 void target_recovery_fini(struct obd_device *obd)
2714 {
2715         class_disconnect_exports(obd);
2716         target_stop_recovery_thread(obd);
2717         target_cleanup_recovery(obd);
2718 }
2719 EXPORT_SYMBOL(target_recovery_fini);
2720
2721 static enum hrtimer_restart target_recovery_expired(struct hrtimer *timer)
2722 {
2723         struct obd_device *obd = container_of(timer, struct obd_device,
2724                                               obd_recovery_timer);
2725
2726         CDEBUG(D_HA,
2727                "%s: recovery timed out; %d clients are still in recovery after %llu seconds (%d clients connected)\n",
2728                obd->obd_name, atomic_read(&obd->obd_lock_replay_clients),
2729                ktime_get_real_seconds() - obd->obd_recovery_start,
2730                atomic_read(&obd->obd_connected_clients));
2731
2732         obd->obd_recovery_expired = 1;
2733         wake_up(&obd->obd_next_transno_waitq);
2734         return HRTIMER_NORESTART;
2735 }
2736
2737 void target_recovery_init(struct lu_target *lut, svc_handler_t handler)
2738 {
2739         struct obd_device *obd = lut->lut_obd;
2740
2741         if (lut->lut_bottom->dd_rdonly)
2742                 return;
2743
2744         if (atomic_read(&obd->obd_max_recoverable_clients) == 0) {
2745                 /** Update server last boot epoch */
2746                 tgt_boot_epoch_update(lut);
2747                 return;
2748         }
2749
2750         CDEBUG(D_HA, "RECOVERY: service %s, %d recoverable clients, "
2751                "last_transno %llu\n", obd->obd_name,
2752                atomic_read(&obd->obd_max_recoverable_clients),
2753                obd->obd_last_committed);
2754         LASSERT(obd->obd_stopping == 0);
2755         obd->obd_next_recovery_transno = obd->obd_last_committed + 1;
2756         obd->obd_recovery_start = 0;
2757         obd->obd_recovery_end = 0;
2758
2759         hrtimer_init(&obd->obd_recovery_timer, CLOCK_MONOTONIC,
2760                      HRTIMER_MODE_ABS);
2761         obd->obd_recovery_timer.function = &target_recovery_expired;
2762         target_start_recovery_thread(lut, handler);
2763 }
2764 EXPORT_SYMBOL(target_recovery_init);
2765
2766 static int target_process_req_flags(struct obd_device *obd,
2767                                     struct ptlrpc_request *req)
2768 {
2769         struct obd_export *exp = req->rq_export;
2770         LASSERT(exp != NULL);
2771         if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REQ_REPLAY_DONE) {
2772                 /* client declares he's ready to replay locks */
2773                 spin_lock(&exp->exp_lock);
2774                 if (exp->exp_req_replay_needed) {
2775                         exp->exp_req_replay_needed = 0;
2776                         spin_unlock(&exp->exp_lock);
2777
2778                         LASSERT_ATOMIC_POS(&obd->obd_req_replay_clients);
2779                         atomic_dec(&obd->obd_req_replay_clients);
2780                 } else {
2781                         spin_unlock(&exp->exp_lock);
2782                 }
2783         }
2784         if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_LOCK_REPLAY_DONE) {
2785                 /* client declares he's ready to complete recovery
2786                  * so, we put the request on th final queue */
2787                 spin_lock(&exp->exp_lock);
2788                 if (exp->exp_lock_replay_needed) {
2789                         exp->exp_lock_replay_needed = 0;
2790                         spin_unlock(&exp->exp_lock);
2791
2792                         LASSERT_ATOMIC_POS(&obd->obd_lock_replay_clients);
2793                         atomic_dec(&obd->obd_lock_replay_clients);
2794                 } else {
2795                         spin_unlock(&exp->exp_lock);
2796                 }
2797         }
2798         return 0;
2799 }
2800
2801 int target_queue_recovery_request(struct ptlrpc_request *req,
2802                                   struct obd_device *obd)
2803 {
2804         __u64 transno = lustre_msg_get_transno(req->rq_reqmsg);
2805         struct ptlrpc_request *reqiter;
2806         int inserted = 0;
2807         ENTRY;
2808
2809         if (obd->obd_recovery_data.trd_processing_task == current_pid()) {
2810                 /* Processing the queue right now, don't re-add. */
2811                 RETURN(1);
2812         }
2813
2814         target_process_req_flags(obd, req);
2815
2816         if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_LOCK_REPLAY_DONE) {
2817                 if (unlikely(OBD_FAIL_CHECK(OBD_FAIL_TGT_RECOVERY_REQ_RACE))) {
2818                         if (cfs_fail_val == 1) {
2819                                 cfs_race_state = 1;
2820                                 cfs_fail_val = 0;
2821                                 wake_up(&cfs_race_waitq);
2822
2823                                 set_current_state(TASK_INTERRUPTIBLE);
2824                                 schedule_timeout(cfs_time_seconds(1));
2825                         }
2826                 }
2827
2828                 /* client declares he's ready to complete recovery
2829                  * so, we put the request on th final queue */
2830                 target_request_copy_get(req);
2831                 DEBUG_REQ(D_HA, req, "queue final req");
2832                 wake_up(&obd->obd_next_transno_waitq);
2833                 spin_lock(&obd->obd_recovery_task_lock);
2834                 if (obd->obd_recovering) {
2835                         struct ptlrpc_request *tmp;
2836                         struct ptlrpc_request *duplicate = NULL;
2837
2838                         if (likely(!req->rq_export->exp_replay_done)) {
2839                                 req->rq_export->exp_replay_done = 1;
2840                                 list_add_tail(&req->rq_list,
2841                                               &obd->obd_final_req_queue);
2842                                 spin_unlock(&obd->obd_recovery_task_lock);
2843                                 RETURN(0);
2844                         }
2845
2846                         /* XXX O(n), but only happens if final ping is
2847                          * timed out, probably reorganize the list as
2848                          * a hash list later */
2849                         list_for_each_entry_safe(reqiter, tmp,
2850                                                  &obd->obd_final_req_queue,
2851                                                  rq_list) {
2852                                 if (reqiter->rq_export == req->rq_export) {
2853                                         list_del_init(&reqiter->rq_list);
2854                                         duplicate = reqiter;
2855                                         break;
2856                                 }
2857                         }
2858
2859                         list_add_tail(&req->rq_list,
2860                                       &obd->obd_final_req_queue);
2861                         req->rq_export->exp_replay_done = 1;
2862                         spin_unlock(&obd->obd_recovery_task_lock);
2863
2864                         if (duplicate != NULL) {
2865                                 DEBUG_REQ(D_HA, duplicate,
2866                                           "put prev final req\n");
2867                                 target_request_copy_put(duplicate);
2868                         }
2869                         RETURN(0);
2870                 } else {
2871                         spin_unlock(&obd->obd_recovery_task_lock);
2872                         target_request_copy_put(req);
2873                         RETURN(obd->obd_stopping ? -ENOTCONN : 1);
2874                 }
2875         }
2876         if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REQ_REPLAY_DONE) {
2877                 /* client declares he's ready to replay locks */
2878                 target_request_copy_get(req);
2879                 DEBUG_REQ(D_HA, req, "queue lock replay req");
2880                 wake_up(&obd->obd_next_transno_waitq);
2881                 spin_lock(&obd->obd_recovery_task_lock);
2882                 LASSERT(obd->obd_recovering);
2883                 /* usually due to recovery abort */
2884                 if (!req->rq_export->exp_in_recovery) {
2885                         spin_unlock(&obd->obd_recovery_task_lock);
2886                         target_request_copy_put(req);
2887                         RETURN(-ENOTCONN);
2888                 }
2889                 LASSERT(req->rq_export->exp_lock_replay_needed);
2890                 list_add_tail(&req->rq_list, &obd->obd_lock_replay_queue);
2891                 spin_unlock(&obd->obd_recovery_task_lock);
2892                 RETURN(0);
2893         }
2894
2895         /* CAVEAT EMPTOR: The incoming request message has been swabbed
2896          * (i.e. buflens etc are in my own byte order), but type-dependent
2897          * buffers (eg mdt_body, ost_body etc) have NOT been swabbed. */
2898
2899         if (!transno) {
2900                 INIT_LIST_HEAD(&req->rq_list);
2901                 DEBUG_REQ(D_HA, req, "not queueing");
2902                 RETURN(1);
2903         }
2904
2905         /* If we're processing the queue, we want don't want to queue this
2906          * message.
2907          *
2908          * Also, if this request has a transno less than the one we're waiting
2909          * for, we should process it now.  It could (and currently always will)
2910          * be an open request for a descriptor that was opened some time ago.
2911          *
2912          * Also, a resent, replayed request that has already been
2913          * handled will pass through here and be processed immediately.
2914          */
2915         CDEBUG(D_HA, "Next recovery transno: %llu"
2916                ", current: %llu, replaying\n",
2917                obd->obd_next_recovery_transno, transno);
2918
2919         /* If the request has been replayed by update replay, then sends this
2920          * request to the recovery thread (replay_request_or_update()), where
2921          * it will be handled */
2922         spin_lock(&obd->obd_recovery_task_lock);
2923         if (transno < obd->obd_next_recovery_transno &&
2924             !is_req_replayed_by_update(req)) {
2925                 /* Processing the queue right now, don't re-add. */
2926                 LASSERT(list_empty(&req->rq_list));
2927                 spin_unlock(&obd->obd_recovery_task_lock);
2928                 RETURN(1);
2929         }
2930         spin_unlock(&obd->obd_recovery_task_lock);
2931
2932         if (OBD_FAIL_CHECK(OBD_FAIL_TGT_REPLAY_DROP))
2933                 RETURN(0);
2934
2935         target_request_copy_get(req);
2936         if (!req->rq_export->exp_in_recovery) {
2937                 target_request_copy_put(req);
2938                 RETURN(-ENOTCONN);
2939         }
2940         LASSERT(req->rq_export->exp_req_replay_needed);
2941
2942         if (target_exp_enqueue_req_replay(req)) {
2943                 DEBUG_REQ(D_ERROR, req, "dropping resent queued req");
2944                 target_request_copy_put(req);
2945                 RETURN(0);
2946         }
2947
2948         /* XXX O(n^2) */
2949         spin_lock(&obd->obd_recovery_task_lock);
2950         LASSERT(obd->obd_recovering);
2951         list_for_each_entry(reqiter, &obd->obd_req_replay_queue, rq_list) {
2952                 if (lustre_msg_get_transno(reqiter->rq_reqmsg) > transno) {
2953                         list_add_tail(&req->rq_list, &reqiter->rq_list);
2954                         inserted = 1;
2955                         goto added;
2956                 }
2957
2958                 if (unlikely(lustre_msg_get_transno(reqiter->rq_reqmsg) ==
2959                              transno)) {
2960                         DEBUG_REQ(D_ERROR, req, "dropping replay: transno "
2961                                   "has been claimed by another client");
2962                         spin_unlock(&obd->obd_recovery_task_lock);
2963                         target_exp_dequeue_req_replay(req);
2964                         target_request_copy_put(req);
2965                         RETURN(0);
2966                 }
2967         }
2968 added:
2969         if (!inserted)
2970                 list_add_tail(&req->rq_list, &obd->obd_req_replay_queue);
2971
2972         obd->obd_requests_queued_for_recovery++;
2973         spin_unlock(&obd->obd_recovery_task_lock);
2974         wake_up(&obd->obd_next_transno_waitq);
2975         RETURN(0);
2976 }
2977
2978 void target_committed_to_req(struct ptlrpc_request *req)
2979 {
2980         struct obd_export *exp = req->rq_export;
2981
2982         if (!exp->exp_obd->obd_no_transno && req->rq_repmsg != NULL)
2983                 lustre_msg_set_last_committed(req->rq_repmsg,
2984                                               exp->exp_last_committed);
2985         else
2986                 DEBUG_REQ(D_IOCTL, req, "not sending last_committed update (%d/"
2987                           "%d)", exp->exp_obd->obd_no_transno,
2988                           req->rq_repmsg == NULL);
2989
2990         CDEBUG(D_INFO, "last_committed %llu, transno %llu, xid %llu\n",
2991                exp->exp_last_committed, req->rq_transno, req->rq_xid);
2992 }
2993
2994 #endif /* HAVE_SERVER_SUPPORT */
2995
2996 /**
2997  * Packs current SLV and Limit into \a req.
2998  */
2999 int target_pack_pool_reply(struct ptlrpc_request *req)
3000 {
3001         struct obd_device *obd;
3002         ENTRY;
3003
3004         /* Check that we still have all structures alive as this may
3005          * be some late RPC at shutdown time. */
3006         if (unlikely(!req->rq_export || !req->rq_export->exp_obd ||
3007                      !exp_connect_lru_resize(req->rq_export))) {
3008                 lustre_msg_set_slv(req->rq_repmsg, 0);
3009                 lustre_msg_set_limit(req->rq_repmsg, 0);
3010                 RETURN(0);
3011         }
3012
3013         /* OBD is alive here as export is alive, which we checked above. */
3014         obd = req->rq_export->exp_obd;
3015
3016         read_lock(&obd->obd_pool_lock);
3017         lustre_msg_set_slv(req->rq_repmsg, obd->obd_pool_slv);
3018         lustre_msg_set_limit(req->rq_repmsg, obd->obd_pool_limit);
3019         read_unlock(&obd->obd_pool_lock);
3020
3021         RETURN(0);
3022 }
3023
3024 static int target_send_reply_msg(struct ptlrpc_request *req,
3025                                  int rc, int fail_id)
3026 {
3027         if (OBD_FAIL_CHECK_ORSET(fail_id & ~OBD_FAIL_ONCE, OBD_FAIL_ONCE)) {
3028                 DEBUG_REQ(D_ERROR, req, "dropping reply");
3029                 return -ECOMM;
3030         }
3031         /* We can have a null rq_reqmsg in the event of bad signature or
3032          * no context when unwrapping */
3033         if (req->rq_reqmsg &&
3034             unlikely(lustre_msg_get_opc(req->rq_reqmsg) == MDS_REINT &&
3035             OBD_FAIL_CHECK(OBD_FAIL_MDS_REINT_MULTI_NET_REP)))
3036                 return -ECOMM;
3037
3038         if (unlikely(rc)) {
3039                 DEBUG_REQ(D_NET, req, "processing error (%d)", rc);
3040                 req->rq_status = rc;
3041                 return ptlrpc_send_error(req, 1);
3042         } else {
3043                 DEBUG_REQ(D_NET, req, "sending reply");
3044         }
3045
3046         return ptlrpc_send_reply(req, PTLRPC_REPLY_MAYBE_DIFFICULT);
3047 }
3048
3049 void target_send_reply(struct ptlrpc_request *req, int rc, int fail_id)
3050 {
3051         struct ptlrpc_service_part *svcpt;
3052         int                        netrc;
3053         struct ptlrpc_reply_state *rs;
3054         struct obd_export         *exp;
3055         ENTRY;
3056
3057         if (req->rq_no_reply) {
3058                 EXIT;
3059                 return;
3060         }
3061
3062         svcpt = req->rq_rqbd->rqbd_svcpt;
3063         rs = req->rq_reply_state;
3064         if (rs == NULL || !rs->rs_difficult) {
3065                 /* no notifiers */
3066                 target_send_reply_msg (req, rc, fail_id);
3067                 EXIT;
3068                 return;
3069         }
3070
3071         /* must be an export if locks saved */
3072         LASSERT(req->rq_export != NULL);
3073         /* req/reply consistent */
3074         LASSERT(rs->rs_svcpt == svcpt);
3075
3076         /* "fresh" reply */
3077         LASSERT(!rs->rs_scheduled);
3078         LASSERT(!rs->rs_scheduled_ever);
3079         LASSERT(!rs->rs_handled);
3080         LASSERT(!rs->rs_on_net);
3081         LASSERT(rs->rs_export == NULL);
3082         LASSERT(list_empty(&rs->rs_obd_list));
3083         LASSERT(list_empty(&rs->rs_exp_list));
3084
3085         exp = class_export_get(req->rq_export);
3086
3087         /* disable reply scheduling while I'm setting up */
3088         rs->rs_scheduled = 1;
3089         rs->rs_on_net    = 1;
3090         rs->rs_xid       = req->rq_xid;
3091         rs->rs_transno   = req->rq_transno;
3092         rs->rs_export    = exp;
3093         rs->rs_opc       = lustre_msg_get_opc(req->rq_reqmsg);
3094
3095         spin_lock(&exp->exp_uncommitted_replies_lock);
3096         CDEBUG(D_NET, "rs transno = %llu, last committed = %llu\n",
3097                rs->rs_transno, exp->exp_last_committed);
3098         if (rs->rs_transno > exp->exp_last_committed) {
3099                 /* not committed already */
3100                 list_add_tail(&rs->rs_obd_list,
3101                                   &exp->exp_uncommitted_replies);
3102         }
3103         spin_unlock(&exp->exp_uncommitted_replies_lock);
3104
3105         spin_lock(&exp->exp_lock);
3106         list_add_tail(&rs->rs_exp_list, &exp->exp_outstanding_replies);
3107         spin_unlock(&exp->exp_lock);
3108
3109         netrc = target_send_reply_msg(req, rc, fail_id);
3110
3111         spin_lock(&svcpt->scp_rep_lock);
3112
3113         atomic_inc(&svcpt->scp_nreps_difficult);
3114
3115         if (netrc != 0) {
3116                 /* error sending: reply is off the net.  Also we need +1
3117                  * reply ref until ptlrpc_handle_rs() is done
3118                  * with the reply state (if the send was successful, there
3119                  * would have been +1 ref for the net, which
3120                  * reply_out_callback leaves alone) */
3121                 rs->rs_on_net = 0;
3122                 ptlrpc_rs_addref(rs);
3123         }
3124
3125         spin_lock(&rs->rs_lock);
3126         if (rs->rs_transno <= exp->exp_last_committed ||
3127             (!rs->rs_on_net && !rs->rs_no_ack) ||
3128             list_empty(&rs->rs_exp_list) ||     /* completed already */
3129             list_empty(&rs->rs_obd_list)) {
3130                 CDEBUG(D_HA, "Schedule reply immediately\n");
3131                 ptlrpc_dispatch_difficult_reply(rs);
3132         } else {
3133                 list_add(&rs->rs_list, &svcpt->scp_rep_active);
3134                 rs->rs_scheduled = 0;   /* allow notifier to schedule */
3135         }
3136         spin_unlock(&rs->rs_lock);
3137         spin_unlock(&svcpt->scp_rep_lock);
3138         EXIT;
3139 }
3140
3141 enum ldlm_mode lck_compat_array[] = {
3142         [LCK_EX]    = LCK_COMPAT_EX,
3143         [LCK_PW]    = LCK_COMPAT_PW,
3144         [LCK_PR]    = LCK_COMPAT_PR,
3145         [LCK_CW]    = LCK_COMPAT_CW,
3146         [LCK_CR]    = LCK_COMPAT_CR,
3147         [LCK_NL]    = LCK_COMPAT_NL,
3148         [LCK_GROUP] = LCK_COMPAT_GROUP,
3149         [LCK_COS]   = LCK_COMPAT_COS,
3150 };
3151
3152 /**
3153  * Rather arbitrary mapping from LDLM error codes to errno values. This should
3154  * not escape to the user level.
3155  */
3156 int ldlm_error2errno(enum ldlm_error error)
3157 {
3158         int result;
3159
3160         switch (error) {
3161         case ELDLM_OK:
3162         case ELDLM_LOCK_MATCHED:
3163                 result = 0;
3164                 break;
3165         case ELDLM_LOCK_CHANGED:
3166                 result = -ESTALE;
3167                 break;
3168         case ELDLM_LOCK_ABORTED:
3169                 result = -ENAVAIL;
3170                 break;
3171         case ELDLM_LOCK_REPLACED:
3172                 result = -ESRCH;
3173                 break;
3174         case ELDLM_NO_LOCK_DATA:
3175                 result = -ENOENT;
3176                 break;
3177         case ELDLM_NAMESPACE_EXISTS:
3178                 result = -EEXIST;
3179                 break;
3180         case ELDLM_BAD_NAMESPACE:
3181                 result = -EBADF;
3182                 break;
3183         default:
3184                 if (((int)error) < 0) { /* cast to signed type */
3185                         result = error; /* as ldlm_error can be unsigned */
3186                 } else {
3187                         CERROR("Invalid DLM result code: %d\n", error);
3188                         result = -EPROTO;
3189                 }
3190         }
3191         return result;
3192 }
3193 EXPORT_SYMBOL(ldlm_error2errno);
3194
3195 /**
3196  * Dual to ldlm_error2errno(): maps errno values back to enum ldlm_error.
3197  */
3198 enum ldlm_error ldlm_errno2error(int err_no)
3199 {
3200         int error;
3201
3202         switch (err_no) {
3203         case 0:
3204                 error = ELDLM_OK;
3205                 break;
3206         case -ESTALE:
3207                 error = ELDLM_LOCK_CHANGED;
3208                 break;
3209         case -ENAVAIL:
3210                 error = ELDLM_LOCK_ABORTED;
3211                 break;
3212         case -ESRCH:
3213                 error = ELDLM_LOCK_REPLACED;
3214                 break;
3215         case -ENOENT:
3216                 error = ELDLM_NO_LOCK_DATA;
3217                 break;
3218         case -EEXIST:
3219                 error = ELDLM_NAMESPACE_EXISTS;
3220                 break;
3221         case -EBADF:
3222                 error = ELDLM_BAD_NAMESPACE;
3223                 break;
3224         default:
3225                 error = err_no;
3226         }
3227         return error;
3228 }
3229
3230 #if LUSTRE_TRACKS_LOCK_EXP_REFS
3231 void ldlm_dump_export_locks(struct obd_export *exp)
3232 {
3233         spin_lock(&exp->exp_locks_list_guard);
3234         if (!list_empty(&exp->exp_locks_list)) {
3235                 struct ldlm_lock *lock;
3236
3237                 CERROR("dumping locks for export %p,"
3238                        "ignore if the unmount doesn't hang\n", exp);
3239                 list_for_each_entry(lock, &exp->exp_locks_list,
3240                                         l_exp_refs_link)
3241                         LDLM_ERROR(lock, "lock:");
3242         }
3243         spin_unlock(&exp->exp_locks_list_guard);
3244 }
3245 #endif
3246
3247 #ifdef HAVE_SERVER_SUPPORT
3248 static int target_bulk_timeout(void *data)
3249 {
3250         ENTRY;
3251         /* We don't fail the connection here, because having the export
3252          * killed makes the (vital) call to commitrw very sad.
3253          */
3254         RETURN(1);
3255 }
3256
3257 static inline const char *bulk2type(struct ptlrpc_request *req)
3258 {
3259         if (req->rq_bulk_read)
3260                 return "READ";
3261         if (req->rq_bulk_write)
3262                 return "WRITE";
3263         return "UNKNOWN";
3264 }
3265
3266 int target_bulk_io(struct obd_export *exp, struct ptlrpc_bulk_desc *desc,
3267                    struct l_wait_info *lwi)
3268 {
3269         struct ptlrpc_request *req = desc->bd_req;
3270         time64_t start = ktime_get_seconds();
3271         time64_t deadline;
3272         int rc = 0;
3273
3274         ENTRY;
3275
3276         /* If there is eviction in progress, wait for it to finish. */
3277         if (unlikely(atomic_read(&exp->exp_obd->obd_evict_inprogress))) {
3278                 *lwi = LWI_INTR(NULL, NULL);
3279                 rc = l_wait_event(exp->exp_obd->obd_evict_inprogress_waitq,
3280                                   !atomic_read(&exp->exp_obd->
3281                                                    obd_evict_inprogress),
3282                                   lwi);
3283         }
3284
3285         /* Check if client was evicted or reconnected already. */
3286         if (exp->exp_failed ||
3287             exp->exp_conn_cnt > lustre_msg_get_conn_cnt(req->rq_reqmsg)) {
3288                 rc = -ENOTCONN;
3289         } else {
3290                 if (req->rq_bulk_read)
3291                         rc = sptlrpc_svc_wrap_bulk(req, desc);
3292
3293                 if (OCD_HAS_FLAG(&exp->exp_connect_data, BULK_MBITS))
3294                         req->rq_mbits = lustre_msg_get_mbits(req->rq_reqmsg);
3295                 else /* old version, bulk matchbits is rq_xid */
3296                         req->rq_mbits = req->rq_xid;
3297
3298                 if (rc == 0)
3299                         rc = ptlrpc_start_bulk_transfer(desc);
3300         }
3301
3302         if (rc < 0) {
3303                 DEBUG_REQ(D_ERROR, req, "bulk %s failed: rc %d",
3304                           bulk2type(req), rc);
3305                 RETURN(rc);
3306         }
3307
3308         if (OBD_FAIL_CHECK(OBD_FAIL_MDS_SENDPAGE)) {
3309                 ptlrpc_abort_bulk(desc);
3310                 RETURN(0);
3311         }
3312
3313         /* limit actual bulk transfer to bulk_timeout seconds */
3314         deadline = start + bulk_timeout;
3315         if (deadline > req->rq_deadline)
3316                 deadline = req->rq_deadline;
3317
3318         do {
3319                 time64_t timeoutl = deadline - ktime_get_seconds();
3320                 long timeout_jiffies = timeoutl <= 0 ?
3321                                        1 : cfs_time_seconds(timeoutl);
3322                 time64_t rq_deadline;
3323
3324                 *lwi = LWI_TIMEOUT_INTERVAL(timeout_jiffies,
3325                                             cfs_time_seconds(1),
3326                                             target_bulk_timeout, desc);
3327                 rc = l_wait_event(desc->bd_waitq,
3328                                   !ptlrpc_server_bulk_active(desc) ||
3329                                   exp->exp_failed ||
3330                                   exp->exp_conn_cnt >
3331                                   lustre_msg_get_conn_cnt(req->rq_reqmsg),
3332                                   lwi);
3333                 LASSERT(rc == 0 || rc == -ETIMEDOUT);
3334                 /* Wait again if we changed rq_deadline. */
3335                 rq_deadline = READ_ONCE(req->rq_deadline);
3336                 deadline = start + bulk_timeout;
3337                 if (deadline > rq_deadline)
3338                         deadline = rq_deadline;
3339         } while (rc == -ETIMEDOUT &&
3340                  deadline > ktime_get_seconds());
3341
3342         if (rc == -ETIMEDOUT) {
3343                 DEBUG_REQ(D_ERROR, req, "timeout on bulk %s after %lld%+llds",
3344                           bulk2type(req), deadline - start,
3345                           ktime_get_real_seconds() - deadline);
3346                 ptlrpc_abort_bulk(desc);
3347         } else if (exp->exp_failed) {
3348                 DEBUG_REQ(D_ERROR, req, "Eviction on bulk %s",
3349                           bulk2type(req));
3350                 rc = -ENOTCONN;
3351                 ptlrpc_abort_bulk(desc);
3352         } else if (exp->exp_conn_cnt >
3353                    lustre_msg_get_conn_cnt(req->rq_reqmsg)) {
3354                 DEBUG_REQ(D_ERROR, req, "Reconnect on bulk %s",
3355                           bulk2type(req));
3356                 /* We don't reply anyway. */
3357                 rc = -ETIMEDOUT;
3358                 ptlrpc_abort_bulk(desc);
3359         } else if (desc->bd_failure) {
3360                 DEBUG_REQ(D_ERROR, req, "network error on bulk %s",
3361                           bulk2type(req));
3362                 /* XXX should this be a different errno? */
3363                 rc = -ETIMEDOUT;
3364         } else {
3365                 if (req->rq_bulk_write)
3366                         rc = sptlrpc_svc_unwrap_bulk(req, desc);
3367                 if (rc == 0 && desc->bd_nob_transferred != desc->bd_nob) {
3368                         DEBUG_REQ(D_ERROR, req, "truncated bulk %s %d(%d)",
3369                                   bulk2type(req), desc->bd_nob_transferred,
3370                                   desc->bd_nob);
3371                         /* XXX should this be a different errno? */
3372                         rc = -ETIMEDOUT;
3373                 }
3374         }
3375
3376         RETURN(rc);
3377 }
3378 EXPORT_SYMBOL(target_bulk_io);
3379
3380 #endif /* HAVE_SERVER_SUPPORT */