lustre/ldlm/ldlm_lib.c

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