4 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
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.
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).
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.sun.com/software/products/lustre/docs/GPLv2.pdf
20 * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
21 * CA 95054 USA or visit www.sun.com if you need additional information or
27 * Copyright (c) 2003, 2010, Oracle and/or its affiliates. All rights reserved.
28 * Use is subject to license terms.
30 * Copyright (c) 2010, 2014, Intel Corporation.
33 * This file is part of Lustre, http://www.lustre.org/
34 * Lustre is a trademark of Sun Microsystems, Inc.
38 * This file deals with various client/target related logic including recovery.
40 * TODO: This code more logically belongs in the ptlrpc module than in ldlm and
44 #define DEBUG_SUBSYSTEM S_LDLM
46 #include <libcfs/libcfs.h>
48 #include <obd_class.h>
49 #include <lustre_dlm.h>
50 #include <lustre_net.h>
51 #include <lustre_sec.h>
52 #include "ldlm_internal.h"
54 /* @priority: If non-zero, move the selected connection to the list head.
55 * @create: If zero, only search in existing connections.
57 static int import_set_conn(struct obd_import *imp, struct obd_uuid *uuid,
58 int priority, int create)
60 struct ptlrpc_connection *ptlrpc_conn;
61 struct obd_import_conn *imp_conn = NULL, *item;
65 if (!create && !priority) {
66 CDEBUG(D_HA, "Nothing to do\n");
70 ptlrpc_conn = ptlrpc_uuid_to_connection(uuid);
72 CDEBUG(D_HA, "can't find connection %s\n", uuid->uuid);
77 OBD_ALLOC(imp_conn, sizeof(*imp_conn));
79 GOTO(out_put, rc = -ENOMEM);
83 spin_lock(&imp->imp_lock);
84 list_for_each_entry(item, &imp->imp_conn_list, oic_item) {
85 if (obd_uuid_equals(uuid, &item->oic_uuid)) {
87 list_del(&item->oic_item);
88 list_add(&item->oic_item,
90 item->oic_last_attempt = 0;
92 CDEBUG(D_HA, "imp %p@%s: found existing conn %s%s\n",
93 imp, imp->imp_obd->obd_name, uuid->uuid,
94 (priority ? ", moved to head" : ""));
95 spin_unlock(&imp->imp_lock);
96 GOTO(out_free, rc = 0);
99 /* No existing import connection found for \a uuid. */
101 imp_conn->oic_conn = ptlrpc_conn;
102 imp_conn->oic_uuid = *uuid;
103 imp_conn->oic_last_attempt = 0;
105 list_add(&imp_conn->oic_item, &imp->imp_conn_list);
107 list_add_tail(&imp_conn->oic_item,
108 &imp->imp_conn_list);
109 CDEBUG(D_HA, "imp %p@%s: add connection %s at %s\n",
110 imp, imp->imp_obd->obd_name, uuid->uuid,
111 (priority ? "head" : "tail"));
113 spin_unlock(&imp->imp_lock);
114 GOTO(out_free, rc = -ENOENT);
117 spin_unlock(&imp->imp_lock);
121 OBD_FREE(imp_conn, sizeof(*imp_conn));
123 ptlrpc_connection_put(ptlrpc_conn);
127 int import_set_conn_priority(struct obd_import *imp, struct obd_uuid *uuid)
129 return import_set_conn(imp, uuid, 1, 0);
132 int client_import_add_conn(struct obd_import *imp, struct obd_uuid *uuid,
135 return import_set_conn(imp, uuid, priority, 1);
137 EXPORT_SYMBOL(client_import_add_conn);
139 int client_import_del_conn(struct obd_import *imp, struct obd_uuid *uuid)
141 struct obd_import_conn *imp_conn;
142 struct obd_export *dlmexp;
146 spin_lock(&imp->imp_lock);
147 if (list_empty(&imp->imp_conn_list)) {
148 LASSERT(!imp->imp_connection);
152 list_for_each_entry(imp_conn, &imp->imp_conn_list, oic_item) {
153 if (!obd_uuid_equals(uuid, &imp_conn->oic_uuid))
155 LASSERT(imp_conn->oic_conn);
157 if (imp_conn == imp->imp_conn_current) {
158 LASSERT(imp_conn->oic_conn == imp->imp_connection);
160 if (imp->imp_state != LUSTRE_IMP_CLOSED &&
161 imp->imp_state != LUSTRE_IMP_DISCON) {
162 CERROR("can't remove current connection\n");
163 GOTO(out, rc = -EBUSY);
166 ptlrpc_connection_put(imp->imp_connection);
167 imp->imp_connection = NULL;
169 dlmexp = class_conn2export(&imp->imp_dlm_handle);
170 if (dlmexp && dlmexp->exp_connection) {
171 LASSERT(dlmexp->exp_connection ==
173 ptlrpc_connection_put(dlmexp->exp_connection);
174 dlmexp->exp_connection = NULL;
178 list_del(&imp_conn->oic_item);
179 ptlrpc_connection_put(imp_conn->oic_conn);
180 OBD_FREE(imp_conn, sizeof(*imp_conn));
181 CDEBUG(D_HA, "imp %p@%s: remove connection %s\n",
182 imp, imp->imp_obd->obd_name, uuid->uuid);
187 spin_unlock(&imp->imp_lock);
189 CERROR("connection %s not found\n", uuid->uuid);
192 EXPORT_SYMBOL(client_import_del_conn);
195 * Find conn UUID by peer NID. \a peer is a server NID. This function is used
196 * to find a conn uuid of \a imp which can reach \a peer.
198 int client_import_find_conn(struct obd_import *imp, lnet_nid_t peer,
199 struct obd_uuid *uuid)
201 struct obd_import_conn *conn;
205 spin_lock(&imp->imp_lock);
206 list_for_each_entry(conn, &imp->imp_conn_list, oic_item) {
207 /* Check if conn UUID does have this peer NID. */
208 if (class_check_uuid(&conn->oic_uuid, peer)) {
209 *uuid = conn->oic_uuid;
214 spin_unlock(&imp->imp_lock);
217 EXPORT_SYMBOL(client_import_find_conn);
219 void client_destroy_import(struct obd_import *imp)
221 /* Drop security policy instance after all RPCs have finished/aborted
222 * to let all busy contexts be released. */
223 class_import_get(imp);
224 class_destroy_import(imp);
225 sptlrpc_import_sec_put(imp);
226 class_import_put(imp);
228 EXPORT_SYMBOL(client_destroy_import);
231 * Check whether or not the OSC is on MDT.
234 * setup 0:{fsname}-OSTxxxx-osc[-MDTxxxx] 1:lustre-OST0000_UUID 2:NID
236 * setup 0:{fsname}-OSTxxxx-osc 1:lustre-OST0000_UUID 2:NID
239 static int osc_on_mdt(char *obdname)
243 ptr = strrchr(obdname, '-');
247 if (strncmp(ptr + 1, "MDT", 3) == 0)
253 /* Configure an RPC client OBD device.
258 * 3 - inactive-on-startup
260 int client_obd_setup(struct obd_device *obddev, struct lustre_cfg *lcfg)
262 struct client_obd *cli = &obddev->u.cli;
263 struct obd_import *imp;
264 struct obd_uuid server_uuid;
265 int rq_portal, rp_portal, connect_op;
266 char *name = obddev->obd_type->typ_name;
267 ldlm_ns_type_t ns_type = LDLM_NS_TYPE_UNKNOWN;
271 /* In a more perfect world, we would hang a ptlrpc_client off of
272 * obd_type and just use the values from there. */
273 if (!strcmp(name, LUSTRE_OSC_NAME)) {
274 rq_portal = OST_REQUEST_PORTAL;
275 rp_portal = OSC_REPLY_PORTAL;
276 connect_op = OST_CONNECT;
277 cli->cl_sp_me = LUSTRE_SP_CLI;
278 cli->cl_sp_to = LUSTRE_SP_OST;
279 ns_type = LDLM_NS_TYPE_OSC;
280 } else if (!strcmp(name, LUSTRE_MDC_NAME) ||
281 !strcmp(name, LUSTRE_LWP_NAME)) {
282 rq_portal = MDS_REQUEST_PORTAL;
283 rp_portal = MDC_REPLY_PORTAL;
284 connect_op = MDS_CONNECT;
285 cli->cl_sp_me = LUSTRE_SP_CLI;
286 cli->cl_sp_to = LUSTRE_SP_MDT;
287 ns_type = LDLM_NS_TYPE_MDC;
288 } else if (!strcmp(name, LUSTRE_OSP_NAME)) {
289 if (strstr(lustre_cfg_buf(lcfg, 1), "OST") == NULL) {
290 /* OSP_on_MDT for other MDTs */
291 connect_op = MDS_CONNECT;
292 cli->cl_sp_to = LUSTRE_SP_MDT;
293 ns_type = LDLM_NS_TYPE_MDC;
294 rq_portal = OUT_PORTAL;
296 /* OSP on MDT for OST */
297 connect_op = OST_CONNECT;
298 cli->cl_sp_to = LUSTRE_SP_OST;
299 ns_type = LDLM_NS_TYPE_OSC;
300 rq_portal = OST_REQUEST_PORTAL;
302 rp_portal = OSC_REPLY_PORTAL;
303 cli->cl_sp_me = LUSTRE_SP_CLI;
304 } else if (!strcmp(name, LUSTRE_MGC_NAME)) {
305 rq_portal = MGS_REQUEST_PORTAL;
306 rp_portal = MGC_REPLY_PORTAL;
307 connect_op = MGS_CONNECT;
308 cli->cl_sp_me = LUSTRE_SP_MGC;
309 cli->cl_sp_to = LUSTRE_SP_MGS;
310 cli->cl_flvr_mgc.sf_rpc = SPTLRPC_FLVR_INVALID;
311 ns_type = LDLM_NS_TYPE_MGC;
313 CERROR("unknown client OBD type \"%s\", can't setup\n",
318 if (LUSTRE_CFG_BUFLEN(lcfg, 1) < 1) {
319 CERROR("requires a TARGET UUID\n");
323 if (LUSTRE_CFG_BUFLEN(lcfg, 1) > 37) {
324 CERROR("client UUID must be less than 38 characters\n");
328 if (LUSTRE_CFG_BUFLEN(lcfg, 2) < 1) {
329 CERROR("setup requires a SERVER UUID\n");
333 if (LUSTRE_CFG_BUFLEN(lcfg, 2) > 37) {
334 CERROR("target UUID must be less than 38 characters\n");
338 init_rwsem(&cli->cl_sem);
339 mutex_init(&cli->cl_mgc_mutex);
340 cli->cl_conn_count = 0;
341 memcpy(server_uuid.uuid, lustre_cfg_buf(lcfg, 2),
342 min_t(unsigned int, LUSTRE_CFG_BUFLEN(lcfg, 2),
343 sizeof(server_uuid)));
345 cli->cl_dirty_pages = 0;
346 cli->cl_avail_grant = 0;
347 /* FIXME: Should limit this for the sum of all cl_dirty_max_pages. */
348 /* cl_dirty_max_pages may be changed at connect time in
349 * ptlrpc_connect_interpret(). */
350 client_adjust_max_dirty(cli);
351 INIT_LIST_HEAD(&cli->cl_cache_waiters);
352 INIT_LIST_HEAD(&cli->cl_loi_ready_list);
353 INIT_LIST_HEAD(&cli->cl_loi_hp_ready_list);
354 INIT_LIST_HEAD(&cli->cl_loi_write_list);
355 INIT_LIST_HEAD(&cli->cl_loi_read_list);
356 spin_lock_init(&cli->cl_loi_list_lock);
357 atomic_set(&cli->cl_pending_w_pages, 0);
358 atomic_set(&cli->cl_pending_r_pages, 0);
359 cli->cl_r_in_flight = 0;
360 cli->cl_w_in_flight = 0;
362 spin_lock_init(&cli->cl_read_rpc_hist.oh_lock);
363 spin_lock_init(&cli->cl_write_rpc_hist.oh_lock);
364 spin_lock_init(&cli->cl_read_page_hist.oh_lock);
365 spin_lock_init(&cli->cl_write_page_hist.oh_lock);
366 spin_lock_init(&cli->cl_read_offset_hist.oh_lock);
367 spin_lock_init(&cli->cl_write_offset_hist.oh_lock);
370 INIT_LIST_HEAD(&cli->cl_lru_osc);
371 atomic_set(&cli->cl_lru_shrinkers, 0);
372 atomic_long_set(&cli->cl_lru_busy, 0);
373 atomic_long_set(&cli->cl_lru_in_list, 0);
374 INIT_LIST_HEAD(&cli->cl_lru_list);
375 spin_lock_init(&cli->cl_lru_list_lock);
376 atomic_long_set(&cli->cl_unstable_count, 0);
378 init_waitqueue_head(&cli->cl_destroy_waitq);
379 atomic_set(&cli->cl_destroy_in_flight, 0);
380 #ifdef ENABLE_CHECKSUM
381 /* Turn on checksumming by default. */
382 cli->cl_checksum = 1;
384 * The supported checksum types will be worked out at connect time
385 * Set cl_chksum* to CRC32 for now to avoid returning screwed info
388 cli->cl_cksum_type = cli->cl_supp_cksum_types = OBD_CKSUM_CRC32;
390 atomic_set(&cli->cl_resends, OSC_DEFAULT_RESENDS);
392 /* This value may be reduced at connect time in
393 * ptlrpc_connect_interpret() . We initialize it to only
394 * 1MB until we know what the performance looks like.
395 * In the future this should likely be increased. LU-1431 */
396 cli->cl_max_pages_per_rpc = min_t(int, PTLRPC_MAX_BRW_PAGES,
397 LNET_MTU >> PAGE_CACHE_SHIFT);
399 /* set cl_chunkbits default value to PAGE_CACHE_SHIFT,
400 * it will be updated at OSC connection time. */
401 cli->cl_chunkbits = PAGE_CACHE_SHIFT;
403 if (!strcmp(name, LUSTRE_MDC_NAME)) {
404 cli->cl_max_rpcs_in_flight = OBD_MAX_RIF_DEFAULT;
405 } else if (totalram_pages >> (20 - PAGE_CACHE_SHIFT) <= 128 /* MB */) {
406 cli->cl_max_rpcs_in_flight = 2;
407 } else if (totalram_pages >> (20 - PAGE_CACHE_SHIFT) <= 256 /* MB */) {
408 cli->cl_max_rpcs_in_flight = 3;
409 } else if (totalram_pages >> (20 - PAGE_CACHE_SHIFT) <= 512 /* MB */) {
410 cli->cl_max_rpcs_in_flight = 4;
412 if (osc_on_mdt(obddev->obd_name))
413 cli->cl_max_rpcs_in_flight = OBD_MAX_RIF_MAX;
415 cli->cl_max_rpcs_in_flight = OBD_MAX_RIF_DEFAULT;
419 CERROR("ldlm_get_ref failed: %d\n", rc);
423 ptlrpc_init_client(rq_portal, rp_portal, name,
424 &obddev->obd_ldlm_client);
426 imp = class_new_import(obddev);
428 GOTO(err_ldlm, rc = -ENOENT);
429 imp->imp_client = &obddev->obd_ldlm_client;
430 imp->imp_connect_op = connect_op;
431 memcpy(cli->cl_target_uuid.uuid, lustre_cfg_buf(lcfg, 1),
432 LUSTRE_CFG_BUFLEN(lcfg, 1));
433 class_import_put(imp);
435 rc = client_import_add_conn(imp, &server_uuid, 1);
437 CERROR("can't add initial connection\n");
438 GOTO(err_import, rc);
441 cli->cl_import = imp;
442 /* cli->cl_max_mds_{easize,cookiesize} updated by mdc_init_ea_size() */
443 cli->cl_max_mds_easize = sizeof(struct lov_mds_md_v3);
444 cli->cl_max_mds_cookiesize = sizeof(struct llog_cookie);
446 if (LUSTRE_CFG_BUFLEN(lcfg, 3) > 0) {
447 if (!strcmp(lustre_cfg_string(lcfg, 3), "inactive")) {
448 CDEBUG(D_HA, "marking %s %s->%s as inactive\n",
449 name, obddev->obd_name,
450 cli->cl_target_uuid.uuid);
451 spin_lock(&imp->imp_lock);
452 imp->imp_deactive = 1;
453 spin_unlock(&imp->imp_lock);
457 obddev->obd_namespace = ldlm_namespace_new(obddev, obddev->obd_name,
458 LDLM_NAMESPACE_CLIENT,
459 LDLM_NAMESPACE_GREEDY,
461 if (obddev->obd_namespace == NULL) {
462 CERROR("Unable to create client namespace - %s\n",
464 GOTO(err_import, rc = -ENOMEM);
467 cli->cl_qchk_stat = CL_NOT_QUOTACHECKED;
472 class_destroy_import(imp);
479 EXPORT_SYMBOL(client_obd_setup);
481 int client_obd_cleanup(struct obd_device *obddev)
485 ldlm_namespace_free_post(obddev->obd_namespace);
486 obddev->obd_namespace = NULL;
488 obd_cleanup_client_import(obddev);
489 LASSERT(obddev->u.cli.cl_import == NULL);
494 EXPORT_SYMBOL(client_obd_cleanup);
496 /* ->o_connect() method for client side (OSC and MDC and MGC) */
497 int client_connect_import(const struct lu_env *env,
498 struct obd_export **exp,
499 struct obd_device *obd, struct obd_uuid *cluuid,
500 struct obd_connect_data *data, void *localdata)
502 struct client_obd *cli = &obd->u.cli;
503 struct obd_import *imp = cli->cl_import;
504 struct obd_connect_data *ocd;
505 struct lustre_handle conn = { 0 };
510 down_write(&cli->cl_sem);
511 if (cli->cl_conn_count > 0)
512 GOTO(out_sem, rc = -EALREADY);
514 rc = class_connect(&conn, obd, cluuid);
518 cli->cl_conn_count++;
519 *exp = class_conn2export(&conn);
521 LASSERT(obd->obd_namespace);
523 imp->imp_dlm_handle = conn;
524 rc = ptlrpc_init_import(imp);
528 ocd = &imp->imp_connect_data;
531 imp->imp_connect_flags_orig = data->ocd_connect_flags;
534 rc = ptlrpc_connect_import(imp);
536 LASSERT (imp->imp_state == LUSTRE_IMP_DISCON);
539 LASSERT(*exp != NULL && (*exp)->exp_connection);
542 LASSERTF((ocd->ocd_connect_flags & data->ocd_connect_flags) ==
543 ocd->ocd_connect_flags, "old "LPX64", new "LPX64"\n",
544 data->ocd_connect_flags, ocd->ocd_connect_flags);
545 data->ocd_connect_flags = ocd->ocd_connect_flags;
548 ptlrpc_pinger_add_import(imp);
554 cli->cl_conn_count--;
555 class_disconnect(*exp);
559 up_write(&cli->cl_sem);
563 EXPORT_SYMBOL(client_connect_import);
565 int client_disconnect_export(struct obd_export *exp)
567 struct obd_device *obd = class_exp2obd(exp);
568 struct client_obd *cli;
569 struct obd_import *imp;
574 CERROR("invalid export for disconnect: exp %p cookie "LPX64"\n",
575 exp, exp ? exp->exp_handle.h_cookie : -1);
580 imp = cli->cl_import;
582 down_write(&cli->cl_sem);
583 CDEBUG(D_INFO, "disconnect %s - %zu\n", obd->obd_name,
586 if (cli->cl_conn_count == 0) {
587 CERROR("disconnecting disconnected device (%s)\n",
589 GOTO(out_disconnect, rc = -EINVAL);
592 cli->cl_conn_count--;
593 if (cli->cl_conn_count != 0)
594 GOTO(out_disconnect, rc = 0);
596 /* Mark import deactivated now, so we don't try to reconnect if any
597 * of the cleanup RPCs fails (e.g. LDLM cancel, etc). We don't
598 * fully deactivate the import, or that would drop all requests. */
599 spin_lock(&imp->imp_lock);
600 imp->imp_deactive = 1;
601 spin_unlock(&imp->imp_lock);
603 /* Some non-replayable imports (MDS's OSCs) are pinged, so just
604 * delete it regardless. (It's safe to delete an import that was
606 (void)ptlrpc_pinger_del_import(imp);
608 if (obd->obd_namespace != NULL) {
609 /* obd_force == local only */
610 ldlm_cli_cancel_unused(obd->obd_namespace, NULL,
611 obd->obd_force ? LCF_LOCAL : 0, NULL);
612 ldlm_namespace_free_prior(obd->obd_namespace, imp, obd->obd_force);
615 /* There's no need to hold sem while disconnecting an import,
616 * and it may actually cause deadlock in GSS. */
617 up_write(&cli->cl_sem);
618 rc = ptlrpc_disconnect_import(imp, 0);
619 down_write(&cli->cl_sem);
621 ptlrpc_invalidate_import(imp);
626 /* Use server style - class_disconnect should be always called for
628 err = class_disconnect(exp);
632 up_write(&cli->cl_sem);
636 EXPORT_SYMBOL(client_disconnect_export);
638 #ifdef HAVE_SERVER_SUPPORT
639 int server_disconnect_export(struct obd_export *exp)
644 /* Disconnect early so that clients can't keep using export. */
645 rc = class_disconnect(exp);
646 /* Close import to avoid sending any requests. */
647 if (exp->exp_imp_reverse)
648 ptlrpc_cleanup_imp(exp->exp_imp_reverse);
650 if (exp->exp_obd->obd_namespace != NULL)
651 ldlm_cancel_locks_for_export(exp);
653 /* complete all outstanding replies */
654 spin_lock(&exp->exp_lock);
655 while (!list_empty(&exp->exp_outstanding_replies)) {
656 struct ptlrpc_reply_state *rs =
657 list_entry(exp->exp_outstanding_replies.next,
658 struct ptlrpc_reply_state, rs_exp_list);
659 struct ptlrpc_service_part *svcpt = rs->rs_svcpt;
661 spin_lock(&svcpt->scp_rep_lock);
663 list_del_init(&rs->rs_exp_list);
664 spin_lock(&rs->rs_lock);
665 ptlrpc_schedule_difficult_reply(rs);
666 spin_unlock(&rs->rs_lock);
668 spin_unlock(&svcpt->scp_rep_lock);
670 spin_unlock(&exp->exp_lock);
674 EXPORT_SYMBOL(server_disconnect_export);
676 /* --------------------------------------------------------------------------
677 * from old lib/target.c
678 * -------------------------------------------------------------------------- */
680 static int target_handle_reconnect(struct lustre_handle *conn,
681 struct obd_export *exp,
682 struct obd_uuid *cluuid)
686 if (exp->exp_connection && exp->exp_imp_reverse) {
687 struct lustre_handle *hdl;
688 struct obd_device *target;
690 hdl = &exp->exp_imp_reverse->imp_remote_handle;
691 target = exp->exp_obd;
693 /* Might be a re-connect after a partition. */
694 if (!memcmp(&conn->cookie, &hdl->cookie, sizeof conn->cookie)) {
695 if (target->obd_recovering) {
696 int timeout = cfs_duration_sec(cfs_time_sub(
698 &target->obd_recovery_timer),
699 cfs_time_current()));
701 LCONSOLE_WARN("%s: Client %s (at %s) reconnect"
702 "ing, waiting for %d clients in recov"
703 "ery for %d:%.02d\n", target->obd_name,
704 obd_uuid2str(&exp->exp_client_uuid),
705 obd_export_nid2str(exp),
706 target->obd_max_recoverable_clients,
707 timeout / 60, timeout % 60);
709 LCONSOLE_WARN("%s: Client %s (at %s) "
710 "reconnecting\n", target->obd_name,
711 obd_uuid2str(&exp->exp_client_uuid),
712 obd_export_nid2str(exp));
715 conn->cookie = exp->exp_handle.h_cookie;
716 /* target_handle_connect() treats EALREADY and
717 * -EALREADY differently. EALREADY means we are
718 * doing a valid reconnect from the same client. */
721 LCONSOLE_WARN("%s: already connected client %s (at %s) "
722 "with handle "LPX64". Rejecting client "
723 "with the same UUID trying to reconnect "
724 "with handle "LPX64"\n", target->obd_name,
725 obd_uuid2str(&exp->exp_client_uuid),
726 obd_export_nid2str(exp),
727 hdl->cookie, conn->cookie);
728 memset(conn, 0, sizeof *conn);
729 /* target_handle_connect() treats EALREADY and
730 * -EALREADY differently. -EALREADY is an error
731 * (same UUID, different handle). */
736 conn->cookie = exp->exp_handle.h_cookie;
737 CDEBUG(D_HA, "connect export for UUID '%s' at %p, cookie "LPX64"\n",
738 cluuid->uuid, exp, conn->cookie);
742 void target_client_add_cb(struct obd_device *obd, __u64 transno, void *cb_data,
745 struct obd_export *exp = cb_data;
747 CDEBUG(D_RPCTRACE, "%s: committing for initial connect of %s\n",
748 obd->obd_name, exp->exp_client_uuid.uuid);
750 spin_lock(&exp->exp_lock);
751 exp->exp_need_sync = 0;
752 spin_unlock(&exp->exp_lock);
753 class_export_cb_put(exp);
757 check_and_start_recovery_timer(struct obd_device *obd,
758 struct ptlrpc_request *req, int new_client);
760 int target_handle_connect(struct ptlrpc_request *req)
762 struct obd_device *target = NULL, *targref = NULL;
763 struct obd_export *export = NULL;
764 struct obd_import *revimp;
765 struct obd_import *tmp_imp = NULL;
766 struct lustre_handle conn;
767 struct lustre_handle *tmp;
768 struct obd_uuid tgtuuid;
769 struct obd_uuid cluuid;
770 struct obd_uuid remote_uuid;
775 bool mds_conn = false, lw_client = false;
776 struct obd_connect_data *data, *tmpdata;
778 lnet_nid_t *client_nid = NULL;
781 OBD_RACE(OBD_FAIL_TGT_CONN_RACE);
783 str = req_capsule_client_get(&req->rq_pill, &RMF_TGTUUID);
785 DEBUG_REQ(D_ERROR, req, "bad target UUID for connect");
786 GOTO(out, rc = -EINVAL);
789 obd_str2uuid(&tgtuuid, str);
790 target = class_uuid2obd(&tgtuuid);
792 target = class_name2obd(str);
795 deuuidify(str, NULL, &target_start, &target_len);
796 LCONSOLE_ERROR_MSG(0x137, "%s: not available for connect "
797 "from %s (no target). If you are running "
798 "an HA pair check that the target is "
799 "mounted on the other server.\n", str,
800 libcfs_nid2str(req->rq_peer.nid));
801 GOTO(out, rc = -ENODEV);
804 spin_lock(&target->obd_dev_lock);
805 if (target->obd_stopping || !target->obd_set_up) {
806 spin_unlock(&target->obd_dev_lock);
808 deuuidify(str, NULL, &target_start, &target_len);
809 LCONSOLE_INFO("%.*s: Not available for connect from %s (%s)\n",
810 target_len, target_start,
811 libcfs_nid2str(req->rq_peer.nid),
812 (target->obd_stopping ?
813 "stopping" : "not set up"));
814 GOTO(out, rc = -ENODEV);
817 if (target->obd_no_conn) {
818 spin_unlock(&target->obd_dev_lock);
820 CDEBUG(D_INFO, "%s: Temporarily refusing client connection "
821 "from %s\n", target->obd_name,
822 libcfs_nid2str(req->rq_peer.nid));
823 GOTO(out, rc = -EAGAIN);
826 /* Make sure the target isn't cleaned up while we're here. Yes,
827 * there's still a race between the above check and our incref here.
828 * Really, class_uuid2obd should take the ref. */
829 targref = class_incref(target, __FUNCTION__, current);
831 target->obd_conn_inprogress++;
832 spin_unlock(&target->obd_dev_lock);
834 str = req_capsule_client_get(&req->rq_pill, &RMF_CLUUID);
836 DEBUG_REQ(D_ERROR, req, "bad client UUID for connect");
837 GOTO(out, rc = -EINVAL);
840 obd_str2uuid(&cluuid, str);
842 /* XXX Extract a nettype and format accordingly. */
843 switch (sizeof(lnet_nid_t)) {
844 /* NB the casts only avoid compiler warnings. */
846 snprintf(remote_uuid.uuid, sizeof remote_uuid,
847 "NET_"LPX64"_UUID", (__u64)req->rq_peer.nid);
850 snprintf(remote_uuid.uuid, sizeof remote_uuid,
851 "NET_%x_UUID", (__u32)req->rq_peer.nid);
857 tmp = req_capsule_client_get(&req->rq_pill, &RMF_CONN);
859 GOTO(out, rc = -EPROTO);
863 size = req_capsule_get_size(&req->rq_pill, &RMF_CONNECT_DATA,
865 data = req_capsule_client_get(&req->rq_pill, &RMF_CONNECT_DATA);
867 GOTO(out, rc = -EPROTO);
869 rc = req_capsule_server_pack(&req->rq_pill);
873 if (lustre_msg_get_op_flags(req->rq_reqmsg) & MSG_CONNECT_LIBCLIENT) {
874 if (data->ocd_version < LUSTRE_VERSION_CODE -
875 LUSTRE_VERSION_ALLOWED_OFFSET ||
876 data->ocd_version > LUSTRE_VERSION_CODE +
877 LUSTRE_VERSION_ALLOWED_OFFSET) {
878 DEBUG_REQ(D_WARNING, req, "Refusing %s (%d.%d.%d.%d) "
879 "libclient connection attempt",
880 data->ocd_version < LUSTRE_VERSION_CODE ?
882 OBD_OCD_VERSION_MAJOR(data->ocd_version),
883 OBD_OCD_VERSION_MINOR(data->ocd_version),
884 OBD_OCD_VERSION_PATCH(data->ocd_version),
885 OBD_OCD_VERSION_FIX(data->ocd_version));
886 data = req_capsule_server_sized_get(&req->rq_pill,
888 offsetof(typeof(*data), ocd_version) +
889 sizeof(data->ocd_version));
891 data->ocd_connect_flags = OBD_CONNECT_VERSION;
892 data->ocd_version = LUSTRE_VERSION_CODE;
894 GOTO(out, rc = -EPROTO);
898 if (lustre_msg_get_op_flags(req->rq_reqmsg) & MSG_CONNECT_INITIAL) {
899 mds_conn = (data->ocd_connect_flags & OBD_CONNECT_MDS) != 0;
900 lw_client = (data->ocd_connect_flags &
901 OBD_CONNECT_LIGHTWEIGHT) != 0;
903 /* OBD_CONNECT_MNE_SWAB is defined as OBD_CONNECT_MDS_MDS
904 * for Imperative Recovery connection from MGC to MGS.
906 * Via check OBD_CONNECT_FID, we can distinguish whether
907 * the OBD_CONNECT_MDS_MDS/OBD_CONNECT_MNE_SWAB is from
910 (data->ocd_connect_flags & OBD_CONNECT_MDS_MDS) &&
911 (data->ocd_connect_flags & OBD_CONNECT_FID) &&
912 (data->ocd_connect_flags & OBD_CONNECT_VERSION)) {
913 __u32 major = OBD_OCD_VERSION_MAJOR(data->ocd_version);
914 __u32 minor = OBD_OCD_VERSION_MINOR(data->ocd_version);
915 __u32 patch = OBD_OCD_VERSION_PATCH(data->ocd_version);
917 /* We do not support the MDT-MDT interoperations with
918 * different version MDT because of protocol changes. */
919 if (unlikely(major != LUSTRE_MAJOR ||
920 minor != LUSTRE_MINOR ||
921 abs(patch - LUSTRE_PATCH) > 3)) {
922 LCONSOLE_WARN("%s (%u.%u.%u.%u) refused the "
923 "connection from different version MDT "
924 "(%d.%d.%d.%d) %s %s\n",
925 target->obd_name, LUSTRE_MAJOR,
926 LUSTRE_MINOR, LUSTRE_PATCH, LUSTRE_FIX,
928 OBD_OCD_VERSION_FIX(data->ocd_version),
929 libcfs_nid2str(req->rq_peer.nid), str);
931 GOTO(out, rc = -EPROTO);
936 /* lctl gets a backstage, all-access pass. */
937 if (obd_uuid_equals(&cluuid, &target->obd_uuid))
938 goto dont_check_exports;
940 export = cfs_hash_lookup(target->obd_uuid_hash, &cluuid);
944 /* We've found an export in the hash. */
946 spin_lock(&export->exp_lock);
948 if (export->exp_connecting) { /* bug 9635, et. al. */
949 spin_unlock(&export->exp_lock);
950 LCONSOLE_WARN("%s: Export %p already connecting from %s\n",
951 export->exp_obd->obd_name, export,
952 libcfs_nid2str(req->rq_peer.nid));
953 class_export_put(export);
956 } else if ((mds_conn || lw_client) && export->exp_connection != NULL) {
957 spin_unlock(&export->exp_lock);
958 if (req->rq_peer.nid != export->exp_connection->c_peer.nid)
959 /* MDS or LWP reconnected after failover. */
960 LCONSOLE_WARN("%s: Received %s connection from "
961 "%s, removing former export from %s\n",
962 target->obd_name, mds_conn ? "MDS" : "LWP",
963 libcfs_nid2str(req->rq_peer.nid),
964 libcfs_nid2str(export->exp_connection->c_peer.nid));
966 /* New MDS connection from the same NID. */
967 LCONSOLE_WARN("%s: Received new %s connection from "
968 "%s, removing former export from same NID\n",
969 target->obd_name, mds_conn ? "MDS" : "LWP",
970 libcfs_nid2str(req->rq_peer.nid));
971 class_fail_export(export);
972 class_export_put(export);
975 } else if (export->exp_connection != NULL &&
976 req->rq_peer.nid != export->exp_connection->c_peer.nid &&
977 (lustre_msg_get_op_flags(req->rq_reqmsg) &
978 MSG_CONNECT_INITIAL)) {
979 spin_unlock(&export->exp_lock);
980 /* In MDS failover we have static UUID but NID can change. */
981 LCONSOLE_WARN("%s: Client %s seen on new nid %s when "
982 "existing nid %s is already connected\n",
983 target->obd_name, cluuid.uuid,
984 libcfs_nid2str(req->rq_peer.nid),
986 export->exp_connection->c_peer.nid));
988 class_export_put(export);
991 export->exp_connecting = 1;
992 spin_unlock(&export->exp_lock);
993 LASSERT(export->exp_obd == target);
995 rc = target_handle_reconnect(&conn, export, &cluuid);
998 /* If we found an export, we already unlocked. */
1001 OBD_FAIL_TIMEOUT(OBD_FAIL_TGT_DELAY_CONNECT, 2 * obd_timeout);
1002 } else if (req->rq_export == NULL &&
1003 atomic_read(&export->exp_rpc_count) > 0) {
1004 LCONSOLE_WARN("%s: Client %s (at %s) refused connection, "
1005 "still busy with %d references\n",
1006 target->obd_name, cluuid.uuid,
1007 libcfs_nid2str(req->rq_peer.nid),
1008 atomic_read(&export->exp_refcount));
1009 GOTO(out, rc = -EBUSY);
1010 } else if (lustre_msg_get_conn_cnt(req->rq_reqmsg) == 1) {
1011 if (!strstr(cluuid.uuid, "mdt"))
1012 LCONSOLE_WARN("%s: Rejecting reconnect from the "
1013 "known client %s (at %s) because it "
1014 "is indicating it is a new client",
1015 target->obd_name, cluuid.uuid,
1016 libcfs_nid2str(req->rq_peer.nid));
1017 GOTO(out, rc = -EALREADY);
1019 OBD_FAIL_TIMEOUT(OBD_FAIL_TGT_DELAY_RECONNECT, 2 * obd_timeout);
1026 CDEBUG(D_HA, "%s: connection from %s@%s %st"LPU64" exp %p cur %ld last %ld\n",
1027 target->obd_name, cluuid.uuid, libcfs_nid2str(req->rq_peer.nid),
1028 target->obd_recovering ? "recovering/" : "", data->ocd_transno,
1029 export, (long)cfs_time_current_sec(),
1030 export ? (long)export->exp_last_request_time : 0);
1032 /* If this is the first time a client connects, reset the recovery
1033 * timer. Discard lightweight connections which might be local. */
1034 if (!lw_client && rc == 0 && target->obd_recovering)
1035 check_and_start_recovery_timer(target, req, export == NULL);
1037 /* We want to handle EALREADY but *not* -EALREADY from
1038 * target_handle_reconnect(), return reconnection state in a flag. */
1039 if (rc == EALREADY) {
1040 lustre_msg_add_op_flags(req->rq_repmsg, MSG_CONNECT_RECONNECT);
1046 /* Tell the client if we support replayable requests. */
1047 if (target->obd_replayable)
1048 lustre_msg_add_op_flags(req->rq_repmsg, MSG_CONNECT_REPLAYABLE);
1049 client_nid = &req->rq_peer.nid;
1051 if (export == NULL) {
1052 /* allow lightweight connections during recovery */
1053 if (target->obd_recovering && !lw_client) {
1055 int c; /* connected */
1056 int i; /* in progress */
1058 int s; /* stale/evicted */
1060 c = atomic_read(&target->obd_connected_clients);
1061 i = atomic_read(&target->obd_lock_replay_clients);
1062 k = target->obd_max_recoverable_clients;
1063 s = target->obd_stale_clients;
1064 t = cfs_timer_deadline(&target->obd_recovery_timer);
1065 t = cfs_time_sub(t, cfs_time_current());
1066 t = cfs_duration_sec(t);
1067 LCONSOLE_WARN("%s: Denying connection for new client "
1068 "%s (at %s), waiting for all %d known "
1069 "clients (%d recovered, %d in progress, "
1070 "and %d evicted) to recover in %d:%.02d\n",
1071 target->obd_name, cluuid.uuid,
1072 libcfs_nid2str(req->rq_peer.nid), k,
1073 c - i, i, s, (int)t / 60,
1078 rc = obd_connect(req->rq_svc_thread->t_env,
1079 &export, target, &cluuid, data,
1081 if (mds_conn && OBD_FAIL_CHECK(OBD_FAIL_TGT_RCVG_FLAG))
1082 lustre_msg_add_op_flags(req->rq_repmsg,
1083 MSG_CONNECT_RECOVERING);
1085 conn.cookie = export->exp_handle.h_cookie;
1088 rc = obd_reconnect(req->rq_svc_thread->t_env,
1089 export, target, &cluuid, data, client_nid);
1094 LASSERT(target->u.obt.obt_magic == OBT_MAGIC);
1095 data->ocd_instance = target->u.obt.obt_instance;
1097 /* Return only the parts of obd_connect_data that we understand, so the
1098 * client knows that we don't understand the rest. */
1100 tmpsize = req_capsule_get_size(&req->rq_pill, &RMF_CONNECT_DATA,
1102 tmpdata = req_capsule_server_get(&req->rq_pill,
1104 /* Don't use struct assignment here, because the client reply
1105 * buffer may be smaller/larger than the local struct
1106 * obd_connect_data. */
1107 memcpy(tmpdata, data, min(tmpsize, size));
1110 /* If all else goes well, this is our RPC return code. */
1113 lustre_msg_set_handle(req->rq_repmsg, &conn);
1115 /* If the client and the server are the same node, we will already
1116 * have an export that really points to the client's DLM export,
1117 * because we have a shared handles table.
1119 * XXX this will go away when shaver stops sending the "connect" handle
1120 * in the real "remote handle" field of the request --phik 24 Apr 2003
1122 ptlrpc_request_change_export(req, export);
1124 spin_lock(&export->exp_lock);
1125 if (export->exp_conn_cnt >= lustre_msg_get_conn_cnt(req->rq_reqmsg)) {
1126 spin_unlock(&export->exp_lock);
1127 CDEBUG(D_RPCTRACE, "%s: %s already connected at greater "
1128 "or equal conn_cnt: %d >= %d\n",
1129 cluuid.uuid, libcfs_nid2str(req->rq_peer.nid),
1130 export->exp_conn_cnt,
1131 lustre_msg_get_conn_cnt(req->rq_reqmsg));
1133 GOTO(out, rc = -EALREADY);
1135 LASSERT(lustre_msg_get_conn_cnt(req->rq_reqmsg) > 0);
1136 export->exp_conn_cnt = lustre_msg_get_conn_cnt(req->rq_reqmsg);
1138 /* Don't evict liblustre clients for not pinging. */
1139 if (lustre_msg_get_op_flags(req->rq_reqmsg) & MSG_CONNECT_LIBCLIENT) {
1140 export->exp_libclient = 1;
1141 spin_unlock(&export->exp_lock);
1143 spin_lock(&target->obd_dev_lock);
1144 list_del_init(&export->exp_obd_chain_timed);
1145 spin_unlock(&target->obd_dev_lock);
1147 spin_unlock(&export->exp_lock);
1150 if (export->exp_connection != NULL) {
1151 /* Check to see if connection came from another NID. */
1152 if ((export->exp_connection->c_peer.nid != req->rq_peer.nid) &&
1153 !hlist_unhashed(&export->exp_nid_hash))
1154 cfs_hash_del(export->exp_obd->obd_nid_hash,
1155 &export->exp_connection->c_peer.nid,
1156 &export->exp_nid_hash);
1158 ptlrpc_connection_put(export->exp_connection);
1161 export->exp_connection = ptlrpc_connection_get(req->rq_peer,
1164 if (hlist_unhashed(&export->exp_nid_hash)) {
1165 cfs_hash_add(export->exp_obd->obd_nid_hash,
1166 &export->exp_connection->c_peer.nid,
1167 &export->exp_nid_hash);
1170 if (target->obd_recovering && !export->exp_in_recovery && !lw_client) {
1172 __u64 transno = data->ocd_transno;
1174 spin_lock(&export->exp_lock);
1175 /* possible race with class_disconnect_stale_exports,
1176 * export may be already in the eviction process */
1177 if (export->exp_failed) {
1178 spin_unlock(&export->exp_lock);
1179 GOTO(out, rc = -ENODEV);
1181 export->exp_in_recovery = 1;
1182 export->exp_req_replay_needed = 1;
1183 export->exp_lock_replay_needed = 1;
1184 spin_unlock(&export->exp_lock);
1186 has_transno = !!(lustre_msg_get_op_flags(req->rq_reqmsg) &
1187 MSG_CONNECT_TRANSNO);
1188 if (has_transno && transno == 0)
1189 CWARN("Connect with zero transno!\n");
1191 if (has_transno && transno > 0 &&
1192 transno < target->obd_next_recovery_transno &&
1193 transno > target->obd_last_committed) {
1194 /* Another way is to use cmpxchg() to be lock-free. */
1195 spin_lock(&target->obd_recovery_task_lock);
1196 if (transno < target->obd_next_recovery_transno)
1197 target->obd_next_recovery_transno = transno;
1198 spin_unlock(&target->obd_recovery_task_lock);
1201 atomic_inc(&target->obd_req_replay_clients);
1202 atomic_inc(&target->obd_lock_replay_clients);
1203 if (atomic_inc_return(&target->obd_connected_clients) ==
1204 target->obd_max_recoverable_clients)
1205 wake_up(&target->obd_next_transno_waitq);
1208 /* Tell the client we're in recovery, when client is involved in it. */
1209 if (target->obd_recovering && !lw_client)
1210 lustre_msg_add_op_flags(req->rq_repmsg, MSG_CONNECT_RECOVERING);
1212 tmp = req_capsule_client_get(&req->rq_pill, &RMF_CONN);
1215 /* Return -ENOTCONN in case of errors to let client reconnect. */
1216 revimp = class_new_import(target);
1217 if (revimp == NULL) {
1218 CERROR("fail to alloc new reverse import.\n");
1219 GOTO(out, rc = -ENOTCONN);
1222 spin_lock(&export->exp_lock);
1223 if (export->exp_imp_reverse != NULL)
1224 /* destroyed import can be still referenced in ctxt */
1225 tmp_imp = export->exp_imp_reverse;
1226 export->exp_imp_reverse = revimp;
1227 spin_unlock(&export->exp_lock);
1229 revimp->imp_connection = ptlrpc_connection_addref(export->exp_connection);
1230 revimp->imp_client = &export->exp_obd->obd_ldlm_client;
1231 revimp->imp_remote_handle = conn;
1232 revimp->imp_dlm_fake = 1;
1233 revimp->imp_state = LUSTRE_IMP_FULL;
1235 /* Unknown versions will be caught in
1236 * ptlrpc_handle_server_req_in->lustre_unpack_msg(). */
1237 revimp->imp_msg_magic = req->rq_reqmsg->lm_magic;
1239 if ((data->ocd_connect_flags & OBD_CONNECT_AT) &&
1240 (revimp->imp_msg_magic != LUSTRE_MSG_MAGIC_V1))
1241 revimp->imp_msghdr_flags |= MSGHDR_AT_SUPPORT;
1243 revimp->imp_msghdr_flags &= ~MSGHDR_AT_SUPPORT;
1245 if ((data->ocd_connect_flags & OBD_CONNECT_FULL20) &&
1246 (revimp->imp_msg_magic != LUSTRE_MSG_MAGIC_V1))
1247 revimp->imp_msghdr_flags |= MSGHDR_CKSUM_INCOMPAT18;
1249 revimp->imp_msghdr_flags &= ~MSGHDR_CKSUM_INCOMPAT18;
1251 rc = sptlrpc_import_sec_adapt(revimp, req->rq_svc_ctx, &req->rq_flvr);
1253 CERROR("Failed to get sec for reverse import: %d\n", rc);
1254 spin_lock(&export->exp_lock);
1255 export->exp_imp_reverse = NULL;
1256 spin_unlock(&export->exp_lock);
1257 class_destroy_import(revimp);
1260 class_import_put(revimp);
1263 if (tmp_imp != NULL)
1264 client_destroy_import(tmp_imp);
1266 spin_lock(&export->exp_lock);
1267 export->exp_connecting = 0;
1268 spin_unlock(&export->exp_lock);
1270 class_export_put(export);
1273 spin_lock(&target->obd_dev_lock);
1274 target->obd_conn_inprogress--;
1275 spin_unlock(&target->obd_dev_lock);
1277 class_decref(targref, __func__, current);
1280 req->rq_status = rc;
1284 int target_handle_disconnect(struct ptlrpc_request *req)
1289 rc = req_capsule_server_pack(&req->rq_pill);
1293 /* Keep the rq_export around so we can send the reply. */
1294 req->rq_status = obd_disconnect(class_export_get(req->rq_export));
1299 void target_destroy_export(struct obd_export *exp)
1301 struct obd_import *imp = NULL;
1302 /* exports created from last_rcvd data, and "fake"
1303 exports created by lctl don't have an import */
1304 spin_lock(&exp->exp_lock);
1305 if (exp->exp_imp_reverse != NULL) {
1306 imp = exp->exp_imp_reverse;
1307 exp->exp_imp_reverse = NULL;
1309 spin_unlock(&exp->exp_lock);
1311 client_destroy_import(imp);
1313 LASSERT_ATOMIC_ZERO(&exp->exp_locks_count);
1314 LASSERT_ATOMIC_ZERO(&exp->exp_rpc_count);
1315 LASSERT_ATOMIC_ZERO(&exp->exp_cb_count);
1316 LASSERT_ATOMIC_ZERO(&exp->exp_replay_count);
1318 EXPORT_SYMBOL(target_destroy_export);
1321 * Recovery functions
1323 static void target_request_copy_get(struct ptlrpc_request *req)
1325 class_export_rpc_inc(req->rq_export);
1326 LASSERT(list_empty(&req->rq_list));
1327 INIT_LIST_HEAD(&req->rq_replay_list);
1329 /* Increase refcount to keep request in queue. */
1330 atomic_inc(&req->rq_refcount);
1331 /* Let export know it has replays to be handled. */
1332 atomic_inc(&req->rq_export->exp_replay_count);
1335 static void target_request_copy_put(struct ptlrpc_request *req)
1337 LASSERT(list_empty(&req->rq_replay_list));
1338 LASSERT_ATOMIC_POS(&req->rq_export->exp_replay_count);
1340 atomic_dec(&req->rq_export->exp_replay_count);
1341 class_export_rpc_dec(req->rq_export);
1342 ptlrpc_server_drop_request(req);
1345 static int target_exp_enqueue_req_replay(struct ptlrpc_request *req)
1347 __u64 transno = lustre_msg_get_transno(req->rq_reqmsg);
1348 struct obd_export *exp = req->rq_export;
1349 struct ptlrpc_request *reqiter;
1354 spin_lock(&exp->exp_lock);
1355 list_for_each_entry(reqiter, &exp->exp_req_replay_queue,
1357 if (lustre_msg_get_transno(reqiter->rq_reqmsg) == transno) {
1364 /* We expect it with RESENT and REPLAY flags. */
1365 if ((lustre_msg_get_flags(req->rq_reqmsg) &
1366 (MSG_RESENT | MSG_REPLAY)) != (MSG_RESENT | MSG_REPLAY))
1367 CERROR("invalid flags %x of resent replay\n",
1368 lustre_msg_get_flags(req->rq_reqmsg));
1370 list_add_tail(&req->rq_replay_list,
1371 &exp->exp_req_replay_queue);
1374 spin_unlock(&exp->exp_lock);
1378 static void target_exp_dequeue_req_replay(struct ptlrpc_request *req)
1380 LASSERT(!list_empty(&req->rq_replay_list));
1381 LASSERT(req->rq_export);
1383 spin_lock(&req->rq_export->exp_lock);
1384 list_del_init(&req->rq_replay_list);
1385 spin_unlock(&req->rq_export->exp_lock);
1388 static void target_finish_recovery(struct obd_device *obd)
1392 /* Only log a recovery message when recovery has occurred. */
1393 if (obd->obd_recovery_start) {
1394 time_t elapsed_time = max_t(time_t, 1, cfs_time_current_sec() -
1395 obd->obd_recovery_start);
1396 LCONSOLE_INFO("%s: Recovery over after %d:%.02d, of %d clients "
1397 "%d recovered and %d %s evicted.\n", obd->obd_name,
1398 (int)elapsed_time / 60, (int)elapsed_time % 60,
1399 obd->obd_max_recoverable_clients,
1400 atomic_read(&obd->obd_connected_clients),
1401 obd->obd_stale_clients,
1402 obd->obd_stale_clients == 1 ? "was" : "were");
1405 ldlm_reprocess_all_ns(obd->obd_namespace);
1406 spin_lock(&obd->obd_recovery_task_lock);
1407 if (!list_empty(&obd->obd_req_replay_queue) ||
1408 !list_empty(&obd->obd_lock_replay_queue) ||
1409 !list_empty(&obd->obd_final_req_queue)) {
1410 CERROR("%s: Recovery queues ( %s%s%s) are not empty\n",
1412 list_empty(&obd->obd_req_replay_queue) ? "" : "req ",
1413 list_empty(&obd->obd_lock_replay_queue) ? \
1415 list_empty(&obd->obd_final_req_queue) ? \
1417 spin_unlock(&obd->obd_recovery_task_lock);
1420 spin_unlock(&obd->obd_recovery_task_lock);
1422 obd->obd_recovery_end = cfs_time_current_sec();
1424 /* When recovery finished, cleanup orphans on MDS and OST. */
1425 if (OBT(obd) && OBP(obd, postrecov)) {
1426 int rc = OBP(obd, postrecov)(obd);
1428 LCONSOLE_WARN("%s: Post recovery failed, rc %d\n",
1434 static void abort_req_replay_queue(struct obd_device *obd)
1436 struct ptlrpc_request *req, *n;
1437 struct list_head abort_list;
1439 INIT_LIST_HEAD(&abort_list);
1440 spin_lock(&obd->obd_recovery_task_lock);
1441 list_splice_init(&obd->obd_req_replay_queue, &abort_list);
1442 spin_unlock(&obd->obd_recovery_task_lock);
1443 list_for_each_entry_safe(req, n, &abort_list, rq_list) {
1444 DEBUG_REQ(D_WARNING, req, "aborted:");
1445 req->rq_status = -ENOTCONN;
1446 if (ptlrpc_error(req)) {
1447 DEBUG_REQ(D_ERROR, req,
1448 "failed abort_req_reply; skipping");
1450 target_exp_dequeue_req_replay(req);
1451 target_request_copy_put(req);
1455 static void abort_lock_replay_queue(struct obd_device *obd)
1457 struct ptlrpc_request *req, *n;
1458 struct list_head abort_list;
1460 INIT_LIST_HEAD(&abort_list);
1461 spin_lock(&obd->obd_recovery_task_lock);
1462 list_splice_init(&obd->obd_lock_replay_queue, &abort_list);
1463 spin_unlock(&obd->obd_recovery_task_lock);
1464 list_for_each_entry_safe(req, n, &abort_list, rq_list) {
1465 DEBUG_REQ(D_ERROR, req, "aborted:");
1466 req->rq_status = -ENOTCONN;
1467 if (ptlrpc_error(req)) {
1468 DEBUG_REQ(D_ERROR, req,
1469 "failed abort_lock_reply; skipping");
1471 target_request_copy_put(req);
1475 /* Called from a cleanup function if the device is being cleaned up
1476 forcefully. The exports should all have been disconnected already,
1477 the only thing left to do is
1478 - clear the recovery flags
1480 - free queued requests and replies, but don't send replies
1481 Because the obd_stopping flag is set, no new requests should be received.
1484 void target_cleanup_recovery(struct obd_device *obd)
1486 struct ptlrpc_request *req, *n;
1487 struct list_head clean_list;
1490 INIT_LIST_HEAD(&clean_list);
1491 spin_lock(&obd->obd_dev_lock);
1492 if (!obd->obd_recovering) {
1493 spin_unlock(&obd->obd_dev_lock);
1497 obd->obd_recovering = obd->obd_abort_recovery = 0;
1498 spin_unlock(&obd->obd_dev_lock);
1500 spin_lock(&obd->obd_recovery_task_lock);
1501 target_cancel_recovery_timer(obd);
1502 list_splice_init(&obd->obd_req_replay_queue, &clean_list);
1503 spin_unlock(&obd->obd_recovery_task_lock);
1505 list_for_each_entry_safe(req, n, &clean_list, rq_list) {
1506 LASSERT(req->rq_reply_state == NULL);
1507 target_exp_dequeue_req_replay(req);
1508 target_request_copy_put(req);
1511 spin_lock(&obd->obd_recovery_task_lock);
1512 list_splice_init(&obd->obd_lock_replay_queue, &clean_list);
1513 list_splice_init(&obd->obd_final_req_queue, &clean_list);
1514 spin_unlock(&obd->obd_recovery_task_lock);
1516 list_for_each_entry_safe(req, n, &clean_list, rq_list) {
1517 LASSERT(req->rq_reply_state == NULL);
1518 target_request_copy_put(req);
1523 EXPORT_SYMBOL(target_cleanup_recovery);
1525 /* obd_recovery_task_lock should be held */
1526 void target_cancel_recovery_timer(struct obd_device *obd)
1528 CDEBUG(D_HA, "%s: cancel recovery timer\n", obd->obd_name);
1529 cfs_timer_disarm(&obd->obd_recovery_timer);
1532 static void target_start_recovery_timer(struct obd_device *obd)
1534 if (obd->obd_recovery_start != 0)
1537 spin_lock(&obd->obd_dev_lock);
1538 if (!obd->obd_recovering || obd->obd_abort_recovery) {
1539 spin_unlock(&obd->obd_dev_lock);
1543 LASSERT(obd->obd_recovery_timeout != 0);
1545 if (obd->obd_recovery_start != 0) {
1546 spin_unlock(&obd->obd_dev_lock);
1550 cfs_timer_arm(&obd->obd_recovery_timer,
1551 cfs_time_shift(obd->obd_recovery_timeout));
1552 obd->obd_recovery_start = cfs_time_current_sec();
1553 spin_unlock(&obd->obd_dev_lock);
1555 LCONSOLE_WARN("%s: Will be in recovery for at least %d:%.02d, "
1556 "or until %d client%s reconnect%s\n",
1558 obd->obd_recovery_timeout / 60,
1559 obd->obd_recovery_timeout % 60,
1560 obd->obd_max_recoverable_clients,
1561 (obd->obd_max_recoverable_clients == 1) ? "" : "s",
1562 (obd->obd_max_recoverable_clients == 1) ? "s": "");
1566 * extend recovery window.
1568 * if @extend is true, extend recovery window to have @drt remaining at least;
1569 * otherwise, make sure the recovery timeout value is not less than @drt.
1571 static void extend_recovery_timer(struct obd_device *obd, int drt, bool extend)
1575 cfs_duration_t left;
1578 spin_lock(&obd->obd_dev_lock);
1579 if (!obd->obd_recovering || obd->obd_abort_recovery) {
1580 spin_unlock(&obd->obd_dev_lock);
1583 LASSERT(obd->obd_recovery_start != 0);
1585 now = cfs_time_current_sec();
1586 to = obd->obd_recovery_timeout;
1587 end = obd->obd_recovery_start + to;
1588 left = cfs_time_sub(end, now);
1590 if (extend && (drt > left)) {
1592 } else if (!extend && (drt > to)) {
1596 if (to > obd->obd_recovery_time_hard)
1597 to = obd->obd_recovery_time_hard;
1598 if (obd->obd_recovery_timeout < to) {
1599 obd->obd_recovery_timeout = to;
1600 end = obd->obd_recovery_start + to;
1601 cfs_timer_arm(&obd->obd_recovery_timer,
1602 cfs_time_shift(end - now));
1604 spin_unlock(&obd->obd_dev_lock);
1606 CDEBUG(D_HA, "%s: recovery timer will expire in %u seconds\n",
1607 obd->obd_name, (unsigned)cfs_time_sub(end, now));
1610 /* Reset the timer with each new client connection */
1612 * This timer is actually reconnect_timer, which is for making sure
1613 * the total recovery window is at least as big as my reconnect
1614 * attempt timing. So the initial recovery time_out will be set to
1615 * OBD_RECOVERY_FACTOR * obd_timeout. If the timeout coming
1616 * from client is bigger than this, then the recovery time_out will
1617 * be extended to make sure the client could be reconnected, in the
1618 * process, the timeout from the new client should be ignored.
1622 check_and_start_recovery_timer(struct obd_device *obd,
1623 struct ptlrpc_request *req,
1626 int service_time = lustre_msg_get_service_time(req->rq_reqmsg);
1627 struct obd_device_target *obt = &obd->u.obt;
1629 if (!new_client && service_time)
1630 /* Teach server about old server's estimates, as first guess
1631 * at how long new requests will take. */
1632 at_measured(&req->rq_rqbd->rqbd_svcpt->scp_at_estimate,
1635 target_start_recovery_timer(obd);
1637 /* Convert the service time to RPC timeout,
1638 * and reuse service_time to limit stack usage. */
1639 service_time = at_est2timeout(service_time);
1641 /* We expect other clients to timeout within service_time, then try
1642 * to reconnect, then try the failover server. The max delay between
1643 * connect attempts is SWITCH_MAX + SWITCH_INC + INITIAL. */
1644 service_time += 2 * INITIAL_CONNECT_TIMEOUT;
1646 LASSERT(obt->obt_magic == OBT_MAGIC);
1647 service_time += 2 * (CONNECTION_SWITCH_MAX + CONNECTION_SWITCH_INC);
1648 if (service_time > obd->obd_recovery_timeout && !new_client)
1649 extend_recovery_timer(obd, service_time, false);
1652 /** Health checking routines */
1653 static inline int exp_connect_healthy(struct obd_export *exp)
1655 return (exp->exp_in_recovery);
1658 /** if export done req_replay or has replay in queue */
1659 static inline int exp_req_replay_healthy(struct obd_export *exp)
1661 return (!exp->exp_req_replay_needed ||
1662 atomic_read(&exp->exp_replay_count) > 0);
1664 /** if export done lock_replay or has replay in queue */
1665 static inline int exp_lock_replay_healthy(struct obd_export *exp)
1667 return (!exp->exp_lock_replay_needed ||
1668 atomic_read(&exp->exp_replay_count) > 0);
1671 static inline int exp_vbr_healthy(struct obd_export *exp)
1673 return (!exp->exp_vbr_failed);
1676 static inline int exp_finished(struct obd_export *exp)
1678 return (exp->exp_in_recovery && !exp->exp_lock_replay_needed);
1681 /** Checking routines for recovery */
1682 static int check_for_clients(struct obd_device *obd)
1684 unsigned int clnts = atomic_read(&obd->obd_connected_clients);
1686 if (obd->obd_abort_recovery || obd->obd_recovery_expired)
1688 LASSERT(clnts <= obd->obd_max_recoverable_clients);
1689 return (clnts + obd->obd_stale_clients ==
1690 obd->obd_max_recoverable_clients);
1693 static int check_for_next_transno(struct obd_device *obd)
1695 struct ptlrpc_request *req = NULL;
1696 int wake_up = 0, connected, completed, queue_len;
1697 __u64 next_transno, req_transno;
1700 spin_lock(&obd->obd_recovery_task_lock);
1701 if (!list_empty(&obd->obd_req_replay_queue)) {
1702 req = list_entry(obd->obd_req_replay_queue.next,
1703 struct ptlrpc_request, rq_list);
1704 req_transno = lustre_msg_get_transno(req->rq_reqmsg);
1709 connected = atomic_read(&obd->obd_connected_clients);
1710 completed = connected - atomic_read(&obd->obd_req_replay_clients);
1711 queue_len = obd->obd_requests_queued_for_recovery;
1712 next_transno = obd->obd_next_recovery_transno;
1714 CDEBUG(D_HA, "max: %d, connected: %d, completed: %d, queue_len: %d, "
1715 "req_transno: "LPU64", next_transno: "LPU64"\n",
1716 obd->obd_max_recoverable_clients, connected, completed,
1717 queue_len, req_transno, next_transno);
1719 if (obd->obd_abort_recovery) {
1720 CDEBUG(D_HA, "waking for aborted recovery\n");
1722 } else if (obd->obd_recovery_expired) {
1723 CDEBUG(D_HA, "waking for expired recovery\n");
1725 } else if (req_transno == next_transno) {
1726 CDEBUG(D_HA, "waking for next ("LPD64")\n", next_transno);
1728 } else if (queue_len > 0 &&
1729 queue_len == atomic_read(&obd->obd_req_replay_clients)) {
1731 /** handle gaps occured due to lost reply or VBR */
1732 LASSERTF(req_transno >= next_transno,
1733 "req_transno: "LPU64", next_transno: "LPU64"\n",
1734 req_transno, next_transno);
1735 if (req_transno > obd->obd_last_committed &&
1736 !obd->obd_version_recov)
1739 "%s: waking for gap in transno, VBR is %s (skip: "
1740 LPD64", ql: %d, comp: %d, conn: %d, next: "LPD64
1741 ", last_committed: "LPD64")\n",
1742 obd->obd_name, obd->obd_version_recov ? "ON" : "OFF",
1743 next_transno, queue_len, completed, connected,
1744 req_transno, obd->obd_last_committed);
1745 obd->obd_next_recovery_transno = req_transno;
1747 } else if (atomic_read(&obd->obd_req_replay_clients) == 0) {
1748 CDEBUG(D_HA, "waking for completed recovery\n");
1750 } else if (OBD_FAIL_CHECK(OBD_FAIL_MDS_RECOVERY_ACCEPTS_GAPS)) {
1751 CDEBUG(D_HA, "accepting transno gaps is explicitly allowed"
1752 " by fail_lock, waking up ("LPD64")\n", next_transno);
1753 obd->obd_next_recovery_transno = req_transno;
1756 spin_unlock(&obd->obd_recovery_task_lock);
1760 static int check_for_next_lock(struct obd_device *obd)
1764 spin_lock(&obd->obd_recovery_task_lock);
1765 if (!list_empty(&obd->obd_lock_replay_queue)) {
1766 CDEBUG(D_HA, "waking for next lock\n");
1768 } else if (atomic_read(&obd->obd_lock_replay_clients) == 0) {
1769 CDEBUG(D_HA, "waking for completed lock replay\n");
1771 } else if (obd->obd_abort_recovery) {
1772 CDEBUG(D_HA, "waking for aborted recovery\n");
1774 } else if (obd->obd_recovery_expired) {
1775 CDEBUG(D_HA, "waking for expired recovery\n");
1778 spin_unlock(&obd->obd_recovery_task_lock);
1784 * wait for recovery events,
1785 * check its status with help of check_routine
1786 * evict dead clients via health_check
1788 static int target_recovery_overseer(struct obd_device *obd,
1789 int (*check_routine)(struct obd_device *),
1790 int (*health_check)(struct obd_export *))
1793 if ((obd->obd_recovery_start != 0) && (cfs_time_current_sec() >=
1794 (obd->obd_recovery_start + obd->obd_recovery_time_hard))) {
1795 CWARN("recovery is aborted by hard timeout\n");
1796 obd->obd_abort_recovery = 1;
1799 while (wait_event_timeout(obd->obd_next_transno_waitq,
1801 msecs_to_jiffies(60 * MSEC_PER_SEC)) == 0)
1802 /* wait indefinitely for event, but don't trigger watchdog */;
1804 if (obd->obd_abort_recovery) {
1805 CWARN("recovery is aborted, evict exports in recovery\n");
1806 /** evict exports which didn't finish recovery yet */
1807 class_disconnect_stale_exports(obd, exp_finished);
1809 } else if (obd->obd_recovery_expired) {
1810 obd->obd_recovery_expired = 0;
1811 /** If some clients died being recovered, evict them */
1812 LCONSOLE_WARN("%s: recovery is timed out, "
1813 "evict stale exports\n", obd->obd_name);
1814 /** evict cexports with no replay in queue, they are stalled */
1815 class_disconnect_stale_exports(obd, health_check);
1816 /** continue with VBR */
1817 spin_lock(&obd->obd_dev_lock);
1818 obd->obd_version_recov = 1;
1819 spin_unlock(&obd->obd_dev_lock);
1821 * reset timer, recovery will proceed with versions now,
1822 * timeout is set just to handle reconnection delays
1824 extend_recovery_timer(obd, RECONNECT_DELAY_MAX, true);
1825 /** Wait for recovery events again, after evicting bad clients */
1831 static struct ptlrpc_request *target_next_replay_req(struct obd_device *obd)
1833 struct ptlrpc_request *req = NULL;
1836 CDEBUG(D_HA, "Waiting for transno "LPD64"\n",
1837 obd->obd_next_recovery_transno);
1839 CFS_FAIL_TIMEOUT(OBD_FAIL_TGT_REPLAY_DELAY2, cfs_fail_val);
1840 /** It is needed to extend recovery window above recovery_time_soft.
1841 * Extending is possible only in the end of recovery window
1842 * (see more details in handle_recovery_req).
1844 CFS_FAIL_TIMEOUT_MS(OBD_FAIL_TGT_REPLAY_DELAY, 300);
1846 if (target_recovery_overseer(obd, check_for_next_transno,
1847 exp_req_replay_healthy)) {
1848 abort_req_replay_queue(obd);
1849 abort_lock_replay_queue(obd);
1852 spin_lock(&obd->obd_recovery_task_lock);
1853 if (!list_empty(&obd->obd_req_replay_queue)) {
1854 req = list_entry(obd->obd_req_replay_queue.next,
1855 struct ptlrpc_request, rq_list);
1856 list_del_init(&req->rq_list);
1857 obd->obd_requests_queued_for_recovery--;
1858 spin_unlock(&obd->obd_recovery_task_lock);
1860 spin_unlock(&obd->obd_recovery_task_lock);
1861 LASSERT(list_empty(&obd->obd_req_replay_queue));
1862 LASSERT(atomic_read(&obd->obd_req_replay_clients) == 0);
1863 /** evict exports failed VBR */
1864 class_disconnect_stale_exports(obd, exp_vbr_healthy);
1869 static struct ptlrpc_request *target_next_replay_lock(struct obd_device *obd)
1871 struct ptlrpc_request *req = NULL;
1873 CDEBUG(D_HA, "Waiting for lock\n");
1874 if (target_recovery_overseer(obd, check_for_next_lock,
1875 exp_lock_replay_healthy))
1876 abort_lock_replay_queue(obd);
1878 spin_lock(&obd->obd_recovery_task_lock);
1879 if (!list_empty(&obd->obd_lock_replay_queue)) {
1880 req = list_entry(obd->obd_lock_replay_queue.next,
1881 struct ptlrpc_request, rq_list);
1882 list_del_init(&req->rq_list);
1883 spin_unlock(&obd->obd_recovery_task_lock);
1885 spin_unlock(&obd->obd_recovery_task_lock);
1886 LASSERT(list_empty(&obd->obd_lock_replay_queue));
1887 LASSERT(atomic_read(&obd->obd_lock_replay_clients) == 0);
1888 /** evict exports failed VBR */
1889 class_disconnect_stale_exports(obd, exp_vbr_healthy);
1894 static struct ptlrpc_request *target_next_final_ping(struct obd_device *obd)
1896 struct ptlrpc_request *req = NULL;
1898 spin_lock(&obd->obd_recovery_task_lock);
1899 if (!list_empty(&obd->obd_final_req_queue)) {
1900 req = list_entry(obd->obd_final_req_queue.next,
1901 struct ptlrpc_request, rq_list);
1902 list_del_init(&req->rq_list);
1903 spin_unlock(&obd->obd_recovery_task_lock);
1904 if (req->rq_export->exp_in_recovery) {
1905 spin_lock(&req->rq_export->exp_lock);
1906 req->rq_export->exp_in_recovery = 0;
1907 spin_unlock(&req->rq_export->exp_lock);
1910 spin_unlock(&obd->obd_recovery_task_lock);
1915 static void handle_recovery_req(struct ptlrpc_thread *thread,
1916 struct ptlrpc_request *req,
1917 svc_handler_t handler)
1922 * export can be evicted during recovery, no need to handle replays for
1923 * it after that, discard such request silently
1925 if (req->rq_export->exp_disconnected)
1928 req->rq_session.lc_thread = thread;
1929 req->rq_svc_thread = thread;
1930 req->rq_svc_thread->t_env->le_ses = &req->rq_session;
1932 /* thread context */
1933 lu_context_enter(&thread->t_env->le_ctx);
1935 lu_context_exit(&thread->t_env->le_ctx);
1937 /* don't reset timer for final stage */
1938 if (!exp_finished(req->rq_export)) {
1939 int to = obd_timeout;
1942 * Add request timeout to the recovery time so next request from
1943 * this client may come in recovery time
1946 struct ptlrpc_service_part *svcpt;
1948 svcpt = req->rq_rqbd->rqbd_svcpt;
1949 /* If the server sent early reply for this request,
1950 * the client will recalculate the timeout according to
1951 * current server estimate service time, so we will
1952 * use the maxium timeout here for waiting the client
1953 * sending the next req */
1954 to = max((int)at_est2timeout(
1955 at_get(&svcpt->scp_at_estimate)),
1956 (int)lustre_msg_get_timeout(req->rq_reqmsg));
1957 /* Add 2 net_latency, one for balance rq_deadline
1958 * (see ptl_send_rpc), one for resend the req to server,
1959 * Note: client will pack net_latency in replay req
1960 * (see ptlrpc_replay_req) */
1961 to += 2 * lustre_msg_get_service_time(req->rq_reqmsg);
1963 extend_recovery_timer(class_exp2obd(req->rq_export), to, true);
1968 static int target_recovery_thread(void *arg)
1970 struct lu_target *lut = arg;
1971 struct obd_device *obd = lut->lut_obd;
1972 struct ptlrpc_request *req;
1973 struct target_recovery_data *trd = &obd->obd_recovery_data;
1974 unsigned long delta;
1976 struct ptlrpc_thread *thread = NULL;
1980 unshare_fs_struct();
1981 OBD_ALLOC_PTR(thread);
1987 OBD_FREE_PTR(thread);
1991 rc = lu_context_init(&env->le_ctx, LCT_MD_THREAD | LCT_DT_THREAD);
1993 OBD_FREE_PTR(thread);
1998 thread->t_env = env;
1999 thread->t_id = -1; /* force filter_iobuf_get/put to use local buffers */
2000 env->le_ctx.lc_thread = thread;
2001 tgt_io_thread_init(thread); /* init thread_big_cache for IO requests */
2002 thread->t_watchdog = NULL;
2004 CDEBUG(D_HA, "%s: started recovery thread pid %d\n", obd->obd_name,
2006 trd->trd_processing_task = current_pid();
2008 spin_lock(&obd->obd_dev_lock);
2009 obd->obd_recovering = 1;
2010 spin_unlock(&obd->obd_dev_lock);
2011 complete(&trd->trd_starting);
2013 /* first of all, we have to know the first transno to replay */
2014 if (target_recovery_overseer(obd, check_for_clients,
2015 exp_connect_healthy)) {
2016 abort_req_replay_queue(obd);
2017 abort_lock_replay_queue(obd);
2020 /* next stage: replay requests */
2022 CDEBUG(D_INFO, "1: request replay stage - %d clients from t"LPU64"\n",
2023 atomic_read(&obd->obd_req_replay_clients),
2024 obd->obd_next_recovery_transno);
2025 while ((req = target_next_replay_req(obd))) {
2026 LASSERT(trd->trd_processing_task == current_pid());
2027 DEBUG_REQ(D_HA, req, "processing t"LPD64" from %s",
2028 lustre_msg_get_transno(req->rq_reqmsg),
2029 libcfs_nid2str(req->rq_peer.nid));
2030 handle_recovery_req(thread, req,
2031 trd->trd_recovery_handler);
2033 * bz18031: increase next_recovery_transno before
2034 * target_request_copy_put() will drop exp_rpc reference
2036 spin_lock(&obd->obd_recovery_task_lock);
2037 obd->obd_next_recovery_transno++;
2038 spin_unlock(&obd->obd_recovery_task_lock);
2039 target_exp_dequeue_req_replay(req);
2040 target_request_copy_put(req);
2041 obd->obd_replayed_requests++;
2045 * The second stage: replay locks
2047 CDEBUG(D_INFO, "2: lock replay stage - %d clients\n",
2048 atomic_read(&obd->obd_lock_replay_clients));
2049 while ((req = target_next_replay_lock(obd))) {
2050 LASSERT(trd->trd_processing_task == current_pid());
2051 DEBUG_REQ(D_HA, req, "processing lock from %s: ",
2052 libcfs_nid2str(req->rq_peer.nid));
2053 handle_recovery_req(thread, req,
2054 trd->trd_recovery_handler);
2055 target_request_copy_put(req);
2056 obd->obd_replayed_locks++;
2060 * The third stage: reply on final pings, at this moment all clients
2061 * must have request in final queue
2063 CFS_FAIL_TIMEOUT(OBD_FAIL_TGT_REPLAY_RECONNECT, cfs_fail_val);
2064 CDEBUG(D_INFO, "3: final stage - process recovery completion pings\n");
2065 /** Update server last boot epoch */
2066 tgt_boot_epoch_update(lut);
2067 /* We drop recoverying flag to forward all new requests
2068 * to regular mds_handle() since now */
2069 spin_lock(&obd->obd_dev_lock);
2070 obd->obd_recovering = obd->obd_abort_recovery = 0;
2071 spin_unlock(&obd->obd_dev_lock);
2072 spin_lock(&obd->obd_recovery_task_lock);
2073 target_cancel_recovery_timer(obd);
2074 spin_unlock(&obd->obd_recovery_task_lock);
2075 while ((req = target_next_final_ping(obd))) {
2076 LASSERT(trd->trd_processing_task == current_pid());
2077 DEBUG_REQ(D_HA, req, "processing final ping from %s: ",
2078 libcfs_nid2str(req->rq_peer.nid));
2079 handle_recovery_req(thread, req,
2080 trd->trd_recovery_handler);
2081 /* Because the waiting client can not send ping to server,
2082 * so we need refresh the last_request_time, to avoid the
2083 * export is being evicted */
2084 ptlrpc_update_export_timer(req->rq_export, 0);
2085 target_request_copy_put(req);
2088 delta = jiffies_to_msecs(jiffies - delta) / MSEC_PER_SEC;
2089 CDEBUG(D_INFO,"4: recovery completed in %lus - %d/%d reqs/locks\n",
2090 delta, obd->obd_replayed_requests, obd->obd_replayed_locks);
2091 if (delta > OBD_RECOVERY_TIME_SOFT) {
2092 CWARN("too long recovery - read logs\n");
2093 libcfs_debug_dumplog();
2096 target_finish_recovery(obd);
2098 lu_context_fini(&env->le_ctx);
2099 trd->trd_processing_task = 0;
2100 complete(&trd->trd_finishing);
2102 tgt_io_thread_done(thread);
2103 OBD_FREE_PTR(thread);
2108 static int target_start_recovery_thread(struct lu_target *lut,
2109 svc_handler_t handler)
2111 struct obd_device *obd = lut->lut_obd;
2113 struct target_recovery_data *trd = &obd->obd_recovery_data;
2115 memset(trd, 0, sizeof(*trd));
2116 init_completion(&trd->trd_starting);
2117 init_completion(&trd->trd_finishing);
2118 trd->trd_recovery_handler = handler;
2120 if (!IS_ERR(kthread_run(target_recovery_thread,
2121 lut, "tgt_recov"))) {
2122 wait_for_completion(&trd->trd_starting);
2123 LASSERT(obd->obd_recovering != 0);
2131 void target_stop_recovery_thread(struct obd_device *obd)
2133 if (obd->obd_recovery_data.trd_processing_task > 0) {
2134 struct target_recovery_data *trd = &obd->obd_recovery_data;
2135 /** recovery can be done but postrecovery is not yet */
2136 spin_lock(&obd->obd_dev_lock);
2137 if (obd->obd_recovering) {
2138 CERROR("%s: Aborting recovery\n", obd->obd_name);
2139 obd->obd_abort_recovery = 1;
2140 wake_up(&obd->obd_next_transno_waitq);
2142 spin_unlock(&obd->obd_dev_lock);
2143 wait_for_completion(&trd->trd_finishing);
2146 EXPORT_SYMBOL(target_stop_recovery_thread);
2148 void target_recovery_fini(struct obd_device *obd)
2150 class_disconnect_exports(obd);
2151 target_stop_recovery_thread(obd);
2152 target_cleanup_recovery(obd);
2154 EXPORT_SYMBOL(target_recovery_fini);
2156 static void target_recovery_expired(unsigned long castmeharder)
2158 struct obd_device *obd = (struct obd_device *)castmeharder;
2159 CDEBUG(D_HA, "%s: recovery timed out; %d clients are still in recovery"
2160 " after %lds (%d clients connected)\n",
2161 obd->obd_name, atomic_read(&obd->obd_lock_replay_clients),
2162 cfs_time_current_sec()- obd->obd_recovery_start,
2163 atomic_read(&obd->obd_connected_clients));
2165 obd->obd_recovery_expired = 1;
2166 wake_up(&obd->obd_next_transno_waitq);
2169 void target_recovery_init(struct lu_target *lut, svc_handler_t handler)
2171 struct obd_device *obd = lut->lut_obd;
2172 if (obd->obd_max_recoverable_clients == 0) {
2173 /** Update server last boot epoch */
2174 tgt_boot_epoch_update(lut);
2178 CDEBUG(D_HA, "RECOVERY: service %s, %d recoverable clients, "
2179 "last_transno "LPU64"\n", obd->obd_name,
2180 obd->obd_max_recoverable_clients, obd->obd_last_committed);
2181 LASSERT(obd->obd_stopping == 0);
2182 obd->obd_next_recovery_transno = obd->obd_last_committed + 1;
2183 obd->obd_recovery_start = 0;
2184 obd->obd_recovery_end = 0;
2186 cfs_timer_init(&obd->obd_recovery_timer, target_recovery_expired, obd);
2187 target_start_recovery_thread(lut, handler);
2189 EXPORT_SYMBOL(target_recovery_init);
2192 static int target_process_req_flags(struct obd_device *obd,
2193 struct ptlrpc_request *req)
2195 struct obd_export *exp = req->rq_export;
2196 LASSERT(exp != NULL);
2197 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REQ_REPLAY_DONE) {
2198 /* client declares he's ready to replay locks */
2199 spin_lock(&exp->exp_lock);
2200 if (exp->exp_req_replay_needed) {
2201 exp->exp_req_replay_needed = 0;
2202 spin_unlock(&exp->exp_lock);
2204 LASSERT_ATOMIC_POS(&obd->obd_req_replay_clients);
2205 atomic_dec(&obd->obd_req_replay_clients);
2207 spin_unlock(&exp->exp_lock);
2210 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_LOCK_REPLAY_DONE) {
2211 /* client declares he's ready to complete recovery
2212 * so, we put the request on th final queue */
2213 spin_lock(&exp->exp_lock);
2214 if (exp->exp_lock_replay_needed) {
2215 exp->exp_lock_replay_needed = 0;
2216 spin_unlock(&exp->exp_lock);
2218 LASSERT_ATOMIC_POS(&obd->obd_lock_replay_clients);
2219 atomic_dec(&obd->obd_lock_replay_clients);
2221 spin_unlock(&exp->exp_lock);
2227 int target_queue_recovery_request(struct ptlrpc_request *req,
2228 struct obd_device *obd)
2230 __u64 transno = lustre_msg_get_transno(req->rq_reqmsg);
2231 struct ptlrpc_request *reqiter;
2235 if (obd->obd_recovery_data.trd_processing_task == current_pid()) {
2236 /* Processing the queue right now, don't re-add. */
2240 target_process_req_flags(obd, req);
2242 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_LOCK_REPLAY_DONE) {
2243 /* client declares he's ready to complete recovery
2244 * so, we put the request on th final queue */
2245 target_request_copy_get(req);
2246 DEBUG_REQ(D_HA, req, "queue final req");
2247 wake_up(&obd->obd_next_transno_waitq);
2248 spin_lock(&obd->obd_recovery_task_lock);
2249 if (obd->obd_recovering) {
2250 list_add_tail(&req->rq_list,
2251 &obd->obd_final_req_queue);
2253 spin_unlock(&obd->obd_recovery_task_lock);
2254 target_request_copy_put(req);
2255 RETURN(obd->obd_stopping ? -ENOTCONN : 1);
2257 spin_unlock(&obd->obd_recovery_task_lock);
2260 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REQ_REPLAY_DONE) {
2261 /* client declares he's ready to replay locks */
2262 target_request_copy_get(req);
2263 DEBUG_REQ(D_HA, req, "queue lock replay req");
2264 wake_up(&obd->obd_next_transno_waitq);
2265 spin_lock(&obd->obd_recovery_task_lock);
2266 LASSERT(obd->obd_recovering);
2267 /* usually due to recovery abort */
2268 if (!req->rq_export->exp_in_recovery) {
2269 spin_unlock(&obd->obd_recovery_task_lock);
2270 target_request_copy_put(req);
2273 LASSERT(req->rq_export->exp_lock_replay_needed);
2274 list_add_tail(&req->rq_list, &obd->obd_lock_replay_queue);
2275 spin_unlock(&obd->obd_recovery_task_lock);
2279 /* CAVEAT EMPTOR: The incoming request message has been swabbed
2280 * (i.e. buflens etc are in my own byte order), but type-dependent
2281 * buffers (eg mdt_body, ost_body etc) have NOT been swabbed. */
2284 INIT_LIST_HEAD(&req->rq_list);
2285 DEBUG_REQ(D_HA, req, "not queueing");
2289 /* If we're processing the queue, we want don't want to queue this
2292 * Also, if this request has a transno less than the one we're waiting
2293 * for, we should process it now. It could (and currently always will)
2294 * be an open request for a descriptor that was opened some time ago.
2296 * Also, a resent, replayed request that has already been
2297 * handled will pass through here and be processed immediately.
2299 CDEBUG(D_HA, "Next recovery transno: "LPU64
2300 ", current: "LPU64", replaying\n",
2301 obd->obd_next_recovery_transno, transno);
2302 spin_lock(&obd->obd_recovery_task_lock);
2303 if (transno < obd->obd_next_recovery_transno) {
2304 /* Processing the queue right now, don't re-add. */
2305 LASSERT(list_empty(&req->rq_list));
2306 spin_unlock(&obd->obd_recovery_task_lock);
2309 spin_unlock(&obd->obd_recovery_task_lock);
2311 if (OBD_FAIL_CHECK(OBD_FAIL_TGT_REPLAY_DROP))
2314 target_request_copy_get(req);
2315 if (!req->rq_export->exp_in_recovery) {
2316 target_request_copy_put(req);
2319 LASSERT(req->rq_export->exp_req_replay_needed);
2321 if (target_exp_enqueue_req_replay(req)) {
2322 DEBUG_REQ(D_ERROR, req, "dropping resent queued req");
2323 target_request_copy_put(req);
2328 spin_lock(&obd->obd_recovery_task_lock);
2329 LASSERT(obd->obd_recovering);
2330 list_for_each_entry(reqiter, &obd->obd_req_replay_queue, rq_list) {
2331 if (lustre_msg_get_transno(reqiter->rq_reqmsg) > transno) {
2332 list_add_tail(&req->rq_list, &reqiter->rq_list);
2337 if (unlikely(lustre_msg_get_transno(reqiter->rq_reqmsg) ==
2339 DEBUG_REQ(D_ERROR, req, "dropping replay: transno "
2340 "has been claimed by another client");
2341 spin_unlock(&obd->obd_recovery_task_lock);
2342 target_exp_dequeue_req_replay(req);
2343 target_request_copy_put(req);
2349 list_add_tail(&req->rq_list, &obd->obd_req_replay_queue);
2351 obd->obd_requests_queued_for_recovery++;
2352 spin_unlock(&obd->obd_recovery_task_lock);
2353 wake_up(&obd->obd_next_transno_waitq);
2357 int target_handle_ping(struct ptlrpc_request *req)
2359 obd_ping(req->rq_svc_thread->t_env, req->rq_export);
2360 return req_capsule_server_pack(&req->rq_pill);
2363 void target_committed_to_req(struct ptlrpc_request *req)
2365 struct obd_export *exp = req->rq_export;
2367 if (!exp->exp_obd->obd_no_transno && req->rq_repmsg != NULL)
2368 lustre_msg_set_last_committed(req->rq_repmsg,
2369 exp->exp_last_committed);
2371 DEBUG_REQ(D_IOCTL, req, "not sending last_committed update (%d/"
2372 "%d)", exp->exp_obd->obd_no_transno,
2373 req->rq_repmsg == NULL);
2375 CDEBUG(D_INFO, "last_committed "LPU64", transno "LPU64", xid "LPU64"\n",
2376 exp->exp_last_committed, req->rq_transno, req->rq_xid);
2379 #endif /* HAVE_SERVER_SUPPORT */
2382 * Packs current SLV and Limit into \a req.
2384 int target_pack_pool_reply(struct ptlrpc_request *req)
2386 struct obd_device *obd;
2389 /* Check that we still have all structures alive as this may
2390 * be some late RPC at shutdown time. */
2391 if (unlikely(!req->rq_export || !req->rq_export->exp_obd ||
2392 !exp_connect_lru_resize(req->rq_export))) {
2393 lustre_msg_set_slv(req->rq_repmsg, 0);
2394 lustre_msg_set_limit(req->rq_repmsg, 0);
2398 /* OBD is alive here as export is alive, which we checked above. */
2399 obd = req->rq_export->exp_obd;
2401 read_lock(&obd->obd_pool_lock);
2402 lustre_msg_set_slv(req->rq_repmsg, obd->obd_pool_slv);
2403 lustre_msg_set_limit(req->rq_repmsg, obd->obd_pool_limit);
2404 read_unlock(&obd->obd_pool_lock);
2409 static int target_send_reply_msg(struct ptlrpc_request *req,
2410 int rc, int fail_id)
2412 if (OBD_FAIL_CHECK_ORSET(fail_id & ~OBD_FAIL_ONCE, OBD_FAIL_ONCE)) {
2413 DEBUG_REQ(D_ERROR, req, "dropping reply");
2418 DEBUG_REQ(D_NET, req, "processing error (%d)", rc);
2419 req->rq_status = rc;
2420 return (ptlrpc_send_error(req, 1));
2422 DEBUG_REQ(D_NET, req, "sending reply");
2425 return (ptlrpc_send_reply(req, PTLRPC_REPLY_MAYBE_DIFFICULT));
2428 void target_send_reply(struct ptlrpc_request *req, int rc, int fail_id)
2430 struct ptlrpc_service_part *svcpt;
2432 struct ptlrpc_reply_state *rs;
2433 struct obd_export *exp;
2436 if (req->rq_no_reply) {
2441 svcpt = req->rq_rqbd->rqbd_svcpt;
2442 rs = req->rq_reply_state;
2443 if (rs == NULL || !rs->rs_difficult) {
2445 target_send_reply_msg (req, rc, fail_id);
2450 /* must be an export if locks saved */
2451 LASSERT(req->rq_export != NULL);
2452 /* req/reply consistent */
2453 LASSERT(rs->rs_svcpt == svcpt);
2456 LASSERT(!rs->rs_scheduled);
2457 LASSERT(!rs->rs_scheduled_ever);
2458 LASSERT(!rs->rs_handled);
2459 LASSERT(!rs->rs_on_net);
2460 LASSERT(rs->rs_export == NULL);
2461 LASSERT(list_empty(&rs->rs_obd_list));
2462 LASSERT(list_empty(&rs->rs_exp_list));
2464 exp = class_export_get(req->rq_export);
2466 /* disable reply scheduling while I'm setting up */
2467 rs->rs_scheduled = 1;
2469 rs->rs_xid = req->rq_xid;
2470 rs->rs_transno = req->rq_transno;
2471 rs->rs_export = exp;
2472 rs->rs_opc = lustre_msg_get_opc(req->rq_reqmsg);
2474 spin_lock(&exp->exp_uncommitted_replies_lock);
2475 CDEBUG(D_NET, "rs transno = "LPU64", last committed = "LPU64"\n",
2476 rs->rs_transno, exp->exp_last_committed);
2477 if (rs->rs_transno > exp->exp_last_committed) {
2478 /* not committed already */
2479 list_add_tail(&rs->rs_obd_list,
2480 &exp->exp_uncommitted_replies);
2482 spin_unlock(&exp->exp_uncommitted_replies_lock);
2484 spin_lock(&exp->exp_lock);
2485 list_add_tail(&rs->rs_exp_list, &exp->exp_outstanding_replies);
2486 spin_unlock(&exp->exp_lock);
2488 netrc = target_send_reply_msg(req, rc, fail_id);
2490 spin_lock(&svcpt->scp_rep_lock);
2492 atomic_inc(&svcpt->scp_nreps_difficult);
2495 /* error sending: reply is off the net. Also we need +1
2496 * reply ref until ptlrpc_handle_rs() is done
2497 * with the reply state (if the send was successful, there
2498 * would have been +1 ref for the net, which
2499 * reply_out_callback leaves alone) */
2501 ptlrpc_rs_addref(rs);
2504 spin_lock(&rs->rs_lock);
2505 if (rs->rs_transno <= exp->exp_last_committed ||
2506 (!rs->rs_on_net && !rs->rs_no_ack) ||
2507 list_empty(&rs->rs_exp_list) || /* completed already */
2508 list_empty(&rs->rs_obd_list)) {
2509 CDEBUG(D_HA, "Schedule reply immediately\n");
2510 ptlrpc_dispatch_difficult_reply(rs);
2512 list_add(&rs->rs_list, &svcpt->scp_rep_active);
2513 rs->rs_scheduled = 0; /* allow notifier to schedule */
2515 spin_unlock(&rs->rs_lock);
2516 spin_unlock(&svcpt->scp_rep_lock);
2520 ldlm_mode_t lck_compat_array[] = {
2521 [LCK_EX] = LCK_COMPAT_EX,
2522 [LCK_PW] = LCK_COMPAT_PW,
2523 [LCK_PR] = LCK_COMPAT_PR,
2524 [LCK_CW] = LCK_COMPAT_CW,
2525 [LCK_CR] = LCK_COMPAT_CR,
2526 [LCK_NL] = LCK_COMPAT_NL,
2527 [LCK_GROUP] = LCK_COMPAT_GROUP,
2528 [LCK_COS] = LCK_COMPAT_COS,
2532 * Rather arbitrary mapping from LDLM error codes to errno values. This should
2533 * not escape to the user level.
2535 int ldlm_error2errno(ldlm_error_t error)
2541 case ELDLM_LOCK_MATCHED:
2544 case ELDLM_LOCK_CHANGED:
2547 case ELDLM_LOCK_ABORTED:
2550 case ELDLM_LOCK_REPLACED:
2553 case ELDLM_NO_LOCK_DATA:
2556 case ELDLM_NAMESPACE_EXISTS:
2559 case ELDLM_BAD_NAMESPACE:
2563 if (((int)error) < 0) /* cast to signed type */
2564 result = error; /* as ldlm_error_t can be unsigned */
2566 CERROR("Invalid DLM result code: %d\n", error);
2572 EXPORT_SYMBOL(ldlm_error2errno);
2575 * Dual to ldlm_error2errno(): maps errno values back to ldlm_error_t.
2577 ldlm_error_t ldlm_errno2error(int err_no)
2586 error = ELDLM_LOCK_CHANGED;
2589 error = ELDLM_LOCK_ABORTED;
2592 error = ELDLM_LOCK_REPLACED;
2595 error = ELDLM_NO_LOCK_DATA;
2598 error = ELDLM_NAMESPACE_EXISTS;
2601 error = ELDLM_BAD_NAMESPACE;
2609 #if LUSTRE_TRACKS_LOCK_EXP_REFS
2610 void ldlm_dump_export_locks(struct obd_export *exp)
2612 spin_lock(&exp->exp_locks_list_guard);
2613 if (!list_empty(&exp->exp_locks_list)) {
2614 struct ldlm_lock *lock;
2616 CERROR("dumping locks for export %p,"
2617 "ignore if the unmount doesn't hang\n", exp);
2618 list_for_each_entry(lock, &exp->exp_locks_list,
2620 LDLM_ERROR(lock, "lock:");
2622 spin_unlock(&exp->exp_locks_list_guard);
2626 #ifdef HAVE_SERVER_SUPPORT
2627 static int target_bulk_timeout(void *data)
2630 /* We don't fail the connection here, because having the export
2631 * killed makes the (vital) call to commitrw very sad.
2636 static inline const char *bulk2type(struct ptlrpc_request *req)
2638 if (req->rq_bulk_read)
2640 if (req->rq_bulk_write)
2645 int target_bulk_io(struct obd_export *exp, struct ptlrpc_bulk_desc *desc,
2646 struct l_wait_info *lwi)
2648 struct ptlrpc_request *req = desc->bd_req;
2649 time_t start = cfs_time_current_sec();
2655 /* If there is eviction in progress, wait for it to finish. */
2656 if (unlikely(atomic_read(&exp->exp_obd->obd_evict_inprogress))) {
2657 *lwi = LWI_INTR(NULL, NULL);
2658 rc = l_wait_event(exp->exp_obd->obd_evict_inprogress_waitq,
2659 !atomic_read(&exp->exp_obd->
2660 obd_evict_inprogress),
2664 /* Check if client was evicted or reconnected already. */
2665 if (exp->exp_failed ||
2666 exp->exp_conn_cnt > lustre_msg_get_conn_cnt(req->rq_reqmsg)) {
2669 if (req->rq_bulk_read)
2670 rc = sptlrpc_svc_wrap_bulk(req, desc);
2672 rc = ptlrpc_start_bulk_transfer(desc);
2676 DEBUG_REQ(D_ERROR, req, "bulk %s failed: rc %d",
2677 bulk2type(req), rc);
2681 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_SENDPAGE)) {
2682 ptlrpc_abort_bulk(desc);
2686 /* limit actual bulk transfer to bulk_timeout seconds */
2687 deadline = start + bulk_timeout;
2688 if (deadline > req->rq_deadline)
2689 deadline = req->rq_deadline;
2692 long timeoutl = deadline - cfs_time_current_sec();
2693 cfs_duration_t timeout = timeoutl <= 0 ?
2694 CFS_TICK : cfs_time_seconds(timeoutl);
2697 *lwi = LWI_TIMEOUT_INTERVAL(timeout, cfs_time_seconds(1),
2698 target_bulk_timeout, desc);
2699 rc = l_wait_event(desc->bd_waitq,
2700 !ptlrpc_server_bulk_active(desc) ||
2703 lustre_msg_get_conn_cnt(req->rq_reqmsg),
2705 LASSERT(rc == 0 || rc == -ETIMEDOUT);
2706 /* Wait again if we changed rq_deadline. */
2707 rq_deadline = ACCESS_ONCE(req->rq_deadline);
2708 deadline = start + bulk_timeout;
2709 if (deadline > rq_deadline)
2710 deadline = rq_deadline;
2711 } while ((rc == -ETIMEDOUT) &&
2712 (deadline > cfs_time_current_sec()));
2714 if (rc == -ETIMEDOUT) {
2715 DEBUG_REQ(D_ERROR, req, "timeout on bulk %s after %ld%+lds",
2716 bulk2type(req), deadline - start,
2717 cfs_time_current_sec() - deadline);
2718 ptlrpc_abort_bulk(desc);
2719 } else if (exp->exp_failed) {
2720 DEBUG_REQ(D_ERROR, req, "Eviction on bulk %s",
2723 ptlrpc_abort_bulk(desc);
2724 } else if (exp->exp_conn_cnt >
2725 lustre_msg_get_conn_cnt(req->rq_reqmsg)) {
2726 DEBUG_REQ(D_ERROR, req, "Reconnect on bulk %s",
2728 /* We don't reply anyway. */
2730 ptlrpc_abort_bulk(desc);
2731 } else if (desc->bd_failure) {
2732 DEBUG_REQ(D_ERROR, req, "network error on bulk %s",
2734 /* XXX should this be a different errno? */
2737 if (req->rq_bulk_write)
2738 rc = sptlrpc_svc_unwrap_bulk(req, desc);
2739 if (rc == 0 && desc->bd_nob_transferred != desc->bd_nob) {
2740 DEBUG_REQ(D_ERROR, req, "truncated bulk %s %d(%d)",
2741 bulk2type(req), desc->bd_nob_transferred,
2743 /* XXX should this be a different errno? */
2750 EXPORT_SYMBOL(target_bulk_io);
2752 #endif /* HAVE_SERVER_SUPPORT */