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, 2013, 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 client_obd_list_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 client_obd_list_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 - %d\n", obd->obd_name,
586 if (!cli->cl_conn_count) {
587 CERROR("disconnecting disconnected device (%s)\n",
589 GOTO(out_disconnect, rc = -EINVAL);
592 cli->cl_conn_count--;
593 if (cli->cl_conn_count)
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);
755 EXPORT_SYMBOL(target_client_add_cb);
758 check_and_start_recovery_timer(struct obd_device *obd,
759 struct ptlrpc_request *req, int new_client);
761 int target_handle_connect(struct ptlrpc_request *req)
763 struct obd_device *target = NULL, *targref = NULL;
764 struct obd_export *export = NULL;
765 struct obd_import *revimp;
766 struct obd_import *tmp_imp = NULL;
767 struct lustre_handle conn;
768 struct lustre_handle *tmp;
769 struct obd_uuid tgtuuid;
770 struct obd_uuid cluuid;
771 struct obd_uuid remote_uuid;
776 bool mds_conn = false, lw_client = false;
777 struct obd_connect_data *data, *tmpdata;
779 lnet_nid_t *client_nid = NULL;
782 OBD_RACE(OBD_FAIL_TGT_CONN_RACE);
784 str = req_capsule_client_get(&req->rq_pill, &RMF_TGTUUID);
786 DEBUG_REQ(D_ERROR, req, "bad target UUID for connect");
787 GOTO(out, rc = -EINVAL);
790 obd_str2uuid(&tgtuuid, str);
791 target = class_uuid2obd(&tgtuuid);
793 target = class_name2obd(str);
796 deuuidify(str, NULL, &target_start, &target_len);
797 LCONSOLE_ERROR_MSG(0x137, "%s: not available for connect "
798 "from %s (no target). If you are running "
799 "an HA pair check that the target is "
800 "mounted on the other server.\n", str,
801 libcfs_nid2str(req->rq_peer.nid));
802 GOTO(out, rc = -ENODEV);
805 spin_lock(&target->obd_dev_lock);
806 if (target->obd_stopping || !target->obd_set_up) {
807 spin_unlock(&target->obd_dev_lock);
809 deuuidify(str, NULL, &target_start, &target_len);
810 LCONSOLE_INFO("%.*s: Not available for connect from %s (%s)\n",
811 target_len, target_start,
812 libcfs_nid2str(req->rq_peer.nid),
813 (target->obd_stopping ?
814 "stopping" : "not set up"));
815 GOTO(out, rc = -ENODEV);
818 if (target->obd_no_conn) {
819 spin_unlock(&target->obd_dev_lock);
821 CDEBUG(D_INFO, "%s: Temporarily refusing client connection "
822 "from %s\n", target->obd_name,
823 libcfs_nid2str(req->rq_peer.nid));
824 GOTO(out, rc = -EAGAIN);
827 /* Make sure the target isn't cleaned up while we're here. Yes,
828 * there's still a race between the above check and our incref here.
829 * Really, class_uuid2obd should take the ref. */
830 targref = class_incref(target, __FUNCTION__, current);
832 target->obd_conn_inprogress++;
833 spin_unlock(&target->obd_dev_lock);
835 str = req_capsule_client_get(&req->rq_pill, &RMF_CLUUID);
837 DEBUG_REQ(D_ERROR, req, "bad client UUID for connect");
838 GOTO(out, rc = -EINVAL);
841 obd_str2uuid(&cluuid, str);
843 /* XXX Extract a nettype and format accordingly. */
844 switch (sizeof(lnet_nid_t)) {
845 /* NB the casts only avoid compiler warnings. */
847 snprintf(remote_uuid.uuid, sizeof remote_uuid,
848 "NET_"LPX64"_UUID", (__u64)req->rq_peer.nid);
851 snprintf(remote_uuid.uuid, sizeof remote_uuid,
852 "NET_%x_UUID", (__u32)req->rq_peer.nid);
858 tmp = req_capsule_client_get(&req->rq_pill, &RMF_CONN);
860 GOTO(out, rc = -EPROTO);
864 size = req_capsule_get_size(&req->rq_pill, &RMF_CONNECT_DATA,
866 data = req_capsule_client_get(&req->rq_pill, &RMF_CONNECT_DATA);
868 GOTO(out, rc = -EPROTO);
870 rc = req_capsule_server_pack(&req->rq_pill);
874 if (lustre_msg_get_op_flags(req->rq_reqmsg) & MSG_CONNECT_LIBCLIENT) {
875 if (data->ocd_version < LUSTRE_VERSION_CODE -
876 LUSTRE_VERSION_ALLOWED_OFFSET ||
877 data->ocd_version > LUSTRE_VERSION_CODE +
878 LUSTRE_VERSION_ALLOWED_OFFSET) {
879 DEBUG_REQ(D_WARNING, req, "Refusing %s (%d.%d.%d.%d) "
880 "libclient connection attempt",
881 data->ocd_version < LUSTRE_VERSION_CODE ?
883 OBD_OCD_VERSION_MAJOR(data->ocd_version),
884 OBD_OCD_VERSION_MINOR(data->ocd_version),
885 OBD_OCD_VERSION_PATCH(data->ocd_version),
886 OBD_OCD_VERSION_FIX(data->ocd_version));
887 data = req_capsule_server_sized_get(&req->rq_pill,
889 offsetof(typeof(*data), ocd_version) +
890 sizeof(data->ocd_version));
892 data->ocd_connect_flags = OBD_CONNECT_VERSION;
893 data->ocd_version = LUSTRE_VERSION_CODE;
895 GOTO(out, rc = -EPROTO);
899 if (lustre_msg_get_op_flags(req->rq_reqmsg) & MSG_CONNECT_INITIAL) {
900 mds_conn = (data->ocd_connect_flags & OBD_CONNECT_MDS) != 0;
901 lw_client = (data->ocd_connect_flags &
902 OBD_CONNECT_LIGHTWEIGHT) != 0;
905 /* lctl gets a backstage, all-access pass. */
906 if (obd_uuid_equals(&cluuid, &target->obd_uuid))
907 goto dont_check_exports;
909 export = cfs_hash_lookup(target->obd_uuid_hash, &cluuid);
913 /* We've found an export in the hash. */
915 spin_lock(&export->exp_lock);
917 if (export->exp_connecting) { /* bug 9635, et. al. */
918 spin_unlock(&export->exp_lock);
919 LCONSOLE_WARN("%s: Export %p already connecting from %s\n",
920 export->exp_obd->obd_name, export,
921 libcfs_nid2str(req->rq_peer.nid));
922 class_export_put(export);
925 } else if ((mds_conn || lw_client) && export->exp_connection != NULL) {
926 spin_unlock(&export->exp_lock);
927 if (req->rq_peer.nid != export->exp_connection->c_peer.nid)
928 /* MDS or LWP reconnected after failover. */
929 LCONSOLE_WARN("%s: Received %s connection from "
930 "%s, removing former export from %s\n",
931 target->obd_name, mds_conn ? "MDS" : "LWP",
932 libcfs_nid2str(req->rq_peer.nid),
933 libcfs_nid2str(export->exp_connection->c_peer.nid));
935 /* New MDS connection from the same NID. */
936 LCONSOLE_WARN("%s: Received new %s connection from "
937 "%s, removing former export from same NID\n",
938 target->obd_name, mds_conn ? "MDS" : "LWP",
939 libcfs_nid2str(req->rq_peer.nid));
940 class_fail_export(export);
941 class_export_put(export);
944 } else if (export->exp_connection != NULL &&
945 req->rq_peer.nid != export->exp_connection->c_peer.nid &&
946 (lustre_msg_get_op_flags(req->rq_reqmsg) &
947 MSG_CONNECT_INITIAL)) {
948 spin_unlock(&export->exp_lock);
949 /* In MDS failover we have static UUID but NID can change. */
950 LCONSOLE_WARN("%s: Client %s seen on new nid %s when "
951 "existing nid %s is already connected\n",
952 target->obd_name, cluuid.uuid,
953 libcfs_nid2str(req->rq_peer.nid),
955 export->exp_connection->c_peer.nid));
957 class_export_put(export);
960 export->exp_connecting = 1;
961 spin_unlock(&export->exp_lock);
962 LASSERT(export->exp_obd == target);
964 rc = target_handle_reconnect(&conn, export, &cluuid);
967 /* If we found an export, we already unlocked. */
970 OBD_FAIL_TIMEOUT(OBD_FAIL_TGT_DELAY_CONNECT, 2 * obd_timeout);
971 } else if (req->rq_export == NULL &&
972 atomic_read(&export->exp_rpc_count) > 0) {
973 LCONSOLE_WARN("%s: Client %s (at %s) refused connection, "
974 "still busy with %d references\n",
975 target->obd_name, cluuid.uuid,
976 libcfs_nid2str(req->rq_peer.nid),
977 atomic_read(&export->exp_refcount));
978 GOTO(out, rc = -EBUSY);
979 } else if (lustre_msg_get_conn_cnt(req->rq_reqmsg) == 1) {
980 if (!strstr(cluuid.uuid, "mdt"))
981 LCONSOLE_WARN("%s: Rejecting reconnect from the "
982 "known client %s (at %s) because it "
983 "is indicating it is a new client",
984 target->obd_name, cluuid.uuid,
985 libcfs_nid2str(req->rq_peer.nid));
986 GOTO(out, rc = -EALREADY);
988 OBD_FAIL_TIMEOUT(OBD_FAIL_TGT_DELAY_RECONNECT, 2 * obd_timeout);
995 CDEBUG(D_HA, "%s: connection from %s@%s %st"LPU64" exp %p cur %ld last %ld\n",
996 target->obd_name, cluuid.uuid, libcfs_nid2str(req->rq_peer.nid),
997 target->obd_recovering ? "recovering/" : "", data->ocd_transno,
998 export, (long)cfs_time_current_sec(),
999 export ? (long)export->exp_last_request_time : 0);
1001 /* If this is the first time a client connects, reset the recovery
1002 * timer. Discard lightweight connections which might be local. */
1003 if (!lw_client && rc == 0 && target->obd_recovering)
1004 check_and_start_recovery_timer(target, req, export == NULL);
1006 /* We want to handle EALREADY but *not* -EALREADY from
1007 * target_handle_reconnect(), return reconnection state in a flag. */
1008 if (rc == EALREADY) {
1009 lustre_msg_add_op_flags(req->rq_repmsg, MSG_CONNECT_RECONNECT);
1015 /* Tell the client if we support replayable requests. */
1016 if (target->obd_replayable)
1017 lustre_msg_add_op_flags(req->rq_repmsg, MSG_CONNECT_REPLAYABLE);
1018 client_nid = &req->rq_peer.nid;
1020 if (export == NULL) {
1021 /* allow lightweight connections during recovery */
1022 if (target->obd_recovering && !lw_client) {
1024 int c; /* connected */
1025 int i; /* in progress */
1027 int s; /* stale/evicted */
1029 c = atomic_read(&target->obd_connected_clients);
1030 i = atomic_read(&target->obd_lock_replay_clients);
1031 k = target->obd_max_recoverable_clients;
1032 s = target->obd_stale_clients;
1033 t = cfs_timer_deadline(&target->obd_recovery_timer);
1034 t = cfs_time_sub(t, cfs_time_current());
1035 t = cfs_duration_sec(t);
1036 LCONSOLE_WARN("%s: Denying connection for new client "
1037 "%s (at %s), waiting for all %d known "
1038 "clients (%d recovered, %d in progress, "
1039 "and %d evicted) to recover in %d:%.02d\n",
1040 target->obd_name, cluuid.uuid,
1041 libcfs_nid2str(req->rq_peer.nid), k,
1042 c - i, i, s, (int)t / 60,
1047 rc = obd_connect(req->rq_svc_thread->t_env,
1048 &export, target, &cluuid, data,
1050 if (mds_conn && OBD_FAIL_CHECK(OBD_FAIL_TGT_RCVG_FLAG))
1051 lustre_msg_add_op_flags(req->rq_repmsg,
1052 MSG_CONNECT_RECOVERING);
1054 conn.cookie = export->exp_handle.h_cookie;
1057 rc = obd_reconnect(req->rq_svc_thread->t_env,
1058 export, target, &cluuid, data, client_nid);
1063 #if LUSTRE_VERSION_CODE < OBD_OCD_VERSION(2, 6, 53, 0)
1064 /* 2.2.0 clients always swab nidtbl entries due to a bug, so server
1065 * will do the swabbing for if the client is using the same endianness.
1067 * This fixup is version-limited, because we don't want to carry the
1068 * OBD_CONNECT_MNE_SWAB flag around forever, just so long as we need
1069 * interop with unpatched 2.2 clients. For newer clients, servers
1070 * will never do MNE swabbing, let the client handle that. LU-1644 */
1071 spin_lock(&export->exp_lock);
1072 export->exp_need_mne_swab = !ptlrpc_req_need_swab(req) &&
1073 !(data->ocd_connect_flags & OBD_CONNECT_MNE_SWAB);
1074 spin_unlock(&export->exp_lock);
1077 LASSERT(target->u.obt.obt_magic == OBT_MAGIC);
1078 data->ocd_instance = target->u.obt.obt_instance;
1080 /* Return only the parts of obd_connect_data that we understand, so the
1081 * client knows that we don't understand the rest. */
1083 tmpsize = req_capsule_get_size(&req->rq_pill, &RMF_CONNECT_DATA,
1085 tmpdata = req_capsule_server_get(&req->rq_pill,
1087 /* Don't use struct assignment here, because the client reply
1088 * buffer may be smaller/larger than the local struct
1089 * obd_connect_data. */
1090 memcpy(tmpdata, data, min(tmpsize, size));
1093 /* If all else goes well, this is our RPC return code. */
1096 lustre_msg_set_handle(req->rq_repmsg, &conn);
1098 /* If the client and the server are the same node, we will already
1099 * have an export that really points to the client's DLM export,
1100 * because we have a shared handles table.
1102 * XXX this will go away when shaver stops sending the "connect" handle
1103 * in the real "remote handle" field of the request --phik 24 Apr 2003
1105 ptlrpc_request_change_export(req, export);
1107 spin_lock(&export->exp_lock);
1108 if (export->exp_conn_cnt >= lustre_msg_get_conn_cnt(req->rq_reqmsg)) {
1109 spin_unlock(&export->exp_lock);
1110 CDEBUG(D_RPCTRACE, "%s: %s already connected at greater "
1111 "or equal conn_cnt: %d >= %d\n",
1112 cluuid.uuid, libcfs_nid2str(req->rq_peer.nid),
1113 export->exp_conn_cnt,
1114 lustre_msg_get_conn_cnt(req->rq_reqmsg));
1116 GOTO(out, rc = -EALREADY);
1118 LASSERT(lustre_msg_get_conn_cnt(req->rq_reqmsg) > 0);
1119 export->exp_conn_cnt = lustre_msg_get_conn_cnt(req->rq_reqmsg);
1121 /* Don't evict liblustre clients for not pinging. */
1122 if (lustre_msg_get_op_flags(req->rq_reqmsg) & MSG_CONNECT_LIBCLIENT) {
1123 export->exp_libclient = 1;
1124 spin_unlock(&export->exp_lock);
1126 spin_lock(&target->obd_dev_lock);
1127 list_del_init(&export->exp_obd_chain_timed);
1128 spin_unlock(&target->obd_dev_lock);
1130 spin_unlock(&export->exp_lock);
1133 if (export->exp_connection != NULL) {
1134 /* Check to see if connection came from another NID. */
1135 if ((export->exp_connection->c_peer.nid != req->rq_peer.nid) &&
1136 !hlist_unhashed(&export->exp_nid_hash))
1137 cfs_hash_del(export->exp_obd->obd_nid_hash,
1138 &export->exp_connection->c_peer.nid,
1139 &export->exp_nid_hash);
1141 ptlrpc_connection_put(export->exp_connection);
1144 export->exp_connection = ptlrpc_connection_get(req->rq_peer,
1147 if (hlist_unhashed(&export->exp_nid_hash)) {
1148 cfs_hash_add(export->exp_obd->obd_nid_hash,
1149 &export->exp_connection->c_peer.nid,
1150 &export->exp_nid_hash);
1153 if (target->obd_recovering && !export->exp_in_recovery && !lw_client) {
1155 __u64 transno = data->ocd_transno;
1157 spin_lock(&export->exp_lock);
1158 /* possible race with class_disconnect_stale_exports,
1159 * export may be already in the eviction process */
1160 if (export->exp_failed) {
1161 spin_unlock(&export->exp_lock);
1162 GOTO(out, rc = -ENODEV);
1164 export->exp_in_recovery = 1;
1165 export->exp_req_replay_needed = 1;
1166 export->exp_lock_replay_needed = 1;
1167 spin_unlock(&export->exp_lock);
1169 has_transno = !!(lustre_msg_get_op_flags(req->rq_reqmsg) &
1170 MSG_CONNECT_TRANSNO);
1171 if (has_transno && transno == 0)
1172 CWARN("Connect with zero transno!\n");
1174 if (has_transno && transno > 0 &&
1175 transno < target->obd_next_recovery_transno &&
1176 transno > target->obd_last_committed) {
1177 /* Another way is to use cmpxchg() to be lock-free. */
1178 spin_lock(&target->obd_recovery_task_lock);
1179 if (transno < target->obd_next_recovery_transno)
1180 target->obd_next_recovery_transno = transno;
1181 spin_unlock(&target->obd_recovery_task_lock);
1184 atomic_inc(&target->obd_req_replay_clients);
1185 atomic_inc(&target->obd_lock_replay_clients);
1186 if (atomic_inc_return(&target->obd_connected_clients) ==
1187 target->obd_max_recoverable_clients)
1188 wake_up(&target->obd_next_transno_waitq);
1191 /* Tell the client we're in recovery, when client is involved in it. */
1192 if (target->obd_recovering && !lw_client)
1193 lustre_msg_add_op_flags(req->rq_repmsg, MSG_CONNECT_RECOVERING);
1195 tmp = req_capsule_client_get(&req->rq_pill, &RMF_CONN);
1198 /* Return -ENOTCONN in case of errors to let client reconnect. */
1199 revimp = class_new_import(target);
1200 if (revimp == NULL) {
1201 CERROR("fail to alloc new reverse import.\n");
1202 GOTO(out, rc = -ENOTCONN);
1205 spin_lock(&export->exp_lock);
1206 if (export->exp_imp_reverse != NULL)
1207 /* destroyed import can be still referenced in ctxt */
1208 tmp_imp = export->exp_imp_reverse;
1209 export->exp_imp_reverse = revimp;
1210 spin_unlock(&export->exp_lock);
1212 revimp->imp_connection = ptlrpc_connection_addref(export->exp_connection);
1213 revimp->imp_client = &export->exp_obd->obd_ldlm_client;
1214 revimp->imp_remote_handle = conn;
1215 revimp->imp_dlm_fake = 1;
1216 revimp->imp_state = LUSTRE_IMP_FULL;
1218 /* Unknown versions will be caught in
1219 * ptlrpc_handle_server_req_in->lustre_unpack_msg(). */
1220 revimp->imp_msg_magic = req->rq_reqmsg->lm_magic;
1222 if ((data->ocd_connect_flags & OBD_CONNECT_AT) &&
1223 (revimp->imp_msg_magic != LUSTRE_MSG_MAGIC_V1))
1224 revimp->imp_msghdr_flags |= MSGHDR_AT_SUPPORT;
1226 revimp->imp_msghdr_flags &= ~MSGHDR_AT_SUPPORT;
1228 if ((data->ocd_connect_flags & OBD_CONNECT_FULL20) &&
1229 (revimp->imp_msg_magic != LUSTRE_MSG_MAGIC_V1))
1230 revimp->imp_msghdr_flags |= MSGHDR_CKSUM_INCOMPAT18;
1232 revimp->imp_msghdr_flags &= ~MSGHDR_CKSUM_INCOMPAT18;
1234 rc = sptlrpc_import_sec_adapt(revimp, req->rq_svc_ctx, &req->rq_flvr);
1236 CERROR("Failed to get sec for reverse import: %d\n", rc);
1237 spin_lock(&export->exp_lock);
1238 export->exp_imp_reverse = NULL;
1239 spin_unlock(&export->exp_lock);
1240 class_destroy_import(revimp);
1243 class_import_put(revimp);
1246 if (tmp_imp != NULL)
1247 client_destroy_import(tmp_imp);
1249 spin_lock(&export->exp_lock);
1250 export->exp_connecting = 0;
1251 spin_unlock(&export->exp_lock);
1253 class_export_put(export);
1256 spin_lock(&target->obd_dev_lock);
1257 target->obd_conn_inprogress--;
1258 spin_unlock(&target->obd_dev_lock);
1260 class_decref(targref, __func__, current);
1263 req->rq_status = rc;
1266 EXPORT_SYMBOL(target_handle_connect);
1268 int target_handle_disconnect(struct ptlrpc_request *req)
1273 rc = req_capsule_server_pack(&req->rq_pill);
1277 /* Keep the rq_export around so we can send the reply. */
1278 req->rq_status = obd_disconnect(class_export_get(req->rq_export));
1282 EXPORT_SYMBOL(target_handle_disconnect);
1284 void target_destroy_export(struct obd_export *exp)
1286 struct obd_import *imp = NULL;
1287 /* exports created from last_rcvd data, and "fake"
1288 exports created by lctl don't have an import */
1289 spin_lock(&exp->exp_lock);
1290 if (exp->exp_imp_reverse != NULL) {
1291 imp = exp->exp_imp_reverse;
1292 exp->exp_imp_reverse = NULL;
1294 spin_unlock(&exp->exp_lock);
1296 client_destroy_import(imp);
1298 LASSERT_ATOMIC_ZERO(&exp->exp_locks_count);
1299 LASSERT_ATOMIC_ZERO(&exp->exp_rpc_count);
1300 LASSERT_ATOMIC_ZERO(&exp->exp_cb_count);
1301 LASSERT_ATOMIC_ZERO(&exp->exp_replay_count);
1303 EXPORT_SYMBOL(target_destroy_export);
1306 * Recovery functions
1308 static void target_request_copy_get(struct ptlrpc_request *req)
1310 class_export_rpc_inc(req->rq_export);
1311 LASSERT(list_empty(&req->rq_list));
1312 INIT_LIST_HEAD(&req->rq_replay_list);
1314 /* Increase refcount to keep request in queue. */
1315 atomic_inc(&req->rq_refcount);
1316 /* Let export know it has replays to be handled. */
1317 atomic_inc(&req->rq_export->exp_replay_count);
1320 static void target_request_copy_put(struct ptlrpc_request *req)
1322 LASSERT(list_empty(&req->rq_replay_list));
1323 LASSERT_ATOMIC_POS(&req->rq_export->exp_replay_count);
1325 atomic_dec(&req->rq_export->exp_replay_count);
1326 class_export_rpc_dec(req->rq_export);
1327 ptlrpc_server_drop_request(req);
1330 static int target_exp_enqueue_req_replay(struct ptlrpc_request *req)
1332 __u64 transno = lustre_msg_get_transno(req->rq_reqmsg);
1333 struct obd_export *exp = req->rq_export;
1334 struct ptlrpc_request *reqiter;
1339 spin_lock(&exp->exp_lock);
1340 list_for_each_entry(reqiter, &exp->exp_req_replay_queue,
1342 if (lustre_msg_get_transno(reqiter->rq_reqmsg) == transno) {
1349 /* We expect it with RESENT and REPLAY flags. */
1350 if ((lustre_msg_get_flags(req->rq_reqmsg) &
1351 (MSG_RESENT | MSG_REPLAY)) != (MSG_RESENT | MSG_REPLAY))
1352 CERROR("invalid flags %x of resent replay\n",
1353 lustre_msg_get_flags(req->rq_reqmsg));
1355 list_add_tail(&req->rq_replay_list,
1356 &exp->exp_req_replay_queue);
1359 spin_unlock(&exp->exp_lock);
1363 static void target_exp_dequeue_req_replay(struct ptlrpc_request *req)
1365 LASSERT(!list_empty(&req->rq_replay_list));
1366 LASSERT(req->rq_export);
1368 spin_lock(&req->rq_export->exp_lock);
1369 list_del_init(&req->rq_replay_list);
1370 spin_unlock(&req->rq_export->exp_lock);
1373 static void target_finish_recovery(struct obd_device *obd)
1377 /* Only log a recovery message when recovery has occurred. */
1378 if (obd->obd_recovery_start) {
1379 time_t elapsed_time = max_t(time_t, 1, cfs_time_current_sec() -
1380 obd->obd_recovery_start);
1381 LCONSOLE_INFO("%s: Recovery over after %d:%.02d, of %d clients "
1382 "%d recovered and %d %s evicted.\n", obd->obd_name,
1383 (int)elapsed_time / 60, (int)elapsed_time % 60,
1384 obd->obd_max_recoverable_clients,
1385 atomic_read(&obd->obd_connected_clients),
1386 obd->obd_stale_clients,
1387 obd->obd_stale_clients == 1 ? "was" : "were");
1390 ldlm_reprocess_all_ns(obd->obd_namespace);
1391 spin_lock(&obd->obd_recovery_task_lock);
1392 if (!list_empty(&obd->obd_req_replay_queue) ||
1393 !list_empty(&obd->obd_lock_replay_queue) ||
1394 !list_empty(&obd->obd_final_req_queue)) {
1395 CERROR("%s: Recovery queues ( %s%s%s) are not empty\n",
1397 list_empty(&obd->obd_req_replay_queue) ? "" : "req ",
1398 list_empty(&obd->obd_lock_replay_queue) ? \
1400 list_empty(&obd->obd_final_req_queue) ? \
1402 spin_unlock(&obd->obd_recovery_task_lock);
1405 spin_unlock(&obd->obd_recovery_task_lock);
1407 obd->obd_recovery_end = cfs_time_current_sec();
1409 /* When recovery finished, cleanup orphans on MDS and OST. */
1410 if (OBT(obd) && OBP(obd, postrecov)) {
1411 int rc = OBP(obd, postrecov)(obd);
1413 LCONSOLE_WARN("%s: Post recovery failed, rc %d\n",
1419 static void abort_req_replay_queue(struct obd_device *obd)
1421 struct ptlrpc_request *req, *n;
1422 struct list_head abort_list;
1424 INIT_LIST_HEAD(&abort_list);
1425 spin_lock(&obd->obd_recovery_task_lock);
1426 list_splice_init(&obd->obd_req_replay_queue, &abort_list);
1427 spin_unlock(&obd->obd_recovery_task_lock);
1428 list_for_each_entry_safe(req, n, &abort_list, rq_list) {
1429 DEBUG_REQ(D_WARNING, req, "aborted:");
1430 req->rq_status = -ENOTCONN;
1431 if (ptlrpc_error(req)) {
1432 DEBUG_REQ(D_ERROR, req,
1433 "failed abort_req_reply; skipping");
1435 target_exp_dequeue_req_replay(req);
1436 target_request_copy_put(req);
1440 static void abort_lock_replay_queue(struct obd_device *obd)
1442 struct ptlrpc_request *req, *n;
1443 struct list_head abort_list;
1445 INIT_LIST_HEAD(&abort_list);
1446 spin_lock(&obd->obd_recovery_task_lock);
1447 list_splice_init(&obd->obd_lock_replay_queue, &abort_list);
1448 spin_unlock(&obd->obd_recovery_task_lock);
1449 list_for_each_entry_safe(req, n, &abort_list, rq_list) {
1450 DEBUG_REQ(D_ERROR, req, "aborted:");
1451 req->rq_status = -ENOTCONN;
1452 if (ptlrpc_error(req)) {
1453 DEBUG_REQ(D_ERROR, req,
1454 "failed abort_lock_reply; skipping");
1456 target_request_copy_put(req);
1460 /* Called from a cleanup function if the device is being cleaned up
1461 forcefully. The exports should all have been disconnected already,
1462 the only thing left to do is
1463 - clear the recovery flags
1465 - free queued requests and replies, but don't send replies
1466 Because the obd_stopping flag is set, no new requests should be received.
1469 void target_cleanup_recovery(struct obd_device *obd)
1471 struct ptlrpc_request *req, *n;
1472 struct list_head clean_list;
1475 INIT_LIST_HEAD(&clean_list);
1476 spin_lock(&obd->obd_dev_lock);
1477 if (!obd->obd_recovering) {
1478 spin_unlock(&obd->obd_dev_lock);
1482 obd->obd_recovering = obd->obd_abort_recovery = 0;
1483 spin_unlock(&obd->obd_dev_lock);
1485 spin_lock(&obd->obd_recovery_task_lock);
1486 target_cancel_recovery_timer(obd);
1487 list_splice_init(&obd->obd_req_replay_queue, &clean_list);
1488 spin_unlock(&obd->obd_recovery_task_lock);
1490 list_for_each_entry_safe(req, n, &clean_list, rq_list) {
1491 LASSERT(req->rq_reply_state == 0);
1492 target_exp_dequeue_req_replay(req);
1493 target_request_copy_put(req);
1496 spin_lock(&obd->obd_recovery_task_lock);
1497 list_splice_init(&obd->obd_lock_replay_queue, &clean_list);
1498 list_splice_init(&obd->obd_final_req_queue, &clean_list);
1499 spin_unlock(&obd->obd_recovery_task_lock);
1501 list_for_each_entry_safe(req, n, &clean_list, rq_list) {
1502 LASSERT(req->rq_reply_state == 0);
1503 target_request_copy_put(req);
1508 EXPORT_SYMBOL(target_cleanup_recovery);
1510 /* obd_recovery_task_lock should be held */
1511 void target_cancel_recovery_timer(struct obd_device *obd)
1513 CDEBUG(D_HA, "%s: cancel recovery timer\n", obd->obd_name);
1514 cfs_timer_disarm(&obd->obd_recovery_timer);
1516 EXPORT_SYMBOL(target_cancel_recovery_timer);
1518 static void target_start_recovery_timer(struct obd_device *obd)
1520 if (obd->obd_recovery_start != 0)
1523 spin_lock(&obd->obd_dev_lock);
1524 if (!obd->obd_recovering || obd->obd_abort_recovery) {
1525 spin_unlock(&obd->obd_dev_lock);
1529 LASSERT(obd->obd_recovery_timeout != 0);
1531 if (obd->obd_recovery_start != 0) {
1532 spin_unlock(&obd->obd_dev_lock);
1536 cfs_timer_arm(&obd->obd_recovery_timer,
1537 cfs_time_shift(obd->obd_recovery_timeout));
1538 obd->obd_recovery_start = cfs_time_current_sec();
1539 spin_unlock(&obd->obd_dev_lock);
1541 LCONSOLE_WARN("%s: Will be in recovery for at least %d:%.02d, "
1542 "or until %d client%s reconnect%s\n",
1544 obd->obd_recovery_timeout / 60,
1545 obd->obd_recovery_timeout % 60,
1546 obd->obd_max_recoverable_clients,
1547 (obd->obd_max_recoverable_clients == 1) ? "" : "s",
1548 (obd->obd_max_recoverable_clients == 1) ? "s": "");
1552 * extend recovery window.
1554 * if @extend is true, extend recovery window to have @drt remaining at least;
1555 * otherwise, make sure the recovery timeout value is not less than @drt.
1557 static void extend_recovery_timer(struct obd_device *obd, int drt, bool extend)
1561 cfs_duration_t left;
1564 spin_lock(&obd->obd_dev_lock);
1565 if (!obd->obd_recovering || obd->obd_abort_recovery) {
1566 spin_unlock(&obd->obd_dev_lock);
1569 LASSERT(obd->obd_recovery_start != 0);
1571 now = cfs_time_current_sec();
1572 to = obd->obd_recovery_timeout;
1573 end = obd->obd_recovery_start + to;
1574 left = cfs_time_sub(end, now);
1576 if (extend && (drt > left)) {
1578 } else if (!extend && (drt > to)) {
1580 /* reduce drt by already passed time */
1581 drt -= obd->obd_recovery_timeout - left;
1584 if (to > obd->obd_recovery_time_hard)
1585 to = obd->obd_recovery_time_hard;
1586 if (obd->obd_recovery_timeout < to ||
1587 obd->obd_recovery_timeout == obd->obd_recovery_time_hard) {
1588 obd->obd_recovery_timeout = to;
1589 cfs_timer_arm(&obd->obd_recovery_timer,
1590 cfs_time_shift(drt));
1592 spin_unlock(&obd->obd_dev_lock);
1594 CDEBUG(D_HA, "%s: recovery timer will expire in %u seconds\n",
1595 obd->obd_name, (unsigned)drt);
1598 /* Reset the timer with each new client connection */
1600 * This timer is actually reconnect_timer, which is for making sure
1601 * the total recovery window is at least as big as my reconnect
1602 * attempt timing. So the initial recovery time_out will be set to
1603 * OBD_RECOVERY_FACTOR * obd_timeout. If the timeout coming
1604 * from client is bigger than this, then the recovery time_out will
1605 * be extended to make sure the client could be reconnected, in the
1606 * process, the timeout from the new client should be ignored.
1610 check_and_start_recovery_timer(struct obd_device *obd,
1611 struct ptlrpc_request *req,
1614 int service_time = lustre_msg_get_service_time(req->rq_reqmsg);
1615 struct obd_device_target *obt = &obd->u.obt;
1617 if (!new_client && service_time)
1618 /* Teach server about old server's estimates, as first guess
1619 * at how long new requests will take. */
1620 at_measured(&req->rq_rqbd->rqbd_svcpt->scp_at_estimate,
1623 target_start_recovery_timer(obd);
1625 /* Convert the service time to RPC timeout,
1626 * and reuse service_time to limit stack usage. */
1627 service_time = at_est2timeout(service_time);
1629 /* We expect other clients to timeout within service_time, then try
1630 * to reconnect, then try the failover server. The max delay between
1631 * connect attempts is SWITCH_MAX + SWITCH_INC + INITIAL. */
1632 service_time += 2 * INITIAL_CONNECT_TIMEOUT;
1634 LASSERT(obt->obt_magic == OBT_MAGIC);
1635 service_time += 2 * (CONNECTION_SWITCH_MAX + CONNECTION_SWITCH_INC);
1636 if (service_time > obd->obd_recovery_timeout && !new_client)
1637 extend_recovery_timer(obd, service_time, false);
1640 /** Health checking routines */
1641 static inline int exp_connect_healthy(struct obd_export *exp)
1643 return (exp->exp_in_recovery);
1646 /** if export done req_replay or has replay in queue */
1647 static inline int exp_req_replay_healthy(struct obd_export *exp)
1649 return (!exp->exp_req_replay_needed ||
1650 atomic_read(&exp->exp_replay_count) > 0);
1652 /** if export done lock_replay or has replay in queue */
1653 static inline int exp_lock_replay_healthy(struct obd_export *exp)
1655 return (!exp->exp_lock_replay_needed ||
1656 atomic_read(&exp->exp_replay_count) > 0);
1659 static inline int exp_vbr_healthy(struct obd_export *exp)
1661 return (!exp->exp_vbr_failed);
1664 static inline int exp_finished(struct obd_export *exp)
1666 return (exp->exp_in_recovery && !exp->exp_lock_replay_needed);
1669 /** Checking routines for recovery */
1670 static int check_for_clients(struct obd_device *obd)
1672 unsigned int clnts = atomic_read(&obd->obd_connected_clients);
1674 if (obd->obd_abort_recovery || obd->obd_recovery_expired)
1676 LASSERT(clnts <= obd->obd_max_recoverable_clients);
1677 return (clnts + obd->obd_stale_clients ==
1678 obd->obd_max_recoverable_clients);
1681 static int check_for_next_transno(struct obd_device *obd)
1683 struct ptlrpc_request *req = NULL;
1684 int wake_up = 0, connected, completed, queue_len;
1685 __u64 next_transno, req_transno;
1688 spin_lock(&obd->obd_recovery_task_lock);
1689 if (!list_empty(&obd->obd_req_replay_queue)) {
1690 req = list_entry(obd->obd_req_replay_queue.next,
1691 struct ptlrpc_request, rq_list);
1692 req_transno = lustre_msg_get_transno(req->rq_reqmsg);
1697 connected = atomic_read(&obd->obd_connected_clients);
1698 completed = connected - atomic_read(&obd->obd_req_replay_clients);
1699 queue_len = obd->obd_requests_queued_for_recovery;
1700 next_transno = obd->obd_next_recovery_transno;
1702 CDEBUG(D_HA, "max: %d, connected: %d, completed: %d, queue_len: %d, "
1703 "req_transno: "LPU64", next_transno: "LPU64"\n",
1704 obd->obd_max_recoverable_clients, connected, completed,
1705 queue_len, req_transno, next_transno);
1707 if (obd->obd_abort_recovery) {
1708 CDEBUG(D_HA, "waking for aborted recovery\n");
1710 } else if (obd->obd_recovery_expired) {
1711 CDEBUG(D_HA, "waking for expired recovery\n");
1713 } else if (req_transno == next_transno) {
1714 CDEBUG(D_HA, "waking for next ("LPD64")\n", next_transno);
1716 } else if (queue_len > 0 &&
1717 queue_len == atomic_read(&obd->obd_req_replay_clients)) {
1719 /** handle gaps occured due to lost reply or VBR */
1720 LASSERTF(req_transno >= next_transno,
1721 "req_transno: "LPU64", next_transno: "LPU64"\n",
1722 req_transno, next_transno);
1723 if (req_transno > obd->obd_last_committed &&
1724 !obd->obd_version_recov)
1727 "%s: waking for gap in transno, VBR is %s (skip: "
1728 LPD64", ql: %d, comp: %d, conn: %d, next: "LPD64
1729 ", last_committed: "LPD64")\n",
1730 obd->obd_name, obd->obd_version_recov ? "ON" : "OFF",
1731 next_transno, queue_len, completed, connected,
1732 req_transno, obd->obd_last_committed);
1733 obd->obd_next_recovery_transno = req_transno;
1735 } else if (atomic_read(&obd->obd_req_replay_clients) == 0) {
1736 CDEBUG(D_HA, "waking for completed recovery\n");
1738 } else if (OBD_FAIL_CHECK(OBD_FAIL_MDS_RECOVERY_ACCEPTS_GAPS)) {
1739 CDEBUG(D_HA, "accepting transno gaps is explicitly allowed"
1740 " by fail_lock, waking up ("LPD64")\n", next_transno);
1741 obd->obd_next_recovery_transno = req_transno;
1744 spin_unlock(&obd->obd_recovery_task_lock);
1748 static int check_for_next_lock(struct obd_device *obd)
1752 spin_lock(&obd->obd_recovery_task_lock);
1753 if (!list_empty(&obd->obd_lock_replay_queue)) {
1754 CDEBUG(D_HA, "waking for next lock\n");
1756 } else if (atomic_read(&obd->obd_lock_replay_clients) == 0) {
1757 CDEBUG(D_HA, "waking for completed lock replay\n");
1759 } else if (obd->obd_abort_recovery) {
1760 CDEBUG(D_HA, "waking for aborted recovery\n");
1762 } else if (obd->obd_recovery_expired) {
1763 CDEBUG(D_HA, "waking for expired recovery\n");
1766 spin_unlock(&obd->obd_recovery_task_lock);
1772 * wait for recovery events,
1773 * check its status with help of check_routine
1774 * evict dead clients via health_check
1776 static int target_recovery_overseer(struct obd_device *obd,
1777 int (*check_routine)(struct obd_device *),
1778 int (*health_check)(struct obd_export *))
1781 wait_event(obd->obd_next_transno_waitq, check_routine(obd));
1782 if (obd->obd_abort_recovery) {
1783 CWARN("recovery is aborted, evict exports in recovery\n");
1784 /** evict exports which didn't finish recovery yet */
1785 class_disconnect_stale_exports(obd, exp_finished);
1787 } else if (obd->obd_recovery_expired) {
1788 obd->obd_recovery_expired = 0;
1789 /** If some clients died being recovered, evict them */
1790 LCONSOLE_WARN("%s: recovery is timed out, "
1791 "evict stale exports\n", obd->obd_name);
1792 /** evict cexports with no replay in queue, they are stalled */
1793 class_disconnect_stale_exports(obd, health_check);
1794 /** continue with VBR */
1795 spin_lock(&obd->obd_dev_lock);
1796 obd->obd_version_recov = 1;
1797 spin_unlock(&obd->obd_dev_lock);
1799 * reset timer, recovery will proceed with versions now,
1800 * timeout is set just to handle reconnection delays
1802 extend_recovery_timer(obd, RECONNECT_DELAY_MAX, true);
1803 /** Wait for recovery events again, after evicting bad clients */
1809 static struct ptlrpc_request *target_next_replay_req(struct obd_device *obd)
1811 struct ptlrpc_request *req = NULL;
1814 CDEBUG(D_HA, "Waiting for transno "LPD64"\n",
1815 obd->obd_next_recovery_transno);
1817 CFS_FAIL_TIMEOUT(OBD_FAIL_TGT_REPLAY_DELAY2, cfs_fail_val);
1819 if (target_recovery_overseer(obd, check_for_next_transno,
1820 exp_req_replay_healthy)) {
1821 abort_req_replay_queue(obd);
1822 abort_lock_replay_queue(obd);
1825 spin_lock(&obd->obd_recovery_task_lock);
1826 if (!list_empty(&obd->obd_req_replay_queue)) {
1827 req = list_entry(obd->obd_req_replay_queue.next,
1828 struct ptlrpc_request, rq_list);
1829 list_del_init(&req->rq_list);
1830 obd->obd_requests_queued_for_recovery--;
1831 spin_unlock(&obd->obd_recovery_task_lock);
1833 spin_unlock(&obd->obd_recovery_task_lock);
1834 LASSERT(list_empty(&obd->obd_req_replay_queue));
1835 LASSERT(atomic_read(&obd->obd_req_replay_clients) == 0);
1836 /** evict exports failed VBR */
1837 class_disconnect_stale_exports(obd, exp_vbr_healthy);
1842 static struct ptlrpc_request *target_next_replay_lock(struct obd_device *obd)
1844 struct ptlrpc_request *req = NULL;
1846 CDEBUG(D_HA, "Waiting for lock\n");
1847 if (target_recovery_overseer(obd, check_for_next_lock,
1848 exp_lock_replay_healthy))
1849 abort_lock_replay_queue(obd);
1851 spin_lock(&obd->obd_recovery_task_lock);
1852 if (!list_empty(&obd->obd_lock_replay_queue)) {
1853 req = list_entry(obd->obd_lock_replay_queue.next,
1854 struct ptlrpc_request, rq_list);
1855 list_del_init(&req->rq_list);
1856 spin_unlock(&obd->obd_recovery_task_lock);
1858 spin_unlock(&obd->obd_recovery_task_lock);
1859 LASSERT(list_empty(&obd->obd_lock_replay_queue));
1860 LASSERT(atomic_read(&obd->obd_lock_replay_clients) == 0);
1861 /** evict exports failed VBR */
1862 class_disconnect_stale_exports(obd, exp_vbr_healthy);
1867 static struct ptlrpc_request *target_next_final_ping(struct obd_device *obd)
1869 struct ptlrpc_request *req = NULL;
1871 spin_lock(&obd->obd_recovery_task_lock);
1872 if (!list_empty(&obd->obd_final_req_queue)) {
1873 req = list_entry(obd->obd_final_req_queue.next,
1874 struct ptlrpc_request, rq_list);
1875 list_del_init(&req->rq_list);
1876 spin_unlock(&obd->obd_recovery_task_lock);
1877 if (req->rq_export->exp_in_recovery) {
1878 spin_lock(&req->rq_export->exp_lock);
1879 req->rq_export->exp_in_recovery = 0;
1880 spin_unlock(&req->rq_export->exp_lock);
1883 spin_unlock(&obd->obd_recovery_task_lock);
1888 static void handle_recovery_req(struct ptlrpc_thread *thread,
1889 struct ptlrpc_request *req,
1890 svc_handler_t handler)
1895 * export can be evicted during recovery, no need to handle replays for
1896 * it after that, discard such request silently
1898 if (req->rq_export->exp_disconnected)
1901 req->rq_session.lc_thread = thread;
1902 req->rq_svc_thread = thread;
1903 req->rq_svc_thread->t_env->le_ses = &req->rq_session;
1905 /* thread context */
1906 lu_context_enter(&thread->t_env->le_ctx);
1908 lu_context_exit(&thread->t_env->le_ctx);
1910 /* don't reset timer for final stage */
1911 if (!exp_finished(req->rq_export)) {
1912 int to = obd_timeout;
1915 * Add request timeout to the recovery time so next request from
1916 * this client may come in recovery time
1919 struct ptlrpc_service_part *svcpt;
1921 svcpt = req->rq_rqbd->rqbd_svcpt;
1922 /* If the server sent early reply for this request,
1923 * the client will recalculate the timeout according to
1924 * current server estimate service time, so we will
1925 * use the maxium timeout here for waiting the client
1926 * sending the next req */
1927 to = max((int)at_est2timeout(
1928 at_get(&svcpt->scp_at_estimate)),
1929 (int)lustre_msg_get_timeout(req->rq_reqmsg));
1930 /* Add 2 net_latency, one for balance rq_deadline
1931 * (see ptl_send_rpc), one for resend the req to server,
1932 * Note: client will pack net_latency in replay req
1933 * (see ptlrpc_replay_req) */
1934 to += 2 * lustre_msg_get_service_time(req->rq_reqmsg);
1936 extend_recovery_timer(class_exp2obd(req->rq_export), to, true);
1941 static int target_recovery_thread(void *arg)
1943 struct lu_target *lut = arg;
1944 struct obd_device *obd = lut->lut_obd;
1945 struct ptlrpc_request *req;
1946 struct target_recovery_data *trd = &obd->obd_recovery_data;
1947 unsigned long delta;
1949 struct ptlrpc_thread *thread = NULL;
1953 unshare_fs_struct();
1954 OBD_ALLOC_PTR(thread);
1960 OBD_FREE_PTR(thread);
1964 rc = lu_context_init(&env->le_ctx, LCT_MD_THREAD | LCT_DT_THREAD);
1966 OBD_FREE_PTR(thread);
1971 thread->t_env = env;
1972 thread->t_id = -1; /* force filter_iobuf_get/put to use local buffers */
1973 env->le_ctx.lc_thread = thread;
1974 tgt_io_thread_init(thread); /* init thread_big_cache for IO requests */
1975 thread->t_watchdog = NULL;
1977 CDEBUG(D_HA, "%s: started recovery thread pid %d\n", obd->obd_name,
1979 trd->trd_processing_task = current_pid();
1981 spin_lock(&obd->obd_dev_lock);
1982 obd->obd_recovering = 1;
1983 spin_unlock(&obd->obd_dev_lock);
1984 complete(&trd->trd_starting);
1986 /* first of all, we have to know the first transno to replay */
1987 if (target_recovery_overseer(obd, check_for_clients,
1988 exp_connect_healthy)) {
1989 abort_req_replay_queue(obd);
1990 abort_lock_replay_queue(obd);
1993 /* next stage: replay requests */
1995 CDEBUG(D_INFO, "1: request replay stage - %d clients from t"LPU64"\n",
1996 atomic_read(&obd->obd_req_replay_clients),
1997 obd->obd_next_recovery_transno);
1998 while ((req = target_next_replay_req(obd))) {
1999 LASSERT(trd->trd_processing_task == current_pid());
2000 DEBUG_REQ(D_HA, req, "processing t"LPD64" from %s",
2001 lustre_msg_get_transno(req->rq_reqmsg),
2002 libcfs_nid2str(req->rq_peer.nid));
2003 handle_recovery_req(thread, req,
2004 trd->trd_recovery_handler);
2006 * bz18031: increase next_recovery_transno before
2007 * target_request_copy_put() will drop exp_rpc reference
2009 spin_lock(&obd->obd_recovery_task_lock);
2010 obd->obd_next_recovery_transno++;
2011 spin_unlock(&obd->obd_recovery_task_lock);
2012 target_exp_dequeue_req_replay(req);
2013 target_request_copy_put(req);
2014 obd->obd_replayed_requests++;
2018 * The second stage: replay locks
2020 CDEBUG(D_INFO, "2: lock replay stage - %d clients\n",
2021 atomic_read(&obd->obd_lock_replay_clients));
2022 while ((req = target_next_replay_lock(obd))) {
2023 LASSERT(trd->trd_processing_task == current_pid());
2024 DEBUG_REQ(D_HA, req, "processing lock from %s: ",
2025 libcfs_nid2str(req->rq_peer.nid));
2026 handle_recovery_req(thread, req,
2027 trd->trd_recovery_handler);
2028 target_request_copy_put(req);
2029 obd->obd_replayed_locks++;
2033 * The third stage: reply on final pings, at this moment all clients
2034 * must have request in final queue
2036 CDEBUG(D_INFO, "3: final stage - process recovery completion pings\n");
2037 /** Update server last boot epoch */
2038 tgt_boot_epoch_update(lut);
2039 /* We drop recoverying flag to forward all new requests
2040 * to regular mds_handle() since now */
2041 spin_lock(&obd->obd_dev_lock);
2042 obd->obd_recovering = obd->obd_abort_recovery = 0;
2043 spin_unlock(&obd->obd_dev_lock);
2044 spin_lock(&obd->obd_recovery_task_lock);
2045 target_cancel_recovery_timer(obd);
2046 spin_unlock(&obd->obd_recovery_task_lock);
2047 while ((req = target_next_final_ping(obd))) {
2048 LASSERT(trd->trd_processing_task == current_pid());
2049 DEBUG_REQ(D_HA, req, "processing final ping from %s: ",
2050 libcfs_nid2str(req->rq_peer.nid));
2051 handle_recovery_req(thread, req,
2052 trd->trd_recovery_handler);
2053 /* Because the waiting client can not send ping to server,
2054 * so we need refresh the last_request_time, to avoid the
2055 * export is being evicted */
2056 ptlrpc_update_export_timer(req->rq_export, 0);
2057 target_request_copy_put(req);
2060 delta = (jiffies - delta) / HZ;
2061 CDEBUG(D_INFO,"4: recovery completed in %lus - %d/%d reqs/locks\n",
2062 delta, obd->obd_replayed_requests, obd->obd_replayed_locks);
2063 if (delta > OBD_RECOVERY_TIME_SOFT) {
2064 CWARN("too long recovery - read logs\n");
2065 libcfs_debug_dumplog();
2068 target_finish_recovery(obd);
2070 lu_context_fini(&env->le_ctx);
2071 trd->trd_processing_task = 0;
2072 complete(&trd->trd_finishing);
2074 tgt_io_thread_done(thread);
2075 OBD_FREE_PTR(thread);
2080 static int target_start_recovery_thread(struct lu_target *lut,
2081 svc_handler_t handler)
2083 struct obd_device *obd = lut->lut_obd;
2085 struct target_recovery_data *trd = &obd->obd_recovery_data;
2087 memset(trd, 0, sizeof(*trd));
2088 init_completion(&trd->trd_starting);
2089 init_completion(&trd->trd_finishing);
2090 trd->trd_recovery_handler = handler;
2092 if (!IS_ERR(kthread_run(target_recovery_thread,
2093 lut, "tgt_recov"))) {
2094 wait_for_completion(&trd->trd_starting);
2095 LASSERT(obd->obd_recovering != 0);
2103 void target_stop_recovery_thread(struct obd_device *obd)
2105 if (obd->obd_recovery_data.trd_processing_task > 0) {
2106 struct target_recovery_data *trd = &obd->obd_recovery_data;
2107 /** recovery can be done but postrecovery is not yet */
2108 spin_lock(&obd->obd_dev_lock);
2109 if (obd->obd_recovering) {
2110 CERROR("%s: Aborting recovery\n", obd->obd_name);
2111 obd->obd_abort_recovery = 1;
2112 wake_up(&obd->obd_next_transno_waitq);
2114 spin_unlock(&obd->obd_dev_lock);
2115 wait_for_completion(&trd->trd_finishing);
2118 EXPORT_SYMBOL(target_stop_recovery_thread);
2120 void target_recovery_fini(struct obd_device *obd)
2122 class_disconnect_exports(obd);
2123 target_stop_recovery_thread(obd);
2124 target_cleanup_recovery(obd);
2126 EXPORT_SYMBOL(target_recovery_fini);
2128 static void target_recovery_expired(unsigned long castmeharder)
2130 struct obd_device *obd = (struct obd_device *)castmeharder;
2131 CDEBUG(D_HA, "%s: recovery timed out; %d clients are still in recovery"
2132 " after %lds (%d clients connected)\n",
2133 obd->obd_name, atomic_read(&obd->obd_lock_replay_clients),
2134 cfs_time_current_sec()- obd->obd_recovery_start,
2135 atomic_read(&obd->obd_connected_clients));
2137 obd->obd_recovery_expired = 1;
2138 wake_up(&obd->obd_next_transno_waitq);
2141 void target_recovery_init(struct lu_target *lut, svc_handler_t handler)
2143 struct obd_device *obd = lut->lut_obd;
2144 if (obd->obd_max_recoverable_clients == 0) {
2145 /** Update server last boot epoch */
2146 tgt_boot_epoch_update(lut);
2150 CDEBUG(D_HA, "RECOVERY: service %s, %d recoverable clients, "
2151 "last_transno "LPU64"\n", obd->obd_name,
2152 obd->obd_max_recoverable_clients, obd->obd_last_committed);
2153 LASSERT(obd->obd_stopping == 0);
2154 obd->obd_next_recovery_transno = obd->obd_last_committed + 1;
2155 obd->obd_recovery_start = 0;
2156 obd->obd_recovery_end = 0;
2158 cfs_timer_init(&obd->obd_recovery_timer, target_recovery_expired, obd);
2159 target_start_recovery_thread(lut, handler);
2161 EXPORT_SYMBOL(target_recovery_init);
2164 static int target_process_req_flags(struct obd_device *obd,
2165 struct ptlrpc_request *req)
2167 struct obd_export *exp = req->rq_export;
2168 LASSERT(exp != NULL);
2169 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REQ_REPLAY_DONE) {
2170 /* client declares he's ready to replay locks */
2171 spin_lock(&exp->exp_lock);
2172 if (exp->exp_req_replay_needed) {
2173 exp->exp_req_replay_needed = 0;
2174 spin_unlock(&exp->exp_lock);
2176 LASSERT_ATOMIC_POS(&obd->obd_req_replay_clients);
2177 atomic_dec(&obd->obd_req_replay_clients);
2179 spin_unlock(&exp->exp_lock);
2182 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_LOCK_REPLAY_DONE) {
2183 /* client declares he's ready to complete recovery
2184 * so, we put the request on th final queue */
2185 spin_lock(&exp->exp_lock);
2186 if (exp->exp_lock_replay_needed) {
2187 exp->exp_lock_replay_needed = 0;
2188 spin_unlock(&exp->exp_lock);
2190 LASSERT_ATOMIC_POS(&obd->obd_lock_replay_clients);
2191 atomic_dec(&obd->obd_lock_replay_clients);
2193 spin_unlock(&exp->exp_lock);
2199 int target_queue_recovery_request(struct ptlrpc_request *req,
2200 struct obd_device *obd)
2202 __u64 transno = lustre_msg_get_transno(req->rq_reqmsg);
2203 struct ptlrpc_request *reqiter;
2207 if (obd->obd_recovery_data.trd_processing_task == current_pid()) {
2208 /* Processing the queue right now, don't re-add. */
2212 target_process_req_flags(obd, req);
2214 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_LOCK_REPLAY_DONE) {
2215 /* client declares he's ready to complete recovery
2216 * so, we put the request on th final queue */
2217 target_request_copy_get(req);
2218 DEBUG_REQ(D_HA, req, "queue final req");
2219 wake_up(&obd->obd_next_transno_waitq);
2220 spin_lock(&obd->obd_recovery_task_lock);
2221 if (obd->obd_recovering) {
2222 list_add_tail(&req->rq_list,
2223 &obd->obd_final_req_queue);
2225 spin_unlock(&obd->obd_recovery_task_lock);
2226 target_request_copy_put(req);
2227 RETURN(obd->obd_stopping ? -ENOTCONN : 1);
2229 spin_unlock(&obd->obd_recovery_task_lock);
2232 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REQ_REPLAY_DONE) {
2233 /* client declares he's ready to replay locks */
2234 target_request_copy_get(req);
2235 DEBUG_REQ(D_HA, req, "queue lock replay req");
2236 wake_up(&obd->obd_next_transno_waitq);
2237 spin_lock(&obd->obd_recovery_task_lock);
2238 LASSERT(obd->obd_recovering);
2239 /* usually due to recovery abort */
2240 if (!req->rq_export->exp_in_recovery) {
2241 spin_unlock(&obd->obd_recovery_task_lock);
2242 target_request_copy_put(req);
2245 LASSERT(req->rq_export->exp_lock_replay_needed);
2246 list_add_tail(&req->rq_list, &obd->obd_lock_replay_queue);
2247 spin_unlock(&obd->obd_recovery_task_lock);
2251 /* CAVEAT EMPTOR: The incoming request message has been swabbed
2252 * (i.e. buflens etc are in my own byte order), but type-dependent
2253 * buffers (eg mdt_body, ost_body etc) have NOT been swabbed. */
2256 INIT_LIST_HEAD(&req->rq_list);
2257 DEBUG_REQ(D_HA, req, "not queueing");
2261 /* If we're processing the queue, we want don't want to queue this
2264 * Also, if this request has a transno less than the one we're waiting
2265 * for, we should process it now. It could (and currently always will)
2266 * be an open request for a descriptor that was opened some time ago.
2268 * Also, a resent, replayed request that has already been
2269 * handled will pass through here and be processed immediately.
2271 CDEBUG(D_HA, "Next recovery transno: "LPU64
2272 ", current: "LPU64", replaying\n",
2273 obd->obd_next_recovery_transno, transno);
2274 spin_lock(&obd->obd_recovery_task_lock);
2275 if (transno < obd->obd_next_recovery_transno) {
2276 /* Processing the queue right now, don't re-add. */
2277 LASSERT(list_empty(&req->rq_list));
2278 spin_unlock(&obd->obd_recovery_task_lock);
2281 spin_unlock(&obd->obd_recovery_task_lock);
2283 if (OBD_FAIL_CHECK(OBD_FAIL_TGT_REPLAY_DROP))
2286 target_request_copy_get(req);
2287 if (!req->rq_export->exp_in_recovery) {
2288 target_request_copy_put(req);
2291 LASSERT(req->rq_export->exp_req_replay_needed);
2293 if (target_exp_enqueue_req_replay(req)) {
2294 DEBUG_REQ(D_ERROR, req, "dropping resent queued req");
2295 target_request_copy_put(req);
2300 spin_lock(&obd->obd_recovery_task_lock);
2301 LASSERT(obd->obd_recovering);
2302 list_for_each_entry(reqiter, &obd->obd_req_replay_queue, rq_list) {
2303 if (lustre_msg_get_transno(reqiter->rq_reqmsg) > transno) {
2304 list_add_tail(&req->rq_list, &reqiter->rq_list);
2309 if (unlikely(lustre_msg_get_transno(reqiter->rq_reqmsg) ==
2311 DEBUG_REQ(D_ERROR, req, "dropping replay: transno "
2312 "has been claimed by another client");
2313 spin_unlock(&obd->obd_recovery_task_lock);
2314 target_exp_dequeue_req_replay(req);
2315 target_request_copy_put(req);
2321 list_add_tail(&req->rq_list, &obd->obd_req_replay_queue);
2323 obd->obd_requests_queued_for_recovery++;
2324 spin_unlock(&obd->obd_recovery_task_lock);
2325 wake_up(&obd->obd_next_transno_waitq);
2328 EXPORT_SYMBOL(target_queue_recovery_request);
2330 int target_handle_ping(struct ptlrpc_request *req)
2332 obd_ping(req->rq_svc_thread->t_env, req->rq_export);
2333 return req_capsule_server_pack(&req->rq_pill);
2335 EXPORT_SYMBOL(target_handle_ping);
2337 void target_committed_to_req(struct ptlrpc_request *req)
2339 struct obd_export *exp = req->rq_export;
2341 if (!exp->exp_obd->obd_no_transno && req->rq_repmsg != NULL)
2342 lustre_msg_set_last_committed(req->rq_repmsg,
2343 exp->exp_last_committed);
2345 DEBUG_REQ(D_IOCTL, req, "not sending last_committed update (%d/"
2346 "%d)", exp->exp_obd->obd_no_transno,
2347 req->rq_repmsg == NULL);
2349 CDEBUG(D_INFO, "last_committed "LPU64", transno "LPU64", xid "LPU64"\n",
2350 exp->exp_last_committed, req->rq_transno, req->rq_xid);
2352 EXPORT_SYMBOL(target_committed_to_req);
2354 #endif /* HAVE_SERVER_SUPPORT */
2357 * Packs current SLV and Limit into \a req.
2359 int target_pack_pool_reply(struct ptlrpc_request *req)
2361 struct obd_device *obd;
2364 /* Check that we still have all structures alive as this may
2365 * be some late RPC at shutdown time. */
2366 if (unlikely(!req->rq_export || !req->rq_export->exp_obd ||
2367 !exp_connect_lru_resize(req->rq_export))) {
2368 lustre_msg_set_slv(req->rq_repmsg, 0);
2369 lustre_msg_set_limit(req->rq_repmsg, 0);
2373 /* OBD is alive here as export is alive, which we checked above. */
2374 obd = req->rq_export->exp_obd;
2376 read_lock(&obd->obd_pool_lock);
2377 lustre_msg_set_slv(req->rq_repmsg, obd->obd_pool_slv);
2378 lustre_msg_set_limit(req->rq_repmsg, obd->obd_pool_limit);
2379 read_unlock(&obd->obd_pool_lock);
2383 EXPORT_SYMBOL(target_pack_pool_reply);
2385 int target_send_reply_msg(struct ptlrpc_request *req, int rc, int fail_id)
2387 if (OBD_FAIL_CHECK_ORSET(fail_id & ~OBD_FAIL_ONCE, OBD_FAIL_ONCE)) {
2388 DEBUG_REQ(D_ERROR, req, "dropping reply");
2393 DEBUG_REQ(D_NET, req, "processing error (%d)", rc);
2394 req->rq_status = rc;
2395 return (ptlrpc_send_error(req, 1));
2397 DEBUG_REQ(D_NET, req, "sending reply");
2400 return (ptlrpc_send_reply(req, PTLRPC_REPLY_MAYBE_DIFFICULT));
2403 void target_send_reply(struct ptlrpc_request *req, int rc, int fail_id)
2405 struct ptlrpc_service_part *svcpt;
2407 struct ptlrpc_reply_state *rs;
2408 struct obd_export *exp;
2411 if (req->rq_no_reply) {
2416 svcpt = req->rq_rqbd->rqbd_svcpt;
2417 rs = req->rq_reply_state;
2418 if (rs == NULL || !rs->rs_difficult) {
2420 target_send_reply_msg (req, rc, fail_id);
2425 /* must be an export if locks saved */
2426 LASSERT(req->rq_export != NULL);
2427 /* req/reply consistent */
2428 LASSERT(rs->rs_svcpt == svcpt);
2431 LASSERT(!rs->rs_scheduled);
2432 LASSERT(!rs->rs_scheduled_ever);
2433 LASSERT(!rs->rs_handled);
2434 LASSERT(!rs->rs_on_net);
2435 LASSERT(rs->rs_export == NULL);
2436 LASSERT(list_empty(&rs->rs_obd_list));
2437 LASSERT(list_empty(&rs->rs_exp_list));
2439 exp = class_export_get(req->rq_export);
2441 /* disable reply scheduling while I'm setting up */
2442 rs->rs_scheduled = 1;
2444 rs->rs_xid = req->rq_xid;
2445 rs->rs_transno = req->rq_transno;
2446 rs->rs_export = exp;
2447 rs->rs_opc = lustre_msg_get_opc(req->rq_reqmsg);
2449 spin_lock(&exp->exp_uncommitted_replies_lock);
2450 CDEBUG(D_NET, "rs transno = "LPU64", last committed = "LPU64"\n",
2451 rs->rs_transno, exp->exp_last_committed);
2452 if (rs->rs_transno > exp->exp_last_committed) {
2453 /* not committed already */
2454 list_add_tail(&rs->rs_obd_list,
2455 &exp->exp_uncommitted_replies);
2457 spin_unlock(&exp->exp_uncommitted_replies_lock);
2459 spin_lock(&exp->exp_lock);
2460 list_add_tail(&rs->rs_exp_list, &exp->exp_outstanding_replies);
2461 spin_unlock(&exp->exp_lock);
2463 netrc = target_send_reply_msg(req, rc, fail_id);
2465 spin_lock(&svcpt->scp_rep_lock);
2467 atomic_inc(&svcpt->scp_nreps_difficult);
2470 /* error sending: reply is off the net. Also we need +1
2471 * reply ref until ptlrpc_handle_rs() is done
2472 * with the reply state (if the send was successful, there
2473 * would have been +1 ref for the net, which
2474 * reply_out_callback leaves alone) */
2476 ptlrpc_rs_addref(rs);
2479 spin_lock(&rs->rs_lock);
2480 if (rs->rs_transno <= exp->exp_last_committed ||
2481 (!rs->rs_on_net && !rs->rs_no_ack) ||
2482 list_empty(&rs->rs_exp_list) || /* completed already */
2483 list_empty(&rs->rs_obd_list)) {
2484 CDEBUG(D_HA, "Schedule reply immediately\n");
2485 ptlrpc_dispatch_difficult_reply(rs);
2487 list_add(&rs->rs_list, &svcpt->scp_rep_active);
2488 rs->rs_scheduled = 0; /* allow notifier to schedule */
2490 spin_unlock(&rs->rs_lock);
2491 spin_unlock(&svcpt->scp_rep_lock);
2494 EXPORT_SYMBOL(target_send_reply);
2496 ldlm_mode_t lck_compat_array[] = {
2497 [LCK_EX] = LCK_COMPAT_EX,
2498 [LCK_PW] = LCK_COMPAT_PW,
2499 [LCK_PR] = LCK_COMPAT_PR,
2500 [LCK_CW] = LCK_COMPAT_CW,
2501 [LCK_CR] = LCK_COMPAT_CR,
2502 [LCK_NL] = LCK_COMPAT_NL,
2503 [LCK_GROUP] = LCK_COMPAT_GROUP,
2504 [LCK_COS] = LCK_COMPAT_COS,
2508 * Rather arbitrary mapping from LDLM error codes to errno values. This should
2509 * not escape to the user level.
2511 int ldlm_error2errno(ldlm_error_t error)
2517 case ELDLM_LOCK_MATCHED:
2520 case ELDLM_LOCK_CHANGED:
2523 case ELDLM_LOCK_ABORTED:
2526 case ELDLM_LOCK_REPLACED:
2529 case ELDLM_NO_LOCK_DATA:
2532 case ELDLM_NAMESPACE_EXISTS:
2535 case ELDLM_BAD_NAMESPACE:
2539 if (((int)error) < 0) /* cast to signed type */
2540 result = error; /* as ldlm_error_t can be unsigned */
2542 CERROR("Invalid DLM result code: %d\n", error);
2548 EXPORT_SYMBOL(ldlm_error2errno);
2551 * Dual to ldlm_error2errno(): maps errno values back to ldlm_error_t.
2553 ldlm_error_t ldlm_errno2error(int err_no)
2562 error = ELDLM_LOCK_CHANGED;
2565 error = ELDLM_LOCK_ABORTED;
2568 error = ELDLM_LOCK_REPLACED;
2571 error = ELDLM_NO_LOCK_DATA;
2574 error = ELDLM_NAMESPACE_EXISTS;
2577 error = ELDLM_BAD_NAMESPACE;
2584 EXPORT_SYMBOL(ldlm_errno2error);
2586 #if LUSTRE_TRACKS_LOCK_EXP_REFS
2587 void ldlm_dump_export_locks(struct obd_export *exp)
2589 spin_lock(&exp->exp_locks_list_guard);
2590 if (!list_empty(&exp->exp_locks_list)) {
2591 struct ldlm_lock *lock;
2593 CERROR("dumping locks for export %p,"
2594 "ignore if the unmount doesn't hang\n", exp);
2595 list_for_each_entry(lock, &exp->exp_locks_list,
2597 LDLM_ERROR(lock, "lock:");
2599 spin_unlock(&exp->exp_locks_list_guard);
2603 #ifdef HAVE_SERVER_SUPPORT
2604 static int target_bulk_timeout(void *data)
2607 /* We don't fail the connection here, because having the export
2608 * killed makes the (vital) call to commitrw very sad.
2613 static inline char *bulk2type(struct ptlrpc_bulk_desc *desc)
2615 return desc->bd_type == BULK_GET_SINK ? "GET" : "PUT";
2618 int target_bulk_io(struct obd_export *exp, struct ptlrpc_bulk_desc *desc,
2619 struct l_wait_info *lwi)
2621 struct ptlrpc_request *req = desc->bd_req;
2622 time_t start = cfs_time_current_sec();
2628 /* If there is eviction in progress, wait for it to finish. */
2629 if (unlikely(atomic_read(&exp->exp_obd->obd_evict_inprogress))) {
2630 *lwi = LWI_INTR(NULL, NULL);
2631 rc = l_wait_event(exp->exp_obd->obd_evict_inprogress_waitq,
2632 !atomic_read(&exp->exp_obd->
2633 obd_evict_inprogress),
2637 /* Check if client was evicted or reconnected already. */
2638 if (exp->exp_failed ||
2639 exp->exp_conn_cnt > lustre_msg_get_conn_cnt(req->rq_reqmsg)) {
2642 if (desc->bd_type == BULK_PUT_SINK)
2643 rc = sptlrpc_svc_wrap_bulk(req, desc);
2645 rc = ptlrpc_start_bulk_transfer(desc);
2649 DEBUG_REQ(D_ERROR, req, "bulk %s failed: rc %d",
2650 bulk2type(desc), rc);
2654 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_SENDPAGE)) {
2655 ptlrpc_abort_bulk(desc);
2659 /* limit actual bulk transfer to bulk_timeout seconds */
2660 deadline = start + bulk_timeout;
2661 if (deadline > req->rq_deadline)
2662 deadline = req->rq_deadline;
2665 long timeoutl = deadline - cfs_time_current_sec();
2666 cfs_duration_t timeout = timeoutl <= 0 ?
2667 CFS_TICK : cfs_time_seconds(timeoutl);
2669 *lwi = LWI_TIMEOUT_INTERVAL(timeout, cfs_time_seconds(1),
2670 target_bulk_timeout, desc);
2671 rc = l_wait_event(desc->bd_waitq,
2672 !ptlrpc_server_bulk_active(desc) ||
2675 lustre_msg_get_conn_cnt(req->rq_reqmsg),
2677 LASSERT(rc == 0 || rc == -ETIMEDOUT);
2678 /* Wait again if we changed rq_deadline. */
2679 deadline = start + bulk_timeout;
2680 if (deadline > req->rq_deadline)
2681 deadline = req->rq_deadline;
2682 } while ((rc == -ETIMEDOUT) &&
2683 (deadline > cfs_time_current_sec()));
2685 if (rc == -ETIMEDOUT) {
2686 DEBUG_REQ(D_ERROR, req, "timeout on bulk %s after %ld%+lds",
2687 bulk2type(desc), deadline - start,
2688 cfs_time_current_sec() - deadline);
2689 ptlrpc_abort_bulk(desc);
2690 } else if (exp->exp_failed) {
2691 DEBUG_REQ(D_ERROR, req, "Eviction on bulk %s",
2694 ptlrpc_abort_bulk(desc);
2695 } else if (exp->exp_conn_cnt >
2696 lustre_msg_get_conn_cnt(req->rq_reqmsg)) {
2697 DEBUG_REQ(D_ERROR, req, "Reconnect on bulk %s",
2699 /* We don't reply anyway. */
2701 ptlrpc_abort_bulk(desc);
2702 } else if (desc->bd_failure ||
2703 desc->bd_nob_transferred != desc->bd_nob) {
2704 DEBUG_REQ(D_ERROR, req, "%s bulk %s %d(%d)",
2705 desc->bd_failure ? "network error on" : "truncated",
2706 bulk2type(desc), desc->bd_nob_transferred,
2708 /* XXX Should this be a different errno? */
2710 } else if (desc->bd_type == BULK_GET_SINK) {
2711 rc = sptlrpc_svc_unwrap_bulk(req, desc);
2716 EXPORT_SYMBOL(target_bulk_io);
2718 #endif /* HAVE_SERVER_SUPPORT */