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 <linux/kthread.h>
47 #include <libcfs/libcfs.h>
49 #include <obd_class.h>
50 #include <lustre_dlm.h>
51 #include <lustre_net.h>
52 #include <lustre_sec.h>
53 #include "ldlm_internal.h"
55 /* @priority: If non-zero, move the selected connection to the list head.
56 * @create: If zero, only search in existing connections.
58 static int import_set_conn(struct obd_import *imp, struct obd_uuid *uuid,
59 int priority, int create)
61 struct ptlrpc_connection *ptlrpc_conn;
62 struct obd_import_conn *imp_conn = NULL, *item;
66 if (!create && !priority) {
67 CDEBUG(D_HA, "Nothing to do\n");
71 ptlrpc_conn = ptlrpc_uuid_to_connection(uuid);
73 CDEBUG(D_HA, "can't find connection %s\n", uuid->uuid);
78 OBD_ALLOC(imp_conn, sizeof(*imp_conn));
80 GOTO(out_put, rc = -ENOMEM);
84 spin_lock(&imp->imp_lock);
85 list_for_each_entry(item, &imp->imp_conn_list, oic_item) {
86 if (obd_uuid_equals(uuid, &item->oic_uuid)) {
88 list_del(&item->oic_item);
89 list_add(&item->oic_item,
91 item->oic_last_attempt = 0;
93 CDEBUG(D_HA, "imp %p@%s: found existing conn %s%s\n",
94 imp, imp->imp_obd->obd_name, uuid->uuid,
95 (priority ? ", moved to head" : ""));
96 spin_unlock(&imp->imp_lock);
97 GOTO(out_free, rc = 0);
100 /* No existing import connection found for \a uuid. */
102 imp_conn->oic_conn = ptlrpc_conn;
103 imp_conn->oic_uuid = *uuid;
104 imp_conn->oic_last_attempt = 0;
106 list_add(&imp_conn->oic_item, &imp->imp_conn_list);
108 list_add_tail(&imp_conn->oic_item,
109 &imp->imp_conn_list);
110 CDEBUG(D_HA, "imp %p@%s: add connection %s at %s\n",
111 imp, imp->imp_obd->obd_name, uuid->uuid,
112 (priority ? "head" : "tail"));
114 spin_unlock(&imp->imp_lock);
115 GOTO(out_free, rc = -ENOENT);
118 spin_unlock(&imp->imp_lock);
122 OBD_FREE(imp_conn, sizeof(*imp_conn));
124 ptlrpc_connection_put(ptlrpc_conn);
128 int import_set_conn_priority(struct obd_import *imp, struct obd_uuid *uuid)
130 return import_set_conn(imp, uuid, 1, 0);
133 int client_import_add_conn(struct obd_import *imp, struct obd_uuid *uuid,
136 return import_set_conn(imp, uuid, priority, 1);
138 EXPORT_SYMBOL(client_import_add_conn);
140 int client_import_del_conn(struct obd_import *imp, struct obd_uuid *uuid)
142 struct obd_import_conn *imp_conn;
143 struct obd_export *dlmexp;
147 spin_lock(&imp->imp_lock);
148 if (list_empty(&imp->imp_conn_list)) {
149 LASSERT(!imp->imp_connection);
153 list_for_each_entry(imp_conn, &imp->imp_conn_list, oic_item) {
154 if (!obd_uuid_equals(uuid, &imp_conn->oic_uuid))
156 LASSERT(imp_conn->oic_conn);
158 if (imp_conn == imp->imp_conn_current) {
159 LASSERT(imp_conn->oic_conn == imp->imp_connection);
161 if (imp->imp_state != LUSTRE_IMP_CLOSED &&
162 imp->imp_state != LUSTRE_IMP_DISCON) {
163 CERROR("can't remove current connection\n");
164 GOTO(out, rc = -EBUSY);
167 ptlrpc_connection_put(imp->imp_connection);
168 imp->imp_connection = NULL;
170 dlmexp = class_conn2export(&imp->imp_dlm_handle);
171 if (dlmexp && dlmexp->exp_connection) {
172 LASSERT(dlmexp->exp_connection ==
174 ptlrpc_connection_put(dlmexp->exp_connection);
175 dlmexp->exp_connection = NULL;
179 list_del(&imp_conn->oic_item);
180 ptlrpc_connection_put(imp_conn->oic_conn);
181 OBD_FREE(imp_conn, sizeof(*imp_conn));
182 CDEBUG(D_HA, "imp %p@%s: remove connection %s\n",
183 imp, imp->imp_obd->obd_name, uuid->uuid);
188 spin_unlock(&imp->imp_lock);
190 CERROR("connection %s not found\n", uuid->uuid);
193 EXPORT_SYMBOL(client_import_del_conn);
196 * Find conn UUID by peer NID. \a peer is a server NID. This function is used
197 * to find a conn uuid of \a imp which can reach \a peer.
199 int client_import_find_conn(struct obd_import *imp, lnet_nid_t peer,
200 struct obd_uuid *uuid)
202 struct obd_import_conn *conn;
206 spin_lock(&imp->imp_lock);
207 list_for_each_entry(conn, &imp->imp_conn_list, oic_item) {
208 /* Check if conn UUID does have this peer NID. */
209 if (class_check_uuid(&conn->oic_uuid, peer)) {
210 *uuid = conn->oic_uuid;
215 spin_unlock(&imp->imp_lock);
218 EXPORT_SYMBOL(client_import_find_conn);
220 void client_destroy_import(struct obd_import *imp)
222 /* Drop security policy instance after all RPCs have finished/aborted
223 * to let all busy contexts be released. */
224 class_import_get(imp);
225 class_destroy_import(imp);
226 sptlrpc_import_sec_put(imp);
227 class_import_put(imp);
229 EXPORT_SYMBOL(client_destroy_import);
232 * Check whether or not the OSC is on MDT.
235 * setup 0:{fsname}-OSTxxxx-osc[-MDTxxxx] 1:lustre-OST0000_UUID 2:NID
237 * setup 0:{fsname}-OSTxxxx-osc 1:lustre-OST0000_UUID 2:NID
240 static int osc_on_mdt(char *obdname)
244 ptr = strrchr(obdname, '-');
248 if (strncmp(ptr + 1, "MDT", 3) == 0)
254 /* Configure an RPC client OBD device.
259 * 3 - inactive-on-startup
261 int client_obd_setup(struct obd_device *obddev, struct lustre_cfg *lcfg)
263 struct client_obd *cli = &obddev->u.cli;
264 struct obd_import *imp;
265 struct obd_uuid server_uuid;
266 int rq_portal, rp_portal, connect_op;
267 char *name = obddev->obd_type->typ_name;
268 ldlm_ns_type_t ns_type = LDLM_NS_TYPE_UNKNOWN;
272 /* In a more perfect world, we would hang a ptlrpc_client off of
273 * obd_type and just use the values from there. */
274 if (!strcmp(name, LUSTRE_OSC_NAME)) {
275 rq_portal = OST_REQUEST_PORTAL;
276 rp_portal = OSC_REPLY_PORTAL;
277 connect_op = OST_CONNECT;
278 cli->cl_sp_me = LUSTRE_SP_CLI;
279 cli->cl_sp_to = LUSTRE_SP_OST;
280 ns_type = LDLM_NS_TYPE_OSC;
281 } else if (!strcmp(name, LUSTRE_MDC_NAME) ||
282 !strcmp(name, LUSTRE_LWP_NAME)) {
283 rq_portal = MDS_REQUEST_PORTAL;
284 rp_portal = MDC_REPLY_PORTAL;
285 connect_op = MDS_CONNECT;
286 cli->cl_sp_me = LUSTRE_SP_CLI;
287 cli->cl_sp_to = LUSTRE_SP_MDT;
288 ns_type = LDLM_NS_TYPE_MDC;
289 } else if (!strcmp(name, LUSTRE_OSP_NAME)) {
290 if (strstr(lustre_cfg_buf(lcfg, 1), "OST") == NULL) {
291 /* OSP_on_MDT for other MDTs */
292 connect_op = MDS_CONNECT;
293 cli->cl_sp_to = LUSTRE_SP_MDT;
294 ns_type = LDLM_NS_TYPE_MDC;
295 rq_portal = OUT_PORTAL;
297 /* OSP on MDT for OST */
298 connect_op = OST_CONNECT;
299 cli->cl_sp_to = LUSTRE_SP_OST;
300 ns_type = LDLM_NS_TYPE_OSC;
301 rq_portal = OST_REQUEST_PORTAL;
303 rp_portal = OSC_REPLY_PORTAL;
304 cli->cl_sp_me = LUSTRE_SP_CLI;
305 } else if (!strcmp(name, LUSTRE_MGC_NAME)) {
306 rq_portal = MGS_REQUEST_PORTAL;
307 rp_portal = MGC_REPLY_PORTAL;
308 connect_op = MGS_CONNECT;
309 cli->cl_sp_me = LUSTRE_SP_MGC;
310 cli->cl_sp_to = LUSTRE_SP_MGS;
311 cli->cl_flvr_mgc.sf_rpc = SPTLRPC_FLVR_INVALID;
312 ns_type = LDLM_NS_TYPE_MGC;
314 CERROR("unknown client OBD type \"%s\", can't setup\n",
319 if (LUSTRE_CFG_BUFLEN(lcfg, 1) < 1) {
320 CERROR("requires a TARGET UUID\n");
324 if (LUSTRE_CFG_BUFLEN(lcfg, 1) > 37) {
325 CERROR("client UUID must be less than 38 characters\n");
329 if (LUSTRE_CFG_BUFLEN(lcfg, 2) < 1) {
330 CERROR("setup requires a SERVER UUID\n");
334 if (LUSTRE_CFG_BUFLEN(lcfg, 2) > 37) {
335 CERROR("target UUID must be less than 38 characters\n");
339 init_rwsem(&cli->cl_sem);
340 mutex_init(&cli->cl_mgc_mutex);
341 cli->cl_conn_count = 0;
342 memcpy(server_uuid.uuid, lustre_cfg_buf(lcfg, 2),
343 min_t(unsigned int, LUSTRE_CFG_BUFLEN(lcfg, 2),
344 sizeof(server_uuid)));
346 cli->cl_dirty_pages = 0;
347 cli->cl_avail_grant = 0;
348 /* FIXME: Should limit this for the sum of all cl_dirty_max_pages. */
349 /* cl_dirty_max_pages may be changed at connect time in
350 * ptlrpc_connect_interpret(). */
351 client_adjust_max_dirty(cli);
352 INIT_LIST_HEAD(&cli->cl_cache_waiters);
353 INIT_LIST_HEAD(&cli->cl_loi_ready_list);
354 INIT_LIST_HEAD(&cli->cl_loi_hp_ready_list);
355 INIT_LIST_HEAD(&cli->cl_loi_write_list);
356 INIT_LIST_HEAD(&cli->cl_loi_read_list);
357 spin_lock_init(&cli->cl_loi_list_lock);
358 atomic_set(&cli->cl_pending_w_pages, 0);
359 atomic_set(&cli->cl_pending_r_pages, 0);
360 cli->cl_r_in_flight = 0;
361 cli->cl_w_in_flight = 0;
363 spin_lock_init(&cli->cl_read_rpc_hist.oh_lock);
364 spin_lock_init(&cli->cl_write_rpc_hist.oh_lock);
365 spin_lock_init(&cli->cl_read_page_hist.oh_lock);
366 spin_lock_init(&cli->cl_write_page_hist.oh_lock);
367 spin_lock_init(&cli->cl_read_offset_hist.oh_lock);
368 spin_lock_init(&cli->cl_write_offset_hist.oh_lock);
371 INIT_LIST_HEAD(&cli->cl_lru_osc);
372 atomic_set(&cli->cl_lru_shrinkers, 0);
373 atomic_long_set(&cli->cl_lru_busy, 0);
374 atomic_long_set(&cli->cl_lru_in_list, 0);
375 INIT_LIST_HEAD(&cli->cl_lru_list);
376 spin_lock_init(&cli->cl_lru_list_lock);
377 atomic_long_set(&cli->cl_unstable_count, 0);
379 init_waitqueue_head(&cli->cl_destroy_waitq);
380 atomic_set(&cli->cl_destroy_in_flight, 0);
381 #ifdef ENABLE_CHECKSUM
382 /* Turn on checksumming by default. */
383 cli->cl_checksum = 1;
385 * The supported checksum types will be worked out at connect time
386 * Set cl_chksum* to CRC32 for now to avoid returning screwed info
389 cli->cl_cksum_type = cli->cl_supp_cksum_types = OBD_CKSUM_CRC32;
391 atomic_set(&cli->cl_resends, OSC_DEFAULT_RESENDS);
393 /* This value may be reduced at connect time in
394 * ptlrpc_connect_interpret() . We initialize it to only
395 * 1MB until we know what the performance looks like.
396 * In the future this should likely be increased. LU-1431 */
397 cli->cl_max_pages_per_rpc = min_t(int, PTLRPC_MAX_BRW_PAGES,
398 LNET_MTU >> PAGE_CACHE_SHIFT);
400 /* set cl_chunkbits default value to PAGE_CACHE_SHIFT,
401 * it will be updated at OSC connection time. */
402 cli->cl_chunkbits = PAGE_CACHE_SHIFT;
404 if (!strcmp(name, LUSTRE_MDC_NAME)) {
405 cli->cl_max_rpcs_in_flight = OBD_MAX_RIF_DEFAULT;
406 } else if (totalram_pages >> (20 - PAGE_CACHE_SHIFT) <= 128 /* MB */) {
407 cli->cl_max_rpcs_in_flight = 2;
408 } else if (totalram_pages >> (20 - PAGE_CACHE_SHIFT) <= 256 /* MB */) {
409 cli->cl_max_rpcs_in_flight = 3;
410 } else if (totalram_pages >> (20 - PAGE_CACHE_SHIFT) <= 512 /* MB */) {
411 cli->cl_max_rpcs_in_flight = 4;
413 if (osc_on_mdt(obddev->obd_name))
414 cli->cl_max_rpcs_in_flight = OBD_MAX_RIF_MAX;
416 cli->cl_max_rpcs_in_flight = OBD_MAX_RIF_DEFAULT;
420 CERROR("ldlm_get_ref failed: %d\n", rc);
424 ptlrpc_init_client(rq_portal, rp_portal, name,
425 &obddev->obd_ldlm_client);
427 imp = class_new_import(obddev);
429 GOTO(err_ldlm, rc = -ENOENT);
430 imp->imp_client = &obddev->obd_ldlm_client;
431 imp->imp_connect_op = connect_op;
432 memcpy(cli->cl_target_uuid.uuid, lustre_cfg_buf(lcfg, 1),
433 LUSTRE_CFG_BUFLEN(lcfg, 1));
434 class_import_put(imp);
436 rc = client_import_add_conn(imp, &server_uuid, 1);
438 CERROR("can't add initial connection\n");
439 GOTO(err_import, rc);
442 cli->cl_import = imp;
443 /* cli->cl_max_mds_easize updated by mdc_init_ea_size() */
444 cli->cl_max_mds_easize = sizeof(struct lov_mds_md_v3);
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);
470 class_destroy_import(imp);
477 EXPORT_SYMBOL(client_obd_setup);
479 int client_obd_cleanup(struct obd_device *obddev)
483 ldlm_namespace_free_post(obddev->obd_namespace);
484 obddev->obd_namespace = NULL;
486 obd_cleanup_client_import(obddev);
487 LASSERT(obddev->u.cli.cl_import == NULL);
492 EXPORT_SYMBOL(client_obd_cleanup);
494 /* ->o_connect() method for client side (OSC and MDC and MGC) */
495 int client_connect_import(const struct lu_env *env,
496 struct obd_export **exp,
497 struct obd_device *obd, struct obd_uuid *cluuid,
498 struct obd_connect_data *data, void *localdata)
500 struct client_obd *cli = &obd->u.cli;
501 struct obd_import *imp = cli->cl_import;
502 struct obd_connect_data *ocd;
503 struct lustre_handle conn = { 0 };
508 down_write(&cli->cl_sem);
509 if (cli->cl_conn_count > 0)
510 GOTO(out_sem, rc = -EALREADY);
512 rc = class_connect(&conn, obd, cluuid);
516 cli->cl_conn_count++;
517 *exp = class_conn2export(&conn);
519 LASSERT(obd->obd_namespace);
521 imp->imp_dlm_handle = conn;
522 rc = ptlrpc_init_import(imp);
526 ocd = &imp->imp_connect_data;
529 imp->imp_connect_flags_orig = data->ocd_connect_flags;
532 rc = ptlrpc_connect_import(imp);
534 LASSERT (imp->imp_state == LUSTRE_IMP_DISCON);
537 LASSERT(*exp != NULL && (*exp)->exp_connection);
540 LASSERTF((ocd->ocd_connect_flags & data->ocd_connect_flags) ==
541 ocd->ocd_connect_flags, "old "LPX64", new "LPX64"\n",
542 data->ocd_connect_flags, ocd->ocd_connect_flags);
543 data->ocd_connect_flags = ocd->ocd_connect_flags;
546 ptlrpc_pinger_add_import(imp);
552 cli->cl_conn_count--;
553 class_disconnect(*exp);
557 up_write(&cli->cl_sem);
561 EXPORT_SYMBOL(client_connect_import);
563 int client_disconnect_export(struct obd_export *exp)
565 struct obd_device *obd = class_exp2obd(exp);
566 struct client_obd *cli;
567 struct obd_import *imp;
572 CERROR("invalid export for disconnect: exp %p cookie "LPX64"\n",
573 exp, exp ? exp->exp_handle.h_cookie : -1);
578 imp = cli->cl_import;
580 down_write(&cli->cl_sem);
581 CDEBUG(D_INFO, "disconnect %s - %zu\n", obd->obd_name,
584 if (cli->cl_conn_count == 0) {
585 CERROR("disconnecting disconnected device (%s)\n",
587 GOTO(out_disconnect, rc = -EINVAL);
590 cli->cl_conn_count--;
591 if (cli->cl_conn_count != 0)
592 GOTO(out_disconnect, rc = 0);
594 /* Mark import deactivated now, so we don't try to reconnect if any
595 * of the cleanup RPCs fails (e.g. LDLM cancel, etc). We don't
596 * fully deactivate the import, or that would drop all requests. */
597 spin_lock(&imp->imp_lock);
598 imp->imp_deactive = 1;
599 spin_unlock(&imp->imp_lock);
601 /* Some non-replayable imports (MDS's OSCs) are pinged, so just
602 * delete it regardless. (It's safe to delete an import that was
604 (void)ptlrpc_pinger_del_import(imp);
606 if (obd->obd_namespace != NULL) {
607 /* obd_force == local only */
608 ldlm_cli_cancel_unused(obd->obd_namespace, NULL,
609 obd->obd_force ? LCF_LOCAL : 0, NULL);
610 ldlm_namespace_free_prior(obd->obd_namespace, imp, obd->obd_force);
613 /* There's no need to hold sem while disconnecting an import,
614 * and it may actually cause deadlock in GSS. */
615 up_write(&cli->cl_sem);
616 rc = ptlrpc_disconnect_import(imp, 0);
617 down_write(&cli->cl_sem);
619 ptlrpc_invalidate_import(imp);
624 /* Use server style - class_disconnect should be always called for
626 err = class_disconnect(exp);
630 up_write(&cli->cl_sem);
634 EXPORT_SYMBOL(client_disconnect_export);
636 #ifdef HAVE_SERVER_SUPPORT
637 int server_disconnect_export(struct obd_export *exp)
642 /* Disconnect early so that clients can't keep using export. */
643 rc = class_disconnect(exp);
644 /* Close import to avoid sending any requests. */
645 if (exp->exp_imp_reverse)
646 ptlrpc_cleanup_imp(exp->exp_imp_reverse);
648 if (exp->exp_obd->obd_namespace != NULL)
649 ldlm_cancel_locks_for_export(exp);
651 /* complete all outstanding replies */
652 spin_lock(&exp->exp_lock);
653 while (!list_empty(&exp->exp_outstanding_replies)) {
654 struct ptlrpc_reply_state *rs =
655 list_entry(exp->exp_outstanding_replies.next,
656 struct ptlrpc_reply_state, rs_exp_list);
657 struct ptlrpc_service_part *svcpt = rs->rs_svcpt;
659 spin_lock(&svcpt->scp_rep_lock);
661 list_del_init(&rs->rs_exp_list);
662 spin_lock(&rs->rs_lock);
663 ptlrpc_schedule_difficult_reply(rs);
664 spin_unlock(&rs->rs_lock);
666 spin_unlock(&svcpt->scp_rep_lock);
668 spin_unlock(&exp->exp_lock);
672 EXPORT_SYMBOL(server_disconnect_export);
674 /* --------------------------------------------------------------------------
675 * from old lib/target.c
676 * -------------------------------------------------------------------------- */
678 static int target_handle_reconnect(struct lustre_handle *conn,
679 struct obd_export *exp,
680 struct obd_uuid *cluuid)
684 if (exp->exp_connection && exp->exp_imp_reverse) {
685 struct lustre_handle *hdl;
686 struct obd_device *target;
688 hdl = &exp->exp_imp_reverse->imp_remote_handle;
689 target = exp->exp_obd;
691 /* Might be a re-connect after a partition. */
692 if (!memcmp(&conn->cookie, &hdl->cookie, sizeof conn->cookie)) {
693 if (target->obd_recovering) {
694 int timeout = cfs_duration_sec(cfs_time_sub(
696 &target->obd_recovery_timer),
697 cfs_time_current()));
699 LCONSOLE_WARN("%s: Client %s (at %s) reconnect"
700 "ing, waiting for %d clients in recov"
701 "ery for %d:%.02d\n", target->obd_name,
702 obd_uuid2str(&exp->exp_client_uuid),
703 obd_export_nid2str(exp),
704 target->obd_max_recoverable_clients,
705 timeout / 60, timeout % 60);
707 LCONSOLE_WARN("%s: Client %s (at %s) "
708 "reconnecting\n", target->obd_name,
709 obd_uuid2str(&exp->exp_client_uuid),
710 obd_export_nid2str(exp));
713 conn->cookie = exp->exp_handle.h_cookie;
714 /* target_handle_connect() treats EALREADY and
715 * -EALREADY differently. EALREADY means we are
716 * doing a valid reconnect from the same client. */
719 LCONSOLE_WARN("%s: already connected client %s (at %s) "
720 "with handle "LPX64". Rejecting client "
721 "with the same UUID trying to reconnect "
722 "with handle "LPX64"\n", target->obd_name,
723 obd_uuid2str(&exp->exp_client_uuid),
724 obd_export_nid2str(exp),
725 hdl->cookie, conn->cookie);
726 memset(conn, 0, sizeof *conn);
727 /* target_handle_connect() treats EALREADY and
728 * -EALREADY differently. -EALREADY is an error
729 * (same UUID, different handle). */
734 conn->cookie = exp->exp_handle.h_cookie;
735 CDEBUG(D_HA, "connect export for UUID '%s' at %p, cookie "LPX64"\n",
736 cluuid->uuid, exp, conn->cookie);
740 void target_client_add_cb(struct obd_device *obd, __u64 transno, void *cb_data,
743 struct obd_export *exp = cb_data;
745 CDEBUG(D_RPCTRACE, "%s: committing for initial connect of %s\n",
746 obd->obd_name, exp->exp_client_uuid.uuid);
748 spin_lock(&exp->exp_lock);
749 exp->exp_need_sync = 0;
750 spin_unlock(&exp->exp_lock);
751 class_export_cb_put(exp);
755 check_and_start_recovery_timer(struct obd_device *obd,
756 struct ptlrpc_request *req, int new_client);
758 int target_handle_connect(struct ptlrpc_request *req)
760 struct obd_device *target = NULL, *targref = NULL;
761 struct obd_export *export = NULL;
762 struct obd_import *revimp;
763 struct obd_import *tmp_imp = NULL;
764 struct lustre_handle conn;
765 struct lustre_handle *tmp;
766 struct obd_uuid tgtuuid;
767 struct obd_uuid cluuid;
768 struct obd_uuid remote_uuid;
773 bool mds_conn = false, lw_client = false;
774 bool mds_mds_conn = false;
775 bool new_mds_mds_conn = 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_VERSION_CODE < OBD_OCD_VERSION(3, 0, 53, 0)
874 /* Don't allow clients to connect that are using old 1.8 format
875 * protocol conventions (LUSTRE_MSG_MAGIC_v1, !MSGHDR_CKSUM_INCOMPAT18,
876 * ldlm_flock_policy_wire format, MDT_ATTR_xTIME_SET, etc). The
877 * FULL20 flag should be set on all connections since 2.0, but no
878 * longer affects behaviour.
880 * Later this check will be disabled and the flag can be retired
881 * completely once interop with 3.0 is no longer needed.
883 if (!(data->ocd_connect_flags & OBD_CONNECT_FULL20))
884 GOTO(out, rc = -EPROTO);
887 if (lustre_msg_get_op_flags(req->rq_reqmsg) & MSG_CONNECT_LIBCLIENT) {
888 if (data->ocd_version < LUSTRE_VERSION_CODE -
889 LUSTRE_VERSION_ALLOWED_OFFSET ||
890 data->ocd_version > LUSTRE_VERSION_CODE +
891 LUSTRE_VERSION_ALLOWED_OFFSET) {
892 DEBUG_REQ(D_WARNING, req, "Refusing %s (%d.%d.%d.%d) "
893 "libclient connection attempt",
894 data->ocd_version < LUSTRE_VERSION_CODE ?
896 OBD_OCD_VERSION_MAJOR(data->ocd_version),
897 OBD_OCD_VERSION_MINOR(data->ocd_version),
898 OBD_OCD_VERSION_PATCH(data->ocd_version),
899 OBD_OCD_VERSION_FIX(data->ocd_version));
900 data = req_capsule_server_sized_get(&req->rq_pill,
902 offsetof(typeof(*data), ocd_version) +
903 sizeof(data->ocd_version));
905 data->ocd_connect_flags = OBD_CONNECT_VERSION;
906 data->ocd_version = LUSTRE_VERSION_CODE;
908 GOTO(out, rc = -EPROTO);
912 /* Note: lw_client is needed in MDS-MDS failover during update log
913 * processing, so we needs to allow lw_client to be connected at
914 * anytime, instead of only the initial connection */
915 lw_client = (data->ocd_connect_flags & OBD_CONNECT_LIGHTWEIGHT) != 0;
917 if (lustre_msg_get_op_flags(req->rq_reqmsg) & MSG_CONNECT_INITIAL) {
918 mds_conn = (data->ocd_connect_flags & OBD_CONNECT_MDS) != 0;
919 mds_mds_conn = (data->ocd_connect_flags &
920 OBD_CONNECT_MDS_MDS) != 0;
922 /* OBD_CONNECT_MNE_SWAB is defined as OBD_CONNECT_MDS_MDS
923 * for Imperative Recovery connection from MGC to MGS.
925 * Via check OBD_CONNECT_FID, we can distinguish whether
926 * the OBD_CONNECT_MDS_MDS/OBD_CONNECT_MNE_SWAB is from
929 (data->ocd_connect_flags & OBD_CONNECT_MDS_MDS) &&
930 (data->ocd_connect_flags & OBD_CONNECT_FID) &&
931 (data->ocd_connect_flags & OBD_CONNECT_VERSION)) {
932 __u32 major = OBD_OCD_VERSION_MAJOR(data->ocd_version);
933 __u32 minor = OBD_OCD_VERSION_MINOR(data->ocd_version);
934 __u32 patch = OBD_OCD_VERSION_PATCH(data->ocd_version);
936 /* We do not support the MDT-MDT interoperations with
937 * different version MDT because of protocol changes. */
938 if (unlikely(major != LUSTRE_MAJOR ||
939 minor != LUSTRE_MINOR ||
940 abs(patch - LUSTRE_PATCH) > 3)) {
941 LCONSOLE_WARN("%s (%u.%u.%u.%u) refused the "
942 "connection from different version MDT "
943 "(%d.%d.%d.%d) %s %s\n",
944 target->obd_name, LUSTRE_MAJOR,
945 LUSTRE_MINOR, LUSTRE_PATCH, LUSTRE_FIX,
947 OBD_OCD_VERSION_FIX(data->ocd_version),
948 libcfs_nid2str(req->rq_peer.nid), str);
950 GOTO(out, rc = -EPROTO);
955 /* lctl gets a backstage, all-access pass. */
956 if (obd_uuid_equals(&cluuid, &target->obd_uuid))
957 goto dont_check_exports;
959 export = cfs_hash_lookup(target->obd_uuid_hash, &cluuid);
963 /* We've found an export in the hash. */
965 spin_lock(&export->exp_lock);
967 if (export->exp_connecting) { /* bug 9635, et. al. */
968 spin_unlock(&export->exp_lock);
969 LCONSOLE_WARN("%s: Export %p already connecting from %s\n",
970 export->exp_obd->obd_name, export,
971 libcfs_nid2str(req->rq_peer.nid));
972 class_export_put(export);
975 } else if ((mds_conn || lw_client) && export->exp_connection != NULL) {
976 spin_unlock(&export->exp_lock);
977 if (req->rq_peer.nid != export->exp_connection->c_peer.nid)
978 /* MDS or LWP reconnected after failover. */
979 LCONSOLE_WARN("%s: Received %s connection from "
980 "%s, removing former export from %s\n",
981 target->obd_name, mds_conn ? "MDS" : "LWP",
982 libcfs_nid2str(req->rq_peer.nid),
983 libcfs_nid2str(export->exp_connection->c_peer.nid));
985 /* New MDS connection from the same NID. */
986 LCONSOLE_WARN("%s: Received new %s connection from "
987 "%s, removing former export from same NID\n",
988 target->obd_name, mds_conn ? "MDS" : "LWP",
989 libcfs_nid2str(req->rq_peer.nid));
990 class_fail_export(export);
991 class_export_put(export);
994 } else if (export->exp_connection != NULL &&
995 req->rq_peer.nid != export->exp_connection->c_peer.nid &&
996 (lustre_msg_get_op_flags(req->rq_reqmsg) &
997 MSG_CONNECT_INITIAL)) {
998 spin_unlock(&export->exp_lock);
999 /* In MDS failover we have static UUID but NID can change. */
1000 LCONSOLE_WARN("%s: Client %s seen on new nid %s when "
1001 "existing nid %s is already connected\n",
1002 target->obd_name, cluuid.uuid,
1003 libcfs_nid2str(req->rq_peer.nid),
1005 export->exp_connection->c_peer.nid));
1007 class_export_put(export);
1010 export->exp_connecting = 1;
1011 spin_unlock(&export->exp_lock);
1012 LASSERT(export->exp_obd == target);
1014 rc = target_handle_reconnect(&conn, export, &cluuid);
1017 /* If we found an export, we already unlocked. */
1020 OBD_FAIL_TIMEOUT(OBD_FAIL_TGT_DELAY_CONNECT, 2 * obd_timeout);
1021 } else if (req->rq_export == NULL &&
1022 atomic_read(&export->exp_rpc_count) > 0) {
1023 LCONSOLE_WARN("%s: Client %s (at %s) refused connection, "
1024 "still busy with %d references\n",
1025 target->obd_name, cluuid.uuid,
1026 libcfs_nid2str(req->rq_peer.nid),
1027 atomic_read(&export->exp_refcount));
1028 GOTO(out, rc = -EBUSY);
1029 } else if (lustre_msg_get_conn_cnt(req->rq_reqmsg) == 1) {
1030 if (!strstr(cluuid.uuid, "mdt"))
1031 LCONSOLE_WARN("%s: Rejecting reconnect from the "
1032 "known client %s (at %s) because it "
1033 "is indicating it is a new client",
1034 target->obd_name, cluuid.uuid,
1035 libcfs_nid2str(req->rq_peer.nid));
1036 GOTO(out, rc = -EALREADY);
1038 OBD_FAIL_TIMEOUT(OBD_FAIL_TGT_DELAY_RECONNECT, 2 * obd_timeout);
1045 CDEBUG(D_HA, "%s: connection from %s@%s %st"LPU64" exp %p cur %ld last %ld\n",
1046 target->obd_name, cluuid.uuid, libcfs_nid2str(req->rq_peer.nid),
1047 target->obd_recovering ? "recovering/" : "", data->ocd_transno,
1048 export, (long)cfs_time_current_sec(),
1049 export ? (long)export->exp_last_request_time : 0);
1051 /* If this is the first time a client connects, reset the recovery
1052 * timer. Discard lightweight connections which might be local. */
1053 if (!lw_client && rc == 0 && target->obd_recovering)
1054 check_and_start_recovery_timer(target, req, export == NULL);
1056 /* We want to handle EALREADY but *not* -EALREADY from
1057 * target_handle_reconnect(), return reconnection state in a flag. */
1058 if (rc == EALREADY) {
1059 lustre_msg_add_op_flags(req->rq_repmsg, MSG_CONNECT_RECONNECT);
1065 /* Tell the client if we support replayable requests. */
1066 if (target->obd_replayable)
1067 lustre_msg_add_op_flags(req->rq_repmsg, MSG_CONNECT_REPLAYABLE);
1068 client_nid = &req->rq_peer.nid;
1070 if (export == NULL) {
1071 /* allow lightweight connections during recovery */
1072 /* allow "new" MDT to be connected during recovery, since we
1073 * need retrieve recovery update records from it */
1074 if (target->obd_recovering && !lw_client && !mds_mds_conn) {
1076 int c; /* connected */
1077 int i; /* in progress */
1079 int s; /* stale/evicted */
1081 c = atomic_read(&target->obd_connected_clients);
1082 i = atomic_read(&target->obd_lock_replay_clients);
1083 k = target->obd_max_recoverable_clients;
1084 s = target->obd_stale_clients;
1085 t = cfs_timer_deadline(&target->obd_recovery_timer);
1086 t = cfs_time_sub(t, cfs_time_current());
1087 t = cfs_duration_sec(t);
1088 LCONSOLE_WARN("%s: Denying connection for new client %s"
1089 "(at %s), waiting for %d known clients "
1090 "(%d recovered, %d in progress, and %d "
1091 "evicted) to recover in %d:%.02d\n",
1092 target->obd_name, cluuid.uuid,
1093 libcfs_nid2str(req->rq_peer.nid), k,
1094 c - i, i, s, (int)t / 60,
1099 rc = obd_connect(req->rq_svc_thread->t_env,
1100 &export, target, &cluuid, data,
1102 if (mds_conn && OBD_FAIL_CHECK(OBD_FAIL_TGT_RCVG_FLAG))
1103 lustre_msg_add_op_flags(req->rq_repmsg,
1104 MSG_CONNECT_RECOVERING);
1106 conn.cookie = export->exp_handle.h_cookie;
1109 new_mds_mds_conn = true;
1112 rc = obd_reconnect(req->rq_svc_thread->t_env,
1113 export, target, &cluuid, data, client_nid);
1118 LASSERT(target->u.obt.obt_magic == OBT_MAGIC);
1119 data->ocd_instance = target->u.obt.obt_instance;
1121 /* Return only the parts of obd_connect_data that we understand, so the
1122 * client knows that we don't understand the rest. */
1124 tmpsize = req_capsule_get_size(&req->rq_pill, &RMF_CONNECT_DATA,
1126 tmpdata = req_capsule_server_get(&req->rq_pill,
1128 /* Don't use struct assignment here, because the client reply
1129 * buffer may be smaller/larger than the local struct
1130 * obd_connect_data. */
1131 memcpy(tmpdata, data, min(tmpsize, size));
1134 /* If all else goes well, this is our RPC return code. */
1137 lustre_msg_set_handle(req->rq_repmsg, &conn);
1139 /* If the client and the server are the same node, we will already
1140 * have an export that really points to the client's DLM export,
1141 * because we have a shared handles table.
1143 * XXX this will go away when shaver stops sending the "connect" handle
1144 * in the real "remote handle" field of the request --phik 24 Apr 2003
1146 ptlrpc_request_change_export(req, export);
1148 spin_lock(&export->exp_lock);
1149 if (export->exp_conn_cnt >= lustre_msg_get_conn_cnt(req->rq_reqmsg)) {
1150 spin_unlock(&export->exp_lock);
1151 CDEBUG(D_RPCTRACE, "%s: %s already connected at greater "
1152 "or equal conn_cnt: %d >= %d\n",
1153 cluuid.uuid, libcfs_nid2str(req->rq_peer.nid),
1154 export->exp_conn_cnt,
1155 lustre_msg_get_conn_cnt(req->rq_reqmsg));
1157 GOTO(out, rc = -EALREADY);
1159 LASSERT(lustre_msg_get_conn_cnt(req->rq_reqmsg) > 0);
1160 export->exp_conn_cnt = lustre_msg_get_conn_cnt(req->rq_reqmsg);
1162 /* Don't evict liblustre clients for not pinging. */
1163 if (lustre_msg_get_op_flags(req->rq_reqmsg) & MSG_CONNECT_LIBCLIENT) {
1164 export->exp_libclient = 1;
1165 spin_unlock(&export->exp_lock);
1167 spin_lock(&target->obd_dev_lock);
1168 list_del_init(&export->exp_obd_chain_timed);
1169 spin_unlock(&target->obd_dev_lock);
1171 spin_unlock(&export->exp_lock);
1174 if (export->exp_connection != NULL) {
1175 /* Check to see if connection came from another NID. */
1176 if ((export->exp_connection->c_peer.nid != req->rq_peer.nid) &&
1177 !hlist_unhashed(&export->exp_nid_hash))
1178 cfs_hash_del(export->exp_obd->obd_nid_hash,
1179 &export->exp_connection->c_peer.nid,
1180 &export->exp_nid_hash);
1182 ptlrpc_connection_put(export->exp_connection);
1185 export->exp_connection = ptlrpc_connection_get(req->rq_peer,
1188 if (hlist_unhashed(&export->exp_nid_hash)) {
1189 cfs_hash_add(export->exp_obd->obd_nid_hash,
1190 &export->exp_connection->c_peer.nid,
1191 &export->exp_nid_hash);
1194 if (target->obd_recovering && !export->exp_in_recovery && !lw_client) {
1196 __u64 transno = data->ocd_transno;
1198 spin_lock(&export->exp_lock);
1199 /* possible race with class_disconnect_stale_exports,
1200 * export may be already in the eviction process */
1201 if (export->exp_failed) {
1202 spin_unlock(&export->exp_lock);
1203 GOTO(out, rc = -ENODEV);
1205 export->exp_in_recovery = 1;
1206 export->exp_req_replay_needed = 1;
1207 export->exp_lock_replay_needed = 1;
1208 spin_unlock(&export->exp_lock);
1210 has_transno = !!(lustre_msg_get_op_flags(req->rq_reqmsg) &
1211 MSG_CONNECT_TRANSNO);
1212 if (has_transno && transno == 0)
1213 CWARN("Connect with zero transno!\n");
1215 if (has_transno && transno > 0 &&
1216 transno < target->obd_next_recovery_transno &&
1217 transno > target->obd_last_committed) {
1218 /* Another way is to use cmpxchg() to be lock-free. */
1219 spin_lock(&target->obd_recovery_task_lock);
1220 if (transno < target->obd_next_recovery_transno)
1221 target->obd_next_recovery_transno = transno;
1222 spin_unlock(&target->obd_recovery_task_lock);
1225 atomic_inc(&target->obd_req_replay_clients);
1226 atomic_inc(&target->obd_lock_replay_clients);
1227 /* Note: MDS-MDS connection is allowed to be connected during
1228 * recovery, no matter if the exports needs to be recoveried.
1229 * Because we need retrieve updates logs from all other MDTs.
1230 * So if the MDS-MDS export is new, obd_max_recoverable_clients
1231 * also needs to be increased to match other recovery checking
1233 if (new_mds_mds_conn)
1234 target->obd_max_recoverable_clients++;
1235 if (atomic_inc_return(&target->obd_connected_clients) ==
1236 target->obd_max_recoverable_clients)
1237 wake_up(&target->obd_next_transno_waitq);
1240 /* Tell the client we're in recovery, when client is involved in it. */
1241 if (target->obd_recovering && !lw_client)
1242 lustre_msg_add_op_flags(req->rq_repmsg, MSG_CONNECT_RECOVERING);
1244 tmp = req_capsule_client_get(&req->rq_pill, &RMF_CONN);
1247 /* Return -ENOTCONN in case of errors to let client reconnect. */
1248 revimp = class_new_import(target);
1249 if (revimp == NULL) {
1250 CERROR("fail to alloc new reverse import.\n");
1251 GOTO(out, rc = -ENOTCONN);
1254 spin_lock(&export->exp_lock);
1255 if (export->exp_imp_reverse != NULL)
1256 /* destroyed import can be still referenced in ctxt */
1257 tmp_imp = export->exp_imp_reverse;
1258 export->exp_imp_reverse = revimp;
1259 spin_unlock(&export->exp_lock);
1261 revimp->imp_connection = ptlrpc_connection_addref(export->exp_connection);
1262 revimp->imp_client = &export->exp_obd->obd_ldlm_client;
1263 revimp->imp_remote_handle = conn;
1264 revimp->imp_dlm_fake = 1;
1265 revimp->imp_state = LUSTRE_IMP_FULL;
1267 /* Unknown versions will be caught in
1268 * ptlrpc_handle_server_req_in->lustre_unpack_msg(). */
1269 revimp->imp_msg_magic = req->rq_reqmsg->lm_magic;
1271 if (data->ocd_connect_flags & OBD_CONNECT_AT)
1272 revimp->imp_msghdr_flags |= MSGHDR_AT_SUPPORT;
1274 revimp->imp_msghdr_flags &= ~MSGHDR_AT_SUPPORT;
1276 revimp->imp_msghdr_flags |= MSGHDR_CKSUM_INCOMPAT18;
1278 rc = sptlrpc_import_sec_adapt(revimp, req->rq_svc_ctx, &req->rq_flvr);
1280 CERROR("Failed to get sec for reverse import: %d\n", rc);
1281 spin_lock(&export->exp_lock);
1282 export->exp_imp_reverse = NULL;
1283 spin_unlock(&export->exp_lock);
1284 class_destroy_import(revimp);
1287 class_import_put(revimp);
1290 if (tmp_imp != NULL)
1291 client_destroy_import(tmp_imp);
1293 spin_lock(&export->exp_lock);
1294 export->exp_connecting = 0;
1295 spin_unlock(&export->exp_lock);
1297 class_export_put(export);
1300 spin_lock(&target->obd_dev_lock);
1301 target->obd_conn_inprogress--;
1302 spin_unlock(&target->obd_dev_lock);
1304 class_decref(targref, __func__, current);
1307 req->rq_status = rc;
1311 int target_handle_disconnect(struct ptlrpc_request *req)
1316 rc = req_capsule_server_pack(&req->rq_pill);
1320 /* Keep the rq_export around so we can send the reply. */
1321 req->rq_status = obd_disconnect(class_export_get(req->rq_export));
1326 void target_destroy_export(struct obd_export *exp)
1328 struct obd_import *imp = NULL;
1329 /* exports created from last_rcvd data, and "fake"
1330 exports created by lctl don't have an import */
1331 spin_lock(&exp->exp_lock);
1332 if (exp->exp_imp_reverse != NULL) {
1333 imp = exp->exp_imp_reverse;
1334 exp->exp_imp_reverse = NULL;
1336 spin_unlock(&exp->exp_lock);
1338 client_destroy_import(imp);
1340 LASSERT_ATOMIC_ZERO(&exp->exp_locks_count);
1341 LASSERT_ATOMIC_ZERO(&exp->exp_rpc_count);
1342 LASSERT_ATOMIC_ZERO(&exp->exp_cb_count);
1343 LASSERT_ATOMIC_ZERO(&exp->exp_replay_count);
1345 EXPORT_SYMBOL(target_destroy_export);
1348 * Recovery functions
1350 static void target_request_copy_get(struct ptlrpc_request *req)
1352 class_export_rpc_inc(req->rq_export);
1353 LASSERT(list_empty(&req->rq_list));
1354 INIT_LIST_HEAD(&req->rq_replay_list);
1356 /* Increase refcount to keep request in queue. */
1357 atomic_inc(&req->rq_refcount);
1358 /* Let export know it has replays to be handled. */
1359 atomic_inc(&req->rq_export->exp_replay_count);
1362 static void target_request_copy_put(struct ptlrpc_request *req)
1364 LASSERT(list_empty(&req->rq_replay_list));
1365 LASSERT_ATOMIC_POS(&req->rq_export->exp_replay_count);
1367 atomic_dec(&req->rq_export->exp_replay_count);
1368 class_export_rpc_dec(req->rq_export);
1369 ptlrpc_server_drop_request(req);
1372 static int target_exp_enqueue_req_replay(struct ptlrpc_request *req)
1374 __u64 transno = lustre_msg_get_transno(req->rq_reqmsg);
1375 struct obd_export *exp = req->rq_export;
1376 struct ptlrpc_request *reqiter;
1381 spin_lock(&exp->exp_lock);
1382 list_for_each_entry(reqiter, &exp->exp_req_replay_queue,
1384 if (lustre_msg_get_transno(reqiter->rq_reqmsg) == transno) {
1391 /* We expect it with RESENT and REPLAY flags. */
1392 if ((lustre_msg_get_flags(req->rq_reqmsg) &
1393 (MSG_RESENT | MSG_REPLAY)) != (MSG_RESENT | MSG_REPLAY))
1394 CERROR("invalid flags %x of resent replay\n",
1395 lustre_msg_get_flags(req->rq_reqmsg));
1397 list_add_tail(&req->rq_replay_list,
1398 &exp->exp_req_replay_queue);
1401 spin_unlock(&exp->exp_lock);
1405 static void target_exp_dequeue_req_replay(struct ptlrpc_request *req)
1407 LASSERT(!list_empty(&req->rq_replay_list));
1408 LASSERT(req->rq_export);
1410 spin_lock(&req->rq_export->exp_lock);
1411 list_del_init(&req->rq_replay_list);
1412 spin_unlock(&req->rq_export->exp_lock);
1415 static void target_finish_recovery(struct lu_target *lut)
1417 struct obd_device *obd = lut->lut_obd;
1420 /* Only log a recovery message when recovery has occurred. */
1421 if (obd->obd_recovery_start) {
1422 time_t elapsed_time = max_t(time_t, 1, cfs_time_current_sec() -
1423 obd->obd_recovery_start);
1424 LCONSOLE_INFO("%s: Recovery over after %d:%.02d, of %d clients "
1425 "%d recovered and %d %s evicted.\n", obd->obd_name,
1426 (int)elapsed_time / 60, (int)elapsed_time % 60,
1427 obd->obd_max_recoverable_clients,
1428 atomic_read(&obd->obd_connected_clients),
1429 obd->obd_stale_clients,
1430 obd->obd_stale_clients == 1 ? "was" : "were");
1433 ldlm_reprocess_all_ns(obd->obd_namespace);
1434 spin_lock(&obd->obd_recovery_task_lock);
1435 if (!list_empty(&obd->obd_req_replay_queue) ||
1436 !list_empty(&obd->obd_lock_replay_queue) ||
1437 !list_empty(&obd->obd_final_req_queue)) {
1438 CERROR("%s: Recovery queues ( %s%s%s) are not empty\n",
1440 list_empty(&obd->obd_req_replay_queue) ? "" : "req ",
1441 list_empty(&obd->obd_lock_replay_queue) ? \
1443 list_empty(&obd->obd_final_req_queue) ? \
1445 spin_unlock(&obd->obd_recovery_task_lock);
1448 spin_unlock(&obd->obd_recovery_task_lock);
1450 if (lut->lut_tdtd != NULL &&
1451 !list_empty(&lut->lut_tdtd->tdtd_replay_list))
1452 dtrq_list_dump(lut->lut_tdtd, D_ERROR);
1454 obd->obd_recovery_end = cfs_time_current_sec();
1456 /* When recovery finished, cleanup orphans on MDS and OST. */
1457 if (OBT(obd) && OBP(obd, postrecov)) {
1458 int rc = OBP(obd, postrecov)(obd);
1460 LCONSOLE_WARN("%s: Post recovery failed, rc %d\n",
1466 static void abort_req_replay_queue(struct obd_device *obd)
1468 struct ptlrpc_request *req, *n;
1469 struct list_head abort_list;
1471 INIT_LIST_HEAD(&abort_list);
1472 spin_lock(&obd->obd_recovery_task_lock);
1473 list_splice_init(&obd->obd_req_replay_queue, &abort_list);
1474 spin_unlock(&obd->obd_recovery_task_lock);
1475 list_for_each_entry_safe(req, n, &abort_list, rq_list) {
1476 DEBUG_REQ(D_WARNING, req, "aborted:");
1477 req->rq_status = -ENOTCONN;
1478 if (ptlrpc_error(req)) {
1479 DEBUG_REQ(D_ERROR, req,
1480 "failed abort_req_reply; skipping");
1482 target_exp_dequeue_req_replay(req);
1483 target_request_copy_put(req);
1487 static void abort_lock_replay_queue(struct obd_device *obd)
1489 struct ptlrpc_request *req, *n;
1490 struct list_head abort_list;
1492 INIT_LIST_HEAD(&abort_list);
1493 spin_lock(&obd->obd_recovery_task_lock);
1494 list_splice_init(&obd->obd_lock_replay_queue, &abort_list);
1495 spin_unlock(&obd->obd_recovery_task_lock);
1496 list_for_each_entry_safe(req, n, &abort_list, rq_list) {
1497 DEBUG_REQ(D_ERROR, req, "aborted:");
1498 req->rq_status = -ENOTCONN;
1499 if (ptlrpc_error(req)) {
1500 DEBUG_REQ(D_ERROR, req,
1501 "failed abort_lock_reply; skipping");
1503 target_request_copy_put(req);
1507 /* Called from a cleanup function if the device is being cleaned up
1508 forcefully. The exports should all have been disconnected already,
1509 the only thing left to do is
1510 - clear the recovery flags
1512 - free queued requests and replies, but don't send replies
1513 Because the obd_stopping flag is set, no new requests should be received.
1516 void target_cleanup_recovery(struct obd_device *obd)
1518 struct ptlrpc_request *req, *n;
1519 struct list_head clean_list;
1522 INIT_LIST_HEAD(&clean_list);
1523 spin_lock(&obd->obd_dev_lock);
1524 if (!obd->obd_recovering) {
1525 spin_unlock(&obd->obd_dev_lock);
1529 obd->obd_recovering = obd->obd_abort_recovery = 0;
1530 obd->obd_force_abort_recovery = 0;
1531 spin_unlock(&obd->obd_dev_lock);
1533 spin_lock(&obd->obd_recovery_task_lock);
1534 target_cancel_recovery_timer(obd);
1535 list_splice_init(&obd->obd_req_replay_queue, &clean_list);
1536 spin_unlock(&obd->obd_recovery_task_lock);
1538 list_for_each_entry_safe(req, n, &clean_list, rq_list) {
1539 LASSERT(req->rq_reply_state == NULL);
1540 target_exp_dequeue_req_replay(req);
1541 target_request_copy_put(req);
1544 spin_lock(&obd->obd_recovery_task_lock);
1545 list_splice_init(&obd->obd_lock_replay_queue, &clean_list);
1546 list_splice_init(&obd->obd_final_req_queue, &clean_list);
1547 spin_unlock(&obd->obd_recovery_task_lock);
1549 list_for_each_entry_safe(req, n, &clean_list, rq_list) {
1550 LASSERT(req->rq_reply_state == NULL);
1551 target_request_copy_put(req);
1556 EXPORT_SYMBOL(target_cleanup_recovery);
1558 /* obd_recovery_task_lock should be held */
1559 void target_cancel_recovery_timer(struct obd_device *obd)
1561 CDEBUG(D_HA, "%s: cancel recovery timer\n", obd->obd_name);
1562 cfs_timer_disarm(&obd->obd_recovery_timer);
1565 static void target_start_recovery_timer(struct obd_device *obd)
1567 if (obd->obd_recovery_start != 0)
1570 spin_lock(&obd->obd_dev_lock);
1571 if (!obd->obd_recovering || obd->obd_abort_recovery ||
1572 obd->obd_force_abort_recovery) {
1573 spin_unlock(&obd->obd_dev_lock);
1577 LASSERT(obd->obd_recovery_timeout != 0);
1579 if (obd->obd_recovery_start != 0) {
1580 spin_unlock(&obd->obd_dev_lock);
1584 cfs_timer_arm(&obd->obd_recovery_timer,
1585 cfs_time_shift(obd->obd_recovery_timeout));
1586 obd->obd_recovery_start = cfs_time_current_sec();
1587 spin_unlock(&obd->obd_dev_lock);
1589 LCONSOLE_WARN("%s: Will be in recovery for at least %d:%.02d, "
1590 "or until %d client%s reconnect%s\n",
1592 obd->obd_recovery_timeout / 60,
1593 obd->obd_recovery_timeout % 60,
1594 obd->obd_max_recoverable_clients,
1595 (obd->obd_max_recoverable_clients == 1) ? "" : "s",
1596 (obd->obd_max_recoverable_clients == 1) ? "s": "");
1600 * extend recovery window.
1602 * if @extend is true, extend recovery window to have @drt remaining at least;
1603 * otherwise, make sure the recovery timeout value is not less than @drt.
1605 static void extend_recovery_timer(struct obd_device *obd, int drt, bool extend)
1609 cfs_duration_t left;
1612 spin_lock(&obd->obd_dev_lock);
1613 if (!obd->obd_recovering || obd->obd_abort_recovery ||
1614 obd->obd_force_abort_recovery) {
1615 spin_unlock(&obd->obd_dev_lock);
1618 LASSERT(obd->obd_recovery_start != 0);
1620 now = cfs_time_current_sec();
1621 to = obd->obd_recovery_timeout;
1622 end = obd->obd_recovery_start + to;
1623 left = cfs_time_sub(end, now);
1625 if (extend && (drt > left)) {
1627 } else if (!extend && (drt > to)) {
1631 if (to > obd->obd_recovery_time_hard)
1632 to = obd->obd_recovery_time_hard;
1633 if (obd->obd_recovery_timeout < to) {
1634 obd->obd_recovery_timeout = to;
1635 end = obd->obd_recovery_start + to;
1636 cfs_timer_arm(&obd->obd_recovery_timer,
1637 cfs_time_shift(end - now));
1639 spin_unlock(&obd->obd_dev_lock);
1641 CDEBUG(D_HA, "%s: recovery timer will expire in %u seconds\n",
1642 obd->obd_name, (unsigned)cfs_time_sub(end, now));
1645 /* Reset the timer with each new client connection */
1647 * This timer is actually reconnect_timer, which is for making sure
1648 * the total recovery window is at least as big as my reconnect
1649 * attempt timing. So the initial recovery time_out will be set to
1650 * OBD_RECOVERY_FACTOR * obd_timeout. If the timeout coming
1651 * from client is bigger than this, then the recovery time_out will
1652 * be extended to make sure the client could be reconnected, in the
1653 * process, the timeout from the new client should be ignored.
1657 check_and_start_recovery_timer(struct obd_device *obd,
1658 struct ptlrpc_request *req,
1661 int service_time = lustre_msg_get_service_time(req->rq_reqmsg);
1662 struct obd_device_target *obt = &obd->u.obt;
1664 if (!new_client && service_time)
1665 /* Teach server about old server's estimates, as first guess
1666 * at how long new requests will take. */
1667 at_measured(&req->rq_rqbd->rqbd_svcpt->scp_at_estimate,
1670 target_start_recovery_timer(obd);
1672 /* Convert the service time to RPC timeout,
1673 * and reuse service_time to limit stack usage. */
1674 service_time = at_est2timeout(service_time);
1676 /* We expect other clients to timeout within service_time, then try
1677 * to reconnect, then try the failover server. The max delay between
1678 * connect attempts is SWITCH_MAX + SWITCH_INC + INITIAL. */
1679 service_time += 2 * INITIAL_CONNECT_TIMEOUT;
1681 LASSERT(obt->obt_magic == OBT_MAGIC);
1682 service_time += 2 * (CONNECTION_SWITCH_MAX + CONNECTION_SWITCH_INC);
1683 if (service_time > obd->obd_recovery_timeout && !new_client)
1684 extend_recovery_timer(obd, service_time, false);
1687 /** Health checking routines */
1688 static inline int exp_connect_healthy(struct obd_export *exp)
1690 return (exp->exp_in_recovery);
1693 /** if export done req_replay or has replay in queue */
1694 static inline int exp_req_replay_healthy(struct obd_export *exp)
1696 return (!exp->exp_req_replay_needed ||
1697 atomic_read(&exp->exp_replay_count) > 0);
1699 /** if export done lock_replay or has replay in queue */
1700 static inline int exp_lock_replay_healthy(struct obd_export *exp)
1702 return (!exp->exp_lock_replay_needed ||
1703 atomic_read(&exp->exp_replay_count) > 0);
1706 static inline int exp_vbr_healthy(struct obd_export *exp)
1708 return (!exp->exp_vbr_failed);
1711 static inline int exp_finished(struct obd_export *exp)
1713 return (exp->exp_in_recovery && !exp->exp_lock_replay_needed);
1716 static int check_for_next_transno(struct lu_target *lut)
1718 struct ptlrpc_request *req = NULL;
1719 struct obd_device *obd = lut->lut_obd;
1720 int wake_up = 0, connected, completed, queue_len;
1721 __u64 req_transno = 0;
1722 __u64 update_transno = 0;
1723 __u64 next_transno = 0;
1726 spin_lock(&obd->obd_recovery_task_lock);
1727 if (!list_empty(&obd->obd_req_replay_queue)) {
1728 req = list_entry(obd->obd_req_replay_queue.next,
1729 struct ptlrpc_request, rq_list);
1730 req_transno = lustre_msg_get_transno(req->rq_reqmsg);
1733 if (lut->lut_tdtd != NULL) {
1734 struct target_distribute_txn_data *tdtd;
1735 __u64 update_transno;
1737 tdtd = lut->lut_tdtd;
1738 update_transno = distribute_txn_get_next_transno(lut->lut_tdtd);
1741 connected = atomic_read(&obd->obd_connected_clients);
1742 completed = connected - atomic_read(&obd->obd_req_replay_clients);
1743 queue_len = obd->obd_requests_queued_for_recovery;
1744 next_transno = obd->obd_next_recovery_transno;
1746 CDEBUG(D_HA, "max: %d, connected: %d, completed: %d, queue_len: %d, "
1747 "req_transno: "LPU64", next_transno: "LPU64"\n",
1748 obd->obd_max_recoverable_clients, connected, completed,
1749 queue_len, req_transno, next_transno);
1751 if (obd->obd_abort_recovery || obd->obd_force_abort_recovery) {
1752 CDEBUG(D_HA, "waking for aborted recovery\n");
1754 } else if (obd->obd_recovery_expired) {
1755 CDEBUG(D_HA, "waking for expired recovery\n");
1757 } else if (req_transno == next_transno || (update_transno != 0 &&
1758 update_transno <= next_transno)) {
1759 CDEBUG(D_HA, "waking for next ("LPD64")\n", next_transno);
1761 } else if (queue_len > 0 &&
1762 queue_len == atomic_read(&obd->obd_req_replay_clients)) {
1764 /** handle gaps occured due to lost reply or VBR */
1765 LASSERTF(req_transno >= next_transno,
1766 "req_transno: "LPU64", next_transno: "LPU64"\n",
1767 req_transno, next_transno);
1768 if (req_transno > obd->obd_last_committed &&
1769 !obd->obd_version_recov)
1772 "%s: waking for gap in transno, VBR is %s (skip: "
1773 LPD64", ql: %d, comp: %d, conn: %d, next: "LPD64
1774 ", next_update "LPD64" last_committed: "LPD64")\n",
1775 obd->obd_name, obd->obd_version_recov ? "ON" : "OFF",
1776 next_transno, queue_len, completed, connected,
1777 req_transno, update_transno, obd->obd_last_committed);
1778 obd->obd_next_recovery_transno = req_transno;
1780 } else if (atomic_read(&obd->obd_req_replay_clients) == 0) {
1781 CDEBUG(D_HA, "waking for completed recovery\n");
1783 } else if (OBD_FAIL_CHECK(OBD_FAIL_MDS_RECOVERY_ACCEPTS_GAPS)) {
1784 CDEBUG(D_HA, "accepting transno gaps is explicitly allowed"
1785 " by fail_lock, waking up ("LPD64")\n", next_transno);
1786 obd->obd_next_recovery_transno = req_transno;
1789 spin_unlock(&obd->obd_recovery_task_lock);
1793 static int check_for_next_lock(struct lu_target *lut)
1795 struct obd_device *obd = lut->lut_obd;
1798 spin_lock(&obd->obd_recovery_task_lock);
1799 if (!list_empty(&obd->obd_lock_replay_queue)) {
1800 CDEBUG(D_HA, "waking for next lock\n");
1802 } else if (atomic_read(&obd->obd_lock_replay_clients) == 0) {
1803 CDEBUG(D_HA, "waking for completed lock replay\n");
1805 } else if (obd->obd_abort_recovery || obd->obd_force_abort_recovery) {
1806 CDEBUG(D_HA, "waking for aborted recovery\n");
1808 } else if (obd->obd_recovery_expired) {
1809 CDEBUG(D_HA, "waking for expired recovery\n");
1812 spin_unlock(&obd->obd_recovery_task_lock);
1818 * wait for recovery events,
1819 * check its status with help of check_routine
1820 * evict dead clients via health_check
1822 static int target_recovery_overseer(struct lu_target *lut,
1823 int (*check_routine)(struct lu_target *),
1824 int (*health_check)(struct obd_export *))
1826 struct obd_device *obd = lut->lut_obd;
1828 if ((obd->obd_recovery_start != 0) && (cfs_time_current_sec() >=
1829 (obd->obd_recovery_start + obd->obd_recovery_time_hard))) {
1830 CWARN("recovery is aborted by hard timeout\n");
1831 obd->obd_abort_recovery = 1;
1834 while (wait_event_timeout(obd->obd_next_transno_waitq,
1836 msecs_to_jiffies(60 * MSEC_PER_SEC)) == 0)
1837 /* wait indefinitely for event, but don't trigger watchdog */;
1839 if (obd->obd_abort_recovery || obd->obd_force_abort_recovery) {
1840 CWARN("recovery is aborted, evict exports in recovery\n");
1841 /** evict exports which didn't finish recovery yet */
1842 class_disconnect_stale_exports(obd, exp_finished);
1844 } else if (obd->obd_recovery_expired) {
1845 obd->obd_recovery_expired = 0;
1846 /** If some clients died being recovered, evict them */
1847 LCONSOLE_WARN("%s: recovery is timed out, "
1848 "evict stale exports\n", obd->obd_name);
1849 /** evict cexports with no replay in queue, they are stalled */
1850 class_disconnect_stale_exports(obd, health_check);
1851 /** continue with VBR */
1852 spin_lock(&obd->obd_dev_lock);
1853 obd->obd_version_recov = 1;
1854 spin_unlock(&obd->obd_dev_lock);
1856 * reset timer, recovery will proceed with versions now,
1857 * timeout is set just to handle reconnection delays
1859 extend_recovery_timer(obd, RECONNECT_DELAY_MAX, true);
1860 /** Wait for recovery events again, after evicting bad clients */
1866 static struct ptlrpc_request *target_next_replay_lock(struct lu_target *lut)
1868 struct obd_device *obd = lut->lut_obd;
1869 struct ptlrpc_request *req = NULL;
1871 CDEBUG(D_HA, "Waiting for lock\n");
1872 if (target_recovery_overseer(lut, check_for_next_lock,
1873 exp_lock_replay_healthy))
1874 abort_lock_replay_queue(obd);
1876 spin_lock(&obd->obd_recovery_task_lock);
1877 if (!list_empty(&obd->obd_lock_replay_queue)) {
1878 req = list_entry(obd->obd_lock_replay_queue.next,
1879 struct ptlrpc_request, rq_list);
1880 list_del_init(&req->rq_list);
1881 spin_unlock(&obd->obd_recovery_task_lock);
1883 spin_unlock(&obd->obd_recovery_task_lock);
1884 LASSERT(list_empty(&obd->obd_lock_replay_queue));
1885 LASSERT(atomic_read(&obd->obd_lock_replay_clients) == 0);
1886 /** evict exports failed VBR */
1887 class_disconnect_stale_exports(obd, exp_vbr_healthy);
1892 static struct ptlrpc_request *target_next_final_ping(struct obd_device *obd)
1894 struct ptlrpc_request *req = NULL;
1896 spin_lock(&obd->obd_recovery_task_lock);
1897 if (!list_empty(&obd->obd_final_req_queue)) {
1898 req = list_entry(obd->obd_final_req_queue.next,
1899 struct ptlrpc_request, rq_list);
1900 list_del_init(&req->rq_list);
1901 spin_unlock(&obd->obd_recovery_task_lock);
1902 if (req->rq_export->exp_in_recovery) {
1903 spin_lock(&req->rq_export->exp_lock);
1904 req->rq_export->exp_in_recovery = 0;
1905 spin_unlock(&req->rq_export->exp_lock);
1908 spin_unlock(&obd->obd_recovery_task_lock);
1913 static void handle_recovery_req(struct ptlrpc_thread *thread,
1914 struct ptlrpc_request *req,
1915 svc_handler_t handler)
1920 * export can be evicted during recovery, no need to handle replays for
1921 * it after that, discard such request silently
1923 if (req->rq_export->exp_disconnected)
1926 req->rq_session.lc_thread = thread;
1927 req->rq_svc_thread = thread;
1928 req->rq_svc_thread->t_env->le_ses = &req->rq_session;
1930 /* thread context */
1931 lu_context_enter(&thread->t_env->le_ctx);
1933 lu_context_exit(&thread->t_env->le_ctx);
1935 /* don't reset timer for final stage */
1936 if (!exp_finished(req->rq_export)) {
1937 int to = obd_timeout;
1940 * Add request timeout to the recovery time so next request from
1941 * this client may come in recovery time
1944 struct ptlrpc_service_part *svcpt;
1946 svcpt = req->rq_rqbd->rqbd_svcpt;
1947 /* If the server sent early reply for this request,
1948 * the client will recalculate the timeout according to
1949 * current server estimate service time, so we will
1950 * use the maxium timeout here for waiting the client
1951 * sending the next req */
1952 to = max((int)at_est2timeout(
1953 at_get(&svcpt->scp_at_estimate)),
1954 (int)lustre_msg_get_timeout(req->rq_reqmsg));
1955 /* Add 2 net_latency, one for balance rq_deadline
1956 * (see ptl_send_rpc), one for resend the req to server,
1957 * Note: client will pack net_latency in replay req
1958 * (see ptlrpc_replay_req) */
1959 to += 2 * lustre_msg_get_service_time(req->rq_reqmsg);
1961 extend_recovery_timer(class_exp2obd(req->rq_export), to, true);
1966 /** Checking routines for recovery */
1967 static int check_for_recovery_ready(struct lu_target *lut)
1969 struct obd_device *obd = lut->lut_obd;
1970 unsigned int clnts = atomic_read(&obd->obd_connected_clients);
1972 CDEBUG(D_HA, "connected %d stale %d max_recoverable_clients %d"
1973 " abort %d expired %d\n", clnts, obd->obd_stale_clients,
1974 obd->obd_max_recoverable_clients, obd->obd_abort_recovery,
1975 obd->obd_recovery_expired);
1977 if (obd->obd_force_abort_recovery)
1980 if (!obd->obd_abort_recovery && !obd->obd_recovery_expired) {
1981 LASSERT(clnts <= obd->obd_max_recoverable_clients);
1982 if (clnts + obd->obd_stale_clients <
1983 obd->obd_max_recoverable_clients)
1987 if (lut->lut_tdtd != NULL) {
1988 if (!lut->lut_tdtd->tdtd_replay_ready) {
1989 /* Let's extend recovery timer, in case the recovery
1990 * timer expired, and some clients got evicted */
1991 extend_recovery_timer(obd, obd->obd_recovery_timeout,
1995 dtrq_list_dump(lut->lut_tdtd, D_HA);
2003 REQUEST_RECOVERY = 1,
2004 UPDATE_RECOVERY = 2,
2007 static __u64 get_next_replay_req_transno(struct obd_device *obd)
2011 if (!list_empty(&obd->obd_req_replay_queue)) {
2012 struct ptlrpc_request *req;
2014 req = list_entry(obd->obd_req_replay_queue.next,
2015 struct ptlrpc_request, rq_list);
2016 transno = lustre_msg_get_transno(req->rq_reqmsg);
2021 __u64 get_next_transno(struct lu_target *lut, int *type)
2023 struct obd_device *obd = lut->lut_obd;
2024 struct target_distribute_txn_data *tdtd = lut->lut_tdtd;
2026 __u64 update_transno;
2029 transno = get_next_replay_req_transno(obd);
2031 *type = REQUEST_RECOVERY;
2036 update_transno = distribute_txn_get_next_transno(tdtd);
2037 if (transno == 0 || (transno >= update_transno &&
2038 update_transno != 0)) {
2039 transno = update_transno;
2041 *type = UPDATE_RECOVERY;
2048 * drop duplicate replay request
2050 * Because the operation has been replayed by update recovery, the request
2051 * with the same transno will be dropped and also notify the client to send
2052 * next replay request.
2054 * \param[in] env execution environment
2055 * \param[in] obd failover obd device
2056 * \param[in] req request to be dropped
2058 static void drop_duplicate_replay_req(struct lu_env *env,
2059 struct obd_device *obd,
2060 struct ptlrpc_request *req)
2062 DEBUG_REQ(D_HA, req, "remove t"LPD64" from %s because of duplicate"
2063 " update records are found.\n",
2064 lustre_msg_get_transno(req->rq_reqmsg),
2065 libcfs_nid2str(req->rq_peer.nid));
2067 /* Right now, only for MDS reint operation update replay and
2068 * normal request replay can have the same transno */
2069 if (lustre_msg_get_opc(req->rq_reqmsg) == MDS_REINT) {
2070 req_capsule_set(&req->rq_pill, &RQF_MDS_REINT);
2071 req->rq_status = req_capsule_server_pack(&req->rq_pill);
2072 if (likely(req->rq_export))
2073 target_committed_to_req(req);
2074 lustre_msg_set_transno(req->rq_repmsg, req->rq_transno);
2075 target_send_reply(req, req->rq_status, 0);
2077 DEBUG_REQ(D_ERROR, req, "wrong opc" "from %s\n",
2078 libcfs_nid2str(req->rq_peer.nid));
2080 target_exp_dequeue_req_replay(req);
2081 target_request_copy_put(req);
2082 obd->obd_replayed_requests++;
2086 * Update last_rcvd of the update
2088 * Because update recovery might update the last_rcvd by updates, i.e.
2089 * it will not update the last_rcvd information in memory, so we need
2090 * refresh these information in memory after update recovery.
2092 * \param[in] obd obd_device under recoverying.
2093 * \param[in] dtrq the update replay requests being replayed.
2095 static void target_update_lcd(struct lu_env *env, struct lu_target *lut,
2096 struct distribute_txn_replay_req *dtrq)
2098 struct obd_device *obd = lut->lut_obd;
2099 struct obd_export *export;
2100 struct tg_export_data *ted;
2101 struct distribute_txn_replay_req_sub *dtrqs;
2102 struct seq_server_site *site;
2103 struct update_records *ur;
2104 const struct lu_fid *fid;
2105 struct update_ops *ops;
2106 struct update_params *params;
2107 struct update_op *op;
2110 struct lsd_client_data *lcd = NULL;
2112 /* if Updates has been executed(committed) on the recovery target,
2113 * i.e. the updates is not being executed on the target, so we do
2114 * not need update it in memory */
2115 site = lu_site2seq(obd->obd_lu_dev->ld_site);
2116 mdt_index = site->ss_node_id;
2117 dtrqs = dtrq_sub_lookup(dtrq, mdt_index);
2121 if (dtrq->dtrq_lur == NULL)
2124 /* Find the update last_rcvd record */
2125 fid = lu_object_fid(&lut->lut_last_rcvd->do_lu);
2126 ur = &dtrq->dtrq_lur->lur_update_rec;
2128 params = update_records_get_params(ur);
2129 for (i = 0, op = &ops->uops_op[0]; i < ur->ur_update_count;
2130 i++, op = update_op_next_op(op)) {
2135 if (!lu_fid_eq(&op->uop_fid, fid))
2138 if (op->uop_type != OUT_WRITE)
2141 buf = update_params_get_param_buf(params, op->uop_params_off[1],
2142 ur->ur_param_count, NULL);
2146 pos = le64_to_cpu(*(__u64 *)buf);
2150 buf = update_params_get_param_buf(params, op->uop_params_off[0],
2151 ur->ur_param_count, &size);
2155 if (size != sizeof(*lcd))
2160 if (lcd == NULL || lcd->lcd_uuid[0] == '\0')
2163 /* locate the export then update the exp_target_data if needed */
2164 export = cfs_hash_lookup(obd->obd_uuid_hash, lcd->lcd_uuid);
2168 ted = &export->exp_target_data;
2169 if (lcd->lcd_last_xid > ted->ted_lcd->lcd_last_xid) {
2170 CDEBUG(D_HA, "%s update xid from "LPU64" to "LPU64"\n",
2171 lut->lut_obd->obd_name, ted->ted_lcd->lcd_last_xid,
2173 ted->ted_lcd->lcd_last_xid = lcd->lcd_last_xid;
2174 ted->ted_lcd->lcd_last_result = lcd->lcd_last_result;
2176 class_export_put(export);
2179 static void replay_request_or_update(struct lu_env *env,
2180 struct lu_target *lut,
2181 struct target_recovery_data *trd,
2182 struct ptlrpc_thread *thread)
2184 struct obd_device *obd = lut->lut_obd;
2185 struct ptlrpc_request *req = NULL;
2190 CDEBUG(D_HA, "Waiting for transno "LPD64"\n",
2191 obd->obd_next_recovery_transno);
2193 /* Replay all of request and update by transno */
2195 struct target_distribute_txn_data *tdtd = lut->lut_tdtd;
2197 CFS_FAIL_TIMEOUT(OBD_FAIL_TGT_REPLAY_DELAY2, cfs_fail_val);
2199 /** It is needed to extend recovery window above
2200 * recovery_time_soft. Extending is possible only in the
2201 * end of recovery window (see more details in
2202 * handle_recovery_req()).
2204 CFS_FAIL_TIMEOUT_MS(OBD_FAIL_TGT_REPLAY_DELAY, 300);
2206 if (target_recovery_overseer(lut, check_for_next_transno,
2207 exp_req_replay_healthy)) {
2208 abort_req_replay_queue(obd);
2209 abort_lock_replay_queue(obd);
2212 spin_lock(&obd->obd_recovery_task_lock);
2213 transno = get_next_transno(lut, &type);
2214 if (type == REQUEST_RECOVERY && tdtd != NULL &&
2215 transno == tdtd->tdtd_last_update_transno) {
2216 /* Drop replay request from client side, if the
2217 * replay has been executed by update with the
2219 req = list_entry(obd->obd_req_replay_queue.next,
2220 struct ptlrpc_request, rq_list);
2221 list_del_init(&req->rq_list);
2222 obd->obd_requests_queued_for_recovery--;
2223 spin_unlock(&obd->obd_recovery_task_lock);
2224 drop_duplicate_replay_req(env, obd, req);
2225 } else if (type == REQUEST_RECOVERY && transno != 0) {
2226 req = list_entry(obd->obd_req_replay_queue.next,
2227 struct ptlrpc_request, rq_list);
2228 list_del_init(&req->rq_list);
2229 obd->obd_requests_queued_for_recovery--;
2230 spin_unlock(&obd->obd_recovery_task_lock);
2231 LASSERT(trd->trd_processing_task == current_pid());
2232 DEBUG_REQ(D_HA, req, "processing t"LPD64" from %s",
2233 lustre_msg_get_transno(req->rq_reqmsg),
2234 libcfs_nid2str(req->rq_peer.nid));
2236 handle_recovery_req(thread, req,
2237 trd->trd_recovery_handler);
2239 * bz18031: increase next_recovery_transno before
2240 * target_request_copy_put() will drop exp_rpc reference
2242 spin_lock(&obd->obd_recovery_task_lock);
2243 obd->obd_next_recovery_transno++;
2244 spin_unlock(&obd->obd_recovery_task_lock);
2245 target_exp_dequeue_req_replay(req);
2246 target_request_copy_put(req);
2247 obd->obd_replayed_requests++;
2248 } else if (type == UPDATE_RECOVERY && transno != 0) {
2249 struct distribute_txn_replay_req *dtrq;
2251 spin_unlock(&obd->obd_recovery_task_lock);
2253 LASSERT(tdtd != NULL);
2254 dtrq = distribute_txn_get_next_req(tdtd);
2255 lu_context_enter(&thread->t_env->le_ctx);
2256 tdtd->tdtd_replay_handler(env, tdtd, dtrq);
2257 lu_context_exit(&thread->t_env->le_ctx);
2258 extend_recovery_timer(obd, obd_timeout, true);
2259 LASSERT(tdtd->tdtd_last_update_transno <= transno);
2260 tdtd->tdtd_last_update_transno = transno;
2261 spin_lock(&obd->obd_recovery_task_lock);
2262 if (transno > obd->obd_next_recovery_transno)
2263 obd->obd_next_recovery_transno = transno;
2264 spin_unlock(&obd->obd_recovery_task_lock);
2265 target_update_lcd(env, lut, dtrq);
2268 spin_unlock(&obd->obd_recovery_task_lock);
2269 LASSERT(list_empty(&obd->obd_req_replay_queue));
2270 LASSERT(atomic_read(&obd->obd_req_replay_clients) == 0);
2271 /** evict exports failed VBR */
2272 class_disconnect_stale_exports(obd, exp_vbr_healthy);
2278 static int target_recovery_thread(void *arg)
2280 struct lu_target *lut = arg;
2281 struct obd_device *obd = lut->lut_obd;
2282 struct ptlrpc_request *req;
2283 struct target_recovery_data *trd = &obd->obd_recovery_data;
2284 unsigned long delta;
2286 struct ptlrpc_thread *thread = NULL;
2290 unshare_fs_struct();
2291 OBD_ALLOC_PTR(thread);
2297 OBD_FREE_PTR(thread);
2301 rc = lu_context_init(&env->le_ctx, LCT_MD_THREAD | LCT_DT_THREAD);
2303 OBD_FREE_PTR(thread);
2308 thread->t_env = env;
2309 thread->t_id = -1; /* force filter_iobuf_get/put to use local buffers */
2310 env->le_ctx.lc_thread = thread;
2311 tgt_io_thread_init(thread); /* init thread_big_cache for IO requests */
2312 thread->t_watchdog = NULL;
2314 CDEBUG(D_HA, "%s: started recovery thread pid %d\n", obd->obd_name,
2316 trd->trd_processing_task = current_pid();
2318 spin_lock(&obd->obd_dev_lock);
2319 obd->obd_recovering = 1;
2320 spin_unlock(&obd->obd_dev_lock);
2321 complete(&trd->trd_starting);
2323 /* first of all, we have to know the first transno to replay */
2324 if (target_recovery_overseer(lut, check_for_recovery_ready,
2325 exp_connect_healthy)) {
2326 abort_req_replay_queue(obd);
2327 abort_lock_replay_queue(obd);
2328 if (lut->lut_tdtd != NULL)
2329 dtrq_list_destroy(lut->lut_tdtd);
2332 /* next stage: replay requests or update */
2334 CDEBUG(D_INFO, "1: request replay stage - %d clients from t"LPU64"\n",
2335 atomic_read(&obd->obd_req_replay_clients),
2336 obd->obd_next_recovery_transno);
2337 replay_request_or_update(env, lut, trd, thread);
2340 * The second stage: replay locks
2342 CDEBUG(D_INFO, "2: lock replay stage - %d clients\n",
2343 atomic_read(&obd->obd_lock_replay_clients));
2344 while ((req = target_next_replay_lock(lut))) {
2345 LASSERT(trd->trd_processing_task == current_pid());
2346 DEBUG_REQ(D_HA, req, "processing lock from %s: ",
2347 libcfs_nid2str(req->rq_peer.nid));
2348 handle_recovery_req(thread, req,
2349 trd->trd_recovery_handler);
2350 target_request_copy_put(req);
2351 obd->obd_replayed_locks++;
2355 * The third stage: reply on final pings, at this moment all clients
2356 * must have request in final queue
2358 CFS_FAIL_TIMEOUT(OBD_FAIL_TGT_REPLAY_RECONNECT, cfs_fail_val);
2359 CDEBUG(D_INFO, "3: final stage - process recovery completion pings\n");
2360 /** Update server last boot epoch */
2361 tgt_boot_epoch_update(lut);
2362 /* We drop recoverying flag to forward all new requests
2363 * to regular mds_handle() since now */
2364 spin_lock(&obd->obd_dev_lock);
2365 obd->obd_recovering = obd->obd_abort_recovery = 0;
2366 spin_unlock(&obd->obd_dev_lock);
2367 spin_lock(&obd->obd_recovery_task_lock);
2368 target_cancel_recovery_timer(obd);
2369 spin_unlock(&obd->obd_recovery_task_lock);
2370 while ((req = target_next_final_ping(obd))) {
2371 LASSERT(trd->trd_processing_task == current_pid());
2372 DEBUG_REQ(D_HA, req, "processing final ping from %s: ",
2373 libcfs_nid2str(req->rq_peer.nid));
2374 handle_recovery_req(thread, req,
2375 trd->trd_recovery_handler);
2376 /* Because the waiting client can not send ping to server,
2377 * so we need refresh the last_request_time, to avoid the
2378 * export is being evicted */
2379 ptlrpc_update_export_timer(req->rq_export, 0);
2380 target_request_copy_put(req);
2383 delta = jiffies_to_msecs(jiffies - delta) / MSEC_PER_SEC;
2384 CDEBUG(D_INFO,"4: recovery completed in %lus - %d/%d reqs/locks\n",
2385 delta, obd->obd_replayed_requests, obd->obd_replayed_locks);
2386 if (delta > OBD_RECOVERY_TIME_SOFT) {
2387 CWARN("too long recovery - read logs\n");
2388 libcfs_debug_dumplog();
2391 target_finish_recovery(lut);
2393 lu_context_fini(&env->le_ctx);
2394 trd->trd_processing_task = 0;
2395 complete(&trd->trd_finishing);
2397 tgt_io_thread_done(thread);
2398 OBD_FREE_PTR(thread);
2403 static int target_start_recovery_thread(struct lu_target *lut,
2404 svc_handler_t handler)
2406 struct obd_device *obd = lut->lut_obd;
2408 struct target_recovery_data *trd = &obd->obd_recovery_data;
2410 memset(trd, 0, sizeof(*trd));
2411 init_completion(&trd->trd_starting);
2412 init_completion(&trd->trd_finishing);
2413 trd->trd_recovery_handler = handler;
2415 if (!IS_ERR(kthread_run(target_recovery_thread,
2416 lut, "tgt_recov"))) {
2417 wait_for_completion(&trd->trd_starting);
2418 LASSERT(obd->obd_recovering != 0);
2426 void target_stop_recovery_thread(struct obd_device *obd)
2428 if (obd->obd_recovery_data.trd_processing_task > 0) {
2429 struct target_recovery_data *trd = &obd->obd_recovery_data;
2430 /** recovery can be done but postrecovery is not yet */
2431 spin_lock(&obd->obd_dev_lock);
2432 if (obd->obd_recovering) {
2433 CERROR("%s: Aborting recovery\n", obd->obd_name);
2434 obd->obd_abort_recovery = 1;
2435 wake_up(&obd->obd_next_transno_waitq);
2437 spin_unlock(&obd->obd_dev_lock);
2438 wait_for_completion(&trd->trd_finishing);
2441 EXPORT_SYMBOL(target_stop_recovery_thread);
2443 void target_recovery_fini(struct obd_device *obd)
2445 class_disconnect_exports(obd);
2446 target_stop_recovery_thread(obd);
2447 target_cleanup_recovery(obd);
2449 EXPORT_SYMBOL(target_recovery_fini);
2451 static void target_recovery_expired(unsigned long castmeharder)
2453 struct obd_device *obd = (struct obd_device *)castmeharder;
2454 CDEBUG(D_HA, "%s: recovery timed out; %d clients are still in recovery"
2455 " after %lds (%d clients connected)\n",
2456 obd->obd_name, atomic_read(&obd->obd_lock_replay_clients),
2457 cfs_time_current_sec()- obd->obd_recovery_start,
2458 atomic_read(&obd->obd_connected_clients));
2460 obd->obd_recovery_expired = 1;
2461 wake_up(&obd->obd_next_transno_waitq);
2464 void target_recovery_init(struct lu_target *lut, svc_handler_t handler)
2466 struct obd_device *obd = lut->lut_obd;
2468 if (obd->obd_max_recoverable_clients == 0) {
2469 /** Update server last boot epoch */
2470 tgt_boot_epoch_update(lut);
2474 CDEBUG(D_HA, "RECOVERY: service %s, %d recoverable clients, "
2475 "last_transno "LPU64"\n", obd->obd_name,
2476 obd->obd_max_recoverable_clients, obd->obd_last_committed);
2477 LASSERT(obd->obd_stopping == 0);
2478 obd->obd_next_recovery_transno = obd->obd_last_committed + 1;
2479 obd->obd_recovery_start = 0;
2480 obd->obd_recovery_end = 0;
2482 cfs_timer_init(&obd->obd_recovery_timer, target_recovery_expired, obd);
2483 target_start_recovery_thread(lut, handler);
2485 EXPORT_SYMBOL(target_recovery_init);
2487 static int target_process_req_flags(struct obd_device *obd,
2488 struct ptlrpc_request *req)
2490 struct obd_export *exp = req->rq_export;
2491 LASSERT(exp != NULL);
2492 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REQ_REPLAY_DONE) {
2493 /* client declares he's ready to replay locks */
2494 spin_lock(&exp->exp_lock);
2495 if (exp->exp_req_replay_needed) {
2496 exp->exp_req_replay_needed = 0;
2497 spin_unlock(&exp->exp_lock);
2499 LASSERT_ATOMIC_POS(&obd->obd_req_replay_clients);
2500 atomic_dec(&obd->obd_req_replay_clients);
2502 spin_unlock(&exp->exp_lock);
2505 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_LOCK_REPLAY_DONE) {
2506 /* client declares he's ready to complete recovery
2507 * so, we put the request on th final queue */
2508 spin_lock(&exp->exp_lock);
2509 if (exp->exp_lock_replay_needed) {
2510 exp->exp_lock_replay_needed = 0;
2511 spin_unlock(&exp->exp_lock);
2513 LASSERT_ATOMIC_POS(&obd->obd_lock_replay_clients);
2514 atomic_dec(&obd->obd_lock_replay_clients);
2516 spin_unlock(&exp->exp_lock);
2522 int target_queue_recovery_request(struct ptlrpc_request *req,
2523 struct obd_device *obd)
2525 __u64 transno = lustre_msg_get_transno(req->rq_reqmsg);
2526 struct ptlrpc_request *reqiter;
2530 if (obd->obd_recovery_data.trd_processing_task == current_pid()) {
2531 /* Processing the queue right now, don't re-add. */
2535 target_process_req_flags(obd, req);
2537 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_LOCK_REPLAY_DONE) {
2538 /* client declares he's ready to complete recovery
2539 * so, we put the request on th final queue */
2540 target_request_copy_get(req);
2541 DEBUG_REQ(D_HA, req, "queue final req");
2542 wake_up(&obd->obd_next_transno_waitq);
2543 spin_lock(&obd->obd_recovery_task_lock);
2544 if (obd->obd_recovering) {
2545 list_add_tail(&req->rq_list,
2546 &obd->obd_final_req_queue);
2548 spin_unlock(&obd->obd_recovery_task_lock);
2549 target_request_copy_put(req);
2550 RETURN(obd->obd_stopping ? -ENOTCONN : 1);
2552 spin_unlock(&obd->obd_recovery_task_lock);
2555 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REQ_REPLAY_DONE) {
2556 /* client declares he's ready to replay locks */
2557 target_request_copy_get(req);
2558 DEBUG_REQ(D_HA, req, "queue lock replay req");
2559 wake_up(&obd->obd_next_transno_waitq);
2560 spin_lock(&obd->obd_recovery_task_lock);
2561 LASSERT(obd->obd_recovering);
2562 /* usually due to recovery abort */
2563 if (!req->rq_export->exp_in_recovery) {
2564 spin_unlock(&obd->obd_recovery_task_lock);
2565 target_request_copy_put(req);
2568 LASSERT(req->rq_export->exp_lock_replay_needed);
2569 list_add_tail(&req->rq_list, &obd->obd_lock_replay_queue);
2570 spin_unlock(&obd->obd_recovery_task_lock);
2574 /* CAVEAT EMPTOR: The incoming request message has been swabbed
2575 * (i.e. buflens etc are in my own byte order), but type-dependent
2576 * buffers (eg mdt_body, ost_body etc) have NOT been swabbed. */
2579 INIT_LIST_HEAD(&req->rq_list);
2580 DEBUG_REQ(D_HA, req, "not queueing");
2584 /* If we're processing the queue, we want don't want to queue this
2587 * Also, if this request has a transno less than the one we're waiting
2588 * for, we should process it now. It could (and currently always will)
2589 * be an open request for a descriptor that was opened some time ago.
2591 * Also, a resent, replayed request that has already been
2592 * handled will pass through here and be processed immediately.
2594 CDEBUG(D_HA, "Next recovery transno: "LPU64
2595 ", current: "LPU64", replaying\n",
2596 obd->obd_next_recovery_transno, transno);
2597 spin_lock(&obd->obd_recovery_task_lock);
2598 if (transno < obd->obd_next_recovery_transno) {
2599 /* Processing the queue right now, don't re-add. */
2600 LASSERT(list_empty(&req->rq_list));
2601 spin_unlock(&obd->obd_recovery_task_lock);
2604 spin_unlock(&obd->obd_recovery_task_lock);
2606 if (OBD_FAIL_CHECK(OBD_FAIL_TGT_REPLAY_DROP))
2609 target_request_copy_get(req);
2610 if (!req->rq_export->exp_in_recovery) {
2611 target_request_copy_put(req);
2614 LASSERT(req->rq_export->exp_req_replay_needed);
2616 if (target_exp_enqueue_req_replay(req)) {
2617 DEBUG_REQ(D_ERROR, req, "dropping resent queued req");
2618 target_request_copy_put(req);
2623 spin_lock(&obd->obd_recovery_task_lock);
2624 LASSERT(obd->obd_recovering);
2625 list_for_each_entry(reqiter, &obd->obd_req_replay_queue, rq_list) {
2626 if (lustre_msg_get_transno(reqiter->rq_reqmsg) > transno) {
2627 list_add_tail(&req->rq_list, &reqiter->rq_list);
2632 if (unlikely(lustre_msg_get_transno(reqiter->rq_reqmsg) ==
2634 DEBUG_REQ(D_ERROR, req, "dropping replay: transno "
2635 "has been claimed by another client");
2636 spin_unlock(&obd->obd_recovery_task_lock);
2637 target_exp_dequeue_req_replay(req);
2638 target_request_copy_put(req);
2644 list_add_tail(&req->rq_list, &obd->obd_req_replay_queue);
2646 obd->obd_requests_queued_for_recovery++;
2647 spin_unlock(&obd->obd_recovery_task_lock);
2648 wake_up(&obd->obd_next_transno_waitq);
2652 int target_handle_ping(struct ptlrpc_request *req)
2654 obd_ping(req->rq_svc_thread->t_env, req->rq_export);
2655 return req_capsule_server_pack(&req->rq_pill);
2658 void target_committed_to_req(struct ptlrpc_request *req)
2660 struct obd_export *exp = req->rq_export;
2662 if (!exp->exp_obd->obd_no_transno && req->rq_repmsg != NULL)
2663 lustre_msg_set_last_committed(req->rq_repmsg,
2664 exp->exp_last_committed);
2666 DEBUG_REQ(D_IOCTL, req, "not sending last_committed update (%d/"
2667 "%d)", exp->exp_obd->obd_no_transno,
2668 req->rq_repmsg == NULL);
2670 CDEBUG(D_INFO, "last_committed "LPU64", transno "LPU64", xid "LPU64"\n",
2671 exp->exp_last_committed, req->rq_transno, req->rq_xid);
2674 #endif /* HAVE_SERVER_SUPPORT */
2677 * Packs current SLV and Limit into \a req.
2679 int target_pack_pool_reply(struct ptlrpc_request *req)
2681 struct obd_device *obd;
2684 /* Check that we still have all structures alive as this may
2685 * be some late RPC at shutdown time. */
2686 if (unlikely(!req->rq_export || !req->rq_export->exp_obd ||
2687 !exp_connect_lru_resize(req->rq_export))) {
2688 lustre_msg_set_slv(req->rq_repmsg, 0);
2689 lustre_msg_set_limit(req->rq_repmsg, 0);
2693 /* OBD is alive here as export is alive, which we checked above. */
2694 obd = req->rq_export->exp_obd;
2696 read_lock(&obd->obd_pool_lock);
2697 lustre_msg_set_slv(req->rq_repmsg, obd->obd_pool_slv);
2698 lustre_msg_set_limit(req->rq_repmsg, obd->obd_pool_limit);
2699 read_unlock(&obd->obd_pool_lock);
2704 static int target_send_reply_msg(struct ptlrpc_request *req,
2705 int rc, int fail_id)
2707 if (OBD_FAIL_CHECK_ORSET(fail_id & ~OBD_FAIL_ONCE, OBD_FAIL_ONCE)) {
2708 DEBUG_REQ(D_ERROR, req, "dropping reply");
2713 DEBUG_REQ(D_NET, req, "processing error (%d)", rc);
2714 req->rq_status = rc;
2715 return (ptlrpc_send_error(req, 1));
2717 DEBUG_REQ(D_NET, req, "sending reply");
2720 return (ptlrpc_send_reply(req, PTLRPC_REPLY_MAYBE_DIFFICULT));
2723 void target_send_reply(struct ptlrpc_request *req, int rc, int fail_id)
2725 struct ptlrpc_service_part *svcpt;
2727 struct ptlrpc_reply_state *rs;
2728 struct obd_export *exp;
2731 if (req->rq_no_reply) {
2736 svcpt = req->rq_rqbd->rqbd_svcpt;
2737 rs = req->rq_reply_state;
2738 if (rs == NULL || !rs->rs_difficult) {
2740 target_send_reply_msg (req, rc, fail_id);
2745 /* must be an export if locks saved */
2746 LASSERT(req->rq_export != NULL);
2747 /* req/reply consistent */
2748 LASSERT(rs->rs_svcpt == svcpt);
2751 LASSERT(!rs->rs_scheduled);
2752 LASSERT(!rs->rs_scheduled_ever);
2753 LASSERT(!rs->rs_handled);
2754 LASSERT(!rs->rs_on_net);
2755 LASSERT(rs->rs_export == NULL);
2756 LASSERT(list_empty(&rs->rs_obd_list));
2757 LASSERT(list_empty(&rs->rs_exp_list));
2759 exp = class_export_get(req->rq_export);
2761 /* disable reply scheduling while I'm setting up */
2762 rs->rs_scheduled = 1;
2764 rs->rs_xid = req->rq_xid;
2765 rs->rs_transno = req->rq_transno;
2766 rs->rs_export = exp;
2767 rs->rs_opc = lustre_msg_get_opc(req->rq_reqmsg);
2769 spin_lock(&exp->exp_uncommitted_replies_lock);
2770 CDEBUG(D_NET, "rs transno = "LPU64", last committed = "LPU64"\n",
2771 rs->rs_transno, exp->exp_last_committed);
2772 if (rs->rs_transno > exp->exp_last_committed) {
2773 /* not committed already */
2774 list_add_tail(&rs->rs_obd_list,
2775 &exp->exp_uncommitted_replies);
2777 spin_unlock(&exp->exp_uncommitted_replies_lock);
2779 spin_lock(&exp->exp_lock);
2780 list_add_tail(&rs->rs_exp_list, &exp->exp_outstanding_replies);
2781 spin_unlock(&exp->exp_lock);
2783 netrc = target_send_reply_msg(req, rc, fail_id);
2785 spin_lock(&svcpt->scp_rep_lock);
2787 atomic_inc(&svcpt->scp_nreps_difficult);
2790 /* error sending: reply is off the net. Also we need +1
2791 * reply ref until ptlrpc_handle_rs() is done
2792 * with the reply state (if the send was successful, there
2793 * would have been +1 ref for the net, which
2794 * reply_out_callback leaves alone) */
2796 ptlrpc_rs_addref(rs);
2799 spin_lock(&rs->rs_lock);
2800 if (rs->rs_transno <= exp->exp_last_committed ||
2801 (!rs->rs_on_net && !rs->rs_no_ack) ||
2802 list_empty(&rs->rs_exp_list) || /* completed already */
2803 list_empty(&rs->rs_obd_list)) {
2804 CDEBUG(D_HA, "Schedule reply immediately\n");
2805 ptlrpc_dispatch_difficult_reply(rs);
2807 list_add(&rs->rs_list, &svcpt->scp_rep_active);
2808 rs->rs_scheduled = 0; /* allow notifier to schedule */
2810 spin_unlock(&rs->rs_lock);
2811 spin_unlock(&svcpt->scp_rep_lock);
2815 ldlm_mode_t lck_compat_array[] = {
2816 [LCK_EX] = LCK_COMPAT_EX,
2817 [LCK_PW] = LCK_COMPAT_PW,
2818 [LCK_PR] = LCK_COMPAT_PR,
2819 [LCK_CW] = LCK_COMPAT_CW,
2820 [LCK_CR] = LCK_COMPAT_CR,
2821 [LCK_NL] = LCK_COMPAT_NL,
2822 [LCK_GROUP] = LCK_COMPAT_GROUP,
2823 [LCK_COS] = LCK_COMPAT_COS,
2827 * Rather arbitrary mapping from LDLM error codes to errno values. This should
2828 * not escape to the user level.
2830 int ldlm_error2errno(ldlm_error_t error)
2836 case ELDLM_LOCK_MATCHED:
2839 case ELDLM_LOCK_CHANGED:
2842 case ELDLM_LOCK_ABORTED:
2845 case ELDLM_LOCK_REPLACED:
2848 case ELDLM_NO_LOCK_DATA:
2851 case ELDLM_NAMESPACE_EXISTS:
2854 case ELDLM_BAD_NAMESPACE:
2858 if (((int)error) < 0) /* cast to signed type */
2859 result = error; /* as ldlm_error_t can be unsigned */
2861 CERROR("Invalid DLM result code: %d\n", error);
2867 EXPORT_SYMBOL(ldlm_error2errno);
2870 * Dual to ldlm_error2errno(): maps errno values back to ldlm_error_t.
2872 ldlm_error_t ldlm_errno2error(int err_no)
2881 error = ELDLM_LOCK_CHANGED;
2884 error = ELDLM_LOCK_ABORTED;
2887 error = ELDLM_LOCK_REPLACED;
2890 error = ELDLM_NO_LOCK_DATA;
2893 error = ELDLM_NAMESPACE_EXISTS;
2896 error = ELDLM_BAD_NAMESPACE;
2904 #if LUSTRE_TRACKS_LOCK_EXP_REFS
2905 void ldlm_dump_export_locks(struct obd_export *exp)
2907 spin_lock(&exp->exp_locks_list_guard);
2908 if (!list_empty(&exp->exp_locks_list)) {
2909 struct ldlm_lock *lock;
2911 CERROR("dumping locks for export %p,"
2912 "ignore if the unmount doesn't hang\n", exp);
2913 list_for_each_entry(lock, &exp->exp_locks_list,
2915 LDLM_ERROR(lock, "lock:");
2917 spin_unlock(&exp->exp_locks_list_guard);
2921 #ifdef HAVE_SERVER_SUPPORT
2922 static int target_bulk_timeout(void *data)
2925 /* We don't fail the connection here, because having the export
2926 * killed makes the (vital) call to commitrw very sad.
2931 static inline const char *bulk2type(struct ptlrpc_request *req)
2933 if (req->rq_bulk_read)
2935 if (req->rq_bulk_write)
2940 int target_bulk_io(struct obd_export *exp, struct ptlrpc_bulk_desc *desc,
2941 struct l_wait_info *lwi)
2943 struct ptlrpc_request *req = desc->bd_req;
2944 time_t start = cfs_time_current_sec();
2950 /* If there is eviction in progress, wait for it to finish. */
2951 if (unlikely(atomic_read(&exp->exp_obd->obd_evict_inprogress))) {
2952 *lwi = LWI_INTR(NULL, NULL);
2953 rc = l_wait_event(exp->exp_obd->obd_evict_inprogress_waitq,
2954 !atomic_read(&exp->exp_obd->
2955 obd_evict_inprogress),
2959 /* Check if client was evicted or reconnected already. */
2960 if (exp->exp_failed ||
2961 exp->exp_conn_cnt > lustre_msg_get_conn_cnt(req->rq_reqmsg)) {
2964 if (req->rq_bulk_read)
2965 rc = sptlrpc_svc_wrap_bulk(req, desc);
2967 rc = ptlrpc_start_bulk_transfer(desc);
2971 DEBUG_REQ(D_ERROR, req, "bulk %s failed: rc %d",
2972 bulk2type(req), rc);
2976 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_SENDPAGE)) {
2977 ptlrpc_abort_bulk(desc);
2981 /* limit actual bulk transfer to bulk_timeout seconds */
2982 deadline = start + bulk_timeout;
2983 if (deadline > req->rq_deadline)
2984 deadline = req->rq_deadline;
2987 long timeoutl = deadline - cfs_time_current_sec();
2988 cfs_duration_t timeout = timeoutl <= 0 ?
2989 CFS_TICK : cfs_time_seconds(timeoutl);
2992 *lwi = LWI_TIMEOUT_INTERVAL(timeout, cfs_time_seconds(1),
2993 target_bulk_timeout, desc);
2994 rc = l_wait_event(desc->bd_waitq,
2995 !ptlrpc_server_bulk_active(desc) ||
2998 lustre_msg_get_conn_cnt(req->rq_reqmsg),
3000 LASSERT(rc == 0 || rc == -ETIMEDOUT);
3001 /* Wait again if we changed rq_deadline. */
3002 rq_deadline = ACCESS_ONCE(req->rq_deadline);
3003 deadline = start + bulk_timeout;
3004 if (deadline > rq_deadline)
3005 deadline = rq_deadline;
3006 } while ((rc == -ETIMEDOUT) &&
3007 (deadline > cfs_time_current_sec()));
3009 if (rc == -ETIMEDOUT) {
3010 DEBUG_REQ(D_ERROR, req, "timeout on bulk %s after %ld%+lds",
3011 bulk2type(req), deadline - start,
3012 cfs_time_current_sec() - deadline);
3013 ptlrpc_abort_bulk(desc);
3014 } else if (exp->exp_failed) {
3015 DEBUG_REQ(D_ERROR, req, "Eviction on bulk %s",
3018 ptlrpc_abort_bulk(desc);
3019 } else if (exp->exp_conn_cnt >
3020 lustre_msg_get_conn_cnt(req->rq_reqmsg)) {
3021 DEBUG_REQ(D_ERROR, req, "Reconnect on bulk %s",
3023 /* We don't reply anyway. */
3025 ptlrpc_abort_bulk(desc);
3026 } else if (desc->bd_failure) {
3027 DEBUG_REQ(D_ERROR, req, "network error on bulk %s",
3029 /* XXX should this be a different errno? */
3032 if (req->rq_bulk_write)
3033 rc = sptlrpc_svc_unwrap_bulk(req, desc);
3034 if (rc == 0 && desc->bd_nob_transferred != desc->bd_nob) {
3035 DEBUG_REQ(D_ERROR, req, "truncated bulk %s %d(%d)",
3036 bulk2type(req), desc->bd_nob_transferred,
3038 /* XXX should this be a different errno? */
3045 EXPORT_SYMBOL(target_bulk_io);
3047 #endif /* HAVE_SERVER_SUPPORT */