1 /* -*- mode: c; c-basic-offset: 8; indent-tabs-mode: nil; -*-
2 * vim:expandtab:shiftwidth=8:tabstop=8:
6 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
8 * This program is free software; you can redistribute it and/or modify
9 * it under the terms of the GNU General Public License version 2 only,
10 * as published by the Free Software Foundation.
12 * This program is distributed in the hope that it will be useful, but
13 * WITHOUT ANY WARRANTY; without even the implied warranty of
14 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
15 * General Public License version 2 for more details (a copy is included
16 * in the LICENSE file that accompanied this code).
18 * You should have received a copy of the GNU General Public License
19 * version 2 along with this program; If not, see
20 * http://www.sun.com/software/products/lustre/docs/GPLv2.pdf
22 * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
23 * CA 95054 USA or visit www.sun.com if you need additional information or
29 * Copyright (c) 2003, 2010, Oracle and/or its affiliates. All rights reserved.
30 * Use is subject to license terms.
32 * Copyright (c) 2011, 2012, Whamcloud, Inc.
35 * This file is part of Lustre, http://www.lustre.org/
36 * Lustre is a trademark of Sun Microsystems, Inc.
40 # define EXPORT_SYMTAB
42 #define DEBUG_SUBSYSTEM S_LDLM
45 # include <libcfs/libcfs.h>
47 # include <liblustre.h>
50 #include <lustre_mds.h>
51 #include <lustre_dlm.h>
52 #include <lustre_net.h>
53 #include <lustre_sec.h>
54 #include "ldlm_internal.h"
56 /* @priority: if non-zero, move the selected to the list head
57 * @create: if zero, only search in existed connections
59 static int import_set_conn(struct obd_import *imp, struct obd_uuid *uuid,
60 int priority, int create)
62 struct ptlrpc_connection *ptlrpc_conn;
63 struct obd_import_conn *imp_conn = NULL, *item;
67 if (!create && !priority) {
68 CDEBUG(D_HA, "Nothing to do\n");
72 ptlrpc_conn = ptlrpc_uuid_to_connection(uuid);
74 CDEBUG(D_HA, "can't find connection %s\n", uuid->uuid);
79 OBD_ALLOC(imp_conn, sizeof(*imp_conn));
81 GOTO(out_put, rc = -ENOMEM);
85 cfs_spin_lock(&imp->imp_lock);
86 cfs_list_for_each_entry(item, &imp->imp_conn_list, oic_item) {
87 if (obd_uuid_equals(uuid, &item->oic_uuid)) {
89 cfs_list_del(&item->oic_item);
90 cfs_list_add(&item->oic_item,
92 item->oic_last_attempt = 0;
94 CDEBUG(D_HA, "imp %p@%s: found existing conn %s%s\n",
95 imp, imp->imp_obd->obd_name, uuid->uuid,
96 (priority ? ", moved to head" : ""));
97 cfs_spin_unlock(&imp->imp_lock);
98 GOTO(out_free, rc = 0);
103 imp_conn->oic_conn = ptlrpc_conn;
104 imp_conn->oic_uuid = *uuid;
105 imp_conn->oic_last_attempt = 0;
107 cfs_list_add(&imp_conn->oic_item, &imp->imp_conn_list);
109 cfs_list_add_tail(&imp_conn->oic_item,
110 &imp->imp_conn_list);
111 CDEBUG(D_HA, "imp %p@%s: add connection %s at %s\n",
112 imp, imp->imp_obd->obd_name, uuid->uuid,
113 (priority ? "head" : "tail"));
115 cfs_spin_unlock(&imp->imp_lock);
116 GOTO(out_free, rc = -ENOENT);
119 cfs_spin_unlock(&imp->imp_lock);
123 OBD_FREE(imp_conn, sizeof(*imp_conn));
125 ptlrpc_connection_put(ptlrpc_conn);
129 int import_set_conn_priority(struct obd_import *imp, struct obd_uuid *uuid)
131 return import_set_conn(imp, uuid, 1, 0);
134 int client_import_add_conn(struct obd_import *imp, struct obd_uuid *uuid,
137 return import_set_conn(imp, uuid, priority, 1);
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 cfs_spin_lock(&imp->imp_lock);
148 if (cfs_list_empty(&imp->imp_conn_list)) {
149 LASSERT(!imp->imp_connection);
153 cfs_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 /* is current conn? */
159 if (imp_conn == imp->imp_conn_current) {
160 LASSERT(imp_conn->oic_conn == imp->imp_connection);
162 if (imp->imp_state != LUSTRE_IMP_CLOSED &&
163 imp->imp_state != LUSTRE_IMP_DISCON) {
164 CERROR("can't remove current connection\n");
165 GOTO(out, rc = -EBUSY);
168 ptlrpc_connection_put(imp->imp_connection);
169 imp->imp_connection = NULL;
171 dlmexp = class_conn2export(&imp->imp_dlm_handle);
172 if (dlmexp && dlmexp->exp_connection) {
173 LASSERT(dlmexp->exp_connection ==
175 ptlrpc_connection_put(dlmexp->exp_connection);
176 dlmexp->exp_connection = NULL;
180 cfs_list_del(&imp_conn->oic_item);
181 ptlrpc_connection_put(imp_conn->oic_conn);
182 OBD_FREE(imp_conn, sizeof(*imp_conn));
183 CDEBUG(D_HA, "imp %p@%s: remove connection %s\n",
184 imp, imp->imp_obd->obd_name, uuid->uuid);
189 cfs_spin_unlock(&imp->imp_lock);
191 CERROR("connection %s not found\n", uuid->uuid);
196 * Find conn uuid by peer nid. @peer is a server nid. This function is used
197 * to find a conn uuid of @imp which can reach @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 cfs_spin_lock(&imp->imp_lock);
207 cfs_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 cfs_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 rpc 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);
231 /* configure an RPC client OBD device
236 * 3 - inactive-on-startup
238 int client_obd_setup(struct obd_device *obddev, struct lustre_cfg *lcfg)
240 struct client_obd *cli = &obddev->u.cli;
241 struct obd_import *imp;
242 struct obd_uuid server_uuid;
243 int rq_portal, rp_portal, connect_op;
244 char *name = obddev->obd_type->typ_name;
245 ldlm_ns_type_t ns_type = LDLM_NS_TYPE_UNKNOWN;
249 /* In a more perfect world, we would hang a ptlrpc_client off of
250 * obd_type and just use the values from there. */
251 if (!strcmp(name, LUSTRE_OSC_NAME)) {
252 rq_portal = OST_REQUEST_PORTAL;
253 rp_portal = OSC_REPLY_PORTAL;
254 connect_op = OST_CONNECT;
255 cli->cl_sp_me = LUSTRE_SP_CLI;
256 cli->cl_sp_to = LUSTRE_SP_OST;
257 ns_type = LDLM_NS_TYPE_OSC;
259 } else if (!strcmp(name, LUSTRE_MDC_NAME)) {
260 rq_portal = MDS_REQUEST_PORTAL;
261 rp_portal = MDC_REPLY_PORTAL;
262 connect_op = MDS_CONNECT;
263 cli->cl_sp_me = LUSTRE_SP_CLI;
264 cli->cl_sp_to = LUSTRE_SP_MDT;
265 ns_type = LDLM_NS_TYPE_MDC;
267 } else if (!strcmp(name, LUSTRE_MGC_NAME)) {
268 rq_portal = MGS_REQUEST_PORTAL;
269 rp_portal = MGC_REPLY_PORTAL;
270 connect_op = MGS_CONNECT;
271 cli->cl_sp_me = LUSTRE_SP_MGC;
272 cli->cl_sp_to = LUSTRE_SP_MGS;
273 cli->cl_flvr_mgc.sf_rpc = SPTLRPC_FLVR_INVALID;
274 ns_type = LDLM_NS_TYPE_MGC;
277 CERROR("unknown client OBD type \"%s\", can't setup\n",
282 if (LUSTRE_CFG_BUFLEN(lcfg, 1) < 1) {
283 CERROR("requires a TARGET UUID\n");
287 if (LUSTRE_CFG_BUFLEN(lcfg, 1) > 37) {
288 CERROR("client UUID must be less than 38 characters\n");
292 if (LUSTRE_CFG_BUFLEN(lcfg, 2) < 1) {
293 CERROR("setup requires a SERVER UUID\n");
297 if (LUSTRE_CFG_BUFLEN(lcfg, 2) > 37) {
298 CERROR("target UUID must be less than 38 characters\n");
302 cfs_init_rwsem(&cli->cl_sem);
303 cfs_sema_init(&cli->cl_mgc_sem, 1);
304 cli->cl_conn_count = 0;
305 memcpy(server_uuid.uuid, lustre_cfg_buf(lcfg, 2),
306 min_t(unsigned int, LUSTRE_CFG_BUFLEN(lcfg, 2),
307 sizeof(server_uuid)));
310 cli->cl_avail_grant = 0;
311 /* FIXME: should limit this for the sum of all cl_dirty_max */
312 cli->cl_dirty_max = OSC_MAX_DIRTY_DEFAULT * 1024 * 1024;
313 if (cli->cl_dirty_max >> CFS_PAGE_SHIFT > cfs_num_physpages / 8)
314 cli->cl_dirty_max = cfs_num_physpages << (CFS_PAGE_SHIFT - 3);
315 CFS_INIT_LIST_HEAD(&cli->cl_cache_waiters);
316 CFS_INIT_LIST_HEAD(&cli->cl_loi_ready_list);
317 CFS_INIT_LIST_HEAD(&cli->cl_loi_hp_ready_list);
318 CFS_INIT_LIST_HEAD(&cli->cl_loi_write_list);
319 CFS_INIT_LIST_HEAD(&cli->cl_loi_read_list);
320 client_obd_list_lock_init(&cli->cl_loi_list_lock);
321 cli->cl_r_in_flight = 0;
322 cli->cl_w_in_flight = 0;
324 cfs_spin_lock_init(&cli->cl_read_rpc_hist.oh_lock);
325 cfs_spin_lock_init(&cli->cl_write_rpc_hist.oh_lock);
326 cfs_spin_lock_init(&cli->cl_read_page_hist.oh_lock);
327 cfs_spin_lock_init(&cli->cl_write_page_hist.oh_lock);
328 cfs_spin_lock_init(&cli->cl_read_offset_hist.oh_lock);
329 cfs_spin_lock_init(&cli->cl_write_offset_hist.oh_lock);
330 cfs_waitq_init(&cli->cl_destroy_waitq);
331 cfs_atomic_set(&cli->cl_destroy_in_flight, 0);
332 #ifdef ENABLE_CHECKSUM
333 /* Turn on checksumming by default. */
334 cli->cl_checksum = 1;
336 * The supported checksum types will be worked out at connect time
337 * Set cl_chksum* to CRC32 for now to avoid returning screwed info
340 cli->cl_cksum_type = cli->cl_supp_cksum_types = OBD_CKSUM_CRC32;
342 cfs_atomic_set(&cli->cl_resends, OSC_DEFAULT_RESENDS);
344 /* This value may be changed at connect time in
345 ptlrpc_connect_interpret. */
346 cli->cl_max_pages_per_rpc = min((int)PTLRPC_MAX_BRW_PAGES,
347 (int)(1024 * 1024 >> CFS_PAGE_SHIFT));
349 if (!strcmp(name, LUSTRE_MDC_NAME)) {
350 cli->cl_max_rpcs_in_flight = MDC_MAX_RIF_DEFAULT;
351 } else if (cfs_num_physpages >> (20 - CFS_PAGE_SHIFT) <= 128 /* MB */) {
352 cli->cl_max_rpcs_in_flight = 2;
353 } else if (cfs_num_physpages >> (20 - CFS_PAGE_SHIFT) <= 256 /* MB */) {
354 cli->cl_max_rpcs_in_flight = 3;
355 } else if (cfs_num_physpages >> (20 - CFS_PAGE_SHIFT) <= 512 /* MB */) {
356 cli->cl_max_rpcs_in_flight = 4;
358 cli->cl_max_rpcs_in_flight = OSC_MAX_RIF_DEFAULT;
363 CERROR("ldlm_get_ref failed: %d\n", rc);
367 ptlrpc_init_client(rq_portal, rp_portal, name,
368 &obddev->obd_ldlm_client);
370 imp = class_new_import(obddev);
372 GOTO(err_ldlm, rc = -ENOENT);
373 imp->imp_client = &obddev->obd_ldlm_client;
374 imp->imp_connect_op = connect_op;
375 CFS_INIT_LIST_HEAD(&imp->imp_pinger_chain);
376 memcpy(cli->cl_target_uuid.uuid, lustre_cfg_buf(lcfg, 1),
377 LUSTRE_CFG_BUFLEN(lcfg, 1));
378 class_import_put(imp);
380 rc = client_import_add_conn(imp, &server_uuid, 1);
382 CERROR("can't add initial connection\n");
383 GOTO(err_import, rc);
386 cli->cl_import = imp;
387 /* cli->cl_max_mds_{easize,cookiesize} updated by mdc_init_ea_size() */
388 cli->cl_max_mds_easize = sizeof(struct lov_mds_md_v3);
389 cli->cl_max_mds_cookiesize = sizeof(struct llog_cookie);
391 if (LUSTRE_CFG_BUFLEN(lcfg, 3) > 0) {
392 if (!strcmp(lustre_cfg_string(lcfg, 3), "inactive")) {
393 CDEBUG(D_HA, "marking %s %s->%s as inactive\n",
394 name, obddev->obd_name,
395 cli->cl_target_uuid.uuid);
396 cfs_spin_lock(&imp->imp_lock);
397 imp->imp_deactive = 1;
398 cfs_spin_unlock(&imp->imp_lock);
402 obddev->obd_namespace = ldlm_namespace_new(obddev, obddev->obd_name,
403 LDLM_NAMESPACE_CLIENT,
404 LDLM_NAMESPACE_GREEDY,
406 if (obddev->obd_namespace == NULL) {
407 CERROR("Unable to create client namespace - %s\n",
409 GOTO(err_import, rc = -ENOMEM);
412 cli->cl_qchk_stat = CL_NOT_QUOTACHECKED;
417 class_destroy_import(imp);
425 int client_obd_cleanup(struct obd_device *obddev)
429 ldlm_namespace_free_post(obddev->obd_namespace);
430 obddev->obd_namespace = NULL;
432 LASSERT(obddev->u.cli.cl_import == NULL);
438 /* ->o_connect() method for client side (OSC and MDC and MGC) */
439 int client_connect_import(const struct lu_env *env,
440 struct obd_export **exp,
441 struct obd_device *obd, struct obd_uuid *cluuid,
442 struct obd_connect_data *data, void *localdata)
444 struct client_obd *cli = &obd->u.cli;
445 struct obd_import *imp = cli->cl_import;
446 struct obd_connect_data *ocd;
447 struct lustre_handle conn = { 0 };
452 cfs_down_write(&cli->cl_sem);
453 if (cli->cl_conn_count > 0 )
454 GOTO(out_sem, rc = -EALREADY);
456 rc = class_connect(&conn, obd, cluuid);
460 cli->cl_conn_count++;
461 *exp = class_conn2export(&conn);
463 LASSERT(obd->obd_namespace);
465 imp->imp_dlm_handle = conn;
466 rc = ptlrpc_init_import(imp);
470 ocd = &imp->imp_connect_data;
473 imp->imp_connect_flags_orig = data->ocd_connect_flags;
476 rc = ptlrpc_connect_import(imp);
478 LASSERT (imp->imp_state == LUSTRE_IMP_DISCON);
481 LASSERT((*exp)->exp_connection);
484 LASSERTF((ocd->ocd_connect_flags & data->ocd_connect_flags) ==
485 ocd->ocd_connect_flags, "old "LPX64", new "LPX64"\n",
486 data->ocd_connect_flags, ocd->ocd_connect_flags);
487 data->ocd_connect_flags = ocd->ocd_connect_flags;
490 ptlrpc_pinger_add_import(imp);
496 cli->cl_conn_count--;
497 class_disconnect(*exp);
501 cfs_up_write(&cli->cl_sem);
506 int client_disconnect_export(struct obd_export *exp)
508 struct obd_device *obd = class_exp2obd(exp);
509 struct client_obd *cli;
510 struct obd_import *imp;
515 CERROR("invalid export for disconnect: exp %p cookie "LPX64"\n",
516 exp, exp ? exp->exp_handle.h_cookie : -1);
521 imp = cli->cl_import;
523 cfs_down_write(&cli->cl_sem);
524 CDEBUG(D_INFO, "disconnect %s - %d\n", obd->obd_name,
527 if (!cli->cl_conn_count) {
528 CERROR("disconnecting disconnected device (%s)\n",
530 GOTO(out_disconnect, rc = -EINVAL);
533 cli->cl_conn_count--;
534 if (cli->cl_conn_count)
535 GOTO(out_disconnect, rc = 0);
537 /* Mark import deactivated now, so we don't try to reconnect if any
538 * of the cleanup RPCs fails (e.g. ldlm cancel, etc). We don't
539 * fully deactivate the import, or that would drop all requests. */
540 cfs_spin_lock(&imp->imp_lock);
541 imp->imp_deactive = 1;
542 cfs_spin_unlock(&imp->imp_lock);
544 /* Some non-replayable imports (MDS's OSCs) are pinged, so just
545 * delete it regardless. (It's safe to delete an import that was
547 (void)ptlrpc_pinger_del_import(imp);
549 if (obd->obd_namespace != NULL) {
550 /* obd_force == local only */
551 ldlm_cli_cancel_unused(obd->obd_namespace, NULL,
552 obd->obd_force ? LCF_LOCAL : 0, NULL);
553 ldlm_namespace_free_prior(obd->obd_namespace, imp, obd->obd_force);
557 * there's no need to hold sem during disconnecting an import,
558 * and actually it may cause deadlock in gss.
560 cfs_up_write(&cli->cl_sem);
561 rc = ptlrpc_disconnect_import(imp, 0);
562 cfs_down_write(&cli->cl_sem);
564 ptlrpc_invalidate_import(imp);
569 /* use server style - class_disconnect should be always called for
571 err = class_disconnect(exp);
575 cfs_up_write(&cli->cl_sem);
580 int server_disconnect_export(struct obd_export *exp)
585 /* Disconnect early so that clients can't keep using export */
586 rc = class_disconnect(exp);
587 /* close import for avoid sending any requests */
588 if (exp->exp_imp_reverse)
589 ptlrpc_cleanup_imp(exp->exp_imp_reverse);
591 if (exp->exp_obd->obd_namespace != NULL)
592 ldlm_cancel_locks_for_export(exp);
594 /* complete all outstanding replies */
595 cfs_spin_lock(&exp->exp_lock);
596 while (!cfs_list_empty(&exp->exp_outstanding_replies)) {
597 struct ptlrpc_reply_state *rs =
598 cfs_list_entry(exp->exp_outstanding_replies.next,
599 struct ptlrpc_reply_state, rs_exp_list);
600 struct ptlrpc_service *svc = rs->rs_service;
602 cfs_spin_lock(&svc->srv_rs_lock);
603 cfs_list_del_init(&rs->rs_exp_list);
604 cfs_spin_lock(&rs->rs_lock);
605 ptlrpc_schedule_difficult_reply(rs);
606 cfs_spin_unlock(&rs->rs_lock);
607 cfs_spin_unlock(&svc->srv_rs_lock);
609 cfs_spin_unlock(&exp->exp_lock);
614 /* --------------------------------------------------------------------------
615 * from old lib/target.c
616 * -------------------------------------------------------------------------- */
618 static int target_handle_reconnect(struct lustre_handle *conn,
619 struct obd_export *exp,
620 struct obd_uuid *cluuid)
624 if (exp->exp_connection && exp->exp_imp_reverse) {
625 struct lustre_handle *hdl;
626 struct obd_device *target;
628 hdl = &exp->exp_imp_reverse->imp_remote_handle;
629 target = exp->exp_obd;
631 /* Might be a re-connect after a partition. */
632 if (!memcmp(&conn->cookie, &hdl->cookie, sizeof conn->cookie)) {
633 if (target->obd_recovering)
634 LCONSOLE_WARN("%s: Client %s (at %s) reconnect"
635 "ing, waiting for %d clients in "
636 "recovery for %lds\n", target->obd_name,
637 obd_uuid2str(&exp->exp_client_uuid),
638 obd_export_nid2str(exp),
639 target->obd_max_recoverable_clients,
640 cfs_duration_sec(cfs_time_sub(
642 &target->obd_recovery_timer),
643 cfs_time_current())));
645 LCONSOLE_WARN("%s: Client %s (at %s) "
646 "reconnecting\n", target->obd_name,
647 obd_uuid2str(&exp->exp_client_uuid),
648 obd_export_nid2str(exp));
650 conn->cookie = exp->exp_handle.h_cookie;
651 /* target_handle_connect() treats EALREADY and
652 * -EALREADY differently. EALREADY means we are
653 * doing a valid reconnect from the same client. */
656 LCONSOLE_WARN("%s: The server has already connected "
657 "client %s (at %s) with handle " LPX64
658 ", rejecting a client with the same "
659 "uuid trying to reconnect with "
660 "handle " LPX64, target->obd_name,
661 obd_uuid2str(&exp->exp_client_uuid),
662 obd_export_nid2str(exp),
663 hdl->cookie, conn->cookie);
664 memset(conn, 0, sizeof *conn);
665 /* target_handle_connect() treats EALREADY and
666 * -EALREADY differently. -EALREADY is an error
667 * (same UUID, different handle). */
672 conn->cookie = exp->exp_handle.h_cookie;
673 CDEBUG(D_HA, "connect export for UUID '%s' at %p, cookie "LPX64"\n",
674 cluuid->uuid, exp, conn->cookie);
678 void target_client_add_cb(struct obd_device *obd, __u64 transno, void *cb_data,
681 struct obd_export *exp = cb_data;
683 CDEBUG(D_RPCTRACE, "%s: committing for initial connect of %s\n",
684 obd->obd_name, exp->exp_client_uuid.uuid);
686 cfs_spin_lock(&exp->exp_lock);
687 exp->exp_need_sync = 0;
688 cfs_spin_unlock(&exp->exp_lock);
689 class_export_cb_put(exp);
691 EXPORT_SYMBOL(target_client_add_cb);
695 check_and_start_recovery_timer(struct obd_device *obd,
696 struct ptlrpc_request *req, int new_client);
699 check_and_start_recovery_timer(struct obd_device *obd,
700 struct ptlrpc_request *req, int new_client)
705 int target_handle_connect(struct ptlrpc_request *req)
707 struct obd_device *target, *targref = NULL;
708 struct obd_export *export = NULL;
709 struct obd_import *revimp;
710 struct lustre_handle conn;
711 struct lustre_handle *tmp;
712 struct obd_uuid tgtuuid;
713 struct obd_uuid cluuid;
714 struct obd_uuid remote_uuid;
720 struct obd_connect_data *data, *tmpdata;
722 lnet_nid_t *client_nid = NULL;
725 OBD_RACE(OBD_FAIL_TGT_CONN_RACE);
727 str = req_capsule_client_get(&req->rq_pill, &RMF_TGTUUID);
729 DEBUG_REQ(D_ERROR, req, "bad target UUID for connect");
730 GOTO(out, rc = -EINVAL);
733 obd_str2uuid(&tgtuuid, str);
734 target = class_uuid2obd(&tgtuuid);
736 target = class_name2obd(str);
738 if (!target || target->obd_stopping || !target->obd_set_up) {
739 deuuidify(str, NULL, &target_start, &target_len);
740 LCONSOLE_ERROR_MSG(0x137, "%.*s: Not available for connect "
741 "from %s (%s)\n", target_len, target_start,
742 libcfs_nid2str(req->rq_peer.nid), !target ?
743 "no target" : (target->obd_stopping ?
744 "stopping" : "not set up"));
745 GOTO(out, rc = -ENODEV);
748 if (target->obd_no_conn) {
749 LCONSOLE_WARN("%s: Temporarily refusing client connection "
750 "from %s\n", target->obd_name,
751 libcfs_nid2str(req->rq_peer.nid));
752 GOTO(out, rc = -EAGAIN);
755 /* Make sure the target isn't cleaned up while we're here. Yes,
756 there's still a race between the above check and our incref here.
757 Really, class_uuid2obd should take the ref. */
758 targref = class_incref(target, __FUNCTION__, cfs_current());
761 str = req_capsule_client_get(&req->rq_pill, &RMF_CLUUID);
763 DEBUG_REQ(D_ERROR, req, "bad client UUID for connect");
764 GOTO(out, rc = -EINVAL);
767 obd_str2uuid(&cluuid, str);
769 /* XXX extract a nettype and format accordingly */
770 switch (sizeof(lnet_nid_t)) {
771 /* NB the casts only avoid compiler warnings */
773 snprintf(remote_uuid.uuid, sizeof remote_uuid,
774 "NET_"LPX64"_UUID", (__u64)req->rq_peer.nid);
777 snprintf(remote_uuid.uuid, sizeof remote_uuid,
778 "NET_%x_UUID", (__u32)req->rq_peer.nid);
784 tmp = req_capsule_client_get(&req->rq_pill, &RMF_CONN);
786 GOTO(out, rc = -EPROTO);
790 size = req_capsule_get_size(&req->rq_pill, &RMF_CONNECT_DATA,
792 data = req_capsule_client_get(&req->rq_pill, &RMF_CONNECT_DATA);
794 GOTO(out, rc = -EPROTO);
796 rc = req_capsule_server_pack(&req->rq_pill);
800 if (lustre_msg_get_op_flags(req->rq_reqmsg) & MSG_CONNECT_LIBCLIENT) {
802 DEBUG_REQ(D_WARNING, req, "Refusing old (unversioned) "
803 "libclient connection attempt");
804 GOTO(out, rc = -EPROTO);
805 } else if (data->ocd_version < LUSTRE_VERSION_CODE -
806 LUSTRE_VERSION_ALLOWED_OFFSET ||
807 data->ocd_version > LUSTRE_VERSION_CODE +
808 LUSTRE_VERSION_ALLOWED_OFFSET) {
809 DEBUG_REQ(D_WARNING, req, "Refusing %s (%d.%d.%d.%d) "
810 "libclient connection attempt",
811 data->ocd_version < LUSTRE_VERSION_CODE ?
813 OBD_OCD_VERSION_MAJOR(data->ocd_version),
814 OBD_OCD_VERSION_MINOR(data->ocd_version),
815 OBD_OCD_VERSION_PATCH(data->ocd_version),
816 OBD_OCD_VERSION_FIX(data->ocd_version));
817 data = req_capsule_server_sized_get(&req->rq_pill,
819 offsetof(typeof(*data), ocd_version) +
820 sizeof(data->ocd_version));
822 data->ocd_connect_flags = OBD_CONNECT_VERSION;
823 data->ocd_version = LUSTRE_VERSION_CODE;
825 GOTO(out, rc = -EPROTO);
829 if ((lustre_msg_get_op_flags(req->rq_reqmsg) & MSG_CONNECT_INITIAL) &&
830 (data->ocd_connect_flags & OBD_CONNECT_MDS))
833 /* lctl gets a backstage, all-access pass. */
834 if (obd_uuid_equals(&cluuid, &target->obd_uuid))
835 goto dont_check_exports;
837 export = cfs_hash_lookup(target->obd_uuid_hash, &cluuid);
841 /* we've found an export in the hash */
842 if (export->exp_connecting) { /* bug 9635, et. al. */
843 LCONSOLE_WARN("%s: Export %p already connecting from %s\n",
844 export->exp_obd->obd_name, export,
845 libcfs_nid2str(req->rq_peer.nid));
846 class_export_put(export);
849 } else if (mds_conn && export->exp_connection) {
850 if (req->rq_peer.nid != export->exp_connection->c_peer.nid)
851 /* mds reconnected after failover */
852 LCONSOLE_WARN("%s: Received MDS connection from "
853 "%s, removing former export from %s\n",
854 target->obd_name, libcfs_nid2str(req->rq_peer.nid),
855 libcfs_nid2str(export->exp_connection->c_peer.nid));
857 /* new mds connection from the same nid */
858 LCONSOLE_WARN("%s: Received new MDS connection from "
859 "%s, removing former export from same NID\n",
860 target->obd_name, libcfs_nid2str(req->rq_peer.nid));
861 class_fail_export(export);
862 class_export_put(export);
865 } else if (export->exp_connection != NULL &&
866 req->rq_peer.nid != export->exp_connection->c_peer.nid &&
867 (lustre_msg_get_op_flags(req->rq_reqmsg) &
868 MSG_CONNECT_INITIAL)) {
869 /* in mds failover we have static uuid but nid can be
871 LCONSOLE_WARN("%s: Client %s seen on new nid %s when "
872 "existing nid %s is already connected\n",
873 target->obd_name, cluuid.uuid,
874 libcfs_nid2str(req->rq_peer.nid),
876 export->exp_connection->c_peer.nid));
878 class_export_put(export);
881 cfs_spin_lock(&export->exp_lock);
882 export->exp_connecting = 1;
883 cfs_spin_unlock(&export->exp_lock);
884 LASSERT(export->exp_obd == target);
886 rc = target_handle_reconnect(&conn, export, &cluuid);
889 /* If we found an export, we already unlocked. */
892 OBD_FAIL_TIMEOUT(OBD_FAIL_TGT_DELAY_CONNECT, 2 * obd_timeout);
893 } else if (req->rq_export == NULL &&
894 cfs_atomic_read(&export->exp_rpc_count) > 0) {
895 LCONSOLE_WARN("%s: Client %s (at %s) refused connection, "
896 "still busy with %d references\n",
897 target->obd_name, cluuid.uuid,
898 libcfs_nid2str(req->rq_peer.nid),
899 cfs_atomic_read(&export->exp_refcount));
900 GOTO(out, rc = -EBUSY);
901 } else if (req->rq_export != NULL &&
902 (cfs_atomic_read(&export->exp_rpc_count) > 1)) {
903 /* the current connect rpc has increased exp_rpc_count */
904 LCONSOLE_WARN("%s: Client %s (at %s) refused reconnection, "
905 "still busy with %d active RPCs\n",
906 target->obd_name, cluuid.uuid,
907 libcfs_nid2str(req->rq_peer.nid),
908 cfs_atomic_read(&export->exp_rpc_count) - 1);
909 cfs_spin_lock(&export->exp_lock);
910 if (req->rq_export->exp_conn_cnt <
911 lustre_msg_get_conn_cnt(req->rq_reqmsg))
912 /* try to abort active requests */
913 req->rq_export->exp_abort_active_req = 1;
914 cfs_spin_unlock(&export->exp_lock);
915 GOTO(out, rc = -EBUSY);
916 } else if (lustre_msg_get_conn_cnt(req->rq_reqmsg) == 1) {
917 if (!strstr(cluuid.uuid, "mdt"))
918 LCONSOLE_WARN("%s: Rejecting reconnect from the "
919 "known client %s (at %s) because it "
920 "is indicating it is a new client",
921 target->obd_name, cluuid.uuid,
922 libcfs_nid2str(req->rq_peer.nid));
923 GOTO(out, rc = -EALREADY);
925 OBD_FAIL_TIMEOUT(OBD_FAIL_TGT_DELAY_RECONNECT, 2 * obd_timeout);
932 CDEBUG(D_HA, "%s: connection from %s@%s %st"LPU64" exp %p cur %ld last %ld\n",
933 target->obd_name, cluuid.uuid, libcfs_nid2str(req->rq_peer.nid),
934 target->obd_recovering ? "recovering/" : "", data->ocd_transno,
935 export, (long)cfs_time_current_sec(),
936 export ? (long)export->exp_last_request_time : 0);
938 /* If this is the first time a client connects, reset the recovery
940 if (rc == 0 && target->obd_recovering)
941 check_and_start_recovery_timer(target, req, export == NULL);
943 /* We want to handle EALREADY but *not* -EALREADY from
944 * target_handle_reconnect(), return reconnection state in a flag */
945 if (rc == EALREADY) {
946 lustre_msg_add_op_flags(req->rq_repmsg, MSG_CONNECT_RECONNECT);
952 /* Tell the client if we support replayable requests */
953 if (target->obd_replayable)
954 lustre_msg_add_op_flags(req->rq_repmsg, MSG_CONNECT_REPLAYABLE);
955 client_nid = &req->rq_peer.nid;
957 if (export == NULL) {
958 if (target->obd_recovering) {
961 t = cfs_timer_deadline(&target->obd_recovery_timer);
962 t = cfs_time_sub(t, cfs_time_current());
963 LCONSOLE_WARN("%s: Denying connection for new client "
964 "%s (at %s), waiting for %d clients in "
965 "recovery for "CFS_TIME_T"s\n",
967 libcfs_nid2str(req->rq_peer.nid),
969 cfs_atomic_read(&target-> \
970 obd_lock_replay_clients),
971 cfs_duration_sec(t));
975 rc = obd_connect(req->rq_svc_thread->t_env,
976 &export, target, &cluuid, data,
979 conn.cookie = export->exp_handle.h_cookie;
980 /* LU-1092 reconnect put export refcount in the
981 * end, connect needs take one here too. */
982 class_export_get(export);
986 rc = obd_reconnect(req->rq_svc_thread->t_env,
987 export, target, &cluuid, data, client_nid);
989 /* prevous done via class_conn2export */
990 class_export_get(export);
995 LASSERT(target->u.obt.obt_magic == OBT_MAGIC);
996 data->ocd_instance = target->u.obt.obt_instance;
998 /* Return only the parts of obd_connect_data that we understand, so the
999 * client knows that we don't understand the rest. */
1001 tmpsize = req_capsule_get_size(&req->rq_pill, &RMF_CONNECT_DATA,
1003 tmpdata = req_capsule_server_get(&req->rq_pill,
1005 /* Don't use struct assignment here, because the client reply
1006 * buffer may be smaller/larger than the local struct
1007 * obd_connect_data. */
1008 memcpy(tmpdata, data, min(tmpsize, size));
1011 /* If all else goes well, this is our RPC return code. */
1014 lustre_msg_set_handle(req->rq_repmsg, &conn);
1016 /* If the client and the server are the same node, we will already
1017 * have an export that really points to the client's DLM export,
1018 * because we have a shared handles table.
1020 * XXX this will go away when shaver stops sending the "connect" handle
1021 * in the real "remote handle" field of the request --phik 24 Apr 2003
1023 if (req->rq_export != NULL)
1024 class_export_put(req->rq_export);
1026 req->rq_export = export;
1028 cfs_spin_lock(&export->exp_lock);
1029 if (export->exp_conn_cnt >= lustre_msg_get_conn_cnt(req->rq_reqmsg)) {
1030 cfs_spin_unlock(&export->exp_lock);
1031 CDEBUG(D_RPCTRACE, "%s: %s already connected at higher "
1032 "conn_cnt: %d > %d\n",
1033 cluuid.uuid, libcfs_nid2str(req->rq_peer.nid),
1034 export->exp_conn_cnt,
1035 lustre_msg_get_conn_cnt(req->rq_reqmsg));
1037 GOTO(out, rc = -EALREADY);
1039 LASSERT(lustre_msg_get_conn_cnt(req->rq_reqmsg) > 0);
1040 export->exp_conn_cnt = lustre_msg_get_conn_cnt(req->rq_reqmsg);
1041 export->exp_abort_active_req = 0;
1043 /* request from liblustre? Don't evict it for not pinging. */
1044 if (lustre_msg_get_op_flags(req->rq_reqmsg) & MSG_CONNECT_LIBCLIENT) {
1045 export->exp_libclient = 1;
1046 cfs_spin_unlock(&export->exp_lock);
1048 cfs_spin_lock(&target->obd_dev_lock);
1049 cfs_list_del_init(&export->exp_obd_chain_timed);
1050 cfs_spin_unlock(&target->obd_dev_lock);
1052 cfs_spin_unlock(&export->exp_lock);
1055 if (export->exp_connection != NULL) {
1056 /* Check to see if connection came from another NID */
1057 if ((export->exp_connection->c_peer.nid != req->rq_peer.nid) &&
1058 !cfs_hlist_unhashed(&export->exp_nid_hash))
1059 cfs_hash_del(export->exp_obd->obd_nid_hash,
1060 &export->exp_connection->c_peer.nid,
1061 &export->exp_nid_hash);
1063 ptlrpc_connection_put(export->exp_connection);
1066 export->exp_connection = ptlrpc_connection_get(req->rq_peer,
1069 if (cfs_hlist_unhashed(&export->exp_nid_hash)) {
1070 cfs_hash_add(export->exp_obd->obd_nid_hash,
1071 &export->exp_connection->c_peer.nid,
1072 &export->exp_nid_hash);
1075 class_disconnect->class_export_recovery_cleanup() race
1077 if (target->obd_recovering && !export->exp_in_recovery) {
1079 __u64 transno = data->ocd_transno;
1081 cfs_spin_lock(&export->exp_lock);
1082 export->exp_in_recovery = 1;
1083 export->exp_req_replay_needed = 1;
1084 export->exp_lock_replay_needed = 1;
1085 cfs_spin_unlock(&export->exp_lock);
1087 has_transno = !!(lustre_msg_get_op_flags(req->rq_reqmsg) &
1088 MSG_CONNECT_TRANSNO);
1089 if (has_transno && transno == 0)
1090 CWARN("Connect with zero transno!\n");
1092 if (has_transno && transno > 0 &&
1093 transno < target->obd_next_recovery_transno &&
1094 transno > target->obd_last_committed) {
1095 /* another way is to use cmpxchg() so it will be
1097 cfs_spin_lock(&target->obd_recovery_task_lock);
1098 if (transno < target->obd_next_recovery_transno)
1099 target->obd_next_recovery_transno = transno;
1100 cfs_spin_unlock(&target->obd_recovery_task_lock);
1103 cfs_atomic_inc(&target->obd_req_replay_clients);
1104 cfs_atomic_inc(&target->obd_lock_replay_clients);
1105 if (cfs_atomic_inc_return(&target->obd_connected_clients) ==
1106 target->obd_max_recoverable_clients)
1107 cfs_waitq_signal(&target->obd_next_transno_waitq);
1110 /* Tell the client we're in recovery, when client is involved in it. */
1111 if (target->obd_recovering)
1112 lustre_msg_add_op_flags(req->rq_repmsg, MSG_CONNECT_RECOVERING);
1114 tmp = req_capsule_client_get(&req->rq_pill, &RMF_CONN);
1117 if (export->exp_imp_reverse != NULL) {
1118 /* destroyed import can be still referenced in ctxt */
1119 obd_set_info_async(export, sizeof(KEY_REVIMP_UPD),
1120 KEY_REVIMP_UPD, 0, NULL, NULL);
1122 client_destroy_import(export->exp_imp_reverse);
1125 /* for the rest part, we return -ENOTCONN in case of errors
1126 * in order to let client initialize connection again.
1128 revimp = export->exp_imp_reverse = class_new_import(target);
1130 CERROR("fail to alloc new reverse import.\n");
1131 GOTO(out, rc = -ENOTCONN);
1134 revimp->imp_connection = ptlrpc_connection_addref(export->exp_connection);
1135 revimp->imp_client = &export->exp_obd->obd_ldlm_client;
1136 revimp->imp_remote_handle = conn;
1137 revimp->imp_dlm_fake = 1;
1138 revimp->imp_state = LUSTRE_IMP_FULL;
1140 /* unknown versions will be caught in
1141 * ptlrpc_handle_server_req_in->lustre_unpack_msg() */
1142 revimp->imp_msg_magic = req->rq_reqmsg->lm_magic;
1144 if ((export->exp_connect_flags & OBD_CONNECT_AT) &&
1145 (revimp->imp_msg_magic != LUSTRE_MSG_MAGIC_V1))
1146 revimp->imp_msghdr_flags |= MSGHDR_AT_SUPPORT;
1148 revimp->imp_msghdr_flags &= ~MSGHDR_AT_SUPPORT;
1150 if ((export->exp_connect_flags & OBD_CONNECT_FULL20) &&
1151 (revimp->imp_msg_magic != LUSTRE_MSG_MAGIC_V1))
1152 revimp->imp_msghdr_flags |= MSGHDR_CKSUM_INCOMPAT18;
1154 revimp->imp_msghdr_flags &= ~MSGHDR_CKSUM_INCOMPAT18;
1156 rc = sptlrpc_import_sec_adapt(revimp, req->rq_svc_ctx, &req->rq_flvr);
1158 CERROR("Failed to get sec for reverse import: %d\n", rc);
1159 export->exp_imp_reverse = NULL;
1160 class_destroy_import(revimp);
1163 class_import_put(revimp);
1166 cfs_spin_lock(&export->exp_lock);
1167 export->exp_connecting = 0;
1168 cfs_spin_unlock(&export->exp_lock);
1170 class_export_put(export);
1173 class_decref(targref, __FUNCTION__, cfs_current());
1175 req->rq_status = rc;
1179 int target_handle_disconnect(struct ptlrpc_request *req)
1184 rc = req_capsule_server_pack(&req->rq_pill);
1188 /* keep the rq_export around so we can send the reply */
1189 req->rq_status = obd_disconnect(class_export_get(req->rq_export));
1194 void target_destroy_export(struct obd_export *exp)
1196 /* exports created from last_rcvd data, and "fake"
1197 exports created by lctl don't have an import */
1198 if (exp->exp_imp_reverse != NULL)
1199 client_destroy_import(exp->exp_imp_reverse);
1201 LASSERT_ATOMIC_ZERO(&exp->exp_locks_count);
1202 LASSERT_ATOMIC_ZERO(&exp->exp_rpc_count);
1203 LASSERT_ATOMIC_ZERO(&exp->exp_cb_count);
1204 LASSERT_ATOMIC_ZERO(&exp->exp_replay_count);
1208 * Recovery functions
1210 static void target_request_copy_get(struct ptlrpc_request *req)
1212 class_export_rpc_get(req->rq_export);
1213 LASSERT(cfs_list_empty(&req->rq_list));
1214 CFS_INIT_LIST_HEAD(&req->rq_replay_list);
1216 /* increase refcount to keep request in queue */
1217 cfs_atomic_inc(&req->rq_refcount);
1218 /** let export know it has replays to be handled */
1219 cfs_atomic_inc(&req->rq_export->exp_replay_count);
1222 static void target_request_copy_put(struct ptlrpc_request *req)
1224 LASSERT(cfs_list_empty(&req->rq_replay_list));
1225 LASSERT_ATOMIC_POS(&req->rq_export->exp_replay_count);
1227 cfs_atomic_dec(&req->rq_export->exp_replay_count);
1228 class_export_rpc_put(req->rq_export);
1229 ptlrpc_server_drop_request(req);
1232 static int target_exp_enqueue_req_replay(struct ptlrpc_request *req)
1234 __u64 transno = lustre_msg_get_transno(req->rq_reqmsg);
1235 struct obd_export *exp = req->rq_export;
1236 struct ptlrpc_request *reqiter;
1241 cfs_spin_lock(&exp->exp_lock);
1242 cfs_list_for_each_entry(reqiter, &exp->exp_req_replay_queue,
1244 if (lustre_msg_get_transno(reqiter->rq_reqmsg) == transno) {
1251 /* we expect it with RESENT and REPLAY flags */
1252 if ((lustre_msg_get_flags(req->rq_reqmsg) &
1253 (MSG_RESENT | MSG_REPLAY)) != (MSG_RESENT | MSG_REPLAY))
1254 CERROR("invalid flags %x of resent replay\n",
1255 lustre_msg_get_flags(req->rq_reqmsg));
1257 cfs_list_add_tail(&req->rq_replay_list,
1258 &exp->exp_req_replay_queue);
1261 cfs_spin_unlock(&exp->exp_lock);
1265 static void target_exp_dequeue_req_replay(struct ptlrpc_request *req)
1267 LASSERT(!cfs_list_empty(&req->rq_replay_list));
1268 LASSERT(req->rq_export);
1270 cfs_spin_lock(&req->rq_export->exp_lock);
1271 cfs_list_del_init(&req->rq_replay_list);
1272 cfs_spin_unlock(&req->rq_export->exp_lock);
1276 static void target_finish_recovery(struct obd_device *obd)
1278 time_t elapsed_time = max_t(time_t, 1, cfs_time_current_sec() -
1279 obd->obd_recovery_start);
1282 LCONSOLE_INFO("%s: Recovery over after %d:%.02d, of %d clients "
1283 "%d recovered and %d %s evicted.\n", obd->obd_name,
1284 (int)elapsed_time / 60, (int)elapsed_time % 60,
1285 obd->obd_max_recoverable_clients,
1286 cfs_atomic_read(&obd->obd_connected_clients),
1287 obd->obd_stale_clients,
1288 obd->obd_stale_clients == 1 ? "was" : "were");
1290 ldlm_reprocess_all_ns(obd->obd_namespace);
1291 cfs_spin_lock(&obd->obd_recovery_task_lock);
1292 if (!cfs_list_empty(&obd->obd_req_replay_queue) ||
1293 !cfs_list_empty(&obd->obd_lock_replay_queue) ||
1294 !cfs_list_empty(&obd->obd_final_req_queue)) {
1295 CERROR("%s: Recovery queues ( %s%s%s) are not empty\n",
1297 cfs_list_empty(&obd->obd_req_replay_queue) ? "" : "req ",
1298 cfs_list_empty(&obd->obd_lock_replay_queue) ? \
1300 cfs_list_empty(&obd->obd_final_req_queue) ? \
1302 cfs_spin_unlock(&obd->obd_recovery_task_lock);
1305 cfs_spin_unlock(&obd->obd_recovery_task_lock);
1307 obd->obd_recovery_end = cfs_time_current_sec();
1309 /* when recovery finished, cleanup orphans on mds and ost */
1310 if (OBT(obd) && OBP(obd, postrecov)) {
1311 int rc = OBP(obd, postrecov)(obd);
1313 LCONSOLE_WARN("%s: Post recovery failed, rc %d\n",
1319 static void abort_req_replay_queue(struct obd_device *obd)
1321 struct ptlrpc_request *req, *n;
1322 cfs_list_t abort_list;
1324 CFS_INIT_LIST_HEAD(&abort_list);
1325 cfs_spin_lock(&obd->obd_recovery_task_lock);
1326 cfs_list_splice_init(&obd->obd_req_replay_queue, &abort_list);
1327 cfs_spin_unlock(&obd->obd_recovery_task_lock);
1328 cfs_list_for_each_entry_safe(req, n, &abort_list, rq_list) {
1329 DEBUG_REQ(D_WARNING, req, "aborted:");
1330 req->rq_status = -ENOTCONN;
1331 if (ptlrpc_error(req)) {
1332 DEBUG_REQ(D_ERROR, req,
1333 "failed abort_req_reply; skipping");
1335 target_exp_dequeue_req_replay(req);
1336 target_request_copy_put(req);
1340 static void abort_lock_replay_queue(struct obd_device *obd)
1342 struct ptlrpc_request *req, *n;
1343 cfs_list_t abort_list;
1345 CFS_INIT_LIST_HEAD(&abort_list);
1346 cfs_spin_lock(&obd->obd_recovery_task_lock);
1347 cfs_list_splice_init(&obd->obd_lock_replay_queue, &abort_list);
1348 cfs_spin_unlock(&obd->obd_recovery_task_lock);
1349 cfs_list_for_each_entry_safe(req, n, &abort_list, rq_list){
1350 DEBUG_REQ(D_ERROR, req, "aborted:");
1351 req->rq_status = -ENOTCONN;
1352 if (ptlrpc_error(req)) {
1353 DEBUG_REQ(D_ERROR, req,
1354 "failed abort_lock_reply; skipping");
1356 target_request_copy_put(req);
1360 /* Called from a cleanup function if the device is being cleaned up
1361 forcefully. The exports should all have been disconnected already,
1362 the only thing left to do is
1363 - clear the recovery flags
1365 - free queued requests and replies, but don't send replies
1366 Because the obd_stopping flag is set, no new requests should be received.
1369 void target_cleanup_recovery(struct obd_device *obd)
1371 struct ptlrpc_request *req, *n;
1372 cfs_list_t clean_list;
1375 CFS_INIT_LIST_HEAD(&clean_list);
1376 cfs_spin_lock(&obd->obd_dev_lock);
1377 if (!obd->obd_recovering) {
1378 cfs_spin_unlock(&obd->obd_dev_lock);
1382 obd->obd_recovering = obd->obd_abort_recovery = 0;
1383 cfs_spin_unlock(&obd->obd_dev_lock);
1385 cfs_spin_lock(&obd->obd_recovery_task_lock);
1386 target_cancel_recovery_timer(obd);
1387 cfs_list_splice_init(&obd->obd_req_replay_queue, &clean_list);
1388 cfs_spin_unlock(&obd->obd_recovery_task_lock);
1390 cfs_list_for_each_entry_safe(req, n, &clean_list, rq_list) {
1391 LASSERT(req->rq_reply_state == 0);
1392 target_exp_dequeue_req_replay(req);
1393 target_request_copy_put(req);
1396 cfs_spin_lock(&obd->obd_recovery_task_lock);
1397 cfs_list_splice_init(&obd->obd_lock_replay_queue, &clean_list);
1398 cfs_list_splice_init(&obd->obd_final_req_queue, &clean_list);
1399 cfs_spin_unlock(&obd->obd_recovery_task_lock);
1401 cfs_list_for_each_entry_safe(req, n, &clean_list, rq_list){
1402 LASSERT(req->rq_reply_state == 0);
1403 target_request_copy_put(req);
1409 /* obd_recovery_task_lock should be held */
1410 void target_cancel_recovery_timer(struct obd_device *obd)
1412 CDEBUG(D_HA, "%s: cancel recovery timer\n", obd->obd_name);
1413 cfs_timer_disarm(&obd->obd_recovery_timer);
1416 static void target_start_recovery_timer(struct obd_device *obd)
1418 if (obd->obd_recovery_start != 0)
1421 cfs_spin_lock(&obd->obd_dev_lock);
1422 if (!obd->obd_recovering || obd->obd_abort_recovery) {
1423 cfs_spin_unlock(&obd->obd_dev_lock);
1427 LASSERT(obd->obd_recovery_timeout != 0);
1429 if (obd->obd_recovery_start != 0) {
1430 cfs_spin_unlock(&obd->obd_dev_lock);
1434 cfs_timer_arm(&obd->obd_recovery_timer,
1435 cfs_time_shift(obd->obd_recovery_timeout));
1436 obd->obd_recovery_start = cfs_time_current_sec();
1437 cfs_spin_unlock(&obd->obd_dev_lock);
1439 LCONSOLE_WARN("%s: Will be in recovery for at least %d:%.02d, "
1440 "or until %d client%s reconnect%s\n",
1442 obd->obd_recovery_timeout / 60,
1443 obd->obd_recovery_timeout % 60,
1444 obd->obd_max_recoverable_clients,
1445 (obd->obd_max_recoverable_clients == 1) ? "" : "s",
1446 (obd->obd_max_recoverable_clients == 1) ? "s": "");
1450 * extend recovery window.
1452 * if @extend is true, extend recovery window to have @drt remaining at least;
1453 * otherwise, make sure the recovery timeout value is not less than @drt.
1455 static void extend_recovery_timer(struct obd_device *obd, int drt, bool extend)
1459 cfs_duration_t left;
1462 cfs_spin_lock(&obd->obd_dev_lock);
1463 if (!obd->obd_recovering || obd->obd_abort_recovery) {
1464 cfs_spin_unlock(&obd->obd_dev_lock);
1467 LASSERT(obd->obd_recovery_start != 0);
1469 now = cfs_time_current_sec();
1470 to = obd->obd_recovery_timeout;
1471 end = obd->obd_recovery_start + to;
1472 left = cfs_time_sub(end, now);
1474 if (extend && (drt > left)) {
1476 } else if (!extend && (drt > to)) {
1478 /* reduce drt by already passed time */
1479 drt -= obd->obd_recovery_timeout - left;
1482 if (to > obd->obd_recovery_time_hard)
1483 to = obd->obd_recovery_time_hard;
1484 if (obd->obd_recovery_timeout < to) {
1485 obd->obd_recovery_timeout = to;
1486 cfs_timer_arm(&obd->obd_recovery_timer,
1487 cfs_time_shift(drt));
1489 cfs_spin_unlock(&obd->obd_dev_lock);
1491 CDEBUG(D_HA, "%s: recovery timer will expire in %u seconds\n",
1492 obd->obd_name, (unsigned)drt);
1495 /* Reset the timer with each new client connection */
1497 * This timer is actually reconnect_timer, which is for making sure
1498 * the total recovery window is at least as big as my reconnect
1499 * attempt timing. So the initial recovery time_out will be set to
1500 * OBD_RECOVERY_FACTOR * obd_timeout. If the timeout coming
1501 * from client is bigger than this, then the recovery time_out will
1502 * be extended to make sure the client could be reconnected, in the
1503 * process, the timeout from the new client should be ignored.
1507 check_and_start_recovery_timer(struct obd_device *obd,
1508 struct ptlrpc_request *req,
1511 int service_time = lustre_msg_get_service_time(req->rq_reqmsg);
1512 struct obd_device_target *obt = &obd->u.obt;
1513 struct lustre_sb_info *lsi;
1515 if (!new_client && service_time)
1516 /* Teach server about old server's estimates, as first guess
1517 * at how long new requests will take. */
1518 at_measured(&req->rq_rqbd->rqbd_service->srv_at_estimate,
1521 target_start_recovery_timer(obd);
1523 /* convert the service time to rpc timeout,
1524 * reuse service_time to limit stack usage */
1525 service_time = at_est2timeout(service_time);
1527 /* We expect other clients to timeout within service_time, then try
1528 * to reconnect, then try the failover server. The max delay between
1529 * connect attempts is SWITCH_MAX + SWITCH_INC + INITIAL */
1530 service_time += 2 * INITIAL_CONNECT_TIMEOUT;
1532 LASSERT(obt->obt_magic == OBT_MAGIC);
1533 lsi = s2lsi(obt->obt_sb);
1534 if (!(lsi->lsi_flags | LSI_IR_CAPABLE))
1535 service_time += 2 * (CONNECTION_SWITCH_MAX +
1536 CONNECTION_SWITCH_INC);
1537 if (service_time > obd->obd_recovery_timeout && !new_client)
1538 extend_recovery_timer(obd, service_time, false);
1541 /** Health checking routines */
1542 static inline int exp_connect_healthy(struct obd_export *exp)
1544 return (exp->exp_in_recovery);
1547 /** if export done req_replay or has replay in queue */
1548 static inline int exp_req_replay_healthy(struct obd_export *exp)
1550 return (!exp->exp_req_replay_needed ||
1551 cfs_atomic_read(&exp->exp_replay_count) > 0);
1553 /** if export done lock_replay or has replay in queue */
1554 static inline int exp_lock_replay_healthy(struct obd_export *exp)
1556 return (!exp->exp_lock_replay_needed ||
1557 cfs_atomic_read(&exp->exp_replay_count) > 0);
1560 static inline int exp_vbr_healthy(struct obd_export *exp)
1562 return (!exp->exp_vbr_failed);
1565 static inline int exp_finished(struct obd_export *exp)
1567 return (exp->exp_in_recovery && !exp->exp_lock_replay_needed);
1570 /** Checking routines for recovery */
1571 static int check_for_clients(struct obd_device *obd)
1573 unsigned int clnts = cfs_atomic_read(&obd->obd_connected_clients);
1575 if (obd->obd_abort_recovery || obd->obd_recovery_expired)
1577 LASSERT(clnts <= obd->obd_max_recoverable_clients);
1578 if (obd->obd_no_conn == 0 &&
1579 clnts + obd->obd_stale_clients == obd->obd_max_recoverable_clients)
1584 static int check_for_next_transno(struct obd_device *obd)
1586 struct ptlrpc_request *req = NULL;
1587 int wake_up = 0, connected, completed, queue_len;
1588 __u64 next_transno, req_transno;
1591 cfs_spin_lock(&obd->obd_recovery_task_lock);
1592 if (!cfs_list_empty(&obd->obd_req_replay_queue)) {
1593 req = cfs_list_entry(obd->obd_req_replay_queue.next,
1594 struct ptlrpc_request, rq_list);
1595 req_transno = lustre_msg_get_transno(req->rq_reqmsg);
1600 connected = cfs_atomic_read(&obd->obd_connected_clients);
1601 completed = connected - cfs_atomic_read(&obd->obd_req_replay_clients);
1602 queue_len = obd->obd_requests_queued_for_recovery;
1603 next_transno = obd->obd_next_recovery_transno;
1605 CDEBUG(D_HA, "max: %d, connected: %d, completed: %d, queue_len: %d, "
1606 "req_transno: "LPU64", next_transno: "LPU64"\n",
1607 obd->obd_max_recoverable_clients, connected, completed,
1608 queue_len, req_transno, next_transno);
1610 if (obd->obd_abort_recovery) {
1611 CDEBUG(D_HA, "waking for aborted recovery\n");
1613 } else if (obd->obd_recovery_expired) {
1614 CDEBUG(D_HA, "waking for expired recovery\n");
1616 } else if (cfs_atomic_read(&obd->obd_req_replay_clients) == 0) {
1617 CDEBUG(D_HA, "waking for completed recovery\n");
1619 } else if (req_transno == next_transno) {
1620 CDEBUG(D_HA, "waking for next ("LPD64")\n", next_transno);
1622 } else if (queue_len == cfs_atomic_read(&obd->obd_req_replay_clients)) {
1624 /** handle gaps occured due to lost reply or VBR */
1625 LASSERTF(req_transno >= next_transno,
1626 "req_transno: "LPU64", next_transno: "LPU64"\n",
1627 req_transno, next_transno);
1628 if (req_transno > obd->obd_last_committed &&
1629 !obd->obd_version_recov)
1632 "%s: waking for gap in transno, VBR is %s (skip: "
1633 LPD64", ql: %d, comp: %d, conn: %d, next: "LPD64
1634 ", last_committed: "LPD64")\n",
1635 obd->obd_name, obd->obd_version_recov ? "ON" : "OFF",
1636 next_transno, queue_len, completed, connected,
1637 req_transno, obd->obd_last_committed);
1638 obd->obd_next_recovery_transno = req_transno;
1640 } else if (OBD_FAIL_CHECK(OBD_FAIL_MDS_RECOVERY_ACCEPTS_GAPS)) {
1641 CDEBUG(D_HA, "accepting transno gaps is explicitly allowed"
1642 " by fail_lock, waking up ("LPD64")\n", next_transno);
1643 obd->obd_next_recovery_transno = req_transno;
1646 cfs_spin_unlock(&obd->obd_recovery_task_lock);
1650 static int check_for_next_lock(struct obd_device *obd)
1654 cfs_spin_lock(&obd->obd_recovery_task_lock);
1655 if (!cfs_list_empty(&obd->obd_lock_replay_queue)) {
1656 CDEBUG(D_HA, "waking for next lock\n");
1658 } else if (cfs_atomic_read(&obd->obd_lock_replay_clients) == 0) {
1659 CDEBUG(D_HA, "waking for completed lock replay\n");
1661 } else if (obd->obd_abort_recovery) {
1662 CDEBUG(D_HA, "waking for aborted recovery\n");
1664 } else if (obd->obd_recovery_expired) {
1665 CDEBUG(D_HA, "waking for expired recovery\n");
1668 cfs_spin_unlock(&obd->obd_recovery_task_lock);
1674 * wait for recovery events,
1675 * check its status with help of check_routine
1676 * evict dead clients via health_check
1678 static int target_recovery_overseer(struct obd_device *obd,
1679 int (*check_routine)(struct obd_device *),
1680 int (*health_check)(struct obd_export *))
1683 cfs_wait_event(obd->obd_next_transno_waitq, check_routine(obd));
1684 if (obd->obd_abort_recovery) {
1685 CWARN("recovery is aborted, evict exports in recovery\n");
1686 /** evict exports which didn't finish recovery yet */
1687 class_disconnect_stale_exports(obd, exp_finished);
1689 } else if (obd->obd_recovery_expired) {
1690 obd->obd_recovery_expired = 0;
1691 /** If some clients died being recovered, evict them */
1693 "recovery is timed out, evict stale exports\n");
1694 /** evict cexports with no replay in queue, they are stalled */
1695 class_disconnect_stale_exports(obd, health_check);
1696 /** continue with VBR */
1697 cfs_spin_lock(&obd->obd_dev_lock);
1698 obd->obd_version_recov = 1;
1699 cfs_spin_unlock(&obd->obd_dev_lock);
1701 * reset timer, recovery will proceed with versions now,
1702 * timeout is set just to handle reconnection delays
1704 extend_recovery_timer(obd, RECONNECT_DELAY_MAX, true);
1705 /** Wait for recovery events again, after evicting bad clients */
1711 static struct ptlrpc_request *target_next_replay_req(struct obd_device *obd)
1713 struct ptlrpc_request *req = NULL;
1716 CDEBUG(D_HA, "Waiting for transno "LPD64"\n",
1717 obd->obd_next_recovery_transno);
1719 if (target_recovery_overseer(obd, check_for_next_transno,
1720 exp_req_replay_healthy)) {
1721 abort_req_replay_queue(obd);
1722 abort_lock_replay_queue(obd);
1725 cfs_spin_lock(&obd->obd_recovery_task_lock);
1726 if (!cfs_list_empty(&obd->obd_req_replay_queue)) {
1727 req = cfs_list_entry(obd->obd_req_replay_queue.next,
1728 struct ptlrpc_request, rq_list);
1729 cfs_list_del_init(&req->rq_list);
1730 obd->obd_requests_queued_for_recovery--;
1731 cfs_spin_unlock(&obd->obd_recovery_task_lock);
1733 cfs_spin_unlock(&obd->obd_recovery_task_lock);
1734 LASSERT(cfs_list_empty(&obd->obd_req_replay_queue));
1735 LASSERT(cfs_atomic_read(&obd->obd_req_replay_clients) == 0);
1736 /** evict exports failed VBR */
1737 class_disconnect_stale_exports(obd, exp_vbr_healthy);
1742 static struct ptlrpc_request *target_next_replay_lock(struct obd_device *obd)
1744 struct ptlrpc_request *req = NULL;
1746 CDEBUG(D_HA, "Waiting for lock\n");
1747 if (target_recovery_overseer(obd, check_for_next_lock,
1748 exp_lock_replay_healthy))
1749 abort_lock_replay_queue(obd);
1751 cfs_spin_lock(&obd->obd_recovery_task_lock);
1752 if (!cfs_list_empty(&obd->obd_lock_replay_queue)) {
1753 req = cfs_list_entry(obd->obd_lock_replay_queue.next,
1754 struct ptlrpc_request, rq_list);
1755 cfs_list_del_init(&req->rq_list);
1756 cfs_spin_unlock(&obd->obd_recovery_task_lock);
1758 cfs_spin_unlock(&obd->obd_recovery_task_lock);
1759 LASSERT(cfs_list_empty(&obd->obd_lock_replay_queue));
1760 LASSERT(cfs_atomic_read(&obd->obd_lock_replay_clients) == 0);
1761 /** evict exports failed VBR */
1762 class_disconnect_stale_exports(obd, exp_vbr_healthy);
1767 static struct ptlrpc_request *target_next_final_ping(struct obd_device *obd)
1769 struct ptlrpc_request *req = NULL;
1771 cfs_spin_lock(&obd->obd_recovery_task_lock);
1772 if (!cfs_list_empty(&obd->obd_final_req_queue)) {
1773 req = cfs_list_entry(obd->obd_final_req_queue.next,
1774 struct ptlrpc_request, rq_list);
1775 cfs_list_del_init(&req->rq_list);
1776 cfs_spin_unlock(&obd->obd_recovery_task_lock);
1777 if (req->rq_export->exp_in_recovery) {
1778 cfs_spin_lock(&req->rq_export->exp_lock);
1779 req->rq_export->exp_in_recovery = 0;
1780 cfs_spin_unlock(&req->rq_export->exp_lock);
1783 cfs_spin_unlock(&obd->obd_recovery_task_lock);
1788 static int handle_recovery_req(struct ptlrpc_thread *thread,
1789 struct ptlrpc_request *req,
1790 svc_handler_t handler)
1796 * export can be evicted during recovery, no need to handle replays for
1797 * it after that, discard such request silently
1799 if (req->rq_export->exp_disconnected)
1800 GOTO(reqcopy_put, rc = 0);
1802 rc = lu_context_init(&req->rq_recov_session, LCT_SESSION);
1804 CERROR("Failure to initialize session: %d\n", rc);
1805 GOTO(reqcopy_put, rc);
1808 req->rq_recov_session.lc_thread = thread;
1809 lu_context_enter(&req->rq_recov_session);
1810 req->rq_svc_thread = thread;
1811 req->rq_svc_thread->t_env->le_ses = &req->rq_recov_session;
1813 /* thread context */
1814 lu_context_enter(&thread->t_env->le_ctx);
1816 lu_context_exit(&thread->t_env->le_ctx);
1818 lu_context_exit(&req->rq_recov_session);
1819 lu_context_fini(&req->rq_recov_session);
1820 /* don't reset timer for final stage */
1821 if (!exp_finished(req->rq_export)) {
1822 int to = obd_timeout;
1825 * Add request timeout to the recovery time so next request from
1826 * this client may come in recovery time
1829 struct ptlrpc_service *svc = req->rq_rqbd->rqbd_service;
1830 /* If the server sent early reply for this request,
1831 * the client will recalculate the timeout according to
1832 * current server estimate service time, so we will
1833 * use the maxium timeout here for waiting the client
1834 * sending the next req */
1835 to = max((int)at_est2timeout(
1836 at_get(&svc->srv_at_estimate)),
1837 (int)lustre_msg_get_timeout(req->rq_reqmsg));
1838 /* Add net_latency (see ptlrpc_replay_req) */
1839 to += lustre_msg_get_service_time(req->rq_reqmsg);
1841 extend_recovery_timer(class_exp2obd(req->rq_export), to, true);
1847 static int target_recovery_thread(void *arg)
1849 struct lu_target *lut = arg;
1850 struct obd_device *obd = lut->lut_obd;
1851 struct ptlrpc_request *req;
1852 struct target_recovery_data *trd = &obd->obd_recovery_data;
1853 unsigned long delta;
1854 unsigned long flags;
1856 struct ptlrpc_thread *thread = NULL;
1860 cfs_daemonize_ctxt("tgt_recov");
1862 SIGNAL_MASK_LOCK(current, flags);
1863 sigfillset(¤t->blocked);
1865 SIGNAL_MASK_UNLOCK(current, flags);
1867 OBD_ALLOC_PTR(thread);
1873 OBD_FREE_PTR(thread);
1877 rc = lu_context_init(&env->le_ctx, LCT_MD_THREAD);
1879 OBD_FREE_PTR(thread);
1884 thread->t_env = env;
1885 thread->t_id = -1; /* force filter_iobuf_get/put to use local buffers */
1886 env->le_ctx.lc_thread = thread;
1887 thread->t_data = NULL;
1888 thread->t_watchdog = NULL;
1890 CDEBUG(D_HA, "%s: started recovery thread pid %d\n", obd->obd_name,
1892 trd->trd_processing_task = cfs_curproc_pid();
1894 cfs_spin_lock(&obd->obd_dev_lock);
1895 obd->obd_recovering = 1;
1896 cfs_spin_unlock(&obd->obd_dev_lock);
1897 cfs_complete(&trd->trd_starting);
1899 /* first of all, we have to know the first transno to replay */
1900 if (target_recovery_overseer(obd, check_for_clients,
1901 exp_connect_healthy)) {
1902 abort_req_replay_queue(obd);
1903 abort_lock_replay_queue(obd);
1906 /* next stage: replay requests */
1908 CDEBUG(D_INFO, "1: request replay stage - %d clients from t"LPU64"\n",
1909 cfs_atomic_read(&obd->obd_req_replay_clients),
1910 obd->obd_next_recovery_transno);
1911 while ((req = target_next_replay_req(obd))) {
1912 LASSERT(trd->trd_processing_task == cfs_curproc_pid());
1913 DEBUG_REQ(D_HA, req, "processing t"LPD64" from %s",
1914 lustre_msg_get_transno(req->rq_reqmsg),
1915 libcfs_nid2str(req->rq_peer.nid));
1916 handle_recovery_req(thread, req,
1917 trd->trd_recovery_handler);
1919 * bz18031: increase next_recovery_transno before
1920 * target_request_copy_put() will drop exp_rpc reference
1922 cfs_spin_lock(&obd->obd_recovery_task_lock);
1923 obd->obd_next_recovery_transno++;
1924 cfs_spin_unlock(&obd->obd_recovery_task_lock);
1925 target_exp_dequeue_req_replay(req);
1926 target_request_copy_put(req);
1927 obd->obd_replayed_requests++;
1931 * The second stage: replay locks
1933 CDEBUG(D_INFO, "2: lock replay stage - %d clients\n",
1934 cfs_atomic_read(&obd->obd_lock_replay_clients));
1935 while ((req = target_next_replay_lock(obd))) {
1936 LASSERT(trd->trd_processing_task == cfs_curproc_pid());
1937 DEBUG_REQ(D_HA, req, "processing lock from %s: ",
1938 libcfs_nid2str(req->rq_peer.nid));
1939 handle_recovery_req(thread, req,
1940 trd->trd_recovery_handler);
1941 target_request_copy_put(req);
1942 obd->obd_replayed_locks++;
1946 * The third stage: reply on final pings, at this moment all clients
1947 * must have request in final queue
1949 CDEBUG(D_INFO, "3: final stage - process recovery completion pings\n");
1950 /** Update server last boot epoch */
1951 lut_boot_epoch_update(lut);
1952 /* We drop recoverying flag to forward all new requests
1953 * to regular mds_handle() since now */
1954 cfs_spin_lock(&obd->obd_dev_lock);
1955 obd->obd_recovering = obd->obd_abort_recovery = 0;
1956 cfs_spin_unlock(&obd->obd_dev_lock);
1957 cfs_spin_lock(&obd->obd_recovery_task_lock);
1958 target_cancel_recovery_timer(obd);
1959 cfs_spin_unlock(&obd->obd_recovery_task_lock);
1960 while ((req = target_next_final_ping(obd))) {
1961 LASSERT(trd->trd_processing_task == cfs_curproc_pid());
1962 DEBUG_REQ(D_HA, req, "processing final ping from %s: ",
1963 libcfs_nid2str(req->rq_peer.nid));
1964 handle_recovery_req(thread, req,
1965 trd->trd_recovery_handler);
1966 target_request_copy_put(req);
1969 delta = (jiffies - delta) / CFS_HZ;
1970 CDEBUG(D_INFO,"4: recovery completed in %lus - %d/%d reqs/locks\n",
1971 delta, obd->obd_replayed_requests, obd->obd_replayed_locks);
1972 if (delta > OBD_RECOVERY_TIME_SOFT) {
1973 CWARN("too long recovery - read logs\n");
1974 libcfs_debug_dumplog();
1977 target_finish_recovery(obd);
1979 lu_context_fini(&env->le_ctx);
1980 trd->trd_processing_task = 0;
1981 cfs_complete(&trd->trd_finishing);
1983 OBD_FREE_PTR(thread);
1988 static int target_start_recovery_thread(struct lu_target *lut,
1989 svc_handler_t handler)
1991 struct obd_device *obd = lut->lut_obd;
1993 struct target_recovery_data *trd = &obd->obd_recovery_data;
1995 memset(trd, 0, sizeof(*trd));
1996 cfs_init_completion(&trd->trd_starting);
1997 cfs_init_completion(&trd->trd_finishing);
1998 trd->trd_recovery_handler = handler;
2000 if (cfs_create_thread(target_recovery_thread, lut, 0) > 0) {
2001 cfs_wait_for_completion(&trd->trd_starting);
2002 LASSERT(obd->obd_recovering != 0);
2009 void target_stop_recovery_thread(struct obd_device *obd)
2011 if (obd->obd_recovery_data.trd_processing_task > 0) {
2012 struct target_recovery_data *trd = &obd->obd_recovery_data;
2013 /** recovery can be done but postrecovery is not yet */
2014 cfs_spin_lock(&obd->obd_dev_lock);
2015 if (obd->obd_recovering) {
2016 CERROR("%s: Aborting recovery\n", obd->obd_name);
2017 obd->obd_abort_recovery = 1;
2018 cfs_waitq_signal(&obd->obd_next_transno_waitq);
2020 cfs_spin_unlock(&obd->obd_dev_lock);
2021 cfs_wait_for_completion(&trd->trd_finishing);
2025 void target_recovery_fini(struct obd_device *obd)
2027 class_disconnect_exports(obd);
2028 target_stop_recovery_thread(obd);
2029 target_cleanup_recovery(obd);
2031 EXPORT_SYMBOL(target_recovery_fini);
2033 static void target_recovery_expired(unsigned long castmeharder)
2035 struct obd_device *obd = (struct obd_device *)castmeharder;
2036 CDEBUG(D_HA, "%s: recovery timed out; %d clients are still in recovery"
2037 " after %lds (%d clients connected)\n",
2038 obd->obd_name, cfs_atomic_read(&obd->obd_lock_replay_clients),
2039 cfs_time_current_sec()- obd->obd_recovery_start,
2040 cfs_atomic_read(&obd->obd_connected_clients));
2042 obd->obd_recovery_expired = 1;
2043 cfs_waitq_signal(&obd->obd_next_transno_waitq);
2046 void target_recovery_init(struct lu_target *lut, svc_handler_t handler)
2048 struct obd_device *obd = lut->lut_obd;
2049 if (obd->obd_max_recoverable_clients == 0) {
2050 /** Update server last boot epoch */
2051 lut_boot_epoch_update(lut);
2055 CWARN("RECOVERY: service %s, %d recoverable clients, "
2056 "last_transno "LPU64"\n", obd->obd_name,
2057 obd->obd_max_recoverable_clients, obd->obd_last_committed);
2058 LASSERT(obd->obd_stopping == 0);
2059 obd->obd_next_recovery_transno = obd->obd_last_committed + 1;
2060 obd->obd_recovery_start = 0;
2061 obd->obd_recovery_end = 0;
2063 cfs_timer_init(&obd->obd_recovery_timer, target_recovery_expired, obd);
2064 target_start_recovery_thread(lut, handler);
2066 EXPORT_SYMBOL(target_recovery_init);
2070 static int target_process_req_flags(struct obd_device *obd,
2071 struct ptlrpc_request *req)
2073 struct obd_export *exp = req->rq_export;
2074 LASSERT(exp != NULL);
2075 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REQ_REPLAY_DONE) {
2076 /* client declares he's ready to replay locks */
2077 cfs_spin_lock(&exp->exp_lock);
2078 if (exp->exp_req_replay_needed) {
2079 exp->exp_req_replay_needed = 0;
2080 cfs_spin_unlock(&exp->exp_lock);
2082 LASSERT_ATOMIC_POS(&obd->obd_req_replay_clients);
2083 cfs_atomic_dec(&obd->obd_req_replay_clients);
2085 cfs_spin_unlock(&exp->exp_lock);
2088 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_LOCK_REPLAY_DONE) {
2089 /* client declares he's ready to complete recovery
2090 * so, we put the request on th final queue */
2091 cfs_spin_lock(&exp->exp_lock);
2092 if (exp->exp_lock_replay_needed) {
2093 exp->exp_lock_replay_needed = 0;
2094 cfs_spin_unlock(&exp->exp_lock);
2096 LASSERT_ATOMIC_POS(&obd->obd_lock_replay_clients);
2097 cfs_atomic_dec(&obd->obd_lock_replay_clients);
2099 cfs_spin_unlock(&exp->exp_lock);
2105 int target_queue_recovery_request(struct ptlrpc_request *req,
2106 struct obd_device *obd)
2110 __u64 transno = lustre_msg_get_transno(req->rq_reqmsg);
2113 if (obd->obd_recovery_data.trd_processing_task == cfs_curproc_pid()) {
2114 /* Processing the queue right now, don't re-add. */
2118 target_process_req_flags(obd, req);
2120 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_LOCK_REPLAY_DONE) {
2121 /* client declares he's ready to complete recovery
2122 * so, we put the request on th final queue */
2123 target_request_copy_get(req);
2124 DEBUG_REQ(D_HA, req, "queue final req");
2125 cfs_waitq_signal(&obd->obd_next_transno_waitq);
2126 cfs_spin_lock(&obd->obd_recovery_task_lock);
2127 if (obd->obd_recovering) {
2128 cfs_list_add_tail(&req->rq_list,
2129 &obd->obd_final_req_queue);
2131 cfs_spin_unlock(&obd->obd_recovery_task_lock);
2132 target_request_copy_put(req);
2133 RETURN(obd->obd_stopping ? -ENOTCONN : 1);
2135 cfs_spin_unlock(&obd->obd_recovery_task_lock);
2138 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REQ_REPLAY_DONE) {
2139 /* client declares he's ready to replay locks */
2140 target_request_copy_get(req);
2141 DEBUG_REQ(D_HA, req, "queue lock replay req");
2142 cfs_waitq_signal(&obd->obd_next_transno_waitq);
2143 cfs_spin_lock(&obd->obd_recovery_task_lock);
2144 LASSERT(obd->obd_recovering);
2145 /* usually due to recovery abort */
2146 if (!req->rq_export->exp_in_recovery) {
2147 cfs_spin_unlock(&obd->obd_recovery_task_lock);
2148 target_request_copy_put(req);
2151 LASSERT(req->rq_export->exp_lock_replay_needed);
2152 cfs_list_add_tail(&req->rq_list, &obd->obd_lock_replay_queue);
2153 cfs_spin_unlock(&obd->obd_recovery_task_lock);
2157 /* CAVEAT EMPTOR: The incoming request message has been swabbed
2158 * (i.e. buflens etc are in my own byte order), but type-dependent
2159 * buffers (eg mdt_body, ost_body etc) have NOT been swabbed. */
2162 CFS_INIT_LIST_HEAD(&req->rq_list);
2163 DEBUG_REQ(D_HA, req, "not queueing");
2167 /* If we're processing the queue, we want don't want to queue this
2170 * Also, if this request has a transno less than the one we're waiting
2171 * for, we should process it now. It could (and currently always will)
2172 * be an open request for a descriptor that was opened some time ago.
2174 * Also, a resent, replayed request that has already been
2175 * handled will pass through here and be processed immediately.
2177 CWARN("Next recovery transno: "LPU64", current: "LPU64", replaying\n",
2178 obd->obd_next_recovery_transno, transno);
2179 cfs_spin_lock(&obd->obd_recovery_task_lock);
2180 if (transno < obd->obd_next_recovery_transno) {
2181 /* Processing the queue right now, don't re-add. */
2182 LASSERT(cfs_list_empty(&req->rq_list));
2183 cfs_spin_unlock(&obd->obd_recovery_task_lock);
2186 cfs_spin_unlock(&obd->obd_recovery_task_lock);
2188 if (OBD_FAIL_CHECK(OBD_FAIL_TGT_REPLAY_DROP))
2191 target_request_copy_get(req);
2192 if (!req->rq_export->exp_in_recovery) {
2193 target_request_copy_put(req);
2196 LASSERT(req->rq_export->exp_req_replay_needed);
2198 if (target_exp_enqueue_req_replay(req)) {
2199 DEBUG_REQ(D_ERROR, req, "dropping resent queued req");
2200 target_request_copy_put(req);
2205 cfs_spin_lock(&obd->obd_recovery_task_lock);
2206 LASSERT(obd->obd_recovering);
2207 cfs_list_for_each(tmp, &obd->obd_req_replay_queue) {
2208 struct ptlrpc_request *reqiter =
2209 cfs_list_entry(tmp, struct ptlrpc_request, rq_list);
2211 if (lustre_msg_get_transno(reqiter->rq_reqmsg) > transno) {
2212 cfs_list_add_tail(&req->rq_list, &reqiter->rq_list);
2217 if (unlikely(lustre_msg_get_transno(reqiter->rq_reqmsg) ==
2219 DEBUG_REQ(D_ERROR, req, "dropping replay: transno "
2220 "has been claimed by another client");
2221 cfs_spin_unlock(&obd->obd_recovery_task_lock);
2222 target_exp_dequeue_req_replay(req);
2223 target_request_copy_put(req);
2229 cfs_list_add_tail(&req->rq_list, &obd->obd_req_replay_queue);
2231 obd->obd_requests_queued_for_recovery++;
2232 cfs_spin_unlock(&obd->obd_recovery_task_lock);
2233 cfs_waitq_signal(&obd->obd_next_transno_waitq);
2238 * Packs current SLV and Limit into \a req.
2240 int target_pack_pool_reply(struct ptlrpc_request *req)
2242 struct obd_device *obd;
2246 * Check that we still have all structures alive as this may
2247 * be some late rpc in shutdown time.
2249 if (unlikely(!req->rq_export || !req->rq_export->exp_obd ||
2250 !exp_connect_lru_resize(req->rq_export))) {
2251 lustre_msg_set_slv(req->rq_repmsg, 0);
2252 lustre_msg_set_limit(req->rq_repmsg, 0);
2257 * OBD is alive here as export is alive, which we checked above.
2259 obd = req->rq_export->exp_obd;
2261 cfs_read_lock(&obd->obd_pool_lock);
2262 lustre_msg_set_slv(req->rq_repmsg, obd->obd_pool_slv);
2263 lustre_msg_set_limit(req->rq_repmsg, obd->obd_pool_limit);
2264 cfs_read_unlock(&obd->obd_pool_lock);
2269 int target_send_reply_msg(struct ptlrpc_request *req, int rc, int fail_id)
2271 if (OBD_FAIL_CHECK_ORSET(fail_id & ~OBD_FAIL_ONCE, OBD_FAIL_ONCE)) {
2272 DEBUG_REQ(D_ERROR, req, "dropping reply");
2277 DEBUG_REQ(D_NET, req, "processing error (%d)", rc);
2278 req->rq_status = rc;
2279 return (ptlrpc_send_error(req, 1));
2281 DEBUG_REQ(D_NET, req, "sending reply");
2284 return (ptlrpc_send_reply(req, PTLRPC_REPLY_MAYBE_DIFFICULT));
2287 void target_send_reply(struct ptlrpc_request *req, int rc, int fail_id)
2290 struct ptlrpc_reply_state *rs;
2291 struct obd_export *exp;
2292 struct ptlrpc_service *svc;
2295 if (req->rq_no_reply) {
2300 svc = req->rq_rqbd->rqbd_service;
2301 rs = req->rq_reply_state;
2302 if (rs == NULL || !rs->rs_difficult) {
2304 target_send_reply_msg (req, rc, fail_id);
2309 /* must be an export if locks saved */
2310 LASSERT (req->rq_export != NULL);
2311 /* req/reply consistent */
2312 LASSERT (rs->rs_service == svc);
2315 LASSERT (!rs->rs_scheduled);
2316 LASSERT (!rs->rs_scheduled_ever);
2317 LASSERT (!rs->rs_handled);
2318 LASSERT (!rs->rs_on_net);
2319 LASSERT (rs->rs_export == NULL);
2320 LASSERT (cfs_list_empty(&rs->rs_obd_list));
2321 LASSERT (cfs_list_empty(&rs->rs_exp_list));
2323 exp = class_export_get (req->rq_export);
2325 /* disable reply scheduling while I'm setting up */
2326 rs->rs_scheduled = 1;
2328 rs->rs_xid = req->rq_xid;
2329 rs->rs_transno = req->rq_transno;
2330 rs->rs_export = exp;
2331 rs->rs_opc = lustre_msg_get_opc(rs->rs_msg);
2333 cfs_spin_lock(&exp->exp_uncommitted_replies_lock);
2334 CDEBUG(D_NET, "rs transno = "LPU64", last committed = "LPU64"\n",
2335 rs->rs_transno, exp->exp_last_committed);
2336 if (rs->rs_transno > exp->exp_last_committed) {
2337 /* not committed already */
2338 cfs_list_add_tail(&rs->rs_obd_list,
2339 &exp->exp_uncommitted_replies);
2341 cfs_spin_unlock (&exp->exp_uncommitted_replies_lock);
2343 cfs_spin_lock(&exp->exp_lock);
2344 cfs_list_add_tail(&rs->rs_exp_list, &exp->exp_outstanding_replies);
2345 cfs_spin_unlock(&exp->exp_lock);
2347 netrc = target_send_reply_msg (req, rc, fail_id);
2349 cfs_spin_lock(&svc->srv_rs_lock);
2351 cfs_atomic_inc(&svc->srv_n_difficult_replies);
2354 /* error sending: reply is off the net. Also we need +1
2355 * reply ref until ptlrpc_handle_rs() is done
2356 * with the reply state (if the send was successful, there
2357 * would have been +1 ref for the net, which
2358 * reply_out_callback leaves alone) */
2360 ptlrpc_rs_addref(rs);
2363 cfs_spin_lock(&rs->rs_lock);
2364 if (rs->rs_transno <= exp->exp_last_committed ||
2365 (!rs->rs_on_net && !rs->rs_no_ack) ||
2366 cfs_list_empty(&rs->rs_exp_list) || /* completed already */
2367 cfs_list_empty(&rs->rs_obd_list)) {
2368 CDEBUG(D_HA, "Schedule reply immediately\n");
2369 ptlrpc_dispatch_difficult_reply(rs);
2371 cfs_list_add (&rs->rs_list, &svc->srv_active_replies);
2372 rs->rs_scheduled = 0; /* allow notifier to schedule */
2374 cfs_spin_unlock(&rs->rs_lock);
2375 cfs_spin_unlock(&svc->srv_rs_lock);
2379 int target_handle_ping(struct ptlrpc_request *req)
2381 obd_ping(req->rq_export);
2382 return req_capsule_server_pack(&req->rq_pill);
2385 void target_committed_to_req(struct ptlrpc_request *req)
2387 struct obd_export *exp = req->rq_export;
2389 if (!exp->exp_obd->obd_no_transno && req->rq_repmsg != NULL)
2390 lustre_msg_set_last_committed(req->rq_repmsg,
2391 exp->exp_last_committed);
2393 DEBUG_REQ(D_IOCTL, req, "not sending last_committed update (%d/"
2394 "%d)", exp->exp_obd->obd_no_transno,
2395 req->rq_repmsg == NULL);
2397 CDEBUG(D_INFO, "last_committed "LPU64", transno "LPU64", xid "LPU64"\n",
2398 exp->exp_last_committed, req->rq_transno, req->rq_xid);
2400 EXPORT_SYMBOL(target_committed_to_req);
2402 int target_handle_qc_callback(struct ptlrpc_request *req)
2404 struct obd_quotactl *oqctl;
2405 struct client_obd *cli = &req->rq_export->exp_obd->u.cli;
2407 oqctl = req_capsule_client_get(&req->rq_pill, &RMF_OBD_QUOTACTL);
2408 if (oqctl == NULL) {
2409 CERROR("Can't unpack obd_quotactl\n");
2413 cli->cl_qchk_stat = oqctl->qc_stat;
2418 #ifdef HAVE_QUOTA_SUPPORT
2419 int target_handle_dqacq_callback(struct ptlrpc_request *req)
2422 struct obd_device *obd = req->rq_export->exp_obd;
2423 struct obd_device *master_obd = NULL, *lov_obd = NULL;
2424 struct obd_device_target *obt;
2425 struct lustre_quota_ctxt *qctxt;
2426 struct qunit_data *qdata = NULL;
2430 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_DROP_QUOTA_REQ))
2433 rc = req_capsule_server_pack(&req->rq_pill);
2435 CERROR("packing reply failed!: rc = %d\n", rc);
2439 LASSERT(req->rq_export);
2441 qdata = quota_get_qdata(req, QUOTA_REQUEST, QUOTA_EXPORT);
2442 if (IS_ERR(qdata)) {
2443 rc = PTR_ERR(qdata);
2444 CDEBUG(D_ERROR, "Can't unpack qunit_data(rc: %d)\n", rc);
2445 req->rq_status = rc;
2449 /* we use the observer */
2450 if (obd_pin_observer(obd, &lov_obd) ||
2451 obd_pin_observer(lov_obd, &master_obd)) {
2452 CERROR("Can't find the observer, it is recovering\n");
2453 req->rq_status = -EAGAIN;
2457 obt = &master_obd->u.obt;
2458 qctxt = &obt->obt_qctxt;
2460 if (!qctxt->lqc_setup || !qctxt->lqc_valid) {
2461 /* quota_type has not been processed yet, return EAGAIN
2462 * until we know whether or not quotas are supposed to
2464 CDEBUG(D_QUOTA, "quota_type not processed yet, return "
2466 req->rq_status = -EAGAIN;
2470 cfs_down_read(&obt->obt_rwsem);
2471 if (qctxt->lqc_lqs_hash == NULL) {
2472 cfs_up_read(&obt->obt_rwsem);
2473 /* quota_type has not been processed yet, return EAGAIN
2474 * until we know whether or not quotas are supposed to
2476 CDEBUG(D_QUOTA, "quota_ctxt is not ready yet, return "
2478 req->rq_status = -EAGAIN;
2482 LASSERT(qctxt->lqc_handler);
2483 rc = qctxt->lqc_handler(master_obd, qdata,
2484 lustre_msg_get_opc(req->rq_reqmsg));
2485 cfs_up_read(&obt->obt_rwsem);
2486 if (rc && rc != -EDQUOT)
2487 CDEBUG(rc == -EBUSY ? D_QUOTA : D_ERROR,
2488 "dqacq/dqrel failed! (rc:%d)\n", rc);
2489 req->rq_status = rc;
2491 rc = quota_copy_qdata(req, qdata, QUOTA_REPLY, QUOTA_EXPORT);
2493 CERROR("Can't pack qunit_data(rc: %d)\n", rc);
2497 /* Block the quota req. b=14840 */
2498 OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_BLOCK_QUOTA_REQ, obd_timeout);
2503 obd_unpin_observer(lov_obd);
2505 obd_unpin_observer(obd);
2507 rc = ptlrpc_reply(req);
2511 #endif /* !__KERNEL__ */
2513 #endif /* HAVE_QUOTA_SUPPORT */
2515 ldlm_mode_t lck_compat_array[] = {
2516 [LCK_EX] LCK_COMPAT_EX,
2517 [LCK_PW] LCK_COMPAT_PW,
2518 [LCK_PR] LCK_COMPAT_PR,
2519 [LCK_CW] LCK_COMPAT_CW,
2520 [LCK_CR] LCK_COMPAT_CR,
2521 [LCK_NL] LCK_COMPAT_NL,
2522 [LCK_GROUP] LCK_COMPAT_GROUP,
2523 [LCK_COS] LCK_COMPAT_COS,
2527 * Rather arbitrary mapping from LDLM error codes to errno values. This should
2528 * not escape to the user level.
2530 int ldlm_error2errno(ldlm_error_t error)
2538 case ELDLM_LOCK_CHANGED:
2541 case ELDLM_LOCK_ABORTED:
2544 case ELDLM_LOCK_REPLACED:
2547 case ELDLM_NO_LOCK_DATA:
2550 case ELDLM_NAMESPACE_EXISTS:
2553 case ELDLM_BAD_NAMESPACE:
2557 if (((int)error) < 0) /* cast to signed type */
2558 result = error; /* as ldlm_error_t can be unsigned */
2560 CERROR("Invalid DLM result code: %d\n", error);
2566 EXPORT_SYMBOL(ldlm_error2errno);
2569 * Dual to ldlm_error2errno(): maps errno values back to ldlm_error_t.
2571 ldlm_error_t ldlm_errno2error(int err_no)
2580 error = ELDLM_LOCK_CHANGED;
2583 error = ELDLM_LOCK_ABORTED;
2586 error = ELDLM_LOCK_REPLACED;
2589 error = ELDLM_NO_LOCK_DATA;
2592 error = ELDLM_NAMESPACE_EXISTS;
2595 error = ELDLM_BAD_NAMESPACE;
2602 EXPORT_SYMBOL(ldlm_errno2error);
2604 #if LUSTRE_TRACKS_LOCK_EXP_REFS
2605 void ldlm_dump_export_locks(struct obd_export *exp)
2607 cfs_spin_lock(&exp->exp_locks_list_guard);
2608 if (!cfs_list_empty(&exp->exp_locks_list)) {
2609 struct ldlm_lock *lock;
2611 CERROR("dumping locks for export %p,"
2612 "ignore if the unmount doesn't hang\n", exp);
2613 cfs_list_for_each_entry(lock, &exp->exp_locks_list, l_exp_refs_link)
2614 ldlm_lock_dump(D_ERROR, lock, 0);
2616 cfs_spin_unlock(&exp->exp_locks_list_guard);
2620 static int target_bulk_timeout(void *data)
2623 /* We don't fail the connection here, because having the export
2624 * killed makes the (vital) call to commitrw very sad.
2629 static inline char *bulk2type(struct ptlrpc_bulk_desc *desc)
2631 return desc->bd_type == BULK_GET_SINK ? "GET" : "PUT";
2634 int target_bulk_io(struct obd_export *exp, struct ptlrpc_bulk_desc *desc,
2635 struct l_wait_info *lwi)
2637 struct ptlrpc_request *req = desc->bd_req;
2641 /* Check if there is eviction in progress, and if so, wait for
2643 if (unlikely(cfs_atomic_read(&exp->exp_obd->obd_evict_inprogress))) {
2644 *lwi = LWI_INTR(NULL, NULL);
2645 rc = l_wait_event(exp->exp_obd->obd_evict_inprogress_waitq,
2646 !cfs_atomic_read(&exp->exp_obd->
2647 obd_evict_inprogress),
2651 /* Check if client was evicted or tried to reconnect already */
2652 if (exp->exp_failed || exp->exp_abort_active_req) {
2655 if (desc->bd_type == BULK_PUT_SINK)
2656 rc = sptlrpc_svc_wrap_bulk(req, desc);
2658 rc = ptlrpc_start_bulk_transfer(desc);
2661 if (rc == 0 && OBD_FAIL_CHECK(OBD_FAIL_MDS_SENDPAGE)) {
2662 ptlrpc_abort_bulk(desc);
2663 } else if (rc == 0) {
2664 time_t start = cfs_time_current_sec();
2666 long timeoutl = req->rq_deadline - cfs_time_current_sec();
2667 cfs_duration_t timeout = timeoutl <= 0 ?
2668 CFS_TICK : cfs_time_seconds(timeoutl);
2669 *lwi = LWI_TIMEOUT_INTERVAL(timeout,
2670 cfs_time_seconds(1),
2671 target_bulk_timeout,
2673 rc = l_wait_event(desc->bd_waitq,
2674 !ptlrpc_server_bulk_active(desc) ||
2676 exp->exp_abort_active_req,
2678 LASSERT(rc == 0 || rc == -ETIMEDOUT);
2679 /* Wait again if we changed deadline */
2680 } while ((rc == -ETIMEDOUT) &&
2681 (req->rq_deadline > cfs_time_current_sec()));
2683 if (rc == -ETIMEDOUT) {
2684 DEBUG_REQ(D_ERROR, req,
2685 "timeout on bulk %s after %ld%+lds",
2687 req->rq_deadline - start,
2688 cfs_time_current_sec() -
2690 ptlrpc_abort_bulk(desc);
2691 } else if (exp->exp_failed) {
2692 DEBUG_REQ(D_ERROR, req, "Eviction on bulk %s",
2695 ptlrpc_abort_bulk(desc);
2696 } else if (exp->exp_abort_active_req) {
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_success ||
2703 desc->bd_nob_transferred != desc->bd_nob) {
2704 DEBUG_REQ(D_ERROR, req, "%s bulk %s %d(%d)",
2706 "truncated" : "network error on",
2708 desc->bd_nob_transferred,
2710 /* XXX should this be a different errno? */
2712 } else if (desc->bd_type == BULK_GET_SINK) {
2713 rc = sptlrpc_svc_unwrap_bulk(req, desc);
2716 DEBUG_REQ(D_ERROR, req, "bulk %s failed: rc %d",
2717 bulk2type(desc), rc);
2722 EXPORT_SYMBOL(target_bulk_io);