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) 2011, 2012, Whamcloud, Inc.
33 * This file is part of Lustre, http://www.lustre.org/
34 * Lustre is a trademark of Sun Microsystems, Inc.
38 # define EXPORT_SYMTAB
40 #define DEBUG_SUBSYSTEM S_LDLM
43 # include <libcfs/libcfs.h>
45 # include <liblustre.h>
48 #include <lustre_mds.h>
49 #include <lustre_dlm.h>
50 #include <lustre_net.h>
51 #include <lustre_sec.h>
52 #include "ldlm_internal.h"
54 /* @priority: if non-zero, move the selected to the list head
55 * @create: if zero, only search in existed connections
57 static int import_set_conn(struct obd_import *imp, struct obd_uuid *uuid,
58 int priority, int create)
60 struct ptlrpc_connection *ptlrpc_conn;
61 struct obd_import_conn *imp_conn = NULL, *item;
65 if (!create && !priority) {
66 CDEBUG(D_HA, "Nothing to do\n");
70 ptlrpc_conn = ptlrpc_uuid_to_connection(uuid);
72 CDEBUG(D_HA, "can't find connection %s\n", uuid->uuid);
77 OBD_ALLOC(imp_conn, sizeof(*imp_conn));
79 GOTO(out_put, rc = -ENOMEM);
83 cfs_spin_lock(&imp->imp_lock);
84 cfs_list_for_each_entry(item, &imp->imp_conn_list, oic_item) {
85 if (obd_uuid_equals(uuid, &item->oic_uuid)) {
87 cfs_list_del(&item->oic_item);
88 cfs_list_add(&item->oic_item,
90 item->oic_last_attempt = 0;
92 CDEBUG(D_HA, "imp %p@%s: found existing conn %s%s\n",
93 imp, imp->imp_obd->obd_name, uuid->uuid,
94 (priority ? ", moved to head" : ""));
95 cfs_spin_unlock(&imp->imp_lock);
96 GOTO(out_free, rc = 0);
101 imp_conn->oic_conn = ptlrpc_conn;
102 imp_conn->oic_uuid = *uuid;
103 imp_conn->oic_last_attempt = 0;
105 cfs_list_add(&imp_conn->oic_item, &imp->imp_conn_list);
107 cfs_list_add_tail(&imp_conn->oic_item,
108 &imp->imp_conn_list);
109 CDEBUG(D_HA, "imp %p@%s: add connection %s at %s\n",
110 imp, imp->imp_obd->obd_name, uuid->uuid,
111 (priority ? "head" : "tail"));
113 cfs_spin_unlock(&imp->imp_lock);
114 GOTO(out_free, rc = -ENOENT);
117 cfs_spin_unlock(&imp->imp_lock);
121 OBD_FREE(imp_conn, sizeof(*imp_conn));
123 ptlrpc_connection_put(ptlrpc_conn);
127 int import_set_conn_priority(struct obd_import *imp, struct obd_uuid *uuid)
129 return import_set_conn(imp, uuid, 1, 0);
132 int client_import_add_conn(struct obd_import *imp, struct obd_uuid *uuid,
135 return import_set_conn(imp, uuid, priority, 1);
138 int client_import_del_conn(struct obd_import *imp, struct obd_uuid *uuid)
140 struct obd_import_conn *imp_conn;
141 struct obd_export *dlmexp;
145 cfs_spin_lock(&imp->imp_lock);
146 if (cfs_list_empty(&imp->imp_conn_list)) {
147 LASSERT(!imp->imp_connection);
151 cfs_list_for_each_entry(imp_conn, &imp->imp_conn_list, oic_item) {
152 if (!obd_uuid_equals(uuid, &imp_conn->oic_uuid))
154 LASSERT(imp_conn->oic_conn);
156 /* is current conn? */
157 if (imp_conn == imp->imp_conn_current) {
158 LASSERT(imp_conn->oic_conn == imp->imp_connection);
160 if (imp->imp_state != LUSTRE_IMP_CLOSED &&
161 imp->imp_state != LUSTRE_IMP_DISCON) {
162 CERROR("can't remove current connection\n");
163 GOTO(out, rc = -EBUSY);
166 ptlrpc_connection_put(imp->imp_connection);
167 imp->imp_connection = NULL;
169 dlmexp = class_conn2export(&imp->imp_dlm_handle);
170 if (dlmexp && dlmexp->exp_connection) {
171 LASSERT(dlmexp->exp_connection ==
173 ptlrpc_connection_put(dlmexp->exp_connection);
174 dlmexp->exp_connection = NULL;
178 cfs_list_del(&imp_conn->oic_item);
179 ptlrpc_connection_put(imp_conn->oic_conn);
180 OBD_FREE(imp_conn, sizeof(*imp_conn));
181 CDEBUG(D_HA, "imp %p@%s: remove connection %s\n",
182 imp, imp->imp_obd->obd_name, uuid->uuid);
187 cfs_spin_unlock(&imp->imp_lock);
189 CERROR("connection %s not found\n", uuid->uuid);
194 * Find conn uuid by peer nid. @peer is a server nid. This function is used
195 * to find a conn uuid of @imp which can reach @peer.
197 int client_import_find_conn(struct obd_import *imp, lnet_nid_t peer,
198 struct obd_uuid *uuid)
200 struct obd_import_conn *conn;
204 cfs_spin_lock(&imp->imp_lock);
205 cfs_list_for_each_entry(conn, &imp->imp_conn_list, oic_item) {
206 /* check if conn uuid does have this peer nid */
207 if (class_check_uuid(&conn->oic_uuid, peer)) {
208 *uuid = conn->oic_uuid;
213 cfs_spin_unlock(&imp->imp_lock);
216 EXPORT_SYMBOL(client_import_find_conn);
218 void client_destroy_import(struct obd_import *imp)
220 /* drop security policy instance after all rpc finished/aborted
221 * to let all busy contexts be released. */
222 class_import_get(imp);
223 class_destroy_import(imp);
224 sptlrpc_import_sec_put(imp);
225 class_import_put(imp);
227 EXPORT_SYMBOL(client_destroy_import);
229 /* configure an RPC client OBD device
234 * 3 - inactive-on-startup
236 int client_obd_setup(struct obd_device *obddev, struct lustre_cfg *lcfg)
238 struct client_obd *cli = &obddev->u.cli;
239 struct obd_import *imp;
240 struct obd_uuid server_uuid;
241 int rq_portal, rp_portal, connect_op;
242 char *name = obddev->obd_type->typ_name;
243 ldlm_ns_type_t ns_type = LDLM_NS_TYPE_UNKNOWN;
247 /* In a more perfect world, we would hang a ptlrpc_client off of
248 * obd_type and just use the values from there. */
249 if (!strcmp(name, LUSTRE_OSC_NAME)) {
250 rq_portal = OST_REQUEST_PORTAL;
251 rp_portal = OSC_REPLY_PORTAL;
252 connect_op = OST_CONNECT;
253 cli->cl_sp_me = LUSTRE_SP_CLI;
254 cli->cl_sp_to = LUSTRE_SP_OST;
255 ns_type = LDLM_NS_TYPE_OSC;
257 } else if (!strcmp(name, LUSTRE_MDC_NAME)) {
258 rq_portal = MDS_REQUEST_PORTAL;
259 rp_portal = MDC_REPLY_PORTAL;
260 connect_op = MDS_CONNECT;
261 cli->cl_sp_me = LUSTRE_SP_CLI;
262 cli->cl_sp_to = LUSTRE_SP_MDT;
263 ns_type = LDLM_NS_TYPE_MDC;
265 } else if (!strcmp(name, LUSTRE_MGC_NAME)) {
266 rq_portal = MGS_REQUEST_PORTAL;
267 rp_portal = MGC_REPLY_PORTAL;
268 connect_op = MGS_CONNECT;
269 cli->cl_sp_me = LUSTRE_SP_MGC;
270 cli->cl_sp_to = LUSTRE_SP_MGS;
271 cli->cl_flvr_mgc.sf_rpc = SPTLRPC_FLVR_INVALID;
272 ns_type = LDLM_NS_TYPE_MGC;
275 CERROR("unknown client OBD type \"%s\", can't setup\n",
280 if (LUSTRE_CFG_BUFLEN(lcfg, 1) < 1) {
281 CERROR("requires a TARGET UUID\n");
285 if (LUSTRE_CFG_BUFLEN(lcfg, 1) > 37) {
286 CERROR("client UUID must be less than 38 characters\n");
290 if (LUSTRE_CFG_BUFLEN(lcfg, 2) < 1) {
291 CERROR("setup requires a SERVER UUID\n");
295 if (LUSTRE_CFG_BUFLEN(lcfg, 2) > 37) {
296 CERROR("target UUID must be less than 38 characters\n");
300 cfs_init_rwsem(&cli->cl_sem);
301 cfs_sema_init(&cli->cl_mgc_sem, 1);
302 cli->cl_conn_count = 0;
303 memcpy(server_uuid.uuid, lustre_cfg_buf(lcfg, 2),
304 min_t(unsigned int, LUSTRE_CFG_BUFLEN(lcfg, 2),
305 sizeof(server_uuid)));
308 cli->cl_avail_grant = 0;
309 /* FIXME: should limit this for the sum of all cl_dirty_max */
310 cli->cl_dirty_max = OSC_MAX_DIRTY_DEFAULT * 1024 * 1024;
311 if (cli->cl_dirty_max >> CFS_PAGE_SHIFT > cfs_num_physpages / 8)
312 cli->cl_dirty_max = cfs_num_physpages << (CFS_PAGE_SHIFT - 3);
313 CFS_INIT_LIST_HEAD(&cli->cl_cache_waiters);
314 CFS_INIT_LIST_HEAD(&cli->cl_loi_ready_list);
315 CFS_INIT_LIST_HEAD(&cli->cl_loi_hp_ready_list);
316 CFS_INIT_LIST_HEAD(&cli->cl_loi_write_list);
317 CFS_INIT_LIST_HEAD(&cli->cl_loi_read_list);
318 client_obd_list_lock_init(&cli->cl_loi_list_lock);
319 cli->cl_r_in_flight = 0;
320 cli->cl_w_in_flight = 0;
322 cfs_spin_lock_init(&cli->cl_read_rpc_hist.oh_lock);
323 cfs_spin_lock_init(&cli->cl_write_rpc_hist.oh_lock);
324 cfs_spin_lock_init(&cli->cl_read_page_hist.oh_lock);
325 cfs_spin_lock_init(&cli->cl_write_page_hist.oh_lock);
326 cfs_spin_lock_init(&cli->cl_read_offset_hist.oh_lock);
327 cfs_spin_lock_init(&cli->cl_write_offset_hist.oh_lock);
328 cfs_waitq_init(&cli->cl_destroy_waitq);
329 cfs_atomic_set(&cli->cl_destroy_in_flight, 0);
330 #ifdef ENABLE_CHECKSUM
331 /* Turn on checksumming by default. */
332 cli->cl_checksum = 1;
334 * The supported checksum types will be worked out at connect time
335 * Set cl_chksum* to CRC32 for now to avoid returning screwed info
338 cli->cl_cksum_type = cli->cl_supp_cksum_types = OBD_CKSUM_CRC32;
340 cfs_atomic_set(&cli->cl_resends, OSC_DEFAULT_RESENDS);
342 /* This value may be changed at connect time in
343 ptlrpc_connect_interpret. */
344 cli->cl_max_pages_per_rpc = min((int)PTLRPC_MAX_BRW_PAGES,
345 (int)(1024 * 1024 >> CFS_PAGE_SHIFT));
347 if (!strcmp(name, LUSTRE_MDC_NAME)) {
348 cli->cl_max_rpcs_in_flight = MDC_MAX_RIF_DEFAULT;
349 } else if (cfs_num_physpages >> (20 - CFS_PAGE_SHIFT) <= 128 /* MB */) {
350 cli->cl_max_rpcs_in_flight = 2;
351 } else if (cfs_num_physpages >> (20 - CFS_PAGE_SHIFT) <= 256 /* MB */) {
352 cli->cl_max_rpcs_in_flight = 3;
353 } else if (cfs_num_physpages >> (20 - CFS_PAGE_SHIFT) <= 512 /* MB */) {
354 cli->cl_max_rpcs_in_flight = 4;
356 cli->cl_max_rpcs_in_flight = OSC_MAX_RIF_DEFAULT;
361 CERROR("ldlm_get_ref failed: %d\n", rc);
365 ptlrpc_init_client(rq_portal, rp_portal, name,
366 &obddev->obd_ldlm_client);
368 imp = class_new_import(obddev);
370 GOTO(err_ldlm, rc = -ENOENT);
371 imp->imp_client = &obddev->obd_ldlm_client;
372 imp->imp_connect_op = connect_op;
373 CFS_INIT_LIST_HEAD(&imp->imp_pinger_chain);
374 memcpy(cli->cl_target_uuid.uuid, lustre_cfg_buf(lcfg, 1),
375 LUSTRE_CFG_BUFLEN(lcfg, 1));
376 class_import_put(imp);
378 rc = client_import_add_conn(imp, &server_uuid, 1);
380 CERROR("can't add initial connection\n");
381 GOTO(err_import, rc);
384 cli->cl_import = imp;
385 /* cli->cl_max_mds_{easize,cookiesize} updated by mdc_init_ea_size() */
386 cli->cl_max_mds_easize = sizeof(struct lov_mds_md_v3);
387 cli->cl_max_mds_cookiesize = sizeof(struct llog_cookie);
389 if (LUSTRE_CFG_BUFLEN(lcfg, 3) > 0) {
390 if (!strcmp(lustre_cfg_string(lcfg, 3), "inactive")) {
391 CDEBUG(D_HA, "marking %s %s->%s as inactive\n",
392 name, obddev->obd_name,
393 cli->cl_target_uuid.uuid);
394 cfs_spin_lock(&imp->imp_lock);
395 imp->imp_deactive = 1;
396 cfs_spin_unlock(&imp->imp_lock);
400 obddev->obd_namespace = ldlm_namespace_new(obddev, obddev->obd_name,
401 LDLM_NAMESPACE_CLIENT,
402 LDLM_NAMESPACE_GREEDY,
404 if (obddev->obd_namespace == NULL) {
405 CERROR("Unable to create client namespace - %s\n",
407 GOTO(err_import, rc = -ENOMEM);
410 cli->cl_qchk_stat = CL_NOT_QUOTACHECKED;
415 class_destroy_import(imp);
423 int client_obd_cleanup(struct obd_device *obddev)
427 ldlm_namespace_free_post(obddev->obd_namespace);
428 obddev->obd_namespace = NULL;
430 LASSERT(obddev->u.cli.cl_import == NULL);
436 /* ->o_connect() method for client side (OSC and MDC and MGC) */
437 int client_connect_import(const struct lu_env *env,
438 struct obd_export **exp,
439 struct obd_device *obd, struct obd_uuid *cluuid,
440 struct obd_connect_data *data, void *localdata)
442 struct client_obd *cli = &obd->u.cli;
443 struct obd_import *imp = cli->cl_import;
444 struct obd_connect_data *ocd;
445 struct lustre_handle conn = { 0 };
450 cfs_down_write(&cli->cl_sem);
451 if (cli->cl_conn_count > 0 )
452 GOTO(out_sem, rc = -EALREADY);
454 rc = class_connect(&conn, obd, cluuid);
458 cli->cl_conn_count++;
459 *exp = class_conn2export(&conn);
461 LASSERT(obd->obd_namespace);
463 imp->imp_dlm_handle = conn;
464 rc = ptlrpc_init_import(imp);
468 ocd = &imp->imp_connect_data;
471 imp->imp_connect_flags_orig = data->ocd_connect_flags;
474 rc = ptlrpc_connect_import(imp);
476 LASSERT (imp->imp_state == LUSTRE_IMP_DISCON);
479 LASSERT((*exp)->exp_connection);
482 LASSERTF((ocd->ocd_connect_flags & data->ocd_connect_flags) ==
483 ocd->ocd_connect_flags, "old "LPX64", new "LPX64"\n",
484 data->ocd_connect_flags, ocd->ocd_connect_flags);
485 data->ocd_connect_flags = ocd->ocd_connect_flags;
488 ptlrpc_pinger_add_import(imp);
494 cli->cl_conn_count--;
495 class_disconnect(*exp);
499 cfs_up_write(&cli->cl_sem);
504 int client_disconnect_export(struct obd_export *exp)
506 struct obd_device *obd = class_exp2obd(exp);
507 struct client_obd *cli;
508 struct obd_import *imp;
513 CERROR("invalid export for disconnect: exp %p cookie "LPX64"\n",
514 exp, exp ? exp->exp_handle.h_cookie : -1);
519 imp = cli->cl_import;
521 cfs_down_write(&cli->cl_sem);
522 CDEBUG(D_INFO, "disconnect %s - %d\n", obd->obd_name,
525 if (!cli->cl_conn_count) {
526 CERROR("disconnecting disconnected device (%s)\n",
528 GOTO(out_disconnect, rc = -EINVAL);
531 cli->cl_conn_count--;
532 if (cli->cl_conn_count)
533 GOTO(out_disconnect, rc = 0);
535 /* Mark import deactivated now, so we don't try to reconnect if any
536 * of the cleanup RPCs fails (e.g. ldlm cancel, etc). We don't
537 * fully deactivate the import, or that would drop all requests. */
538 cfs_spin_lock(&imp->imp_lock);
539 imp->imp_deactive = 1;
540 cfs_spin_unlock(&imp->imp_lock);
542 /* Some non-replayable imports (MDS's OSCs) are pinged, so just
543 * delete it regardless. (It's safe to delete an import that was
545 (void)ptlrpc_pinger_del_import(imp);
547 if (obd->obd_namespace != NULL) {
548 /* obd_force == local only */
549 ldlm_cli_cancel_unused(obd->obd_namespace, NULL,
550 obd->obd_force ? LCF_LOCAL : 0, NULL);
551 ldlm_namespace_free_prior(obd->obd_namespace, imp, obd->obd_force);
555 * there's no need to hold sem during disconnecting an import,
556 * and actually it may cause deadlock in gss.
558 cfs_up_write(&cli->cl_sem);
559 rc = ptlrpc_disconnect_import(imp, 0);
560 cfs_down_write(&cli->cl_sem);
562 ptlrpc_invalidate_import(imp);
567 /* use server style - class_disconnect should be always called for
569 err = class_disconnect(exp);
573 cfs_up_write(&cli->cl_sem);
578 #ifdef HAVE_SERVER_SUPPORT
579 int server_disconnect_export(struct obd_export *exp)
584 /* Disconnect early so that clients can't keep using export */
585 rc = class_disconnect(exp);
586 /* close import for avoid sending any requests */
587 if (exp->exp_imp_reverse)
588 ptlrpc_cleanup_imp(exp->exp_imp_reverse);
590 if (exp->exp_obd->obd_namespace != NULL)
591 ldlm_cancel_locks_for_export(exp);
593 /* complete all outstanding replies */
594 cfs_spin_lock(&exp->exp_lock);
595 while (!cfs_list_empty(&exp->exp_outstanding_replies)) {
596 struct ptlrpc_reply_state *rs =
597 cfs_list_entry(exp->exp_outstanding_replies.next,
598 struct ptlrpc_reply_state, rs_exp_list);
599 struct ptlrpc_service *svc = rs->rs_service;
601 cfs_spin_lock(&svc->srv_rs_lock);
602 cfs_list_del_init(&rs->rs_exp_list);
603 cfs_spin_lock(&rs->rs_lock);
604 ptlrpc_schedule_difficult_reply(rs);
605 cfs_spin_unlock(&rs->rs_lock);
606 cfs_spin_unlock(&svc->srv_rs_lock);
608 cfs_spin_unlock(&exp->exp_lock);
613 /* --------------------------------------------------------------------------
614 * from old lib/target.c
615 * -------------------------------------------------------------------------- */
617 static int target_handle_reconnect(struct lustre_handle *conn,
618 struct obd_export *exp,
619 struct obd_uuid *cluuid)
623 if (exp->exp_connection && exp->exp_imp_reverse) {
624 struct lustre_handle *hdl;
625 struct obd_device *target;
627 hdl = &exp->exp_imp_reverse->imp_remote_handle;
628 target = exp->exp_obd;
630 /* Might be a re-connect after a partition. */
631 if (!memcmp(&conn->cookie, &hdl->cookie, sizeof conn->cookie)) {
632 if (target->obd_recovering) {
633 int timeout = cfs_duration_sec(cfs_time_sub(
635 &target->obd_recovery_timer),
636 cfs_time_current()));
638 LCONSOLE_WARN("%s: Client %s (at %s) reconnect"
639 "ing, waiting for %d clients in recov"
640 "ery for %d:%.02d\n", target->obd_name,
641 obd_uuid2str(&exp->exp_client_uuid),
642 obd_export_nid2str(exp),
643 target->obd_max_recoverable_clients,
644 timeout / 60, timeout % 60);
646 LCONSOLE_WARN("%s: Client %s (at %s) "
647 "reconnecting\n", target->obd_name,
648 obd_uuid2str(&exp->exp_client_uuid),
649 obd_export_nid2str(exp));
652 conn->cookie = exp->exp_handle.h_cookie;
653 /* target_handle_connect() treats EALREADY and
654 * -EALREADY differently. EALREADY means we are
655 * doing a valid reconnect from the same client. */
658 LCONSOLE_WARN("%s: The server has already connected "
659 "client %s (at %s) with handle " LPX64
660 ", rejecting a client with the same "
661 "uuid trying to reconnect with "
662 "handle " LPX64, target->obd_name,
663 obd_uuid2str(&exp->exp_client_uuid),
664 obd_export_nid2str(exp),
665 hdl->cookie, conn->cookie);
666 memset(conn, 0, sizeof *conn);
667 /* target_handle_connect() treats EALREADY and
668 * -EALREADY differently. -EALREADY is an error
669 * (same UUID, different handle). */
674 conn->cookie = exp->exp_handle.h_cookie;
675 CDEBUG(D_HA, "connect export for UUID '%s' at %p, cookie "LPX64"\n",
676 cluuid->uuid, exp, conn->cookie);
680 void target_client_add_cb(struct obd_device *obd, __u64 transno, void *cb_data,
683 struct obd_export *exp = cb_data;
685 CDEBUG(D_RPCTRACE, "%s: committing for initial connect of %s\n",
686 obd->obd_name, exp->exp_client_uuid.uuid);
688 cfs_spin_lock(&exp->exp_lock);
689 exp->exp_need_sync = 0;
690 cfs_spin_unlock(&exp->exp_lock);
691 class_export_cb_put(exp);
693 EXPORT_SYMBOL(target_client_add_cb);
697 check_and_start_recovery_timer(struct obd_device *obd,
698 struct ptlrpc_request *req, int new_client);
701 check_and_start_recovery_timer(struct obd_device *obd,
702 struct ptlrpc_request *req, int new_client)
707 int target_handle_connect(struct ptlrpc_request *req)
709 struct obd_device *target, *targref = NULL;
710 struct obd_export *export = NULL;
711 struct obd_import *revimp;
712 struct lustre_handle conn;
713 struct lustre_handle *tmp;
714 struct obd_uuid tgtuuid;
715 struct obd_uuid cluuid;
716 struct obd_uuid remote_uuid;
722 struct obd_connect_data *data, *tmpdata;
724 lnet_nid_t *client_nid = NULL;
727 OBD_RACE(OBD_FAIL_TGT_CONN_RACE);
729 str = req_capsule_client_get(&req->rq_pill, &RMF_TGTUUID);
731 DEBUG_REQ(D_ERROR, req, "bad target UUID for connect");
732 GOTO(out, rc = -EINVAL);
735 obd_str2uuid(&tgtuuid, str);
736 target = class_uuid2obd(&tgtuuid);
738 target = class_name2obd(str);
740 if (!target || target->obd_stopping || !target->obd_set_up) {
741 deuuidify(str, NULL, &target_start, &target_len);
742 LCONSOLE_ERROR_MSG(0x137, "%.*s: Not available for connect "
743 "from %s (%s)\n", target_len, target_start,
744 libcfs_nid2str(req->rq_peer.nid), !target ?
745 "no target" : (target->obd_stopping ?
746 "stopping" : "not set up"));
747 GOTO(out, rc = -ENODEV);
750 if (target->obd_no_conn) {
751 LCONSOLE_WARN("%s: Temporarily refusing client connection "
752 "from %s\n", target->obd_name,
753 libcfs_nid2str(req->rq_peer.nid));
754 GOTO(out, rc = -EAGAIN);
757 /* Make sure the target isn't cleaned up while we're here. Yes,
758 there's still a race between the above check and our incref here.
759 Really, class_uuid2obd should take the ref. */
760 targref = class_incref(target, __FUNCTION__, cfs_current());
763 str = req_capsule_client_get(&req->rq_pill, &RMF_CLUUID);
765 DEBUG_REQ(D_ERROR, req, "bad client UUID for connect");
766 GOTO(out, rc = -EINVAL);
769 obd_str2uuid(&cluuid, str);
771 /* XXX extract a nettype and format accordingly */
772 switch (sizeof(lnet_nid_t)) {
773 /* NB the casts only avoid compiler warnings */
775 snprintf(remote_uuid.uuid, sizeof remote_uuid,
776 "NET_"LPX64"_UUID", (__u64)req->rq_peer.nid);
779 snprintf(remote_uuid.uuid, sizeof remote_uuid,
780 "NET_%x_UUID", (__u32)req->rq_peer.nid);
786 tmp = req_capsule_client_get(&req->rq_pill, &RMF_CONN);
788 GOTO(out, rc = -EPROTO);
792 size = req_capsule_get_size(&req->rq_pill, &RMF_CONNECT_DATA,
794 data = req_capsule_client_get(&req->rq_pill, &RMF_CONNECT_DATA);
796 GOTO(out, rc = -EPROTO);
798 rc = req_capsule_server_pack(&req->rq_pill);
802 if (lustre_msg_get_op_flags(req->rq_reqmsg) & MSG_CONNECT_LIBCLIENT) {
804 DEBUG_REQ(D_WARNING, req, "Refusing old (unversioned) "
805 "libclient connection attempt");
806 GOTO(out, rc = -EPROTO);
807 } else if (data->ocd_version < LUSTRE_VERSION_CODE -
808 LUSTRE_VERSION_ALLOWED_OFFSET ||
809 data->ocd_version > LUSTRE_VERSION_CODE +
810 LUSTRE_VERSION_ALLOWED_OFFSET) {
811 DEBUG_REQ(D_WARNING, req, "Refusing %s (%d.%d.%d.%d) "
812 "libclient connection attempt",
813 data->ocd_version < LUSTRE_VERSION_CODE ?
815 OBD_OCD_VERSION_MAJOR(data->ocd_version),
816 OBD_OCD_VERSION_MINOR(data->ocd_version),
817 OBD_OCD_VERSION_PATCH(data->ocd_version),
818 OBD_OCD_VERSION_FIX(data->ocd_version));
819 data = req_capsule_server_sized_get(&req->rq_pill,
821 offsetof(typeof(*data), ocd_version) +
822 sizeof(data->ocd_version));
824 data->ocd_connect_flags = OBD_CONNECT_VERSION;
825 data->ocd_version = LUSTRE_VERSION_CODE;
827 GOTO(out, rc = -EPROTO);
831 if ((lustre_msg_get_op_flags(req->rq_reqmsg) & MSG_CONNECT_INITIAL) &&
832 (data->ocd_connect_flags & OBD_CONNECT_MDS))
835 /* lctl gets a backstage, all-access pass. */
836 if (obd_uuid_equals(&cluuid, &target->obd_uuid))
837 goto dont_check_exports;
839 export = cfs_hash_lookup(target->obd_uuid_hash, &cluuid);
843 /* we've found an export in the hash */
844 if (export->exp_connecting) { /* bug 9635, et. al. */
845 LCONSOLE_WARN("%s: Export %p already connecting from %s\n",
846 export->exp_obd->obd_name, export,
847 libcfs_nid2str(req->rq_peer.nid));
848 class_export_put(export);
851 } else if (mds_conn && export->exp_connection) {
852 if (req->rq_peer.nid != export->exp_connection->c_peer.nid)
853 /* mds reconnected after failover */
854 LCONSOLE_WARN("%s: Received MDS connection from "
855 "%s, removing former export from %s\n",
856 target->obd_name, libcfs_nid2str(req->rq_peer.nid),
857 libcfs_nid2str(export->exp_connection->c_peer.nid));
859 /* new mds connection from the same nid */
860 LCONSOLE_WARN("%s: Received new MDS connection from "
861 "%s, removing former export from same NID\n",
862 target->obd_name, libcfs_nid2str(req->rq_peer.nid));
863 class_fail_export(export);
864 class_export_put(export);
867 } else if (export->exp_connection != NULL &&
868 req->rq_peer.nid != export->exp_connection->c_peer.nid &&
869 (lustre_msg_get_op_flags(req->rq_reqmsg) &
870 MSG_CONNECT_INITIAL)) {
871 /* in mds failover we have static uuid but nid can be
873 LCONSOLE_WARN("%s: Client %s seen on new nid %s when "
874 "existing nid %s is already connected\n",
875 target->obd_name, cluuid.uuid,
876 libcfs_nid2str(req->rq_peer.nid),
878 export->exp_connection->c_peer.nid));
880 class_export_put(export);
883 cfs_spin_lock(&export->exp_lock);
884 export->exp_connecting = 1;
885 cfs_spin_unlock(&export->exp_lock);
886 LASSERT(export->exp_obd == target);
888 rc = target_handle_reconnect(&conn, export, &cluuid);
891 /* If we found an export, we already unlocked. */
894 OBD_FAIL_TIMEOUT(OBD_FAIL_TGT_DELAY_CONNECT, 2 * obd_timeout);
895 } else if (req->rq_export == NULL &&
896 cfs_atomic_read(&export->exp_rpc_count) > 0) {
897 LCONSOLE_WARN("%s: Client %s (at %s) refused connection, "
898 "still busy with %d references\n",
899 target->obd_name, cluuid.uuid,
900 libcfs_nid2str(req->rq_peer.nid),
901 cfs_atomic_read(&export->exp_refcount));
902 GOTO(out, rc = -EBUSY);
903 } else if (req->rq_export != NULL &&
904 (cfs_atomic_read(&export->exp_rpc_count) > 1)) {
905 /* the current connect rpc has increased exp_rpc_count */
906 LCONSOLE_WARN("%s: Client %s (at %s) refused reconnection, "
907 "still busy with %d active RPCs\n",
908 target->obd_name, cluuid.uuid,
909 libcfs_nid2str(req->rq_peer.nid),
910 cfs_atomic_read(&export->exp_rpc_count) - 1);
911 cfs_spin_lock(&export->exp_lock);
912 if (req->rq_export->exp_conn_cnt <
913 lustre_msg_get_conn_cnt(req->rq_reqmsg))
914 /* try to abort active requests */
915 req->rq_export->exp_abort_active_req = 1;
916 cfs_spin_unlock(&export->exp_lock);
917 GOTO(out, rc = -EBUSY);
918 } else if (lustre_msg_get_conn_cnt(req->rq_reqmsg) == 1) {
919 if (!strstr(cluuid.uuid, "mdt"))
920 LCONSOLE_WARN("%s: Rejecting reconnect from the "
921 "known client %s (at %s) because it "
922 "is indicating it is a new client",
923 target->obd_name, cluuid.uuid,
924 libcfs_nid2str(req->rq_peer.nid));
925 GOTO(out, rc = -EALREADY);
927 OBD_FAIL_TIMEOUT(OBD_FAIL_TGT_DELAY_RECONNECT, 2 * obd_timeout);
934 CDEBUG(D_HA, "%s: connection from %s@%s %st"LPU64" exp %p cur %ld last %ld\n",
935 target->obd_name, cluuid.uuid, libcfs_nid2str(req->rq_peer.nid),
936 target->obd_recovering ? "recovering/" : "", data->ocd_transno,
937 export, (long)cfs_time_current_sec(),
938 export ? (long)export->exp_last_request_time : 0);
940 /* If this is the first time a client connects, reset the recovery
942 if (rc == 0 && target->obd_recovering)
943 check_and_start_recovery_timer(target, req, export == NULL);
945 /* We want to handle EALREADY but *not* -EALREADY from
946 * target_handle_reconnect(), return reconnection state in a flag */
947 if (rc == EALREADY) {
948 lustre_msg_add_op_flags(req->rq_repmsg, MSG_CONNECT_RECONNECT);
954 /* Tell the client if we support replayable requests */
955 if (target->obd_replayable)
956 lustre_msg_add_op_flags(req->rq_repmsg, MSG_CONNECT_REPLAYABLE);
957 client_nid = &req->rq_peer.nid;
959 if (export == NULL) {
960 if (target->obd_recovering) {
963 t = cfs_timer_deadline(&target->obd_recovery_timer);
964 t = cfs_time_sub(t, cfs_time_current());
965 t = cfs_duration_sec(t);
966 LCONSOLE_WARN("%s: Denying connection for new client "
967 "%s (at %s), waiting for %d clients in "
968 "recovery for %d:%.02d\n",
970 libcfs_nid2str(req->rq_peer.nid),
972 cfs_atomic_read(&target-> \
973 obd_lock_replay_clients),
974 (int)t / 60, (int)t % 60);
978 rc = obd_connect(req->rq_svc_thread->t_env,
979 &export, target, &cluuid, data,
982 conn.cookie = export->exp_handle.h_cookie;
983 /* LU-1092 reconnect put export refcount in the
984 * end, connect needs take one here too. */
985 class_export_get(export);
989 rc = obd_reconnect(req->rq_svc_thread->t_env,
990 export, target, &cluuid, data, client_nid);
992 /* prevous done via class_conn2export */
993 class_export_get(export);
998 LASSERT(target->u.obt.obt_magic == OBT_MAGIC);
999 data->ocd_instance = target->u.obt.obt_instance;
1001 /* Return only the parts of obd_connect_data that we understand, so the
1002 * client knows that we don't understand the rest. */
1004 tmpsize = req_capsule_get_size(&req->rq_pill, &RMF_CONNECT_DATA,
1006 tmpdata = req_capsule_server_get(&req->rq_pill,
1008 /* Don't use struct assignment here, because the client reply
1009 * buffer may be smaller/larger than the local struct
1010 * obd_connect_data. */
1011 memcpy(tmpdata, data, min(tmpsize, size));
1014 /* If all else goes well, this is our RPC return code. */
1017 lustre_msg_set_handle(req->rq_repmsg, &conn);
1019 /* If the client and the server are the same node, we will already
1020 * have an export that really points to the client's DLM export,
1021 * because we have a shared handles table.
1023 * XXX this will go away when shaver stops sending the "connect" handle
1024 * in the real "remote handle" field of the request --phik 24 Apr 2003
1026 if (req->rq_export != NULL)
1027 class_export_put(req->rq_export);
1029 req->rq_export = export;
1031 cfs_spin_lock(&export->exp_lock);
1032 if (export->exp_conn_cnt >= lustre_msg_get_conn_cnt(req->rq_reqmsg)) {
1033 cfs_spin_unlock(&export->exp_lock);
1034 CDEBUG(D_RPCTRACE, "%s: %s already connected at higher "
1035 "conn_cnt: %d > %d\n",
1036 cluuid.uuid, libcfs_nid2str(req->rq_peer.nid),
1037 export->exp_conn_cnt,
1038 lustre_msg_get_conn_cnt(req->rq_reqmsg));
1040 GOTO(out, rc = -EALREADY);
1042 LASSERT(lustre_msg_get_conn_cnt(req->rq_reqmsg) > 0);
1043 export->exp_conn_cnt = lustre_msg_get_conn_cnt(req->rq_reqmsg);
1044 export->exp_abort_active_req = 0;
1046 /* request from liblustre? Don't evict it for not pinging. */
1047 if (lustre_msg_get_op_flags(req->rq_reqmsg) & MSG_CONNECT_LIBCLIENT) {
1048 export->exp_libclient = 1;
1049 cfs_spin_unlock(&export->exp_lock);
1051 cfs_spin_lock(&target->obd_dev_lock);
1052 cfs_list_del_init(&export->exp_obd_chain_timed);
1053 cfs_spin_unlock(&target->obd_dev_lock);
1055 cfs_spin_unlock(&export->exp_lock);
1058 if (export->exp_connection != NULL) {
1059 /* Check to see if connection came from another NID */
1060 if ((export->exp_connection->c_peer.nid != req->rq_peer.nid) &&
1061 !cfs_hlist_unhashed(&export->exp_nid_hash))
1062 cfs_hash_del(export->exp_obd->obd_nid_hash,
1063 &export->exp_connection->c_peer.nid,
1064 &export->exp_nid_hash);
1066 ptlrpc_connection_put(export->exp_connection);
1069 export->exp_connection = ptlrpc_connection_get(req->rq_peer,
1072 if (cfs_hlist_unhashed(&export->exp_nid_hash)) {
1073 cfs_hash_add(export->exp_obd->obd_nid_hash,
1074 &export->exp_connection->c_peer.nid,
1075 &export->exp_nid_hash);
1078 class_disconnect->class_export_recovery_cleanup() race
1080 if (target->obd_recovering && !export->exp_in_recovery) {
1082 __u64 transno = data->ocd_transno;
1084 cfs_spin_lock(&export->exp_lock);
1085 export->exp_in_recovery = 1;
1086 export->exp_req_replay_needed = 1;
1087 export->exp_lock_replay_needed = 1;
1088 cfs_spin_unlock(&export->exp_lock);
1090 has_transno = !!(lustre_msg_get_op_flags(req->rq_reqmsg) &
1091 MSG_CONNECT_TRANSNO);
1092 if (has_transno && transno == 0)
1093 CWARN("Connect with zero transno!\n");
1095 if (has_transno && transno > 0 &&
1096 transno < target->obd_next_recovery_transno &&
1097 transno > target->obd_last_committed) {
1098 /* another way is to use cmpxchg() so it will be
1100 cfs_spin_lock(&target->obd_recovery_task_lock);
1101 if (transno < target->obd_next_recovery_transno)
1102 target->obd_next_recovery_transno = transno;
1103 cfs_spin_unlock(&target->obd_recovery_task_lock);
1106 cfs_atomic_inc(&target->obd_req_replay_clients);
1107 cfs_atomic_inc(&target->obd_lock_replay_clients);
1108 if (cfs_atomic_inc_return(&target->obd_connected_clients) ==
1109 target->obd_max_recoverable_clients)
1110 cfs_waitq_signal(&target->obd_next_transno_waitq);
1113 /* Tell the client we're in recovery, when client is involved in it. */
1114 if (target->obd_recovering)
1115 lustre_msg_add_op_flags(req->rq_repmsg, MSG_CONNECT_RECOVERING);
1117 tmp = req_capsule_client_get(&req->rq_pill, &RMF_CONN);
1120 if (export->exp_imp_reverse != NULL) {
1121 /* destroyed import can be still referenced in ctxt */
1122 obd_set_info_async(export, sizeof(KEY_REVIMP_UPD),
1123 KEY_REVIMP_UPD, 0, NULL, NULL);
1125 client_destroy_import(export->exp_imp_reverse);
1128 /* for the rest part, we return -ENOTCONN in case of errors
1129 * in order to let client initialize connection again.
1131 revimp = export->exp_imp_reverse = class_new_import(target);
1133 CERROR("fail to alloc new reverse import.\n");
1134 GOTO(out, rc = -ENOTCONN);
1137 revimp->imp_connection = ptlrpc_connection_addref(export->exp_connection);
1138 revimp->imp_client = &export->exp_obd->obd_ldlm_client;
1139 revimp->imp_remote_handle = conn;
1140 revimp->imp_dlm_fake = 1;
1141 revimp->imp_state = LUSTRE_IMP_FULL;
1143 /* unknown versions will be caught in
1144 * ptlrpc_handle_server_req_in->lustre_unpack_msg() */
1145 revimp->imp_msg_magic = req->rq_reqmsg->lm_magic;
1147 if ((export->exp_connect_flags & OBD_CONNECT_AT) &&
1148 (revimp->imp_msg_magic != LUSTRE_MSG_MAGIC_V1))
1149 revimp->imp_msghdr_flags |= MSGHDR_AT_SUPPORT;
1151 revimp->imp_msghdr_flags &= ~MSGHDR_AT_SUPPORT;
1153 if ((export->exp_connect_flags & OBD_CONNECT_FULL20) &&
1154 (revimp->imp_msg_magic != LUSTRE_MSG_MAGIC_V1))
1155 revimp->imp_msghdr_flags |= MSGHDR_CKSUM_INCOMPAT18;
1157 revimp->imp_msghdr_flags &= ~MSGHDR_CKSUM_INCOMPAT18;
1159 rc = sptlrpc_import_sec_adapt(revimp, req->rq_svc_ctx, &req->rq_flvr);
1161 CERROR("Failed to get sec for reverse import: %d\n", rc);
1162 export->exp_imp_reverse = NULL;
1163 class_destroy_import(revimp);
1166 class_import_put(revimp);
1169 cfs_spin_lock(&export->exp_lock);
1170 export->exp_connecting = 0;
1171 cfs_spin_unlock(&export->exp_lock);
1173 class_export_put(export);
1176 class_decref(targref, __FUNCTION__, cfs_current());
1178 req->rq_status = rc;
1182 int target_handle_disconnect(struct ptlrpc_request *req)
1187 rc = req_capsule_server_pack(&req->rq_pill);
1191 /* keep the rq_export around so we can send the reply */
1192 req->rq_status = obd_disconnect(class_export_get(req->rq_export));
1197 void target_destroy_export(struct obd_export *exp)
1199 /* exports created from last_rcvd data, and "fake"
1200 exports created by lctl don't have an import */
1201 if (exp->exp_imp_reverse != NULL)
1202 client_destroy_import(exp->exp_imp_reverse);
1204 LASSERT_ATOMIC_ZERO(&exp->exp_locks_count);
1205 LASSERT_ATOMIC_ZERO(&exp->exp_rpc_count);
1206 LASSERT_ATOMIC_ZERO(&exp->exp_cb_count);
1207 LASSERT_ATOMIC_ZERO(&exp->exp_replay_count);
1211 * Recovery functions
1213 static void target_request_copy_get(struct ptlrpc_request *req)
1215 class_export_rpc_get(req->rq_export);
1216 LASSERT(cfs_list_empty(&req->rq_list));
1217 CFS_INIT_LIST_HEAD(&req->rq_replay_list);
1219 /* increase refcount to keep request in queue */
1220 cfs_atomic_inc(&req->rq_refcount);
1221 /** let export know it has replays to be handled */
1222 cfs_atomic_inc(&req->rq_export->exp_replay_count);
1225 static void target_request_copy_put(struct ptlrpc_request *req)
1227 LASSERT(cfs_list_empty(&req->rq_replay_list));
1228 LASSERT_ATOMIC_POS(&req->rq_export->exp_replay_count);
1230 cfs_atomic_dec(&req->rq_export->exp_replay_count);
1231 class_export_rpc_put(req->rq_export);
1232 ptlrpc_server_drop_request(req);
1235 static int target_exp_enqueue_req_replay(struct ptlrpc_request *req)
1237 __u64 transno = lustre_msg_get_transno(req->rq_reqmsg);
1238 struct obd_export *exp = req->rq_export;
1239 struct ptlrpc_request *reqiter;
1244 cfs_spin_lock(&exp->exp_lock);
1245 cfs_list_for_each_entry(reqiter, &exp->exp_req_replay_queue,
1247 if (lustre_msg_get_transno(reqiter->rq_reqmsg) == transno) {
1254 /* we expect it with RESENT and REPLAY flags */
1255 if ((lustre_msg_get_flags(req->rq_reqmsg) &
1256 (MSG_RESENT | MSG_REPLAY)) != (MSG_RESENT | MSG_REPLAY))
1257 CERROR("invalid flags %x of resent replay\n",
1258 lustre_msg_get_flags(req->rq_reqmsg));
1260 cfs_list_add_tail(&req->rq_replay_list,
1261 &exp->exp_req_replay_queue);
1264 cfs_spin_unlock(&exp->exp_lock);
1268 static void target_exp_dequeue_req_replay(struct ptlrpc_request *req)
1270 LASSERT(!cfs_list_empty(&req->rq_replay_list));
1271 LASSERT(req->rq_export);
1273 cfs_spin_lock(&req->rq_export->exp_lock);
1274 cfs_list_del_init(&req->rq_replay_list);
1275 cfs_spin_unlock(&req->rq_export->exp_lock);
1279 static void target_finish_recovery(struct obd_device *obd)
1281 time_t elapsed_time = max_t(time_t, 1, cfs_time_current_sec() -
1282 obd->obd_recovery_start);
1285 LCONSOLE_INFO("%s: Recovery over after %d:%.02d, of %d clients "
1286 "%d recovered and %d %s evicted.\n", obd->obd_name,
1287 (int)elapsed_time / 60, (int)elapsed_time % 60,
1288 obd->obd_max_recoverable_clients,
1289 cfs_atomic_read(&obd->obd_connected_clients),
1290 obd->obd_stale_clients,
1291 obd->obd_stale_clients == 1 ? "was" : "were");
1293 ldlm_reprocess_all_ns(obd->obd_namespace);
1294 cfs_spin_lock(&obd->obd_recovery_task_lock);
1295 if (!cfs_list_empty(&obd->obd_req_replay_queue) ||
1296 !cfs_list_empty(&obd->obd_lock_replay_queue) ||
1297 !cfs_list_empty(&obd->obd_final_req_queue)) {
1298 CERROR("%s: Recovery queues ( %s%s%s) are not empty\n",
1300 cfs_list_empty(&obd->obd_req_replay_queue) ? "" : "req ",
1301 cfs_list_empty(&obd->obd_lock_replay_queue) ? \
1303 cfs_list_empty(&obd->obd_final_req_queue) ? \
1305 cfs_spin_unlock(&obd->obd_recovery_task_lock);
1308 cfs_spin_unlock(&obd->obd_recovery_task_lock);
1310 obd->obd_recovery_end = cfs_time_current_sec();
1312 /* when recovery finished, cleanup orphans on mds and ost */
1313 if (OBT(obd) && OBP(obd, postrecov)) {
1314 int rc = OBP(obd, postrecov)(obd);
1316 LCONSOLE_WARN("%s: Post recovery failed, rc %d\n",
1322 static void abort_req_replay_queue(struct obd_device *obd)
1324 struct ptlrpc_request *req, *n;
1325 cfs_list_t abort_list;
1327 CFS_INIT_LIST_HEAD(&abort_list);
1328 cfs_spin_lock(&obd->obd_recovery_task_lock);
1329 cfs_list_splice_init(&obd->obd_req_replay_queue, &abort_list);
1330 cfs_spin_unlock(&obd->obd_recovery_task_lock);
1331 cfs_list_for_each_entry_safe(req, n, &abort_list, rq_list) {
1332 DEBUG_REQ(D_WARNING, req, "aborted:");
1333 req->rq_status = -ENOTCONN;
1334 if (ptlrpc_error(req)) {
1335 DEBUG_REQ(D_ERROR, req,
1336 "failed abort_req_reply; skipping");
1338 target_exp_dequeue_req_replay(req);
1339 target_request_copy_put(req);
1343 static void abort_lock_replay_queue(struct obd_device *obd)
1345 struct ptlrpc_request *req, *n;
1346 cfs_list_t abort_list;
1348 CFS_INIT_LIST_HEAD(&abort_list);
1349 cfs_spin_lock(&obd->obd_recovery_task_lock);
1350 cfs_list_splice_init(&obd->obd_lock_replay_queue, &abort_list);
1351 cfs_spin_unlock(&obd->obd_recovery_task_lock);
1352 cfs_list_for_each_entry_safe(req, n, &abort_list, rq_list){
1353 DEBUG_REQ(D_ERROR, req, "aborted:");
1354 req->rq_status = -ENOTCONN;
1355 if (ptlrpc_error(req)) {
1356 DEBUG_REQ(D_ERROR, req,
1357 "failed abort_lock_reply; skipping");
1359 target_request_copy_put(req);
1363 /* Called from a cleanup function if the device is being cleaned up
1364 forcefully. The exports should all have been disconnected already,
1365 the only thing left to do is
1366 - clear the recovery flags
1368 - free queued requests and replies, but don't send replies
1369 Because the obd_stopping flag is set, no new requests should be received.
1372 void target_cleanup_recovery(struct obd_device *obd)
1374 struct ptlrpc_request *req, *n;
1375 cfs_list_t clean_list;
1378 CFS_INIT_LIST_HEAD(&clean_list);
1379 cfs_spin_lock(&obd->obd_dev_lock);
1380 if (!obd->obd_recovering) {
1381 cfs_spin_unlock(&obd->obd_dev_lock);
1385 obd->obd_recovering = obd->obd_abort_recovery = 0;
1386 cfs_spin_unlock(&obd->obd_dev_lock);
1388 cfs_spin_lock(&obd->obd_recovery_task_lock);
1389 target_cancel_recovery_timer(obd);
1390 cfs_list_splice_init(&obd->obd_req_replay_queue, &clean_list);
1391 cfs_spin_unlock(&obd->obd_recovery_task_lock);
1393 cfs_list_for_each_entry_safe(req, n, &clean_list, rq_list) {
1394 LASSERT(req->rq_reply_state == 0);
1395 target_exp_dequeue_req_replay(req);
1396 target_request_copy_put(req);
1399 cfs_spin_lock(&obd->obd_recovery_task_lock);
1400 cfs_list_splice_init(&obd->obd_lock_replay_queue, &clean_list);
1401 cfs_list_splice_init(&obd->obd_final_req_queue, &clean_list);
1402 cfs_spin_unlock(&obd->obd_recovery_task_lock);
1404 cfs_list_for_each_entry_safe(req, n, &clean_list, rq_list){
1405 LASSERT(req->rq_reply_state == 0);
1406 target_request_copy_put(req);
1412 /* obd_recovery_task_lock should be held */
1413 void target_cancel_recovery_timer(struct obd_device *obd)
1415 CDEBUG(D_HA, "%s: cancel recovery timer\n", obd->obd_name);
1416 cfs_timer_disarm(&obd->obd_recovery_timer);
1419 static void target_start_recovery_timer(struct obd_device *obd)
1421 if (obd->obd_recovery_start != 0)
1424 cfs_spin_lock(&obd->obd_dev_lock);
1425 if (!obd->obd_recovering || obd->obd_abort_recovery) {
1426 cfs_spin_unlock(&obd->obd_dev_lock);
1430 LASSERT(obd->obd_recovery_timeout != 0);
1432 if (obd->obd_recovery_start != 0) {
1433 cfs_spin_unlock(&obd->obd_dev_lock);
1437 cfs_timer_arm(&obd->obd_recovery_timer,
1438 cfs_time_shift(obd->obd_recovery_timeout));
1439 obd->obd_recovery_start = cfs_time_current_sec();
1440 cfs_spin_unlock(&obd->obd_dev_lock);
1442 LCONSOLE_WARN("%s: Will be in recovery for at least %d:%.02d, "
1443 "or until %d client%s reconnect%s\n",
1445 obd->obd_recovery_timeout / 60,
1446 obd->obd_recovery_timeout % 60,
1447 obd->obd_max_recoverable_clients,
1448 (obd->obd_max_recoverable_clients == 1) ? "" : "s",
1449 (obd->obd_max_recoverable_clients == 1) ? "s": "");
1453 * extend recovery window.
1455 * if @extend is true, extend recovery window to have @drt remaining at least;
1456 * otherwise, make sure the recovery timeout value is not less than @drt.
1458 static void extend_recovery_timer(struct obd_device *obd, int drt, bool extend)
1462 cfs_duration_t left;
1465 cfs_spin_lock(&obd->obd_dev_lock);
1466 if (!obd->obd_recovering || obd->obd_abort_recovery) {
1467 cfs_spin_unlock(&obd->obd_dev_lock);
1470 LASSERT(obd->obd_recovery_start != 0);
1472 now = cfs_time_current_sec();
1473 to = obd->obd_recovery_timeout;
1474 end = obd->obd_recovery_start + to;
1475 left = cfs_time_sub(end, now);
1477 if (extend && (drt > left)) {
1479 } else if (!extend && (drt > to)) {
1481 /* reduce drt by already passed time */
1482 drt -= obd->obd_recovery_timeout - left;
1485 if (to > obd->obd_recovery_time_hard)
1486 to = obd->obd_recovery_time_hard;
1487 if (obd->obd_recovery_timeout < to) {
1488 obd->obd_recovery_timeout = to;
1489 cfs_timer_arm(&obd->obd_recovery_timer,
1490 cfs_time_shift(drt));
1492 cfs_spin_unlock(&obd->obd_dev_lock);
1494 CDEBUG(D_HA, "%s: recovery timer will expire in %u seconds\n",
1495 obd->obd_name, (unsigned)drt);
1498 /* Reset the timer with each new client connection */
1500 * This timer is actually reconnect_timer, which is for making sure
1501 * the total recovery window is at least as big as my reconnect
1502 * attempt timing. So the initial recovery time_out will be set to
1503 * OBD_RECOVERY_FACTOR * obd_timeout. If the timeout coming
1504 * from client is bigger than this, then the recovery time_out will
1505 * be extended to make sure the client could be reconnected, in the
1506 * process, the timeout from the new client should be ignored.
1510 check_and_start_recovery_timer(struct obd_device *obd,
1511 struct ptlrpc_request *req,
1514 int service_time = lustre_msg_get_service_time(req->rq_reqmsg);
1515 struct obd_device_target *obt = &obd->u.obt;
1516 struct lustre_sb_info *lsi;
1518 if (!new_client && service_time)
1519 /* Teach server about old server's estimates, as first guess
1520 * at how long new requests will take. */
1521 at_measured(&req->rq_rqbd->rqbd_service->srv_at_estimate,
1524 target_start_recovery_timer(obd);
1526 /* convert the service time to rpc timeout,
1527 * reuse service_time to limit stack usage */
1528 service_time = at_est2timeout(service_time);
1530 /* We expect other clients to timeout within service_time, then try
1531 * to reconnect, then try the failover server. The max delay between
1532 * connect attempts is SWITCH_MAX + SWITCH_INC + INITIAL */
1533 service_time += 2 * INITIAL_CONNECT_TIMEOUT;
1535 LASSERT(obt->obt_magic == OBT_MAGIC);
1536 lsi = s2lsi(obt->obt_sb);
1537 if (!(lsi->lsi_flags | LSI_IR_CAPABLE))
1538 service_time += 2 * (CONNECTION_SWITCH_MAX +
1539 CONNECTION_SWITCH_INC);
1540 if (service_time > obd->obd_recovery_timeout && !new_client)
1541 extend_recovery_timer(obd, service_time, false);
1544 /** Health checking routines */
1545 static inline int exp_connect_healthy(struct obd_export *exp)
1547 return (exp->exp_in_recovery);
1550 /** if export done req_replay or has replay in queue */
1551 static inline int exp_req_replay_healthy(struct obd_export *exp)
1553 return (!exp->exp_req_replay_needed ||
1554 cfs_atomic_read(&exp->exp_replay_count) > 0);
1556 /** if export done lock_replay or has replay in queue */
1557 static inline int exp_lock_replay_healthy(struct obd_export *exp)
1559 return (!exp->exp_lock_replay_needed ||
1560 cfs_atomic_read(&exp->exp_replay_count) > 0);
1563 static inline int exp_vbr_healthy(struct obd_export *exp)
1565 return (!exp->exp_vbr_failed);
1568 static inline int exp_finished(struct obd_export *exp)
1570 return (exp->exp_in_recovery && !exp->exp_lock_replay_needed);
1573 /** Checking routines for recovery */
1574 static int check_for_clients(struct obd_device *obd)
1576 unsigned int clnts = cfs_atomic_read(&obd->obd_connected_clients);
1578 if (obd->obd_abort_recovery || obd->obd_recovery_expired)
1580 LASSERT(clnts <= obd->obd_max_recoverable_clients);
1581 if (obd->obd_no_conn == 0 &&
1582 clnts + obd->obd_stale_clients == obd->obd_max_recoverable_clients)
1587 static int check_for_next_transno(struct obd_device *obd)
1589 struct ptlrpc_request *req = NULL;
1590 int wake_up = 0, connected, completed, queue_len;
1591 __u64 next_transno, req_transno;
1594 cfs_spin_lock(&obd->obd_recovery_task_lock);
1595 if (!cfs_list_empty(&obd->obd_req_replay_queue)) {
1596 req = cfs_list_entry(obd->obd_req_replay_queue.next,
1597 struct ptlrpc_request, rq_list);
1598 req_transno = lustre_msg_get_transno(req->rq_reqmsg);
1603 connected = cfs_atomic_read(&obd->obd_connected_clients);
1604 completed = connected - cfs_atomic_read(&obd->obd_req_replay_clients);
1605 queue_len = obd->obd_requests_queued_for_recovery;
1606 next_transno = obd->obd_next_recovery_transno;
1608 CDEBUG(D_HA, "max: %d, connected: %d, completed: %d, queue_len: %d, "
1609 "req_transno: "LPU64", next_transno: "LPU64"\n",
1610 obd->obd_max_recoverable_clients, connected, completed,
1611 queue_len, req_transno, next_transno);
1613 if (obd->obd_abort_recovery) {
1614 CDEBUG(D_HA, "waking for aborted recovery\n");
1616 } else if (obd->obd_recovery_expired) {
1617 CDEBUG(D_HA, "waking for expired recovery\n");
1619 } else if (cfs_atomic_read(&obd->obd_req_replay_clients) == 0) {
1620 CDEBUG(D_HA, "waking for completed recovery\n");
1622 } else if (req_transno == next_transno) {
1623 CDEBUG(D_HA, "waking for next ("LPD64")\n", next_transno);
1625 } else if (queue_len == cfs_atomic_read(&obd->obd_req_replay_clients)) {
1627 /** handle gaps occured due to lost reply or VBR */
1628 LASSERTF(req_transno >= next_transno,
1629 "req_transno: "LPU64", next_transno: "LPU64"\n",
1630 req_transno, next_transno);
1631 if (req_transno > obd->obd_last_committed &&
1632 !obd->obd_version_recov)
1635 "%s: waking for gap in transno, VBR is %s (skip: "
1636 LPD64", ql: %d, comp: %d, conn: %d, next: "LPD64
1637 ", last_committed: "LPD64")\n",
1638 obd->obd_name, obd->obd_version_recov ? "ON" : "OFF",
1639 next_transno, queue_len, completed, connected,
1640 req_transno, obd->obd_last_committed);
1641 obd->obd_next_recovery_transno = req_transno;
1643 } else if (OBD_FAIL_CHECK(OBD_FAIL_MDS_RECOVERY_ACCEPTS_GAPS)) {
1644 CDEBUG(D_HA, "accepting transno gaps is explicitly allowed"
1645 " by fail_lock, waking up ("LPD64")\n", next_transno);
1646 obd->obd_next_recovery_transno = req_transno;
1649 cfs_spin_unlock(&obd->obd_recovery_task_lock);
1653 static int check_for_next_lock(struct obd_device *obd)
1657 cfs_spin_lock(&obd->obd_recovery_task_lock);
1658 if (!cfs_list_empty(&obd->obd_lock_replay_queue)) {
1659 CDEBUG(D_HA, "waking for next lock\n");
1661 } else if (cfs_atomic_read(&obd->obd_lock_replay_clients) == 0) {
1662 CDEBUG(D_HA, "waking for completed lock replay\n");
1664 } else if (obd->obd_abort_recovery) {
1665 CDEBUG(D_HA, "waking for aborted recovery\n");
1667 } else if (obd->obd_recovery_expired) {
1668 CDEBUG(D_HA, "waking for expired recovery\n");
1671 cfs_spin_unlock(&obd->obd_recovery_task_lock);
1677 * wait for recovery events,
1678 * check its status with help of check_routine
1679 * evict dead clients via health_check
1681 static int target_recovery_overseer(struct obd_device *obd,
1682 int (*check_routine)(struct obd_device *),
1683 int (*health_check)(struct obd_export *))
1686 cfs_wait_event(obd->obd_next_transno_waitq, check_routine(obd));
1687 if (obd->obd_abort_recovery) {
1688 CWARN("recovery is aborted, evict exports in recovery\n");
1689 /** evict exports which didn't finish recovery yet */
1690 class_disconnect_stale_exports(obd, exp_finished);
1692 } else if (obd->obd_recovery_expired) {
1693 obd->obd_recovery_expired = 0;
1694 /** If some clients died being recovered, evict them */
1695 LCONSOLE_WARN("%s: recovery is timed out, "
1696 "evict stale exports\n", obd->obd_name);
1697 /** evict cexports with no replay in queue, they are stalled */
1698 class_disconnect_stale_exports(obd, health_check);
1699 /** continue with VBR */
1700 cfs_spin_lock(&obd->obd_dev_lock);
1701 obd->obd_version_recov = 1;
1702 cfs_spin_unlock(&obd->obd_dev_lock);
1704 * reset timer, recovery will proceed with versions now,
1705 * timeout is set just to handle reconnection delays
1707 extend_recovery_timer(obd, RECONNECT_DELAY_MAX, true);
1708 /** Wait for recovery events again, after evicting bad clients */
1714 static struct ptlrpc_request *target_next_replay_req(struct obd_device *obd)
1716 struct ptlrpc_request *req = NULL;
1719 CDEBUG(D_HA, "Waiting for transno "LPD64"\n",
1720 obd->obd_next_recovery_transno);
1722 if (target_recovery_overseer(obd, check_for_next_transno,
1723 exp_req_replay_healthy)) {
1724 abort_req_replay_queue(obd);
1725 abort_lock_replay_queue(obd);
1728 cfs_spin_lock(&obd->obd_recovery_task_lock);
1729 if (!cfs_list_empty(&obd->obd_req_replay_queue)) {
1730 req = cfs_list_entry(obd->obd_req_replay_queue.next,
1731 struct ptlrpc_request, rq_list);
1732 cfs_list_del_init(&req->rq_list);
1733 obd->obd_requests_queued_for_recovery--;
1734 cfs_spin_unlock(&obd->obd_recovery_task_lock);
1736 cfs_spin_unlock(&obd->obd_recovery_task_lock);
1737 LASSERT(cfs_list_empty(&obd->obd_req_replay_queue));
1738 LASSERT(cfs_atomic_read(&obd->obd_req_replay_clients) == 0);
1739 /** evict exports failed VBR */
1740 class_disconnect_stale_exports(obd, exp_vbr_healthy);
1745 static struct ptlrpc_request *target_next_replay_lock(struct obd_device *obd)
1747 struct ptlrpc_request *req = NULL;
1749 CDEBUG(D_HA, "Waiting for lock\n");
1750 if (target_recovery_overseer(obd, check_for_next_lock,
1751 exp_lock_replay_healthy))
1752 abort_lock_replay_queue(obd);
1754 cfs_spin_lock(&obd->obd_recovery_task_lock);
1755 if (!cfs_list_empty(&obd->obd_lock_replay_queue)) {
1756 req = cfs_list_entry(obd->obd_lock_replay_queue.next,
1757 struct ptlrpc_request, rq_list);
1758 cfs_list_del_init(&req->rq_list);
1759 cfs_spin_unlock(&obd->obd_recovery_task_lock);
1761 cfs_spin_unlock(&obd->obd_recovery_task_lock);
1762 LASSERT(cfs_list_empty(&obd->obd_lock_replay_queue));
1763 LASSERT(cfs_atomic_read(&obd->obd_lock_replay_clients) == 0);
1764 /** evict exports failed VBR */
1765 class_disconnect_stale_exports(obd, exp_vbr_healthy);
1770 static struct ptlrpc_request *target_next_final_ping(struct obd_device *obd)
1772 struct ptlrpc_request *req = NULL;
1774 cfs_spin_lock(&obd->obd_recovery_task_lock);
1775 if (!cfs_list_empty(&obd->obd_final_req_queue)) {
1776 req = cfs_list_entry(obd->obd_final_req_queue.next,
1777 struct ptlrpc_request, rq_list);
1778 cfs_list_del_init(&req->rq_list);
1779 cfs_spin_unlock(&obd->obd_recovery_task_lock);
1780 if (req->rq_export->exp_in_recovery) {
1781 cfs_spin_lock(&req->rq_export->exp_lock);
1782 req->rq_export->exp_in_recovery = 0;
1783 cfs_spin_unlock(&req->rq_export->exp_lock);
1786 cfs_spin_unlock(&obd->obd_recovery_task_lock);
1791 static int handle_recovery_req(struct ptlrpc_thread *thread,
1792 struct ptlrpc_request *req,
1793 svc_handler_t handler)
1799 * export can be evicted during recovery, no need to handle replays for
1800 * it after that, discard such request silently
1802 if (req->rq_export->exp_disconnected)
1803 GOTO(reqcopy_put, rc = 0);
1805 rc = lu_context_init(&req->rq_recov_session, LCT_SESSION);
1807 CERROR("Failure to initialize session: %d\n", rc);
1808 GOTO(reqcopy_put, rc);
1811 req->rq_recov_session.lc_thread = thread;
1812 lu_context_enter(&req->rq_recov_session);
1813 req->rq_svc_thread = thread;
1814 req->rq_svc_thread->t_env->le_ses = &req->rq_recov_session;
1816 /* thread context */
1817 lu_context_enter(&thread->t_env->le_ctx);
1819 lu_context_exit(&thread->t_env->le_ctx);
1821 lu_context_exit(&req->rq_recov_session);
1822 lu_context_fini(&req->rq_recov_session);
1823 /* don't reset timer for final stage */
1824 if (!exp_finished(req->rq_export)) {
1825 int to = obd_timeout;
1828 * Add request timeout to the recovery time so next request from
1829 * this client may come in recovery time
1832 struct ptlrpc_service *svc = req->rq_rqbd->rqbd_service;
1833 /* If the server sent early reply for this request,
1834 * the client will recalculate the timeout according to
1835 * current server estimate service time, so we will
1836 * use the maxium timeout here for waiting the client
1837 * sending the next req */
1838 to = max((int)at_est2timeout(
1839 at_get(&svc->srv_at_estimate)),
1840 (int)lustre_msg_get_timeout(req->rq_reqmsg));
1841 /* Add net_latency (see ptlrpc_replay_req) */
1842 to += lustre_msg_get_service_time(req->rq_reqmsg);
1844 extend_recovery_timer(class_exp2obd(req->rq_export), to, true);
1850 static int target_recovery_thread(void *arg)
1852 struct lu_target *lut = arg;
1853 struct obd_device *obd = lut->lut_obd;
1854 struct ptlrpc_request *req;
1855 struct target_recovery_data *trd = &obd->obd_recovery_data;
1856 unsigned long delta;
1857 unsigned long flags;
1859 struct ptlrpc_thread *thread = NULL;
1863 cfs_daemonize_ctxt("tgt_recov");
1865 SIGNAL_MASK_LOCK(current, flags);
1866 sigfillset(¤t->blocked);
1868 SIGNAL_MASK_UNLOCK(current, flags);
1870 OBD_ALLOC_PTR(thread);
1876 OBD_FREE_PTR(thread);
1880 rc = lu_context_init(&env->le_ctx, LCT_MD_THREAD | LCT_DT_THREAD);
1882 OBD_FREE_PTR(thread);
1887 thread->t_env = env;
1888 thread->t_id = -1; /* force filter_iobuf_get/put to use local buffers */
1889 env->le_ctx.lc_thread = thread;
1890 thread->t_data = NULL;
1891 thread->t_watchdog = NULL;
1893 CDEBUG(D_HA, "%s: started recovery thread pid %d\n", obd->obd_name,
1895 trd->trd_processing_task = cfs_curproc_pid();
1897 cfs_spin_lock(&obd->obd_dev_lock);
1898 obd->obd_recovering = 1;
1899 cfs_spin_unlock(&obd->obd_dev_lock);
1900 cfs_complete(&trd->trd_starting);
1902 /* first of all, we have to know the first transno to replay */
1903 if (target_recovery_overseer(obd, check_for_clients,
1904 exp_connect_healthy)) {
1905 abort_req_replay_queue(obd);
1906 abort_lock_replay_queue(obd);
1909 /* next stage: replay requests */
1911 CDEBUG(D_INFO, "1: request replay stage - %d clients from t"LPU64"\n",
1912 cfs_atomic_read(&obd->obd_req_replay_clients),
1913 obd->obd_next_recovery_transno);
1914 while ((req = target_next_replay_req(obd))) {
1915 LASSERT(trd->trd_processing_task == cfs_curproc_pid());
1916 DEBUG_REQ(D_HA, req, "processing t"LPD64" from %s",
1917 lustre_msg_get_transno(req->rq_reqmsg),
1918 libcfs_nid2str(req->rq_peer.nid));
1919 handle_recovery_req(thread, req,
1920 trd->trd_recovery_handler);
1922 * bz18031: increase next_recovery_transno before
1923 * target_request_copy_put() will drop exp_rpc reference
1925 cfs_spin_lock(&obd->obd_recovery_task_lock);
1926 obd->obd_next_recovery_transno++;
1927 cfs_spin_unlock(&obd->obd_recovery_task_lock);
1928 target_exp_dequeue_req_replay(req);
1929 target_request_copy_put(req);
1930 obd->obd_replayed_requests++;
1934 * The second stage: replay locks
1936 CDEBUG(D_INFO, "2: lock replay stage - %d clients\n",
1937 cfs_atomic_read(&obd->obd_lock_replay_clients));
1938 while ((req = target_next_replay_lock(obd))) {
1939 LASSERT(trd->trd_processing_task == cfs_curproc_pid());
1940 DEBUG_REQ(D_HA, req, "processing lock from %s: ",
1941 libcfs_nid2str(req->rq_peer.nid));
1942 handle_recovery_req(thread, req,
1943 trd->trd_recovery_handler);
1944 target_request_copy_put(req);
1945 obd->obd_replayed_locks++;
1949 * The third stage: reply on final pings, at this moment all clients
1950 * must have request in final queue
1952 CDEBUG(D_INFO, "3: final stage - process recovery completion pings\n");
1953 /** Update server last boot epoch */
1954 lut_boot_epoch_update(lut);
1955 /* We drop recoverying flag to forward all new requests
1956 * to regular mds_handle() since now */
1957 cfs_spin_lock(&obd->obd_dev_lock);
1958 obd->obd_recovering = obd->obd_abort_recovery = 0;
1959 cfs_spin_unlock(&obd->obd_dev_lock);
1960 cfs_spin_lock(&obd->obd_recovery_task_lock);
1961 target_cancel_recovery_timer(obd);
1962 cfs_spin_unlock(&obd->obd_recovery_task_lock);
1963 while ((req = target_next_final_ping(obd))) {
1964 LASSERT(trd->trd_processing_task == cfs_curproc_pid());
1965 DEBUG_REQ(D_HA, req, "processing final ping from %s: ",
1966 libcfs_nid2str(req->rq_peer.nid));
1967 handle_recovery_req(thread, req,
1968 trd->trd_recovery_handler);
1969 target_request_copy_put(req);
1972 delta = (jiffies - delta) / CFS_HZ;
1973 CDEBUG(D_INFO,"4: recovery completed in %lus - %d/%d reqs/locks\n",
1974 delta, obd->obd_replayed_requests, obd->obd_replayed_locks);
1975 if (delta > OBD_RECOVERY_TIME_SOFT) {
1976 CWARN("too long recovery - read logs\n");
1977 libcfs_debug_dumplog();
1980 target_finish_recovery(obd);
1982 lu_context_fini(&env->le_ctx);
1983 trd->trd_processing_task = 0;
1984 cfs_complete(&trd->trd_finishing);
1986 OBD_FREE_PTR(thread);
1991 static int target_start_recovery_thread(struct lu_target *lut,
1992 svc_handler_t handler)
1994 struct obd_device *obd = lut->lut_obd;
1996 struct target_recovery_data *trd = &obd->obd_recovery_data;
1998 memset(trd, 0, sizeof(*trd));
1999 cfs_init_completion(&trd->trd_starting);
2000 cfs_init_completion(&trd->trd_finishing);
2001 trd->trd_recovery_handler = handler;
2003 if (cfs_create_thread(target_recovery_thread, lut, 0) > 0) {
2004 cfs_wait_for_completion(&trd->trd_starting);
2005 LASSERT(obd->obd_recovering != 0);
2012 void target_stop_recovery_thread(struct obd_device *obd)
2014 if (obd->obd_recovery_data.trd_processing_task > 0) {
2015 struct target_recovery_data *trd = &obd->obd_recovery_data;
2016 /** recovery can be done but postrecovery is not yet */
2017 cfs_spin_lock(&obd->obd_dev_lock);
2018 if (obd->obd_recovering) {
2019 CERROR("%s: Aborting recovery\n", obd->obd_name);
2020 obd->obd_abort_recovery = 1;
2021 cfs_waitq_signal(&obd->obd_next_transno_waitq);
2023 cfs_spin_unlock(&obd->obd_dev_lock);
2024 cfs_wait_for_completion(&trd->trd_finishing);
2028 void target_recovery_fini(struct obd_device *obd)
2030 class_disconnect_exports(obd);
2031 target_stop_recovery_thread(obd);
2032 target_cleanup_recovery(obd);
2034 EXPORT_SYMBOL(target_recovery_fini);
2036 static void target_recovery_expired(unsigned long castmeharder)
2038 struct obd_device *obd = (struct obd_device *)castmeharder;
2039 CDEBUG(D_HA, "%s: recovery timed out; %d clients are still in recovery"
2040 " after %lds (%d clients connected)\n",
2041 obd->obd_name, cfs_atomic_read(&obd->obd_lock_replay_clients),
2042 cfs_time_current_sec()- obd->obd_recovery_start,
2043 cfs_atomic_read(&obd->obd_connected_clients));
2045 obd->obd_recovery_expired = 1;
2046 cfs_waitq_signal(&obd->obd_next_transno_waitq);
2049 void target_recovery_init(struct lu_target *lut, svc_handler_t handler)
2051 struct obd_device *obd = lut->lut_obd;
2052 if (obd->obd_max_recoverable_clients == 0) {
2053 /** Update server last boot epoch */
2054 lut_boot_epoch_update(lut);
2058 CWARN("RECOVERY: service %s, %d recoverable clients, "
2059 "last_transno "LPU64"\n", obd->obd_name,
2060 obd->obd_max_recoverable_clients, obd->obd_last_committed);
2061 LASSERT(obd->obd_stopping == 0);
2062 obd->obd_next_recovery_transno = obd->obd_last_committed + 1;
2063 obd->obd_recovery_start = 0;
2064 obd->obd_recovery_end = 0;
2066 cfs_timer_init(&obd->obd_recovery_timer, target_recovery_expired, obd);
2067 target_start_recovery_thread(lut, handler);
2069 EXPORT_SYMBOL(target_recovery_init);
2071 #endif /* __KERNEL__ */
2073 static int target_process_req_flags(struct obd_device *obd,
2074 struct ptlrpc_request *req)
2076 struct obd_export *exp = req->rq_export;
2077 LASSERT(exp != NULL);
2078 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REQ_REPLAY_DONE) {
2079 /* client declares he's ready to replay locks */
2080 cfs_spin_lock(&exp->exp_lock);
2081 if (exp->exp_req_replay_needed) {
2082 exp->exp_req_replay_needed = 0;
2083 cfs_spin_unlock(&exp->exp_lock);
2085 LASSERT_ATOMIC_POS(&obd->obd_req_replay_clients);
2086 cfs_atomic_dec(&obd->obd_req_replay_clients);
2088 cfs_spin_unlock(&exp->exp_lock);
2091 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_LOCK_REPLAY_DONE) {
2092 /* client declares he's ready to complete recovery
2093 * so, we put the request on th final queue */
2094 cfs_spin_lock(&exp->exp_lock);
2095 if (exp->exp_lock_replay_needed) {
2096 exp->exp_lock_replay_needed = 0;
2097 cfs_spin_unlock(&exp->exp_lock);
2099 LASSERT_ATOMIC_POS(&obd->obd_lock_replay_clients);
2100 cfs_atomic_dec(&obd->obd_lock_replay_clients);
2102 cfs_spin_unlock(&exp->exp_lock);
2108 int target_queue_recovery_request(struct ptlrpc_request *req,
2109 struct obd_device *obd)
2113 __u64 transno = lustre_msg_get_transno(req->rq_reqmsg);
2116 if (obd->obd_recovery_data.trd_processing_task == cfs_curproc_pid()) {
2117 /* Processing the queue right now, don't re-add. */
2121 target_process_req_flags(obd, req);
2123 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_LOCK_REPLAY_DONE) {
2124 /* client declares he's ready to complete recovery
2125 * so, we put the request on th final queue */
2126 target_request_copy_get(req);
2127 DEBUG_REQ(D_HA, req, "queue final req");
2128 cfs_waitq_signal(&obd->obd_next_transno_waitq);
2129 cfs_spin_lock(&obd->obd_recovery_task_lock);
2130 if (obd->obd_recovering) {
2131 cfs_list_add_tail(&req->rq_list,
2132 &obd->obd_final_req_queue);
2134 cfs_spin_unlock(&obd->obd_recovery_task_lock);
2135 target_request_copy_put(req);
2136 RETURN(obd->obd_stopping ? -ENOTCONN : 1);
2138 cfs_spin_unlock(&obd->obd_recovery_task_lock);
2141 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REQ_REPLAY_DONE) {
2142 /* client declares he's ready to replay locks */
2143 target_request_copy_get(req);
2144 DEBUG_REQ(D_HA, req, "queue lock replay req");
2145 cfs_waitq_signal(&obd->obd_next_transno_waitq);
2146 cfs_spin_lock(&obd->obd_recovery_task_lock);
2147 LASSERT(obd->obd_recovering);
2148 /* usually due to recovery abort */
2149 if (!req->rq_export->exp_in_recovery) {
2150 cfs_spin_unlock(&obd->obd_recovery_task_lock);
2151 target_request_copy_put(req);
2154 LASSERT(req->rq_export->exp_lock_replay_needed);
2155 cfs_list_add_tail(&req->rq_list, &obd->obd_lock_replay_queue);
2156 cfs_spin_unlock(&obd->obd_recovery_task_lock);
2160 /* CAVEAT EMPTOR: The incoming request message has been swabbed
2161 * (i.e. buflens etc are in my own byte order), but type-dependent
2162 * buffers (eg mdt_body, ost_body etc) have NOT been swabbed. */
2165 CFS_INIT_LIST_HEAD(&req->rq_list);
2166 DEBUG_REQ(D_HA, req, "not queueing");
2170 /* If we're processing the queue, we want don't want to queue this
2173 * Also, if this request has a transno less than the one we're waiting
2174 * for, we should process it now. It could (and currently always will)
2175 * be an open request for a descriptor that was opened some time ago.
2177 * Also, a resent, replayed request that has already been
2178 * handled will pass through here and be processed immediately.
2180 CDEBUG(D_HA, "Next recovery transno: "LPU64
2181 ", current: "LPU64", replaying\n",
2182 obd->obd_next_recovery_transno, transno);
2183 cfs_spin_lock(&obd->obd_recovery_task_lock);
2184 if (transno < obd->obd_next_recovery_transno) {
2185 /* Processing the queue right now, don't re-add. */
2186 LASSERT(cfs_list_empty(&req->rq_list));
2187 cfs_spin_unlock(&obd->obd_recovery_task_lock);
2190 cfs_spin_unlock(&obd->obd_recovery_task_lock);
2192 if (OBD_FAIL_CHECK(OBD_FAIL_TGT_REPLAY_DROP))
2195 target_request_copy_get(req);
2196 if (!req->rq_export->exp_in_recovery) {
2197 target_request_copy_put(req);
2200 LASSERT(req->rq_export->exp_req_replay_needed);
2202 if (target_exp_enqueue_req_replay(req)) {
2203 DEBUG_REQ(D_ERROR, req, "dropping resent queued req");
2204 target_request_copy_put(req);
2209 cfs_spin_lock(&obd->obd_recovery_task_lock);
2210 LASSERT(obd->obd_recovering);
2211 cfs_list_for_each(tmp, &obd->obd_req_replay_queue) {
2212 struct ptlrpc_request *reqiter =
2213 cfs_list_entry(tmp, struct ptlrpc_request, rq_list);
2215 if (lustre_msg_get_transno(reqiter->rq_reqmsg) > transno) {
2216 cfs_list_add_tail(&req->rq_list, &reqiter->rq_list);
2221 if (unlikely(lustre_msg_get_transno(reqiter->rq_reqmsg) ==
2223 DEBUG_REQ(D_ERROR, req, "dropping replay: transno "
2224 "has been claimed by another client");
2225 cfs_spin_unlock(&obd->obd_recovery_task_lock);
2226 target_exp_dequeue_req_replay(req);
2227 target_request_copy_put(req);
2233 cfs_list_add_tail(&req->rq_list, &obd->obd_req_replay_queue);
2235 obd->obd_requests_queued_for_recovery++;
2236 cfs_spin_unlock(&obd->obd_recovery_task_lock);
2237 cfs_waitq_signal(&obd->obd_next_transno_waitq);
2241 int target_handle_ping(struct ptlrpc_request *req)
2243 obd_ping(req->rq_export);
2244 return req_capsule_server_pack(&req->rq_pill);
2247 void target_committed_to_req(struct ptlrpc_request *req)
2249 struct obd_export *exp = req->rq_export;
2251 if (!exp->exp_obd->obd_no_transno && req->rq_repmsg != NULL)
2252 lustre_msg_set_last_committed(req->rq_repmsg,
2253 exp->exp_last_committed);
2255 DEBUG_REQ(D_IOCTL, req, "not sending last_committed update (%d/"
2256 "%d)", exp->exp_obd->obd_no_transno,
2257 req->rq_repmsg == NULL);
2259 CDEBUG(D_INFO, "last_committed "LPU64", transno "LPU64", xid "LPU64"\n",
2260 exp->exp_last_committed, req->rq_transno, req->rq_xid);
2262 EXPORT_SYMBOL(target_committed_to_req);
2264 #endif /* HAVE_SERVER_SUPPORT */
2267 * Packs current SLV and Limit into \a req.
2269 int target_pack_pool_reply(struct ptlrpc_request *req)
2271 struct obd_device *obd;
2275 * Check that we still have all structures alive as this may
2276 * be some late rpc in shutdown time.
2278 if (unlikely(!req->rq_export || !req->rq_export->exp_obd ||
2279 !exp_connect_lru_resize(req->rq_export))) {
2280 lustre_msg_set_slv(req->rq_repmsg, 0);
2281 lustre_msg_set_limit(req->rq_repmsg, 0);
2286 * OBD is alive here as export is alive, which we checked above.
2288 obd = req->rq_export->exp_obd;
2290 cfs_read_lock(&obd->obd_pool_lock);
2291 lustre_msg_set_slv(req->rq_repmsg, obd->obd_pool_slv);
2292 lustre_msg_set_limit(req->rq_repmsg, obd->obd_pool_limit);
2293 cfs_read_unlock(&obd->obd_pool_lock);
2298 int target_send_reply_msg(struct ptlrpc_request *req, int rc, int fail_id)
2300 if (OBD_FAIL_CHECK_ORSET(fail_id & ~OBD_FAIL_ONCE, OBD_FAIL_ONCE)) {
2301 DEBUG_REQ(D_ERROR, req, "dropping reply");
2306 DEBUG_REQ(D_NET, req, "processing error (%d)", rc);
2307 req->rq_status = rc;
2308 return (ptlrpc_send_error(req, 1));
2310 DEBUG_REQ(D_NET, req, "sending reply");
2313 return (ptlrpc_send_reply(req, PTLRPC_REPLY_MAYBE_DIFFICULT));
2316 void target_send_reply(struct ptlrpc_request *req, int rc, int fail_id)
2319 struct ptlrpc_reply_state *rs;
2320 struct obd_export *exp;
2321 struct ptlrpc_service *svc;
2324 if (req->rq_no_reply) {
2329 svc = req->rq_rqbd->rqbd_service;
2330 rs = req->rq_reply_state;
2331 if (rs == NULL || !rs->rs_difficult) {
2333 target_send_reply_msg (req, rc, fail_id);
2338 /* must be an export if locks saved */
2339 LASSERT (req->rq_export != NULL);
2340 /* req/reply consistent */
2341 LASSERT (rs->rs_service == svc);
2344 LASSERT (!rs->rs_scheduled);
2345 LASSERT (!rs->rs_scheduled_ever);
2346 LASSERT (!rs->rs_handled);
2347 LASSERT (!rs->rs_on_net);
2348 LASSERT (rs->rs_export == NULL);
2349 LASSERT (cfs_list_empty(&rs->rs_obd_list));
2350 LASSERT (cfs_list_empty(&rs->rs_exp_list));
2352 exp = class_export_get (req->rq_export);
2354 /* disable reply scheduling while I'm setting up */
2355 rs->rs_scheduled = 1;
2357 rs->rs_xid = req->rq_xid;
2358 rs->rs_transno = req->rq_transno;
2359 rs->rs_export = exp;
2360 rs->rs_opc = lustre_msg_get_opc(rs->rs_msg);
2362 cfs_spin_lock(&exp->exp_uncommitted_replies_lock);
2363 CDEBUG(D_NET, "rs transno = "LPU64", last committed = "LPU64"\n",
2364 rs->rs_transno, exp->exp_last_committed);
2365 if (rs->rs_transno > exp->exp_last_committed) {
2366 /* not committed already */
2367 cfs_list_add_tail(&rs->rs_obd_list,
2368 &exp->exp_uncommitted_replies);
2370 cfs_spin_unlock (&exp->exp_uncommitted_replies_lock);
2372 cfs_spin_lock(&exp->exp_lock);
2373 cfs_list_add_tail(&rs->rs_exp_list, &exp->exp_outstanding_replies);
2374 cfs_spin_unlock(&exp->exp_lock);
2376 netrc = target_send_reply_msg (req, rc, fail_id);
2378 cfs_spin_lock(&svc->srv_rs_lock);
2380 cfs_atomic_inc(&svc->srv_n_difficult_replies);
2383 /* error sending: reply is off the net. Also we need +1
2384 * reply ref until ptlrpc_handle_rs() is done
2385 * with the reply state (if the send was successful, there
2386 * would have been +1 ref for the net, which
2387 * reply_out_callback leaves alone) */
2389 ptlrpc_rs_addref(rs);
2392 cfs_spin_lock(&rs->rs_lock);
2393 if (rs->rs_transno <= exp->exp_last_committed ||
2394 (!rs->rs_on_net && !rs->rs_no_ack) ||
2395 cfs_list_empty(&rs->rs_exp_list) || /* completed already */
2396 cfs_list_empty(&rs->rs_obd_list)) {
2397 CDEBUG(D_HA, "Schedule reply immediately\n");
2398 ptlrpc_dispatch_difficult_reply(rs);
2400 cfs_list_add (&rs->rs_list, &svc->srv_active_replies);
2401 rs->rs_scheduled = 0; /* allow notifier to schedule */
2403 cfs_spin_unlock(&rs->rs_lock);
2404 cfs_spin_unlock(&svc->srv_rs_lock);
2408 int target_handle_qc_callback(struct ptlrpc_request *req)
2410 struct obd_quotactl *oqctl;
2411 struct client_obd *cli = &req->rq_export->exp_obd->u.cli;
2413 oqctl = req_capsule_client_get(&req->rq_pill, &RMF_OBD_QUOTACTL);
2414 if (oqctl == NULL) {
2415 CERROR("Can't unpack obd_quotactl\n");
2419 cli->cl_qchk_stat = oqctl->qc_stat;
2424 #ifdef HAVE_QUOTA_SUPPORT
2425 int target_handle_dqacq_callback(struct ptlrpc_request *req)
2428 struct obd_device *obd = req->rq_export->exp_obd;
2429 struct obd_device *master_obd = NULL, *lov_obd = NULL;
2430 struct obd_device_target *obt;
2431 struct lustre_quota_ctxt *qctxt;
2432 struct qunit_data *qdata = NULL;
2436 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_DROP_QUOTA_REQ))
2439 rc = req_capsule_server_pack(&req->rq_pill);
2441 CERROR("packing reply failed!: rc = %d\n", rc);
2445 LASSERT(req->rq_export);
2447 qdata = quota_get_qdata(req, QUOTA_REQUEST, QUOTA_EXPORT);
2448 if (IS_ERR(qdata)) {
2449 rc = PTR_ERR(qdata);
2450 CDEBUG(D_ERROR, "Can't unpack qunit_data(rc: %d)\n", rc);
2451 req->rq_status = rc;
2455 /* we use the observer */
2456 if (obd_pin_observer(obd, &lov_obd) ||
2457 obd_pin_observer(lov_obd, &master_obd)) {
2458 CERROR("Can't find the observer, it is recovering\n");
2459 req->rq_status = -EAGAIN;
2463 obt = &master_obd->u.obt;
2464 qctxt = &obt->obt_qctxt;
2466 if (!qctxt->lqc_setup || !qctxt->lqc_valid) {
2467 /* quota_type has not been processed yet, return EAGAIN
2468 * until we know whether or not quotas are supposed to
2470 CDEBUG(D_QUOTA, "quota_type not processed yet, return "
2472 req->rq_status = -EAGAIN;
2476 cfs_down_read(&obt->obt_rwsem);
2477 if (qctxt->lqc_lqs_hash == NULL) {
2478 cfs_up_read(&obt->obt_rwsem);
2479 /* quota_type has not been processed yet, return EAGAIN
2480 * until we know whether or not quotas are supposed to
2482 CDEBUG(D_QUOTA, "quota_ctxt is not ready yet, return "
2484 req->rq_status = -EAGAIN;
2488 LASSERT(qctxt->lqc_handler);
2489 rc = qctxt->lqc_handler(master_obd, qdata,
2490 lustre_msg_get_opc(req->rq_reqmsg));
2491 cfs_up_read(&obt->obt_rwsem);
2492 if (rc && rc != -EDQUOT)
2493 CDEBUG(rc == -EBUSY ? D_QUOTA : D_ERROR,
2494 "dqacq/dqrel failed! (rc:%d)\n", rc);
2495 req->rq_status = rc;
2497 rc = quota_copy_qdata(req, qdata, QUOTA_REPLY, QUOTA_EXPORT);
2499 CERROR("Can't pack qunit_data(rc: %d)\n", rc);
2503 /* Block the quota req. b=14840 */
2504 OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_BLOCK_QUOTA_REQ, obd_timeout);
2509 obd_unpin_observer(lov_obd);
2511 obd_unpin_observer(obd);
2513 rc = ptlrpc_reply(req);
2517 #endif /* !__KERNEL__ */
2519 #endif /* HAVE_QUOTA_SUPPORT */
2521 ldlm_mode_t lck_compat_array[] = {
2522 [LCK_EX] LCK_COMPAT_EX,
2523 [LCK_PW] LCK_COMPAT_PW,
2524 [LCK_PR] LCK_COMPAT_PR,
2525 [LCK_CW] LCK_COMPAT_CW,
2526 [LCK_CR] LCK_COMPAT_CR,
2527 [LCK_NL] LCK_COMPAT_NL,
2528 [LCK_GROUP] LCK_COMPAT_GROUP,
2529 [LCK_COS] LCK_COMPAT_COS,
2533 * Rather arbitrary mapping from LDLM error codes to errno values. This should
2534 * not escape to the user level.
2536 int ldlm_error2errno(ldlm_error_t error)
2544 case ELDLM_LOCK_CHANGED:
2547 case ELDLM_LOCK_ABORTED:
2550 case ELDLM_LOCK_REPLACED:
2553 case ELDLM_NO_LOCK_DATA:
2556 case ELDLM_NAMESPACE_EXISTS:
2559 case ELDLM_BAD_NAMESPACE:
2563 if (((int)error) < 0) /* cast to signed type */
2564 result = error; /* as ldlm_error_t can be unsigned */
2566 CERROR("Invalid DLM result code: %d\n", error);
2572 EXPORT_SYMBOL(ldlm_error2errno);
2575 * Dual to ldlm_error2errno(): maps errno values back to ldlm_error_t.
2577 ldlm_error_t ldlm_errno2error(int err_no)
2586 error = ELDLM_LOCK_CHANGED;
2589 error = ELDLM_LOCK_ABORTED;
2592 error = ELDLM_LOCK_REPLACED;
2595 error = ELDLM_NO_LOCK_DATA;
2598 error = ELDLM_NAMESPACE_EXISTS;
2601 error = ELDLM_BAD_NAMESPACE;
2608 EXPORT_SYMBOL(ldlm_errno2error);
2610 #if LUSTRE_TRACKS_LOCK_EXP_REFS
2611 void ldlm_dump_export_locks(struct obd_export *exp)
2613 cfs_spin_lock(&exp->exp_locks_list_guard);
2614 if (!cfs_list_empty(&exp->exp_locks_list)) {
2615 struct ldlm_lock *lock;
2617 CERROR("dumping locks for export %p,"
2618 "ignore if the unmount doesn't hang\n", exp);
2619 cfs_list_for_each_entry(lock, &exp->exp_locks_list, l_exp_refs_link)
2620 ldlm_lock_dump(D_ERROR, lock, 0);
2622 cfs_spin_unlock(&exp->exp_locks_list_guard);
2626 #ifdef HAVE_SERVER_SUPPORT
2627 static int target_bulk_timeout(void *data)
2630 /* We don't fail the connection here, because having the export
2631 * killed makes the (vital) call to commitrw very sad.
2636 static inline char *bulk2type(struct ptlrpc_bulk_desc *desc)
2638 return desc->bd_type == BULK_GET_SINK ? "GET" : "PUT";
2641 int target_bulk_io(struct obd_export *exp, struct ptlrpc_bulk_desc *desc,
2642 struct l_wait_info *lwi)
2644 struct ptlrpc_request *req = desc->bd_req;
2648 /* Check if there is eviction in progress, and if so, wait for
2650 if (unlikely(cfs_atomic_read(&exp->exp_obd->obd_evict_inprogress))) {
2651 *lwi = LWI_INTR(NULL, NULL);
2652 rc = l_wait_event(exp->exp_obd->obd_evict_inprogress_waitq,
2653 !cfs_atomic_read(&exp->exp_obd->
2654 obd_evict_inprogress),
2658 /* Check if client was evicted or tried to reconnect already */
2659 if (exp->exp_failed || exp->exp_abort_active_req) {
2662 if (desc->bd_type == BULK_PUT_SINK)
2663 rc = sptlrpc_svc_wrap_bulk(req, desc);
2665 rc = ptlrpc_start_bulk_transfer(desc);
2668 if (rc == 0 && OBD_FAIL_CHECK(OBD_FAIL_MDS_SENDPAGE)) {
2669 ptlrpc_abort_bulk(desc);
2670 } else if (rc == 0) {
2671 time_t start = cfs_time_current_sec();
2673 long timeoutl = req->rq_deadline - cfs_time_current_sec();
2674 cfs_duration_t timeout = timeoutl <= 0 ?
2675 CFS_TICK : cfs_time_seconds(timeoutl);
2676 *lwi = LWI_TIMEOUT_INTERVAL(timeout,
2677 cfs_time_seconds(1),
2678 target_bulk_timeout,
2680 rc = l_wait_event(desc->bd_waitq,
2681 !ptlrpc_server_bulk_active(desc) ||
2683 exp->exp_abort_active_req,
2685 LASSERT(rc == 0 || rc == -ETIMEDOUT);
2686 /* Wait again if we changed deadline */
2687 } while ((rc == -ETIMEDOUT) &&
2688 (req->rq_deadline > cfs_time_current_sec()));
2690 if (rc == -ETIMEDOUT) {
2691 DEBUG_REQ(D_ERROR, req,
2692 "timeout on bulk %s after %ld%+lds",
2694 req->rq_deadline - start,
2695 cfs_time_current_sec() -
2697 ptlrpc_abort_bulk(desc);
2698 } else if (exp->exp_failed) {
2699 DEBUG_REQ(D_ERROR, req, "Eviction on bulk %s",
2702 ptlrpc_abort_bulk(desc);
2703 } else if (exp->exp_abort_active_req) {
2704 DEBUG_REQ(D_ERROR, req, "Reconnect on bulk %s",
2706 /* we don't reply anyway */
2708 ptlrpc_abort_bulk(desc);
2709 } else if (!desc->bd_success ||
2710 desc->bd_nob_transferred != desc->bd_nob) {
2711 DEBUG_REQ(D_ERROR, req, "%s bulk %s %d(%d)",
2713 "truncated" : "network error on",
2715 desc->bd_nob_transferred,
2717 /* XXX should this be a different errno? */
2719 } else if (desc->bd_type == BULK_GET_SINK) {
2720 rc = sptlrpc_svc_unwrap_bulk(req, desc);
2723 DEBUG_REQ(D_ERROR, req, "bulk %s failed: rc %d",
2724 bulk2type(desc), rc);
2729 EXPORT_SYMBOL(target_bulk_io);
2731 #endif /* HAVE_SERVER_SUPPORT */