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 #ifdef HAVE_SERVER_SUPPORT
581 int server_disconnect_export(struct obd_export *exp)
586 /* Disconnect early so that clients can't keep using export */
587 rc = class_disconnect(exp);
588 /* close import for avoid sending any requests */
589 if (exp->exp_imp_reverse)
590 ptlrpc_cleanup_imp(exp->exp_imp_reverse);
592 if (exp->exp_obd->obd_namespace != NULL)
593 ldlm_cancel_locks_for_export(exp);
595 /* complete all outstanding replies */
596 cfs_spin_lock(&exp->exp_lock);
597 while (!cfs_list_empty(&exp->exp_outstanding_replies)) {
598 struct ptlrpc_reply_state *rs =
599 cfs_list_entry(exp->exp_outstanding_replies.next,
600 struct ptlrpc_reply_state, rs_exp_list);
601 struct ptlrpc_service *svc = rs->rs_service;
603 cfs_spin_lock(&svc->srv_rs_lock);
604 cfs_list_del_init(&rs->rs_exp_list);
605 cfs_spin_lock(&rs->rs_lock);
606 ptlrpc_schedule_difficult_reply(rs);
607 cfs_spin_unlock(&rs->rs_lock);
608 cfs_spin_unlock(&svc->srv_rs_lock);
610 cfs_spin_unlock(&exp->exp_lock);
615 /* --------------------------------------------------------------------------
616 * from old lib/target.c
617 * -------------------------------------------------------------------------- */
619 static int target_handle_reconnect(struct lustre_handle *conn,
620 struct obd_export *exp,
621 struct obd_uuid *cluuid)
625 if (exp->exp_connection && exp->exp_imp_reverse) {
626 struct lustre_handle *hdl;
627 struct obd_device *target;
629 hdl = &exp->exp_imp_reverse->imp_remote_handle;
630 target = exp->exp_obd;
632 /* Might be a re-connect after a partition. */
633 if (!memcmp(&conn->cookie, &hdl->cookie, sizeof conn->cookie)) {
634 if (target->obd_recovering) {
635 int timeout = cfs_duration_sec(cfs_time_sub(
637 &target->obd_recovery_timer),
638 cfs_time_current()));
640 LCONSOLE_WARN("%s: Client %s (at %s) reconnect"
641 "ing, waiting for %d clients in recov"
642 "ery for %d:%.02d\n", target->obd_name,
643 obd_uuid2str(&exp->exp_client_uuid),
644 obd_export_nid2str(exp),
645 target->obd_max_recoverable_clients,
646 timeout / 60, timeout % 60);
648 LCONSOLE_WARN("%s: Client %s (at %s) "
649 "reconnecting\n", target->obd_name,
650 obd_uuid2str(&exp->exp_client_uuid),
651 obd_export_nid2str(exp));
654 conn->cookie = exp->exp_handle.h_cookie;
655 /* target_handle_connect() treats EALREADY and
656 * -EALREADY differently. EALREADY means we are
657 * doing a valid reconnect from the same client. */
660 LCONSOLE_WARN("%s: The server has already connected "
661 "client %s (at %s) with handle " LPX64
662 ", rejecting a client with the same "
663 "uuid trying to reconnect with "
664 "handle " LPX64, target->obd_name,
665 obd_uuid2str(&exp->exp_client_uuid),
666 obd_export_nid2str(exp),
667 hdl->cookie, conn->cookie);
668 memset(conn, 0, sizeof *conn);
669 /* target_handle_connect() treats EALREADY and
670 * -EALREADY differently. -EALREADY is an error
671 * (same UUID, different handle). */
676 conn->cookie = exp->exp_handle.h_cookie;
677 CDEBUG(D_HA, "connect export for UUID '%s' at %p, cookie "LPX64"\n",
678 cluuid->uuid, exp, conn->cookie);
682 void target_client_add_cb(struct obd_device *obd, __u64 transno, void *cb_data,
685 struct obd_export *exp = cb_data;
687 CDEBUG(D_RPCTRACE, "%s: committing for initial connect of %s\n",
688 obd->obd_name, exp->exp_client_uuid.uuid);
690 cfs_spin_lock(&exp->exp_lock);
691 exp->exp_need_sync = 0;
692 cfs_spin_unlock(&exp->exp_lock);
693 class_export_cb_put(exp);
695 EXPORT_SYMBOL(target_client_add_cb);
699 check_and_start_recovery_timer(struct obd_device *obd,
700 struct ptlrpc_request *req, int new_client);
703 check_and_start_recovery_timer(struct obd_device *obd,
704 struct ptlrpc_request *req, int new_client)
709 int target_handle_connect(struct ptlrpc_request *req)
711 struct obd_device *target, *targref = NULL;
712 struct obd_export *export = NULL;
713 struct obd_import *revimp;
714 struct lustre_handle conn;
715 struct lustre_handle *tmp;
716 struct obd_uuid tgtuuid;
717 struct obd_uuid cluuid;
718 struct obd_uuid remote_uuid;
724 struct obd_connect_data *data, *tmpdata;
726 lnet_nid_t *client_nid = NULL;
729 OBD_RACE(OBD_FAIL_TGT_CONN_RACE);
731 str = req_capsule_client_get(&req->rq_pill, &RMF_TGTUUID);
733 DEBUG_REQ(D_ERROR, req, "bad target UUID for connect");
734 GOTO(out, rc = -EINVAL);
737 obd_str2uuid(&tgtuuid, str);
738 target = class_uuid2obd(&tgtuuid);
740 target = class_name2obd(str);
742 if (!target || target->obd_stopping || !target->obd_set_up) {
743 deuuidify(str, NULL, &target_start, &target_len);
744 LCONSOLE_ERROR_MSG(0x137, "%.*s: Not available for connect "
745 "from %s (%s)\n", target_len, target_start,
746 libcfs_nid2str(req->rq_peer.nid), !target ?
747 "no target" : (target->obd_stopping ?
748 "stopping" : "not set up"));
749 GOTO(out, rc = -ENODEV);
752 if (target->obd_no_conn) {
753 LCONSOLE_WARN("%s: Temporarily refusing client connection "
754 "from %s\n", target->obd_name,
755 libcfs_nid2str(req->rq_peer.nid));
756 GOTO(out, rc = -EAGAIN);
759 /* Make sure the target isn't cleaned up while we're here. Yes,
760 there's still a race between the above check and our incref here.
761 Really, class_uuid2obd should take the ref. */
762 targref = class_incref(target, __FUNCTION__, cfs_current());
765 str = req_capsule_client_get(&req->rq_pill, &RMF_CLUUID);
767 DEBUG_REQ(D_ERROR, req, "bad client UUID for connect");
768 GOTO(out, rc = -EINVAL);
771 obd_str2uuid(&cluuid, str);
773 /* XXX extract a nettype and format accordingly */
774 switch (sizeof(lnet_nid_t)) {
775 /* NB the casts only avoid compiler warnings */
777 snprintf(remote_uuid.uuid, sizeof remote_uuid,
778 "NET_"LPX64"_UUID", (__u64)req->rq_peer.nid);
781 snprintf(remote_uuid.uuid, sizeof remote_uuid,
782 "NET_%x_UUID", (__u32)req->rq_peer.nid);
788 tmp = req_capsule_client_get(&req->rq_pill, &RMF_CONN);
790 GOTO(out, rc = -EPROTO);
794 size = req_capsule_get_size(&req->rq_pill, &RMF_CONNECT_DATA,
796 data = req_capsule_client_get(&req->rq_pill, &RMF_CONNECT_DATA);
798 GOTO(out, rc = -EPROTO);
800 rc = req_capsule_server_pack(&req->rq_pill);
804 if (lustre_msg_get_op_flags(req->rq_reqmsg) & MSG_CONNECT_LIBCLIENT) {
806 DEBUG_REQ(D_WARNING, req, "Refusing old (unversioned) "
807 "libclient connection attempt");
808 GOTO(out, rc = -EPROTO);
809 } else if (data->ocd_version < LUSTRE_VERSION_CODE -
810 LUSTRE_VERSION_ALLOWED_OFFSET ||
811 data->ocd_version > LUSTRE_VERSION_CODE +
812 LUSTRE_VERSION_ALLOWED_OFFSET) {
813 DEBUG_REQ(D_WARNING, req, "Refusing %s (%d.%d.%d.%d) "
814 "libclient connection attempt",
815 data->ocd_version < LUSTRE_VERSION_CODE ?
817 OBD_OCD_VERSION_MAJOR(data->ocd_version),
818 OBD_OCD_VERSION_MINOR(data->ocd_version),
819 OBD_OCD_VERSION_PATCH(data->ocd_version),
820 OBD_OCD_VERSION_FIX(data->ocd_version));
821 data = req_capsule_server_sized_get(&req->rq_pill,
823 offsetof(typeof(*data), ocd_version) +
824 sizeof(data->ocd_version));
826 data->ocd_connect_flags = OBD_CONNECT_VERSION;
827 data->ocd_version = LUSTRE_VERSION_CODE;
829 GOTO(out, rc = -EPROTO);
833 if ((lustre_msg_get_op_flags(req->rq_reqmsg) & MSG_CONNECT_INITIAL) &&
834 (data->ocd_connect_flags & OBD_CONNECT_MDS))
837 /* lctl gets a backstage, all-access pass. */
838 if (obd_uuid_equals(&cluuid, &target->obd_uuid))
839 goto dont_check_exports;
841 export = cfs_hash_lookup(target->obd_uuid_hash, &cluuid);
845 /* we've found an export in the hash */
846 if (export->exp_connecting) { /* bug 9635, et. al. */
847 LCONSOLE_WARN("%s: Export %p already connecting from %s\n",
848 export->exp_obd->obd_name, export,
849 libcfs_nid2str(req->rq_peer.nid));
850 class_export_put(export);
853 } else if (mds_conn && export->exp_connection) {
854 if (req->rq_peer.nid != export->exp_connection->c_peer.nid)
855 /* mds reconnected after failover */
856 LCONSOLE_WARN("%s: Received MDS connection from "
857 "%s, removing former export from %s\n",
858 target->obd_name, libcfs_nid2str(req->rq_peer.nid),
859 libcfs_nid2str(export->exp_connection->c_peer.nid));
861 /* new mds connection from the same nid */
862 LCONSOLE_WARN("%s: Received new MDS connection from "
863 "%s, removing former export from same NID\n",
864 target->obd_name, libcfs_nid2str(req->rq_peer.nid));
865 class_fail_export(export);
866 class_export_put(export);
869 } else if (export->exp_connection != NULL &&
870 req->rq_peer.nid != export->exp_connection->c_peer.nid &&
871 (lustre_msg_get_op_flags(req->rq_reqmsg) &
872 MSG_CONNECT_INITIAL)) {
873 /* in mds failover we have static uuid but nid can be
875 LCONSOLE_WARN("%s: Client %s seen on new nid %s when "
876 "existing nid %s is already connected\n",
877 target->obd_name, cluuid.uuid,
878 libcfs_nid2str(req->rq_peer.nid),
880 export->exp_connection->c_peer.nid));
882 class_export_put(export);
885 cfs_spin_lock(&export->exp_lock);
886 export->exp_connecting = 1;
887 cfs_spin_unlock(&export->exp_lock);
888 LASSERT(export->exp_obd == target);
890 rc = target_handle_reconnect(&conn, export, &cluuid);
893 /* If we found an export, we already unlocked. */
896 OBD_FAIL_TIMEOUT(OBD_FAIL_TGT_DELAY_CONNECT, 2 * obd_timeout);
897 } else if (req->rq_export == NULL &&
898 cfs_atomic_read(&export->exp_rpc_count) > 0) {
899 LCONSOLE_WARN("%s: Client %s (at %s) refused connection, "
900 "still busy with %d references\n",
901 target->obd_name, cluuid.uuid,
902 libcfs_nid2str(req->rq_peer.nid),
903 cfs_atomic_read(&export->exp_refcount));
904 GOTO(out, rc = -EBUSY);
905 } else if (req->rq_export != NULL &&
906 (cfs_atomic_read(&export->exp_rpc_count) > 1)) {
907 /* the current connect rpc has increased exp_rpc_count */
908 LCONSOLE_WARN("%s: Client %s (at %s) refused reconnection, "
909 "still busy with %d active RPCs\n",
910 target->obd_name, cluuid.uuid,
911 libcfs_nid2str(req->rq_peer.nid),
912 cfs_atomic_read(&export->exp_rpc_count) - 1);
913 cfs_spin_lock(&export->exp_lock);
914 if (req->rq_export->exp_conn_cnt <
915 lustre_msg_get_conn_cnt(req->rq_reqmsg))
916 /* try to abort active requests */
917 req->rq_export->exp_abort_active_req = 1;
918 cfs_spin_unlock(&export->exp_lock);
919 GOTO(out, rc = -EBUSY);
920 } else if (lustre_msg_get_conn_cnt(req->rq_reqmsg) == 1) {
921 if (!strstr(cluuid.uuid, "mdt"))
922 LCONSOLE_WARN("%s: Rejecting reconnect from the "
923 "known client %s (at %s) because it "
924 "is indicating it is a new client",
925 target->obd_name, cluuid.uuid,
926 libcfs_nid2str(req->rq_peer.nid));
927 GOTO(out, rc = -EALREADY);
929 OBD_FAIL_TIMEOUT(OBD_FAIL_TGT_DELAY_RECONNECT, 2 * obd_timeout);
936 CDEBUG(D_HA, "%s: connection from %s@%s %st"LPU64" exp %p cur %ld last %ld\n",
937 target->obd_name, cluuid.uuid, libcfs_nid2str(req->rq_peer.nid),
938 target->obd_recovering ? "recovering/" : "", data->ocd_transno,
939 export, (long)cfs_time_current_sec(),
940 export ? (long)export->exp_last_request_time : 0);
942 /* If this is the first time a client connects, reset the recovery
944 if (rc == 0 && target->obd_recovering)
945 check_and_start_recovery_timer(target, req, export == NULL);
947 /* We want to handle EALREADY but *not* -EALREADY from
948 * target_handle_reconnect(), return reconnection state in a flag */
949 if (rc == EALREADY) {
950 lustre_msg_add_op_flags(req->rq_repmsg, MSG_CONNECT_RECONNECT);
956 /* Tell the client if we support replayable requests */
957 if (target->obd_replayable)
958 lustre_msg_add_op_flags(req->rq_repmsg, MSG_CONNECT_REPLAYABLE);
959 client_nid = &req->rq_peer.nid;
961 if (export == NULL) {
962 if (target->obd_recovering) {
965 t = cfs_timer_deadline(&target->obd_recovery_timer);
966 t = cfs_time_sub(t, cfs_time_current());
967 t = cfs_duration_sec(t);
968 LCONSOLE_WARN("%s: Denying connection for new client "
969 "%s (at %s), waiting for %d clients in "
970 "recovery for %d:%.02d\n",
972 libcfs_nid2str(req->rq_peer.nid),
974 cfs_atomic_read(&target-> \
975 obd_lock_replay_clients),
976 (int)t / 60, (int)t % 60);
980 rc = obd_connect(req->rq_svc_thread->t_env,
981 &export, target, &cluuid, data,
984 conn.cookie = export->exp_handle.h_cookie;
985 /* LU-1092 reconnect put export refcount in the
986 * end, connect needs take one here too. */
987 class_export_get(export);
991 rc = obd_reconnect(req->rq_svc_thread->t_env,
992 export, target, &cluuid, data, client_nid);
994 /* prevous done via class_conn2export */
995 class_export_get(export);
1000 LASSERT(target->u.obt.obt_magic == OBT_MAGIC);
1001 data->ocd_instance = target->u.obt.obt_instance;
1003 /* Return only the parts of obd_connect_data that we understand, so the
1004 * client knows that we don't understand the rest. */
1006 tmpsize = req_capsule_get_size(&req->rq_pill, &RMF_CONNECT_DATA,
1008 tmpdata = req_capsule_server_get(&req->rq_pill,
1010 /* Don't use struct assignment here, because the client reply
1011 * buffer may be smaller/larger than the local struct
1012 * obd_connect_data. */
1013 memcpy(tmpdata, data, min(tmpsize, size));
1016 /* If all else goes well, this is our RPC return code. */
1019 lustre_msg_set_handle(req->rq_repmsg, &conn);
1021 /* If the client and the server are the same node, we will already
1022 * have an export that really points to the client's DLM export,
1023 * because we have a shared handles table.
1025 * XXX this will go away when shaver stops sending the "connect" handle
1026 * in the real "remote handle" field of the request --phik 24 Apr 2003
1028 if (req->rq_export != NULL)
1029 class_export_put(req->rq_export);
1031 req->rq_export = export;
1033 cfs_spin_lock(&export->exp_lock);
1034 if (export->exp_conn_cnt >= lustre_msg_get_conn_cnt(req->rq_reqmsg)) {
1035 cfs_spin_unlock(&export->exp_lock);
1036 CDEBUG(D_RPCTRACE, "%s: %s already connected at higher "
1037 "conn_cnt: %d > %d\n",
1038 cluuid.uuid, libcfs_nid2str(req->rq_peer.nid),
1039 export->exp_conn_cnt,
1040 lustre_msg_get_conn_cnt(req->rq_reqmsg));
1042 GOTO(out, rc = -EALREADY);
1044 LASSERT(lustre_msg_get_conn_cnt(req->rq_reqmsg) > 0);
1045 export->exp_conn_cnt = lustre_msg_get_conn_cnt(req->rq_reqmsg);
1046 export->exp_abort_active_req = 0;
1048 /* request from liblustre? Don't evict it for not pinging. */
1049 if (lustre_msg_get_op_flags(req->rq_reqmsg) & MSG_CONNECT_LIBCLIENT) {
1050 export->exp_libclient = 1;
1051 cfs_spin_unlock(&export->exp_lock);
1053 cfs_spin_lock(&target->obd_dev_lock);
1054 cfs_list_del_init(&export->exp_obd_chain_timed);
1055 cfs_spin_unlock(&target->obd_dev_lock);
1057 cfs_spin_unlock(&export->exp_lock);
1060 if (export->exp_connection != NULL) {
1061 /* Check to see if connection came from another NID */
1062 if ((export->exp_connection->c_peer.nid != req->rq_peer.nid) &&
1063 !cfs_hlist_unhashed(&export->exp_nid_hash))
1064 cfs_hash_del(export->exp_obd->obd_nid_hash,
1065 &export->exp_connection->c_peer.nid,
1066 &export->exp_nid_hash);
1068 ptlrpc_connection_put(export->exp_connection);
1071 export->exp_connection = ptlrpc_connection_get(req->rq_peer,
1074 if (cfs_hlist_unhashed(&export->exp_nid_hash)) {
1075 cfs_hash_add(export->exp_obd->obd_nid_hash,
1076 &export->exp_connection->c_peer.nid,
1077 &export->exp_nid_hash);
1080 class_disconnect->class_export_recovery_cleanup() race
1082 if (target->obd_recovering && !export->exp_in_recovery) {
1084 __u64 transno = data->ocd_transno;
1086 cfs_spin_lock(&export->exp_lock);
1087 export->exp_in_recovery = 1;
1088 export->exp_req_replay_needed = 1;
1089 export->exp_lock_replay_needed = 1;
1090 cfs_spin_unlock(&export->exp_lock);
1092 has_transno = !!(lustre_msg_get_op_flags(req->rq_reqmsg) &
1093 MSG_CONNECT_TRANSNO);
1094 if (has_transno && transno == 0)
1095 CWARN("Connect with zero transno!\n");
1097 if (has_transno && transno > 0 &&
1098 transno < target->obd_next_recovery_transno &&
1099 transno > target->obd_last_committed) {
1100 /* another way is to use cmpxchg() so it will be
1102 cfs_spin_lock(&target->obd_recovery_task_lock);
1103 if (transno < target->obd_next_recovery_transno)
1104 target->obd_next_recovery_transno = transno;
1105 cfs_spin_unlock(&target->obd_recovery_task_lock);
1108 cfs_atomic_inc(&target->obd_req_replay_clients);
1109 cfs_atomic_inc(&target->obd_lock_replay_clients);
1110 if (cfs_atomic_inc_return(&target->obd_connected_clients) ==
1111 target->obd_max_recoverable_clients)
1112 cfs_waitq_signal(&target->obd_next_transno_waitq);
1115 /* Tell the client we're in recovery, when client is involved in it. */
1116 if (target->obd_recovering)
1117 lustre_msg_add_op_flags(req->rq_repmsg, MSG_CONNECT_RECOVERING);
1119 tmp = req_capsule_client_get(&req->rq_pill, &RMF_CONN);
1122 if (export->exp_imp_reverse != NULL) {
1123 /* destroyed import can be still referenced in ctxt */
1124 obd_set_info_async(export, sizeof(KEY_REVIMP_UPD),
1125 KEY_REVIMP_UPD, 0, NULL, NULL);
1127 client_destroy_import(export->exp_imp_reverse);
1130 /* for the rest part, we return -ENOTCONN in case of errors
1131 * in order to let client initialize connection again.
1133 revimp = export->exp_imp_reverse = class_new_import(target);
1135 CERROR("fail to alloc new reverse import.\n");
1136 GOTO(out, rc = -ENOTCONN);
1139 revimp->imp_connection = ptlrpc_connection_addref(export->exp_connection);
1140 revimp->imp_client = &export->exp_obd->obd_ldlm_client;
1141 revimp->imp_remote_handle = conn;
1142 revimp->imp_dlm_fake = 1;
1143 revimp->imp_state = LUSTRE_IMP_FULL;
1145 /* unknown versions will be caught in
1146 * ptlrpc_handle_server_req_in->lustre_unpack_msg() */
1147 revimp->imp_msg_magic = req->rq_reqmsg->lm_magic;
1149 if ((export->exp_connect_flags & OBD_CONNECT_AT) &&
1150 (revimp->imp_msg_magic != LUSTRE_MSG_MAGIC_V1))
1151 revimp->imp_msghdr_flags |= MSGHDR_AT_SUPPORT;
1153 revimp->imp_msghdr_flags &= ~MSGHDR_AT_SUPPORT;
1155 if ((export->exp_connect_flags & OBD_CONNECT_FULL20) &&
1156 (revimp->imp_msg_magic != LUSTRE_MSG_MAGIC_V1))
1157 revimp->imp_msghdr_flags |= MSGHDR_CKSUM_INCOMPAT18;
1159 revimp->imp_msghdr_flags &= ~MSGHDR_CKSUM_INCOMPAT18;
1161 rc = sptlrpc_import_sec_adapt(revimp, req->rq_svc_ctx, &req->rq_flvr);
1163 CERROR("Failed to get sec for reverse import: %d\n", rc);
1164 export->exp_imp_reverse = NULL;
1165 class_destroy_import(revimp);
1168 class_import_put(revimp);
1171 cfs_spin_lock(&export->exp_lock);
1172 export->exp_connecting = 0;
1173 cfs_spin_unlock(&export->exp_lock);
1175 class_export_put(export);
1178 class_decref(targref, __FUNCTION__, cfs_current());
1180 req->rq_status = rc;
1184 int target_handle_disconnect(struct ptlrpc_request *req)
1189 rc = req_capsule_server_pack(&req->rq_pill);
1193 /* keep the rq_export around so we can send the reply */
1194 req->rq_status = obd_disconnect(class_export_get(req->rq_export));
1199 void target_destroy_export(struct obd_export *exp)
1201 /* exports created from last_rcvd data, and "fake"
1202 exports created by lctl don't have an import */
1203 if (exp->exp_imp_reverse != NULL)
1204 client_destroy_import(exp->exp_imp_reverse);
1206 LASSERT_ATOMIC_ZERO(&exp->exp_locks_count);
1207 LASSERT_ATOMIC_ZERO(&exp->exp_rpc_count);
1208 LASSERT_ATOMIC_ZERO(&exp->exp_cb_count);
1209 LASSERT_ATOMIC_ZERO(&exp->exp_replay_count);
1213 * Recovery functions
1215 static void target_request_copy_get(struct ptlrpc_request *req)
1217 class_export_rpc_get(req->rq_export);
1218 LASSERT(cfs_list_empty(&req->rq_list));
1219 CFS_INIT_LIST_HEAD(&req->rq_replay_list);
1221 /* increase refcount to keep request in queue */
1222 cfs_atomic_inc(&req->rq_refcount);
1223 /** let export know it has replays to be handled */
1224 cfs_atomic_inc(&req->rq_export->exp_replay_count);
1227 static void target_request_copy_put(struct ptlrpc_request *req)
1229 LASSERT(cfs_list_empty(&req->rq_replay_list));
1230 LASSERT_ATOMIC_POS(&req->rq_export->exp_replay_count);
1232 cfs_atomic_dec(&req->rq_export->exp_replay_count);
1233 class_export_rpc_put(req->rq_export);
1234 ptlrpc_server_drop_request(req);
1237 static int target_exp_enqueue_req_replay(struct ptlrpc_request *req)
1239 __u64 transno = lustre_msg_get_transno(req->rq_reqmsg);
1240 struct obd_export *exp = req->rq_export;
1241 struct ptlrpc_request *reqiter;
1246 cfs_spin_lock(&exp->exp_lock);
1247 cfs_list_for_each_entry(reqiter, &exp->exp_req_replay_queue,
1249 if (lustre_msg_get_transno(reqiter->rq_reqmsg) == transno) {
1256 /* we expect it with RESENT and REPLAY flags */
1257 if ((lustre_msg_get_flags(req->rq_reqmsg) &
1258 (MSG_RESENT | MSG_REPLAY)) != (MSG_RESENT | MSG_REPLAY))
1259 CERROR("invalid flags %x of resent replay\n",
1260 lustre_msg_get_flags(req->rq_reqmsg));
1262 cfs_list_add_tail(&req->rq_replay_list,
1263 &exp->exp_req_replay_queue);
1266 cfs_spin_unlock(&exp->exp_lock);
1270 static void target_exp_dequeue_req_replay(struct ptlrpc_request *req)
1272 LASSERT(!cfs_list_empty(&req->rq_replay_list));
1273 LASSERT(req->rq_export);
1275 cfs_spin_lock(&req->rq_export->exp_lock);
1276 cfs_list_del_init(&req->rq_replay_list);
1277 cfs_spin_unlock(&req->rq_export->exp_lock);
1281 static void target_finish_recovery(struct obd_device *obd)
1283 time_t elapsed_time = max_t(time_t, 1, cfs_time_current_sec() -
1284 obd->obd_recovery_start);
1287 LCONSOLE_INFO("%s: Recovery over after %d:%.02d, of %d clients "
1288 "%d recovered and %d %s evicted.\n", obd->obd_name,
1289 (int)elapsed_time / 60, (int)elapsed_time % 60,
1290 obd->obd_max_recoverable_clients,
1291 cfs_atomic_read(&obd->obd_connected_clients),
1292 obd->obd_stale_clients,
1293 obd->obd_stale_clients == 1 ? "was" : "were");
1295 ldlm_reprocess_all_ns(obd->obd_namespace);
1296 cfs_spin_lock(&obd->obd_recovery_task_lock);
1297 if (!cfs_list_empty(&obd->obd_req_replay_queue) ||
1298 !cfs_list_empty(&obd->obd_lock_replay_queue) ||
1299 !cfs_list_empty(&obd->obd_final_req_queue)) {
1300 CERROR("%s: Recovery queues ( %s%s%s) are not empty\n",
1302 cfs_list_empty(&obd->obd_req_replay_queue) ? "" : "req ",
1303 cfs_list_empty(&obd->obd_lock_replay_queue) ? \
1305 cfs_list_empty(&obd->obd_final_req_queue) ? \
1307 cfs_spin_unlock(&obd->obd_recovery_task_lock);
1310 cfs_spin_unlock(&obd->obd_recovery_task_lock);
1312 obd->obd_recovery_end = cfs_time_current_sec();
1314 /* when recovery finished, cleanup orphans on mds and ost */
1315 if (OBT(obd) && OBP(obd, postrecov)) {
1316 int rc = OBP(obd, postrecov)(obd);
1318 LCONSOLE_WARN("%s: Post recovery failed, rc %d\n",
1324 static void abort_req_replay_queue(struct obd_device *obd)
1326 struct ptlrpc_request *req, *n;
1327 cfs_list_t abort_list;
1329 CFS_INIT_LIST_HEAD(&abort_list);
1330 cfs_spin_lock(&obd->obd_recovery_task_lock);
1331 cfs_list_splice_init(&obd->obd_req_replay_queue, &abort_list);
1332 cfs_spin_unlock(&obd->obd_recovery_task_lock);
1333 cfs_list_for_each_entry_safe(req, n, &abort_list, rq_list) {
1334 DEBUG_REQ(D_WARNING, req, "aborted:");
1335 req->rq_status = -ENOTCONN;
1336 if (ptlrpc_error(req)) {
1337 DEBUG_REQ(D_ERROR, req,
1338 "failed abort_req_reply; skipping");
1340 target_exp_dequeue_req_replay(req);
1341 target_request_copy_put(req);
1345 static void abort_lock_replay_queue(struct obd_device *obd)
1347 struct ptlrpc_request *req, *n;
1348 cfs_list_t abort_list;
1350 CFS_INIT_LIST_HEAD(&abort_list);
1351 cfs_spin_lock(&obd->obd_recovery_task_lock);
1352 cfs_list_splice_init(&obd->obd_lock_replay_queue, &abort_list);
1353 cfs_spin_unlock(&obd->obd_recovery_task_lock);
1354 cfs_list_for_each_entry_safe(req, n, &abort_list, rq_list){
1355 DEBUG_REQ(D_ERROR, req, "aborted:");
1356 req->rq_status = -ENOTCONN;
1357 if (ptlrpc_error(req)) {
1358 DEBUG_REQ(D_ERROR, req,
1359 "failed abort_lock_reply; skipping");
1361 target_request_copy_put(req);
1365 /* Called from a cleanup function if the device is being cleaned up
1366 forcefully. The exports should all have been disconnected already,
1367 the only thing left to do is
1368 - clear the recovery flags
1370 - free queued requests and replies, but don't send replies
1371 Because the obd_stopping flag is set, no new requests should be received.
1374 void target_cleanup_recovery(struct obd_device *obd)
1376 struct ptlrpc_request *req, *n;
1377 cfs_list_t clean_list;
1380 CFS_INIT_LIST_HEAD(&clean_list);
1381 cfs_spin_lock(&obd->obd_dev_lock);
1382 if (!obd->obd_recovering) {
1383 cfs_spin_unlock(&obd->obd_dev_lock);
1387 obd->obd_recovering = obd->obd_abort_recovery = 0;
1388 cfs_spin_unlock(&obd->obd_dev_lock);
1390 cfs_spin_lock(&obd->obd_recovery_task_lock);
1391 target_cancel_recovery_timer(obd);
1392 cfs_list_splice_init(&obd->obd_req_replay_queue, &clean_list);
1393 cfs_spin_unlock(&obd->obd_recovery_task_lock);
1395 cfs_list_for_each_entry_safe(req, n, &clean_list, rq_list) {
1396 LASSERT(req->rq_reply_state == 0);
1397 target_exp_dequeue_req_replay(req);
1398 target_request_copy_put(req);
1401 cfs_spin_lock(&obd->obd_recovery_task_lock);
1402 cfs_list_splice_init(&obd->obd_lock_replay_queue, &clean_list);
1403 cfs_list_splice_init(&obd->obd_final_req_queue, &clean_list);
1404 cfs_spin_unlock(&obd->obd_recovery_task_lock);
1406 cfs_list_for_each_entry_safe(req, n, &clean_list, rq_list){
1407 LASSERT(req->rq_reply_state == 0);
1408 target_request_copy_put(req);
1414 /* obd_recovery_task_lock should be held */
1415 void target_cancel_recovery_timer(struct obd_device *obd)
1417 CDEBUG(D_HA, "%s: cancel recovery timer\n", obd->obd_name);
1418 cfs_timer_disarm(&obd->obd_recovery_timer);
1421 static void target_start_recovery_timer(struct obd_device *obd)
1423 if (obd->obd_recovery_start != 0)
1426 cfs_spin_lock(&obd->obd_dev_lock);
1427 if (!obd->obd_recovering || obd->obd_abort_recovery) {
1428 cfs_spin_unlock(&obd->obd_dev_lock);
1432 LASSERT(obd->obd_recovery_timeout != 0);
1434 if (obd->obd_recovery_start != 0) {
1435 cfs_spin_unlock(&obd->obd_dev_lock);
1439 cfs_timer_arm(&obd->obd_recovery_timer,
1440 cfs_time_shift(obd->obd_recovery_timeout));
1441 obd->obd_recovery_start = cfs_time_current_sec();
1442 cfs_spin_unlock(&obd->obd_dev_lock);
1444 LCONSOLE_WARN("%s: Will be in recovery for at least %d:%.02d, "
1445 "or until %d client%s reconnect%s\n",
1447 obd->obd_recovery_timeout / 60,
1448 obd->obd_recovery_timeout % 60,
1449 obd->obd_max_recoverable_clients,
1450 (obd->obd_max_recoverable_clients == 1) ? "" : "s",
1451 (obd->obd_max_recoverable_clients == 1) ? "s": "");
1455 * extend recovery window.
1457 * if @extend is true, extend recovery window to have @drt remaining at least;
1458 * otherwise, make sure the recovery timeout value is not less than @drt.
1460 static void extend_recovery_timer(struct obd_device *obd, int drt, bool extend)
1464 cfs_duration_t left;
1467 cfs_spin_lock(&obd->obd_dev_lock);
1468 if (!obd->obd_recovering || obd->obd_abort_recovery) {
1469 cfs_spin_unlock(&obd->obd_dev_lock);
1472 LASSERT(obd->obd_recovery_start != 0);
1474 now = cfs_time_current_sec();
1475 to = obd->obd_recovery_timeout;
1476 end = obd->obd_recovery_start + to;
1477 left = cfs_time_sub(end, now);
1479 if (extend && (drt > left)) {
1481 } else if (!extend && (drt > to)) {
1483 /* reduce drt by already passed time */
1484 drt -= obd->obd_recovery_timeout - left;
1487 if (to > obd->obd_recovery_time_hard)
1488 to = obd->obd_recovery_time_hard;
1489 if (obd->obd_recovery_timeout < to) {
1490 obd->obd_recovery_timeout = to;
1491 cfs_timer_arm(&obd->obd_recovery_timer,
1492 cfs_time_shift(drt));
1494 cfs_spin_unlock(&obd->obd_dev_lock);
1496 CDEBUG(D_HA, "%s: recovery timer will expire in %u seconds\n",
1497 obd->obd_name, (unsigned)drt);
1500 /* Reset the timer with each new client connection */
1502 * This timer is actually reconnect_timer, which is for making sure
1503 * the total recovery window is at least as big as my reconnect
1504 * attempt timing. So the initial recovery time_out will be set to
1505 * OBD_RECOVERY_FACTOR * obd_timeout. If the timeout coming
1506 * from client is bigger than this, then the recovery time_out will
1507 * be extended to make sure the client could be reconnected, in the
1508 * process, the timeout from the new client should be ignored.
1512 check_and_start_recovery_timer(struct obd_device *obd,
1513 struct ptlrpc_request *req,
1516 int service_time = lustre_msg_get_service_time(req->rq_reqmsg);
1517 struct obd_device_target *obt = &obd->u.obt;
1518 struct lustre_sb_info *lsi;
1520 if (!new_client && service_time)
1521 /* Teach server about old server's estimates, as first guess
1522 * at how long new requests will take. */
1523 at_measured(&req->rq_rqbd->rqbd_service->srv_at_estimate,
1526 target_start_recovery_timer(obd);
1528 /* convert the service time to rpc timeout,
1529 * reuse service_time to limit stack usage */
1530 service_time = at_est2timeout(service_time);
1532 /* We expect other clients to timeout within service_time, then try
1533 * to reconnect, then try the failover server. The max delay between
1534 * connect attempts is SWITCH_MAX + SWITCH_INC + INITIAL */
1535 service_time += 2 * INITIAL_CONNECT_TIMEOUT;
1537 LASSERT(obt->obt_magic == OBT_MAGIC);
1538 lsi = s2lsi(obt->obt_sb);
1539 if (!(lsi->lsi_flags | LSI_IR_CAPABLE))
1540 service_time += 2 * (CONNECTION_SWITCH_MAX +
1541 CONNECTION_SWITCH_INC);
1542 if (service_time > obd->obd_recovery_timeout && !new_client)
1543 extend_recovery_timer(obd, service_time, false);
1546 /** Health checking routines */
1547 static inline int exp_connect_healthy(struct obd_export *exp)
1549 return (exp->exp_in_recovery);
1552 /** if export done req_replay or has replay in queue */
1553 static inline int exp_req_replay_healthy(struct obd_export *exp)
1555 return (!exp->exp_req_replay_needed ||
1556 cfs_atomic_read(&exp->exp_replay_count) > 0);
1558 /** if export done lock_replay or has replay in queue */
1559 static inline int exp_lock_replay_healthy(struct obd_export *exp)
1561 return (!exp->exp_lock_replay_needed ||
1562 cfs_atomic_read(&exp->exp_replay_count) > 0);
1565 static inline int exp_vbr_healthy(struct obd_export *exp)
1567 return (!exp->exp_vbr_failed);
1570 static inline int exp_finished(struct obd_export *exp)
1572 return (exp->exp_in_recovery && !exp->exp_lock_replay_needed);
1575 /** Checking routines for recovery */
1576 static int check_for_clients(struct obd_device *obd)
1578 unsigned int clnts = cfs_atomic_read(&obd->obd_connected_clients);
1580 if (obd->obd_abort_recovery || obd->obd_recovery_expired)
1582 LASSERT(clnts <= obd->obd_max_recoverable_clients);
1583 if (obd->obd_no_conn == 0 &&
1584 clnts + obd->obd_stale_clients == obd->obd_max_recoverable_clients)
1589 static int check_for_next_transno(struct obd_device *obd)
1591 struct ptlrpc_request *req = NULL;
1592 int wake_up = 0, connected, completed, queue_len;
1593 __u64 next_transno, req_transno;
1596 cfs_spin_lock(&obd->obd_recovery_task_lock);
1597 if (!cfs_list_empty(&obd->obd_req_replay_queue)) {
1598 req = cfs_list_entry(obd->obd_req_replay_queue.next,
1599 struct ptlrpc_request, rq_list);
1600 req_transno = lustre_msg_get_transno(req->rq_reqmsg);
1605 connected = cfs_atomic_read(&obd->obd_connected_clients);
1606 completed = connected - cfs_atomic_read(&obd->obd_req_replay_clients);
1607 queue_len = obd->obd_requests_queued_for_recovery;
1608 next_transno = obd->obd_next_recovery_transno;
1610 CDEBUG(D_HA, "max: %d, connected: %d, completed: %d, queue_len: %d, "
1611 "req_transno: "LPU64", next_transno: "LPU64"\n",
1612 obd->obd_max_recoverable_clients, connected, completed,
1613 queue_len, req_transno, next_transno);
1615 if (obd->obd_abort_recovery) {
1616 CDEBUG(D_HA, "waking for aborted recovery\n");
1618 } else if (obd->obd_recovery_expired) {
1619 CDEBUG(D_HA, "waking for expired recovery\n");
1621 } else if (cfs_atomic_read(&obd->obd_req_replay_clients) == 0) {
1622 CDEBUG(D_HA, "waking for completed recovery\n");
1624 } else if (req_transno == next_transno) {
1625 CDEBUG(D_HA, "waking for next ("LPD64")\n", next_transno);
1627 } else if (queue_len == cfs_atomic_read(&obd->obd_req_replay_clients)) {
1629 /** handle gaps occured due to lost reply or VBR */
1630 LASSERTF(req_transno >= next_transno,
1631 "req_transno: "LPU64", next_transno: "LPU64"\n",
1632 req_transno, next_transno);
1633 if (req_transno > obd->obd_last_committed &&
1634 !obd->obd_version_recov)
1637 "%s: waking for gap in transno, VBR is %s (skip: "
1638 LPD64", ql: %d, comp: %d, conn: %d, next: "LPD64
1639 ", last_committed: "LPD64")\n",
1640 obd->obd_name, obd->obd_version_recov ? "ON" : "OFF",
1641 next_transno, queue_len, completed, connected,
1642 req_transno, obd->obd_last_committed);
1643 obd->obd_next_recovery_transno = req_transno;
1645 } else if (OBD_FAIL_CHECK(OBD_FAIL_MDS_RECOVERY_ACCEPTS_GAPS)) {
1646 CDEBUG(D_HA, "accepting transno gaps is explicitly allowed"
1647 " by fail_lock, waking up ("LPD64")\n", next_transno);
1648 obd->obd_next_recovery_transno = req_transno;
1651 cfs_spin_unlock(&obd->obd_recovery_task_lock);
1655 static int check_for_next_lock(struct obd_device *obd)
1659 cfs_spin_lock(&obd->obd_recovery_task_lock);
1660 if (!cfs_list_empty(&obd->obd_lock_replay_queue)) {
1661 CDEBUG(D_HA, "waking for next lock\n");
1663 } else if (cfs_atomic_read(&obd->obd_lock_replay_clients) == 0) {
1664 CDEBUG(D_HA, "waking for completed lock replay\n");
1666 } else if (obd->obd_abort_recovery) {
1667 CDEBUG(D_HA, "waking for aborted recovery\n");
1669 } else if (obd->obd_recovery_expired) {
1670 CDEBUG(D_HA, "waking for expired recovery\n");
1673 cfs_spin_unlock(&obd->obd_recovery_task_lock);
1679 * wait for recovery events,
1680 * check its status with help of check_routine
1681 * evict dead clients via health_check
1683 static int target_recovery_overseer(struct obd_device *obd,
1684 int (*check_routine)(struct obd_device *),
1685 int (*health_check)(struct obd_export *))
1688 cfs_wait_event(obd->obd_next_transno_waitq, check_routine(obd));
1689 if (obd->obd_abort_recovery) {
1690 CWARN("recovery is aborted, evict exports in recovery\n");
1691 /** evict exports which didn't finish recovery yet */
1692 class_disconnect_stale_exports(obd, exp_finished);
1694 } else if (obd->obd_recovery_expired) {
1695 obd->obd_recovery_expired = 0;
1696 /** If some clients died being recovered, evict them */
1698 "recovery is timed out, evict stale exports\n");
1699 /** evict cexports with no replay in queue, they are stalled */
1700 class_disconnect_stale_exports(obd, health_check);
1701 /** continue with VBR */
1702 cfs_spin_lock(&obd->obd_dev_lock);
1703 obd->obd_version_recov = 1;
1704 cfs_spin_unlock(&obd->obd_dev_lock);
1706 * reset timer, recovery will proceed with versions now,
1707 * timeout is set just to handle reconnection delays
1709 extend_recovery_timer(obd, RECONNECT_DELAY_MAX, true);
1710 /** Wait for recovery events again, after evicting bad clients */
1716 static struct ptlrpc_request *target_next_replay_req(struct obd_device *obd)
1718 struct ptlrpc_request *req = NULL;
1721 CDEBUG(D_HA, "Waiting for transno "LPD64"\n",
1722 obd->obd_next_recovery_transno);
1724 if (target_recovery_overseer(obd, check_for_next_transno,
1725 exp_req_replay_healthy)) {
1726 abort_req_replay_queue(obd);
1727 abort_lock_replay_queue(obd);
1730 cfs_spin_lock(&obd->obd_recovery_task_lock);
1731 if (!cfs_list_empty(&obd->obd_req_replay_queue)) {
1732 req = cfs_list_entry(obd->obd_req_replay_queue.next,
1733 struct ptlrpc_request, rq_list);
1734 cfs_list_del_init(&req->rq_list);
1735 obd->obd_requests_queued_for_recovery--;
1736 cfs_spin_unlock(&obd->obd_recovery_task_lock);
1738 cfs_spin_unlock(&obd->obd_recovery_task_lock);
1739 LASSERT(cfs_list_empty(&obd->obd_req_replay_queue));
1740 LASSERT(cfs_atomic_read(&obd->obd_req_replay_clients) == 0);
1741 /** evict exports failed VBR */
1742 class_disconnect_stale_exports(obd, exp_vbr_healthy);
1747 static struct ptlrpc_request *target_next_replay_lock(struct obd_device *obd)
1749 struct ptlrpc_request *req = NULL;
1751 CDEBUG(D_HA, "Waiting for lock\n");
1752 if (target_recovery_overseer(obd, check_for_next_lock,
1753 exp_lock_replay_healthy))
1754 abort_lock_replay_queue(obd);
1756 cfs_spin_lock(&obd->obd_recovery_task_lock);
1757 if (!cfs_list_empty(&obd->obd_lock_replay_queue)) {
1758 req = cfs_list_entry(obd->obd_lock_replay_queue.next,
1759 struct ptlrpc_request, rq_list);
1760 cfs_list_del_init(&req->rq_list);
1761 cfs_spin_unlock(&obd->obd_recovery_task_lock);
1763 cfs_spin_unlock(&obd->obd_recovery_task_lock);
1764 LASSERT(cfs_list_empty(&obd->obd_lock_replay_queue));
1765 LASSERT(cfs_atomic_read(&obd->obd_lock_replay_clients) == 0);
1766 /** evict exports failed VBR */
1767 class_disconnect_stale_exports(obd, exp_vbr_healthy);
1772 static struct ptlrpc_request *target_next_final_ping(struct obd_device *obd)
1774 struct ptlrpc_request *req = NULL;
1776 cfs_spin_lock(&obd->obd_recovery_task_lock);
1777 if (!cfs_list_empty(&obd->obd_final_req_queue)) {
1778 req = cfs_list_entry(obd->obd_final_req_queue.next,
1779 struct ptlrpc_request, rq_list);
1780 cfs_list_del_init(&req->rq_list);
1781 cfs_spin_unlock(&obd->obd_recovery_task_lock);
1782 if (req->rq_export->exp_in_recovery) {
1783 cfs_spin_lock(&req->rq_export->exp_lock);
1784 req->rq_export->exp_in_recovery = 0;
1785 cfs_spin_unlock(&req->rq_export->exp_lock);
1788 cfs_spin_unlock(&obd->obd_recovery_task_lock);
1793 static int handle_recovery_req(struct ptlrpc_thread *thread,
1794 struct ptlrpc_request *req,
1795 svc_handler_t handler)
1801 * export can be evicted during recovery, no need to handle replays for
1802 * it after that, discard such request silently
1804 if (req->rq_export->exp_disconnected)
1805 GOTO(reqcopy_put, rc = 0);
1807 rc = lu_context_init(&req->rq_recov_session, LCT_SESSION);
1809 CERROR("Failure to initialize session: %d\n", rc);
1810 GOTO(reqcopy_put, rc);
1813 req->rq_recov_session.lc_thread = thread;
1814 lu_context_enter(&req->rq_recov_session);
1815 req->rq_svc_thread = thread;
1816 req->rq_svc_thread->t_env->le_ses = &req->rq_recov_session;
1818 /* thread context */
1819 lu_context_enter(&thread->t_env->le_ctx);
1821 lu_context_exit(&thread->t_env->le_ctx);
1823 lu_context_exit(&req->rq_recov_session);
1824 lu_context_fini(&req->rq_recov_session);
1825 /* don't reset timer for final stage */
1826 if (!exp_finished(req->rq_export)) {
1827 int to = obd_timeout;
1830 * Add request timeout to the recovery time so next request from
1831 * this client may come in recovery time
1834 struct ptlrpc_service *svc = req->rq_rqbd->rqbd_service;
1835 /* If the server sent early reply for this request,
1836 * the client will recalculate the timeout according to
1837 * current server estimate service time, so we will
1838 * use the maxium timeout here for waiting the client
1839 * sending the next req */
1840 to = max((int)at_est2timeout(
1841 at_get(&svc->srv_at_estimate)),
1842 (int)lustre_msg_get_timeout(req->rq_reqmsg));
1843 /* Add net_latency (see ptlrpc_replay_req) */
1844 to += lustre_msg_get_service_time(req->rq_reqmsg);
1846 extend_recovery_timer(class_exp2obd(req->rq_export), to, true);
1852 static int target_recovery_thread(void *arg)
1854 struct lu_target *lut = arg;
1855 struct obd_device *obd = lut->lut_obd;
1856 struct ptlrpc_request *req;
1857 struct target_recovery_data *trd = &obd->obd_recovery_data;
1858 unsigned long delta;
1859 unsigned long flags;
1861 struct ptlrpc_thread *thread = NULL;
1865 cfs_daemonize_ctxt("tgt_recov");
1867 SIGNAL_MASK_LOCK(current, flags);
1868 sigfillset(¤t->blocked);
1870 SIGNAL_MASK_UNLOCK(current, flags);
1872 OBD_ALLOC_PTR(thread);
1878 OBD_FREE_PTR(thread);
1882 rc = lu_context_init(&env->le_ctx, LCT_MD_THREAD);
1884 OBD_FREE_PTR(thread);
1889 thread->t_env = env;
1890 thread->t_id = -1; /* force filter_iobuf_get/put to use local buffers */
1891 env->le_ctx.lc_thread = thread;
1892 thread->t_data = NULL;
1893 thread->t_watchdog = NULL;
1895 CDEBUG(D_HA, "%s: started recovery thread pid %d\n", obd->obd_name,
1897 trd->trd_processing_task = cfs_curproc_pid();
1899 cfs_spin_lock(&obd->obd_dev_lock);
1900 obd->obd_recovering = 1;
1901 cfs_spin_unlock(&obd->obd_dev_lock);
1902 cfs_complete(&trd->trd_starting);
1904 /* first of all, we have to know the first transno to replay */
1905 if (target_recovery_overseer(obd, check_for_clients,
1906 exp_connect_healthy)) {
1907 abort_req_replay_queue(obd);
1908 abort_lock_replay_queue(obd);
1911 /* next stage: replay requests */
1913 CDEBUG(D_INFO, "1: request replay stage - %d clients from t"LPU64"\n",
1914 cfs_atomic_read(&obd->obd_req_replay_clients),
1915 obd->obd_next_recovery_transno);
1916 while ((req = target_next_replay_req(obd))) {
1917 LASSERT(trd->trd_processing_task == cfs_curproc_pid());
1918 DEBUG_REQ(D_HA, req, "processing t"LPD64" from %s",
1919 lustre_msg_get_transno(req->rq_reqmsg),
1920 libcfs_nid2str(req->rq_peer.nid));
1921 handle_recovery_req(thread, req,
1922 trd->trd_recovery_handler);
1924 * bz18031: increase next_recovery_transno before
1925 * target_request_copy_put() will drop exp_rpc reference
1927 cfs_spin_lock(&obd->obd_recovery_task_lock);
1928 obd->obd_next_recovery_transno++;
1929 cfs_spin_unlock(&obd->obd_recovery_task_lock);
1930 target_exp_dequeue_req_replay(req);
1931 target_request_copy_put(req);
1932 obd->obd_replayed_requests++;
1936 * The second stage: replay locks
1938 CDEBUG(D_INFO, "2: lock replay stage - %d clients\n",
1939 cfs_atomic_read(&obd->obd_lock_replay_clients));
1940 while ((req = target_next_replay_lock(obd))) {
1941 LASSERT(trd->trd_processing_task == cfs_curproc_pid());
1942 DEBUG_REQ(D_HA, req, "processing lock from %s: ",
1943 libcfs_nid2str(req->rq_peer.nid));
1944 handle_recovery_req(thread, req,
1945 trd->trd_recovery_handler);
1946 target_request_copy_put(req);
1947 obd->obd_replayed_locks++;
1951 * The third stage: reply on final pings, at this moment all clients
1952 * must have request in final queue
1954 CDEBUG(D_INFO, "3: final stage - process recovery completion pings\n");
1955 /** Update server last boot epoch */
1956 lut_boot_epoch_update(lut);
1957 /* We drop recoverying flag to forward all new requests
1958 * to regular mds_handle() since now */
1959 cfs_spin_lock(&obd->obd_dev_lock);
1960 obd->obd_recovering = obd->obd_abort_recovery = 0;
1961 cfs_spin_unlock(&obd->obd_dev_lock);
1962 cfs_spin_lock(&obd->obd_recovery_task_lock);
1963 target_cancel_recovery_timer(obd);
1964 cfs_spin_unlock(&obd->obd_recovery_task_lock);
1965 while ((req = target_next_final_ping(obd))) {
1966 LASSERT(trd->trd_processing_task == cfs_curproc_pid());
1967 DEBUG_REQ(D_HA, req, "processing final ping from %s: ",
1968 libcfs_nid2str(req->rq_peer.nid));
1969 handle_recovery_req(thread, req,
1970 trd->trd_recovery_handler);
1971 target_request_copy_put(req);
1974 delta = (jiffies - delta) / CFS_HZ;
1975 CDEBUG(D_INFO,"4: recovery completed in %lus - %d/%d reqs/locks\n",
1976 delta, obd->obd_replayed_requests, obd->obd_replayed_locks);
1977 if (delta > OBD_RECOVERY_TIME_SOFT) {
1978 CWARN("too long recovery - read logs\n");
1979 libcfs_debug_dumplog();
1982 target_finish_recovery(obd);
1984 lu_context_fini(&env->le_ctx);
1985 trd->trd_processing_task = 0;
1986 cfs_complete(&trd->trd_finishing);
1988 OBD_FREE_PTR(thread);
1993 static int target_start_recovery_thread(struct lu_target *lut,
1994 svc_handler_t handler)
1996 struct obd_device *obd = lut->lut_obd;
1998 struct target_recovery_data *trd = &obd->obd_recovery_data;
2000 memset(trd, 0, sizeof(*trd));
2001 cfs_init_completion(&trd->trd_starting);
2002 cfs_init_completion(&trd->trd_finishing);
2003 trd->trd_recovery_handler = handler;
2005 if (cfs_create_thread(target_recovery_thread, lut, 0) > 0) {
2006 cfs_wait_for_completion(&trd->trd_starting);
2007 LASSERT(obd->obd_recovering != 0);
2014 void target_stop_recovery_thread(struct obd_device *obd)
2016 if (obd->obd_recovery_data.trd_processing_task > 0) {
2017 struct target_recovery_data *trd = &obd->obd_recovery_data;
2018 /** recovery can be done but postrecovery is not yet */
2019 cfs_spin_lock(&obd->obd_dev_lock);
2020 if (obd->obd_recovering) {
2021 CERROR("%s: Aborting recovery\n", obd->obd_name);
2022 obd->obd_abort_recovery = 1;
2023 cfs_waitq_signal(&obd->obd_next_transno_waitq);
2025 cfs_spin_unlock(&obd->obd_dev_lock);
2026 cfs_wait_for_completion(&trd->trd_finishing);
2030 void target_recovery_fini(struct obd_device *obd)
2032 class_disconnect_exports(obd);
2033 target_stop_recovery_thread(obd);
2034 target_cleanup_recovery(obd);
2036 EXPORT_SYMBOL(target_recovery_fini);
2038 static void target_recovery_expired(unsigned long castmeharder)
2040 struct obd_device *obd = (struct obd_device *)castmeharder;
2041 CDEBUG(D_HA, "%s: recovery timed out; %d clients are still in recovery"
2042 " after %lds (%d clients connected)\n",
2043 obd->obd_name, cfs_atomic_read(&obd->obd_lock_replay_clients),
2044 cfs_time_current_sec()- obd->obd_recovery_start,
2045 cfs_atomic_read(&obd->obd_connected_clients));
2047 obd->obd_recovery_expired = 1;
2048 cfs_waitq_signal(&obd->obd_next_transno_waitq);
2051 void target_recovery_init(struct lu_target *lut, svc_handler_t handler)
2053 struct obd_device *obd = lut->lut_obd;
2054 if (obd->obd_max_recoverable_clients == 0) {
2055 /** Update server last boot epoch */
2056 lut_boot_epoch_update(lut);
2060 CWARN("RECOVERY: service %s, %d recoverable clients, "
2061 "last_transno "LPU64"\n", obd->obd_name,
2062 obd->obd_max_recoverable_clients, obd->obd_last_committed);
2063 LASSERT(obd->obd_stopping == 0);
2064 obd->obd_next_recovery_transno = obd->obd_last_committed + 1;
2065 obd->obd_recovery_start = 0;
2066 obd->obd_recovery_end = 0;
2068 cfs_timer_init(&obd->obd_recovery_timer, target_recovery_expired, obd);
2069 target_start_recovery_thread(lut, handler);
2071 EXPORT_SYMBOL(target_recovery_init);
2073 #endif /* __KERNEL__ */
2075 static int target_process_req_flags(struct obd_device *obd,
2076 struct ptlrpc_request *req)
2078 struct obd_export *exp = req->rq_export;
2079 LASSERT(exp != NULL);
2080 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REQ_REPLAY_DONE) {
2081 /* client declares he's ready to replay locks */
2082 cfs_spin_lock(&exp->exp_lock);
2083 if (exp->exp_req_replay_needed) {
2084 exp->exp_req_replay_needed = 0;
2085 cfs_spin_unlock(&exp->exp_lock);
2087 LASSERT_ATOMIC_POS(&obd->obd_req_replay_clients);
2088 cfs_atomic_dec(&obd->obd_req_replay_clients);
2090 cfs_spin_unlock(&exp->exp_lock);
2093 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_LOCK_REPLAY_DONE) {
2094 /* client declares he's ready to complete recovery
2095 * so, we put the request on th final queue */
2096 cfs_spin_lock(&exp->exp_lock);
2097 if (exp->exp_lock_replay_needed) {
2098 exp->exp_lock_replay_needed = 0;
2099 cfs_spin_unlock(&exp->exp_lock);
2101 LASSERT_ATOMIC_POS(&obd->obd_lock_replay_clients);
2102 cfs_atomic_dec(&obd->obd_lock_replay_clients);
2104 cfs_spin_unlock(&exp->exp_lock);
2110 int target_queue_recovery_request(struct ptlrpc_request *req,
2111 struct obd_device *obd)
2115 __u64 transno = lustre_msg_get_transno(req->rq_reqmsg);
2118 if (obd->obd_recovery_data.trd_processing_task == cfs_curproc_pid()) {
2119 /* Processing the queue right now, don't re-add. */
2123 target_process_req_flags(obd, req);
2125 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_LOCK_REPLAY_DONE) {
2126 /* client declares he's ready to complete recovery
2127 * so, we put the request on th final queue */
2128 target_request_copy_get(req);
2129 DEBUG_REQ(D_HA, req, "queue final req");
2130 cfs_waitq_signal(&obd->obd_next_transno_waitq);
2131 cfs_spin_lock(&obd->obd_recovery_task_lock);
2132 if (obd->obd_recovering) {
2133 cfs_list_add_tail(&req->rq_list,
2134 &obd->obd_final_req_queue);
2136 cfs_spin_unlock(&obd->obd_recovery_task_lock);
2137 target_request_copy_put(req);
2138 RETURN(obd->obd_stopping ? -ENOTCONN : 1);
2140 cfs_spin_unlock(&obd->obd_recovery_task_lock);
2143 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REQ_REPLAY_DONE) {
2144 /* client declares he's ready to replay locks */
2145 target_request_copy_get(req);
2146 DEBUG_REQ(D_HA, req, "queue lock replay req");
2147 cfs_waitq_signal(&obd->obd_next_transno_waitq);
2148 cfs_spin_lock(&obd->obd_recovery_task_lock);
2149 LASSERT(obd->obd_recovering);
2150 /* usually due to recovery abort */
2151 if (!req->rq_export->exp_in_recovery) {
2152 cfs_spin_unlock(&obd->obd_recovery_task_lock);
2153 target_request_copy_put(req);
2156 LASSERT(req->rq_export->exp_lock_replay_needed);
2157 cfs_list_add_tail(&req->rq_list, &obd->obd_lock_replay_queue);
2158 cfs_spin_unlock(&obd->obd_recovery_task_lock);
2162 /* CAVEAT EMPTOR: The incoming request message has been swabbed
2163 * (i.e. buflens etc are in my own byte order), but type-dependent
2164 * buffers (eg mdt_body, ost_body etc) have NOT been swabbed. */
2167 CFS_INIT_LIST_HEAD(&req->rq_list);
2168 DEBUG_REQ(D_HA, req, "not queueing");
2172 /* If we're processing the queue, we want don't want to queue this
2175 * Also, if this request has a transno less than the one we're waiting
2176 * for, we should process it now. It could (and currently always will)
2177 * be an open request for a descriptor that was opened some time ago.
2179 * Also, a resent, replayed request that has already been
2180 * handled will pass through here and be processed immediately.
2182 CWARN("Next recovery transno: "LPU64", current: "LPU64", replaying\n",
2183 obd->obd_next_recovery_transno, transno);
2184 cfs_spin_lock(&obd->obd_recovery_task_lock);
2185 if (transno < obd->obd_next_recovery_transno) {
2186 /* Processing the queue right now, don't re-add. */
2187 LASSERT(cfs_list_empty(&req->rq_list));
2188 cfs_spin_unlock(&obd->obd_recovery_task_lock);
2191 cfs_spin_unlock(&obd->obd_recovery_task_lock);
2193 if (OBD_FAIL_CHECK(OBD_FAIL_TGT_REPLAY_DROP))
2196 target_request_copy_get(req);
2197 if (!req->rq_export->exp_in_recovery) {
2198 target_request_copy_put(req);
2201 LASSERT(req->rq_export->exp_req_replay_needed);
2203 if (target_exp_enqueue_req_replay(req)) {
2204 DEBUG_REQ(D_ERROR, req, "dropping resent queued req");
2205 target_request_copy_put(req);
2210 cfs_spin_lock(&obd->obd_recovery_task_lock);
2211 LASSERT(obd->obd_recovering);
2212 cfs_list_for_each(tmp, &obd->obd_req_replay_queue) {
2213 struct ptlrpc_request *reqiter =
2214 cfs_list_entry(tmp, struct ptlrpc_request, rq_list);
2216 if (lustre_msg_get_transno(reqiter->rq_reqmsg) > transno) {
2217 cfs_list_add_tail(&req->rq_list, &reqiter->rq_list);
2222 if (unlikely(lustre_msg_get_transno(reqiter->rq_reqmsg) ==
2224 DEBUG_REQ(D_ERROR, req, "dropping replay: transno "
2225 "has been claimed by another client");
2226 cfs_spin_unlock(&obd->obd_recovery_task_lock);
2227 target_exp_dequeue_req_replay(req);
2228 target_request_copy_put(req);
2234 cfs_list_add_tail(&req->rq_list, &obd->obd_req_replay_queue);
2236 obd->obd_requests_queued_for_recovery++;
2237 cfs_spin_unlock(&obd->obd_recovery_task_lock);
2238 cfs_waitq_signal(&obd->obd_next_transno_waitq);
2242 int target_handle_ping(struct ptlrpc_request *req)
2244 obd_ping(req->rq_export);
2245 return req_capsule_server_pack(&req->rq_pill);
2248 void target_committed_to_req(struct ptlrpc_request *req)
2250 struct obd_export *exp = req->rq_export;
2252 if (!exp->exp_obd->obd_no_transno && req->rq_repmsg != NULL)
2253 lustre_msg_set_last_committed(req->rq_repmsg,
2254 exp->exp_last_committed);
2256 DEBUG_REQ(D_IOCTL, req, "not sending last_committed update (%d/"
2257 "%d)", exp->exp_obd->obd_no_transno,
2258 req->rq_repmsg == NULL);
2260 CDEBUG(D_INFO, "last_committed "LPU64", transno "LPU64", xid "LPU64"\n",
2261 exp->exp_last_committed, req->rq_transno, req->rq_xid);
2263 EXPORT_SYMBOL(target_committed_to_req);
2265 #endif /* HAVE_SERVER_SUPPORT */
2268 * Packs current SLV and Limit into \a req.
2270 int target_pack_pool_reply(struct ptlrpc_request *req)
2272 struct obd_device *obd;
2276 * Check that we still have all structures alive as this may
2277 * be some late rpc in shutdown time.
2279 if (unlikely(!req->rq_export || !req->rq_export->exp_obd ||
2280 !exp_connect_lru_resize(req->rq_export))) {
2281 lustre_msg_set_slv(req->rq_repmsg, 0);
2282 lustre_msg_set_limit(req->rq_repmsg, 0);
2287 * OBD is alive here as export is alive, which we checked above.
2289 obd = req->rq_export->exp_obd;
2291 cfs_read_lock(&obd->obd_pool_lock);
2292 lustre_msg_set_slv(req->rq_repmsg, obd->obd_pool_slv);
2293 lustre_msg_set_limit(req->rq_repmsg, obd->obd_pool_limit);
2294 cfs_read_unlock(&obd->obd_pool_lock);
2299 int target_send_reply_msg(struct ptlrpc_request *req, int rc, int fail_id)
2301 if (OBD_FAIL_CHECK_ORSET(fail_id & ~OBD_FAIL_ONCE, OBD_FAIL_ONCE)) {
2302 DEBUG_REQ(D_ERROR, req, "dropping reply");
2307 DEBUG_REQ(D_NET, req, "processing error (%d)", rc);
2308 req->rq_status = rc;
2309 return (ptlrpc_send_error(req, 1));
2311 DEBUG_REQ(D_NET, req, "sending reply");
2314 return (ptlrpc_send_reply(req, PTLRPC_REPLY_MAYBE_DIFFICULT));
2317 void target_send_reply(struct ptlrpc_request *req, int rc, int fail_id)
2320 struct ptlrpc_reply_state *rs;
2321 struct obd_export *exp;
2322 struct ptlrpc_service *svc;
2325 if (req->rq_no_reply) {
2330 svc = req->rq_rqbd->rqbd_service;
2331 rs = req->rq_reply_state;
2332 if (rs == NULL || !rs->rs_difficult) {
2334 target_send_reply_msg (req, rc, fail_id);
2339 /* must be an export if locks saved */
2340 LASSERT (req->rq_export != NULL);
2341 /* req/reply consistent */
2342 LASSERT (rs->rs_service == svc);
2345 LASSERT (!rs->rs_scheduled);
2346 LASSERT (!rs->rs_scheduled_ever);
2347 LASSERT (!rs->rs_handled);
2348 LASSERT (!rs->rs_on_net);
2349 LASSERT (rs->rs_export == NULL);
2350 LASSERT (cfs_list_empty(&rs->rs_obd_list));
2351 LASSERT (cfs_list_empty(&rs->rs_exp_list));
2353 exp = class_export_get (req->rq_export);
2355 /* disable reply scheduling while I'm setting up */
2356 rs->rs_scheduled = 1;
2358 rs->rs_xid = req->rq_xid;
2359 rs->rs_transno = req->rq_transno;
2360 rs->rs_export = exp;
2361 rs->rs_opc = lustre_msg_get_opc(rs->rs_msg);
2363 cfs_spin_lock(&exp->exp_uncommitted_replies_lock);
2364 CDEBUG(D_NET, "rs transno = "LPU64", last committed = "LPU64"\n",
2365 rs->rs_transno, exp->exp_last_committed);
2366 if (rs->rs_transno > exp->exp_last_committed) {
2367 /* not committed already */
2368 cfs_list_add_tail(&rs->rs_obd_list,
2369 &exp->exp_uncommitted_replies);
2371 cfs_spin_unlock (&exp->exp_uncommitted_replies_lock);
2373 cfs_spin_lock(&exp->exp_lock);
2374 cfs_list_add_tail(&rs->rs_exp_list, &exp->exp_outstanding_replies);
2375 cfs_spin_unlock(&exp->exp_lock);
2377 netrc = target_send_reply_msg (req, rc, fail_id);
2379 cfs_spin_lock(&svc->srv_rs_lock);
2381 cfs_atomic_inc(&svc->srv_n_difficult_replies);
2384 /* error sending: reply is off the net. Also we need +1
2385 * reply ref until ptlrpc_handle_rs() is done
2386 * with the reply state (if the send was successful, there
2387 * would have been +1 ref for the net, which
2388 * reply_out_callback leaves alone) */
2390 ptlrpc_rs_addref(rs);
2393 cfs_spin_lock(&rs->rs_lock);
2394 if (rs->rs_transno <= exp->exp_last_committed ||
2395 (!rs->rs_on_net && !rs->rs_no_ack) ||
2396 cfs_list_empty(&rs->rs_exp_list) || /* completed already */
2397 cfs_list_empty(&rs->rs_obd_list)) {
2398 CDEBUG(D_HA, "Schedule reply immediately\n");
2399 ptlrpc_dispatch_difficult_reply(rs);
2401 cfs_list_add (&rs->rs_list, &svc->srv_active_replies);
2402 rs->rs_scheduled = 0; /* allow notifier to schedule */
2404 cfs_spin_unlock(&rs->rs_lock);
2405 cfs_spin_unlock(&svc->srv_rs_lock);
2409 int target_handle_qc_callback(struct ptlrpc_request *req)
2411 struct obd_quotactl *oqctl;
2412 struct client_obd *cli = &req->rq_export->exp_obd->u.cli;
2414 oqctl = req_capsule_client_get(&req->rq_pill, &RMF_OBD_QUOTACTL);
2415 if (oqctl == NULL) {
2416 CERROR("Can't unpack obd_quotactl\n");
2420 cli->cl_qchk_stat = oqctl->qc_stat;
2425 #ifdef HAVE_QUOTA_SUPPORT
2426 int target_handle_dqacq_callback(struct ptlrpc_request *req)
2429 struct obd_device *obd = req->rq_export->exp_obd;
2430 struct obd_device *master_obd = NULL, *lov_obd = NULL;
2431 struct obd_device_target *obt;
2432 struct lustre_quota_ctxt *qctxt;
2433 struct qunit_data *qdata = NULL;
2437 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_DROP_QUOTA_REQ))
2440 rc = req_capsule_server_pack(&req->rq_pill);
2442 CERROR("packing reply failed!: rc = %d\n", rc);
2446 LASSERT(req->rq_export);
2448 qdata = quota_get_qdata(req, QUOTA_REQUEST, QUOTA_EXPORT);
2449 if (IS_ERR(qdata)) {
2450 rc = PTR_ERR(qdata);
2451 CDEBUG(D_ERROR, "Can't unpack qunit_data(rc: %d)\n", rc);
2452 req->rq_status = rc;
2456 /* we use the observer */
2457 if (obd_pin_observer(obd, &lov_obd) ||
2458 obd_pin_observer(lov_obd, &master_obd)) {
2459 CERROR("Can't find the observer, it is recovering\n");
2460 req->rq_status = -EAGAIN;
2464 obt = &master_obd->u.obt;
2465 qctxt = &obt->obt_qctxt;
2467 if (!qctxt->lqc_setup || !qctxt->lqc_valid) {
2468 /* quota_type has not been processed yet, return EAGAIN
2469 * until we know whether or not quotas are supposed to
2471 CDEBUG(D_QUOTA, "quota_type not processed yet, return "
2473 req->rq_status = -EAGAIN;
2477 cfs_down_read(&obt->obt_rwsem);
2478 if (qctxt->lqc_lqs_hash == NULL) {
2479 cfs_up_read(&obt->obt_rwsem);
2480 /* quota_type has not been processed yet, return EAGAIN
2481 * until we know whether or not quotas are supposed to
2483 CDEBUG(D_QUOTA, "quota_ctxt is not ready yet, return "
2485 req->rq_status = -EAGAIN;
2489 LASSERT(qctxt->lqc_handler);
2490 rc = qctxt->lqc_handler(master_obd, qdata,
2491 lustre_msg_get_opc(req->rq_reqmsg));
2492 cfs_up_read(&obt->obt_rwsem);
2493 if (rc && rc != -EDQUOT)
2494 CDEBUG(rc == -EBUSY ? D_QUOTA : D_ERROR,
2495 "dqacq/dqrel failed! (rc:%d)\n", rc);
2496 req->rq_status = rc;
2498 rc = quota_copy_qdata(req, qdata, QUOTA_REPLY, QUOTA_EXPORT);
2500 CERROR("Can't pack qunit_data(rc: %d)\n", rc);
2504 /* Block the quota req. b=14840 */
2505 OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_BLOCK_QUOTA_REQ, obd_timeout);
2510 obd_unpin_observer(lov_obd);
2512 obd_unpin_observer(obd);
2514 rc = ptlrpc_reply(req);
2518 #endif /* !__KERNEL__ */
2520 #endif /* HAVE_QUOTA_SUPPORT */
2522 ldlm_mode_t lck_compat_array[] = {
2523 [LCK_EX] LCK_COMPAT_EX,
2524 [LCK_PW] LCK_COMPAT_PW,
2525 [LCK_PR] LCK_COMPAT_PR,
2526 [LCK_CW] LCK_COMPAT_CW,
2527 [LCK_CR] LCK_COMPAT_CR,
2528 [LCK_NL] LCK_COMPAT_NL,
2529 [LCK_GROUP] LCK_COMPAT_GROUP,
2530 [LCK_COS] LCK_COMPAT_COS,
2534 * Rather arbitrary mapping from LDLM error codes to errno values. This should
2535 * not escape to the user level.
2537 int ldlm_error2errno(ldlm_error_t error)
2545 case ELDLM_LOCK_CHANGED:
2548 case ELDLM_LOCK_ABORTED:
2551 case ELDLM_LOCK_REPLACED:
2554 case ELDLM_NO_LOCK_DATA:
2557 case ELDLM_NAMESPACE_EXISTS:
2560 case ELDLM_BAD_NAMESPACE:
2564 if (((int)error) < 0) /* cast to signed type */
2565 result = error; /* as ldlm_error_t can be unsigned */
2567 CERROR("Invalid DLM result code: %d\n", error);
2573 EXPORT_SYMBOL(ldlm_error2errno);
2576 * Dual to ldlm_error2errno(): maps errno values back to ldlm_error_t.
2578 ldlm_error_t ldlm_errno2error(int err_no)
2587 error = ELDLM_LOCK_CHANGED;
2590 error = ELDLM_LOCK_ABORTED;
2593 error = ELDLM_LOCK_REPLACED;
2596 error = ELDLM_NO_LOCK_DATA;
2599 error = ELDLM_NAMESPACE_EXISTS;
2602 error = ELDLM_BAD_NAMESPACE;
2609 EXPORT_SYMBOL(ldlm_errno2error);
2611 #if LUSTRE_TRACKS_LOCK_EXP_REFS
2612 void ldlm_dump_export_locks(struct obd_export *exp)
2614 cfs_spin_lock(&exp->exp_locks_list_guard);
2615 if (!cfs_list_empty(&exp->exp_locks_list)) {
2616 struct ldlm_lock *lock;
2618 CERROR("dumping locks for export %p,"
2619 "ignore if the unmount doesn't hang\n", exp);
2620 cfs_list_for_each_entry(lock, &exp->exp_locks_list, l_exp_refs_link)
2621 ldlm_lock_dump(D_ERROR, lock, 0);
2623 cfs_spin_unlock(&exp->exp_locks_list_guard);
2627 #ifdef HAVE_SERVER_SUPPORT
2628 static int target_bulk_timeout(void *data)
2631 /* We don't fail the connection here, because having the export
2632 * killed makes the (vital) call to commitrw very sad.
2637 static inline char *bulk2type(struct ptlrpc_bulk_desc *desc)
2639 return desc->bd_type == BULK_GET_SINK ? "GET" : "PUT";
2642 int target_bulk_io(struct obd_export *exp, struct ptlrpc_bulk_desc *desc,
2643 struct l_wait_info *lwi)
2645 struct ptlrpc_request *req = desc->bd_req;
2649 /* Check if there is eviction in progress, and if so, wait for
2651 if (unlikely(cfs_atomic_read(&exp->exp_obd->obd_evict_inprogress))) {
2652 *lwi = LWI_INTR(NULL, NULL);
2653 rc = l_wait_event(exp->exp_obd->obd_evict_inprogress_waitq,
2654 !cfs_atomic_read(&exp->exp_obd->
2655 obd_evict_inprogress),
2659 /* Check if client was evicted or tried to reconnect already */
2660 if (exp->exp_failed || exp->exp_abort_active_req) {
2663 if (desc->bd_type == BULK_PUT_SINK)
2664 rc = sptlrpc_svc_wrap_bulk(req, desc);
2666 rc = ptlrpc_start_bulk_transfer(desc);
2669 if (rc == 0 && OBD_FAIL_CHECK(OBD_FAIL_MDS_SENDPAGE)) {
2670 ptlrpc_abort_bulk(desc);
2671 } else if (rc == 0) {
2672 time_t start = cfs_time_current_sec();
2674 long timeoutl = req->rq_deadline - cfs_time_current_sec();
2675 cfs_duration_t timeout = timeoutl <= 0 ?
2676 CFS_TICK : cfs_time_seconds(timeoutl);
2677 *lwi = LWI_TIMEOUT_INTERVAL(timeout,
2678 cfs_time_seconds(1),
2679 target_bulk_timeout,
2681 rc = l_wait_event(desc->bd_waitq,
2682 !ptlrpc_server_bulk_active(desc) ||
2684 exp->exp_abort_active_req,
2686 LASSERT(rc == 0 || rc == -ETIMEDOUT);
2687 /* Wait again if we changed deadline */
2688 } while ((rc == -ETIMEDOUT) &&
2689 (req->rq_deadline > cfs_time_current_sec()));
2691 if (rc == -ETIMEDOUT) {
2692 DEBUG_REQ(D_ERROR, req,
2693 "timeout on bulk %s after %ld%+lds",
2695 req->rq_deadline - start,
2696 cfs_time_current_sec() -
2698 ptlrpc_abort_bulk(desc);
2699 } else if (exp->exp_failed) {
2700 DEBUG_REQ(D_ERROR, req, "Eviction on bulk %s",
2703 ptlrpc_abort_bulk(desc);
2704 } else if (exp->exp_abort_active_req) {
2705 DEBUG_REQ(D_ERROR, req, "Reconnect on bulk %s",
2707 /* we don't reply anyway */
2709 ptlrpc_abort_bulk(desc);
2710 } else if (!desc->bd_success ||
2711 desc->bd_nob_transferred != desc->bd_nob) {
2712 DEBUG_REQ(D_ERROR, req, "%s bulk %s %d(%d)",
2714 "truncated" : "network error on",
2716 desc->bd_nob_transferred,
2718 /* XXX should this be a different errno? */
2720 } else if (desc->bd_type == BULK_GET_SINK) {
2721 rc = sptlrpc_svc_unwrap_bulk(req, desc);
2724 DEBUG_REQ(D_ERROR, req, "bulk %s failed: rc %d",
2725 bulk2type(desc), rc);
2730 EXPORT_SYMBOL(target_bulk_io);
2732 #endif /* HAVE_SERVER_SUPPORT */