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.
33 * Copyright (c) 2011 Whamcloud, Inc.
36 * This file is part of Lustre, http://www.lustre.org/
37 * Lustre is a trademark of Sun Microsystems, Inc.
41 # define EXPORT_SYMTAB
43 #define DEBUG_SUBSYSTEM S_LDLM
46 # include <libcfs/libcfs.h>
48 # include <liblustre.h>
51 #include <lustre_mds.h>
52 #include <lustre_dlm.h>
53 #include <lustre_net.h>
54 #include <lustre_sec.h>
55 #include "ldlm_internal.h"
57 /* @priority: if non-zero, move the selected to the list head
58 * @create: if zero, only search in existed connections
60 static int import_set_conn(struct obd_import *imp, struct obd_uuid *uuid,
61 int priority, int create)
63 struct ptlrpc_connection *ptlrpc_conn;
64 struct obd_import_conn *imp_conn = NULL, *item;
68 if (!create && !priority) {
69 CDEBUG(D_HA, "Nothing to do\n");
73 ptlrpc_conn = ptlrpc_uuid_to_connection(uuid);
75 CDEBUG(D_HA, "can't find connection %s\n", uuid->uuid);
80 OBD_ALLOC(imp_conn, sizeof(*imp_conn));
82 GOTO(out_put, rc = -ENOMEM);
86 cfs_spin_lock(&imp->imp_lock);
87 cfs_list_for_each_entry(item, &imp->imp_conn_list, oic_item) {
88 if (obd_uuid_equals(uuid, &item->oic_uuid)) {
90 cfs_list_del(&item->oic_item);
91 cfs_list_add(&item->oic_item,
93 item->oic_last_attempt = 0;
95 CDEBUG(D_HA, "imp %p@%s: found existing conn %s%s\n",
96 imp, imp->imp_obd->obd_name, uuid->uuid,
97 (priority ? ", moved to head" : ""));
98 cfs_spin_unlock(&imp->imp_lock);
99 GOTO(out_free, rc = 0);
104 imp_conn->oic_conn = ptlrpc_conn;
105 imp_conn->oic_uuid = *uuid;
106 imp_conn->oic_last_attempt = 0;
108 cfs_list_add(&imp_conn->oic_item, &imp->imp_conn_list);
110 cfs_list_add_tail(&imp_conn->oic_item,
111 &imp->imp_conn_list);
112 CDEBUG(D_HA, "imp %p@%s: add connection %s at %s\n",
113 imp, imp->imp_obd->obd_name, uuid->uuid,
114 (priority ? "head" : "tail"));
116 cfs_spin_unlock(&imp->imp_lock);
117 GOTO(out_free, rc = -ENOENT);
120 cfs_spin_unlock(&imp->imp_lock);
124 OBD_FREE(imp_conn, sizeof(*imp_conn));
126 ptlrpc_connection_put(ptlrpc_conn);
130 int import_set_conn_priority(struct obd_import *imp, struct obd_uuid *uuid)
132 return import_set_conn(imp, uuid, 1, 0);
135 int client_import_add_conn(struct obd_import *imp, struct obd_uuid *uuid,
138 return import_set_conn(imp, uuid, priority, 1);
141 int client_import_del_conn(struct obd_import *imp, struct obd_uuid *uuid)
143 struct obd_import_conn *imp_conn;
144 struct obd_export *dlmexp;
148 cfs_spin_lock(&imp->imp_lock);
149 if (cfs_list_empty(&imp->imp_conn_list)) {
150 LASSERT(!imp->imp_connection);
154 cfs_list_for_each_entry(imp_conn, &imp->imp_conn_list, oic_item) {
155 if (!obd_uuid_equals(uuid, &imp_conn->oic_uuid))
157 LASSERT(imp_conn->oic_conn);
159 /* is current conn? */
160 if (imp_conn == imp->imp_conn_current) {
161 LASSERT(imp_conn->oic_conn == imp->imp_connection);
163 if (imp->imp_state != LUSTRE_IMP_CLOSED &&
164 imp->imp_state != LUSTRE_IMP_DISCON) {
165 CERROR("can't remove current connection\n");
166 GOTO(out, rc = -EBUSY);
169 ptlrpc_connection_put(imp->imp_connection);
170 imp->imp_connection = NULL;
172 dlmexp = class_conn2export(&imp->imp_dlm_handle);
173 if (dlmexp && dlmexp->exp_connection) {
174 LASSERT(dlmexp->exp_connection ==
176 ptlrpc_connection_put(dlmexp->exp_connection);
177 dlmexp->exp_connection = NULL;
181 cfs_list_del(&imp_conn->oic_item);
182 ptlrpc_connection_put(imp_conn->oic_conn);
183 OBD_FREE(imp_conn, sizeof(*imp_conn));
184 CDEBUG(D_HA, "imp %p@%s: remove connection %s\n",
185 imp, imp->imp_obd->obd_name, uuid->uuid);
190 cfs_spin_unlock(&imp->imp_lock);
192 CERROR("connection %s not found\n", uuid->uuid);
197 * Find conn uuid by peer nid. @peer is a server nid. This function is used
198 * to find a conn uuid of @imp which can reach @peer.
200 int client_import_find_conn(struct obd_import *imp, lnet_nid_t peer,
201 struct obd_uuid *uuid)
203 struct obd_import_conn *conn;
207 cfs_spin_lock(&imp->imp_lock);
208 cfs_list_for_each_entry(conn, &imp->imp_conn_list, oic_item) {
209 /* check if conn uuid does have this peer nid */
210 if (class_check_uuid(&conn->oic_uuid, peer)) {
211 *uuid = conn->oic_uuid;
216 cfs_spin_unlock(&imp->imp_lock);
219 EXPORT_SYMBOL(client_import_find_conn);
221 void client_destroy_import(struct obd_import *imp)
223 /* drop security policy instance after all rpc finished/aborted
224 * to let all busy contexts be released. */
225 class_import_get(imp);
226 class_destroy_import(imp);
227 sptlrpc_import_sec_put(imp);
228 class_import_put(imp);
230 EXPORT_SYMBOL(client_destroy_import);
232 /* configure an RPC client OBD device
237 * 3 - inactive-on-startup
239 int client_obd_setup(struct obd_device *obddev, struct lustre_cfg *lcfg)
241 struct client_obd *cli = &obddev->u.cli;
242 struct obd_import *imp;
243 struct obd_uuid server_uuid;
244 int rq_portal, rp_portal, connect_op;
245 char *name = obddev->obd_type->typ_name;
246 ldlm_ns_type_t ns_type = LDLM_NS_TYPE_UNKNOWN;
250 /* In a more perfect world, we would hang a ptlrpc_client off of
251 * obd_type and just use the values from there. */
252 if (!strcmp(name, LUSTRE_OSC_NAME)) {
253 rq_portal = OST_REQUEST_PORTAL;
254 rp_portal = OSC_REPLY_PORTAL;
255 connect_op = OST_CONNECT;
256 cli->cl_sp_me = LUSTRE_SP_CLI;
257 cli->cl_sp_to = LUSTRE_SP_OST;
258 ns_type = LDLM_NS_TYPE_OSC;
260 } else if (!strcmp(name, LUSTRE_MDC_NAME)) {
261 rq_portal = MDS_REQUEST_PORTAL;
262 rp_portal = MDC_REPLY_PORTAL;
263 connect_op = MDS_CONNECT;
264 cli->cl_sp_me = LUSTRE_SP_CLI;
265 cli->cl_sp_to = LUSTRE_SP_MDT;
266 ns_type = LDLM_NS_TYPE_MDC;
268 } else if (!strcmp(name, LUSTRE_MGC_NAME)) {
269 rq_portal = MGS_REQUEST_PORTAL;
270 rp_portal = MGC_REPLY_PORTAL;
271 connect_op = MGS_CONNECT;
272 cli->cl_sp_me = LUSTRE_SP_MGC;
273 cli->cl_sp_to = LUSTRE_SP_MGS;
274 cli->cl_flvr_mgc.sf_rpc = SPTLRPC_FLVR_INVALID;
275 ns_type = LDLM_NS_TYPE_MGC;
278 CERROR("unknown client OBD type \"%s\", can't setup\n",
283 if (LUSTRE_CFG_BUFLEN(lcfg, 1) < 1) {
284 CERROR("requires a TARGET UUID\n");
288 if (LUSTRE_CFG_BUFLEN(lcfg, 1) > 37) {
289 CERROR("client UUID must be less than 38 characters\n");
293 if (LUSTRE_CFG_BUFLEN(lcfg, 2) < 1) {
294 CERROR("setup requires a SERVER UUID\n");
298 if (LUSTRE_CFG_BUFLEN(lcfg, 2) > 37) {
299 CERROR("target UUID must be less than 38 characters\n");
303 cfs_init_rwsem(&cli->cl_sem);
304 cfs_sema_init(&cli->cl_mgc_sem, 1);
305 cli->cl_conn_count = 0;
306 memcpy(server_uuid.uuid, lustre_cfg_buf(lcfg, 2),
307 min_t(unsigned int, LUSTRE_CFG_BUFLEN(lcfg, 2),
308 sizeof(server_uuid)));
311 cli->cl_avail_grant = 0;
312 /* FIXME: should limit this for the sum of all cl_dirty_max */
313 cli->cl_dirty_max = OSC_MAX_DIRTY_DEFAULT * 1024 * 1024;
314 if (cli->cl_dirty_max >> CFS_PAGE_SHIFT > cfs_num_physpages / 8)
315 cli->cl_dirty_max = cfs_num_physpages << (CFS_PAGE_SHIFT - 3);
316 CFS_INIT_LIST_HEAD(&cli->cl_cache_waiters);
317 CFS_INIT_LIST_HEAD(&cli->cl_loi_ready_list);
318 CFS_INIT_LIST_HEAD(&cli->cl_loi_hp_ready_list);
319 CFS_INIT_LIST_HEAD(&cli->cl_loi_write_list);
320 CFS_INIT_LIST_HEAD(&cli->cl_loi_read_list);
321 client_obd_list_lock_init(&cli->cl_loi_list_lock);
322 cli->cl_r_in_flight = 0;
323 cli->cl_w_in_flight = 0;
325 cfs_spin_lock_init(&cli->cl_read_rpc_hist.oh_lock);
326 cfs_spin_lock_init(&cli->cl_write_rpc_hist.oh_lock);
327 cfs_spin_lock_init(&cli->cl_read_page_hist.oh_lock);
328 cfs_spin_lock_init(&cli->cl_write_page_hist.oh_lock);
329 cfs_spin_lock_init(&cli->cl_read_offset_hist.oh_lock);
330 cfs_spin_lock_init(&cli->cl_write_offset_hist.oh_lock);
331 cfs_waitq_init(&cli->cl_destroy_waitq);
332 cfs_atomic_set(&cli->cl_destroy_in_flight, 0);
333 #ifdef ENABLE_CHECKSUM
334 /* Turn on checksumming by default. */
335 cli->cl_checksum = 1;
337 * The supported checksum types will be worked out at connect time
338 * Set cl_chksum* to CRC32 for now to avoid returning screwed info
341 cli->cl_cksum_type = cli->cl_supp_cksum_types = OBD_CKSUM_CRC32;
343 cfs_atomic_set(&cli->cl_resends, OSC_DEFAULT_RESENDS);
345 /* This value may be changed at connect time in
346 ptlrpc_connect_interpret. */
347 cli->cl_max_pages_per_rpc = min((int)PTLRPC_MAX_BRW_PAGES,
348 (int)(1024 * 1024 >> CFS_PAGE_SHIFT));
350 if (!strcmp(name, LUSTRE_MDC_NAME)) {
351 cli->cl_max_rpcs_in_flight = MDC_MAX_RIF_DEFAULT;
352 } else if (cfs_num_physpages >> (20 - CFS_PAGE_SHIFT) <= 128 /* MB */) {
353 cli->cl_max_rpcs_in_flight = 2;
354 } else if (cfs_num_physpages >> (20 - CFS_PAGE_SHIFT) <= 256 /* MB */) {
355 cli->cl_max_rpcs_in_flight = 3;
356 } else if (cfs_num_physpages >> (20 - CFS_PAGE_SHIFT) <= 512 /* MB */) {
357 cli->cl_max_rpcs_in_flight = 4;
359 cli->cl_max_rpcs_in_flight = OSC_MAX_RIF_DEFAULT;
364 CERROR("ldlm_get_ref failed: %d\n", rc);
368 ptlrpc_init_client(rq_portal, rp_portal, name,
369 &obddev->obd_ldlm_client);
371 imp = class_new_import(obddev);
373 GOTO(err_ldlm, rc = -ENOENT);
374 imp->imp_client = &obddev->obd_ldlm_client;
375 imp->imp_connect_op = connect_op;
376 CFS_INIT_LIST_HEAD(&imp->imp_pinger_chain);
377 memcpy(cli->cl_target_uuid.uuid, lustre_cfg_buf(lcfg, 1),
378 LUSTRE_CFG_BUFLEN(lcfg, 1));
379 class_import_put(imp);
381 rc = client_import_add_conn(imp, &server_uuid, 1);
383 CERROR("can't add initial connection\n");
384 GOTO(err_import, rc);
387 cli->cl_import = imp;
388 /* cli->cl_max_mds_{easize,cookiesize} updated by mdc_init_ea_size() */
389 cli->cl_max_mds_easize = sizeof(struct lov_mds_md_v3);
390 cli->cl_max_mds_cookiesize = sizeof(struct llog_cookie);
392 if (LUSTRE_CFG_BUFLEN(lcfg, 3) > 0) {
393 if (!strcmp(lustre_cfg_string(lcfg, 3), "inactive")) {
394 CDEBUG(D_HA, "marking %s %s->%s as inactive\n",
395 name, obddev->obd_name,
396 cli->cl_target_uuid.uuid);
397 cfs_spin_lock(&imp->imp_lock);
398 imp->imp_deactive = 1;
399 cfs_spin_unlock(&imp->imp_lock);
403 obddev->obd_namespace = ldlm_namespace_new(obddev, obddev->obd_name,
404 LDLM_NAMESPACE_CLIENT,
405 LDLM_NAMESPACE_GREEDY,
407 if (obddev->obd_namespace == NULL) {
408 CERROR("Unable to create client namespace - %s\n",
410 GOTO(err_import, rc = -ENOMEM);
413 cli->cl_qchk_stat = CL_NOT_QUOTACHECKED;
418 class_destroy_import(imp);
426 int client_obd_cleanup(struct obd_device *obddev)
430 ldlm_namespace_free_post(obddev->obd_namespace);
431 obddev->obd_namespace = NULL;
433 LASSERT(obddev->u.cli.cl_import == NULL);
439 /* ->o_connect() method for client side (OSC and MDC and MGC) */
440 int client_connect_import(const struct lu_env *env,
441 struct obd_export **exp,
442 struct obd_device *obd, struct obd_uuid *cluuid,
443 struct obd_connect_data *data, void *localdata)
445 struct client_obd *cli = &obd->u.cli;
446 struct obd_import *imp = cli->cl_import;
447 struct obd_connect_data *ocd;
448 struct lustre_handle conn = { 0 };
453 cfs_down_write(&cli->cl_sem);
454 if (cli->cl_conn_count > 0 )
455 GOTO(out_sem, rc = -EALREADY);
457 rc = class_connect(&conn, obd, cluuid);
461 cli->cl_conn_count++;
462 *exp = class_conn2export(&conn);
464 LASSERT(obd->obd_namespace);
466 imp->imp_dlm_handle = conn;
467 rc = ptlrpc_init_import(imp);
471 ocd = &imp->imp_connect_data;
474 imp->imp_connect_flags_orig = data->ocd_connect_flags;
477 rc = ptlrpc_connect_import(imp, NULL);
479 LASSERT (imp->imp_state == LUSTRE_IMP_DISCON);
482 LASSERT((*exp)->exp_connection);
485 LASSERTF((ocd->ocd_connect_flags & data->ocd_connect_flags) ==
486 ocd->ocd_connect_flags, "old "LPX64", new "LPX64"\n",
487 data->ocd_connect_flags, ocd->ocd_connect_flags);
488 data->ocd_connect_flags = ocd->ocd_connect_flags;
491 ptlrpc_pinger_add_import(imp);
497 cli->cl_conn_count--;
498 class_disconnect(*exp);
502 cfs_up_write(&cli->cl_sem);
507 int client_disconnect_export(struct obd_export *exp)
509 struct obd_device *obd = class_exp2obd(exp);
510 struct client_obd *cli;
511 struct obd_import *imp;
516 CERROR("invalid export for disconnect: exp %p cookie "LPX64"\n",
517 exp, exp ? exp->exp_handle.h_cookie : -1);
522 imp = cli->cl_import;
524 cfs_down_write(&cli->cl_sem);
525 CDEBUG(D_INFO, "disconnect %s - %d\n", obd->obd_name,
528 if (!cli->cl_conn_count) {
529 CERROR("disconnecting disconnected device (%s)\n",
531 GOTO(out_disconnect, rc = -EINVAL);
534 cli->cl_conn_count--;
535 if (cli->cl_conn_count)
536 GOTO(out_disconnect, rc = 0);
538 /* Mark import deactivated now, so we don't try to reconnect if any
539 * of the cleanup RPCs fails (e.g. ldlm cancel, etc). We don't
540 * fully deactivate the import, or that would drop all requests. */
541 cfs_spin_lock(&imp->imp_lock);
542 imp->imp_deactive = 1;
543 cfs_spin_unlock(&imp->imp_lock);
545 /* Some non-replayable imports (MDS's OSCs) are pinged, so just
546 * delete it regardless. (It's safe to delete an import that was
548 (void)ptlrpc_pinger_del_import(imp);
550 if (obd->obd_namespace != NULL) {
551 /* obd_force == local only */
552 ldlm_cli_cancel_unused(obd->obd_namespace, NULL,
553 obd->obd_force ? LCF_LOCAL : 0, NULL);
554 ldlm_namespace_free_prior(obd->obd_namespace, imp, obd->obd_force);
558 * there's no need to hold sem during disconnecting an import,
559 * and actually it may cause deadlock in gss.
561 cfs_up_write(&cli->cl_sem);
562 rc = ptlrpc_disconnect_import(imp, 0);
563 cfs_down_write(&cli->cl_sem);
565 ptlrpc_invalidate_import(imp);
570 /* use server style - class_disconnect should be always called for
572 err = class_disconnect(exp);
576 cfs_up_write(&cli->cl_sem);
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: already connected client %s (at %s) "
661 "with handle "LPX64". Rejecting client "
662 "with the same UUID trying to reconnect "
663 "with handle "LPX64"\n", target->obd_name,
664 obd_uuid2str(&exp->exp_client_uuid),
665 obd_export_nid2str(exp),
666 hdl->cookie, conn->cookie);
667 memset(conn, 0, sizeof *conn);
668 /* target_handle_connect() treats EALREADY and
669 * -EALREADY differently. -EALREADY is an error
670 * (same UUID, different handle). */
675 conn->cookie = exp->exp_handle.h_cookie;
676 CDEBUG(D_HA, "connect export for UUID '%s' at %p, cookie "LPX64"\n",
677 cluuid->uuid, exp, conn->cookie);
681 void target_client_add_cb(struct obd_device *obd, __u64 transno, void *cb_data,
684 struct obd_export *exp = cb_data;
686 CDEBUG(D_RPCTRACE, "%s: committing for initial connect of %s\n",
687 obd->obd_name, exp->exp_client_uuid.uuid);
689 cfs_spin_lock(&exp->exp_lock);
690 exp->exp_need_sync = 0;
691 cfs_spin_unlock(&exp->exp_lock);
692 class_export_cb_put(exp);
694 EXPORT_SYMBOL(target_client_add_cb);
697 target_start_and_reset_recovery_timer(struct obd_device *obd,
698 struct ptlrpc_request *req,
701 int target_handle_connect(struct ptlrpc_request *req)
703 struct obd_device *target = NULL, *targref = NULL;
704 struct obd_export *export = NULL;
705 struct obd_import *revimp;
706 struct lustre_handle conn;
707 struct lustre_handle *tmp;
708 struct obd_uuid tgtuuid;
709 struct obd_uuid cluuid;
710 struct obd_uuid remote_uuid;
716 struct obd_connect_data *data, *tmpdata;
718 lnet_nid_t *client_nid = NULL;
721 OBD_RACE(OBD_FAIL_TGT_CONN_RACE);
723 str = req_capsule_client_get(&req->rq_pill, &RMF_TGTUUID);
725 DEBUG_REQ(D_ERROR, req, "bad target UUID for connect");
726 GOTO(out, rc = -EINVAL);
729 obd_str2uuid(&tgtuuid, str);
730 target = class_uuid2obd(&tgtuuid);
732 target = class_name2obd(str);
735 deuuidify(str, NULL, &target_start, &target_len);
736 LCONSOLE_ERROR_MSG(0x137, "UUID '%s' is not available for "
737 "connect (no target)\n", str);
738 GOTO(out, rc = -ENODEV);
741 cfs_spin_lock(&target->obd_dev_lock);
742 if (target->obd_stopping || !target->obd_set_up) {
743 cfs_spin_unlock(&target->obd_dev_lock);
745 deuuidify(str, NULL, &target_start, &target_len);
746 LCONSOLE_ERROR_MSG(0x137, "%.*s: Not available for connect "
747 "from %s (%s)\n", target_len, target_start,
748 libcfs_nid2str(req->rq_peer.nid),
749 (target->obd_stopping ?
750 "stopping" : "not set up"));
751 GOTO(out, rc = -ENODEV);
754 if (target->obd_no_conn) {
755 cfs_spin_unlock(&target->obd_dev_lock);
757 LCONSOLE_WARN("%s: Temporarily refusing client connection "
758 "from %s\n", target->obd_name,
759 libcfs_nid2str(req->rq_peer.nid));
760 GOTO(out, rc = -EAGAIN);
763 /* Make sure the target isn't cleaned up while we're here. Yes,
764 there's still a race between the above check and our incref here.
765 Really, class_uuid2obd should take the ref. */
766 targref = class_incref(target, __FUNCTION__, cfs_current());
768 target->obd_conn_inprogress++;
769 cfs_spin_unlock(&target->obd_dev_lock);
771 str = req_capsule_client_get(&req->rq_pill, &RMF_CLUUID);
773 DEBUG_REQ(D_ERROR, req, "bad client UUID for connect");
774 GOTO(out, rc = -EINVAL);
777 obd_str2uuid(&cluuid, str);
779 /* XXX extract a nettype and format accordingly */
780 switch (sizeof(lnet_nid_t)) {
781 /* NB the casts only avoid compiler warnings */
783 snprintf(remote_uuid.uuid, sizeof remote_uuid,
784 "NET_"LPX64"_UUID", (__u64)req->rq_peer.nid);
787 snprintf(remote_uuid.uuid, sizeof remote_uuid,
788 "NET_%x_UUID", (__u32)req->rq_peer.nid);
794 tmp = req_capsule_client_get(&req->rq_pill, &RMF_CONN);
796 GOTO(out, rc = -EPROTO);
800 size = req_capsule_get_size(&req->rq_pill, &RMF_CONNECT_DATA,
802 data = req_capsule_client_get(&req->rq_pill, &RMF_CONNECT_DATA);
804 GOTO(out, rc = -EPROTO);
806 rc = req_capsule_server_pack(&req->rq_pill);
810 if (lustre_msg_get_op_flags(req->rq_reqmsg) & MSG_CONNECT_LIBCLIENT) {
812 DEBUG_REQ(D_WARNING, req, "Refusing old (unversioned) "
813 "libclient connection attempt");
814 GOTO(out, rc = -EPROTO);
815 } else if (data->ocd_version < LUSTRE_VERSION_CODE -
816 LUSTRE_VERSION_ALLOWED_OFFSET ||
817 data->ocd_version > LUSTRE_VERSION_CODE +
818 LUSTRE_VERSION_ALLOWED_OFFSET) {
819 DEBUG_REQ(D_WARNING, req, "Refusing %s (%d.%d.%d.%d) "
820 "libclient connection attempt",
821 data->ocd_version < LUSTRE_VERSION_CODE ?
823 OBD_OCD_VERSION_MAJOR(data->ocd_version),
824 OBD_OCD_VERSION_MINOR(data->ocd_version),
825 OBD_OCD_VERSION_PATCH(data->ocd_version),
826 OBD_OCD_VERSION_FIX(data->ocd_version));
827 data = req_capsule_server_sized_get(&req->rq_pill,
829 offsetof(typeof(*data), ocd_version) +
830 sizeof(data->ocd_version));
832 data->ocd_connect_flags = OBD_CONNECT_VERSION;
833 data->ocd_version = LUSTRE_VERSION_CODE;
835 GOTO(out, rc = -EPROTO);
839 if ((lustre_msg_get_op_flags(req->rq_reqmsg) & MSG_CONNECT_INITIAL) &&
840 (data->ocd_connect_flags & OBD_CONNECT_MDS))
843 /* lctl gets a backstage, all-access pass. */
844 if (obd_uuid_equals(&cluuid, &target->obd_uuid))
845 goto dont_check_exports;
847 export = cfs_hash_lookup(target->obd_uuid_hash, &cluuid);
851 /* we've found an export in the hash */
852 if (export->exp_connecting) { /* bug 9635, et. al. */
853 LCONSOLE_WARN("%s: Export %p already connecting from %s\n",
854 export->exp_obd->obd_name, export,
855 libcfs_nid2str(req->rq_peer.nid));
856 class_export_put(export);
859 } else if (mds_conn && export->exp_connection) {
860 if (req->rq_peer.nid != export->exp_connection->c_peer.nid)
861 /* mds reconnected after failover */
862 LCONSOLE_WARN("%s: Received MDS connection from "
863 "%s, removing former export from %s\n",
864 target->obd_name, libcfs_nid2str(req->rq_peer.nid),
865 libcfs_nid2str(export->exp_connection->c_peer.nid));
867 /* new mds connection from the same nid */
868 LCONSOLE_WARN("%s: Received new MDS connection from "
869 "%s, removing former export from same NID\n",
870 target->obd_name, libcfs_nid2str(req->rq_peer.nid));
871 class_fail_export(export);
872 class_export_put(export);
875 } else if (export->exp_connection != NULL &&
876 req->rq_peer.nid != export->exp_connection->c_peer.nid &&
877 (lustre_msg_get_op_flags(req->rq_reqmsg) &
878 MSG_CONNECT_INITIAL)) {
879 /* in mds failover we have static uuid but nid can be
881 LCONSOLE_WARN("%s: Client %s seen on new nid %s when "
882 "existing nid %s is already connected\n",
883 target->obd_name, cluuid.uuid,
884 libcfs_nid2str(req->rq_peer.nid),
886 export->exp_connection->c_peer.nid));
888 class_export_put(export);
891 cfs_spin_lock(&export->exp_lock);
892 export->exp_connecting = 1;
893 cfs_spin_unlock(&export->exp_lock);
894 LASSERT(export->exp_obd == target);
896 rc = target_handle_reconnect(&conn, export, &cluuid);
899 /* If we found an export, we already unlocked. */
902 OBD_FAIL_TIMEOUT(OBD_FAIL_TGT_DELAY_CONNECT, 2 * obd_timeout);
903 } else if (req->rq_export == NULL &&
904 cfs_atomic_read(&export->exp_rpc_count) > 0) {
905 LCONSOLE_WARN("%s: Client %s (at %s) refused connection, "
906 "still busy with %d references\n",
907 target->obd_name, cluuid.uuid,
908 libcfs_nid2str(req->rq_peer.nid),
909 cfs_atomic_read(&export->exp_refcount));
910 GOTO(out, rc = -EBUSY);
911 } else if (req->rq_export != NULL &&
912 (cfs_atomic_read(&export->exp_rpc_count) > 1)) {
913 /* the current connect rpc has increased exp_rpc_count */
914 LCONSOLE_WARN("%s: Client %s (at %s) refused reconnection, "
915 "still busy with %d active RPCs\n",
916 target->obd_name, cluuid.uuid,
917 libcfs_nid2str(req->rq_peer.nid),
918 cfs_atomic_read(&export->exp_rpc_count) - 1);
919 cfs_spin_lock(&export->exp_lock);
920 if (req->rq_export->exp_conn_cnt <
921 lustre_msg_get_conn_cnt(req->rq_reqmsg))
922 /* try to abort active requests */
923 req->rq_export->exp_abort_active_req = 1;
924 cfs_spin_unlock(&export->exp_lock);
925 GOTO(out, rc = -EBUSY);
926 } else if (lustre_msg_get_conn_cnt(req->rq_reqmsg) == 1) {
927 if (!strstr(cluuid.uuid, "mdt"))
928 LCONSOLE_WARN("%s: Rejecting reconnect from the "
929 "known client %s (at %s) because it "
930 "is indicating it is a new client",
931 target->obd_name, cluuid.uuid,
932 libcfs_nid2str(req->rq_peer.nid));
933 GOTO(out, rc = -EALREADY);
935 OBD_FAIL_TIMEOUT(OBD_FAIL_TGT_DELAY_RECONNECT, 2 * obd_timeout);
942 CWARN("%s: connection from %s@%s %st"LPU64" exp %p cur %ld last %ld\n",
943 target->obd_name, cluuid.uuid, libcfs_nid2str(req->rq_peer.nid),
944 target->obd_recovering ? "recovering/" : "", data->ocd_transno,
945 export, (long)cfs_time_current_sec(),
946 export ? (long)export->exp_last_request_time : 0);
948 /* If this is the first time a client connects,
949 * reset the recovery timer */
950 if (rc == 0 && target->obd_recovering)
951 target_start_and_reset_recovery_timer(target, req, !export);
953 /* We want to handle EALREADY but *not* -EALREADY from
954 * target_handle_reconnect(), return reconnection state in a flag */
955 if (rc == EALREADY) {
956 lustre_msg_add_op_flags(req->rq_repmsg, MSG_CONNECT_RECONNECT);
962 /* Tell the client if we support replayable requests */
963 if (target->obd_replayable)
964 lustre_msg_add_op_flags(req->rq_repmsg, MSG_CONNECT_REPLAYABLE);
965 client_nid = &req->rq_peer.nid;
967 if (export == NULL) {
968 if (target->obd_recovering) {
971 t = cfs_timer_deadline(&target->obd_recovery_timer);
972 t = cfs_time_sub(t, cfs_time_current());
973 t = cfs_duration_sec(t);
974 LCONSOLE_WARN("%s: Denying connection for new client "
975 "%s (at %s), waiting for %d clients in "
976 "recovery for %d:%.02d\n",
978 libcfs_nid2str(req->rq_peer.nid),
980 cfs_atomic_read(&target-> \
981 obd_lock_replay_clients),
982 (int)t / 60, (int)t % 60);
986 rc = obd_connect(req->rq_svc_thread->t_env,
987 &export, target, &cluuid, data,
990 conn.cookie = export->exp_handle.h_cookie;
993 rc = obd_reconnect(req->rq_svc_thread->t_env,
994 export, target, &cluuid, data, client_nid);
998 /* Return only the parts of obd_connect_data that we understand, so the
999 * client knows that we don't understand the rest. */
1001 tmpsize = req_capsule_get_size(&req->rq_pill, &RMF_CONNECT_DATA,
1003 tmpdata = req_capsule_server_get(&req->rq_pill,
1005 /* Don't use struct assignment here, because the client reply
1006 * buffer may be smaller/larger than the local struct
1007 * obd_connect_data. */
1008 memcpy(tmpdata, data, min(tmpsize, size));
1011 /* If all else goes well, this is our RPC return code. */
1014 lustre_msg_set_handle(req->rq_repmsg, &conn);
1016 /* If the client and the server are the same node, we will already
1017 * have an export that really points to the client's DLM export,
1018 * because we have a shared handles table.
1020 * XXX this will go away when shaver stops sending the "connect" handle
1021 * in the real "remote handle" field of the request --phik 24 Apr 2003
1023 if (req->rq_export != NULL)
1024 class_export_put(req->rq_export);
1026 /* request takes one export refcount */
1027 req->rq_export = class_export_get(export);
1029 cfs_spin_lock(&export->exp_lock);
1030 if (export->exp_conn_cnt >= lustre_msg_get_conn_cnt(req->rq_reqmsg)) {
1031 cfs_spin_unlock(&export->exp_lock);
1032 CDEBUG(D_RPCTRACE, "%s: %s already connected at higher "
1033 "conn_cnt: %d > %d\n",
1034 cluuid.uuid, libcfs_nid2str(req->rq_peer.nid),
1035 export->exp_conn_cnt,
1036 lustre_msg_get_conn_cnt(req->rq_reqmsg));
1038 GOTO(out, rc = -EALREADY);
1040 LASSERT(lustre_msg_get_conn_cnt(req->rq_reqmsg) > 0);
1041 export->exp_conn_cnt = lustre_msg_get_conn_cnt(req->rq_reqmsg);
1042 export->exp_abort_active_req = 0;
1044 /* request from liblustre? Don't evict it for not pinging. */
1045 if (lustre_msg_get_op_flags(req->rq_reqmsg) & MSG_CONNECT_LIBCLIENT) {
1046 export->exp_libclient = 1;
1047 cfs_spin_unlock(&export->exp_lock);
1049 cfs_spin_lock(&target->obd_dev_lock);
1050 cfs_list_del_init(&export->exp_obd_chain_timed);
1051 cfs_spin_unlock(&target->obd_dev_lock);
1053 cfs_spin_unlock(&export->exp_lock);
1056 if (export->exp_connection != NULL) {
1057 /* Check to see if connection came from another NID */
1058 if ((export->exp_connection->c_peer.nid != req->rq_peer.nid) &&
1059 !cfs_hlist_unhashed(&export->exp_nid_hash))
1060 cfs_hash_del(export->exp_obd->obd_nid_hash,
1061 &export->exp_connection->c_peer.nid,
1062 &export->exp_nid_hash);
1064 ptlrpc_connection_put(export->exp_connection);
1067 export->exp_connection = ptlrpc_connection_get(req->rq_peer,
1070 if (cfs_hlist_unhashed(&export->exp_nid_hash)) {
1071 cfs_hash_add(export->exp_obd->obd_nid_hash,
1072 &export->exp_connection->c_peer.nid,
1073 &export->exp_nid_hash);
1076 class_disconnect->class_export_recovery_cleanup() race
1078 cfs_spin_lock(&target->obd_recovery_task_lock);
1079 if (target->obd_recovering && !export->exp_in_recovery &&
1080 !export->exp_disconnected) {
1081 cfs_spin_lock(&export->exp_lock);
1082 /* possible race with class_disconnect_stale_exports,
1083 * export may be already in the eviction process */
1084 if (export->exp_failed) {
1085 cfs_spin_unlock(&export->exp_lock);
1086 GOTO(out, rc = -ENODEV);
1088 export->exp_in_recovery = 1;
1089 export->exp_req_replay_needed = 1;
1090 export->exp_lock_replay_needed = 1;
1091 cfs_spin_unlock(&export->exp_lock);
1092 if ((lustre_msg_get_op_flags(req->rq_reqmsg) & MSG_CONNECT_TRANSNO)
1093 && (data->ocd_transno == 0))
1094 CWARN("Connect with zero transno!\n");
1096 if ((lustre_msg_get_op_flags(req->rq_reqmsg) & MSG_CONNECT_TRANSNO)
1097 && data->ocd_transno < target->obd_next_recovery_transno &&
1098 data->ocd_transno > target->obd_last_committed)
1099 target->obd_next_recovery_transno = data->ocd_transno;
1100 target->obd_connected_clients++;
1101 cfs_atomic_inc(&target->obd_req_replay_clients);
1102 cfs_atomic_inc(&target->obd_lock_replay_clients);
1103 if (target->obd_connected_clients ==
1104 target->obd_max_recoverable_clients)
1105 cfs_waitq_signal(&target->obd_next_transno_waitq);
1107 cfs_spin_unlock(&target->obd_recovery_task_lock);
1109 /* Tell the client we're in recovery, when client is involved in it. */
1110 if (target->obd_recovering)
1111 lustre_msg_add_op_flags(req->rq_repmsg, MSG_CONNECT_RECOVERING);
1113 tmp = req_capsule_client_get(&req->rq_pill, &RMF_CONN);
1116 if (export->exp_imp_reverse != NULL) {
1117 /* destroyed import can be still referenced in ctxt */
1118 obd_set_info_async(export, sizeof(KEY_REVIMP_UPD),
1119 KEY_REVIMP_UPD, 0, NULL, NULL);
1121 client_destroy_import(export->exp_imp_reverse);
1124 /* for the rest part, we return -ENOTCONN in case of errors
1125 * in order to let client initialize connection again.
1127 revimp = export->exp_imp_reverse = class_new_import(target);
1129 CERROR("fail to alloc new reverse import.\n");
1130 GOTO(out, rc = -ENOTCONN);
1133 revimp->imp_connection = ptlrpc_connection_addref(export->exp_connection);
1134 revimp->imp_client = &export->exp_obd->obd_ldlm_client;
1135 revimp->imp_remote_handle = conn;
1136 revimp->imp_dlm_fake = 1;
1137 revimp->imp_state = LUSTRE_IMP_FULL;
1139 /* unknown versions will be caught in
1140 * ptlrpc_handle_server_req_in->lustre_unpack_msg() */
1141 revimp->imp_msg_magic = req->rq_reqmsg->lm_magic;
1143 if ((export->exp_connect_flags & OBD_CONNECT_AT) &&
1144 (revimp->imp_msg_magic != LUSTRE_MSG_MAGIC_V1))
1145 revimp->imp_msghdr_flags |= MSGHDR_AT_SUPPORT;
1147 revimp->imp_msghdr_flags &= ~MSGHDR_AT_SUPPORT;
1149 if ((export->exp_connect_flags & OBD_CONNECT_FULL20) &&
1150 (revimp->imp_msg_magic != LUSTRE_MSG_MAGIC_V1))
1151 revimp->imp_msghdr_flags |= MSGHDR_CKSUM_INCOMPAT18;
1153 revimp->imp_msghdr_flags &= ~MSGHDR_CKSUM_INCOMPAT18;
1155 rc = sptlrpc_import_sec_adapt(revimp, req->rq_svc_ctx, &req->rq_flvr);
1157 CERROR("Failed to get sec for reverse import: %d\n", rc);
1158 export->exp_imp_reverse = NULL;
1159 class_destroy_import(revimp);
1162 class_import_put(revimp);
1165 cfs_spin_lock(&export->exp_lock);
1166 export->exp_connecting = 0;
1167 cfs_spin_unlock(&export->exp_lock);
1169 class_export_put(export);
1172 cfs_spin_lock(&target->obd_dev_lock);
1173 target->obd_conn_inprogress--;
1174 cfs_spin_unlock(&target->obd_dev_lock);
1176 class_decref(targref, __FUNCTION__, cfs_current());
1179 req->rq_status = rc;
1183 int target_handle_disconnect(struct ptlrpc_request *req)
1188 rc = req_capsule_server_pack(&req->rq_pill);
1192 /* keep the rq_export around so we can send the reply */
1193 req->rq_status = obd_disconnect(class_export_get(req->rq_export));
1198 void target_destroy_export(struct obd_export *exp)
1200 /* exports created from last_rcvd data, and "fake"
1201 exports created by lctl don't have an import */
1202 if (exp->exp_imp_reverse != NULL)
1203 client_destroy_import(exp->exp_imp_reverse);
1205 LASSERT_ATOMIC_ZERO(&exp->exp_locks_count);
1206 LASSERT_ATOMIC_ZERO(&exp->exp_rpc_count);
1207 LASSERT_ATOMIC_ZERO(&exp->exp_cb_count);
1208 LASSERT_ATOMIC_ZERO(&exp->exp_replay_count);
1212 * Recovery functions
1214 static void target_request_copy_get(struct ptlrpc_request *req)
1216 class_export_rpc_get(req->rq_export);
1217 LASSERT(cfs_list_empty(&req->rq_list));
1218 CFS_INIT_LIST_HEAD(&req->rq_replay_list);
1220 /* increase refcount to keep request in queue */
1221 cfs_atomic_inc(&req->rq_refcount);
1222 /** let export know it has replays to be handled */
1223 cfs_atomic_inc(&req->rq_export->exp_replay_count);
1226 static void target_request_copy_put(struct ptlrpc_request *req)
1228 LASSERT(cfs_list_empty(&req->rq_replay_list));
1229 LASSERT_ATOMIC_POS(&req->rq_export->exp_replay_count);
1231 cfs_atomic_dec(&req->rq_export->exp_replay_count);
1232 class_export_rpc_put(req->rq_export);
1233 ptlrpc_server_drop_request(req);
1236 static int target_exp_enqueue_req_replay(struct ptlrpc_request *req)
1238 __u64 transno = lustre_msg_get_transno(req->rq_reqmsg);
1239 struct obd_export *exp = req->rq_export;
1240 struct ptlrpc_request *reqiter;
1245 cfs_spin_lock(&exp->exp_lock);
1246 cfs_list_for_each_entry(reqiter, &exp->exp_req_replay_queue,
1248 if (lustre_msg_get_transno(reqiter->rq_reqmsg) == transno) {
1255 /* we expect it with RESENT and REPLAY flags */
1256 if ((lustre_msg_get_flags(req->rq_reqmsg) &
1257 (MSG_RESENT | MSG_REPLAY)) != (MSG_RESENT | MSG_REPLAY))
1258 CERROR("invalid flags %x of resent replay\n",
1259 lustre_msg_get_flags(req->rq_reqmsg));
1261 cfs_list_add_tail(&req->rq_replay_list,
1262 &exp->exp_req_replay_queue);
1265 cfs_spin_unlock(&exp->exp_lock);
1269 static void target_exp_dequeue_req_replay(struct ptlrpc_request *req)
1271 LASSERT(!cfs_list_empty(&req->rq_replay_list));
1272 LASSERT(req->rq_export);
1274 cfs_spin_lock(&req->rq_export->exp_lock);
1275 cfs_list_del_init(&req->rq_replay_list);
1276 cfs_spin_unlock(&req->rq_export->exp_lock);
1280 static void target_finish_recovery(struct obd_device *obd)
1283 LCONSOLE_INFO("%s: sending delayed replies to recovered clients\n",
1286 ldlm_reprocess_all_ns(obd->obd_namespace);
1287 cfs_spin_lock(&obd->obd_recovery_task_lock);
1288 if (!cfs_list_empty(&obd->obd_req_replay_queue) ||
1289 !cfs_list_empty(&obd->obd_lock_replay_queue) ||
1290 !cfs_list_empty(&obd->obd_final_req_queue)) {
1291 CERROR("%s: Recovery queues ( %s%s%s) are not empty\n",
1293 cfs_list_empty(&obd->obd_req_replay_queue) ? "" : "req ",
1294 cfs_list_empty(&obd->obd_lock_replay_queue) ? \
1296 cfs_list_empty(&obd->obd_final_req_queue) ? \
1298 cfs_spin_unlock(&obd->obd_recovery_task_lock);
1301 cfs_spin_unlock(&obd->obd_recovery_task_lock);
1303 obd->obd_recovery_end = cfs_time_current_sec();
1305 /* when recovery finished, cleanup orphans on mds and ost */
1306 if (OBT(obd) && OBP(obd, postrecov)) {
1307 int rc = OBP(obd, postrecov)(obd);
1309 LCONSOLE_WARN("%s: Post recovery failed, rc %d\n",
1315 static void abort_req_replay_queue(struct obd_device *obd)
1317 struct ptlrpc_request *req, *n;
1318 cfs_list_t abort_list;
1320 CFS_INIT_LIST_HEAD(&abort_list);
1321 cfs_spin_lock(&obd->obd_recovery_task_lock);
1322 cfs_list_splice_init(&obd->obd_req_replay_queue, &abort_list);
1323 cfs_spin_unlock(&obd->obd_recovery_task_lock);
1324 cfs_list_for_each_entry_safe(req, n, &abort_list, rq_list) {
1325 DEBUG_REQ(D_WARNING, req, "aborted:");
1326 req->rq_status = -ENOTCONN;
1327 if (ptlrpc_error(req)) {
1328 DEBUG_REQ(D_ERROR, req,
1329 "failed abort_req_reply; skipping");
1331 target_exp_dequeue_req_replay(req);
1332 target_request_copy_put(req);
1336 static void abort_lock_replay_queue(struct obd_device *obd)
1338 struct ptlrpc_request *req, *n;
1339 cfs_list_t abort_list;
1341 CFS_INIT_LIST_HEAD(&abort_list);
1342 cfs_spin_lock(&obd->obd_recovery_task_lock);
1343 cfs_list_splice_init(&obd->obd_lock_replay_queue, &abort_list);
1344 cfs_spin_unlock(&obd->obd_recovery_task_lock);
1345 cfs_list_for_each_entry_safe(req, n, &abort_list, rq_list){
1346 DEBUG_REQ(D_ERROR, req, "aborted:");
1347 req->rq_status = -ENOTCONN;
1348 if (ptlrpc_error(req)) {
1349 DEBUG_REQ(D_ERROR, req,
1350 "failed abort_lock_reply; skipping");
1352 target_request_copy_put(req);
1357 /* Called from a cleanup function if the device is being cleaned up
1358 forcefully. The exports should all have been disconnected already,
1359 the only thing left to do is
1360 - clear the recovery flags
1362 - free queued requests and replies, but don't send replies
1363 Because the obd_stopping flag is set, no new requests should be received.
1366 void target_cleanup_recovery(struct obd_device *obd)
1368 struct ptlrpc_request *req, *n;
1369 cfs_list_t clean_list;
1372 CFS_INIT_LIST_HEAD(&clean_list);
1373 cfs_spin_lock(&obd->obd_dev_lock);
1374 if (!obd->obd_recovering) {
1375 cfs_spin_unlock(&obd->obd_dev_lock);
1379 obd->obd_recovering = obd->obd_abort_recovery = 0;
1380 cfs_spin_unlock(&obd->obd_dev_lock);
1382 cfs_spin_lock(&obd->obd_recovery_task_lock);
1383 target_cancel_recovery_timer(obd);
1384 cfs_list_splice_init(&obd->obd_req_replay_queue, &clean_list);
1385 cfs_spin_unlock(&obd->obd_recovery_task_lock);
1387 cfs_list_for_each_entry_safe(req, n, &clean_list, rq_list) {
1388 LASSERT(req->rq_reply_state == 0);
1389 target_exp_dequeue_req_replay(req);
1390 target_request_copy_put(req);
1393 cfs_spin_lock(&obd->obd_recovery_task_lock);
1394 cfs_list_splice_init(&obd->obd_lock_replay_queue, &clean_list);
1395 cfs_list_splice_init(&obd->obd_final_req_queue, &clean_list);
1396 cfs_spin_unlock(&obd->obd_recovery_task_lock);
1398 cfs_list_for_each_entry_safe(req, n, &clean_list, rq_list){
1399 LASSERT(req->rq_reply_state == 0);
1400 target_request_copy_put(req);
1406 /* obd_recovery_task_lock should be held */
1407 void target_cancel_recovery_timer(struct obd_device *obd)
1409 CDEBUG(D_HA, "%s: cancel recovery timer\n", obd->obd_name);
1410 cfs_timer_disarm(&obd->obd_recovery_timer);
1413 /* extend = 1 means require at least "duration" seconds left in the timer,
1414 extend = 0 means set the total duration (start_recovery_timer) */
1415 static void reset_recovery_timer(struct obd_device *obd, int duration,
1418 cfs_time_t now = cfs_time_current_sec();
1419 cfs_duration_t left;
1421 cfs_spin_lock(&obd->obd_recovery_task_lock);
1422 if (!obd->obd_recovering || obd->obd_abort_recovery) {
1423 cfs_spin_unlock(&obd->obd_recovery_task_lock);
1427 left = cfs_time_sub(obd->obd_recovery_end, now);
1429 if (extend && (duration > left))
1430 obd->obd_recovery_timeout += duration - left;
1431 else if (!extend && (duration > obd->obd_recovery_timeout))
1432 /* Track the client's largest expected replay time */
1433 obd->obd_recovery_timeout = duration;
1435 /* Hard limit of obd_recovery_time_hard which should not happen */
1436 if (obd->obd_recovery_timeout > obd->obd_recovery_time_hard)
1437 obd->obd_recovery_timeout = obd->obd_recovery_time_hard;
1439 obd->obd_recovery_end = obd->obd_recovery_start +
1440 obd->obd_recovery_timeout;
1441 if (cfs_time_before(now, obd->obd_recovery_end)) {
1442 left = cfs_time_sub(obd->obd_recovery_end, now);
1444 cfs_timer_arm(&obd->obd_recovery_timer, cfs_time_shift(left));
1446 cfs_spin_unlock(&obd->obd_recovery_task_lock);
1447 CDEBUG(D_HA, "%s: recovery timer will expire in %u seconds\n",
1448 obd->obd_name, (unsigned)left);
1451 static void check_and_start_recovery_timer(struct obd_device *obd)
1453 cfs_spin_lock(&obd->obd_recovery_task_lock);
1454 if (cfs_timer_is_armed(&obd->obd_recovery_timer)) {
1455 cfs_spin_unlock(&obd->obd_recovery_task_lock);
1458 CDEBUG(D_HA, "%s: starting recovery timer\n", obd->obd_name);
1459 obd->obd_recovery_start = cfs_time_current_sec();
1460 cfs_spin_unlock(&obd->obd_recovery_task_lock);
1462 reset_recovery_timer(obd, obd->obd_recovery_timeout, 0);
1465 /* Reset the timer with each new client connection */
1467 * This timer is actually reconnect_timer, which is for making sure
1468 * the total recovery window is at least as big as my reconnect
1469 * attempt timing. So the initial recovery time_out will be set to
1470 * OBD_RECOVERY_FACTOR * obd_timeout. If the timeout coming
1471 * from client is bigger than this, then the recovery time_out will
1472 * be extended to make sure the client could be reconnected, in the
1473 * process, the timeout from the new client should be ignored.
1477 target_start_and_reset_recovery_timer(struct obd_device *obd,
1478 struct ptlrpc_request *req,
1481 int service_time = lustre_msg_get_service_time(req->rq_reqmsg);
1483 if (!new_client && service_time)
1484 /* Teach server about old server's estimates, as first guess
1485 * at how long new requests will take. */
1486 at_measured(&req->rq_rqbd->rqbd_service->srv_at_estimate,
1489 check_and_start_recovery_timer(obd);
1491 /* convert the service time to rpc timeout,
1492 * reuse service_time to limit stack usage */
1493 service_time = at_est2timeout(service_time);
1495 /* We expect other clients to timeout within service_time, then try
1496 * to reconnect, then try the failover server. The max delay between
1497 * connect attempts is SWITCH_MAX + SWITCH_INC + INITIAL */
1498 service_time += 2 * (CONNECTION_SWITCH_MAX + CONNECTION_SWITCH_INC +
1499 INITIAL_CONNECT_TIMEOUT);
1500 if (service_time > obd->obd_recovery_timeout && !new_client)
1501 reset_recovery_timer(obd, service_time, 0);
1506 /** Health checking routines */
1507 static inline int exp_connect_healthy(struct obd_export *exp)
1509 return (exp->exp_in_recovery);
1512 /** if export done req_replay or has replay in queue */
1513 static inline int exp_req_replay_healthy(struct obd_export *exp)
1515 return (!exp->exp_req_replay_needed ||
1516 cfs_atomic_read(&exp->exp_replay_count) > 0);
1518 /** if export done lock_replay or has replay in queue */
1519 static inline int exp_lock_replay_healthy(struct obd_export *exp)
1521 return (!exp->exp_lock_replay_needed ||
1522 cfs_atomic_read(&exp->exp_replay_count) > 0);
1525 static inline int exp_vbr_healthy(struct obd_export *exp)
1527 return (!exp->exp_vbr_failed);
1530 static inline int exp_finished(struct obd_export *exp)
1532 return (exp->exp_in_recovery && !exp->exp_lock_replay_needed);
1535 /** Checking routines for recovery */
1536 static int check_for_clients(struct obd_device *obd)
1538 if (obd->obd_abort_recovery || obd->obd_recovery_expired)
1540 LASSERT(obd->obd_connected_clients <= obd->obd_max_recoverable_clients);
1541 if (obd->obd_no_conn == 0 &&
1542 obd->obd_connected_clients + obd->obd_stale_clients ==
1543 obd->obd_max_recoverable_clients)
1548 static int check_for_next_transno(struct obd_device *obd)
1550 struct ptlrpc_request *req = NULL;
1551 int wake_up = 0, connected, completed, queue_len;
1552 __u64 next_transno, req_transno;
1555 cfs_spin_lock(&obd->obd_recovery_task_lock);
1556 if (!cfs_list_empty(&obd->obd_req_replay_queue)) {
1557 req = cfs_list_entry(obd->obd_req_replay_queue.next,
1558 struct ptlrpc_request, rq_list);
1559 req_transno = lustre_msg_get_transno(req->rq_reqmsg);
1564 connected = obd->obd_connected_clients;
1565 completed = connected - cfs_atomic_read(&obd->obd_req_replay_clients);
1566 queue_len = obd->obd_requests_queued_for_recovery;
1567 next_transno = obd->obd_next_recovery_transno;
1569 CDEBUG(D_HA, "max: %d, connected: %d, completed: %d, queue_len: %d, "
1570 "req_transno: "LPU64", next_transno: "LPU64"\n",
1571 obd->obd_max_recoverable_clients, connected, completed,
1572 queue_len, req_transno, next_transno);
1574 if (obd->obd_abort_recovery) {
1575 CDEBUG(D_HA, "waking for aborted recovery\n");
1577 } else if (obd->obd_recovery_expired) {
1578 CDEBUG(D_HA, "waking for expired recovery\n");
1580 } else if (cfs_atomic_read(&obd->obd_req_replay_clients) == 0) {
1581 CDEBUG(D_HA, "waking for completed recovery\n");
1583 } else if (req_transno == next_transno) {
1584 CDEBUG(D_HA, "waking for next ("LPD64")\n", next_transno);
1586 } else if (queue_len == cfs_atomic_read(&obd->obd_req_replay_clients)) {
1588 /** handle gaps occured due to lost reply or VBR */
1589 LASSERTF(req_transno >= next_transno,
1590 "req_transno: "LPU64", next_transno: "LPU64"\n",
1591 req_transno, next_transno);
1592 if (req_transno > obd->obd_last_committed &&
1593 !obd->obd_version_recov)
1596 "%s: waking for gap in transno, VBR is %s (skip: "
1597 LPD64", ql: %d, comp: %d, conn: %d, next: "LPD64
1598 ", last_committed: "LPD64")\n",
1599 obd->obd_name, obd->obd_version_recov ? "ON" : "OFF",
1600 next_transno, queue_len, completed, connected,
1601 req_transno, obd->obd_last_committed);
1602 obd->obd_next_recovery_transno = req_transno;
1604 } else if (OBD_FAIL_CHECK(OBD_FAIL_MDS_RECOVERY_ACCEPTS_GAPS)) {
1605 CDEBUG(D_HA, "accepting transno gaps is explicitly allowed"
1606 " by fail_lock, waking up ("LPD64")\n", next_transno);
1607 obd->obd_next_recovery_transno = req_transno;
1610 cfs_spin_unlock(&obd->obd_recovery_task_lock);
1614 static int check_for_next_lock(struct obd_device *obd)
1618 cfs_spin_lock(&obd->obd_recovery_task_lock);
1619 if (!cfs_list_empty(&obd->obd_lock_replay_queue)) {
1620 CDEBUG(D_HA, "waking for next lock\n");
1622 } else if (cfs_atomic_read(&obd->obd_lock_replay_clients) == 0) {
1623 CDEBUG(D_HA, "waking for completed lock replay\n");
1625 } else if (obd->obd_abort_recovery) {
1626 CDEBUG(D_HA, "waking for aborted recovery\n");
1628 } else if (obd->obd_recovery_expired) {
1629 CDEBUG(D_HA, "waking for expired recovery\n");
1632 cfs_spin_unlock(&obd->obd_recovery_task_lock);
1638 * wait for recovery events,
1639 * check its status with help of check_routine
1640 * evict dead clients via health_check
1642 static int target_recovery_overseer(struct obd_device *obd,
1643 int (*check_routine)(struct obd_device *),
1644 int (*health_check)(struct obd_export *))
1647 cfs_wait_event(obd->obd_next_transno_waitq, check_routine(obd));
1648 if (obd->obd_abort_recovery) {
1649 CWARN("recovery is aborted, evict exports in recovery\n");
1650 /** evict exports which didn't finish recovery yet */
1651 class_disconnect_stale_exports(obd, exp_finished);
1653 } else if (obd->obd_recovery_expired) {
1654 obd->obd_recovery_expired = 0;
1655 /** If some clients died being recovered, evict them */
1656 LCONSOLE_WARN("%s: recovery is timed out, "
1657 "evict stale exports\n", obd->obd_name);
1658 /** evict cexports with no replay in queue, they are stalled */
1659 class_disconnect_stale_exports(obd, health_check);
1660 /** continue with VBR */
1661 cfs_spin_lock(&obd->obd_dev_lock);
1662 obd->obd_version_recov = 1;
1663 cfs_spin_unlock(&obd->obd_dev_lock);
1665 * reset timer, recovery will proceed with versions now,
1666 * timeout is set just to handle reconnection delays
1668 reset_recovery_timer(obd, RECONNECT_DELAY_MAX, 1);
1669 /** Wait for recovery events again, after evicting bad clients */
1675 static struct ptlrpc_request *target_next_replay_req(struct obd_device *obd)
1677 struct ptlrpc_request *req = NULL;
1680 CDEBUG(D_HA, "Waiting for transno "LPD64"\n",
1681 obd->obd_next_recovery_transno);
1683 if (target_recovery_overseer(obd, check_for_next_transno,
1684 exp_req_replay_healthy)) {
1685 abort_req_replay_queue(obd);
1686 abort_lock_replay_queue(obd);
1689 cfs_spin_lock(&obd->obd_recovery_task_lock);
1690 if (!cfs_list_empty(&obd->obd_req_replay_queue)) {
1691 req = cfs_list_entry(obd->obd_req_replay_queue.next,
1692 struct ptlrpc_request, rq_list);
1693 cfs_list_del_init(&req->rq_list);
1694 obd->obd_requests_queued_for_recovery--;
1695 cfs_spin_unlock(&obd->obd_recovery_task_lock);
1697 cfs_spin_unlock(&obd->obd_recovery_task_lock);
1698 LASSERT(cfs_list_empty(&obd->obd_req_replay_queue));
1699 LASSERT(cfs_atomic_read(&obd->obd_req_replay_clients) == 0);
1700 /** evict exports failed VBR */
1701 class_disconnect_stale_exports(obd, exp_vbr_healthy);
1706 static struct ptlrpc_request *target_next_replay_lock(struct obd_device *obd)
1708 struct ptlrpc_request *req = NULL;
1710 CDEBUG(D_HA, "Waiting for lock\n");
1711 if (target_recovery_overseer(obd, check_for_next_lock,
1712 exp_lock_replay_healthy))
1713 abort_lock_replay_queue(obd);
1715 cfs_spin_lock(&obd->obd_recovery_task_lock);
1716 if (!cfs_list_empty(&obd->obd_lock_replay_queue)) {
1717 req = cfs_list_entry(obd->obd_lock_replay_queue.next,
1718 struct ptlrpc_request, rq_list);
1719 cfs_list_del_init(&req->rq_list);
1720 cfs_spin_unlock(&obd->obd_recovery_task_lock);
1722 cfs_spin_unlock(&obd->obd_recovery_task_lock);
1723 LASSERT(cfs_list_empty(&obd->obd_lock_replay_queue));
1724 LASSERT(cfs_atomic_read(&obd->obd_lock_replay_clients) == 0);
1725 /** evict exports failed VBR */
1726 class_disconnect_stale_exports(obd, exp_vbr_healthy);
1731 static struct ptlrpc_request *target_next_final_ping(struct obd_device *obd)
1733 struct ptlrpc_request *req = NULL;
1735 cfs_spin_lock(&obd->obd_recovery_task_lock);
1736 if (!cfs_list_empty(&obd->obd_final_req_queue)) {
1737 req = cfs_list_entry(obd->obd_final_req_queue.next,
1738 struct ptlrpc_request, rq_list);
1739 cfs_list_del_init(&req->rq_list);
1740 cfs_spin_unlock(&obd->obd_recovery_task_lock);
1741 if (req->rq_export->exp_in_recovery) {
1742 cfs_spin_lock(&req->rq_export->exp_lock);
1743 req->rq_export->exp_in_recovery = 0;
1744 cfs_spin_unlock(&req->rq_export->exp_lock);
1747 cfs_spin_unlock(&obd->obd_recovery_task_lock);
1752 static int handle_recovery_req(struct ptlrpc_thread *thread,
1753 struct ptlrpc_request *req,
1754 svc_handler_t handler)
1760 * export can be evicted during recovery, no need to handle replays for
1761 * it after that, discard such request silently
1763 if (req->rq_export->exp_disconnected)
1764 GOTO(reqcopy_put, rc = 0);
1766 rc = lu_context_init(&req->rq_recov_session, LCT_SESSION);
1768 CERROR("Failure to initialize session: %d\n", rc);
1769 GOTO(reqcopy_put, rc);
1772 req->rq_recov_session.lc_thread = thread;
1773 lu_context_enter(&req->rq_recov_session);
1774 req->rq_svc_thread = thread;
1775 req->rq_svc_thread->t_env->le_ses = &req->rq_recov_session;
1777 /* thread context */
1778 lu_context_enter(&thread->t_env->le_ctx);
1780 lu_context_exit(&thread->t_env->le_ctx);
1782 lu_context_exit(&req->rq_recov_session);
1783 lu_context_fini(&req->rq_recov_session);
1784 /* don't reset timer for final stage */
1785 if (!exp_finished(req->rq_export)) {
1787 * Add request timeout to the recovery time so next request from
1788 * this client may come in recovery time
1790 reset_recovery_timer(class_exp2obd(req->rq_export),
1791 AT_OFF ? obd_timeout :
1792 lustre_msg_get_timeout(req->rq_reqmsg), 1);
1798 static int target_recovery_thread(void *arg)
1800 struct lu_target *lut = arg;
1801 struct obd_device *obd = lut->lut_obd;
1802 struct ptlrpc_request *req;
1803 struct target_recovery_data *trd = &obd->obd_recovery_data;
1804 unsigned long delta;
1805 unsigned long flags;
1807 struct ptlrpc_thread *thread = NULL;
1811 cfs_daemonize_ctxt("tgt_recov");
1813 SIGNAL_MASK_LOCK(current, flags);
1814 sigfillset(¤t->blocked);
1816 SIGNAL_MASK_UNLOCK(current, flags);
1818 OBD_ALLOC_PTR(thread);
1824 OBD_FREE_PTR(thread);
1828 rc = lu_context_init(&env->le_ctx, LCT_MD_THREAD);
1830 OBD_FREE_PTR(thread);
1835 thread->t_env = env;
1836 thread->t_id = -1; /* force filter_iobuf_get/put to use local buffers */
1837 env->le_ctx.lc_thread = thread;
1838 thread->t_data = NULL;
1839 thread->t_watchdog = NULL;
1841 CDEBUG(D_HA, "%s: started recovery thread pid %d\n", obd->obd_name,
1843 trd->trd_processing_task = cfs_curproc_pid();
1845 cfs_spin_lock(&obd->obd_dev_lock);
1846 obd->obd_recovering = 1;
1847 cfs_spin_unlock(&obd->obd_dev_lock);
1848 cfs_complete(&trd->trd_starting);
1850 /* first of all, we have to know the first transno to replay */
1851 if (target_recovery_overseer(obd, check_for_clients,
1852 exp_connect_healthy)) {
1853 abort_req_replay_queue(obd);
1854 abort_lock_replay_queue(obd);
1857 /* next stage: replay requests */
1859 CDEBUG(D_INFO, "1: request replay stage - %d clients from t"LPU64"\n",
1860 cfs_atomic_read(&obd->obd_req_replay_clients),
1861 obd->obd_next_recovery_transno);
1862 while ((req = target_next_replay_req(obd))) {
1863 LASSERT(trd->trd_processing_task == cfs_curproc_pid());
1864 DEBUG_REQ(D_HA, req, "processing t"LPD64" from %s",
1865 lustre_msg_get_transno(req->rq_reqmsg),
1866 libcfs_nid2str(req->rq_peer.nid));
1867 handle_recovery_req(thread, req,
1868 trd->trd_recovery_handler);
1870 * bz18031: increase next_recovery_transno before
1871 * target_request_copy_put() will drop exp_rpc reference
1873 cfs_spin_lock(&obd->obd_recovery_task_lock);
1874 obd->obd_next_recovery_transno++;
1875 cfs_spin_unlock(&obd->obd_recovery_task_lock);
1876 target_exp_dequeue_req_replay(req);
1877 target_request_copy_put(req);
1878 obd->obd_replayed_requests++;
1882 * The second stage: replay locks
1884 CDEBUG(D_INFO, "2: lock replay stage - %d clients\n",
1885 cfs_atomic_read(&obd->obd_lock_replay_clients));
1886 while ((req = target_next_replay_lock(obd))) {
1887 LASSERT(trd->trd_processing_task == cfs_curproc_pid());
1888 DEBUG_REQ(D_HA, req, "processing lock from %s: ",
1889 libcfs_nid2str(req->rq_peer.nid));
1890 handle_recovery_req(thread, req,
1891 trd->trd_recovery_handler);
1892 target_request_copy_put(req);
1893 obd->obd_replayed_locks++;
1897 * The third stage: reply on final pings, at this moment all clients
1898 * must have request in final queue
1900 CDEBUG(D_INFO, "3: final stage - process recovery completion pings\n");
1901 /** Update server last boot epoch */
1902 lut_boot_epoch_update(lut);
1903 /* We drop recoverying flag to forward all new requests
1904 * to regular mds_handle() since now */
1905 cfs_spin_lock(&obd->obd_dev_lock);
1906 obd->obd_recovering = obd->obd_abort_recovery = 0;
1907 cfs_spin_unlock(&obd->obd_dev_lock);
1908 cfs_spin_lock(&obd->obd_recovery_task_lock);
1909 target_cancel_recovery_timer(obd);
1910 cfs_spin_unlock(&obd->obd_recovery_task_lock);
1911 while ((req = target_next_final_ping(obd))) {
1912 LASSERT(trd->trd_processing_task == cfs_curproc_pid());
1913 DEBUG_REQ(D_HA, req, "processing final ping from %s: ",
1914 libcfs_nid2str(req->rq_peer.nid));
1915 handle_recovery_req(thread, req,
1916 trd->trd_recovery_handler);
1917 target_request_copy_put(req);
1920 delta = (jiffies - delta) / CFS_HZ;
1921 CDEBUG(D_INFO,"4: recovery completed in %lus - %d/%d reqs/locks\n",
1922 delta, obd->obd_replayed_requests, obd->obd_replayed_locks);
1923 if (delta > OBD_RECOVERY_TIME_SOFT) {
1924 CWARN("too long recovery - read logs\n");
1925 libcfs_debug_dumplog();
1928 target_finish_recovery(obd);
1930 lu_context_fini(&env->le_ctx);
1931 trd->trd_processing_task = 0;
1932 cfs_complete(&trd->trd_finishing);
1934 OBD_FREE_PTR(thread);
1939 static int target_start_recovery_thread(struct lu_target *lut,
1940 svc_handler_t handler)
1942 struct obd_device *obd = lut->lut_obd;
1944 struct target_recovery_data *trd = &obd->obd_recovery_data;
1946 memset(trd, 0, sizeof(*trd));
1947 cfs_init_completion(&trd->trd_starting);
1948 cfs_init_completion(&trd->trd_finishing);
1949 trd->trd_recovery_handler = handler;
1951 if (cfs_create_thread(target_recovery_thread, lut, 0) > 0) {
1952 cfs_wait_for_completion(&trd->trd_starting);
1953 LASSERT(obd->obd_recovering != 0);
1960 void target_stop_recovery_thread(struct obd_device *obd)
1962 if (obd->obd_recovery_data.trd_processing_task > 0) {
1963 struct target_recovery_data *trd = &obd->obd_recovery_data;
1964 /** recovery can be done but postrecovery is not yet */
1965 cfs_spin_lock(&obd->obd_dev_lock);
1966 if (obd->obd_recovering) {
1967 CERROR("%s: Aborting recovery\n", obd->obd_name);
1968 obd->obd_abort_recovery = 1;
1969 cfs_waitq_signal(&obd->obd_next_transno_waitq);
1971 cfs_spin_unlock(&obd->obd_dev_lock);
1972 cfs_wait_for_completion(&trd->trd_finishing);
1976 void target_recovery_fini(struct obd_device *obd)
1978 class_disconnect_exports(obd);
1979 target_stop_recovery_thread(obd);
1980 target_cleanup_recovery(obd);
1982 EXPORT_SYMBOL(target_recovery_fini);
1984 static void target_recovery_expired(unsigned long castmeharder)
1986 struct obd_device *obd = (struct obd_device *)castmeharder;
1987 CDEBUG(D_HA, "%s: recovery timed out; %d clients are still in recovery"
1988 " after %lds (%d clients connected)\n",
1989 obd->obd_name, cfs_atomic_read(&obd->obd_lock_replay_clients),
1990 cfs_time_current_sec()- obd->obd_recovery_start,
1991 obd->obd_connected_clients);
1993 obd->obd_recovery_expired = 1;
1994 cfs_waitq_signal(&obd->obd_next_transno_waitq);
1997 void target_recovery_init(struct lu_target *lut, svc_handler_t handler)
1999 struct obd_device *obd = lut->lut_obd;
2000 if (obd->obd_max_recoverable_clients == 0) {
2001 /** Update server last boot epoch */
2002 lut_boot_epoch_update(lut);
2006 CWARN("RECOVERY: service %s, %d recoverable clients, "
2007 "last_transno "LPU64"\n", obd->obd_name,
2008 obd->obd_max_recoverable_clients, obd->obd_last_committed);
2009 LASSERT(obd->obd_stopping == 0);
2010 obd->obd_next_recovery_transno = obd->obd_last_committed + 1;
2011 obd->obd_recovery_start = 0;
2012 obd->obd_recovery_end = 0;
2014 /* both values can be get from mount data already */
2015 if (obd->obd_recovery_timeout == 0)
2016 obd->obd_recovery_timeout = OBD_RECOVERY_TIME_SOFT;
2017 if (obd->obd_recovery_time_hard == 0)
2018 obd->obd_recovery_time_hard = OBD_RECOVERY_TIME_HARD;
2019 cfs_timer_init(&obd->obd_recovery_timer, target_recovery_expired, obd);
2020 target_start_recovery_thread(lut, handler);
2022 EXPORT_SYMBOL(target_recovery_init);
2026 static int target_process_req_flags(struct obd_device *obd,
2027 struct ptlrpc_request *req)
2029 struct obd_export *exp = req->rq_export;
2030 LASSERT(exp != NULL);
2031 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REQ_REPLAY_DONE) {
2032 /* client declares he's ready to replay locks */
2033 cfs_spin_lock(&exp->exp_lock);
2034 if (exp->exp_req_replay_needed) {
2035 exp->exp_req_replay_needed = 0;
2036 cfs_spin_unlock(&exp->exp_lock);
2038 LASSERT_ATOMIC_POS(&obd->obd_req_replay_clients);
2039 cfs_atomic_dec(&obd->obd_req_replay_clients);
2041 cfs_spin_unlock(&exp->exp_lock);
2044 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_LOCK_REPLAY_DONE) {
2045 /* client declares he's ready to complete recovery
2046 * so, we put the request on th final queue */
2047 cfs_spin_lock(&exp->exp_lock);
2048 if (exp->exp_lock_replay_needed) {
2049 exp->exp_lock_replay_needed = 0;
2050 cfs_spin_unlock(&exp->exp_lock);
2052 LASSERT_ATOMIC_POS(&obd->obd_lock_replay_clients);
2053 cfs_atomic_dec(&obd->obd_lock_replay_clients);
2055 cfs_spin_unlock(&exp->exp_lock);
2061 int target_queue_recovery_request(struct ptlrpc_request *req,
2062 struct obd_device *obd)
2066 __u64 transno = lustre_msg_get_transno(req->rq_reqmsg);
2069 if (obd->obd_recovery_data.trd_processing_task == cfs_curproc_pid()) {
2070 /* Processing the queue right now, don't re-add. */
2074 target_process_req_flags(obd, req);
2076 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_LOCK_REPLAY_DONE) {
2077 /* client declares he's ready to complete recovery
2078 * so, we put the request on th final queue */
2079 target_request_copy_get(req);
2080 DEBUG_REQ(D_HA, req, "queue final req");
2081 cfs_waitq_signal(&obd->obd_next_transno_waitq);
2082 cfs_spin_lock(&obd->obd_recovery_task_lock);
2083 if (obd->obd_recovering) {
2084 cfs_list_add_tail(&req->rq_list,
2085 &obd->obd_final_req_queue);
2087 cfs_spin_unlock(&obd->obd_recovery_task_lock);
2088 target_request_copy_put(req);
2089 RETURN(obd->obd_stopping ? -ENOTCONN : 1);
2091 cfs_spin_unlock(&obd->obd_recovery_task_lock);
2094 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REQ_REPLAY_DONE) {
2095 /* client declares he's ready to replay locks */
2096 target_request_copy_get(req);
2097 DEBUG_REQ(D_HA, req, "queue lock replay req");
2098 cfs_waitq_signal(&obd->obd_next_transno_waitq);
2099 cfs_spin_lock(&obd->obd_recovery_task_lock);
2100 LASSERT(obd->obd_recovering);
2101 /* usually due to recovery abort */
2102 if (!req->rq_export->exp_in_recovery) {
2103 cfs_spin_unlock(&obd->obd_recovery_task_lock);
2104 target_request_copy_put(req);
2107 LASSERT(req->rq_export->exp_lock_replay_needed);
2108 cfs_list_add_tail(&req->rq_list, &obd->obd_lock_replay_queue);
2109 cfs_spin_unlock(&obd->obd_recovery_task_lock);
2113 /* CAVEAT EMPTOR: The incoming request message has been swabbed
2114 * (i.e. buflens etc are in my own byte order), but type-dependent
2115 * buffers (eg mds_body, ost_body etc) have NOT been swabbed. */
2118 CFS_INIT_LIST_HEAD(&req->rq_list);
2119 DEBUG_REQ(D_HA, req, "not queueing");
2123 /* If we're processing the queue, we want don't want to queue this
2126 * Also, if this request has a transno less than the one we're waiting
2127 * for, we should process it now. It could (and currently always will)
2128 * be an open request for a descriptor that was opened some time ago.
2130 * Also, a resent, replayed request that has already been
2131 * handled will pass through here and be processed immediately.
2133 CDEBUG(D_HA, "Next recovery transno: "LPU64
2134 ", current: "LPU64", replaying\n",
2135 obd->obd_next_recovery_transno, transno);
2136 cfs_spin_lock(&obd->obd_recovery_task_lock);
2137 if (transno < obd->obd_next_recovery_transno) {
2138 /* Processing the queue right now, don't re-add. */
2139 LASSERT(cfs_list_empty(&req->rq_list));
2140 cfs_spin_unlock(&obd->obd_recovery_task_lock);
2143 cfs_spin_unlock(&obd->obd_recovery_task_lock);
2145 if (OBD_FAIL_CHECK(OBD_FAIL_TGT_REPLAY_DROP))
2148 target_request_copy_get(req);
2149 if (!req->rq_export->exp_in_recovery) {
2150 target_request_copy_put(req);
2153 LASSERT(req->rq_export->exp_req_replay_needed);
2155 if (target_exp_enqueue_req_replay(req)) {
2156 DEBUG_REQ(D_ERROR, req, "dropping resent queued req");
2157 target_request_copy_put(req);
2162 cfs_spin_lock(&obd->obd_recovery_task_lock);
2163 LASSERT(obd->obd_recovering);
2164 cfs_list_for_each(tmp, &obd->obd_req_replay_queue) {
2165 struct ptlrpc_request *reqiter =
2166 cfs_list_entry(tmp, struct ptlrpc_request, rq_list);
2168 if (lustre_msg_get_transno(reqiter->rq_reqmsg) > transno) {
2169 cfs_list_add_tail(&req->rq_list, &reqiter->rq_list);
2174 if (unlikely(lustre_msg_get_transno(reqiter->rq_reqmsg) ==
2176 DEBUG_REQ(D_ERROR, req, "dropping replay: transno "
2177 "has been claimed by another client");
2178 cfs_spin_unlock(&obd->obd_recovery_task_lock);
2179 target_exp_dequeue_req_replay(req);
2180 target_request_copy_put(req);
2186 cfs_list_add_tail(&req->rq_list, &obd->obd_req_replay_queue);
2188 obd->obd_requests_queued_for_recovery++;
2189 cfs_spin_unlock(&obd->obd_recovery_task_lock);
2190 cfs_waitq_signal(&obd->obd_next_transno_waitq);
2195 * Packs current SLV and Limit into \a req.
2197 int target_pack_pool_reply(struct ptlrpc_request *req)
2199 struct obd_device *obd;
2203 * Check that we still have all structures alive as this may
2204 * be some late rpc in shutdown time.
2206 if (unlikely(!req->rq_export || !req->rq_export->exp_obd ||
2207 !exp_connect_lru_resize(req->rq_export))) {
2208 lustre_msg_set_slv(req->rq_repmsg, 0);
2209 lustre_msg_set_limit(req->rq_repmsg, 0);
2214 * OBD is alive here as export is alive, which we checked above.
2216 obd = req->rq_export->exp_obd;
2218 cfs_read_lock(&obd->obd_pool_lock);
2219 lustre_msg_set_slv(req->rq_repmsg, obd->obd_pool_slv);
2220 lustre_msg_set_limit(req->rq_repmsg, obd->obd_pool_limit);
2221 cfs_read_unlock(&obd->obd_pool_lock);
2226 int target_send_reply_msg(struct ptlrpc_request *req, int rc, int fail_id)
2228 if (OBD_FAIL_CHECK_ORSET(fail_id & ~OBD_FAIL_ONCE, OBD_FAIL_ONCE)) {
2229 DEBUG_REQ(D_ERROR, req, "dropping reply");
2234 DEBUG_REQ(D_NET, req, "processing error (%d)", rc);
2235 req->rq_status = rc;
2236 return (ptlrpc_send_error(req, 1));
2238 DEBUG_REQ(D_NET, req, "sending reply");
2241 return (ptlrpc_send_reply(req, PTLRPC_REPLY_MAYBE_DIFFICULT));
2244 void target_send_reply(struct ptlrpc_request *req, int rc, int fail_id)
2247 struct ptlrpc_reply_state *rs;
2248 struct obd_device *obd;
2249 struct obd_export *exp;
2250 struct ptlrpc_service *svc;
2253 if (req->rq_no_reply) {
2258 svc = req->rq_rqbd->rqbd_service;
2259 rs = req->rq_reply_state;
2260 if (rs == NULL || !rs->rs_difficult) {
2262 target_send_reply_msg (req, rc, fail_id);
2267 /* must be an export if locks saved */
2268 LASSERT (req->rq_export != NULL);
2269 /* req/reply consistent */
2270 LASSERT (rs->rs_service == svc);
2273 LASSERT (!rs->rs_scheduled);
2274 LASSERT (!rs->rs_scheduled_ever);
2275 LASSERT (!rs->rs_handled);
2276 LASSERT (!rs->rs_on_net);
2277 LASSERT (rs->rs_export == NULL);
2278 LASSERT (cfs_list_empty(&rs->rs_obd_list));
2279 LASSERT (cfs_list_empty(&rs->rs_exp_list));
2281 exp = class_export_get (req->rq_export);
2284 /* disable reply scheduling while I'm setting up */
2285 rs->rs_scheduled = 1;
2287 rs->rs_xid = req->rq_xid;
2288 rs->rs_transno = req->rq_transno;
2289 rs->rs_export = exp;
2290 rs->rs_opc = lustre_msg_get_opc(rs->rs_msg);
2292 cfs_spin_lock(&exp->exp_uncommitted_replies_lock);
2293 CDEBUG(D_NET, "rs transno = "LPU64", last committed = "LPU64"\n",
2294 rs->rs_transno, exp->exp_last_committed);
2295 if (rs->rs_transno > exp->exp_last_committed) {
2296 /* not committed already */
2297 cfs_list_add_tail(&rs->rs_obd_list,
2298 &exp->exp_uncommitted_replies);
2300 cfs_spin_unlock (&exp->exp_uncommitted_replies_lock);
2302 cfs_spin_lock(&exp->exp_lock);
2303 cfs_list_add_tail(&rs->rs_exp_list, &exp->exp_outstanding_replies);
2304 cfs_spin_unlock(&exp->exp_lock);
2306 netrc = target_send_reply_msg (req, rc, fail_id);
2308 cfs_spin_lock(&svc->srv_rs_lock);
2310 cfs_atomic_inc(&svc->srv_n_difficult_replies);
2313 /* error sending: reply is off the net. Also we need +1
2314 * reply ref until ptlrpc_handle_rs() is done
2315 * with the reply state (if the send was successful, there
2316 * would have been +1 ref for the net, which
2317 * reply_out_callback leaves alone) */
2319 ptlrpc_rs_addref(rs);
2322 cfs_spin_lock(&rs->rs_lock);
2323 if (rs->rs_transno <= exp->exp_last_committed ||
2324 (!rs->rs_on_net && !rs->rs_no_ack) ||
2325 cfs_list_empty(&rs->rs_exp_list) || /* completed already */
2326 cfs_list_empty(&rs->rs_obd_list)) {
2327 CDEBUG(D_HA, "Schedule reply immediately\n");
2328 ptlrpc_dispatch_difficult_reply(rs);
2330 cfs_list_add (&rs->rs_list, &svc->srv_active_replies);
2331 rs->rs_scheduled = 0; /* allow notifier to schedule */
2333 cfs_spin_unlock(&rs->rs_lock);
2334 cfs_spin_unlock(&svc->srv_rs_lock);
2338 int target_handle_ping(struct ptlrpc_request *req)
2340 obd_ping(req->rq_export);
2341 return req_capsule_server_pack(&req->rq_pill);
2344 void target_committed_to_req(struct ptlrpc_request *req)
2346 struct obd_export *exp = req->rq_export;
2348 if (!exp->exp_obd->obd_no_transno && req->rq_repmsg != NULL)
2349 lustre_msg_set_last_committed(req->rq_repmsg,
2350 exp->exp_last_committed);
2352 DEBUG_REQ(D_IOCTL, req, "not sending last_committed update (%d/"
2353 "%d)", exp->exp_obd->obd_no_transno,
2354 req->rq_repmsg == NULL);
2356 CDEBUG(D_INFO, "last_committed "LPU64", transno "LPU64", xid "LPU64"\n",
2357 exp->exp_last_committed, req->rq_transno, req->rq_xid);
2359 EXPORT_SYMBOL(target_committed_to_req);
2361 int target_handle_qc_callback(struct ptlrpc_request *req)
2363 struct obd_quotactl *oqctl;
2364 struct client_obd *cli = &req->rq_export->exp_obd->u.cli;
2366 oqctl = req_capsule_client_get(&req->rq_pill, &RMF_OBD_QUOTACTL);
2367 if (oqctl == NULL) {
2368 CERROR("Can't unpack obd_quotactl\n");
2372 cli->cl_qchk_stat = oqctl->qc_stat;
2377 #ifdef HAVE_QUOTA_SUPPORT
2378 int target_handle_dqacq_callback(struct ptlrpc_request *req)
2381 struct obd_device *obd = req->rq_export->exp_obd;
2382 struct obd_device *master_obd = NULL, *lov_obd = NULL;
2383 struct obd_device_target *obt;
2384 struct lustre_quota_ctxt *qctxt;
2385 struct qunit_data *qdata = NULL;
2389 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_DROP_QUOTA_REQ))
2392 rc = req_capsule_server_pack(&req->rq_pill);
2394 CERROR("packing reply failed!: rc = %d\n", rc);
2398 LASSERT(req->rq_export);
2400 qdata = quota_get_qdata(req, QUOTA_REQUEST, QUOTA_EXPORT);
2401 if (IS_ERR(qdata)) {
2402 rc = PTR_ERR(qdata);
2403 CDEBUG(D_ERROR, "Can't unpack qunit_data(rc: %d)\n", rc);
2404 req->rq_status = rc;
2408 /* we use the observer */
2409 if (obd_pin_observer(obd, &lov_obd) ||
2410 obd_pin_observer(lov_obd, &master_obd)) {
2411 CERROR("Can't find the observer, it is recovering\n");
2412 req->rq_status = -EAGAIN;
2416 obt = &master_obd->u.obt;
2417 qctxt = &obt->obt_qctxt;
2419 if (!qctxt->lqc_setup || !qctxt->lqc_valid) {
2420 /* quota_type has not been processed yet, return EAGAIN
2421 * until we know whether or not quotas are supposed to
2423 CDEBUG(D_QUOTA, "quota_type not processed yet, return "
2425 req->rq_status = -EAGAIN;
2429 cfs_down_read(&obt->obt_rwsem);
2430 if (qctxt->lqc_lqs_hash == NULL) {
2431 cfs_up_read(&obt->obt_rwsem);
2432 /* quota_type has not been processed yet, return EAGAIN
2433 * until we know whether or not quotas are supposed to
2435 CDEBUG(D_QUOTA, "quota_ctxt is not ready yet, return "
2437 req->rq_status = -EAGAIN;
2441 LASSERT(qctxt->lqc_handler);
2442 rc = qctxt->lqc_handler(master_obd, qdata,
2443 lustre_msg_get_opc(req->rq_reqmsg));
2444 cfs_up_read(&obt->obt_rwsem);
2445 if (rc && rc != -EDQUOT)
2446 CDEBUG(rc == -EBUSY ? D_QUOTA : D_ERROR,
2447 "dqacq/dqrel failed! (rc:%d)\n", rc);
2448 req->rq_status = rc;
2450 rc = quota_copy_qdata(req, qdata, QUOTA_REPLY, QUOTA_EXPORT);
2452 CERROR("Can't pack qunit_data(rc: %d)\n", rc);
2456 /* Block the quota req. b=14840 */
2457 OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_BLOCK_QUOTA_REQ, obd_timeout);
2462 obd_unpin_observer(lov_obd);
2464 obd_unpin_observer(obd);
2466 rc = ptlrpc_reply(req);
2470 #endif /* !__KERNEL__ */
2472 #endif /* HAVE_QUOTA_SUPPORT */
2474 ldlm_mode_t lck_compat_array[] = {
2475 [LCK_EX] LCK_COMPAT_EX,
2476 [LCK_PW] LCK_COMPAT_PW,
2477 [LCK_PR] LCK_COMPAT_PR,
2478 [LCK_CW] LCK_COMPAT_CW,
2479 [LCK_CR] LCK_COMPAT_CR,
2480 [LCK_NL] LCK_COMPAT_NL,
2481 [LCK_GROUP] LCK_COMPAT_GROUP,
2482 [LCK_COS] LCK_COMPAT_COS,
2486 * Rather arbitrary mapping from LDLM error codes to errno values. This should
2487 * not escape to the user level.
2489 int ldlm_error2errno(ldlm_error_t error)
2497 case ELDLM_LOCK_CHANGED:
2500 case ELDLM_LOCK_ABORTED:
2503 case ELDLM_LOCK_REPLACED:
2506 case ELDLM_NO_LOCK_DATA:
2509 case ELDLM_NAMESPACE_EXISTS:
2512 case ELDLM_BAD_NAMESPACE:
2516 if (((int)error) < 0) /* cast to signed type */
2517 result = error; /* as ldlm_error_t can be unsigned */
2519 CERROR("Invalid DLM result code: %d\n", error);
2525 EXPORT_SYMBOL(ldlm_error2errno);
2528 * Dual to ldlm_error2errno(): maps errno values back to ldlm_error_t.
2530 ldlm_error_t ldlm_errno2error(int err_no)
2539 error = ELDLM_LOCK_CHANGED;
2542 error = ELDLM_LOCK_ABORTED;
2545 error = ELDLM_LOCK_REPLACED;
2548 error = ELDLM_NO_LOCK_DATA;
2551 error = ELDLM_NAMESPACE_EXISTS;
2554 error = ELDLM_BAD_NAMESPACE;
2561 EXPORT_SYMBOL(ldlm_errno2error);
2563 #if LUSTRE_TRACKS_LOCK_EXP_REFS
2564 void ldlm_dump_export_locks(struct obd_export *exp)
2566 cfs_spin_lock(&exp->exp_locks_list_guard);
2567 if (!cfs_list_empty(&exp->exp_locks_list)) {
2568 struct ldlm_lock *lock;
2570 CERROR("dumping locks for export %p,"
2571 "ignore if the unmount doesn't hang\n", exp);
2572 cfs_list_for_each_entry(lock, &exp->exp_locks_list, l_exp_refs_link)
2573 ldlm_lock_dump(D_ERROR, lock, 0);
2575 cfs_spin_unlock(&exp->exp_locks_list_guard);
2579 static int target_bulk_timeout(void *data)
2582 /* We don't fail the connection here, because having the export
2583 * killed makes the (vital) call to commitrw very sad.
2588 static inline char *bulk2type(struct ptlrpc_bulk_desc *desc)
2590 return desc->bd_type == BULK_GET_SINK ? "GET" : "PUT";
2593 int target_bulk_io(struct obd_export *exp, struct ptlrpc_bulk_desc *desc,
2594 struct l_wait_info *lwi)
2596 struct ptlrpc_request *req = desc->bd_req;
2600 /* Check if there is eviction in progress, and if so, wait for
2602 if (unlikely(cfs_atomic_read(&exp->exp_obd->obd_evict_inprogress))) {
2603 *lwi = LWI_INTR(NULL, NULL);
2604 rc = l_wait_event(exp->exp_obd->obd_evict_inprogress_waitq,
2605 !cfs_atomic_read(&exp->exp_obd->
2606 obd_evict_inprogress),
2610 /* Check if client was evicted or tried to reconnect already */
2611 if (exp->exp_failed || exp->exp_abort_active_req) {
2614 if (desc->bd_type == BULK_PUT_SINK)
2615 rc = sptlrpc_svc_wrap_bulk(req, desc);
2617 rc = ptlrpc_start_bulk_transfer(desc);
2620 if (rc == 0 && OBD_FAIL_CHECK(OBD_FAIL_MDS_SENDPAGE)) {
2621 ptlrpc_abort_bulk(desc);
2622 } else if (rc == 0) {
2623 time_t start = cfs_time_current_sec();
2625 long timeoutl = req->rq_deadline - cfs_time_current_sec();
2626 cfs_duration_t timeout = timeoutl <= 0 ?
2627 CFS_TICK : cfs_time_seconds(timeoutl);
2628 *lwi = LWI_TIMEOUT_INTERVAL(timeout,
2629 cfs_time_seconds(1),
2630 target_bulk_timeout,
2632 rc = l_wait_event(desc->bd_waitq,
2633 !ptlrpc_server_bulk_active(desc) ||
2635 exp->exp_abort_active_req,
2637 LASSERT(rc == 0 || rc == -ETIMEDOUT);
2638 /* Wait again if we changed deadline */
2639 } while ((rc == -ETIMEDOUT) &&
2640 (req->rq_deadline > cfs_time_current_sec()));
2642 if (rc == -ETIMEDOUT) {
2643 DEBUG_REQ(D_ERROR, req,
2644 "timeout on bulk %s after %ld%+lds",
2646 req->rq_deadline - start,
2647 cfs_time_current_sec() -
2649 ptlrpc_abort_bulk(desc);
2650 } else if (exp->exp_failed) {
2651 DEBUG_REQ(D_ERROR, req, "Eviction on bulk %s",
2654 ptlrpc_abort_bulk(desc);
2655 } else if (exp->exp_abort_active_req) {
2656 DEBUG_REQ(D_ERROR, req, "Reconnect on bulk %s",
2658 /* we don't reply anyway */
2660 ptlrpc_abort_bulk(desc);
2661 } else if (!desc->bd_success ||
2662 desc->bd_nob_transferred != desc->bd_nob) {
2663 DEBUG_REQ(D_ERROR, req, "%s bulk %s %d(%d)",
2665 "truncated" : "network error on",
2667 desc->bd_nob_transferred,
2669 /* XXX should this be a different errno? */
2671 } else if (desc->bd_type == BULK_GET_SINK) {
2672 rc = sptlrpc_svc_unwrap_bulk(req, desc);
2675 DEBUG_REQ(D_ERROR, req, "bulk %s failed: rc %d",
2676 bulk2type(desc), rc);
2681 EXPORT_SYMBOL(target_bulk_io);