4 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
6 * This program is free software; you can redistribute it and/or modify
7 * it under the terms of the GNU General Public License version 2 only,
8 * as published by the Free Software Foundation.
10 * This program is distributed in the hope that it will be useful, but
11 * WITHOUT ANY WARRANTY; without even the implied warranty of
12 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
13 * General Public License version 2 for more details (a copy is included
14 * in the LICENSE file that accompanied this code).
16 * You should have received a copy of the GNU General Public License
17 * version 2 along with this program; If not, see
18 * http://www.sun.com/software/products/lustre/docs/GPLv2.pdf
20 * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
21 * CA 95054 USA or visit www.sun.com if you need additional information or
27 * Copyright (c) 2003, 2010, Oracle and/or its affiliates. All rights reserved.
28 * Use is subject to license terms.
30 * Copyright (c) 2010, 2014, Intel Corporation.
33 * This file is part of Lustre, http://www.lustre.org/
34 * Lustre is a trademark of Sun Microsystems, Inc.
38 * This file deals with various client/target related logic including recovery.
40 * TODO: This code more logically belongs in the ptlrpc module than in ldlm and
44 #define DEBUG_SUBSYSTEM S_LDLM
46 #include <linux/kthread.h>
47 #include <libcfs/libcfs.h>
49 #include <obd_class.h>
50 #include <lustre_dlm.h>
51 #include <lustre_net.h>
52 #include <lustre_sec.h>
53 #include "ldlm_internal.h"
55 /* @priority: If non-zero, move the selected connection to the list head.
56 * @create: If zero, only search in existing connections.
58 static int import_set_conn(struct obd_import *imp, struct obd_uuid *uuid,
59 int priority, int create)
61 struct ptlrpc_connection *ptlrpc_conn;
62 struct obd_import_conn *imp_conn = NULL, *item;
66 if (!create && !priority) {
67 CDEBUG(D_HA, "Nothing to do\n");
71 ptlrpc_conn = ptlrpc_uuid_to_connection(uuid);
73 CDEBUG(D_HA, "can't find connection %s\n", uuid->uuid);
78 OBD_ALLOC(imp_conn, sizeof(*imp_conn));
80 GOTO(out_put, rc = -ENOMEM);
84 spin_lock(&imp->imp_lock);
85 list_for_each_entry(item, &imp->imp_conn_list, oic_item) {
86 if (obd_uuid_equals(uuid, &item->oic_uuid)) {
88 list_del(&item->oic_item);
89 list_add(&item->oic_item,
91 item->oic_last_attempt = 0;
93 CDEBUG(D_HA, "imp %p@%s: found existing conn %s%s\n",
94 imp, imp->imp_obd->obd_name, uuid->uuid,
95 (priority ? ", moved to head" : ""));
96 spin_unlock(&imp->imp_lock);
97 GOTO(out_free, rc = 0);
100 /* No existing import connection found for \a uuid. */
102 imp_conn->oic_conn = ptlrpc_conn;
103 imp_conn->oic_uuid = *uuid;
104 imp_conn->oic_last_attempt = 0;
106 list_add(&imp_conn->oic_item, &imp->imp_conn_list);
108 list_add_tail(&imp_conn->oic_item,
109 &imp->imp_conn_list);
110 CDEBUG(D_HA, "imp %p@%s: add connection %s at %s\n",
111 imp, imp->imp_obd->obd_name, uuid->uuid,
112 (priority ? "head" : "tail"));
114 spin_unlock(&imp->imp_lock);
115 GOTO(out_free, rc = -ENOENT);
118 spin_unlock(&imp->imp_lock);
122 OBD_FREE(imp_conn, sizeof(*imp_conn));
124 ptlrpc_connection_put(ptlrpc_conn);
128 int import_set_conn_priority(struct obd_import *imp, struct obd_uuid *uuid)
130 return import_set_conn(imp, uuid, 1, 0);
133 int client_import_add_conn(struct obd_import *imp, struct obd_uuid *uuid,
136 return import_set_conn(imp, uuid, priority, 1);
138 EXPORT_SYMBOL(client_import_add_conn);
140 int client_import_del_conn(struct obd_import *imp, struct obd_uuid *uuid)
142 struct obd_import_conn *imp_conn;
143 struct obd_export *dlmexp;
147 spin_lock(&imp->imp_lock);
148 if (list_empty(&imp->imp_conn_list)) {
149 LASSERT(!imp->imp_connection);
153 list_for_each_entry(imp_conn, &imp->imp_conn_list, oic_item) {
154 if (!obd_uuid_equals(uuid, &imp_conn->oic_uuid))
156 LASSERT(imp_conn->oic_conn);
158 if (imp_conn == imp->imp_conn_current) {
159 LASSERT(imp_conn->oic_conn == imp->imp_connection);
161 if (imp->imp_state != LUSTRE_IMP_CLOSED &&
162 imp->imp_state != LUSTRE_IMP_DISCON) {
163 CERROR("can't remove current connection\n");
164 GOTO(out, rc = -EBUSY);
167 ptlrpc_connection_put(imp->imp_connection);
168 imp->imp_connection = NULL;
170 dlmexp = class_conn2export(&imp->imp_dlm_handle);
171 if (dlmexp && dlmexp->exp_connection) {
172 LASSERT(dlmexp->exp_connection ==
174 ptlrpc_connection_put(dlmexp->exp_connection);
175 dlmexp->exp_connection = NULL;
179 list_del(&imp_conn->oic_item);
180 ptlrpc_connection_put(imp_conn->oic_conn);
181 OBD_FREE(imp_conn, sizeof(*imp_conn));
182 CDEBUG(D_HA, "imp %p@%s: remove connection %s\n",
183 imp, imp->imp_obd->obd_name, uuid->uuid);
188 spin_unlock(&imp->imp_lock);
190 CERROR("connection %s not found\n", uuid->uuid);
193 EXPORT_SYMBOL(client_import_del_conn);
196 * Find conn UUID by peer NID. \a peer is a server NID. This function is used
197 * to find a conn uuid of \a imp which can reach \a peer.
199 int client_import_find_conn(struct obd_import *imp, lnet_nid_t peer,
200 struct obd_uuid *uuid)
202 struct obd_import_conn *conn;
206 spin_lock(&imp->imp_lock);
207 list_for_each_entry(conn, &imp->imp_conn_list, oic_item) {
208 /* Check if conn UUID does have this peer NID. */
209 if (class_check_uuid(&conn->oic_uuid, peer)) {
210 *uuid = conn->oic_uuid;
215 spin_unlock(&imp->imp_lock);
218 EXPORT_SYMBOL(client_import_find_conn);
220 void client_destroy_import(struct obd_import *imp)
222 /* Drop security policy instance after all RPCs have finished/aborted
223 * to let all busy contexts be released. */
224 class_import_get(imp);
225 class_destroy_import(imp);
226 sptlrpc_import_sec_put(imp);
227 class_import_put(imp);
229 EXPORT_SYMBOL(client_destroy_import);
232 * Check whether or not the OSC is on MDT.
235 * setup 0:{fsname}-OSTxxxx-osc[-MDTxxxx] 1:lustre-OST0000_UUID 2:NID
237 * setup 0:{fsname}-OSTxxxx-osc 1:lustre-OST0000_UUID 2:NID
240 static int osc_on_mdt(char *obdname)
244 ptr = strrchr(obdname, '-');
248 if (strncmp(ptr + 1, "MDT", 3) == 0)
254 /* Configure an RPC client OBD device.
259 * 3 - inactive-on-startup
261 int client_obd_setup(struct obd_device *obddev, struct lustre_cfg *lcfg)
263 struct client_obd *cli = &obddev->u.cli;
264 struct obd_import *imp;
265 struct obd_uuid server_uuid;
266 int rq_portal, rp_portal, connect_op;
267 char *name = obddev->obd_type->typ_name;
268 ldlm_ns_type_t ns_type = LDLM_NS_TYPE_UNKNOWN;
270 char *cli_name = lustre_cfg_buf(lcfg, 0);
273 /* In a more perfect world, we would hang a ptlrpc_client off of
274 * obd_type and just use the values from there. */
275 if (!strcmp(name, LUSTRE_OSC_NAME)) {
276 rq_portal = OST_REQUEST_PORTAL;
277 rp_portal = OSC_REPLY_PORTAL;
278 connect_op = OST_CONNECT;
279 cli->cl_sp_me = LUSTRE_SP_CLI;
280 cli->cl_sp_to = LUSTRE_SP_OST;
281 ns_type = LDLM_NS_TYPE_OSC;
282 } else if (!strcmp(name, LUSTRE_MDC_NAME) ||
283 !strcmp(name, LUSTRE_LWP_NAME)) {
284 rq_portal = MDS_REQUEST_PORTAL;
285 rp_portal = MDC_REPLY_PORTAL;
286 connect_op = MDS_CONNECT;
287 if (is_lwp_on_ost(cli_name))
288 cli->cl_sp_me = LUSTRE_SP_OST;
289 else if (is_lwp_on_mdt(cli_name))
290 cli->cl_sp_me = LUSTRE_SP_MDT;
292 cli->cl_sp_me = LUSTRE_SP_CLI;
293 cli->cl_sp_to = LUSTRE_SP_MDT;
294 ns_type = LDLM_NS_TYPE_MDC;
295 } else if (!strcmp(name, LUSTRE_OSP_NAME)) {
296 if (strstr(lustre_cfg_buf(lcfg, 1), "OST") == NULL) {
297 /* OSP_on_MDT for other MDTs */
298 connect_op = MDS_CONNECT;
299 cli->cl_sp_to = LUSTRE_SP_MDT;
300 ns_type = LDLM_NS_TYPE_MDC;
301 rq_portal = OUT_PORTAL;
303 /* OSP on MDT for OST */
304 connect_op = OST_CONNECT;
305 cli->cl_sp_to = LUSTRE_SP_OST;
306 ns_type = LDLM_NS_TYPE_OSC;
307 rq_portal = OST_REQUEST_PORTAL;
309 rp_portal = OSC_REPLY_PORTAL;
310 cli->cl_sp_me = LUSTRE_SP_MDT;
311 } else if (!strcmp(name, LUSTRE_MGC_NAME)) {
312 rq_portal = MGS_REQUEST_PORTAL;
313 rp_portal = MGC_REPLY_PORTAL;
314 connect_op = MGS_CONNECT;
315 cli->cl_sp_me = LUSTRE_SP_MGC;
316 cli->cl_sp_to = LUSTRE_SP_MGS;
317 cli->cl_flvr_mgc.sf_rpc = SPTLRPC_FLVR_INVALID;
318 ns_type = LDLM_NS_TYPE_MGC;
320 CERROR("unknown client OBD type \"%s\", can't setup\n",
325 if (LUSTRE_CFG_BUFLEN(lcfg, 1) < 1) {
326 CERROR("requires a TARGET UUID\n");
330 if (LUSTRE_CFG_BUFLEN(lcfg, 1) > 37) {
331 CERROR("client UUID must be less than 38 characters\n");
335 if (LUSTRE_CFG_BUFLEN(lcfg, 2) < 1) {
336 CERROR("setup requires a SERVER UUID\n");
340 if (LUSTRE_CFG_BUFLEN(lcfg, 2) > 37) {
341 CERROR("target UUID must be less than 38 characters\n");
345 init_rwsem(&cli->cl_sem);
346 mutex_init(&cli->cl_mgc_mutex);
347 cli->cl_conn_count = 0;
348 memcpy(server_uuid.uuid, lustre_cfg_buf(lcfg, 2),
349 min_t(unsigned int, LUSTRE_CFG_BUFLEN(lcfg, 2),
350 sizeof(server_uuid)));
352 cli->cl_dirty_pages = 0;
353 cli->cl_avail_grant = 0;
354 /* FIXME: Should limit this for the sum of all cl_dirty_max_pages. */
355 /* cl_dirty_max_pages may be changed at connect time in
356 * ptlrpc_connect_interpret(). */
357 client_adjust_max_dirty(cli);
358 INIT_LIST_HEAD(&cli->cl_cache_waiters);
359 INIT_LIST_HEAD(&cli->cl_loi_ready_list);
360 INIT_LIST_HEAD(&cli->cl_loi_hp_ready_list);
361 INIT_LIST_HEAD(&cli->cl_loi_write_list);
362 INIT_LIST_HEAD(&cli->cl_loi_read_list);
363 spin_lock_init(&cli->cl_loi_list_lock);
364 atomic_set(&cli->cl_pending_w_pages, 0);
365 atomic_set(&cli->cl_pending_r_pages, 0);
366 cli->cl_r_in_flight = 0;
367 cli->cl_w_in_flight = 0;
369 spin_lock_init(&cli->cl_read_rpc_hist.oh_lock);
370 spin_lock_init(&cli->cl_write_rpc_hist.oh_lock);
371 spin_lock_init(&cli->cl_read_page_hist.oh_lock);
372 spin_lock_init(&cli->cl_write_page_hist.oh_lock);
373 spin_lock_init(&cli->cl_read_offset_hist.oh_lock);
374 spin_lock_init(&cli->cl_write_offset_hist.oh_lock);
377 INIT_LIST_HEAD(&cli->cl_lru_osc);
378 atomic_set(&cli->cl_lru_shrinkers, 0);
379 atomic_long_set(&cli->cl_lru_busy, 0);
380 atomic_long_set(&cli->cl_lru_in_list, 0);
381 INIT_LIST_HEAD(&cli->cl_lru_list);
382 spin_lock_init(&cli->cl_lru_list_lock);
383 atomic_long_set(&cli->cl_unstable_count, 0);
385 init_waitqueue_head(&cli->cl_destroy_waitq);
386 atomic_set(&cli->cl_destroy_in_flight, 0);
387 #ifdef ENABLE_CHECKSUM
388 /* Turn on checksumming by default. */
389 cli->cl_checksum = 1;
391 * The supported checksum types will be worked out at connect time
392 * Set cl_chksum* to CRC32 for now to avoid returning screwed info
395 cli->cl_cksum_type = cli->cl_supp_cksum_types = OBD_CKSUM_CRC32;
397 atomic_set(&cli->cl_resends, OSC_DEFAULT_RESENDS);
399 /* This value may be reduced at connect time in
400 * ptlrpc_connect_interpret() . We initialize it to only
401 * 1MB until we know what the performance looks like.
402 * In the future this should likely be increased. LU-1431 */
403 cli->cl_max_pages_per_rpc = min_t(int, PTLRPC_MAX_BRW_PAGES,
404 LNET_MTU >> PAGE_CACHE_SHIFT);
406 /* set cl_chunkbits default value to PAGE_CACHE_SHIFT,
407 * it will be updated at OSC connection time. */
408 cli->cl_chunkbits = PAGE_CACHE_SHIFT;
410 if (!strcmp(name, LUSTRE_MDC_NAME)) {
411 cli->cl_max_rpcs_in_flight = OBD_MAX_RIF_DEFAULT;
412 } else if (totalram_pages >> (20 - PAGE_CACHE_SHIFT) <= 128 /* MB */) {
413 cli->cl_max_rpcs_in_flight = 2;
414 } else if (totalram_pages >> (20 - PAGE_CACHE_SHIFT) <= 256 /* MB */) {
415 cli->cl_max_rpcs_in_flight = 3;
416 } else if (totalram_pages >> (20 - PAGE_CACHE_SHIFT) <= 512 /* MB */) {
417 cli->cl_max_rpcs_in_flight = 4;
419 if (osc_on_mdt(obddev->obd_name))
420 cli->cl_max_rpcs_in_flight = OBD_MAX_RIF_MAX;
422 cli->cl_max_rpcs_in_flight = OBD_MAX_RIF_DEFAULT;
425 spin_lock_init(&cli->cl_mod_rpcs_lock);
426 spin_lock_init(&cli->cl_mod_rpcs_hist.oh_lock);
427 cli->cl_max_mod_rpcs_in_flight = 0;
428 cli->cl_mod_rpcs_in_flight = 0;
429 cli->cl_close_rpcs_in_flight = 0;
430 init_waitqueue_head(&cli->cl_mod_rpcs_waitq);
431 cli->cl_mod_tag_bitmap = NULL;
433 if (connect_op == MDS_CONNECT) {
434 cli->cl_max_mod_rpcs_in_flight = cli->cl_max_rpcs_in_flight - 1;
435 OBD_ALLOC(cli->cl_mod_tag_bitmap,
436 BITS_TO_LONGS(OBD_MAX_RIF_MAX) * sizeof(long));
437 if (cli->cl_mod_tag_bitmap == NULL)
438 GOTO(err, rc = -ENOMEM);
443 CERROR("ldlm_get_ref failed: %d\n", rc);
447 ptlrpc_init_client(rq_portal, rp_portal, name,
448 &obddev->obd_ldlm_client);
450 imp = class_new_import(obddev);
452 GOTO(err_ldlm, rc = -ENOENT);
453 imp->imp_client = &obddev->obd_ldlm_client;
454 imp->imp_connect_op = connect_op;
455 memcpy(cli->cl_target_uuid.uuid, lustre_cfg_buf(lcfg, 1),
456 LUSTRE_CFG_BUFLEN(lcfg, 1));
457 class_import_put(imp);
459 rc = client_import_add_conn(imp, &server_uuid, 1);
461 CERROR("can't add initial connection\n");
462 GOTO(err_import, rc);
465 cli->cl_import = imp;
466 /* cli->cl_max_mds_easize updated by mdc_init_ea_size() */
467 cli->cl_max_mds_easize = sizeof(struct lov_mds_md_v3);
469 if (LUSTRE_CFG_BUFLEN(lcfg, 3) > 0) {
470 if (!strcmp(lustre_cfg_string(lcfg, 3), "inactive")) {
471 CDEBUG(D_HA, "marking %s %s->%s as inactive\n",
472 name, obddev->obd_name,
473 cli->cl_target_uuid.uuid);
474 spin_lock(&imp->imp_lock);
475 imp->imp_deactive = 1;
476 spin_unlock(&imp->imp_lock);
480 obddev->obd_namespace = ldlm_namespace_new(obddev, obddev->obd_name,
481 LDLM_NAMESPACE_CLIENT,
482 LDLM_NAMESPACE_GREEDY,
484 if (obddev->obd_namespace == NULL) {
485 CERROR("Unable to create client namespace - %s\n",
487 GOTO(err_import, rc = -ENOMEM);
493 class_destroy_import(imp);
497 if (cli->cl_mod_tag_bitmap != NULL)
498 OBD_FREE(cli->cl_mod_tag_bitmap,
499 BITS_TO_LONGS(OBD_MAX_RIF_MAX) * sizeof(long));
500 cli->cl_mod_tag_bitmap = NULL;
504 EXPORT_SYMBOL(client_obd_setup);
506 int client_obd_cleanup(struct obd_device *obddev)
508 struct client_obd *cli = &obddev->u.cli;
511 ldlm_namespace_free_post(obddev->obd_namespace);
512 obddev->obd_namespace = NULL;
514 obd_cleanup_client_import(obddev);
515 LASSERT(obddev->u.cli.cl_import == NULL);
519 if (cli->cl_mod_tag_bitmap != NULL)
520 OBD_FREE(cli->cl_mod_tag_bitmap,
521 BITS_TO_LONGS(OBD_MAX_RIF_MAX) * sizeof(long));
522 cli->cl_mod_tag_bitmap = NULL;
526 EXPORT_SYMBOL(client_obd_cleanup);
528 /* ->o_connect() method for client side (OSC and MDC and MGC) */
529 int client_connect_import(const struct lu_env *env,
530 struct obd_export **exp,
531 struct obd_device *obd, struct obd_uuid *cluuid,
532 struct obd_connect_data *data, void *localdata)
534 struct client_obd *cli = &obd->u.cli;
535 struct obd_import *imp = cli->cl_import;
536 struct obd_connect_data *ocd;
537 struct lustre_handle conn = { 0 };
543 down_write(&cli->cl_sem);
544 if (cli->cl_conn_count > 0)
545 GOTO(out_sem, rc = -EALREADY);
547 rc = class_connect(&conn, obd, cluuid);
551 cli->cl_conn_count++;
552 *exp = class_conn2export(&conn);
554 LASSERT(obd->obd_namespace);
556 imp->imp_dlm_handle = conn;
557 rc = ptlrpc_init_import(imp);
561 ocd = &imp->imp_connect_data;
564 is_mdc = strncmp(imp->imp_obd->obd_type->typ_name,
565 LUSTRE_MDC_NAME, 3) == 0;
567 data->ocd_connect_flags |= OBD_CONNECT_MULTIMODRPCS;
568 imp->imp_connect_flags_orig = data->ocd_connect_flags;
571 rc = ptlrpc_connect_import(imp);
573 LASSERT (imp->imp_state == LUSTRE_IMP_DISCON);
576 LASSERT(*exp != NULL && (*exp)->exp_connection);
579 LASSERTF((ocd->ocd_connect_flags & data->ocd_connect_flags) ==
580 ocd->ocd_connect_flags, "old "LPX64", new "LPX64"\n",
581 data->ocd_connect_flags, ocd->ocd_connect_flags);
582 data->ocd_connect_flags = ocd->ocd_connect_flags;
583 /* clear the flag as it was not set and is not known
586 data->ocd_connect_flags &= ~OBD_CONNECT_MULTIMODRPCS;
589 ptlrpc_pinger_add_import(imp);
595 cli->cl_conn_count--;
596 class_disconnect(*exp);
600 up_write(&cli->cl_sem);
604 EXPORT_SYMBOL(client_connect_import);
606 int client_disconnect_export(struct obd_export *exp)
608 struct obd_device *obd = class_exp2obd(exp);
609 struct client_obd *cli;
610 struct obd_import *imp;
615 CERROR("invalid export for disconnect: exp %p cookie "LPX64"\n",
616 exp, exp ? exp->exp_handle.h_cookie : -1);
621 imp = cli->cl_import;
623 down_write(&cli->cl_sem);
624 CDEBUG(D_INFO, "disconnect %s - %zu\n", obd->obd_name,
627 if (cli->cl_conn_count == 0) {
628 CERROR("disconnecting disconnected device (%s)\n",
630 GOTO(out_disconnect, rc = -EINVAL);
633 cli->cl_conn_count--;
634 if (cli->cl_conn_count != 0)
635 GOTO(out_disconnect, rc = 0);
637 /* Mark import deactivated now, so we don't try to reconnect if any
638 * of the cleanup RPCs fails (e.g. LDLM cancel, etc). We don't
639 * fully deactivate the import, or that would drop all requests. */
640 spin_lock(&imp->imp_lock);
641 imp->imp_deactive = 1;
642 spin_unlock(&imp->imp_lock);
644 /* Some non-replayable imports (MDS's OSCs) are pinged, so just
645 * delete it regardless. (It's safe to delete an import that was
647 (void)ptlrpc_pinger_del_import(imp);
649 if (obd->obd_namespace != NULL) {
650 /* obd_force == local only */
651 ldlm_cli_cancel_unused(obd->obd_namespace, NULL,
652 obd->obd_force ? LCF_LOCAL : 0, NULL);
653 ldlm_namespace_free_prior(obd->obd_namespace, imp, obd->obd_force);
656 /* There's no need to hold sem while disconnecting an import,
657 * and it may actually cause deadlock in GSS. */
658 up_write(&cli->cl_sem);
659 rc = ptlrpc_disconnect_import(imp, 0);
660 down_write(&cli->cl_sem);
662 ptlrpc_invalidate_import(imp);
667 /* Use server style - class_disconnect should be always called for
669 err = class_disconnect(exp);
673 up_write(&cli->cl_sem);
677 EXPORT_SYMBOL(client_disconnect_export);
679 #ifdef HAVE_SERVER_SUPPORT
680 int server_disconnect_export(struct obd_export *exp)
685 /* Disconnect early so that clients can't keep using export. */
686 rc = class_disconnect(exp);
687 /* Close import to avoid sending any requests. */
688 if (exp->exp_imp_reverse)
689 ptlrpc_cleanup_imp(exp->exp_imp_reverse);
691 if (exp->exp_obd->obd_namespace != NULL)
692 ldlm_cancel_locks_for_export(exp);
694 /* complete all outstanding replies */
695 spin_lock(&exp->exp_lock);
696 while (!list_empty(&exp->exp_outstanding_replies)) {
697 struct ptlrpc_reply_state *rs =
698 list_entry(exp->exp_outstanding_replies.next,
699 struct ptlrpc_reply_state, rs_exp_list);
700 struct ptlrpc_service_part *svcpt = rs->rs_svcpt;
702 spin_lock(&svcpt->scp_rep_lock);
704 list_del_init(&rs->rs_exp_list);
705 spin_lock(&rs->rs_lock);
706 ptlrpc_schedule_difficult_reply(rs);
707 spin_unlock(&rs->rs_lock);
709 spin_unlock(&svcpt->scp_rep_lock);
711 spin_unlock(&exp->exp_lock);
715 EXPORT_SYMBOL(server_disconnect_export);
717 /* --------------------------------------------------------------------------
718 * from old lib/target.c
719 * -------------------------------------------------------------------------- */
721 static int target_handle_reconnect(struct lustre_handle *conn,
722 struct obd_export *exp,
723 struct obd_uuid *cluuid)
727 if (exp->exp_connection && exp->exp_imp_reverse) {
728 struct lustre_handle *hdl;
729 struct obd_device *target;
731 hdl = &exp->exp_imp_reverse->imp_remote_handle;
732 target = exp->exp_obd;
734 /* Might be a re-connect after a partition. */
735 if (!memcmp(&conn->cookie, &hdl->cookie, sizeof conn->cookie)) {
736 if (target->obd_recovering) {
737 int timeout = cfs_duration_sec(cfs_time_sub(
739 &target->obd_recovery_timer),
740 cfs_time_current()));
742 LCONSOLE_WARN("%s: Client %s (at %s) reconnect"
743 "ing, waiting for %d clients in recov"
744 "ery for %d:%.02d\n", target->obd_name,
745 obd_uuid2str(&exp->exp_client_uuid),
746 obd_export_nid2str(exp),
747 target->obd_max_recoverable_clients,
748 timeout / 60, timeout % 60);
750 LCONSOLE_WARN("%s: Client %s (at %s) "
751 "reconnecting\n", target->obd_name,
752 obd_uuid2str(&exp->exp_client_uuid),
753 obd_export_nid2str(exp));
756 conn->cookie = exp->exp_handle.h_cookie;
757 /* target_handle_connect() treats EALREADY and
758 * -EALREADY differently. EALREADY means we are
759 * doing a valid reconnect from the same client. */
762 LCONSOLE_WARN("%s: already connected client %s (at %s) "
763 "with handle "LPX64". Rejecting client "
764 "with the same UUID trying to reconnect "
765 "with handle "LPX64"\n", target->obd_name,
766 obd_uuid2str(&exp->exp_client_uuid),
767 obd_export_nid2str(exp),
768 hdl->cookie, conn->cookie);
769 memset(conn, 0, sizeof *conn);
770 /* target_handle_connect() treats EALREADY and
771 * -EALREADY differently. -EALREADY is an error
772 * (same UUID, different handle). */
777 conn->cookie = exp->exp_handle.h_cookie;
778 CDEBUG(D_HA, "connect export for UUID '%s' at %p, cookie "LPX64"\n",
779 cluuid->uuid, exp, conn->cookie);
783 void target_client_add_cb(struct obd_device *obd, __u64 transno, void *cb_data,
786 struct obd_export *exp = cb_data;
788 CDEBUG(D_RPCTRACE, "%s: committing for initial connect of %s\n",
789 obd->obd_name, exp->exp_client_uuid.uuid);
791 spin_lock(&exp->exp_lock);
792 exp->exp_need_sync = 0;
793 spin_unlock(&exp->exp_lock);
794 class_export_cb_put(exp);
798 check_and_start_recovery_timer(struct obd_device *obd,
799 struct ptlrpc_request *req, int new_client);
801 int target_handle_connect(struct ptlrpc_request *req)
803 struct obd_device *target = NULL, *targref = NULL;
804 struct obd_export *export = NULL;
805 struct obd_import *revimp;
806 struct obd_import *tmp_imp = NULL;
807 struct lustre_handle conn;
808 struct lustre_handle *tmp;
809 struct obd_uuid tgtuuid;
810 struct obd_uuid cluuid;
811 struct obd_uuid remote_uuid;
816 bool mds_conn = false, lw_client = false;
817 bool mds_mds_conn = false;
818 bool new_mds_mds_conn = false;
819 struct obd_connect_data *data, *tmpdata;
821 lnet_nid_t *client_nid = NULL;
824 OBD_RACE(OBD_FAIL_TGT_CONN_RACE);
826 str = req_capsule_client_get(&req->rq_pill, &RMF_TGTUUID);
828 DEBUG_REQ(D_ERROR, req, "bad target UUID for connect");
829 GOTO(out, rc = -EINVAL);
832 obd_str2uuid(&tgtuuid, str);
833 target = class_uuid2obd(&tgtuuid);
835 target = class_name2obd(str);
838 deuuidify(str, NULL, &target_start, &target_len);
839 LCONSOLE_ERROR_MSG(0x137, "%s: not available for connect "
840 "from %s (no target). If you are running "
841 "an HA pair check that the target is "
842 "mounted on the other server.\n", str,
843 libcfs_nid2str(req->rq_peer.nid));
844 GOTO(out, rc = -ENODEV);
847 spin_lock(&target->obd_dev_lock);
848 if (target->obd_stopping || !target->obd_set_up) {
849 spin_unlock(&target->obd_dev_lock);
851 deuuidify(str, NULL, &target_start, &target_len);
852 LCONSOLE_INFO("%.*s: Not available for connect from %s (%s)\n",
853 target_len, target_start,
854 libcfs_nid2str(req->rq_peer.nid),
855 (target->obd_stopping ?
856 "stopping" : "not set up"));
857 GOTO(out, rc = -ENODEV);
860 if (target->obd_no_conn) {
861 spin_unlock(&target->obd_dev_lock);
863 CDEBUG(D_INFO, "%s: Temporarily refusing client connection "
864 "from %s\n", target->obd_name,
865 libcfs_nid2str(req->rq_peer.nid));
866 GOTO(out, rc = -EAGAIN);
869 /* Make sure the target isn't cleaned up while we're here. Yes,
870 * there's still a race between the above check and our incref here.
871 * Really, class_uuid2obd should take the ref. */
872 targref = class_incref(target, __FUNCTION__, current);
874 target->obd_conn_inprogress++;
875 spin_unlock(&target->obd_dev_lock);
877 str = req_capsule_client_get(&req->rq_pill, &RMF_CLUUID);
879 DEBUG_REQ(D_ERROR, req, "bad client UUID for connect");
880 GOTO(out, rc = -EINVAL);
883 obd_str2uuid(&cluuid, str);
885 /* XXX Extract a nettype and format accordingly. */
886 switch (sizeof(lnet_nid_t)) {
887 /* NB the casts only avoid compiler warnings. */
889 snprintf(remote_uuid.uuid, sizeof remote_uuid,
890 "NET_"LPX64"_UUID", (__u64)req->rq_peer.nid);
893 snprintf(remote_uuid.uuid, sizeof remote_uuid,
894 "NET_%x_UUID", (__u32)req->rq_peer.nid);
900 tmp = req_capsule_client_get(&req->rq_pill, &RMF_CONN);
902 GOTO(out, rc = -EPROTO);
906 size = req_capsule_get_size(&req->rq_pill, &RMF_CONNECT_DATA,
908 data = req_capsule_client_get(&req->rq_pill, &RMF_CONNECT_DATA);
910 GOTO(out, rc = -EPROTO);
912 rc = req_capsule_server_pack(&req->rq_pill);
916 #if LUSTRE_VERSION_CODE < OBD_OCD_VERSION(3, 0, 53, 0)
917 /* Don't allow clients to connect that are using old 1.8 format
918 * protocol conventions (LUSTRE_MSG_MAGIC_v1, !MSGHDR_CKSUM_INCOMPAT18,
919 * ldlm_flock_policy_wire format, MDT_ATTR_xTIME_SET, etc). The
920 * FULL20 flag should be set on all connections since 2.0, but no
921 * longer affects behaviour.
923 * Later this check will be disabled and the flag can be retired
924 * completely once interop with 3.0 is no longer needed.
926 if (!(data->ocd_connect_flags & OBD_CONNECT_FULL20))
927 GOTO(out, rc = -EPROTO);
930 if (lustre_msg_get_op_flags(req->rq_reqmsg) & MSG_CONNECT_LIBCLIENT) {
931 if (data->ocd_version < LUSTRE_VERSION_CODE -
932 LUSTRE_VERSION_ALLOWED_OFFSET ||
933 data->ocd_version > LUSTRE_VERSION_CODE +
934 LUSTRE_VERSION_ALLOWED_OFFSET) {
935 DEBUG_REQ(D_WARNING, req, "Refusing %s (%d.%d.%d.%d) "
936 "libclient connection attempt",
937 data->ocd_version < LUSTRE_VERSION_CODE ?
939 OBD_OCD_VERSION_MAJOR(data->ocd_version),
940 OBD_OCD_VERSION_MINOR(data->ocd_version),
941 OBD_OCD_VERSION_PATCH(data->ocd_version),
942 OBD_OCD_VERSION_FIX(data->ocd_version));
943 data = req_capsule_server_sized_get(&req->rq_pill,
945 offsetof(typeof(*data), ocd_version) +
946 sizeof(data->ocd_version));
948 data->ocd_connect_flags = OBD_CONNECT_VERSION;
949 data->ocd_version = LUSTRE_VERSION_CODE;
951 GOTO(out, rc = -EPROTO);
955 /* Note: lw_client is needed in MDS-MDS failover during update log
956 * processing, so we needs to allow lw_client to be connected at
957 * anytime, instead of only the initial connection */
958 lw_client = (data->ocd_connect_flags & OBD_CONNECT_LIGHTWEIGHT) != 0;
960 if (lustre_msg_get_op_flags(req->rq_reqmsg) & MSG_CONNECT_INITIAL) {
961 mds_conn = (data->ocd_connect_flags & OBD_CONNECT_MDS) != 0;
962 mds_mds_conn = (data->ocd_connect_flags &
963 OBD_CONNECT_MDS_MDS) != 0;
965 /* OBD_CONNECT_MNE_SWAB is defined as OBD_CONNECT_MDS_MDS
966 * for Imperative Recovery connection from MGC to MGS.
968 * Via check OBD_CONNECT_FID, we can distinguish whether
969 * the OBD_CONNECT_MDS_MDS/OBD_CONNECT_MNE_SWAB is from
972 (data->ocd_connect_flags & OBD_CONNECT_MDS_MDS) &&
973 (data->ocd_connect_flags & OBD_CONNECT_FID) &&
974 (data->ocd_connect_flags & OBD_CONNECT_VERSION)) {
975 __u32 major = OBD_OCD_VERSION_MAJOR(data->ocd_version);
976 __u32 minor = OBD_OCD_VERSION_MINOR(data->ocd_version);
977 __u32 patch = OBD_OCD_VERSION_PATCH(data->ocd_version);
979 /* We do not support the MDT-MDT interoperations with
980 * different version MDT because of protocol changes. */
981 if (unlikely(major != LUSTRE_MAJOR ||
982 minor != LUSTRE_MINOR ||
983 abs(patch - LUSTRE_PATCH) > 3)) {
984 LCONSOLE_WARN("%s (%u.%u.%u.%u) refused the "
985 "connection from different version MDT "
986 "(%d.%d.%d.%d) %s %s\n",
987 target->obd_name, LUSTRE_MAJOR,
988 LUSTRE_MINOR, LUSTRE_PATCH, LUSTRE_FIX,
990 OBD_OCD_VERSION_FIX(data->ocd_version),
991 libcfs_nid2str(req->rq_peer.nid), str);
993 GOTO(out, rc = -EPROTO);
998 /* lctl gets a backstage, all-access pass. */
999 if (obd_uuid_equals(&cluuid, &target->obd_uuid))
1000 goto dont_check_exports;
1002 export = cfs_hash_lookup(target->obd_uuid_hash, &cluuid);
1006 /* We've found an export in the hash. */
1008 spin_lock(&export->exp_lock);
1010 if (export->exp_connecting) { /* bug 9635, et. al. */
1011 spin_unlock(&export->exp_lock);
1012 LCONSOLE_WARN("%s: Export %p already connecting from %s\n",
1013 export->exp_obd->obd_name, export,
1014 libcfs_nid2str(req->rq_peer.nid));
1015 class_export_put(export);
1018 } else if ((mds_conn || lw_client) && export->exp_connection != NULL) {
1019 spin_unlock(&export->exp_lock);
1020 if (req->rq_peer.nid != export->exp_connection->c_peer.nid)
1021 /* MDS or LWP reconnected after failover. */
1022 LCONSOLE_WARN("%s: Received %s connection from "
1023 "%s, removing former export from %s\n",
1024 target->obd_name, mds_conn ? "MDS" : "LWP",
1025 libcfs_nid2str(req->rq_peer.nid),
1026 libcfs_nid2str(export->exp_connection->c_peer.nid));
1028 /* New MDS connection from the same NID. */
1029 LCONSOLE_WARN("%s: Received new %s connection from "
1030 "%s, removing former export from same NID\n",
1031 target->obd_name, mds_conn ? "MDS" : "LWP",
1032 libcfs_nid2str(req->rq_peer.nid));
1033 class_fail_export(export);
1034 class_export_put(export);
1037 } else if (export->exp_connection != NULL &&
1038 req->rq_peer.nid != export->exp_connection->c_peer.nid &&
1039 (lustre_msg_get_op_flags(req->rq_reqmsg) &
1040 MSG_CONNECT_INITIAL)) {
1041 spin_unlock(&export->exp_lock);
1042 /* In MDS failover we have static UUID but NID can change. */
1043 LCONSOLE_WARN("%s: Client %s seen on new nid %s when "
1044 "existing nid %s is already connected\n",
1045 target->obd_name, cluuid.uuid,
1046 libcfs_nid2str(req->rq_peer.nid),
1048 export->exp_connection->c_peer.nid));
1050 class_export_put(export);
1053 export->exp_connecting = 1;
1054 spin_unlock(&export->exp_lock);
1055 LASSERT(export->exp_obd == target);
1057 rc = target_handle_reconnect(&conn, export, &cluuid);
1060 /* If we found an export, we already unlocked. */
1063 OBD_FAIL_TIMEOUT(OBD_FAIL_TGT_DELAY_CONNECT, 2 * obd_timeout);
1064 } else if (req->rq_export == NULL &&
1065 atomic_read(&export->exp_rpc_count) > 0) {
1066 LCONSOLE_WARN("%s: Client %s (at %s) refused connection, "
1067 "still busy with %d references\n",
1068 target->obd_name, cluuid.uuid,
1069 libcfs_nid2str(req->rq_peer.nid),
1070 atomic_read(&export->exp_refcount));
1071 GOTO(out, rc = -EBUSY);
1072 } else if (lustre_msg_get_conn_cnt(req->rq_reqmsg) == 1) {
1073 if (!strstr(cluuid.uuid, "mdt"))
1074 LCONSOLE_WARN("%s: Rejecting reconnect from the "
1075 "known client %s (at %s) because it "
1076 "is indicating it is a new client",
1077 target->obd_name, cluuid.uuid,
1078 libcfs_nid2str(req->rq_peer.nid));
1079 GOTO(out, rc = -EALREADY);
1081 OBD_FAIL_TIMEOUT(OBD_FAIL_TGT_DELAY_RECONNECT, 2 * obd_timeout);
1088 CDEBUG(D_HA, "%s: connection from %s@%s %st"LPU64" exp %p cur %ld last %ld\n",
1089 target->obd_name, cluuid.uuid, libcfs_nid2str(req->rq_peer.nid),
1090 target->obd_recovering ? "recovering/" : "", data->ocd_transno,
1091 export, (long)cfs_time_current_sec(),
1092 export ? (long)export->exp_last_request_time : 0);
1094 /* If this is the first time a client connects, reset the recovery
1095 * timer. Discard lightweight connections which might be local. */
1096 if (!lw_client && rc == 0 && target->obd_recovering)
1097 check_and_start_recovery_timer(target, req, export == NULL);
1099 /* We want to handle EALREADY but *not* -EALREADY from
1100 * target_handle_reconnect(), return reconnection state in a flag. */
1101 if (rc == EALREADY) {
1102 lustre_msg_add_op_flags(req->rq_repmsg, MSG_CONNECT_RECONNECT);
1108 /* Tell the client if we support replayable requests. */
1109 if (target->obd_replayable)
1110 lustre_msg_add_op_flags(req->rq_repmsg, MSG_CONNECT_REPLAYABLE);
1111 client_nid = &req->rq_peer.nid;
1113 if (export == NULL) {
1114 /* allow lightweight connections during recovery */
1115 /* allow "new" MDT to be connected during recovery, since we
1116 * need retrieve recovery update records from it */
1117 if (target->obd_recovering && !lw_client && !mds_mds_conn) {
1119 int c; /* connected */
1120 int i; /* in progress */
1122 int s; /* stale/evicted */
1124 c = atomic_read(&target->obd_connected_clients);
1125 i = atomic_read(&target->obd_lock_replay_clients);
1126 k = target->obd_max_recoverable_clients;
1127 s = target->obd_stale_clients;
1128 t = cfs_timer_deadline(&target->obd_recovery_timer);
1129 t = cfs_time_sub(t, cfs_time_current());
1130 t = cfs_duration_sec(t);
1131 LCONSOLE_WARN("%s: Denying connection for new client %s"
1132 "(at %s), waiting for %d known clients "
1133 "(%d recovered, %d in progress, and %d "
1134 "evicted) to recover in %d:%.02d\n",
1135 target->obd_name, cluuid.uuid,
1136 libcfs_nid2str(req->rq_peer.nid), k,
1137 c - i, i, s, (int)t / 60,
1142 rc = obd_connect(req->rq_svc_thread->t_env,
1143 &export, target, &cluuid, data,
1145 if (mds_conn && OBD_FAIL_CHECK(OBD_FAIL_TGT_RCVG_FLAG))
1146 lustre_msg_add_op_flags(req->rq_repmsg,
1147 MSG_CONNECT_RECOVERING);
1149 conn.cookie = export->exp_handle.h_cookie;
1152 new_mds_mds_conn = true;
1155 rc = obd_reconnect(req->rq_svc_thread->t_env,
1156 export, target, &cluuid, data, client_nid);
1161 LASSERT(target->u.obt.obt_magic == OBT_MAGIC);
1162 data->ocd_instance = target->u.obt.obt_instance;
1164 /* Return only the parts of obd_connect_data that we understand, so the
1165 * client knows that we don't understand the rest. */
1167 tmpsize = req_capsule_get_size(&req->rq_pill, &RMF_CONNECT_DATA,
1169 tmpdata = req_capsule_server_get(&req->rq_pill,
1171 /* Don't use struct assignment here, because the client reply
1172 * buffer may be smaller/larger than the local struct
1173 * obd_connect_data. */
1174 memcpy(tmpdata, data, min(tmpsize, size));
1177 /* If all else goes well, this is our RPC return code. */
1180 lustre_msg_set_handle(req->rq_repmsg, &conn);
1182 /* If the client and the server are the same node, we will already
1183 * have an export that really points to the client's DLM export,
1184 * because we have a shared handles table.
1186 * XXX this will go away when shaver stops sending the "connect" handle
1187 * in the real "remote handle" field of the request --phik 24 Apr 2003
1189 ptlrpc_request_change_export(req, export);
1191 spin_lock(&export->exp_lock);
1192 if (export->exp_conn_cnt >= lustre_msg_get_conn_cnt(req->rq_reqmsg)) {
1193 spin_unlock(&export->exp_lock);
1194 CDEBUG(D_RPCTRACE, "%s: %s already connected at greater "
1195 "or equal conn_cnt: %d >= %d\n",
1196 cluuid.uuid, libcfs_nid2str(req->rq_peer.nid),
1197 export->exp_conn_cnt,
1198 lustre_msg_get_conn_cnt(req->rq_reqmsg));
1200 GOTO(out, rc = -EALREADY);
1202 LASSERT(lustre_msg_get_conn_cnt(req->rq_reqmsg) > 0);
1203 export->exp_conn_cnt = lustre_msg_get_conn_cnt(req->rq_reqmsg);
1205 /* Don't evict liblustre clients for not pinging. */
1206 if (lustre_msg_get_op_flags(req->rq_reqmsg) & MSG_CONNECT_LIBCLIENT) {
1207 export->exp_libclient = 1;
1208 spin_unlock(&export->exp_lock);
1210 spin_lock(&target->obd_dev_lock);
1211 list_del_init(&export->exp_obd_chain_timed);
1212 spin_unlock(&target->obd_dev_lock);
1214 spin_unlock(&export->exp_lock);
1217 if (export->exp_connection != NULL) {
1218 /* Check to see if connection came from another NID. */
1219 if ((export->exp_connection->c_peer.nid != req->rq_peer.nid) &&
1220 !hlist_unhashed(&export->exp_nid_hash))
1221 cfs_hash_del(export->exp_obd->obd_nid_hash,
1222 &export->exp_connection->c_peer.nid,
1223 &export->exp_nid_hash);
1225 ptlrpc_connection_put(export->exp_connection);
1228 export->exp_connection = ptlrpc_connection_get(req->rq_peer,
1231 if (hlist_unhashed(&export->exp_nid_hash)) {
1232 cfs_hash_add(export->exp_obd->obd_nid_hash,
1233 &export->exp_connection->c_peer.nid,
1234 &export->exp_nid_hash);
1237 if (target->obd_recovering && !export->exp_in_recovery && !lw_client) {
1239 __u64 transno = data->ocd_transno;
1241 spin_lock(&export->exp_lock);
1242 /* possible race with class_disconnect_stale_exports,
1243 * export may be already in the eviction process */
1244 if (export->exp_failed) {
1245 spin_unlock(&export->exp_lock);
1246 GOTO(out, rc = -ENODEV);
1248 export->exp_in_recovery = 1;
1249 export->exp_req_replay_needed = 1;
1250 export->exp_lock_replay_needed = 1;
1251 spin_unlock(&export->exp_lock);
1253 has_transno = !!(lustre_msg_get_op_flags(req->rq_reqmsg) &
1254 MSG_CONNECT_TRANSNO);
1255 if (has_transno && transno == 0)
1256 CWARN("Connect with zero transno!\n");
1258 if (has_transno && transno > 0 &&
1259 transno < target->obd_next_recovery_transno &&
1260 transno > target->obd_last_committed) {
1261 /* Another way is to use cmpxchg() to be lock-free. */
1262 spin_lock(&target->obd_recovery_task_lock);
1263 if (transno < target->obd_next_recovery_transno)
1264 target->obd_next_recovery_transno = transno;
1265 spin_unlock(&target->obd_recovery_task_lock);
1268 atomic_inc(&target->obd_req_replay_clients);
1269 atomic_inc(&target->obd_lock_replay_clients);
1270 /* Note: MDS-MDS connection is allowed to be connected during
1271 * recovery, no matter if the exports needs to be recoveried.
1272 * Because we need retrieve updates logs from all other MDTs.
1273 * So if the MDS-MDS export is new, obd_max_recoverable_clients
1274 * also needs to be increased to match other recovery checking
1276 if (new_mds_mds_conn)
1277 target->obd_max_recoverable_clients++;
1278 if (atomic_inc_return(&target->obd_connected_clients) ==
1279 target->obd_max_recoverable_clients)
1280 wake_up(&target->obd_next_transno_waitq);
1283 /* Tell the client we're in recovery, when client is involved in it. */
1284 if (target->obd_recovering && !lw_client)
1285 lustre_msg_add_op_flags(req->rq_repmsg, MSG_CONNECT_RECOVERING);
1287 tmp = req_capsule_client_get(&req->rq_pill, &RMF_CONN);
1290 /* Return -ENOTCONN in case of errors to let client reconnect. */
1291 revimp = class_new_import(target);
1292 if (revimp == NULL) {
1293 CERROR("fail to alloc new reverse import.\n");
1294 GOTO(out, rc = -ENOTCONN);
1297 spin_lock(&export->exp_lock);
1298 if (export->exp_imp_reverse != NULL)
1299 /* destroyed import can be still referenced in ctxt */
1300 tmp_imp = export->exp_imp_reverse;
1301 export->exp_imp_reverse = revimp;
1302 spin_unlock(&export->exp_lock);
1304 revimp->imp_connection = ptlrpc_connection_addref(export->exp_connection);
1305 revimp->imp_client = &export->exp_obd->obd_ldlm_client;
1306 revimp->imp_remote_handle = conn;
1307 revimp->imp_dlm_fake = 1;
1308 revimp->imp_state = LUSTRE_IMP_FULL;
1310 /* Unknown versions will be caught in
1311 * ptlrpc_handle_server_req_in->lustre_unpack_msg(). */
1312 revimp->imp_msg_magic = req->rq_reqmsg->lm_magic;
1314 if (data->ocd_connect_flags & OBD_CONNECT_AT)
1315 revimp->imp_msghdr_flags |= MSGHDR_AT_SUPPORT;
1317 revimp->imp_msghdr_flags &= ~MSGHDR_AT_SUPPORT;
1319 revimp->imp_msghdr_flags |= MSGHDR_CKSUM_INCOMPAT18;
1321 rc = sptlrpc_import_sec_adapt(revimp, req->rq_svc_ctx, &req->rq_flvr);
1323 CERROR("Failed to get sec for reverse import: %d\n", rc);
1324 spin_lock(&export->exp_lock);
1325 export->exp_imp_reverse = NULL;
1326 spin_unlock(&export->exp_lock);
1327 class_destroy_import(revimp);
1330 class_import_put(revimp);
1333 if (tmp_imp != NULL)
1334 client_destroy_import(tmp_imp);
1336 spin_lock(&export->exp_lock);
1337 export->exp_connecting = 0;
1338 spin_unlock(&export->exp_lock);
1340 class_export_put(export);
1343 spin_lock(&target->obd_dev_lock);
1344 target->obd_conn_inprogress--;
1345 spin_unlock(&target->obd_dev_lock);
1347 class_decref(targref, __func__, current);
1350 req->rq_status = rc;
1354 int target_handle_disconnect(struct ptlrpc_request *req)
1359 rc = req_capsule_server_pack(&req->rq_pill);
1363 /* Keep the rq_export around so we can send the reply. */
1364 req->rq_status = obd_disconnect(class_export_get(req->rq_export));
1369 void target_destroy_export(struct obd_export *exp)
1371 struct obd_import *imp = NULL;
1372 /* exports created from last_rcvd data, and "fake"
1373 exports created by lctl don't have an import */
1374 spin_lock(&exp->exp_lock);
1375 if (exp->exp_imp_reverse != NULL) {
1376 imp = exp->exp_imp_reverse;
1377 exp->exp_imp_reverse = NULL;
1379 spin_unlock(&exp->exp_lock);
1381 client_destroy_import(imp);
1383 LASSERT_ATOMIC_ZERO(&exp->exp_locks_count);
1384 LASSERT_ATOMIC_ZERO(&exp->exp_rpc_count);
1385 LASSERT_ATOMIC_ZERO(&exp->exp_cb_count);
1386 LASSERT_ATOMIC_ZERO(&exp->exp_replay_count);
1388 EXPORT_SYMBOL(target_destroy_export);
1391 * Recovery functions
1393 static void target_request_copy_get(struct ptlrpc_request *req)
1395 class_export_rpc_inc(req->rq_export);
1396 LASSERT(list_empty(&req->rq_list));
1397 INIT_LIST_HEAD(&req->rq_replay_list);
1399 /* Increase refcount to keep request in queue. */
1400 atomic_inc(&req->rq_refcount);
1401 /* Let export know it has replays to be handled. */
1402 atomic_inc(&req->rq_export->exp_replay_count);
1405 static void target_request_copy_put(struct ptlrpc_request *req)
1407 LASSERT(list_empty(&req->rq_replay_list));
1408 LASSERT_ATOMIC_POS(&req->rq_export->exp_replay_count);
1410 atomic_dec(&req->rq_export->exp_replay_count);
1411 class_export_rpc_dec(req->rq_export);
1412 ptlrpc_server_drop_request(req);
1415 static int target_exp_enqueue_req_replay(struct ptlrpc_request *req)
1417 __u64 transno = lustre_msg_get_transno(req->rq_reqmsg);
1418 struct obd_export *exp = req->rq_export;
1419 struct ptlrpc_request *reqiter;
1424 spin_lock(&exp->exp_lock);
1425 list_for_each_entry(reqiter, &exp->exp_req_replay_queue,
1427 if (lustre_msg_get_transno(reqiter->rq_reqmsg) == transno) {
1434 /* We expect it with RESENT and REPLAY flags. */
1435 if ((lustre_msg_get_flags(req->rq_reqmsg) &
1436 (MSG_RESENT | MSG_REPLAY)) != (MSG_RESENT | MSG_REPLAY))
1437 CERROR("invalid flags %x of resent replay\n",
1438 lustre_msg_get_flags(req->rq_reqmsg));
1440 list_add_tail(&req->rq_replay_list,
1441 &exp->exp_req_replay_queue);
1444 spin_unlock(&exp->exp_lock);
1448 static void target_exp_dequeue_req_replay(struct ptlrpc_request *req)
1450 LASSERT(!list_empty(&req->rq_replay_list));
1451 LASSERT(req->rq_export);
1453 spin_lock(&req->rq_export->exp_lock);
1454 list_del_init(&req->rq_replay_list);
1455 spin_unlock(&req->rq_export->exp_lock);
1458 static void target_finish_recovery(struct lu_target *lut)
1460 struct obd_device *obd = lut->lut_obd;
1463 /* Only log a recovery message when recovery has occurred. */
1464 if (obd->obd_recovery_start) {
1465 time_t elapsed_time = max_t(time_t, 1, cfs_time_current_sec() -
1466 obd->obd_recovery_start);
1467 LCONSOLE_INFO("%s: Recovery over after %d:%.02d, of %d clients "
1468 "%d recovered and %d %s evicted.\n", obd->obd_name,
1469 (int)elapsed_time / 60, (int)elapsed_time % 60,
1470 obd->obd_max_recoverable_clients,
1471 atomic_read(&obd->obd_connected_clients),
1472 obd->obd_stale_clients,
1473 obd->obd_stale_clients == 1 ? "was" : "were");
1476 ldlm_reprocess_all_ns(obd->obd_namespace);
1477 spin_lock(&obd->obd_recovery_task_lock);
1478 if (!list_empty(&obd->obd_req_replay_queue) ||
1479 !list_empty(&obd->obd_lock_replay_queue) ||
1480 !list_empty(&obd->obd_final_req_queue)) {
1481 CERROR("%s: Recovery queues ( %s%s%s) are not empty\n",
1483 list_empty(&obd->obd_req_replay_queue) ? "" : "req ",
1484 list_empty(&obd->obd_lock_replay_queue) ? \
1486 list_empty(&obd->obd_final_req_queue) ? \
1488 spin_unlock(&obd->obd_recovery_task_lock);
1491 spin_unlock(&obd->obd_recovery_task_lock);
1493 if (lut->lut_tdtd != NULL &&
1494 !list_empty(&lut->lut_tdtd->tdtd_replay_list))
1495 dtrq_list_dump(lut->lut_tdtd, D_ERROR);
1497 obd->obd_recovery_end = cfs_time_current_sec();
1499 /* When recovery finished, cleanup orphans on MDS and OST. */
1500 if (OBT(obd) && OBP(obd, postrecov)) {
1501 int rc = OBP(obd, postrecov)(obd);
1503 LCONSOLE_WARN("%s: Post recovery failed, rc %d\n",
1509 static void abort_req_replay_queue(struct obd_device *obd)
1511 struct ptlrpc_request *req, *n;
1512 struct list_head abort_list;
1514 INIT_LIST_HEAD(&abort_list);
1515 spin_lock(&obd->obd_recovery_task_lock);
1516 list_splice_init(&obd->obd_req_replay_queue, &abort_list);
1517 spin_unlock(&obd->obd_recovery_task_lock);
1518 list_for_each_entry_safe(req, n, &abort_list, rq_list) {
1519 DEBUG_REQ(D_WARNING, req, "aborted:");
1520 req->rq_status = -ENOTCONN;
1521 if (ptlrpc_error(req)) {
1522 DEBUG_REQ(D_ERROR, req,
1523 "failed abort_req_reply; skipping");
1525 target_exp_dequeue_req_replay(req);
1526 target_request_copy_put(req);
1530 static void abort_lock_replay_queue(struct obd_device *obd)
1532 struct ptlrpc_request *req, *n;
1533 struct list_head abort_list;
1535 INIT_LIST_HEAD(&abort_list);
1536 spin_lock(&obd->obd_recovery_task_lock);
1537 list_splice_init(&obd->obd_lock_replay_queue, &abort_list);
1538 spin_unlock(&obd->obd_recovery_task_lock);
1539 list_for_each_entry_safe(req, n, &abort_list, rq_list) {
1540 DEBUG_REQ(D_ERROR, req, "aborted:");
1541 req->rq_status = -ENOTCONN;
1542 if (ptlrpc_error(req)) {
1543 DEBUG_REQ(D_ERROR, req,
1544 "failed abort_lock_reply; skipping");
1546 target_request_copy_put(req);
1550 /* Called from a cleanup function if the device is being cleaned up
1551 forcefully. The exports should all have been disconnected already,
1552 the only thing left to do is
1553 - clear the recovery flags
1555 - free queued requests and replies, but don't send replies
1556 Because the obd_stopping flag is set, no new requests should be received.
1559 void target_cleanup_recovery(struct obd_device *obd)
1561 struct ptlrpc_request *req, *n;
1562 struct list_head clean_list;
1565 INIT_LIST_HEAD(&clean_list);
1566 spin_lock(&obd->obd_dev_lock);
1567 if (!obd->obd_recovering) {
1568 spin_unlock(&obd->obd_dev_lock);
1572 obd->obd_recovering = obd->obd_abort_recovery = 0;
1573 obd->obd_force_abort_recovery = 0;
1574 spin_unlock(&obd->obd_dev_lock);
1576 spin_lock(&obd->obd_recovery_task_lock);
1577 target_cancel_recovery_timer(obd);
1578 list_splice_init(&obd->obd_req_replay_queue, &clean_list);
1579 spin_unlock(&obd->obd_recovery_task_lock);
1581 list_for_each_entry_safe(req, n, &clean_list, rq_list) {
1582 LASSERT(req->rq_reply_state == NULL);
1583 target_exp_dequeue_req_replay(req);
1584 target_request_copy_put(req);
1587 spin_lock(&obd->obd_recovery_task_lock);
1588 list_splice_init(&obd->obd_lock_replay_queue, &clean_list);
1589 list_splice_init(&obd->obd_final_req_queue, &clean_list);
1590 spin_unlock(&obd->obd_recovery_task_lock);
1592 list_for_each_entry_safe(req, n, &clean_list, rq_list) {
1593 LASSERT(req->rq_reply_state == NULL);
1594 target_request_copy_put(req);
1599 EXPORT_SYMBOL(target_cleanup_recovery);
1601 /* obd_recovery_task_lock should be held */
1602 void target_cancel_recovery_timer(struct obd_device *obd)
1604 CDEBUG(D_HA, "%s: cancel recovery timer\n", obd->obd_name);
1605 cfs_timer_disarm(&obd->obd_recovery_timer);
1608 static void target_start_recovery_timer(struct obd_device *obd)
1610 if (obd->obd_recovery_start != 0)
1613 spin_lock(&obd->obd_dev_lock);
1614 if (!obd->obd_recovering || obd->obd_abort_recovery ||
1615 obd->obd_force_abort_recovery) {
1616 spin_unlock(&obd->obd_dev_lock);
1620 LASSERT(obd->obd_recovery_timeout != 0);
1622 if (obd->obd_recovery_start != 0) {
1623 spin_unlock(&obd->obd_dev_lock);
1627 cfs_timer_arm(&obd->obd_recovery_timer,
1628 cfs_time_shift(obd->obd_recovery_timeout));
1629 obd->obd_recovery_start = cfs_time_current_sec();
1630 spin_unlock(&obd->obd_dev_lock);
1632 LCONSOLE_WARN("%s: Will be in recovery for at least %d:%.02d, "
1633 "or until %d client%s reconnect%s\n",
1635 obd->obd_recovery_timeout / 60,
1636 obd->obd_recovery_timeout % 60,
1637 obd->obd_max_recoverable_clients,
1638 (obd->obd_max_recoverable_clients == 1) ? "" : "s",
1639 (obd->obd_max_recoverable_clients == 1) ? "s": "");
1643 * extend recovery window.
1645 * if @extend is true, extend recovery window to have @drt remaining at least;
1646 * otherwise, make sure the recovery timeout value is not less than @drt.
1648 static void extend_recovery_timer(struct obd_device *obd, int drt, bool extend)
1652 cfs_duration_t left;
1655 spin_lock(&obd->obd_dev_lock);
1656 if (!obd->obd_recovering || obd->obd_abort_recovery ||
1657 obd->obd_force_abort_recovery) {
1658 spin_unlock(&obd->obd_dev_lock);
1661 LASSERT(obd->obd_recovery_start != 0);
1663 now = cfs_time_current_sec();
1664 to = obd->obd_recovery_timeout;
1665 end = obd->obd_recovery_start + to;
1666 left = cfs_time_sub(end, now);
1668 if (extend && (drt > left)) {
1670 } else if (!extend && (drt > to)) {
1674 if (to > obd->obd_recovery_time_hard)
1675 to = obd->obd_recovery_time_hard;
1676 if (obd->obd_recovery_timeout < to) {
1677 obd->obd_recovery_timeout = to;
1678 end = obd->obd_recovery_start + to;
1679 cfs_timer_arm(&obd->obd_recovery_timer,
1680 cfs_time_shift(end - now));
1682 spin_unlock(&obd->obd_dev_lock);
1684 CDEBUG(D_HA, "%s: recovery timer will expire in %u seconds\n",
1685 obd->obd_name, (unsigned)cfs_time_sub(end, now));
1688 /* Reset the timer with each new client connection */
1690 * This timer is actually reconnect_timer, which is for making sure
1691 * the total recovery window is at least as big as my reconnect
1692 * attempt timing. So the initial recovery time_out will be set to
1693 * OBD_RECOVERY_FACTOR * obd_timeout. If the timeout coming
1694 * from client is bigger than this, then the recovery time_out will
1695 * be extended to make sure the client could be reconnected, in the
1696 * process, the timeout from the new client should be ignored.
1700 check_and_start_recovery_timer(struct obd_device *obd,
1701 struct ptlrpc_request *req,
1704 int service_time = lustre_msg_get_service_time(req->rq_reqmsg);
1705 struct obd_device_target *obt = &obd->u.obt;
1707 if (!new_client && service_time)
1708 /* Teach server about old server's estimates, as first guess
1709 * at how long new requests will take. */
1710 at_measured(&req->rq_rqbd->rqbd_svcpt->scp_at_estimate,
1713 target_start_recovery_timer(obd);
1715 /* Convert the service time to RPC timeout,
1716 * and reuse service_time to limit stack usage. */
1717 service_time = at_est2timeout(service_time);
1719 /* We expect other clients to timeout within service_time, then try
1720 * to reconnect, then try the failover server. The max delay between
1721 * connect attempts is SWITCH_MAX + SWITCH_INC + INITIAL. */
1722 service_time += 2 * INITIAL_CONNECT_TIMEOUT;
1724 LASSERT(obt->obt_magic == OBT_MAGIC);
1725 service_time += 2 * (CONNECTION_SWITCH_MAX + CONNECTION_SWITCH_INC);
1726 if (service_time > obd->obd_recovery_timeout && !new_client)
1727 extend_recovery_timer(obd, service_time, false);
1730 /** Health checking routines */
1731 static inline int exp_connect_healthy(struct obd_export *exp)
1733 return (exp->exp_in_recovery);
1736 /** if export done req_replay or has replay in queue */
1737 static inline int exp_req_replay_healthy(struct obd_export *exp)
1739 return (!exp->exp_req_replay_needed ||
1740 atomic_read(&exp->exp_replay_count) > 0);
1742 /** if export done lock_replay or has replay in queue */
1743 static inline int exp_lock_replay_healthy(struct obd_export *exp)
1745 return (!exp->exp_lock_replay_needed ||
1746 atomic_read(&exp->exp_replay_count) > 0);
1749 static inline int exp_vbr_healthy(struct obd_export *exp)
1751 return (!exp->exp_vbr_failed);
1754 static inline int exp_finished(struct obd_export *exp)
1756 return (exp->exp_in_recovery && !exp->exp_lock_replay_needed);
1759 static int check_for_next_transno(struct lu_target *lut)
1761 struct ptlrpc_request *req = NULL;
1762 struct obd_device *obd = lut->lut_obd;
1763 int wake_up = 0, connected, completed, queue_len;
1764 __u64 req_transno = 0;
1765 __u64 update_transno = 0;
1766 __u64 next_transno = 0;
1769 spin_lock(&obd->obd_recovery_task_lock);
1770 if (!list_empty(&obd->obd_req_replay_queue)) {
1771 req = list_entry(obd->obd_req_replay_queue.next,
1772 struct ptlrpc_request, rq_list);
1773 req_transno = lustre_msg_get_transno(req->rq_reqmsg);
1776 if (lut->lut_tdtd != NULL) {
1777 struct target_distribute_txn_data *tdtd;
1779 tdtd = lut->lut_tdtd;
1780 update_transno = distribute_txn_get_next_transno(lut->lut_tdtd);
1783 connected = atomic_read(&obd->obd_connected_clients);
1784 completed = connected - atomic_read(&obd->obd_req_replay_clients);
1785 queue_len = obd->obd_requests_queued_for_recovery;
1786 next_transno = obd->obd_next_recovery_transno;
1788 CDEBUG(D_HA, "max: %d, connected: %d, completed: %d, queue_len: %d, "
1789 "req_transno: "LPU64", next_transno: "LPU64"\n",
1790 obd->obd_max_recoverable_clients, connected, completed,
1791 queue_len, req_transno, next_transno);
1793 if (obd->obd_abort_recovery || obd->obd_force_abort_recovery) {
1794 CDEBUG(D_HA, "waking for aborted recovery\n");
1796 } else if (obd->obd_recovery_expired) {
1797 CDEBUG(D_HA, "waking for expired recovery\n");
1799 } else if (req_transno == next_transno ||
1800 (update_transno != 0 && update_transno <= next_transno)) {
1801 CDEBUG(D_HA, "waking for next ("LPD64")\n", next_transno);
1803 } else if (queue_len > 0 &&
1804 queue_len == atomic_read(&obd->obd_req_replay_clients)) {
1806 /** handle gaps occured due to lost reply or VBR */
1807 LASSERTF(req_transno >= next_transno,
1808 "req_transno: "LPU64", next_transno: "LPU64"\n",
1809 req_transno, next_transno);
1810 if (req_transno > obd->obd_last_committed &&
1811 !obd->obd_version_recov)
1814 "%s: waking for gap in transno, VBR is %s (skip: "
1815 LPD64", ql: %d, comp: %d, conn: %d, next: "LPD64
1816 ", next_update "LPD64" last_committed: "LPD64")\n",
1817 obd->obd_name, obd->obd_version_recov ? "ON" : "OFF",
1818 next_transno, queue_len, completed, connected,
1819 req_transno, update_transno, obd->obd_last_committed);
1820 obd->obd_next_recovery_transno = req_transno;
1822 } else if (atomic_read(&obd->obd_req_replay_clients) == 0) {
1823 CDEBUG(D_HA, "waking for completed recovery\n");
1825 } else if (OBD_FAIL_CHECK(OBD_FAIL_MDS_RECOVERY_ACCEPTS_GAPS)) {
1826 CDEBUG(D_HA, "accepting transno gaps is explicitly allowed"
1827 " by fail_lock, waking up ("LPD64")\n", next_transno);
1828 obd->obd_next_recovery_transno = req_transno;
1831 spin_unlock(&obd->obd_recovery_task_lock);
1835 static int check_for_next_lock(struct lu_target *lut)
1837 struct obd_device *obd = lut->lut_obd;
1840 spin_lock(&obd->obd_recovery_task_lock);
1841 if (!list_empty(&obd->obd_lock_replay_queue)) {
1842 CDEBUG(D_HA, "waking for next lock\n");
1844 } else if (atomic_read(&obd->obd_lock_replay_clients) == 0) {
1845 CDEBUG(D_HA, "waking for completed lock replay\n");
1847 } else if (obd->obd_abort_recovery || obd->obd_force_abort_recovery) {
1848 CDEBUG(D_HA, "waking for aborted recovery\n");
1850 } else if (obd->obd_recovery_expired) {
1851 CDEBUG(D_HA, "waking for expired recovery\n");
1854 spin_unlock(&obd->obd_recovery_task_lock);
1860 * wait for recovery events,
1861 * check its status with help of check_routine
1862 * evict dead clients via health_check
1864 static int target_recovery_overseer(struct lu_target *lut,
1865 int (*check_routine)(struct lu_target *),
1866 int (*health_check)(struct obd_export *))
1868 struct obd_device *obd = lut->lut_obd;
1870 if ((obd->obd_recovery_start != 0) && (cfs_time_current_sec() >=
1871 (obd->obd_recovery_start + obd->obd_recovery_time_hard))) {
1872 CWARN("recovery is aborted by hard timeout\n");
1873 obd->obd_abort_recovery = 1;
1876 while (wait_event_timeout(obd->obd_next_transno_waitq,
1878 msecs_to_jiffies(60 * MSEC_PER_SEC)) == 0)
1879 /* wait indefinitely for event, but don't trigger watchdog */;
1881 if (obd->obd_abort_recovery || obd->obd_force_abort_recovery) {
1882 CWARN("recovery is aborted, evict exports in recovery\n");
1883 /** evict exports which didn't finish recovery yet */
1884 class_disconnect_stale_exports(obd, exp_finished);
1886 } else if (obd->obd_recovery_expired) {
1887 obd->obd_recovery_expired = 0;
1888 /** If some clients died being recovered, evict them */
1889 LCONSOLE_WARN("%s: recovery is timed out, "
1890 "evict stale exports\n", obd->obd_name);
1891 /** evict cexports with no replay in queue, they are stalled */
1892 class_disconnect_stale_exports(obd, health_check);
1893 /** continue with VBR */
1894 spin_lock(&obd->obd_dev_lock);
1895 obd->obd_version_recov = 1;
1896 spin_unlock(&obd->obd_dev_lock);
1898 * reset timer, recovery will proceed with versions now,
1899 * timeout is set just to handle reconnection delays
1901 extend_recovery_timer(obd, RECONNECT_DELAY_MAX, true);
1902 /** Wait for recovery events again, after evicting bad clients */
1908 static struct ptlrpc_request *target_next_replay_lock(struct lu_target *lut)
1910 struct obd_device *obd = lut->lut_obd;
1911 struct ptlrpc_request *req = NULL;
1913 CDEBUG(D_HA, "Waiting for lock\n");
1914 if (target_recovery_overseer(lut, check_for_next_lock,
1915 exp_lock_replay_healthy))
1916 abort_lock_replay_queue(obd);
1918 spin_lock(&obd->obd_recovery_task_lock);
1919 if (!list_empty(&obd->obd_lock_replay_queue)) {
1920 req = list_entry(obd->obd_lock_replay_queue.next,
1921 struct ptlrpc_request, rq_list);
1922 list_del_init(&req->rq_list);
1923 spin_unlock(&obd->obd_recovery_task_lock);
1925 spin_unlock(&obd->obd_recovery_task_lock);
1926 LASSERT(list_empty(&obd->obd_lock_replay_queue));
1927 LASSERT(atomic_read(&obd->obd_lock_replay_clients) == 0);
1928 /** evict exports failed VBR */
1929 class_disconnect_stale_exports(obd, exp_vbr_healthy);
1934 static struct ptlrpc_request *target_next_final_ping(struct obd_device *obd)
1936 struct ptlrpc_request *req = NULL;
1938 spin_lock(&obd->obd_recovery_task_lock);
1939 if (!list_empty(&obd->obd_final_req_queue)) {
1940 req = list_entry(obd->obd_final_req_queue.next,
1941 struct ptlrpc_request, rq_list);
1942 list_del_init(&req->rq_list);
1943 spin_unlock(&obd->obd_recovery_task_lock);
1944 if (req->rq_export->exp_in_recovery) {
1945 spin_lock(&req->rq_export->exp_lock);
1946 req->rq_export->exp_in_recovery = 0;
1947 spin_unlock(&req->rq_export->exp_lock);
1950 spin_unlock(&obd->obd_recovery_task_lock);
1955 static void handle_recovery_req(struct ptlrpc_thread *thread,
1956 struct ptlrpc_request *req,
1957 svc_handler_t handler)
1962 * export can be evicted during recovery, no need to handle replays for
1963 * it after that, discard such request silently
1965 if (req->rq_export->exp_disconnected)
1968 req->rq_session.lc_thread = thread;
1969 req->rq_svc_thread = thread;
1970 req->rq_svc_thread->t_env->le_ses = &req->rq_session;
1972 /* thread context */
1973 lu_context_enter(&thread->t_env->le_ctx);
1975 lu_context_exit(&thread->t_env->le_ctx);
1977 /* don't reset timer for final stage */
1978 if (!exp_finished(req->rq_export)) {
1979 int to = obd_timeout;
1982 * Add request timeout to the recovery time so next request from
1983 * this client may come in recovery time
1986 struct ptlrpc_service_part *svcpt;
1988 svcpt = req->rq_rqbd->rqbd_svcpt;
1989 /* If the server sent early reply for this request,
1990 * the client will recalculate the timeout according to
1991 * current server estimate service time, so we will
1992 * use the maxium timeout here for waiting the client
1993 * sending the next req */
1994 to = max((int)at_est2timeout(
1995 at_get(&svcpt->scp_at_estimate)),
1996 (int)lustre_msg_get_timeout(req->rq_reqmsg));
1997 /* Add 2 net_latency, one for balance rq_deadline
1998 * (see ptl_send_rpc), one for resend the req to server,
1999 * Note: client will pack net_latency in replay req
2000 * (see ptlrpc_replay_req) */
2001 to += 2 * lustre_msg_get_service_time(req->rq_reqmsg);
2003 extend_recovery_timer(class_exp2obd(req->rq_export), to, true);
2008 /** Checking routines for recovery */
2009 static int check_for_recovery_ready(struct lu_target *lut)
2011 struct obd_device *obd = lut->lut_obd;
2012 unsigned int clnts = atomic_read(&obd->obd_connected_clients);
2014 CDEBUG(D_HA, "connected %d stale %d max_recoverable_clients %d"
2015 " abort %d expired %d\n", clnts, obd->obd_stale_clients,
2016 obd->obd_max_recoverable_clients, obd->obd_abort_recovery,
2017 obd->obd_recovery_expired);
2019 if (obd->obd_force_abort_recovery)
2022 if (!obd->obd_abort_recovery && !obd->obd_recovery_expired) {
2023 LASSERT(clnts <= obd->obd_max_recoverable_clients);
2024 if (clnts + obd->obd_stale_clients <
2025 obd->obd_max_recoverable_clients)
2029 if (lut->lut_tdtd != NULL) {
2030 if (!lut->lut_tdtd->tdtd_replay_ready) {
2031 /* Let's extend recovery timer, in case the recovery
2032 * timer expired, and some clients got evicted */
2033 extend_recovery_timer(obd, obd->obd_recovery_timeout,
2037 dtrq_list_dump(lut->lut_tdtd, D_HA);
2045 REQUEST_RECOVERY = 1,
2046 UPDATE_RECOVERY = 2,
2049 static __u64 get_next_replay_req_transno(struct obd_device *obd)
2053 if (!list_empty(&obd->obd_req_replay_queue)) {
2054 struct ptlrpc_request *req;
2056 req = list_entry(obd->obd_req_replay_queue.next,
2057 struct ptlrpc_request, rq_list);
2058 transno = lustre_msg_get_transno(req->rq_reqmsg);
2063 __u64 get_next_transno(struct lu_target *lut, int *type)
2065 struct obd_device *obd = lut->lut_obd;
2066 struct target_distribute_txn_data *tdtd = lut->lut_tdtd;
2068 __u64 update_transno;
2071 transno = get_next_replay_req_transno(obd);
2073 *type = REQUEST_RECOVERY;
2078 update_transno = distribute_txn_get_next_transno(tdtd);
2079 if (transno == 0 || (transno >= update_transno &&
2080 update_transno != 0)) {
2081 transno = update_transno;
2083 *type = UPDATE_RECOVERY;
2090 * drop duplicate replay request
2092 * Because the operation has been replayed by update recovery, the request
2093 * with the same transno will be dropped and also notify the client to send
2094 * next replay request.
2096 * \param[in] env execution environment
2097 * \param[in] obd failover obd device
2098 * \param[in] req request to be dropped
2100 static void drop_duplicate_replay_req(struct lu_env *env,
2101 struct obd_device *obd,
2102 struct ptlrpc_request *req)
2104 DEBUG_REQ(D_HA, req, "remove t"LPD64" from %s because of duplicate"
2105 " update records are found.\n",
2106 lustre_msg_get_transno(req->rq_reqmsg),
2107 libcfs_nid2str(req->rq_peer.nid));
2109 /* Right now, only for MDS reint operation update replay and
2110 * normal request replay can have the same transno */
2111 if (lustre_msg_get_opc(req->rq_reqmsg) == MDS_REINT) {
2112 req_capsule_set(&req->rq_pill, &RQF_MDS_REINT);
2113 req->rq_status = req_capsule_server_pack(&req->rq_pill);
2114 if (likely(req->rq_export))
2115 target_committed_to_req(req);
2116 lustre_msg_set_transno(req->rq_repmsg, req->rq_transno);
2117 target_send_reply(req, req->rq_status, 0);
2119 DEBUG_REQ(D_ERROR, req, "wrong opc" "from %s\n",
2120 libcfs_nid2str(req->rq_peer.nid));
2122 target_exp_dequeue_req_replay(req);
2123 target_request_copy_put(req);
2124 obd->obd_replayed_requests++;
2128 * Update last_rcvd of the update
2130 * Because update recovery might update the last_rcvd by updates, i.e.
2131 * it will not update the last_rcvd information in memory, so we need
2132 * refresh these information in memory after update recovery.
2134 * \param[in] obd obd_device under recoverying.
2135 * \param[in] dtrq the update replay requests being replayed.
2137 static void target_update_lcd(struct lu_env *env, struct lu_target *lut,
2138 struct distribute_txn_replay_req *dtrq)
2140 struct obd_device *obd = lut->lut_obd;
2141 struct obd_export *export;
2142 struct tg_export_data *ted;
2143 struct distribute_txn_replay_req_sub *dtrqs;
2144 struct seq_server_site *site;
2145 struct update_records *ur;
2146 const struct lu_fid *fid;
2147 struct update_ops *ops;
2148 struct update_params *params;
2149 struct update_op *op;
2152 struct lsd_client_data *lcd = NULL;
2154 /* if Updates has been executed(committed) on the recovery target,
2155 * i.e. the updates is not being executed on the target, so we do
2156 * not need update it in memory */
2157 site = lu_site2seq(obd->obd_lu_dev->ld_site);
2158 mdt_index = site->ss_node_id;
2159 dtrqs = dtrq_sub_lookup(dtrq, mdt_index);
2163 if (dtrq->dtrq_lur == NULL)
2166 /* Find the update last_rcvd record */
2167 fid = lu_object_fid(&lut->lut_last_rcvd->do_lu);
2168 ur = &dtrq->dtrq_lur->lur_update_rec;
2170 params = update_records_get_params(ur);
2171 for (i = 0, op = &ops->uops_op[0]; i < ur->ur_update_count;
2172 i++, op = update_op_next_op(op)) {
2177 if (!lu_fid_eq(&op->uop_fid, fid))
2180 if (op->uop_type != OUT_WRITE)
2183 buf = update_params_get_param_buf(params, op->uop_params_off[1],
2184 ur->ur_param_count, NULL);
2188 pos = le64_to_cpu(*(__u64 *)buf);
2192 buf = update_params_get_param_buf(params, op->uop_params_off[0],
2193 ur->ur_param_count, &size);
2197 if (size != sizeof(*lcd))
2202 if (lcd == NULL || lcd->lcd_uuid[0] == '\0')
2205 /* locate the export then update the exp_target_data if needed */
2206 export = cfs_hash_lookup(obd->obd_uuid_hash, lcd->lcd_uuid);
2210 ted = &export->exp_target_data;
2211 if (lcd->lcd_last_xid > ted->ted_lcd->lcd_last_xid) {
2212 CDEBUG(D_HA, "%s update xid from "LPU64" to "LPU64"\n",
2213 lut->lut_obd->obd_name, ted->ted_lcd->lcd_last_xid,
2215 ted->ted_lcd->lcd_last_xid = lcd->lcd_last_xid;
2216 ted->ted_lcd->lcd_last_result = lcd->lcd_last_result;
2218 class_export_put(export);
2221 static void replay_request_or_update(struct lu_env *env,
2222 struct lu_target *lut,
2223 struct target_recovery_data *trd,
2224 struct ptlrpc_thread *thread)
2226 struct obd_device *obd = lut->lut_obd;
2227 struct ptlrpc_request *req = NULL;
2232 CDEBUG(D_HA, "Waiting for transno "LPD64"\n",
2233 obd->obd_next_recovery_transno);
2235 /* Replay all of request and update by transno */
2237 struct target_distribute_txn_data *tdtd = lut->lut_tdtd;
2239 CFS_FAIL_TIMEOUT(OBD_FAIL_TGT_REPLAY_DELAY2, cfs_fail_val);
2241 /** It is needed to extend recovery window above
2242 * recovery_time_soft. Extending is possible only in the
2243 * end of recovery window (see more details in
2244 * handle_recovery_req()).
2246 CFS_FAIL_TIMEOUT_MS(OBD_FAIL_TGT_REPLAY_DELAY, 300);
2248 if (target_recovery_overseer(lut, check_for_next_transno,
2249 exp_req_replay_healthy)) {
2250 abort_req_replay_queue(obd);
2251 abort_lock_replay_queue(obd);
2254 spin_lock(&obd->obd_recovery_task_lock);
2255 transno = get_next_transno(lut, &type);
2256 if (type == REQUEST_RECOVERY && tdtd != NULL &&
2257 transno == tdtd->tdtd_last_update_transno) {
2258 /* Drop replay request from client side, if the
2259 * replay has been executed by update with the
2261 req = list_entry(obd->obd_req_replay_queue.next,
2262 struct ptlrpc_request, rq_list);
2263 list_del_init(&req->rq_list);
2264 obd->obd_requests_queued_for_recovery--;
2265 spin_unlock(&obd->obd_recovery_task_lock);
2266 drop_duplicate_replay_req(env, obd, req);
2267 } else if (type == REQUEST_RECOVERY && transno != 0) {
2268 req = list_entry(obd->obd_req_replay_queue.next,
2269 struct ptlrpc_request, rq_list);
2270 list_del_init(&req->rq_list);
2271 obd->obd_requests_queued_for_recovery--;
2272 spin_unlock(&obd->obd_recovery_task_lock);
2273 LASSERT(trd->trd_processing_task == current_pid());
2274 DEBUG_REQ(D_HA, req, "processing t"LPD64" from %s",
2275 lustre_msg_get_transno(req->rq_reqmsg),
2276 libcfs_nid2str(req->rq_peer.nid));
2278 handle_recovery_req(thread, req,
2279 trd->trd_recovery_handler);
2281 * bz18031: increase next_recovery_transno before
2282 * target_request_copy_put() will drop exp_rpc reference
2284 spin_lock(&obd->obd_recovery_task_lock);
2285 obd->obd_next_recovery_transno++;
2286 spin_unlock(&obd->obd_recovery_task_lock);
2287 target_exp_dequeue_req_replay(req);
2288 target_request_copy_put(req);
2289 obd->obd_replayed_requests++;
2290 } else if (type == UPDATE_RECOVERY && transno != 0) {
2291 struct distribute_txn_replay_req *dtrq;
2293 spin_unlock(&obd->obd_recovery_task_lock);
2295 LASSERT(tdtd != NULL);
2296 dtrq = distribute_txn_get_next_req(tdtd);
2297 lu_context_enter(&thread->t_env->le_ctx);
2298 tdtd->tdtd_replay_handler(env, tdtd, dtrq);
2299 lu_context_exit(&thread->t_env->le_ctx);
2300 extend_recovery_timer(obd, obd_timeout, true);
2301 LASSERT(tdtd->tdtd_last_update_transno <= transno);
2302 tdtd->tdtd_last_update_transno = transno;
2303 spin_lock(&obd->obd_recovery_task_lock);
2304 if (transno > obd->obd_next_recovery_transno)
2305 obd->obd_next_recovery_transno = transno;
2306 spin_unlock(&obd->obd_recovery_task_lock);
2307 target_update_lcd(env, lut, dtrq);
2310 spin_unlock(&obd->obd_recovery_task_lock);
2311 LASSERT(list_empty(&obd->obd_req_replay_queue));
2312 LASSERT(atomic_read(&obd->obd_req_replay_clients) == 0);
2313 /** evict exports failed VBR */
2314 class_disconnect_stale_exports(obd, exp_vbr_healthy);
2320 static int target_recovery_thread(void *arg)
2322 struct lu_target *lut = arg;
2323 struct obd_device *obd = lut->lut_obd;
2324 struct ptlrpc_request *req;
2325 struct target_recovery_data *trd = &obd->obd_recovery_data;
2326 unsigned long delta;
2328 struct ptlrpc_thread *thread = NULL;
2332 unshare_fs_struct();
2333 OBD_ALLOC_PTR(thread);
2339 OBD_FREE_PTR(thread);
2343 rc = lu_context_init(&env->le_ctx, LCT_MD_THREAD | LCT_DT_THREAD);
2345 OBD_FREE_PTR(thread);
2350 thread->t_env = env;
2351 thread->t_id = -1; /* force filter_iobuf_get/put to use local buffers */
2352 env->le_ctx.lc_thread = thread;
2353 tgt_io_thread_init(thread); /* init thread_big_cache for IO requests */
2354 thread->t_watchdog = NULL;
2356 CDEBUG(D_HA, "%s: started recovery thread pid %d\n", obd->obd_name,
2358 trd->trd_processing_task = current_pid();
2360 spin_lock(&obd->obd_dev_lock);
2361 obd->obd_recovering = 1;
2362 spin_unlock(&obd->obd_dev_lock);
2363 complete(&trd->trd_starting);
2365 /* first of all, we have to know the first transno to replay */
2366 if (target_recovery_overseer(lut, check_for_recovery_ready,
2367 exp_connect_healthy)) {
2368 abort_req_replay_queue(obd);
2369 abort_lock_replay_queue(obd);
2370 if (lut->lut_tdtd != NULL)
2371 dtrq_list_destroy(lut->lut_tdtd);
2374 /* next stage: replay requests or update */
2376 CDEBUG(D_INFO, "1: request replay stage - %d clients from t"LPU64"\n",
2377 atomic_read(&obd->obd_req_replay_clients),
2378 obd->obd_next_recovery_transno);
2379 replay_request_or_update(env, lut, trd, thread);
2382 * The second stage: replay locks
2384 CDEBUG(D_INFO, "2: lock replay stage - %d clients\n",
2385 atomic_read(&obd->obd_lock_replay_clients));
2386 while ((req = target_next_replay_lock(lut))) {
2387 LASSERT(trd->trd_processing_task == current_pid());
2388 DEBUG_REQ(D_HA, req, "processing lock from %s: ",
2389 libcfs_nid2str(req->rq_peer.nid));
2390 handle_recovery_req(thread, req,
2391 trd->trd_recovery_handler);
2392 target_request_copy_put(req);
2393 obd->obd_replayed_locks++;
2397 * The third stage: reply on final pings, at this moment all clients
2398 * must have request in final queue
2400 CFS_FAIL_TIMEOUT(OBD_FAIL_TGT_REPLAY_RECONNECT, cfs_fail_val);
2401 CDEBUG(D_INFO, "3: final stage - process recovery completion pings\n");
2402 /** Update server last boot epoch */
2403 tgt_boot_epoch_update(lut);
2404 /* We drop recoverying flag to forward all new requests
2405 * to regular mds_handle() since now */
2406 spin_lock(&obd->obd_dev_lock);
2407 obd->obd_recovering = obd->obd_abort_recovery = 0;
2408 spin_unlock(&obd->obd_dev_lock);
2409 spin_lock(&obd->obd_recovery_task_lock);
2410 target_cancel_recovery_timer(obd);
2411 spin_unlock(&obd->obd_recovery_task_lock);
2412 while ((req = target_next_final_ping(obd))) {
2413 LASSERT(trd->trd_processing_task == current_pid());
2414 DEBUG_REQ(D_HA, req, "processing final ping from %s: ",
2415 libcfs_nid2str(req->rq_peer.nid));
2416 handle_recovery_req(thread, req,
2417 trd->trd_recovery_handler);
2418 /* Because the waiting client can not send ping to server,
2419 * so we need refresh the last_request_time, to avoid the
2420 * export is being evicted */
2421 ptlrpc_update_export_timer(req->rq_export, 0);
2422 target_request_copy_put(req);
2425 delta = jiffies_to_msecs(jiffies - delta) / MSEC_PER_SEC;
2426 CDEBUG(D_INFO,"4: recovery completed in %lus - %d/%d reqs/locks\n",
2427 delta, obd->obd_replayed_requests, obd->obd_replayed_locks);
2428 if (delta > OBD_RECOVERY_TIME_SOFT) {
2429 CWARN("too long recovery - read logs\n");
2430 libcfs_debug_dumplog();
2433 target_finish_recovery(lut);
2435 lu_context_fini(&env->le_ctx);
2436 trd->trd_processing_task = 0;
2437 complete(&trd->trd_finishing);
2439 tgt_io_thread_done(thread);
2440 OBD_FREE_PTR(thread);
2445 static int target_start_recovery_thread(struct lu_target *lut,
2446 svc_handler_t handler)
2448 struct obd_device *obd = lut->lut_obd;
2450 struct target_recovery_data *trd = &obd->obd_recovery_data;
2452 memset(trd, 0, sizeof(*trd));
2453 init_completion(&trd->trd_starting);
2454 init_completion(&trd->trd_finishing);
2455 trd->trd_recovery_handler = handler;
2457 if (!IS_ERR(kthread_run(target_recovery_thread,
2458 lut, "tgt_recov"))) {
2459 wait_for_completion(&trd->trd_starting);
2460 LASSERT(obd->obd_recovering != 0);
2468 void target_stop_recovery_thread(struct obd_device *obd)
2470 if (obd->obd_recovery_data.trd_processing_task > 0) {
2471 struct target_recovery_data *trd = &obd->obd_recovery_data;
2472 /** recovery can be done but postrecovery is not yet */
2473 spin_lock(&obd->obd_dev_lock);
2474 if (obd->obd_recovering) {
2475 CERROR("%s: Aborting recovery\n", obd->obd_name);
2476 obd->obd_abort_recovery = 1;
2477 wake_up(&obd->obd_next_transno_waitq);
2479 spin_unlock(&obd->obd_dev_lock);
2480 wait_for_completion(&trd->trd_finishing);
2483 EXPORT_SYMBOL(target_stop_recovery_thread);
2485 void target_recovery_fini(struct obd_device *obd)
2487 class_disconnect_exports(obd);
2488 target_stop_recovery_thread(obd);
2489 target_cleanup_recovery(obd);
2491 EXPORT_SYMBOL(target_recovery_fini);
2493 static void target_recovery_expired(unsigned long castmeharder)
2495 struct obd_device *obd = (struct obd_device *)castmeharder;
2496 CDEBUG(D_HA, "%s: recovery timed out; %d clients are still in recovery"
2497 " after %lds (%d clients connected)\n",
2498 obd->obd_name, atomic_read(&obd->obd_lock_replay_clients),
2499 cfs_time_current_sec()- obd->obd_recovery_start,
2500 atomic_read(&obd->obd_connected_clients));
2502 obd->obd_recovery_expired = 1;
2503 wake_up(&obd->obd_next_transno_waitq);
2506 void target_recovery_init(struct lu_target *lut, svc_handler_t handler)
2508 struct obd_device *obd = lut->lut_obd;
2510 if (obd->obd_max_recoverable_clients == 0) {
2511 /** Update server last boot epoch */
2512 tgt_boot_epoch_update(lut);
2516 CDEBUG(D_HA, "RECOVERY: service %s, %d recoverable clients, "
2517 "last_transno "LPU64"\n", obd->obd_name,
2518 obd->obd_max_recoverable_clients, obd->obd_last_committed);
2519 LASSERT(obd->obd_stopping == 0);
2520 obd->obd_next_recovery_transno = obd->obd_last_committed + 1;
2521 obd->obd_recovery_start = 0;
2522 obd->obd_recovery_end = 0;
2524 cfs_timer_init(&obd->obd_recovery_timer, target_recovery_expired, obd);
2525 target_start_recovery_thread(lut, handler);
2527 EXPORT_SYMBOL(target_recovery_init);
2529 static int target_process_req_flags(struct obd_device *obd,
2530 struct ptlrpc_request *req)
2532 struct obd_export *exp = req->rq_export;
2533 LASSERT(exp != NULL);
2534 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REQ_REPLAY_DONE) {
2535 /* client declares he's ready to replay locks */
2536 spin_lock(&exp->exp_lock);
2537 if (exp->exp_req_replay_needed) {
2538 exp->exp_req_replay_needed = 0;
2539 spin_unlock(&exp->exp_lock);
2541 LASSERT_ATOMIC_POS(&obd->obd_req_replay_clients);
2542 atomic_dec(&obd->obd_req_replay_clients);
2544 spin_unlock(&exp->exp_lock);
2547 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_LOCK_REPLAY_DONE) {
2548 /* client declares he's ready to complete recovery
2549 * so, we put the request on th final queue */
2550 spin_lock(&exp->exp_lock);
2551 if (exp->exp_lock_replay_needed) {
2552 exp->exp_lock_replay_needed = 0;
2553 spin_unlock(&exp->exp_lock);
2555 LASSERT_ATOMIC_POS(&obd->obd_lock_replay_clients);
2556 atomic_dec(&obd->obd_lock_replay_clients);
2558 spin_unlock(&exp->exp_lock);
2564 int target_queue_recovery_request(struct ptlrpc_request *req,
2565 struct obd_device *obd)
2567 __u64 transno = lustre_msg_get_transno(req->rq_reqmsg);
2568 struct ptlrpc_request *reqiter;
2572 if (obd->obd_recovery_data.trd_processing_task == current_pid()) {
2573 /* Processing the queue right now, don't re-add. */
2577 target_process_req_flags(obd, req);
2579 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_LOCK_REPLAY_DONE) {
2580 /* client declares he's ready to complete recovery
2581 * so, we put the request on th final queue */
2582 target_request_copy_get(req);
2583 DEBUG_REQ(D_HA, req, "queue final req");
2584 wake_up(&obd->obd_next_transno_waitq);
2585 spin_lock(&obd->obd_recovery_task_lock);
2586 if (obd->obd_recovering) {
2587 list_add_tail(&req->rq_list,
2588 &obd->obd_final_req_queue);
2590 spin_unlock(&obd->obd_recovery_task_lock);
2591 target_request_copy_put(req);
2592 RETURN(obd->obd_stopping ? -ENOTCONN : 1);
2594 spin_unlock(&obd->obd_recovery_task_lock);
2597 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REQ_REPLAY_DONE) {
2598 /* client declares he's ready to replay locks */
2599 target_request_copy_get(req);
2600 DEBUG_REQ(D_HA, req, "queue lock replay req");
2601 wake_up(&obd->obd_next_transno_waitq);
2602 spin_lock(&obd->obd_recovery_task_lock);
2603 LASSERT(obd->obd_recovering);
2604 /* usually due to recovery abort */
2605 if (!req->rq_export->exp_in_recovery) {
2606 spin_unlock(&obd->obd_recovery_task_lock);
2607 target_request_copy_put(req);
2610 LASSERT(req->rq_export->exp_lock_replay_needed);
2611 list_add_tail(&req->rq_list, &obd->obd_lock_replay_queue);
2612 spin_unlock(&obd->obd_recovery_task_lock);
2616 /* CAVEAT EMPTOR: The incoming request message has been swabbed
2617 * (i.e. buflens etc are in my own byte order), but type-dependent
2618 * buffers (eg mdt_body, ost_body etc) have NOT been swabbed. */
2621 INIT_LIST_HEAD(&req->rq_list);
2622 DEBUG_REQ(D_HA, req, "not queueing");
2626 /* If we're processing the queue, we want don't want to queue this
2629 * Also, if this request has a transno less than the one we're waiting
2630 * for, we should process it now. It could (and currently always will)
2631 * be an open request for a descriptor that was opened some time ago.
2633 * Also, a resent, replayed request that has already been
2634 * handled will pass through here and be processed immediately.
2636 CDEBUG(D_HA, "Next recovery transno: "LPU64
2637 ", current: "LPU64", replaying\n",
2638 obd->obd_next_recovery_transno, transno);
2639 spin_lock(&obd->obd_recovery_task_lock);
2640 if (transno < obd->obd_next_recovery_transno) {
2641 /* Processing the queue right now, don't re-add. */
2642 LASSERT(list_empty(&req->rq_list));
2643 spin_unlock(&obd->obd_recovery_task_lock);
2646 spin_unlock(&obd->obd_recovery_task_lock);
2648 if (OBD_FAIL_CHECK(OBD_FAIL_TGT_REPLAY_DROP))
2651 target_request_copy_get(req);
2652 if (!req->rq_export->exp_in_recovery) {
2653 target_request_copy_put(req);
2656 LASSERT(req->rq_export->exp_req_replay_needed);
2658 if (target_exp_enqueue_req_replay(req)) {
2659 DEBUG_REQ(D_ERROR, req, "dropping resent queued req");
2660 target_request_copy_put(req);
2665 spin_lock(&obd->obd_recovery_task_lock);
2666 LASSERT(obd->obd_recovering);
2667 list_for_each_entry(reqiter, &obd->obd_req_replay_queue, rq_list) {
2668 if (lustre_msg_get_transno(reqiter->rq_reqmsg) > transno) {
2669 list_add_tail(&req->rq_list, &reqiter->rq_list);
2674 if (unlikely(lustre_msg_get_transno(reqiter->rq_reqmsg) ==
2676 DEBUG_REQ(D_ERROR, req, "dropping replay: transno "
2677 "has been claimed by another client");
2678 spin_unlock(&obd->obd_recovery_task_lock);
2679 target_exp_dequeue_req_replay(req);
2680 target_request_copy_put(req);
2686 list_add_tail(&req->rq_list, &obd->obd_req_replay_queue);
2688 obd->obd_requests_queued_for_recovery++;
2689 spin_unlock(&obd->obd_recovery_task_lock);
2690 wake_up(&obd->obd_next_transno_waitq);
2694 int target_handle_ping(struct ptlrpc_request *req)
2696 obd_ping(req->rq_svc_thread->t_env, req->rq_export);
2697 return req_capsule_server_pack(&req->rq_pill);
2700 void target_committed_to_req(struct ptlrpc_request *req)
2702 struct obd_export *exp = req->rq_export;
2704 if (!exp->exp_obd->obd_no_transno && req->rq_repmsg != NULL)
2705 lustre_msg_set_last_committed(req->rq_repmsg,
2706 exp->exp_last_committed);
2708 DEBUG_REQ(D_IOCTL, req, "not sending last_committed update (%d/"
2709 "%d)", exp->exp_obd->obd_no_transno,
2710 req->rq_repmsg == NULL);
2712 CDEBUG(D_INFO, "last_committed "LPU64", transno "LPU64", xid "LPU64"\n",
2713 exp->exp_last_committed, req->rq_transno, req->rq_xid);
2716 #endif /* HAVE_SERVER_SUPPORT */
2719 * Packs current SLV and Limit into \a req.
2721 int target_pack_pool_reply(struct ptlrpc_request *req)
2723 struct obd_device *obd;
2726 /* Check that we still have all structures alive as this may
2727 * be some late RPC at shutdown time. */
2728 if (unlikely(!req->rq_export || !req->rq_export->exp_obd ||
2729 !exp_connect_lru_resize(req->rq_export))) {
2730 lustre_msg_set_slv(req->rq_repmsg, 0);
2731 lustre_msg_set_limit(req->rq_repmsg, 0);
2735 /* OBD is alive here as export is alive, which we checked above. */
2736 obd = req->rq_export->exp_obd;
2738 read_lock(&obd->obd_pool_lock);
2739 lustre_msg_set_slv(req->rq_repmsg, obd->obd_pool_slv);
2740 lustre_msg_set_limit(req->rq_repmsg, obd->obd_pool_limit);
2741 read_unlock(&obd->obd_pool_lock);
2746 static int target_send_reply_msg(struct ptlrpc_request *req,
2747 int rc, int fail_id)
2749 if (OBD_FAIL_CHECK_ORSET(fail_id & ~OBD_FAIL_ONCE, OBD_FAIL_ONCE)) {
2750 DEBUG_REQ(D_ERROR, req, "dropping reply");
2755 DEBUG_REQ(D_NET, req, "processing error (%d)", rc);
2756 req->rq_status = rc;
2757 return (ptlrpc_send_error(req, 1));
2759 DEBUG_REQ(D_NET, req, "sending reply");
2762 return (ptlrpc_send_reply(req, PTLRPC_REPLY_MAYBE_DIFFICULT));
2765 void target_send_reply(struct ptlrpc_request *req, int rc, int fail_id)
2767 struct ptlrpc_service_part *svcpt;
2769 struct ptlrpc_reply_state *rs;
2770 struct obd_export *exp;
2773 if (req->rq_no_reply) {
2778 svcpt = req->rq_rqbd->rqbd_svcpt;
2779 rs = req->rq_reply_state;
2780 if (rs == NULL || !rs->rs_difficult) {
2782 target_send_reply_msg (req, rc, fail_id);
2787 /* must be an export if locks saved */
2788 LASSERT(req->rq_export != NULL);
2789 /* req/reply consistent */
2790 LASSERT(rs->rs_svcpt == svcpt);
2793 LASSERT(!rs->rs_scheduled);
2794 LASSERT(!rs->rs_scheduled_ever);
2795 LASSERT(!rs->rs_handled);
2796 LASSERT(!rs->rs_on_net);
2797 LASSERT(rs->rs_export == NULL);
2798 LASSERT(list_empty(&rs->rs_obd_list));
2799 LASSERT(list_empty(&rs->rs_exp_list));
2801 exp = class_export_get(req->rq_export);
2803 /* disable reply scheduling while I'm setting up */
2804 rs->rs_scheduled = 1;
2806 rs->rs_xid = req->rq_xid;
2807 rs->rs_transno = req->rq_transno;
2808 rs->rs_export = exp;
2809 rs->rs_opc = lustre_msg_get_opc(req->rq_reqmsg);
2811 spin_lock(&exp->exp_uncommitted_replies_lock);
2812 CDEBUG(D_NET, "rs transno = "LPU64", last committed = "LPU64"\n",
2813 rs->rs_transno, exp->exp_last_committed);
2814 if (rs->rs_transno > exp->exp_last_committed) {
2815 /* not committed already */
2816 list_add_tail(&rs->rs_obd_list,
2817 &exp->exp_uncommitted_replies);
2819 spin_unlock(&exp->exp_uncommitted_replies_lock);
2821 spin_lock(&exp->exp_lock);
2822 list_add_tail(&rs->rs_exp_list, &exp->exp_outstanding_replies);
2823 spin_unlock(&exp->exp_lock);
2825 netrc = target_send_reply_msg(req, rc, fail_id);
2827 spin_lock(&svcpt->scp_rep_lock);
2829 atomic_inc(&svcpt->scp_nreps_difficult);
2832 /* error sending: reply is off the net. Also we need +1
2833 * reply ref until ptlrpc_handle_rs() is done
2834 * with the reply state (if the send was successful, there
2835 * would have been +1 ref for the net, which
2836 * reply_out_callback leaves alone) */
2838 ptlrpc_rs_addref(rs);
2841 spin_lock(&rs->rs_lock);
2842 if (rs->rs_transno <= exp->exp_last_committed ||
2843 (!rs->rs_on_net && !rs->rs_no_ack) ||
2844 list_empty(&rs->rs_exp_list) || /* completed already */
2845 list_empty(&rs->rs_obd_list)) {
2846 CDEBUG(D_HA, "Schedule reply immediately\n");
2847 ptlrpc_dispatch_difficult_reply(rs);
2849 list_add(&rs->rs_list, &svcpt->scp_rep_active);
2850 rs->rs_scheduled = 0; /* allow notifier to schedule */
2852 spin_unlock(&rs->rs_lock);
2853 spin_unlock(&svcpt->scp_rep_lock);
2857 ldlm_mode_t lck_compat_array[] = {
2858 [LCK_EX] = LCK_COMPAT_EX,
2859 [LCK_PW] = LCK_COMPAT_PW,
2860 [LCK_PR] = LCK_COMPAT_PR,
2861 [LCK_CW] = LCK_COMPAT_CW,
2862 [LCK_CR] = LCK_COMPAT_CR,
2863 [LCK_NL] = LCK_COMPAT_NL,
2864 [LCK_GROUP] = LCK_COMPAT_GROUP,
2865 [LCK_COS] = LCK_COMPAT_COS,
2869 * Rather arbitrary mapping from LDLM error codes to errno values. This should
2870 * not escape to the user level.
2872 int ldlm_error2errno(ldlm_error_t error)
2878 case ELDLM_LOCK_MATCHED:
2881 case ELDLM_LOCK_CHANGED:
2884 case ELDLM_LOCK_ABORTED:
2887 case ELDLM_LOCK_REPLACED:
2890 case ELDLM_NO_LOCK_DATA:
2893 case ELDLM_NAMESPACE_EXISTS:
2896 case ELDLM_BAD_NAMESPACE:
2900 if (((int)error) < 0) /* cast to signed type */
2901 result = error; /* as ldlm_error_t can be unsigned */
2903 CERROR("Invalid DLM result code: %d\n", error);
2909 EXPORT_SYMBOL(ldlm_error2errno);
2912 * Dual to ldlm_error2errno(): maps errno values back to ldlm_error_t.
2914 ldlm_error_t ldlm_errno2error(int err_no)
2923 error = ELDLM_LOCK_CHANGED;
2926 error = ELDLM_LOCK_ABORTED;
2929 error = ELDLM_LOCK_REPLACED;
2932 error = ELDLM_NO_LOCK_DATA;
2935 error = ELDLM_NAMESPACE_EXISTS;
2938 error = ELDLM_BAD_NAMESPACE;
2946 #if LUSTRE_TRACKS_LOCK_EXP_REFS
2947 void ldlm_dump_export_locks(struct obd_export *exp)
2949 spin_lock(&exp->exp_locks_list_guard);
2950 if (!list_empty(&exp->exp_locks_list)) {
2951 struct ldlm_lock *lock;
2953 CERROR("dumping locks for export %p,"
2954 "ignore if the unmount doesn't hang\n", exp);
2955 list_for_each_entry(lock, &exp->exp_locks_list,
2957 LDLM_ERROR(lock, "lock:");
2959 spin_unlock(&exp->exp_locks_list_guard);
2963 #ifdef HAVE_SERVER_SUPPORT
2964 static int target_bulk_timeout(void *data)
2967 /* We don't fail the connection here, because having the export
2968 * killed makes the (vital) call to commitrw very sad.
2973 static inline const char *bulk2type(struct ptlrpc_request *req)
2975 if (req->rq_bulk_read)
2977 if (req->rq_bulk_write)
2982 int target_bulk_io(struct obd_export *exp, struct ptlrpc_bulk_desc *desc,
2983 struct l_wait_info *lwi)
2985 struct ptlrpc_request *req = desc->bd_req;
2986 time_t start = cfs_time_current_sec();
2992 /* If there is eviction in progress, wait for it to finish. */
2993 if (unlikely(atomic_read(&exp->exp_obd->obd_evict_inprogress))) {
2994 *lwi = LWI_INTR(NULL, NULL);
2995 rc = l_wait_event(exp->exp_obd->obd_evict_inprogress_waitq,
2996 !atomic_read(&exp->exp_obd->
2997 obd_evict_inprogress),
3001 /* Check if client was evicted or reconnected already. */
3002 if (exp->exp_failed ||
3003 exp->exp_conn_cnt > lustre_msg_get_conn_cnt(req->rq_reqmsg)) {
3006 if (req->rq_bulk_read)
3007 rc = sptlrpc_svc_wrap_bulk(req, desc);
3009 if ((exp->exp_connect_data.ocd_connect_flags &
3010 OBD_CONNECT_BULK_MBITS) != 0)
3011 req->rq_mbits = lustre_msg_get_mbits(req->rq_reqmsg);
3012 else /* old version, bulk matchbits is rq_xid */
3013 req->rq_mbits = req->rq_xid;
3016 rc = ptlrpc_start_bulk_transfer(desc);
3020 DEBUG_REQ(D_ERROR, req, "bulk %s failed: rc %d",
3021 bulk2type(req), rc);
3025 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_SENDPAGE)) {
3026 ptlrpc_abort_bulk(desc);
3030 /* limit actual bulk transfer to bulk_timeout seconds */
3031 deadline = start + bulk_timeout;
3032 if (deadline > req->rq_deadline)
3033 deadline = req->rq_deadline;
3036 long timeoutl = deadline - cfs_time_current_sec();
3037 cfs_duration_t timeout = timeoutl <= 0 ?
3038 CFS_TICK : cfs_time_seconds(timeoutl);
3041 *lwi = LWI_TIMEOUT_INTERVAL(timeout, cfs_time_seconds(1),
3042 target_bulk_timeout, desc);
3043 rc = l_wait_event(desc->bd_waitq,
3044 !ptlrpc_server_bulk_active(desc) ||
3047 lustre_msg_get_conn_cnt(req->rq_reqmsg),
3049 LASSERT(rc == 0 || rc == -ETIMEDOUT);
3050 /* Wait again if we changed rq_deadline. */
3051 rq_deadline = ACCESS_ONCE(req->rq_deadline);
3052 deadline = start + bulk_timeout;
3053 if (deadline > rq_deadline)
3054 deadline = rq_deadline;
3055 } while ((rc == -ETIMEDOUT) &&
3056 (deadline > cfs_time_current_sec()));
3058 if (rc == -ETIMEDOUT) {
3059 DEBUG_REQ(D_ERROR, req, "timeout on bulk %s after %ld%+lds",
3060 bulk2type(req), deadline - start,
3061 cfs_time_current_sec() - deadline);
3062 ptlrpc_abort_bulk(desc);
3063 } else if (exp->exp_failed) {
3064 DEBUG_REQ(D_ERROR, req, "Eviction on bulk %s",
3067 ptlrpc_abort_bulk(desc);
3068 } else if (exp->exp_conn_cnt >
3069 lustre_msg_get_conn_cnt(req->rq_reqmsg)) {
3070 DEBUG_REQ(D_ERROR, req, "Reconnect on bulk %s",
3072 /* We don't reply anyway. */
3074 ptlrpc_abort_bulk(desc);
3075 } else if (desc->bd_failure) {
3076 DEBUG_REQ(D_ERROR, req, "network error on bulk %s",
3078 /* XXX should this be a different errno? */
3081 if (req->rq_bulk_write)
3082 rc = sptlrpc_svc_unwrap_bulk(req, desc);
3083 if (rc == 0 && desc->bd_nob_transferred != desc->bd_nob) {
3084 DEBUG_REQ(D_ERROR, req, "truncated bulk %s %d(%d)",
3085 bulk2type(req), desc->bd_nob_transferred,
3087 /* XXX should this be a different errno? */
3094 EXPORT_SYMBOL(target_bulk_io);
3096 #endif /* HAVE_SERVER_SUPPORT */