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 ldlm_bl_thread_wakeup();
693 /* complete all outstanding replies */
694 spin_lock(&exp->exp_lock);
695 while (!list_empty(&exp->exp_outstanding_replies)) {
696 struct ptlrpc_reply_state *rs =
697 list_entry(exp->exp_outstanding_replies.next,
698 struct ptlrpc_reply_state, rs_exp_list);
699 struct ptlrpc_service_part *svcpt = rs->rs_svcpt;
701 spin_lock(&svcpt->scp_rep_lock);
703 list_del_init(&rs->rs_exp_list);
704 spin_lock(&rs->rs_lock);
705 ptlrpc_schedule_difficult_reply(rs);
706 spin_unlock(&rs->rs_lock);
708 spin_unlock(&svcpt->scp_rep_lock);
710 spin_unlock(&exp->exp_lock);
714 EXPORT_SYMBOL(server_disconnect_export);
716 /* --------------------------------------------------------------------------
717 * from old lib/target.c
718 * -------------------------------------------------------------------------- */
720 static int target_handle_reconnect(struct lustre_handle *conn,
721 struct obd_export *exp,
722 struct obd_uuid *cluuid)
726 if (exp->exp_connection && exp->exp_imp_reverse) {
727 struct lustre_handle *hdl;
728 struct obd_device *target;
730 hdl = &exp->exp_imp_reverse->imp_remote_handle;
731 target = exp->exp_obd;
733 /* Might be a re-connect after a partition. */
734 if (!memcmp(&conn->cookie, &hdl->cookie, sizeof conn->cookie)) {
735 if (target->obd_recovering) {
736 int timeout = cfs_duration_sec(cfs_time_sub(
738 &target->obd_recovery_timer),
739 cfs_time_current()));
741 LCONSOLE_WARN("%s: Client %s (at %s) reconnect"
742 "ing, waiting for %d clients in recov"
743 "ery for %d:%.02d\n", target->obd_name,
744 obd_uuid2str(&exp->exp_client_uuid),
745 obd_export_nid2str(exp),
746 target->obd_max_recoverable_clients,
747 timeout / 60, timeout % 60);
749 LCONSOLE_WARN("%s: Client %s (at %s) "
750 "reconnecting\n", target->obd_name,
751 obd_uuid2str(&exp->exp_client_uuid),
752 obd_export_nid2str(exp));
755 conn->cookie = exp->exp_handle.h_cookie;
756 /* target_handle_connect() treats EALREADY and
757 * -EALREADY differently. EALREADY means we are
758 * doing a valid reconnect from the same client. */
761 LCONSOLE_WARN("%s: already connected client %s (at %s) "
762 "with handle "LPX64". Rejecting client "
763 "with the same UUID trying to reconnect "
764 "with handle "LPX64"\n", target->obd_name,
765 obd_uuid2str(&exp->exp_client_uuid),
766 obd_export_nid2str(exp),
767 hdl->cookie, conn->cookie);
768 memset(conn, 0, sizeof *conn);
769 /* target_handle_connect() treats EALREADY and
770 * -EALREADY differently. -EALREADY is an error
771 * (same UUID, different handle). */
776 conn->cookie = exp->exp_handle.h_cookie;
777 CDEBUG(D_HA, "connect export for UUID '%s' at %p, cookie "LPX64"\n",
778 cluuid->uuid, exp, conn->cookie);
782 void target_client_add_cb(struct obd_device *obd, __u64 transno, void *cb_data,
785 struct obd_export *exp = cb_data;
787 CDEBUG(D_RPCTRACE, "%s: committing for initial connect of %s\n",
788 obd->obd_name, exp->exp_client_uuid.uuid);
790 spin_lock(&exp->exp_lock);
791 exp->exp_need_sync = 0;
792 spin_unlock(&exp->exp_lock);
793 class_export_cb_put(exp);
797 check_and_start_recovery_timer(struct obd_device *obd,
798 struct ptlrpc_request *req, int new_client);
800 int target_handle_connect(struct ptlrpc_request *req)
802 struct obd_device *target = NULL, *targref = NULL;
803 struct obd_export *export = NULL;
804 struct obd_import *revimp;
805 struct obd_import *tmp_imp = NULL;
806 struct lustre_handle conn;
807 struct lustre_handle *tmp;
808 struct obd_uuid tgtuuid;
809 struct obd_uuid cluuid;
810 struct obd_uuid remote_uuid;
815 bool mds_conn = false, lw_client = false;
816 bool mds_mds_conn = false;
817 bool new_mds_mds_conn = false;
818 struct obd_connect_data *data, *tmpdata;
820 lnet_nid_t *client_nid = NULL;
823 OBD_RACE(OBD_FAIL_TGT_CONN_RACE);
825 str = req_capsule_client_get(&req->rq_pill, &RMF_TGTUUID);
827 DEBUG_REQ(D_ERROR, req, "bad target UUID for connect");
828 GOTO(out, rc = -EINVAL);
831 obd_str2uuid(&tgtuuid, str);
832 target = class_uuid2obd(&tgtuuid);
834 target = class_name2obd(str);
837 deuuidify(str, NULL, &target_start, &target_len);
838 LCONSOLE_ERROR_MSG(0x137, "%s: not available for connect "
839 "from %s (no target). If you are running "
840 "an HA pair check that the target is "
841 "mounted on the other server.\n", str,
842 libcfs_nid2str(req->rq_peer.nid));
843 GOTO(out, rc = -ENODEV);
846 spin_lock(&target->obd_dev_lock);
847 if (target->obd_stopping || !target->obd_set_up) {
848 spin_unlock(&target->obd_dev_lock);
850 deuuidify(str, NULL, &target_start, &target_len);
851 LCONSOLE_INFO("%.*s: Not available for connect from %s (%s)\n",
852 target_len, target_start,
853 libcfs_nid2str(req->rq_peer.nid),
854 (target->obd_stopping ?
855 "stopping" : "not set up"));
856 GOTO(out, rc = -ENODEV);
859 if (target->obd_no_conn) {
860 spin_unlock(&target->obd_dev_lock);
862 CDEBUG(D_INFO, "%s: Temporarily refusing client connection "
863 "from %s\n", target->obd_name,
864 libcfs_nid2str(req->rq_peer.nid));
865 GOTO(out, rc = -EAGAIN);
868 /* Make sure the target isn't cleaned up while we're here. Yes,
869 * there's still a race between the above check and our incref here.
870 * Really, class_uuid2obd should take the ref. */
871 targref = class_incref(target, __FUNCTION__, current);
873 target->obd_conn_inprogress++;
874 spin_unlock(&target->obd_dev_lock);
876 str = req_capsule_client_get(&req->rq_pill, &RMF_CLUUID);
878 DEBUG_REQ(D_ERROR, req, "bad client UUID for connect");
879 GOTO(out, rc = -EINVAL);
882 obd_str2uuid(&cluuid, str);
884 /* XXX Extract a nettype and format accordingly. */
885 switch (sizeof(lnet_nid_t)) {
886 /* NB the casts only avoid compiler warnings. */
888 snprintf(remote_uuid.uuid, sizeof remote_uuid,
889 "NET_"LPX64"_UUID", (__u64)req->rq_peer.nid);
892 snprintf(remote_uuid.uuid, sizeof remote_uuid,
893 "NET_%x_UUID", (__u32)req->rq_peer.nid);
899 tmp = req_capsule_client_get(&req->rq_pill, &RMF_CONN);
901 GOTO(out, rc = -EPROTO);
905 size = req_capsule_get_size(&req->rq_pill, &RMF_CONNECT_DATA,
907 data = req_capsule_client_get(&req->rq_pill, &RMF_CONNECT_DATA);
909 GOTO(out, rc = -EPROTO);
911 rc = req_capsule_server_pack(&req->rq_pill);
915 #if LUSTRE_VERSION_CODE < OBD_OCD_VERSION(3, 0, 53, 0)
916 /* Don't allow clients to connect that are using old 1.8 format
917 * protocol conventions (LUSTRE_MSG_MAGIC_v1, !MSGHDR_CKSUM_INCOMPAT18,
918 * ldlm_flock_policy_wire format, MDT_ATTR_xTIME_SET, etc). The
919 * FULL20 flag should be set on all connections since 2.0, but no
920 * longer affects behaviour.
922 * Later this check will be disabled and the flag can be retired
923 * completely once interop with 3.0 is no longer needed.
925 if (!(data->ocd_connect_flags & OBD_CONNECT_FULL20))
926 GOTO(out, rc = -EPROTO);
929 if (lustre_msg_get_op_flags(req->rq_reqmsg) & MSG_CONNECT_LIBCLIENT) {
930 if (data->ocd_version < LUSTRE_VERSION_CODE -
931 LUSTRE_VERSION_ALLOWED_OFFSET ||
932 data->ocd_version > LUSTRE_VERSION_CODE +
933 LUSTRE_VERSION_ALLOWED_OFFSET) {
934 DEBUG_REQ(D_WARNING, req, "Refusing %s (%d.%d.%d.%d) "
935 "libclient connection attempt",
936 data->ocd_version < LUSTRE_VERSION_CODE ?
938 OBD_OCD_VERSION_MAJOR(data->ocd_version),
939 OBD_OCD_VERSION_MINOR(data->ocd_version),
940 OBD_OCD_VERSION_PATCH(data->ocd_version),
941 OBD_OCD_VERSION_FIX(data->ocd_version));
942 data = req_capsule_server_sized_get(&req->rq_pill,
944 offsetof(typeof(*data), ocd_version) +
945 sizeof(data->ocd_version));
947 data->ocd_connect_flags = OBD_CONNECT_VERSION;
948 data->ocd_version = LUSTRE_VERSION_CODE;
950 GOTO(out, rc = -EPROTO);
954 /* Note: lw_client is needed in MDS-MDS failover during update log
955 * processing, so we needs to allow lw_client to be connected at
956 * anytime, instead of only the initial connection */
957 lw_client = (data->ocd_connect_flags & OBD_CONNECT_LIGHTWEIGHT) != 0;
959 if (lustre_msg_get_op_flags(req->rq_reqmsg) & MSG_CONNECT_INITIAL) {
960 mds_conn = (data->ocd_connect_flags & OBD_CONNECT_MDS) != 0;
961 mds_mds_conn = (data->ocd_connect_flags &
962 OBD_CONNECT_MDS_MDS) != 0;
964 /* OBD_CONNECT_MNE_SWAB is defined as OBD_CONNECT_MDS_MDS
965 * for Imperative Recovery connection from MGC to MGS.
967 * Via check OBD_CONNECT_FID, we can distinguish whether
968 * the OBD_CONNECT_MDS_MDS/OBD_CONNECT_MNE_SWAB is from
971 (data->ocd_connect_flags & OBD_CONNECT_MDS_MDS) &&
972 (data->ocd_connect_flags & OBD_CONNECT_FID) &&
973 (data->ocd_connect_flags & OBD_CONNECT_VERSION)) {
974 __u32 major = OBD_OCD_VERSION_MAJOR(data->ocd_version);
975 __u32 minor = OBD_OCD_VERSION_MINOR(data->ocd_version);
976 __u32 patch = OBD_OCD_VERSION_PATCH(data->ocd_version);
978 /* We do not support the MDT-MDT interoperations with
979 * different version MDT because of protocol changes. */
980 if (unlikely(major != LUSTRE_MAJOR ||
981 minor != LUSTRE_MINOR ||
982 abs(patch - LUSTRE_PATCH) > 3)) {
983 LCONSOLE_WARN("%s (%u.%u.%u.%u) refused the "
984 "connection from different version MDT "
985 "(%d.%d.%d.%d) %s %s\n",
986 target->obd_name, LUSTRE_MAJOR,
987 LUSTRE_MINOR, LUSTRE_PATCH, LUSTRE_FIX,
989 OBD_OCD_VERSION_FIX(data->ocd_version),
990 libcfs_nid2str(req->rq_peer.nid), str);
992 GOTO(out, rc = -EPROTO);
997 /* lctl gets a backstage, all-access pass. */
998 if (obd_uuid_equals(&cluuid, &target->obd_uuid))
999 goto dont_check_exports;
1001 export = cfs_hash_lookup(target->obd_uuid_hash, &cluuid);
1005 /* We've found an export in the hash. */
1007 spin_lock(&export->exp_lock);
1009 if (export->exp_connecting) { /* bug 9635, et. al. */
1010 spin_unlock(&export->exp_lock);
1011 LCONSOLE_WARN("%s: Export %p already connecting from %s\n",
1012 export->exp_obd->obd_name, export,
1013 libcfs_nid2str(req->rq_peer.nid));
1014 class_export_put(export);
1017 } else if ((mds_conn || lw_client) && export->exp_connection != NULL) {
1018 spin_unlock(&export->exp_lock);
1019 if (req->rq_peer.nid != export->exp_connection->c_peer.nid)
1020 /* MDS or LWP reconnected after failover. */
1021 LCONSOLE_WARN("%s: Received %s connection from "
1022 "%s, removing former export from %s\n",
1023 target->obd_name, mds_conn ? "MDS" : "LWP",
1024 libcfs_nid2str(req->rq_peer.nid),
1025 libcfs_nid2str(export->exp_connection->c_peer.nid));
1027 /* New MDS connection from the same NID. */
1028 LCONSOLE_WARN("%s: Received new %s connection from "
1029 "%s, removing former export from same NID\n",
1030 target->obd_name, mds_conn ? "MDS" : "LWP",
1031 libcfs_nid2str(req->rq_peer.nid));
1032 class_fail_export(export);
1033 class_export_put(export);
1036 } else if (export->exp_connection != NULL &&
1037 req->rq_peer.nid != export->exp_connection->c_peer.nid &&
1038 (lustre_msg_get_op_flags(req->rq_reqmsg) &
1039 MSG_CONNECT_INITIAL)) {
1040 spin_unlock(&export->exp_lock);
1041 /* In MDS failover we have static UUID but NID can change. */
1042 LCONSOLE_WARN("%s: Client %s seen on new nid %s when "
1043 "existing nid %s is already connected\n",
1044 target->obd_name, cluuid.uuid,
1045 libcfs_nid2str(req->rq_peer.nid),
1047 export->exp_connection->c_peer.nid));
1049 class_export_put(export);
1052 export->exp_connecting = 1;
1053 spin_unlock(&export->exp_lock);
1054 LASSERT(export->exp_obd == target);
1056 rc = target_handle_reconnect(&conn, export, &cluuid);
1059 /* If we found an export, we already unlocked. */
1062 OBD_FAIL_TIMEOUT(OBD_FAIL_TGT_DELAY_CONNECT, 2 * obd_timeout);
1063 } else if (req->rq_export == NULL &&
1064 atomic_read(&export->exp_rpc_count) > 0) {
1065 LCONSOLE_WARN("%s: Client %s (at %s) refused connection, "
1066 "still busy with %d references\n",
1067 target->obd_name, cluuid.uuid,
1068 libcfs_nid2str(req->rq_peer.nid),
1069 atomic_read(&export->exp_refcount));
1070 GOTO(out, rc = -EBUSY);
1071 } else if (lustre_msg_get_conn_cnt(req->rq_reqmsg) == 1) {
1072 if (!strstr(cluuid.uuid, "mdt"))
1073 LCONSOLE_WARN("%s: Rejecting reconnect from the "
1074 "known client %s (at %s) because it "
1075 "is indicating it is a new client",
1076 target->obd_name, cluuid.uuid,
1077 libcfs_nid2str(req->rq_peer.nid));
1078 GOTO(out, rc = -EALREADY);
1080 OBD_FAIL_TIMEOUT(OBD_FAIL_TGT_DELAY_RECONNECT, 2 * obd_timeout);
1087 CDEBUG(D_HA, "%s: connection from %s@%s %st"LPU64" exp %p cur %ld last %ld\n",
1088 target->obd_name, cluuid.uuid, libcfs_nid2str(req->rq_peer.nid),
1089 target->obd_recovering ? "recovering/" : "", data->ocd_transno,
1090 export, (long)cfs_time_current_sec(),
1091 export ? (long)export->exp_last_request_time : 0);
1093 /* If this is the first time a client connects, reset the recovery
1094 * timer. Discard lightweight connections which might be local. */
1095 if (!lw_client && rc == 0 && target->obd_recovering)
1096 check_and_start_recovery_timer(target, req, export == NULL);
1098 /* We want to handle EALREADY but *not* -EALREADY from
1099 * target_handle_reconnect(), return reconnection state in a flag. */
1100 if (rc == EALREADY) {
1101 lustre_msg_add_op_flags(req->rq_repmsg, MSG_CONNECT_RECONNECT);
1107 /* Tell the client if we support replayable requests. */
1108 if (target->obd_replayable)
1109 lustre_msg_add_op_flags(req->rq_repmsg, MSG_CONNECT_REPLAYABLE);
1110 client_nid = &req->rq_peer.nid;
1112 if (export == NULL) {
1113 /* allow lightweight connections during recovery */
1114 /* allow "new" MDT to be connected during recovery, since we
1115 * need retrieve recovery update records from it */
1116 if (target->obd_recovering && !lw_client && !mds_mds_conn) {
1118 int c; /* connected */
1119 int i; /* in progress */
1121 int s; /* stale/evicted */
1123 c = atomic_read(&target->obd_connected_clients);
1124 i = atomic_read(&target->obd_lock_replay_clients);
1125 k = target->obd_max_recoverable_clients;
1126 s = target->obd_stale_clients;
1127 t = cfs_timer_deadline(&target->obd_recovery_timer);
1128 t = cfs_time_sub(t, cfs_time_current());
1129 t = cfs_duration_sec(t);
1130 LCONSOLE_WARN("%s: Denying connection for new client %s"
1131 "(at %s), waiting for %d known clients "
1132 "(%d recovered, %d in progress, and %d "
1133 "evicted) to recover in %d:%.02d\n",
1134 target->obd_name, cluuid.uuid,
1135 libcfs_nid2str(req->rq_peer.nid), k,
1136 c - i, i, s, (int)t / 60,
1141 rc = obd_connect(req->rq_svc_thread->t_env,
1142 &export, target, &cluuid, data,
1144 if (mds_conn && OBD_FAIL_CHECK(OBD_FAIL_TGT_RCVG_FLAG))
1145 lustre_msg_add_op_flags(req->rq_repmsg,
1146 MSG_CONNECT_RECOVERING);
1148 conn.cookie = export->exp_handle.h_cookie;
1151 new_mds_mds_conn = true;
1154 rc = obd_reconnect(req->rq_svc_thread->t_env,
1155 export, target, &cluuid, data, client_nid);
1160 LASSERT(target->u.obt.obt_magic == OBT_MAGIC);
1161 data->ocd_instance = target->u.obt.obt_instance;
1163 /* Return only the parts of obd_connect_data that we understand, so the
1164 * client knows that we don't understand the rest. */
1166 tmpsize = req_capsule_get_size(&req->rq_pill, &RMF_CONNECT_DATA,
1168 tmpdata = req_capsule_server_get(&req->rq_pill,
1170 /* Don't use struct assignment here, because the client reply
1171 * buffer may be smaller/larger than the local struct
1172 * obd_connect_data. */
1173 memcpy(tmpdata, data, min(tmpsize, size));
1176 /* If all else goes well, this is our RPC return code. */
1179 lustre_msg_set_handle(req->rq_repmsg, &conn);
1181 /* If the client and the server are the same node, we will already
1182 * have an export that really points to the client's DLM export,
1183 * because we have a shared handles table.
1185 * XXX this will go away when shaver stops sending the "connect" handle
1186 * in the real "remote handle" field of the request --phik 24 Apr 2003
1188 ptlrpc_request_change_export(req, export);
1190 spin_lock(&export->exp_lock);
1191 if (export->exp_conn_cnt >= lustre_msg_get_conn_cnt(req->rq_reqmsg)) {
1192 spin_unlock(&export->exp_lock);
1193 CDEBUG(D_RPCTRACE, "%s: %s already connected at greater "
1194 "or equal conn_cnt: %d >= %d\n",
1195 cluuid.uuid, libcfs_nid2str(req->rq_peer.nid),
1196 export->exp_conn_cnt,
1197 lustre_msg_get_conn_cnt(req->rq_reqmsg));
1199 GOTO(out, rc = -EALREADY);
1201 LASSERT(lustre_msg_get_conn_cnt(req->rq_reqmsg) > 0);
1202 export->exp_conn_cnt = lustre_msg_get_conn_cnt(req->rq_reqmsg);
1204 /* Don't evict liblustre clients for not pinging. */
1205 if (lustre_msg_get_op_flags(req->rq_reqmsg) & MSG_CONNECT_LIBCLIENT) {
1206 export->exp_libclient = 1;
1207 spin_unlock(&export->exp_lock);
1209 spin_lock(&target->obd_dev_lock);
1210 list_del_init(&export->exp_obd_chain_timed);
1211 spin_unlock(&target->obd_dev_lock);
1213 spin_unlock(&export->exp_lock);
1216 if (export->exp_connection != NULL) {
1217 /* Check to see if connection came from another NID. */
1218 if ((export->exp_connection->c_peer.nid != req->rq_peer.nid) &&
1219 !hlist_unhashed(&export->exp_nid_hash))
1220 cfs_hash_del(export->exp_obd->obd_nid_hash,
1221 &export->exp_connection->c_peer.nid,
1222 &export->exp_nid_hash);
1224 ptlrpc_connection_put(export->exp_connection);
1227 export->exp_connection = ptlrpc_connection_get(req->rq_peer,
1230 if (hlist_unhashed(&export->exp_nid_hash)) {
1231 cfs_hash_add(export->exp_obd->obd_nid_hash,
1232 &export->exp_connection->c_peer.nid,
1233 &export->exp_nid_hash);
1236 if (target->obd_recovering && !export->exp_in_recovery && !lw_client) {
1238 __u64 transno = data->ocd_transno;
1240 spin_lock(&export->exp_lock);
1241 /* possible race with class_disconnect_stale_exports,
1242 * export may be already in the eviction process */
1243 if (export->exp_failed) {
1244 spin_unlock(&export->exp_lock);
1245 GOTO(out, rc = -ENODEV);
1247 export->exp_in_recovery = 1;
1248 export->exp_req_replay_needed = 1;
1249 export->exp_lock_replay_needed = 1;
1250 spin_unlock(&export->exp_lock);
1252 has_transno = !!(lustre_msg_get_op_flags(req->rq_reqmsg) &
1253 MSG_CONNECT_TRANSNO);
1254 if (has_transno && transno == 0)
1255 CWARN("Connect with zero transno!\n");
1257 if (has_transno && transno > 0 &&
1258 transno < target->obd_next_recovery_transno &&
1259 transno > target->obd_last_committed) {
1260 /* Another way is to use cmpxchg() to be lock-free. */
1261 spin_lock(&target->obd_recovery_task_lock);
1262 if (transno < target->obd_next_recovery_transno)
1263 target->obd_next_recovery_transno = transno;
1264 spin_unlock(&target->obd_recovery_task_lock);
1267 atomic_inc(&target->obd_req_replay_clients);
1268 atomic_inc(&target->obd_lock_replay_clients);
1269 /* Note: MDS-MDS connection is allowed to be connected during
1270 * recovery, no matter if the exports needs to be recoveried.
1271 * Because we need retrieve updates logs from all other MDTs.
1272 * So if the MDS-MDS export is new, obd_max_recoverable_clients
1273 * also needs to be increased to match other recovery checking
1275 if (new_mds_mds_conn)
1276 target->obd_max_recoverable_clients++;
1277 if (atomic_inc_return(&target->obd_connected_clients) ==
1278 target->obd_max_recoverable_clients)
1279 wake_up(&target->obd_next_transno_waitq);
1282 /* Tell the client we're in recovery, when client is involved in it. */
1283 if (target->obd_recovering && !lw_client)
1284 lustre_msg_add_op_flags(req->rq_repmsg, MSG_CONNECT_RECOVERING);
1286 tmp = req_capsule_client_get(&req->rq_pill, &RMF_CONN);
1289 /* Return -ENOTCONN in case of errors to let client reconnect. */
1290 revimp = class_new_import(target);
1291 if (revimp == NULL) {
1292 CERROR("fail to alloc new reverse import.\n");
1293 GOTO(out, rc = -ENOTCONN);
1296 spin_lock(&export->exp_lock);
1297 if (export->exp_imp_reverse != NULL)
1298 /* destroyed import can be still referenced in ctxt */
1299 tmp_imp = export->exp_imp_reverse;
1300 export->exp_imp_reverse = revimp;
1301 spin_unlock(&export->exp_lock);
1303 revimp->imp_connection = ptlrpc_connection_addref(export->exp_connection);
1304 revimp->imp_client = &export->exp_obd->obd_ldlm_client;
1305 revimp->imp_remote_handle = conn;
1306 revimp->imp_dlm_fake = 1;
1307 revimp->imp_state = LUSTRE_IMP_FULL;
1309 /* Unknown versions will be caught in
1310 * ptlrpc_handle_server_req_in->lustre_unpack_msg(). */
1311 revimp->imp_msg_magic = req->rq_reqmsg->lm_magic;
1313 if (data->ocd_connect_flags & OBD_CONNECT_AT)
1314 revimp->imp_msghdr_flags |= MSGHDR_AT_SUPPORT;
1316 revimp->imp_msghdr_flags &= ~MSGHDR_AT_SUPPORT;
1318 revimp->imp_msghdr_flags |= MSGHDR_CKSUM_INCOMPAT18;
1320 rc = sptlrpc_import_sec_adapt(revimp, req->rq_svc_ctx, &req->rq_flvr);
1322 CERROR("Failed to get sec for reverse import: %d\n", rc);
1323 spin_lock(&export->exp_lock);
1324 export->exp_imp_reverse = NULL;
1325 spin_unlock(&export->exp_lock);
1326 class_destroy_import(revimp);
1329 class_import_put(revimp);
1332 if (tmp_imp != NULL)
1333 client_destroy_import(tmp_imp);
1335 spin_lock(&export->exp_lock);
1336 export->exp_connecting = 0;
1337 spin_unlock(&export->exp_lock);
1339 class_export_put(export);
1342 spin_lock(&target->obd_dev_lock);
1343 target->obd_conn_inprogress--;
1344 spin_unlock(&target->obd_dev_lock);
1346 class_decref(targref, __func__, current);
1349 req->rq_status = rc;
1353 int target_handle_disconnect(struct ptlrpc_request *req)
1358 rc = req_capsule_server_pack(&req->rq_pill);
1362 /* Keep the rq_export around so we can send the reply. */
1363 req->rq_status = obd_disconnect(class_export_get(req->rq_export));
1368 void target_destroy_export(struct obd_export *exp)
1370 struct obd_import *imp = NULL;
1371 /* exports created from last_rcvd data, and "fake"
1372 exports created by lctl don't have an import */
1373 spin_lock(&exp->exp_lock);
1374 if (exp->exp_imp_reverse != NULL) {
1375 imp = exp->exp_imp_reverse;
1376 exp->exp_imp_reverse = NULL;
1378 spin_unlock(&exp->exp_lock);
1380 client_destroy_import(imp);
1382 LASSERT_ATOMIC_ZERO(&exp->exp_locks_count);
1383 LASSERT_ATOMIC_ZERO(&exp->exp_rpc_count);
1384 LASSERT_ATOMIC_ZERO(&exp->exp_cb_count);
1385 LASSERT_ATOMIC_ZERO(&exp->exp_replay_count);
1387 EXPORT_SYMBOL(target_destroy_export);
1390 * Recovery functions
1392 static void target_request_copy_get(struct ptlrpc_request *req)
1394 class_export_rpc_inc(req->rq_export);
1395 LASSERT(list_empty(&req->rq_list));
1396 INIT_LIST_HEAD(&req->rq_replay_list);
1398 /* Increase refcount to keep request in queue. */
1399 atomic_inc(&req->rq_refcount);
1400 /* Let export know it has replays to be handled. */
1401 atomic_inc(&req->rq_export->exp_replay_count);
1404 static void target_request_copy_put(struct ptlrpc_request *req)
1406 LASSERT(list_empty(&req->rq_replay_list));
1407 LASSERT_ATOMIC_POS(&req->rq_export->exp_replay_count);
1409 atomic_dec(&req->rq_export->exp_replay_count);
1410 class_export_rpc_dec(req->rq_export);
1411 ptlrpc_server_drop_request(req);
1414 static int target_exp_enqueue_req_replay(struct ptlrpc_request *req)
1416 __u64 transno = lustre_msg_get_transno(req->rq_reqmsg);
1417 struct obd_export *exp = req->rq_export;
1418 struct ptlrpc_request *reqiter;
1419 struct ptlrpc_request *dup_req = NULL;
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) {
1435 /* We expect it with RESENT and REPLAY flags. */
1436 if ((lustre_msg_get_flags(req->rq_reqmsg) &
1437 (MSG_RESENT | MSG_REPLAY)) != (MSG_RESENT | MSG_REPLAY))
1438 CERROR("invalid flags %x of resent replay\n",
1439 lustre_msg_get_flags(req->rq_reqmsg));
1441 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY) {
1444 new_conn = lustre_msg_get_conn_cnt(req->rq_reqmsg);
1446 lustre_msg_get_conn_cnt(dup_req->rq_reqmsg))
1447 lustre_msg_set_conn_cnt(dup_req->rq_reqmsg,
1451 list_add_tail(&req->rq_replay_list,
1452 &exp->exp_req_replay_queue);
1455 spin_unlock(&exp->exp_lock);
1459 static void target_exp_dequeue_req_replay(struct ptlrpc_request *req)
1461 LASSERT(!list_empty(&req->rq_replay_list));
1462 LASSERT(req->rq_export);
1464 spin_lock(&req->rq_export->exp_lock);
1465 list_del_init(&req->rq_replay_list);
1466 spin_unlock(&req->rq_export->exp_lock);
1469 static void target_finish_recovery(struct lu_target *lut)
1471 struct obd_device *obd = lut->lut_obd;
1474 /* Only log a recovery message when recovery has occurred. */
1475 if (obd->obd_recovery_start) {
1476 time_t elapsed_time = max_t(time_t, 1, cfs_time_current_sec() -
1477 obd->obd_recovery_start);
1478 LCONSOLE_INFO("%s: Recovery over after %d:%.02d, of %d clients "
1479 "%d recovered and %d %s evicted.\n", obd->obd_name,
1480 (int)elapsed_time / 60, (int)elapsed_time % 60,
1481 obd->obd_max_recoverable_clients,
1482 atomic_read(&obd->obd_connected_clients),
1483 obd->obd_stale_clients,
1484 obd->obd_stale_clients == 1 ? "was" : "were");
1487 ldlm_reprocess_all_ns(obd->obd_namespace);
1488 spin_lock(&obd->obd_recovery_task_lock);
1489 if (!list_empty(&obd->obd_req_replay_queue) ||
1490 !list_empty(&obd->obd_lock_replay_queue) ||
1491 !list_empty(&obd->obd_final_req_queue)) {
1492 CERROR("%s: Recovery queues ( %s%s%s) are not empty\n",
1494 list_empty(&obd->obd_req_replay_queue) ? "" : "req ",
1495 list_empty(&obd->obd_lock_replay_queue) ? \
1497 list_empty(&obd->obd_final_req_queue) ? \
1499 spin_unlock(&obd->obd_recovery_task_lock);
1502 spin_unlock(&obd->obd_recovery_task_lock);
1504 if (lut->lut_tdtd != NULL &&
1505 !list_empty(&lut->lut_tdtd->tdtd_replay_list))
1506 dtrq_list_dump(lut->lut_tdtd, D_ERROR);
1508 obd->obd_recovery_end = cfs_time_current_sec();
1510 /* When recovery finished, cleanup orphans on MDS and OST. */
1511 if (OBT(obd) && OBP(obd, postrecov)) {
1512 int rc = OBP(obd, postrecov)(obd);
1514 LCONSOLE_WARN("%s: Post recovery failed, rc %d\n",
1520 static void abort_req_replay_queue(struct obd_device *obd)
1522 struct ptlrpc_request *req, *n;
1523 struct list_head abort_list;
1525 INIT_LIST_HEAD(&abort_list);
1526 spin_lock(&obd->obd_recovery_task_lock);
1527 list_splice_init(&obd->obd_req_replay_queue, &abort_list);
1528 spin_unlock(&obd->obd_recovery_task_lock);
1529 list_for_each_entry_safe(req, n, &abort_list, rq_list) {
1530 DEBUG_REQ(D_WARNING, req, "aborted:");
1531 req->rq_status = -ENOTCONN;
1532 if (ptlrpc_error(req)) {
1533 DEBUG_REQ(D_ERROR, req,
1534 "failed abort_req_reply; skipping");
1536 target_exp_dequeue_req_replay(req);
1537 target_request_copy_put(req);
1541 static void abort_lock_replay_queue(struct obd_device *obd)
1543 struct ptlrpc_request *req, *n;
1544 struct list_head abort_list;
1546 INIT_LIST_HEAD(&abort_list);
1547 spin_lock(&obd->obd_recovery_task_lock);
1548 list_splice_init(&obd->obd_lock_replay_queue, &abort_list);
1549 spin_unlock(&obd->obd_recovery_task_lock);
1550 list_for_each_entry_safe(req, n, &abort_list, rq_list) {
1551 DEBUG_REQ(D_ERROR, req, "aborted:");
1552 req->rq_status = -ENOTCONN;
1553 if (ptlrpc_error(req)) {
1554 DEBUG_REQ(D_ERROR, req,
1555 "failed abort_lock_reply; skipping");
1557 target_request_copy_put(req);
1561 /* Called from a cleanup function if the device is being cleaned up
1562 forcefully. The exports should all have been disconnected already,
1563 the only thing left to do is
1564 - clear the recovery flags
1566 - free queued requests and replies, but don't send replies
1567 Because the obd_stopping flag is set, no new requests should be received.
1570 void target_cleanup_recovery(struct obd_device *obd)
1572 struct ptlrpc_request *req, *n;
1573 struct list_head clean_list;
1576 INIT_LIST_HEAD(&clean_list);
1577 spin_lock(&obd->obd_dev_lock);
1578 if (!obd->obd_recovering) {
1579 spin_unlock(&obd->obd_dev_lock);
1583 obd->obd_recovering = obd->obd_abort_recovery = 0;
1584 obd->obd_force_abort_recovery = 0;
1585 spin_unlock(&obd->obd_dev_lock);
1587 spin_lock(&obd->obd_recovery_task_lock);
1588 target_cancel_recovery_timer(obd);
1589 list_splice_init(&obd->obd_req_replay_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_exp_dequeue_req_replay(req);
1595 target_request_copy_put(req);
1598 spin_lock(&obd->obd_recovery_task_lock);
1599 list_splice_init(&obd->obd_lock_replay_queue, &clean_list);
1600 list_splice_init(&obd->obd_final_req_queue, &clean_list);
1601 spin_unlock(&obd->obd_recovery_task_lock);
1603 list_for_each_entry_safe(req, n, &clean_list, rq_list) {
1604 LASSERT(req->rq_reply_state == NULL);
1605 target_request_copy_put(req);
1610 EXPORT_SYMBOL(target_cleanup_recovery);
1612 /* obd_recovery_task_lock should be held */
1613 void target_cancel_recovery_timer(struct obd_device *obd)
1615 CDEBUG(D_HA, "%s: cancel recovery timer\n", obd->obd_name);
1616 cfs_timer_disarm(&obd->obd_recovery_timer);
1619 static void target_start_recovery_timer(struct obd_device *obd)
1621 if (obd->obd_recovery_start != 0)
1624 spin_lock(&obd->obd_dev_lock);
1625 if (!obd->obd_recovering || obd->obd_abort_recovery ||
1626 obd->obd_force_abort_recovery) {
1627 spin_unlock(&obd->obd_dev_lock);
1631 LASSERT(obd->obd_recovery_timeout != 0);
1633 if (obd->obd_recovery_start != 0) {
1634 spin_unlock(&obd->obd_dev_lock);
1638 cfs_timer_arm(&obd->obd_recovery_timer,
1639 cfs_time_shift(obd->obd_recovery_timeout));
1640 obd->obd_recovery_start = cfs_time_current_sec();
1641 spin_unlock(&obd->obd_dev_lock);
1643 LCONSOLE_WARN("%s: Will be in recovery for at least %d:%.02d, "
1644 "or until %d client%s reconnect%s\n",
1646 obd->obd_recovery_timeout / 60,
1647 obd->obd_recovery_timeout % 60,
1648 obd->obd_max_recoverable_clients,
1649 (obd->obd_max_recoverable_clients == 1) ? "" : "s",
1650 (obd->obd_max_recoverable_clients == 1) ? "s": "");
1654 * extend recovery window.
1656 * if @extend is true, extend recovery window to have @drt remaining at least;
1657 * otherwise, make sure the recovery timeout value is not less than @drt.
1659 static void extend_recovery_timer(struct obd_device *obd, int drt, bool extend)
1663 cfs_duration_t left;
1666 spin_lock(&obd->obd_dev_lock);
1667 if (!obd->obd_recovering || obd->obd_abort_recovery ||
1668 obd->obd_force_abort_recovery) {
1669 spin_unlock(&obd->obd_dev_lock);
1672 LASSERT(obd->obd_recovery_start != 0);
1674 now = cfs_time_current_sec();
1675 to = obd->obd_recovery_timeout;
1676 end = obd->obd_recovery_start + to;
1677 left = cfs_time_sub(end, now);
1679 if (extend && (drt > left)) {
1681 } else if (!extend && (drt > to)) {
1685 if (to > obd->obd_recovery_time_hard)
1686 to = obd->obd_recovery_time_hard;
1687 if (obd->obd_recovery_timeout < to) {
1688 obd->obd_recovery_timeout = to;
1689 end = obd->obd_recovery_start + to;
1690 cfs_timer_arm(&obd->obd_recovery_timer,
1691 cfs_time_shift(end - now));
1693 spin_unlock(&obd->obd_dev_lock);
1695 CDEBUG(D_HA, "%s: recovery timer will expire in %u seconds\n",
1696 obd->obd_name, (unsigned)cfs_time_sub(end, now));
1699 /* Reset the timer with each new client connection */
1701 * This timer is actually reconnect_timer, which is for making sure
1702 * the total recovery window is at least as big as my reconnect
1703 * attempt timing. So the initial recovery time_out will be set to
1704 * OBD_RECOVERY_FACTOR * obd_timeout. If the timeout coming
1705 * from client is bigger than this, then the recovery time_out will
1706 * be extended to make sure the client could be reconnected, in the
1707 * process, the timeout from the new client should be ignored.
1711 check_and_start_recovery_timer(struct obd_device *obd,
1712 struct ptlrpc_request *req,
1715 int service_time = lustre_msg_get_service_time(req->rq_reqmsg);
1716 struct obd_device_target *obt = &obd->u.obt;
1718 if (!new_client && service_time)
1719 /* Teach server about old server's estimates, as first guess
1720 * at how long new requests will take. */
1721 at_measured(&req->rq_rqbd->rqbd_svcpt->scp_at_estimate,
1724 target_start_recovery_timer(obd);
1726 /* Convert the service time to RPC timeout,
1727 * and reuse service_time to limit stack usage. */
1728 service_time = at_est2timeout(service_time);
1730 /* We expect other clients to timeout within service_time, then try
1731 * to reconnect, then try the failover server. The max delay between
1732 * connect attempts is SWITCH_MAX + SWITCH_INC + INITIAL. */
1733 service_time += 2 * INITIAL_CONNECT_TIMEOUT;
1735 LASSERT(obt->obt_magic == OBT_MAGIC);
1736 service_time += 2 * (CONNECTION_SWITCH_MAX + CONNECTION_SWITCH_INC);
1737 if (service_time > obd->obd_recovery_timeout && !new_client)
1738 extend_recovery_timer(obd, service_time, false);
1741 /** Health checking routines */
1742 static inline int exp_connect_healthy(struct obd_export *exp)
1744 return (exp->exp_in_recovery);
1747 /** if export done req_replay or has replay in queue */
1748 static inline int exp_req_replay_healthy(struct obd_export *exp)
1750 return (!exp->exp_req_replay_needed ||
1751 atomic_read(&exp->exp_replay_count) > 0);
1753 /** if export done lock_replay or has replay in queue */
1754 static inline int exp_lock_replay_healthy(struct obd_export *exp)
1756 return (!exp->exp_lock_replay_needed ||
1757 atomic_read(&exp->exp_replay_count) > 0);
1760 static inline int exp_vbr_healthy(struct obd_export *exp)
1762 return (!exp->exp_vbr_failed);
1765 static inline int exp_finished(struct obd_export *exp)
1767 return (exp->exp_in_recovery && !exp->exp_lock_replay_needed);
1770 static int check_for_next_transno(struct lu_target *lut)
1772 struct ptlrpc_request *req = NULL;
1773 struct obd_device *obd = lut->lut_obd;
1774 int wake_up = 0, connected, completed, queue_len;
1775 __u64 req_transno = 0;
1776 __u64 update_transno = 0;
1777 __u64 next_transno = 0;
1780 spin_lock(&obd->obd_recovery_task_lock);
1781 if (!list_empty(&obd->obd_req_replay_queue)) {
1782 req = list_entry(obd->obd_req_replay_queue.next,
1783 struct ptlrpc_request, rq_list);
1784 req_transno = lustre_msg_get_transno(req->rq_reqmsg);
1787 if (lut->lut_tdtd != NULL) {
1788 struct target_distribute_txn_data *tdtd;
1790 tdtd = lut->lut_tdtd;
1791 update_transno = distribute_txn_get_next_transno(lut->lut_tdtd);
1794 connected = atomic_read(&obd->obd_connected_clients);
1795 completed = connected - atomic_read(&obd->obd_req_replay_clients);
1796 queue_len = obd->obd_requests_queued_for_recovery;
1797 next_transno = obd->obd_next_recovery_transno;
1799 CDEBUG(D_HA, "max: %d, connected: %d, completed: %d, queue_len: %d, "
1800 "req_transno: "LPU64", next_transno: "LPU64"\n",
1801 obd->obd_max_recoverable_clients, connected, completed,
1802 queue_len, req_transno, next_transno);
1804 if (obd->obd_abort_recovery || obd->obd_force_abort_recovery) {
1805 CDEBUG(D_HA, "waking for aborted recovery\n");
1807 } else if (obd->obd_recovery_expired) {
1808 CDEBUG(D_HA, "waking for expired recovery\n");
1810 } else if (req_transno == next_transno ||
1811 (update_transno != 0 && update_transno <= next_transno)) {
1812 CDEBUG(D_HA, "waking for next ("LPD64")\n", next_transno);
1814 } else if (queue_len > 0 &&
1815 queue_len == atomic_read(&obd->obd_req_replay_clients)) {
1817 /** handle gaps occured due to lost reply or VBR */
1818 LASSERTF(req_transno >= next_transno,
1819 "req_transno: "LPU64", next_transno: "LPU64"\n",
1820 req_transno, next_transno);
1821 if (req_transno > obd->obd_last_committed &&
1822 !obd->obd_version_recov)
1825 "%s: waking for gap in transno, VBR is %s (skip: "
1826 LPD64", ql: %d, comp: %d, conn: %d, next: "LPD64
1827 ", next_update "LPD64" last_committed: "LPD64")\n",
1828 obd->obd_name, obd->obd_version_recov ? "ON" : "OFF",
1829 next_transno, queue_len, completed, connected,
1830 req_transno, update_transno, obd->obd_last_committed);
1831 obd->obd_next_recovery_transno = req_transno;
1833 } else if (atomic_read(&obd->obd_req_replay_clients) == 0) {
1834 CDEBUG(D_HA, "waking for completed recovery\n");
1836 } else if (OBD_FAIL_CHECK(OBD_FAIL_MDS_RECOVERY_ACCEPTS_GAPS)) {
1837 CDEBUG(D_HA, "accepting transno gaps is explicitly allowed"
1838 " by fail_lock, waking up ("LPD64")\n", next_transno);
1839 obd->obd_next_recovery_transno = req_transno;
1842 spin_unlock(&obd->obd_recovery_task_lock);
1846 static int check_for_next_lock(struct lu_target *lut)
1848 struct obd_device *obd = lut->lut_obd;
1851 spin_lock(&obd->obd_recovery_task_lock);
1852 if (!list_empty(&obd->obd_lock_replay_queue)) {
1853 CDEBUG(D_HA, "waking for next lock\n");
1855 } else if (atomic_read(&obd->obd_lock_replay_clients) == 0) {
1856 CDEBUG(D_HA, "waking for completed lock replay\n");
1858 } else if (obd->obd_abort_recovery || obd->obd_force_abort_recovery) {
1859 CDEBUG(D_HA, "waking for aborted recovery\n");
1861 } else if (obd->obd_recovery_expired) {
1862 CDEBUG(D_HA, "waking for expired recovery\n");
1865 spin_unlock(&obd->obd_recovery_task_lock);
1871 * wait for recovery events,
1872 * check its status with help of check_routine
1873 * evict dead clients via health_check
1875 static int target_recovery_overseer(struct lu_target *lut,
1876 int (*check_routine)(struct lu_target *),
1877 int (*health_check)(struct obd_export *))
1879 struct obd_device *obd = lut->lut_obd;
1881 if ((obd->obd_recovery_start != 0) && (cfs_time_current_sec() >=
1882 (obd->obd_recovery_start + obd->obd_recovery_time_hard))) {
1883 CWARN("recovery is aborted by hard timeout\n");
1884 obd->obd_abort_recovery = 1;
1887 while (wait_event_timeout(obd->obd_next_transno_waitq,
1889 msecs_to_jiffies(60 * MSEC_PER_SEC)) == 0)
1890 /* wait indefinitely for event, but don't trigger watchdog */;
1892 if (obd->obd_abort_recovery || obd->obd_force_abort_recovery) {
1893 CWARN("recovery is aborted, evict exports in recovery\n");
1894 /** evict exports which didn't finish recovery yet */
1895 class_disconnect_stale_exports(obd, exp_finished);
1897 } else if (obd->obd_recovery_expired) {
1898 obd->obd_recovery_expired = 0;
1899 /** If some clients died being recovered, evict them */
1900 LCONSOLE_WARN("%s: recovery is timed out, "
1901 "evict stale exports\n", obd->obd_name);
1902 /** evict cexports with no replay in queue, they are stalled */
1903 class_disconnect_stale_exports(obd, health_check);
1904 /** continue with VBR */
1905 spin_lock(&obd->obd_dev_lock);
1906 obd->obd_version_recov = 1;
1907 spin_unlock(&obd->obd_dev_lock);
1909 * reset timer, recovery will proceed with versions now,
1910 * timeout is set just to handle reconnection delays
1912 extend_recovery_timer(obd, RECONNECT_DELAY_MAX, true);
1913 /** Wait for recovery events again, after evicting bad clients */
1919 static struct ptlrpc_request *target_next_replay_lock(struct lu_target *lut)
1921 struct obd_device *obd = lut->lut_obd;
1922 struct ptlrpc_request *req = NULL;
1924 CDEBUG(D_HA, "Waiting for lock\n");
1925 if (target_recovery_overseer(lut, check_for_next_lock,
1926 exp_lock_replay_healthy))
1927 abort_lock_replay_queue(obd);
1929 spin_lock(&obd->obd_recovery_task_lock);
1930 if (!list_empty(&obd->obd_lock_replay_queue)) {
1931 req = list_entry(obd->obd_lock_replay_queue.next,
1932 struct ptlrpc_request, rq_list);
1933 list_del_init(&req->rq_list);
1934 spin_unlock(&obd->obd_recovery_task_lock);
1936 spin_unlock(&obd->obd_recovery_task_lock);
1937 LASSERT(list_empty(&obd->obd_lock_replay_queue));
1938 LASSERT(atomic_read(&obd->obd_lock_replay_clients) == 0);
1939 /** evict exports failed VBR */
1940 class_disconnect_stale_exports(obd, exp_vbr_healthy);
1945 static struct ptlrpc_request *target_next_final_ping(struct obd_device *obd)
1947 struct ptlrpc_request *req = NULL;
1949 spin_lock(&obd->obd_recovery_task_lock);
1950 if (!list_empty(&obd->obd_final_req_queue)) {
1951 req = list_entry(obd->obd_final_req_queue.next,
1952 struct ptlrpc_request, rq_list);
1953 list_del_init(&req->rq_list);
1954 spin_unlock(&obd->obd_recovery_task_lock);
1955 if (req->rq_export->exp_in_recovery) {
1956 spin_lock(&req->rq_export->exp_lock);
1957 req->rq_export->exp_in_recovery = 0;
1958 spin_unlock(&req->rq_export->exp_lock);
1961 spin_unlock(&obd->obd_recovery_task_lock);
1966 static void handle_recovery_req(struct ptlrpc_thread *thread,
1967 struct ptlrpc_request *req,
1968 svc_handler_t handler)
1973 * export can be evicted during recovery, no need to handle replays for
1974 * it after that, discard such request silently
1976 if (req->rq_export->exp_disconnected)
1979 req->rq_session.lc_thread = thread;
1980 req->rq_svc_thread = thread;
1981 req->rq_svc_thread->t_env->le_ses = &req->rq_session;
1983 /* thread context */
1984 lu_context_enter(&thread->t_env->le_ctx);
1986 lu_context_exit(&thread->t_env->le_ctx);
1988 /* don't reset timer for final stage */
1989 if (!exp_finished(req->rq_export)) {
1990 int to = obd_timeout;
1993 * Add request timeout to the recovery time so next request from
1994 * this client may come in recovery time
1997 struct ptlrpc_service_part *svcpt;
1999 svcpt = req->rq_rqbd->rqbd_svcpt;
2000 /* If the server sent early reply for this request,
2001 * the client will recalculate the timeout according to
2002 * current server estimate service time, so we will
2003 * use the maxium timeout here for waiting the client
2004 * sending the next req */
2005 to = max((int)at_est2timeout(
2006 at_get(&svcpt->scp_at_estimate)),
2007 (int)lustre_msg_get_timeout(req->rq_reqmsg));
2008 /* Add 2 net_latency, one for balance rq_deadline
2009 * (see ptl_send_rpc), one for resend the req to server,
2010 * Note: client will pack net_latency in replay req
2011 * (see ptlrpc_replay_req) */
2012 to += 2 * lustre_msg_get_service_time(req->rq_reqmsg);
2014 extend_recovery_timer(class_exp2obd(req->rq_export), to, true);
2019 /** Checking routines for recovery */
2020 static int check_for_recovery_ready(struct lu_target *lut)
2022 struct obd_device *obd = lut->lut_obd;
2023 unsigned int clnts = atomic_read(&obd->obd_connected_clients);
2025 CDEBUG(D_HA, "connected %d stale %d max_recoverable_clients %d"
2026 " abort %d expired %d\n", clnts, obd->obd_stale_clients,
2027 obd->obd_max_recoverable_clients, obd->obd_abort_recovery,
2028 obd->obd_recovery_expired);
2030 if (obd->obd_force_abort_recovery)
2033 if (!obd->obd_abort_recovery && !obd->obd_recovery_expired) {
2034 LASSERT(clnts <= obd->obd_max_recoverable_clients);
2035 if (clnts + obd->obd_stale_clients <
2036 obd->obd_max_recoverable_clients)
2040 if (lut->lut_tdtd != NULL) {
2041 if (!lut->lut_tdtd->tdtd_replay_ready) {
2042 /* Let's extend recovery timer, in case the recovery
2043 * timer expired, and some clients got evicted */
2044 extend_recovery_timer(obd, obd->obd_recovery_timeout,
2048 dtrq_list_dump(lut->lut_tdtd, D_HA);
2056 REQUEST_RECOVERY = 1,
2057 UPDATE_RECOVERY = 2,
2060 static __u64 get_next_replay_req_transno(struct obd_device *obd)
2064 if (!list_empty(&obd->obd_req_replay_queue)) {
2065 struct ptlrpc_request *req;
2067 req = list_entry(obd->obd_req_replay_queue.next,
2068 struct ptlrpc_request, rq_list);
2069 transno = lustre_msg_get_transno(req->rq_reqmsg);
2074 __u64 get_next_transno(struct lu_target *lut, int *type)
2076 struct obd_device *obd = lut->lut_obd;
2077 struct target_distribute_txn_data *tdtd = lut->lut_tdtd;
2079 __u64 update_transno;
2082 transno = get_next_replay_req_transno(obd);
2084 *type = REQUEST_RECOVERY;
2089 update_transno = distribute_txn_get_next_transno(tdtd);
2090 if (transno == 0 || (transno >= update_transno &&
2091 update_transno != 0)) {
2092 transno = update_transno;
2094 *type = UPDATE_RECOVERY;
2101 * drop duplicate replay request
2103 * Because the operation has been replayed by update recovery, the request
2104 * with the same transno will be dropped and also notify the client to send
2105 * next replay request.
2107 * \param[in] env execution environment
2108 * \param[in] obd failover obd device
2109 * \param[in] req request to be dropped
2111 static void drop_duplicate_replay_req(struct lu_env *env,
2112 struct obd_device *obd,
2113 struct ptlrpc_request *req)
2115 DEBUG_REQ(D_HA, req, "remove t"LPD64" from %s because of duplicate"
2116 " update records are found.\n",
2117 lustre_msg_get_transno(req->rq_reqmsg),
2118 libcfs_nid2str(req->rq_peer.nid));
2120 /* Right now, only for MDS reint operation update replay and
2121 * normal request replay can have the same transno */
2122 if (lustre_msg_get_opc(req->rq_reqmsg) == MDS_REINT) {
2123 req_capsule_set(&req->rq_pill, &RQF_MDS_REINT);
2124 req->rq_status = req_capsule_server_pack(&req->rq_pill);
2125 if (likely(req->rq_export))
2126 target_committed_to_req(req);
2127 lustre_msg_set_transno(req->rq_repmsg, req->rq_transno);
2128 target_send_reply(req, req->rq_status, 0);
2130 DEBUG_REQ(D_ERROR, req, "wrong opc" "from %s\n",
2131 libcfs_nid2str(req->rq_peer.nid));
2133 target_exp_dequeue_req_replay(req);
2134 target_request_copy_put(req);
2135 obd->obd_replayed_requests++;
2139 * Update last_rcvd of the update
2141 * Because update recovery might update the last_rcvd by updates, i.e.
2142 * it will not update the last_rcvd information in memory, so we need
2143 * refresh these information in memory after update recovery.
2145 * \param[in] obd obd_device under recoverying.
2146 * \param[in] dtrq the update replay requests being replayed.
2148 static void target_update_lcd(struct lu_env *env, struct lu_target *lut,
2149 struct distribute_txn_replay_req *dtrq)
2151 struct obd_device *obd = lut->lut_obd;
2152 struct obd_export *export;
2153 struct tg_export_data *ted;
2154 struct distribute_txn_replay_req_sub *dtrqs;
2155 struct seq_server_site *site;
2156 struct update_records *ur;
2157 const struct lu_fid *fid;
2158 struct update_ops *ops;
2159 struct update_params *params;
2160 struct update_op *op;
2163 struct lsd_client_data *lcd = NULL;
2165 /* if Updates has been executed(committed) on the recovery target,
2166 * i.e. the updates is not being executed on the target, so we do
2167 * not need update it in memory */
2168 site = lu_site2seq(obd->obd_lu_dev->ld_site);
2169 mdt_index = site->ss_node_id;
2170 dtrqs = dtrq_sub_lookup(dtrq, mdt_index);
2174 if (dtrq->dtrq_lur == NULL)
2177 /* Find the update last_rcvd record */
2178 fid = lu_object_fid(&lut->lut_last_rcvd->do_lu);
2179 ur = &dtrq->dtrq_lur->lur_update_rec;
2181 params = update_records_get_params(ur);
2182 for (i = 0, op = &ops->uops_op[0]; i < ur->ur_update_count;
2183 i++, op = update_op_next_op(op)) {
2188 if (!lu_fid_eq(&op->uop_fid, fid))
2191 if (op->uop_type != OUT_WRITE)
2194 buf = update_params_get_param_buf(params, op->uop_params_off[1],
2195 ur->ur_param_count, NULL);
2199 pos = le64_to_cpu(*(__u64 *)buf);
2203 buf = update_params_get_param_buf(params, op->uop_params_off[0],
2204 ur->ur_param_count, &size);
2208 if (size != sizeof(*lcd))
2213 if (lcd == NULL || lcd->lcd_uuid[0] == '\0')
2216 /* locate the export then update the exp_target_data if needed */
2217 export = cfs_hash_lookup(obd->obd_uuid_hash, lcd->lcd_uuid);
2221 ted = &export->exp_target_data;
2222 if (lcd->lcd_last_xid > ted->ted_lcd->lcd_last_xid) {
2223 CDEBUG(D_HA, "%s update xid from "LPU64" to "LPU64"\n",
2224 lut->lut_obd->obd_name, ted->ted_lcd->lcd_last_xid,
2226 ted->ted_lcd->lcd_last_xid = lcd->lcd_last_xid;
2227 ted->ted_lcd->lcd_last_result = lcd->lcd_last_result;
2229 class_export_put(export);
2232 static void replay_request_or_update(struct lu_env *env,
2233 struct lu_target *lut,
2234 struct target_recovery_data *trd,
2235 struct ptlrpc_thread *thread)
2237 struct obd_device *obd = lut->lut_obd;
2238 struct ptlrpc_request *req = NULL;
2243 CDEBUG(D_HA, "Waiting for transno "LPD64"\n",
2244 obd->obd_next_recovery_transno);
2246 /* Replay all of request and update by transno */
2248 struct target_distribute_txn_data *tdtd = lut->lut_tdtd;
2250 CFS_FAIL_TIMEOUT(OBD_FAIL_TGT_REPLAY_DELAY2, cfs_fail_val);
2252 /** It is needed to extend recovery window above
2253 * recovery_time_soft. Extending is possible only in the
2254 * end of recovery window (see more details in
2255 * handle_recovery_req()).
2257 CFS_FAIL_TIMEOUT_MS(OBD_FAIL_TGT_REPLAY_DELAY, 300);
2259 if (target_recovery_overseer(lut, check_for_next_transno,
2260 exp_req_replay_healthy)) {
2261 abort_req_replay_queue(obd);
2262 abort_lock_replay_queue(obd);
2265 spin_lock(&obd->obd_recovery_task_lock);
2266 transno = get_next_transno(lut, &type);
2267 if (type == REQUEST_RECOVERY && tdtd != NULL &&
2268 transno == tdtd->tdtd_last_update_transno) {
2269 /* Drop replay request from client side, if the
2270 * replay has been executed by update with the
2272 req = list_entry(obd->obd_req_replay_queue.next,
2273 struct ptlrpc_request, rq_list);
2274 list_del_init(&req->rq_list);
2275 obd->obd_requests_queued_for_recovery--;
2276 spin_unlock(&obd->obd_recovery_task_lock);
2277 drop_duplicate_replay_req(env, obd, req);
2278 } else if (type == REQUEST_RECOVERY && transno != 0) {
2279 req = list_entry(obd->obd_req_replay_queue.next,
2280 struct ptlrpc_request, rq_list);
2281 list_del_init(&req->rq_list);
2282 obd->obd_requests_queued_for_recovery--;
2283 spin_unlock(&obd->obd_recovery_task_lock);
2284 LASSERT(trd->trd_processing_task == current_pid());
2285 DEBUG_REQ(D_HA, req, "processing t"LPD64" from %s",
2286 lustre_msg_get_transno(req->rq_reqmsg),
2287 libcfs_nid2str(req->rq_peer.nid));
2289 handle_recovery_req(thread, req,
2290 trd->trd_recovery_handler);
2292 * bz18031: increase next_recovery_transno before
2293 * target_request_copy_put() will drop exp_rpc reference
2295 spin_lock(&obd->obd_recovery_task_lock);
2296 obd->obd_next_recovery_transno++;
2297 spin_unlock(&obd->obd_recovery_task_lock);
2298 target_exp_dequeue_req_replay(req);
2299 target_request_copy_put(req);
2300 obd->obd_replayed_requests++;
2301 } else if (type == UPDATE_RECOVERY && transno != 0) {
2302 struct distribute_txn_replay_req *dtrq;
2304 spin_unlock(&obd->obd_recovery_task_lock);
2306 LASSERT(tdtd != NULL);
2307 dtrq = distribute_txn_get_next_req(tdtd);
2308 lu_context_enter(&thread->t_env->le_ctx);
2309 tdtd->tdtd_replay_handler(env, tdtd, dtrq);
2310 lu_context_exit(&thread->t_env->le_ctx);
2311 extend_recovery_timer(obd, obd_timeout, true);
2312 LASSERT(tdtd->tdtd_last_update_transno <= transno);
2313 tdtd->tdtd_last_update_transno = transno;
2314 spin_lock(&obd->obd_recovery_task_lock);
2315 if (transno > obd->obd_next_recovery_transno)
2316 obd->obd_next_recovery_transno = transno;
2317 spin_unlock(&obd->obd_recovery_task_lock);
2318 target_update_lcd(env, lut, dtrq);
2321 spin_unlock(&obd->obd_recovery_task_lock);
2322 LASSERT(list_empty(&obd->obd_req_replay_queue));
2323 LASSERT(atomic_read(&obd->obd_req_replay_clients) == 0);
2324 /** evict exports failed VBR */
2325 class_disconnect_stale_exports(obd, exp_vbr_healthy);
2331 static int target_recovery_thread(void *arg)
2333 struct lu_target *lut = arg;
2334 struct obd_device *obd = lut->lut_obd;
2335 struct ptlrpc_request *req;
2336 struct target_recovery_data *trd = &obd->obd_recovery_data;
2337 unsigned long delta;
2339 struct ptlrpc_thread *thread = NULL;
2343 unshare_fs_struct();
2344 OBD_ALLOC_PTR(thread);
2350 OBD_FREE_PTR(thread);
2354 rc = lu_context_init(&env->le_ctx, LCT_MD_THREAD | LCT_DT_THREAD);
2356 OBD_FREE_PTR(thread);
2361 thread->t_env = env;
2362 thread->t_id = -1; /* force filter_iobuf_get/put to use local buffers */
2363 env->le_ctx.lc_thread = thread;
2364 tgt_io_thread_init(thread); /* init thread_big_cache for IO requests */
2365 thread->t_watchdog = NULL;
2367 CDEBUG(D_HA, "%s: started recovery thread pid %d\n", obd->obd_name,
2369 trd->trd_processing_task = current_pid();
2371 spin_lock(&obd->obd_dev_lock);
2372 obd->obd_recovering = 1;
2373 spin_unlock(&obd->obd_dev_lock);
2374 complete(&trd->trd_starting);
2376 /* first of all, we have to know the first transno to replay */
2377 if (target_recovery_overseer(lut, check_for_recovery_ready,
2378 exp_connect_healthy)) {
2379 abort_req_replay_queue(obd);
2380 abort_lock_replay_queue(obd);
2381 if (lut->lut_tdtd != NULL)
2382 dtrq_list_destroy(lut->lut_tdtd);
2385 /* next stage: replay requests or update */
2387 CDEBUG(D_INFO, "1: request replay stage - %d clients from t"LPU64"\n",
2388 atomic_read(&obd->obd_req_replay_clients),
2389 obd->obd_next_recovery_transno);
2390 replay_request_or_update(env, lut, trd, thread);
2393 * The second stage: replay locks
2395 CDEBUG(D_INFO, "2: lock replay stage - %d clients\n",
2396 atomic_read(&obd->obd_lock_replay_clients));
2397 while ((req = target_next_replay_lock(lut))) {
2398 LASSERT(trd->trd_processing_task == current_pid());
2399 DEBUG_REQ(D_HA, req, "processing lock from %s: ",
2400 libcfs_nid2str(req->rq_peer.nid));
2401 handle_recovery_req(thread, req,
2402 trd->trd_recovery_handler);
2403 target_request_copy_put(req);
2404 obd->obd_replayed_locks++;
2408 * The third stage: reply on final pings, at this moment all clients
2409 * must have request in final queue
2411 CFS_FAIL_TIMEOUT(OBD_FAIL_TGT_REPLAY_RECONNECT, cfs_fail_val);
2412 CDEBUG(D_INFO, "3: final stage - process recovery completion pings\n");
2413 /** Update server last boot epoch */
2414 tgt_boot_epoch_update(lut);
2415 /* We drop recoverying flag to forward all new requests
2416 * to regular mds_handle() since now */
2417 spin_lock(&obd->obd_dev_lock);
2418 obd->obd_recovering = obd->obd_abort_recovery = 0;
2419 spin_unlock(&obd->obd_dev_lock);
2420 spin_lock(&obd->obd_recovery_task_lock);
2421 target_cancel_recovery_timer(obd);
2422 spin_unlock(&obd->obd_recovery_task_lock);
2423 while ((req = target_next_final_ping(obd))) {
2424 LASSERT(trd->trd_processing_task == current_pid());
2425 DEBUG_REQ(D_HA, req, "processing final ping from %s: ",
2426 libcfs_nid2str(req->rq_peer.nid));
2427 handle_recovery_req(thread, req,
2428 trd->trd_recovery_handler);
2429 /* Because the waiting client can not send ping to server,
2430 * so we need refresh the last_request_time, to avoid the
2431 * export is being evicted */
2432 ptlrpc_update_export_timer(req->rq_export, 0);
2433 target_request_copy_put(req);
2436 delta = jiffies_to_msecs(jiffies - delta) / MSEC_PER_SEC;
2437 CDEBUG(D_INFO,"4: recovery completed in %lus - %d/%d reqs/locks\n",
2438 delta, obd->obd_replayed_requests, obd->obd_replayed_locks);
2439 if (delta > OBD_RECOVERY_TIME_SOFT) {
2440 CWARN("too long recovery - read logs\n");
2441 libcfs_debug_dumplog();
2444 target_finish_recovery(lut);
2446 lu_context_fini(&env->le_ctx);
2447 trd->trd_processing_task = 0;
2448 complete(&trd->trd_finishing);
2450 tgt_io_thread_done(thread);
2451 OBD_FREE_PTR(thread);
2456 static int target_start_recovery_thread(struct lu_target *lut,
2457 svc_handler_t handler)
2459 struct obd_device *obd = lut->lut_obd;
2461 struct target_recovery_data *trd = &obd->obd_recovery_data;
2463 memset(trd, 0, sizeof(*trd));
2464 init_completion(&trd->trd_starting);
2465 init_completion(&trd->trd_finishing);
2466 trd->trd_recovery_handler = handler;
2468 if (!IS_ERR(kthread_run(target_recovery_thread,
2469 lut, "tgt_recov"))) {
2470 wait_for_completion(&trd->trd_starting);
2471 LASSERT(obd->obd_recovering != 0);
2479 void target_stop_recovery_thread(struct obd_device *obd)
2481 if (obd->obd_recovery_data.trd_processing_task > 0) {
2482 struct target_recovery_data *trd = &obd->obd_recovery_data;
2483 /** recovery can be done but postrecovery is not yet */
2484 spin_lock(&obd->obd_dev_lock);
2485 if (obd->obd_recovering) {
2486 CERROR("%s: Aborting recovery\n", obd->obd_name);
2487 obd->obd_abort_recovery = 1;
2488 wake_up(&obd->obd_next_transno_waitq);
2490 spin_unlock(&obd->obd_dev_lock);
2491 wait_for_completion(&trd->trd_finishing);
2494 EXPORT_SYMBOL(target_stop_recovery_thread);
2496 void target_recovery_fini(struct obd_device *obd)
2498 class_disconnect_exports(obd);
2499 target_stop_recovery_thread(obd);
2500 target_cleanup_recovery(obd);
2502 EXPORT_SYMBOL(target_recovery_fini);
2504 static void target_recovery_expired(unsigned long castmeharder)
2506 struct obd_device *obd = (struct obd_device *)castmeharder;
2507 CDEBUG(D_HA, "%s: recovery timed out; %d clients are still in recovery"
2508 " after %lds (%d clients connected)\n",
2509 obd->obd_name, atomic_read(&obd->obd_lock_replay_clients),
2510 cfs_time_current_sec()- obd->obd_recovery_start,
2511 atomic_read(&obd->obd_connected_clients));
2513 obd->obd_recovery_expired = 1;
2514 wake_up(&obd->obd_next_transno_waitq);
2517 void target_recovery_init(struct lu_target *lut, svc_handler_t handler)
2519 struct obd_device *obd = lut->lut_obd;
2521 if (obd->obd_max_recoverable_clients == 0) {
2522 /** Update server last boot epoch */
2523 tgt_boot_epoch_update(lut);
2527 CDEBUG(D_HA, "RECOVERY: service %s, %d recoverable clients, "
2528 "last_transno "LPU64"\n", obd->obd_name,
2529 obd->obd_max_recoverable_clients, obd->obd_last_committed);
2530 LASSERT(obd->obd_stopping == 0);
2531 obd->obd_next_recovery_transno = obd->obd_last_committed + 1;
2532 obd->obd_recovery_start = 0;
2533 obd->obd_recovery_end = 0;
2535 cfs_timer_init(&obd->obd_recovery_timer, target_recovery_expired, obd);
2536 target_start_recovery_thread(lut, handler);
2538 EXPORT_SYMBOL(target_recovery_init);
2540 static int target_process_req_flags(struct obd_device *obd,
2541 struct ptlrpc_request *req)
2543 struct obd_export *exp = req->rq_export;
2544 LASSERT(exp != NULL);
2545 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REQ_REPLAY_DONE) {
2546 /* client declares he's ready to replay locks */
2547 spin_lock(&exp->exp_lock);
2548 if (exp->exp_req_replay_needed) {
2549 exp->exp_req_replay_needed = 0;
2550 spin_unlock(&exp->exp_lock);
2552 LASSERT_ATOMIC_POS(&obd->obd_req_replay_clients);
2553 atomic_dec(&obd->obd_req_replay_clients);
2555 spin_unlock(&exp->exp_lock);
2558 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_LOCK_REPLAY_DONE) {
2559 /* client declares he's ready to complete recovery
2560 * so, we put the request on th final queue */
2561 spin_lock(&exp->exp_lock);
2562 if (exp->exp_lock_replay_needed) {
2563 exp->exp_lock_replay_needed = 0;
2564 spin_unlock(&exp->exp_lock);
2566 LASSERT_ATOMIC_POS(&obd->obd_lock_replay_clients);
2567 atomic_dec(&obd->obd_lock_replay_clients);
2569 spin_unlock(&exp->exp_lock);
2575 int target_queue_recovery_request(struct ptlrpc_request *req,
2576 struct obd_device *obd)
2578 __u64 transno = lustre_msg_get_transno(req->rq_reqmsg);
2579 struct ptlrpc_request *reqiter;
2583 if (obd->obd_recovery_data.trd_processing_task == current_pid()) {
2584 /* Processing the queue right now, don't re-add. */
2588 target_process_req_flags(obd, req);
2590 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_LOCK_REPLAY_DONE) {
2591 /* client declares he's ready to complete recovery
2592 * so, we put the request on th final queue */
2593 target_request_copy_get(req);
2594 DEBUG_REQ(D_HA, req, "queue final req");
2595 wake_up(&obd->obd_next_transno_waitq);
2596 spin_lock(&obd->obd_recovery_task_lock);
2597 if (obd->obd_recovering) {
2598 list_add_tail(&req->rq_list,
2599 &obd->obd_final_req_queue);
2601 spin_unlock(&obd->obd_recovery_task_lock);
2602 target_request_copy_put(req);
2603 RETURN(obd->obd_stopping ? -ENOTCONN : 1);
2605 spin_unlock(&obd->obd_recovery_task_lock);
2608 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REQ_REPLAY_DONE) {
2609 /* client declares he's ready to replay locks */
2610 target_request_copy_get(req);
2611 DEBUG_REQ(D_HA, req, "queue lock replay req");
2612 wake_up(&obd->obd_next_transno_waitq);
2613 spin_lock(&obd->obd_recovery_task_lock);
2614 LASSERT(obd->obd_recovering);
2615 /* usually due to recovery abort */
2616 if (!req->rq_export->exp_in_recovery) {
2617 spin_unlock(&obd->obd_recovery_task_lock);
2618 target_request_copy_put(req);
2621 LASSERT(req->rq_export->exp_lock_replay_needed);
2622 list_add_tail(&req->rq_list, &obd->obd_lock_replay_queue);
2623 spin_unlock(&obd->obd_recovery_task_lock);
2627 /* CAVEAT EMPTOR: The incoming request message has been swabbed
2628 * (i.e. buflens etc are in my own byte order), but type-dependent
2629 * buffers (eg mdt_body, ost_body etc) have NOT been swabbed. */
2632 INIT_LIST_HEAD(&req->rq_list);
2633 DEBUG_REQ(D_HA, req, "not queueing");
2637 /* If we're processing the queue, we want don't want to queue this
2640 * Also, if this request has a transno less than the one we're waiting
2641 * for, we should process it now. It could (and currently always will)
2642 * be an open request for a descriptor that was opened some time ago.
2644 * Also, a resent, replayed request that has already been
2645 * handled will pass through here and be processed immediately.
2647 CDEBUG(D_HA, "Next recovery transno: "LPU64
2648 ", current: "LPU64", replaying\n",
2649 obd->obd_next_recovery_transno, transno);
2650 spin_lock(&obd->obd_recovery_task_lock);
2651 if (transno < obd->obd_next_recovery_transno) {
2652 /* Processing the queue right now, don't re-add. */
2653 LASSERT(list_empty(&req->rq_list));
2654 spin_unlock(&obd->obd_recovery_task_lock);
2657 spin_unlock(&obd->obd_recovery_task_lock);
2659 if (OBD_FAIL_CHECK(OBD_FAIL_TGT_REPLAY_DROP))
2662 target_request_copy_get(req);
2663 if (!req->rq_export->exp_in_recovery) {
2664 target_request_copy_put(req);
2667 LASSERT(req->rq_export->exp_req_replay_needed);
2669 if (target_exp_enqueue_req_replay(req)) {
2670 DEBUG_REQ(D_ERROR, req, "dropping resent queued req");
2671 target_request_copy_put(req);
2676 spin_lock(&obd->obd_recovery_task_lock);
2677 LASSERT(obd->obd_recovering);
2678 list_for_each_entry(reqiter, &obd->obd_req_replay_queue, rq_list) {
2679 if (lustre_msg_get_transno(reqiter->rq_reqmsg) > transno) {
2680 list_add_tail(&req->rq_list, &reqiter->rq_list);
2685 if (unlikely(lustre_msg_get_transno(reqiter->rq_reqmsg) ==
2687 DEBUG_REQ(D_ERROR, req, "dropping replay: transno "
2688 "has been claimed by another client");
2689 spin_unlock(&obd->obd_recovery_task_lock);
2690 target_exp_dequeue_req_replay(req);
2691 target_request_copy_put(req);
2697 list_add_tail(&req->rq_list, &obd->obd_req_replay_queue);
2699 obd->obd_requests_queued_for_recovery++;
2700 spin_unlock(&obd->obd_recovery_task_lock);
2701 wake_up(&obd->obd_next_transno_waitq);
2705 int target_handle_ping(struct ptlrpc_request *req)
2707 obd_ping(req->rq_svc_thread->t_env, req->rq_export);
2708 return req_capsule_server_pack(&req->rq_pill);
2711 void target_committed_to_req(struct ptlrpc_request *req)
2713 struct obd_export *exp = req->rq_export;
2715 if (!exp->exp_obd->obd_no_transno && req->rq_repmsg != NULL)
2716 lustre_msg_set_last_committed(req->rq_repmsg,
2717 exp->exp_last_committed);
2719 DEBUG_REQ(D_IOCTL, req, "not sending last_committed update (%d/"
2720 "%d)", exp->exp_obd->obd_no_transno,
2721 req->rq_repmsg == NULL);
2723 CDEBUG(D_INFO, "last_committed "LPU64", transno "LPU64", xid "LPU64"\n",
2724 exp->exp_last_committed, req->rq_transno, req->rq_xid);
2727 #endif /* HAVE_SERVER_SUPPORT */
2730 * Packs current SLV and Limit into \a req.
2732 int target_pack_pool_reply(struct ptlrpc_request *req)
2734 struct obd_device *obd;
2737 /* Check that we still have all structures alive as this may
2738 * be some late RPC at shutdown time. */
2739 if (unlikely(!req->rq_export || !req->rq_export->exp_obd ||
2740 !exp_connect_lru_resize(req->rq_export))) {
2741 lustre_msg_set_slv(req->rq_repmsg, 0);
2742 lustre_msg_set_limit(req->rq_repmsg, 0);
2746 /* OBD is alive here as export is alive, which we checked above. */
2747 obd = req->rq_export->exp_obd;
2749 read_lock(&obd->obd_pool_lock);
2750 lustre_msg_set_slv(req->rq_repmsg, obd->obd_pool_slv);
2751 lustre_msg_set_limit(req->rq_repmsg, obd->obd_pool_limit);
2752 read_unlock(&obd->obd_pool_lock);
2757 static int target_send_reply_msg(struct ptlrpc_request *req,
2758 int rc, int fail_id)
2760 if (OBD_FAIL_CHECK_ORSET(fail_id & ~OBD_FAIL_ONCE, OBD_FAIL_ONCE)) {
2761 DEBUG_REQ(D_ERROR, req, "dropping reply");
2764 if (unlikely(lustre_msg_get_opc(req->rq_reqmsg) == MDS_REINT &&
2765 OBD_FAIL_CHECK(OBD_FAIL_MDS_REINT_MULTI_NET_REP)))
2769 DEBUG_REQ(D_NET, req, "processing error (%d)", rc);
2770 req->rq_status = rc;
2771 return ptlrpc_send_error(req, 1);
2773 DEBUG_REQ(D_NET, req, "sending reply");
2776 return ptlrpc_send_reply(req, PTLRPC_REPLY_MAYBE_DIFFICULT);
2779 void target_send_reply(struct ptlrpc_request *req, int rc, int fail_id)
2781 struct ptlrpc_service_part *svcpt;
2783 struct ptlrpc_reply_state *rs;
2784 struct obd_export *exp;
2787 if (req->rq_no_reply) {
2792 svcpt = req->rq_rqbd->rqbd_svcpt;
2793 rs = req->rq_reply_state;
2794 if (rs == NULL || !rs->rs_difficult) {
2796 target_send_reply_msg (req, rc, fail_id);
2801 /* must be an export if locks saved */
2802 LASSERT(req->rq_export != NULL);
2803 /* req/reply consistent */
2804 LASSERT(rs->rs_svcpt == svcpt);
2807 LASSERT(!rs->rs_scheduled);
2808 LASSERT(!rs->rs_scheduled_ever);
2809 LASSERT(!rs->rs_handled);
2810 LASSERT(!rs->rs_on_net);
2811 LASSERT(rs->rs_export == NULL);
2812 LASSERT(list_empty(&rs->rs_obd_list));
2813 LASSERT(list_empty(&rs->rs_exp_list));
2815 exp = class_export_get(req->rq_export);
2817 /* disable reply scheduling while I'm setting up */
2818 rs->rs_scheduled = 1;
2820 rs->rs_xid = req->rq_xid;
2821 rs->rs_transno = req->rq_transno;
2822 rs->rs_export = exp;
2823 rs->rs_opc = lustre_msg_get_opc(req->rq_reqmsg);
2825 spin_lock(&exp->exp_uncommitted_replies_lock);
2826 CDEBUG(D_NET, "rs transno = "LPU64", last committed = "LPU64"\n",
2827 rs->rs_transno, exp->exp_last_committed);
2828 if (rs->rs_transno > exp->exp_last_committed) {
2829 /* not committed already */
2830 list_add_tail(&rs->rs_obd_list,
2831 &exp->exp_uncommitted_replies);
2833 spin_unlock(&exp->exp_uncommitted_replies_lock);
2835 spin_lock(&exp->exp_lock);
2836 list_add_tail(&rs->rs_exp_list, &exp->exp_outstanding_replies);
2837 spin_unlock(&exp->exp_lock);
2839 netrc = target_send_reply_msg(req, rc, fail_id);
2841 spin_lock(&svcpt->scp_rep_lock);
2843 atomic_inc(&svcpt->scp_nreps_difficult);
2846 /* error sending: reply is off the net. Also we need +1
2847 * reply ref until ptlrpc_handle_rs() is done
2848 * with the reply state (if the send was successful, there
2849 * would have been +1 ref for the net, which
2850 * reply_out_callback leaves alone) */
2852 ptlrpc_rs_addref(rs);
2855 spin_lock(&rs->rs_lock);
2856 if (rs->rs_transno <= exp->exp_last_committed ||
2857 (!rs->rs_on_net && !rs->rs_no_ack) ||
2858 list_empty(&rs->rs_exp_list) || /* completed already */
2859 list_empty(&rs->rs_obd_list)) {
2860 CDEBUG(D_HA, "Schedule reply immediately\n");
2861 ptlrpc_dispatch_difficult_reply(rs);
2863 list_add(&rs->rs_list, &svcpt->scp_rep_active);
2864 rs->rs_scheduled = 0; /* allow notifier to schedule */
2866 spin_unlock(&rs->rs_lock);
2867 spin_unlock(&svcpt->scp_rep_lock);
2871 ldlm_mode_t lck_compat_array[] = {
2872 [LCK_EX] = LCK_COMPAT_EX,
2873 [LCK_PW] = LCK_COMPAT_PW,
2874 [LCK_PR] = LCK_COMPAT_PR,
2875 [LCK_CW] = LCK_COMPAT_CW,
2876 [LCK_CR] = LCK_COMPAT_CR,
2877 [LCK_NL] = LCK_COMPAT_NL,
2878 [LCK_GROUP] = LCK_COMPAT_GROUP,
2879 [LCK_COS] = LCK_COMPAT_COS,
2883 * Rather arbitrary mapping from LDLM error codes to errno values. This should
2884 * not escape to the user level.
2886 int ldlm_error2errno(ldlm_error_t error)
2892 case ELDLM_LOCK_MATCHED:
2895 case ELDLM_LOCK_CHANGED:
2898 case ELDLM_LOCK_ABORTED:
2901 case ELDLM_LOCK_REPLACED:
2904 case ELDLM_NO_LOCK_DATA:
2907 case ELDLM_NAMESPACE_EXISTS:
2910 case ELDLM_BAD_NAMESPACE:
2914 if (((int)error) < 0) /* cast to signed type */
2915 result = error; /* as ldlm_error_t can be unsigned */
2917 CERROR("Invalid DLM result code: %d\n", error);
2923 EXPORT_SYMBOL(ldlm_error2errno);
2926 * Dual to ldlm_error2errno(): maps errno values back to ldlm_error_t.
2928 ldlm_error_t ldlm_errno2error(int err_no)
2937 error = ELDLM_LOCK_CHANGED;
2940 error = ELDLM_LOCK_ABORTED;
2943 error = ELDLM_LOCK_REPLACED;
2946 error = ELDLM_NO_LOCK_DATA;
2949 error = ELDLM_NAMESPACE_EXISTS;
2952 error = ELDLM_BAD_NAMESPACE;
2960 #if LUSTRE_TRACKS_LOCK_EXP_REFS
2961 void ldlm_dump_export_locks(struct obd_export *exp)
2963 spin_lock(&exp->exp_locks_list_guard);
2964 if (!list_empty(&exp->exp_locks_list)) {
2965 struct ldlm_lock *lock;
2967 CERROR("dumping locks for export %p,"
2968 "ignore if the unmount doesn't hang\n", exp);
2969 list_for_each_entry(lock, &exp->exp_locks_list,
2971 LDLM_ERROR(lock, "lock:");
2973 spin_unlock(&exp->exp_locks_list_guard);
2977 #ifdef HAVE_SERVER_SUPPORT
2978 static int target_bulk_timeout(void *data)
2981 /* We don't fail the connection here, because having the export
2982 * killed makes the (vital) call to commitrw very sad.
2987 static inline const char *bulk2type(struct ptlrpc_request *req)
2989 if (req->rq_bulk_read)
2991 if (req->rq_bulk_write)
2996 int target_bulk_io(struct obd_export *exp, struct ptlrpc_bulk_desc *desc,
2997 struct l_wait_info *lwi)
2999 struct ptlrpc_request *req = desc->bd_req;
3000 time_t start = cfs_time_current_sec();
3006 /* If there is eviction in progress, wait for it to finish. */
3007 if (unlikely(atomic_read(&exp->exp_obd->obd_evict_inprogress))) {
3008 *lwi = LWI_INTR(NULL, NULL);
3009 rc = l_wait_event(exp->exp_obd->obd_evict_inprogress_waitq,
3010 !atomic_read(&exp->exp_obd->
3011 obd_evict_inprogress),
3015 /* Check if client was evicted or reconnected already. */
3016 if (exp->exp_failed ||
3017 exp->exp_conn_cnt > lustre_msg_get_conn_cnt(req->rq_reqmsg)) {
3020 if (req->rq_bulk_read)
3021 rc = sptlrpc_svc_wrap_bulk(req, desc);
3023 if ((exp->exp_connect_data.ocd_connect_flags &
3024 OBD_CONNECT_BULK_MBITS) != 0)
3025 req->rq_mbits = lustre_msg_get_mbits(req->rq_reqmsg);
3026 else /* old version, bulk matchbits is rq_xid */
3027 req->rq_mbits = req->rq_xid;
3030 rc = ptlrpc_start_bulk_transfer(desc);
3034 DEBUG_REQ(D_ERROR, req, "bulk %s failed: rc %d",
3035 bulk2type(req), rc);
3039 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_SENDPAGE)) {
3040 ptlrpc_abort_bulk(desc);
3044 /* limit actual bulk transfer to bulk_timeout seconds */
3045 deadline = start + bulk_timeout;
3046 if (deadline > req->rq_deadline)
3047 deadline = req->rq_deadline;
3050 long timeoutl = deadline - cfs_time_current_sec();
3051 cfs_duration_t timeout = timeoutl <= 0 ?
3052 CFS_TICK : cfs_time_seconds(timeoutl);
3055 *lwi = LWI_TIMEOUT_INTERVAL(timeout, cfs_time_seconds(1),
3056 target_bulk_timeout, desc);
3057 rc = l_wait_event(desc->bd_waitq,
3058 !ptlrpc_server_bulk_active(desc) ||
3061 lustre_msg_get_conn_cnt(req->rq_reqmsg),
3063 LASSERT(rc == 0 || rc == -ETIMEDOUT);
3064 /* Wait again if we changed rq_deadline. */
3065 rq_deadline = ACCESS_ONCE(req->rq_deadline);
3066 deadline = start + bulk_timeout;
3067 if (deadline > rq_deadline)
3068 deadline = rq_deadline;
3069 } while ((rc == -ETIMEDOUT) &&
3070 (deadline > cfs_time_current_sec()));
3072 if (rc == -ETIMEDOUT) {
3073 DEBUG_REQ(D_ERROR, req, "timeout on bulk %s after %ld%+lds",
3074 bulk2type(req), deadline - start,
3075 cfs_time_current_sec() - deadline);
3076 ptlrpc_abort_bulk(desc);
3077 } else if (exp->exp_failed) {
3078 DEBUG_REQ(D_ERROR, req, "Eviction on bulk %s",
3081 ptlrpc_abort_bulk(desc);
3082 } else if (exp->exp_conn_cnt >
3083 lustre_msg_get_conn_cnt(req->rq_reqmsg)) {
3084 DEBUG_REQ(D_ERROR, req, "Reconnect on bulk %s",
3086 /* We don't reply anyway. */
3088 ptlrpc_abort_bulk(desc);
3089 } else if (desc->bd_failure) {
3090 DEBUG_REQ(D_ERROR, req, "network error on bulk %s",
3092 /* XXX should this be a different errno? */
3095 if (req->rq_bulk_write)
3096 rc = sptlrpc_svc_unwrap_bulk(req, desc);
3097 if (rc == 0 && desc->bd_nob_transferred != desc->bd_nob) {
3098 DEBUG_REQ(D_ERROR, req, "truncated bulk %s %d(%d)",
3099 bulk2type(req), desc->bd_nob_transferred,
3101 /* XXX should this be a different errno? */
3108 EXPORT_SYMBOL(target_bulk_io);
3110 #endif /* HAVE_SERVER_SUPPORT */