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.gnu.org/licenses/gpl-2.0.html
23 * Copyright (c) 2003, 2010, Oracle and/or its affiliates. All rights reserved.
24 * Use is subject to license terms.
26 * Copyright (c) 2010, 2017, Intel Corporation.
29 * This file is part of Lustre, http://www.lustre.org/
30 * Lustre is a trademark of Sun Microsystems, Inc.
34 * This file deals with various client/target related logic including recovery.
36 * TODO: This code more logically belongs in the ptlrpc module than in ldlm and
40 #define DEBUG_SUBSYSTEM S_LDLM
42 #include <cl_object.h>
43 #include <linux/jiffies.h>
44 #include <linux/kernel.h>
45 #include <linux/kthread.h>
46 #include <libcfs/libcfs.h>
48 #include <obd_class.h>
49 #include <lustre_dlm.h>
50 #include <lustre_net.h>
51 #include <lustre_sec.h>
52 #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;
63 lnet_nid_t nid4refnet = LNET_NID_ANY;
68 if (!create && !priority) {
69 CDEBUG(D_HA, "Nothing to do\n");
73 if (imp->imp_connection &&
74 imp->imp_connection->c_remote_uuid.uuid[0] == 0)
75 /* nid4refnet is used to restrict network connections */
76 nid4refnet = imp->imp_connection->c_self;
77 ptlrpc_conn = ptlrpc_uuid_to_connection(uuid, nid4refnet);
79 CDEBUG(D_HA, "can't find connection %s\n", uuid->uuid);
84 OBD_ALLOC(imp_conn, sizeof(*imp_conn));
86 GOTO(out_put, rc = -ENOMEM);
89 spin_lock(&imp->imp_lock);
90 list_for_each_entry(item, &imp->imp_conn_list, oic_item) {
91 if (obd_uuid_equals(uuid, &item->oic_uuid)) {
93 list_move(&item->oic_item,
95 item->oic_last_attempt = 0;
97 CDEBUG(D_HA, "imp %p@%s: found existing conn %s%s\n",
98 imp, imp->imp_obd->obd_name, uuid->uuid,
99 (priority ? ", moved to head" : ""));
100 spin_unlock(&imp->imp_lock);
101 GOTO(out_free, rc = 0);
104 /* No existing import connection found for \a uuid. */
106 imp_conn->oic_conn = ptlrpc_conn;
107 imp_conn->oic_uuid = *uuid;
108 imp_conn->oic_last_attempt = 0;
110 list_add(&imp_conn->oic_item, &imp->imp_conn_list);
112 list_add_tail(&imp_conn->oic_item,
113 &imp->imp_conn_list);
114 CDEBUG(D_HA, "imp %p@%s: add connection %s at %s\n",
115 imp, imp->imp_obd->obd_name, uuid->uuid,
116 (priority ? "head" : "tail"));
118 spin_unlock(&imp->imp_lock);
119 GOTO(out_free, rc = -ENOENT);
122 spin_unlock(&imp->imp_lock);
126 OBD_FREE(imp_conn, sizeof(*imp_conn));
128 ptlrpc_connection_put(ptlrpc_conn);
132 int import_set_conn_priority(struct obd_import *imp, struct obd_uuid *uuid)
134 return import_set_conn(imp, uuid, 1, 0);
137 int client_import_add_conn(struct obd_import *imp, struct obd_uuid *uuid,
140 return import_set_conn(imp, uuid, priority, 1);
142 EXPORT_SYMBOL(client_import_add_conn);
144 int client_import_dyn_add_conn(struct obd_import *imp, struct obd_uuid *uuid,
145 lnet_nid_t prim_nid, int priority)
147 struct ptlrpc_connection *ptlrpc_conn;
150 ptlrpc_conn = ptlrpc_uuid_to_connection(uuid, prim_nid);
152 const char *str_uuid = obd_uuid2str(uuid);
154 rc = class_add_uuid(str_uuid, prim_nid);
156 CERROR("%s: failed to add UUID '%s': rc = %d\n",
157 imp->imp_obd->obd_name, str_uuid, rc);
161 return import_set_conn(imp, uuid, priority, 1);
163 EXPORT_SYMBOL(client_import_dyn_add_conn);
165 int client_import_add_nids_to_conn(struct obd_import *imp, lnet_nid_t *nids,
166 int nid_count, struct obd_uuid *uuid)
168 struct obd_import_conn *conn;
172 if (nid_count <= 0 || !nids)
175 spin_lock(&imp->imp_lock);
176 list_for_each_entry(conn, &imp->imp_conn_list, oic_item) {
177 if (class_check_uuid(&conn->oic_uuid, nids[0])) {
178 *uuid = conn->oic_uuid;
179 rc = class_add_nids_to_uuid(&conn->oic_uuid, nids,
184 spin_unlock(&imp->imp_lock);
187 EXPORT_SYMBOL(client_import_add_nids_to_conn);
189 int client_import_del_conn(struct obd_import *imp, struct obd_uuid *uuid)
191 struct obd_import_conn *imp_conn;
192 struct obd_export *dlmexp;
197 spin_lock(&imp->imp_lock);
198 if (list_empty(&imp->imp_conn_list)) {
199 LASSERT(!imp->imp_connection);
203 list_for_each_entry(imp_conn, &imp->imp_conn_list, oic_item) {
204 if (!obd_uuid_equals(uuid, &imp_conn->oic_uuid))
206 LASSERT(imp_conn->oic_conn);
208 if (imp_conn == imp->imp_conn_current) {
209 LASSERT(imp_conn->oic_conn == imp->imp_connection);
211 if (imp->imp_state != LUSTRE_IMP_CLOSED &&
212 imp->imp_state != LUSTRE_IMP_DISCON) {
213 CERROR("can't remove current connection\n");
214 GOTO(out, rc = -EBUSY);
217 ptlrpc_connection_put(imp->imp_connection);
218 imp->imp_connection = NULL;
220 dlmexp = class_conn2export(&imp->imp_dlm_handle);
221 if (dlmexp && dlmexp->exp_connection) {
222 LASSERT(dlmexp->exp_connection ==
224 ptlrpc_connection_put(dlmexp->exp_connection);
225 dlmexp->exp_connection = NULL;
229 class_export_put(dlmexp);
232 list_del(&imp_conn->oic_item);
233 ptlrpc_connection_put(imp_conn->oic_conn);
234 OBD_FREE(imp_conn, sizeof(*imp_conn));
235 CDEBUG(D_HA, "imp %p@%s: remove connection %s\n",
236 imp, imp->imp_obd->obd_name, uuid->uuid);
241 spin_unlock(&imp->imp_lock);
243 CERROR("connection %s not found\n", uuid->uuid);
246 EXPORT_SYMBOL(client_import_del_conn);
249 * Find conn UUID by peer NID. \a peer is a server NID. This function is used
250 * to find a conn uuid of \a imp which can reach \a peer.
252 int client_import_find_conn(struct obd_import *imp, lnet_nid_t peer,
253 struct obd_uuid *uuid)
255 struct obd_import_conn *conn;
260 spin_lock(&imp->imp_lock);
261 list_for_each_entry(conn, &imp->imp_conn_list, oic_item) {
262 /* Check if conn UUID does have this peer NID. */
263 if (class_check_uuid(&conn->oic_uuid, peer)) {
264 *uuid = conn->oic_uuid;
269 spin_unlock(&imp->imp_lock);
272 EXPORT_SYMBOL(client_import_find_conn);
274 void client_destroy_import(struct obd_import *imp)
277 * Drop security policy instance after all RPCs have finished/aborted
278 * to let all busy contexts be released.
280 class_import_get(imp);
281 class_destroy_import(imp);
282 sptlrpc_import_sec_put(imp);
283 class_import_put(imp);
285 EXPORT_SYMBOL(client_destroy_import);
288 * Check whether or not the OSC is on MDT.
291 * setup 0:{fsname}-OSTxxxx-osc[-MDTxxxx] 1:lustre-OST0000_UUID 2:NID
293 * setup 0:{fsname}-OSTxxxx-osc 1:lustre-OST0000_UUID 2:NID
296 static int osc_on_mdt(char *obdname)
300 ptr = strrchr(obdname, '-');
304 if (strncmp(ptr + 1, "MDT", 3) == 0)
311 * Configure an RPC client OBD device.
316 * 3 - inactive-on-startup
317 * 4 - restrictive net
319 int client_obd_setup(struct obd_device *obd, struct lustre_cfg *lcfg)
321 struct client_obd *cli = &obd->u.cli;
322 struct obd_import *imp;
323 struct obd_uuid server_uuid;
324 int rq_portal, rp_portal, connect_op;
325 const char *name = obd->obd_type->typ_name;
326 enum ldlm_ns_type ns_type = LDLM_NS_TYPE_UNKNOWN;
327 char *cli_name = lustre_cfg_buf(lcfg, 0);
328 struct ptlrpc_connection fake_conn = { .c_self = 0,
329 .c_remote_uuid.uuid[0] = 0 };
335 * In a more perfect world, we would hang a ptlrpc_client off of
336 * obd_type and just use the values from there.
338 if (!strcmp(name, LUSTRE_OSC_NAME)) {
339 rq_portal = OST_REQUEST_PORTAL;
340 rp_portal = OSC_REPLY_PORTAL;
341 connect_op = OST_CONNECT;
342 cli->cl_sp_me = LUSTRE_SP_CLI;
343 cli->cl_sp_to = LUSTRE_SP_OST;
344 ns_type = LDLM_NS_TYPE_OSC;
345 } else if (!strcmp(name, LUSTRE_MDC_NAME) ||
346 !strcmp(name, LUSTRE_LWP_NAME)) {
347 rq_portal = MDS_REQUEST_PORTAL;
348 rp_portal = MDC_REPLY_PORTAL;
349 connect_op = MDS_CONNECT;
350 if (is_lwp_on_ost(cli_name))
351 cli->cl_sp_me = LUSTRE_SP_OST;
352 else if (is_lwp_on_mdt(cli_name))
353 cli->cl_sp_me = LUSTRE_SP_MDT;
355 cli->cl_sp_me = LUSTRE_SP_CLI;
356 cli->cl_sp_to = LUSTRE_SP_MDT;
357 ns_type = LDLM_NS_TYPE_MDC;
358 } else if (!strcmp(name, LUSTRE_OSP_NAME)) {
359 if (strstr(lustre_cfg_buf(lcfg, 1), "OST") == NULL) {
360 /* OSP_on_MDT for other MDTs */
361 connect_op = MDS_CONNECT;
362 cli->cl_sp_to = LUSTRE_SP_MDT;
363 ns_type = LDLM_NS_TYPE_MDC;
364 rq_portal = OUT_PORTAL;
366 /* OSP on MDT for OST */
367 connect_op = OST_CONNECT;
368 cli->cl_sp_to = LUSTRE_SP_OST;
369 ns_type = LDLM_NS_TYPE_OSC;
370 rq_portal = OST_REQUEST_PORTAL;
372 rp_portal = OSC_REPLY_PORTAL;
373 cli->cl_sp_me = LUSTRE_SP_MDT;
374 } else if (!strcmp(name, LUSTRE_MGC_NAME)) {
375 rq_portal = MGS_REQUEST_PORTAL;
376 rp_portal = MGC_REPLY_PORTAL;
377 connect_op = MGS_CONNECT;
378 cli->cl_sp_me = LUSTRE_SP_MGC;
379 cli->cl_sp_to = LUSTRE_SP_MGS;
380 cli->cl_flvr_mgc.sf_rpc = SPTLRPC_FLVR_INVALID;
381 ns_type = LDLM_NS_TYPE_MGC;
383 CERROR("unknown client OBD type \"%s\", can't setup\n",
388 if (LUSTRE_CFG_BUFLEN(lcfg, 1) < 1) {
389 CERROR("requires a TARGET UUID\n");
393 if (LUSTRE_CFG_BUFLEN(lcfg, 1) > 37) {
394 CERROR("client UUID must be less than 38 characters\n");
398 if (LUSTRE_CFG_BUFLEN(lcfg, 2) < 1) {
399 CERROR("setup requires a SERVER UUID\n");
403 if (LUSTRE_CFG_BUFLEN(lcfg, 2) > 37) {
404 CERROR("target UUID must be less than 38 characters\n");
408 init_rwsem(&cli->cl_sem);
409 mutex_init(&cli->cl_mgc_mutex);
411 init_rwsem(&cli->cl_seq_rwsem);
412 cli->cl_conn_count = 0;
413 memcpy(server_uuid.uuid, lustre_cfg_buf(lcfg, 2),
414 min_t(unsigned int, LUSTRE_CFG_BUFLEN(lcfg, 2),
415 sizeof(server_uuid)));
417 cli->cl_dirty_pages = 0;
418 cli->cl_dirty_max_pages = 0;
419 cli->cl_avail_grant = 0;
420 /* FIXME: Should limit this for the sum of all cl_dirty_max_pages. */
422 * cl_dirty_max_pages may be changed at connect time in
423 * ptlrpc_connect_interpret().
425 client_adjust_max_dirty(cli);
426 init_waitqueue_head(&cli->cl_cache_waiters);
427 INIT_LIST_HEAD(&cli->cl_loi_ready_list);
428 INIT_LIST_HEAD(&cli->cl_loi_hp_ready_list);
429 INIT_LIST_HEAD(&cli->cl_loi_write_list);
430 INIT_LIST_HEAD(&cli->cl_loi_read_list);
431 spin_lock_init(&cli->cl_loi_list_lock);
432 atomic_set(&cli->cl_pending_w_pages, 0);
433 atomic_set(&cli->cl_pending_r_pages, 0);
434 cli->cl_r_in_flight = 0;
435 cli->cl_w_in_flight = 0;
437 spin_lock_init(&cli->cl_read_rpc_hist.oh_lock);
438 spin_lock_init(&cli->cl_write_rpc_hist.oh_lock);
439 spin_lock_init(&cli->cl_read_page_hist.oh_lock);
440 spin_lock_init(&cli->cl_write_page_hist.oh_lock);
441 spin_lock_init(&cli->cl_read_offset_hist.oh_lock);
442 spin_lock_init(&cli->cl_write_offset_hist.oh_lock);
445 INIT_LIST_HEAD(&cli->cl_lru_osc);
446 atomic_set(&cli->cl_lru_shrinkers, 0);
447 atomic_long_set(&cli->cl_lru_busy, 0);
448 atomic_long_set(&cli->cl_lru_in_list, 0);
449 INIT_LIST_HEAD(&cli->cl_lru_list);
450 spin_lock_init(&cli->cl_lru_list_lock);
451 atomic_long_set(&cli->cl_unstable_count, 0);
452 INIT_LIST_HEAD(&cli->cl_shrink_list);
453 INIT_LIST_HEAD(&cli->cl_grant_chain);
455 INIT_LIST_HEAD(&cli->cl_flight_waiters);
456 cli->cl_rpcs_in_flight = 0;
458 init_waitqueue_head(&cli->cl_destroy_waitq);
459 atomic_set(&cli->cl_destroy_in_flight, 0);
462 cli->cl_supp_cksum_types = OBD_CKSUM_CRC32;
463 cli->cl_preferred_cksum_type = 0;
464 #ifdef ENABLE_CHECKSUM
465 /* Turn on checksumming by default. */
466 cli->cl_checksum = 1;
468 * The supported checksum types will be worked out at connect time
469 * Set cl_chksum* to CRC32 for now to avoid returning screwed info
472 cli->cl_cksum_type = cli->cl_supp_cksum_types;
474 atomic_set(&cli->cl_resends, OSC_DEFAULT_RESENDS);
477 * Set it to possible maximum size. It may be reduced by ocd_brw_size
478 * from OFD after connecting.
480 cli->cl_max_pages_per_rpc = PTLRPC_MAX_BRW_PAGES;
482 cli->cl_max_short_io_bytes = OBD_DEF_SHORT_IO_BYTES;
485 * set cl_chunkbits default value to PAGE_SHIFT,
486 * it will be updated at OSC connection time.
488 cli->cl_chunkbits = PAGE_SHIFT;
490 if (!strcmp(name, LUSTRE_MDC_NAME)) {
491 cli->cl_max_rpcs_in_flight = OBD_MAX_RIF_DEFAULT;
492 } else if (cfs_totalram_pages() >> (20 - PAGE_SHIFT) <= 128 /* MB */) {
493 cli->cl_max_rpcs_in_flight = 2;
494 } else if (cfs_totalram_pages() >> (20 - PAGE_SHIFT) <= 256 /* MB */) {
495 cli->cl_max_rpcs_in_flight = 3;
496 } else if (cfs_totalram_pages() >> (20 - PAGE_SHIFT) <= 512 /* MB */) {
497 cli->cl_max_rpcs_in_flight = 4;
499 if (osc_on_mdt(obd->obd_name))
500 cli->cl_max_rpcs_in_flight = OBD_MAX_RIF_MAX;
502 cli->cl_max_rpcs_in_flight = OBD_MAX_RIF_DEFAULT;
505 spin_lock_init(&cli->cl_mod_rpcs_lock);
506 spin_lock_init(&cli->cl_mod_rpcs_hist.oh_lock);
507 cli->cl_max_mod_rpcs_in_flight = 0;
508 cli->cl_mod_rpcs_in_flight = 0;
509 cli->cl_close_rpcs_in_flight = 0;
510 init_waitqueue_head(&cli->cl_mod_rpcs_waitq);
511 cli->cl_mod_tag_bitmap = NULL;
513 INIT_LIST_HEAD(&cli->cl_chg_dev_linkage);
515 if (connect_op == MDS_CONNECT) {
516 cli->cl_max_mod_rpcs_in_flight = cli->cl_max_rpcs_in_flight - 1;
517 OBD_ALLOC(cli->cl_mod_tag_bitmap,
518 BITS_TO_LONGS(OBD_MAX_RIF_MAX) * sizeof(long));
519 if (cli->cl_mod_tag_bitmap == NULL)
520 GOTO(err, rc = -ENOMEM);
525 CERROR("ldlm_get_ref failed: %d\n", rc);
529 ptlrpc_init_client(rq_portal, rp_portal, name,
530 &obd->obd_ldlm_client);
532 imp = class_new_import(obd);
534 GOTO(err_ldlm, rc = -ENOENT);
535 imp->imp_client = &obd->obd_ldlm_client;
536 imp->imp_connect_op = connect_op;
537 memcpy(cli->cl_target_uuid.uuid, lustre_cfg_buf(lcfg, 1),
538 LUSTRE_CFG_BUFLEN(lcfg, 1));
539 class_import_put(imp);
541 if (lustre_cfg_buf(lcfg, 4)) {
542 __u32 refnet = libcfs_str2net(lustre_cfg_string(lcfg, 4));
544 if (refnet == LNET_NET_ANY) {
546 CERROR("%s: bad mount option 'network=%s': rc = %d\n",
547 obd->obd_name, lustre_cfg_string(lcfg, 4),
549 GOTO(err_import, rc);
551 fake_conn.c_self = LNET_MKNID(refnet, 0);
552 imp->imp_connection = &fake_conn;
555 rc = client_import_add_conn(imp, &server_uuid, 1);
557 CERROR("can't add initial connection\n");
558 GOTO(err_import, rc);
560 imp->imp_connection = NULL;
562 cli->cl_import = imp;
563 /* cli->cl_max_mds_easize updated by mdc_init_ea_size() */
564 cli->cl_max_mds_easize = sizeof(struct lov_mds_md_v3);
566 if (LUSTRE_CFG_BUFLEN(lcfg, 3) > 0) {
567 if (!strcmp(lustre_cfg_string(lcfg, 3), "inactive")) {
568 CDEBUG(D_HA, "marking %s %s->%s as inactive\n",
570 cli->cl_target_uuid.uuid);
571 spin_lock(&imp->imp_lock);
572 imp->imp_deactive = 1;
573 spin_unlock(&imp->imp_lock);
577 obd->obd_namespace = ldlm_namespace_new(obd, obd->obd_name,
578 LDLM_NAMESPACE_CLIENT,
579 LDLM_NAMESPACE_GREEDY,
581 if (IS_ERR(obd->obd_namespace)) {
582 rc = PTR_ERR(obd->obd_namespace);
583 CERROR("%s: unable to create client namespace: rc = %d\n",
585 obd->obd_namespace = NULL;
586 GOTO(err_import, rc);
592 class_destroy_import(imp);
596 if (cli->cl_mod_tag_bitmap != NULL)
597 OBD_FREE(cli->cl_mod_tag_bitmap,
598 BITS_TO_LONGS(OBD_MAX_RIF_MAX) * sizeof(long));
599 cli->cl_mod_tag_bitmap = NULL;
603 EXPORT_SYMBOL(client_obd_setup);
605 int client_obd_cleanup(struct obd_device *obd)
607 struct client_obd *cli = &obd->u.cli;
611 ldlm_namespace_free_post(obd->obd_namespace);
612 obd->obd_namespace = NULL;
614 obd_cleanup_client_import(obd);
615 LASSERT(obd->u.cli.cl_import == NULL);
619 if (cli->cl_mod_tag_bitmap != NULL)
620 OBD_FREE(cli->cl_mod_tag_bitmap,
621 BITS_TO_LONGS(OBD_MAX_RIF_MAX) * sizeof(long));
622 cli->cl_mod_tag_bitmap = NULL;
626 EXPORT_SYMBOL(client_obd_cleanup);
628 /* ->o_connect() method for client side (OSC and MDC and MGC) */
629 int client_connect_import(const struct lu_env *env,
630 struct obd_export **exp,
631 struct obd_device *obd, struct obd_uuid *cluuid,
632 struct obd_connect_data *data, void *localdata)
634 struct client_obd *cli = &obd->u.cli;
635 struct obd_import *imp = cli->cl_import;
636 struct obd_connect_data *ocd;
637 struct lustre_handle conn = { 0 };
643 down_write(&cli->cl_sem);
644 if (cli->cl_conn_count > 0)
645 GOTO(out_sem, rc = -EALREADY);
647 rc = class_connect(&conn, obd, cluuid);
651 cli->cl_conn_count++;
652 *exp = class_conn2export(&conn);
654 LASSERT(obd->obd_namespace);
656 imp->imp_dlm_handle = conn;
657 rc = ptlrpc_init_import(imp);
661 ocd = &imp->imp_connect_data;
664 imp->imp_connect_flags_orig = data->ocd_connect_flags;
665 imp->imp_connect_flags2_orig = data->ocd_connect_flags2;
668 rc = ptlrpc_connect_import(imp);
670 LASSERT(imp->imp_state == LUSTRE_IMP_DISCON);
673 LASSERT(*exp != NULL && (*exp)->exp_connection);
676 LASSERTF((ocd->ocd_connect_flags & data->ocd_connect_flags) ==
677 ocd->ocd_connect_flags, "old %#llx, new %#llx\n",
678 data->ocd_connect_flags, ocd->ocd_connect_flags);
679 data->ocd_connect_flags = ocd->ocd_connect_flags;
680 data->ocd_connect_flags2 = ocd->ocd_connect_flags2;
683 ptlrpc_pinger_add_import(imp);
689 cli->cl_conn_count--;
690 class_disconnect(*exp);
694 up_write(&cli->cl_sem);
696 if (!rc && localdata) {
697 LASSERT(cli->cl_cache == NULL); /* only once */
698 cli->cl_cache = (struct cl_client_cache *)localdata;
699 cl_cache_incref(cli->cl_cache);
700 cli->cl_lru_left = &cli->cl_cache->ccc_lru_left;
702 /* add this osc into entity list */
703 LASSERT(list_empty(&cli->cl_lru_osc));
704 spin_lock(&cli->cl_cache->ccc_lru_lock);
705 list_add(&cli->cl_lru_osc, &cli->cl_cache->ccc_lru);
706 spin_unlock(&cli->cl_cache->ccc_lru_lock);
711 EXPORT_SYMBOL(client_connect_import);
713 int client_disconnect_export(struct obd_export *exp)
715 struct obd_device *obd = class_exp2obd(exp);
716 struct client_obd *cli;
717 struct obd_import *imp;
723 CERROR("invalid export for disconnect: exp %p cookie %#llx\n",
724 exp, exp ? exp->exp_handle.h_cookie : -1);
729 imp = cli->cl_import;
731 down_write(&cli->cl_sem);
732 CDEBUG(D_INFO, "disconnect %s - %zu\n", obd->obd_name,
735 if (cli->cl_conn_count == 0) {
736 CERROR("disconnecting disconnected device (%s)\n",
738 GOTO(out_disconnect, rc = -EINVAL);
741 cli->cl_conn_count--;
742 if (cli->cl_conn_count != 0)
743 GOTO(out_disconnect, rc = 0);
746 * Mark import deactivated now, so we don't try to reconnect if any
747 * of the cleanup RPCs fails (e.g. LDLM cancel, etc). We don't
748 * fully deactivate the import, or that would drop all requests.
750 spin_lock(&imp->imp_lock);
751 imp->imp_deactive = 1;
752 spin_unlock(&imp->imp_lock);
755 * Some non-replayable imports (MDS's OSCs) are pinged, so just
756 * delete it regardless. (It's safe to delete an import that was
759 (void)ptlrpc_pinger_del_import(imp);
761 if (obd->obd_namespace != NULL) {
762 /* obd_force == local only */
763 ldlm_cli_cancel_unused(obd->obd_namespace, NULL,
764 obd->obd_force ? LCF_LOCAL : 0, NULL);
765 ldlm_namespace_free_prior(obd->obd_namespace, imp,
770 * There's no need to hold sem while disconnecting an import,
771 * and it may actually cause deadlock in GSS.
773 up_write(&cli->cl_sem);
774 rc = ptlrpc_disconnect_import(imp, 0);
775 down_write(&cli->cl_sem);
777 ptlrpc_invalidate_import(imp);
783 * Use server style - class_disconnect should be always called for
786 err = class_disconnect(exp);
790 up_write(&cli->cl_sem);
794 EXPORT_SYMBOL(client_disconnect_export);
796 #ifdef HAVE_SERVER_SUPPORT
797 int server_disconnect_export(struct obd_export *exp)
803 /* Disconnect early so that clients can't keep using export. */
804 rc = class_disconnect(exp);
805 /* Close import to avoid sending any requests. */
806 if (exp->exp_imp_reverse)
807 ptlrpc_cleanup_imp(exp->exp_imp_reverse);
809 ldlm_bl_thread_wakeup();
811 /* complete all outstanding replies */
812 spin_lock(&exp->exp_lock);
813 while (!list_empty(&exp->exp_outstanding_replies)) {
814 struct ptlrpc_reply_state *rs =
815 list_entry(exp->exp_outstanding_replies.next,
816 struct ptlrpc_reply_state, rs_exp_list);
817 struct ptlrpc_service_part *svcpt = rs->rs_svcpt;
819 spin_lock(&svcpt->scp_rep_lock);
821 list_del_init(&rs->rs_exp_list);
823 spin_lock(&rs->rs_lock);
824 /* clear rs_convert_lock to make sure rs is handled and put */
825 rs->rs_convert_lock = 0;
826 ptlrpc_schedule_difficult_reply(rs);
827 spin_unlock(&rs->rs_lock);
829 spin_unlock(&svcpt->scp_rep_lock);
831 spin_unlock(&exp->exp_lock);
835 EXPORT_SYMBOL(server_disconnect_export);
837 static inline int target_check_recovery_timer(struct obd_device *target)
842 if (!target->obd_recovering || target->obd_recovery_start == 0)
845 remaining = hrtimer_get_remaining(&target->obd_recovery_timer);
846 timeout = ktime_divns(remaining, NSEC_PER_SEC);
850 /* the recovery timer should expire, but it isn't triggered,
851 * it's better to abort the recovery of this target to speed up
852 * the recovery of the whole cluster. */
853 spin_lock(&target->obd_dev_lock);
854 if (target->obd_recovering) {
855 CERROR("%s: Aborting recovery\n", target->obd_name);
856 target->obd_abort_recovery = 1;
857 wake_up(&target->obd_next_transno_waitq);
859 spin_unlock(&target->obd_dev_lock);
864 * --------------------------------------------------------------------------
865 * from old lib/target.c
866 * --------------------------------------------------------------------------
868 static int target_handle_reconnect(struct lustre_handle *conn,
869 struct obd_export *exp,
870 struct obd_uuid *cluuid)
872 struct obd_device *target;
873 struct lustre_handle *hdl;
879 hdl = &exp->exp_imp_reverse->imp_remote_handle;
880 if (!exp->exp_connection || !lustre_handle_is_used(hdl)) {
881 conn->cookie = exp->exp_handle.h_cookie;
883 "connect export for UUID '%s' at %p, cookie %#llx\n",
884 cluuid->uuid, exp, conn->cookie);
888 target = exp->exp_obd;
890 /* Might be a re-connect after a partition. */
891 if (memcmp(&conn->cookie, &hdl->cookie, sizeof(conn->cookie))) {
892 LCONSOLE_WARN("%s: already connected client %s (at %s) with handle %#llx. Rejecting client with the same UUID trying to reconnect with handle %#llx\n",
894 obd_uuid2str(&exp->exp_client_uuid),
895 obd_export_nid2str(exp),
896 hdl->cookie, conn->cookie);
897 memset(conn, 0, sizeof(*conn));
899 * target_handle_connect() treats EALREADY and
900 * -EALREADY differently. -EALREADY is an error
901 * (same UUID, different handle).
906 if (!target->obd_recovering) {
907 LCONSOLE_WARN("%s: Client %s (at %s) reconnecting\n",
908 target->obd_name, obd_uuid2str(&exp->exp_client_uuid),
909 obd_export_nid2str(exp));
910 GOTO(out_already, rc);
913 remaining = hrtimer_get_remaining(&target->obd_recovery_timer);
914 timeout = ktime_divns(remaining, NSEC_PER_SEC);
916 LCONSOLE_WARN("%s: Client %s (at %s) reconnected, waiting for %d clients in recovery for %lld:%.02lld\n",
918 obd_uuid2str(&exp->exp_client_uuid),
919 obd_export_nid2str(exp),
920 atomic_read(&target->obd_max_recoverable_clients),
921 timeout / 60, timeout % 60);
923 struct target_distribute_txn_data *tdtd;
928 target_check_recovery_timer(target);
930 tdtd = class_exp2tgt(exp)->lut_tdtd;
931 if (tdtd && tdtd->tdtd_show_update_logs_retrievers)
932 buf = tdtd->tdtd_show_update_logs_retrievers(
933 tdtd->tdtd_show_retrievers_cbdata,
937 LCONSOLE_WARN("%s: Client %s (at %s) reconnecting, waiting for %d MDTs (%s) in recovery for %lld:%.02lld. Please wait until all MDTs recovered or you may force MDT evicition via 'lctl --device %s abort_recovery.\n",
939 obd_uuid2str(&exp->exp_client_uuid),
940 obd_export_nid2str(exp), count,
941 buf ? buf : "unknown (not enough RAM)",
942 (abs(timeout) + target->obd_recovery_timeout) / 60,
943 (abs(timeout) + target->obd_recovery_timeout) % 60,
946 LCONSOLE_WARN("%s: Recovery already passed deadline %lld:%.02lld. If you do not want to wait more, you may force taget eviction via 'lctl --device %s abort_recovery.\n",
947 target->obd_name, abs(timeout) / 60,
948 abs(timeout) % 60, target->obd_name);
955 conn->cookie = exp->exp_handle.h_cookie;
957 * target_handle_connect() treats EALREADY and
958 * -EALREADY differently. EALREADY means we are
959 * doing a valid reconnect from the same client.
965 check_and_start_recovery_timer(struct obd_device *obd,
966 struct ptlrpc_request *req, int new_client);
969 * update flags for import during reconnect process
971 static int rev_import_flags_update(struct obd_import *revimp,
972 struct ptlrpc_request *req)
975 struct obd_connect_data *data;
977 data = req_capsule_client_get(&req->rq_pill, &RMF_CONNECT_DATA);
979 if (data->ocd_connect_flags & OBD_CONNECT_AT)
980 revimp->imp_msghdr_flags |= MSGHDR_AT_SUPPORT;
982 revimp->imp_msghdr_flags &= ~MSGHDR_AT_SUPPORT;
984 revimp->imp_msghdr_flags |= MSGHDR_CKSUM_INCOMPAT18;
986 rc = sptlrpc_import_sec_adapt(revimp, req->rq_svc_ctx, &req->rq_flvr);
988 CERROR("%s: cannot get reverse import %s security: rc = %d\n",
989 revimp->imp_client->cli_name,
990 libcfs_id2str(req->rq_peer), rc);
998 * Allocate a new reverse import for an export.
1000 * \retval -errno in case error hit
1001 * \retval 0 if reverse import correctly init
1003 int rev_import_init(struct obd_export *export)
1005 struct obd_device *obd = export->exp_obd;
1006 struct obd_import *revimp;
1008 LASSERT(export->exp_imp_reverse == NULL);
1010 revimp = class_new_import(obd);
1014 revimp->imp_remote_handle.cookie = 0ULL;
1015 revimp->imp_client = &obd->obd_ldlm_client;
1016 revimp->imp_dlm_fake = 1;
1018 /* it is safe to connect import in new state as no sends possible */
1019 spin_lock(&export->exp_lock);
1020 export->exp_imp_reverse = revimp;
1021 spin_unlock(&export->exp_lock);
1022 class_import_put(revimp);
1026 EXPORT_SYMBOL(rev_import_init);
1029 * Handle reconnect for an export.
1031 * \param exp export to handle reconnect process
1032 * \param req client reconnect request
1034 * \retval -rc in case securitfy flavor can't be changed
1035 * \retval 0 in case none problems
1037 static int rev_import_reconnect(struct obd_export *exp,
1038 struct ptlrpc_request *req)
1040 struct obd_import *revimp = exp->exp_imp_reverse;
1041 struct lustre_handle *lh;
1044 /* avoid sending a request until import flags are changed */
1045 ptlrpc_import_enter_resend(revimp);
1047 ptlrpc_connection_put(revimp->imp_connection);
1050 * client from recovery don't have a handle so we need to take from
1051 * request. it may produce situation when wrong client connected
1052 * to recovery as we trust a client uuid
1054 lh = req_capsule_client_get(&req->rq_pill, &RMF_CONN);
1055 revimp->imp_remote_handle = *lh;
1058 * unknown versions will be caught in
1059 * ptlrpc_handle_server_req_in->lustre_unpack_msg()
1061 revimp->imp_msg_magic = req->rq_reqmsg->lm_magic;
1063 revimp->imp_connection = ptlrpc_connection_addref(exp->exp_connection);
1065 rc = rev_import_flags_update(revimp, req);
1068 * it is safe to still be in RECOVERY phase as we are not able
1069 * to setup correct security flavor so requests are not able to
1070 * be delivered correctly
1075 /* resend all rpc's via new connection */
1076 return ptlrpc_import_recovery_state_machine(revimp);
1079 int target_handle_connect(struct ptlrpc_request *req)
1081 struct obd_device *target = NULL;
1082 struct obd_export *export = NULL;
1084 * connect handle - filled from target_handle_reconnect in
1087 struct lustre_handle conn;
1088 struct lustre_handle *tmp;
1089 struct obd_uuid cluuid;
1094 bool mds_conn = false, lw_client = false, initial_conn = false;
1095 bool mds_mds_conn = false;
1096 bool new_mds_mds_conn = false;
1097 struct obd_connect_data *data, *tmpdata;
1099 lnet_nid_t *client_nid = NULL;
1103 OBD_RACE(OBD_FAIL_TGT_CONN_RACE);
1105 str = req_capsule_client_get(&req->rq_pill, &RMF_TGTUUID);
1107 DEBUG_REQ(D_ERROR, req, "bad target UUID for connect");
1108 GOTO(out, rc = -EINVAL);
1111 target = class_dev_by_str(str);
1113 deuuidify(str, NULL, &target_start, &target_len);
1114 LCONSOLE_ERROR_MSG(0x137,
1115 "%s: not available for connect from %s (no target). If you are running an HA pair check that the target is mounted on the other server.\n",
1116 str, libcfs_nid2str(req->rq_peer.nid));
1117 GOTO(out, rc = -ENODEV);
1120 spin_lock(&target->obd_dev_lock);
1122 target->obd_conn_inprogress++;
1124 if (target->obd_stopping || !target->obd_set_up) {
1125 spin_unlock(&target->obd_dev_lock);
1127 deuuidify(str, NULL, &target_start, &target_len);
1128 LCONSOLE_INFO("%.*s: Not available for connect from %s (%s)\n",
1129 target_len, target_start,
1130 libcfs_nid2str(req->rq_peer.nid),
1131 (target->obd_stopping ?
1132 "stopping" : "not set up"));
1133 GOTO(out, rc = -ENODEV);
1136 if (target->obd_no_conn) {
1137 spin_unlock(&target->obd_dev_lock);
1140 "%s: Temporarily refusing client connection from %s\n",
1141 target->obd_name, libcfs_nid2str(req->rq_peer.nid));
1142 GOTO(out, rc = -EAGAIN);
1145 spin_unlock(&target->obd_dev_lock);
1147 str = req_capsule_client_get(&req->rq_pill, &RMF_CLUUID);
1149 DEBUG_REQ(D_ERROR, req, "bad client UUID for connect");
1150 GOTO(out, rc = -EINVAL);
1153 obd_str2uuid(&cluuid, str);
1155 tmp = req_capsule_client_get(&req->rq_pill, &RMF_CONN);
1157 GOTO(out, rc = -EPROTO);
1161 size = req_capsule_get_size(&req->rq_pill, &RMF_CONNECT_DATA,
1163 if (size < 0 || size > 8 * sizeof(struct obd_connect_data))
1164 GOTO(out, rc = -EPROTO);
1165 data = req_capsule_client_get(&req->rq_pill, &RMF_CONNECT_DATA);
1167 GOTO(out, rc = -EPROTO);
1169 rc = req_capsule_server_pack(&req->rq_pill);
1173 #if LUSTRE_VERSION_CODE < OBD_OCD_VERSION(3, 0, 53, 0)
1175 * Don't allow clients to connect that are using old 1.8 format
1176 * protocol conventions (LUSTRE_MSG_MAGIC_v1, !MSGHDR_CKSUM_INCOMPAT18,
1177 * ldlm_flock_policy_wire format, MDT_ATTR_xTIME_SET, etc). The
1178 * FULL20 flag should be set on all connections since 2.0, but no
1179 * longer affects behaviour.
1181 * Later this check will be disabled and the flag can be retired
1182 * completely once interop with 3.0 is no longer needed.
1184 if (!(data->ocd_connect_flags & OBD_CONNECT_FULL20))
1185 GOTO(out, rc = -EPROTO);
1188 * Don't allow liblustre clients to connect.
1189 * - testing was disabled in v2_2_50_0-61-g6a75d65
1190 * - building was disabled in v2_5_58_0-28-g7277179
1191 * - client code was deleted in v2_6_50_0-101-gcdfbc72,
1192 * - clients were refused connect for version difference > 0.0.1.32
1194 if (lustre_msg_get_op_flags(req->rq_reqmsg) & MSG_CONNECT_LIBCLIENT) {
1195 DEBUG_REQ(D_WARNING, req, "Refusing libclient connection");
1196 GOTO(out, rc = -EPROTO);
1201 * Note: lw_client is needed in MDS-MDS failover during update log
1202 * processing, so we needs to allow lw_client to be connected at
1203 * anytime, instead of only the initial connection
1205 lw_client = (data->ocd_connect_flags & OBD_CONNECT_LIGHTWEIGHT) != 0;
1207 if (lustre_msg_get_op_flags(req->rq_reqmsg) & MSG_CONNECT_INITIAL) {
1208 initial_conn = true;
1209 mds_conn = (data->ocd_connect_flags & OBD_CONNECT_MDS) != 0;
1210 mds_mds_conn = (data->ocd_connect_flags &
1211 OBD_CONNECT_MDS_MDS) != 0;
1214 * OBD_CONNECT_MNE_SWAB is removed at 2.14
1215 * Checking OBD_CONNECT_FID can be removed in the future.
1217 * Via check OBD_CONNECT_FID, we can distinguish whether
1218 * the OBD_CONNECT_MDS_MDS/OBD_CONNECT_MNE_SWAB is from
1219 * MGC or MDT, since MGC does not use OBD_CONNECT_FID.
1222 (data->ocd_connect_flags & OBD_CONNECT_MDS_MDS) &&
1223 (data->ocd_connect_flags & OBD_CONNECT_FID) &&
1224 (data->ocd_connect_flags & OBD_CONNECT_VERSION)) {
1225 __u32 major = OBD_OCD_VERSION_MAJOR(data->ocd_version);
1226 __u32 minor = OBD_OCD_VERSION_MINOR(data->ocd_version);
1227 __u32 patch = OBD_OCD_VERSION_PATCH(data->ocd_version);
1230 * We do not support the MDT-MDT interoperations with
1231 * different version MDT because of protocol changes.
1233 if (unlikely(major != LUSTRE_MAJOR ||
1234 minor != LUSTRE_MINOR ||
1235 abs(patch - LUSTRE_PATCH) > 3)) {
1236 LCONSOLE_WARN("%s (%u.%u.%u.%u) refused the connection from different version MDT (%d.%d.%d.%d) %s %s\n",
1237 target->obd_name, LUSTRE_MAJOR,
1238 LUSTRE_MINOR, LUSTRE_PATCH,
1239 LUSTRE_FIX, major, minor, patch,
1240 OBD_OCD_VERSION_FIX(data->ocd_version),
1241 libcfs_nid2str(req->rq_peer.nid),
1243 GOTO(out, rc = -EPROTO);
1248 /* lctl gets a backstage, all-access pass. */
1249 if (obd_uuid_equals(&cluuid, &target->obd_uuid))
1250 goto dont_check_exports;
1252 export = obd_uuid_lookup(target, &cluuid);
1256 /* We've found an export in the hash. */
1258 spin_lock(&export->exp_lock);
1260 if (export->exp_connecting) { /* b=9635, et. al. */
1261 spin_unlock(&export->exp_lock);
1262 LCONSOLE_WARN("%s: Export %p already connecting from %s\n",
1263 export->exp_obd->obd_name, export,
1264 libcfs_nid2str(req->rq_peer.nid));
1265 class_export_put(export);
1268 } else if ((mds_conn || (lw_client && initial_conn) ||
1269 data->ocd_connect_flags & OBD_CONNECT_MDS_MDS) &&
1270 export->exp_connection != NULL) {
1271 spin_unlock(&export->exp_lock);
1272 if (req->rq_peer.nid != export->exp_connection->c_peer.nid) {
1273 /* MDS or LWP reconnected after failover. */
1274 LCONSOLE_WARN("%s: Received %s connection from %s, removing former export from %s\n",
1276 mds_conn ? "MDS" : "LWP",
1277 libcfs_nid2str(req->rq_peer.nid),
1278 libcfs_nid2str(export->exp_connection->c_peer.nid));
1280 /* New MDS connection from the same NID. */
1281 LCONSOLE_WARN("%s: Received new %s connection from %s, removing former export from same NID\n",
1283 mds_conn ? "MDS" : "LWP",
1284 libcfs_nid2str(req->rq_peer.nid));
1287 if (req->rq_peer.nid == export->exp_connection->c_peer.nid &&
1288 data->ocd_connect_flags & OBD_CONNECT_MDS_MDS) {
1290 * Because exports between MDTs will always be
1291 * kept, let's do not fail such export if they
1292 * come from the same NID, otherwise it might
1293 * cause eviction between MDTs, which might
1294 * cause namespace inconsistency
1296 spin_lock(&export->exp_lock);
1297 export->exp_connecting = 1;
1298 export->exp_conn_cnt = 0;
1299 spin_unlock(&export->exp_lock);
1300 conn.cookie = export->exp_handle.h_cookie;
1303 class_fail_export(export);
1304 class_export_put(export);
1308 } else if (export->exp_connection != NULL && initial_conn &&
1309 req->rq_peer.nid != export->exp_connection->c_peer.nid) {
1310 spin_unlock(&export->exp_lock);
1311 /* In MDS failover we have static UUID but NID can change. */
1312 LCONSOLE_WARN("%s: Client %s seen on new nid %s when existing nid %s is already connected\n",
1313 target->obd_name, cluuid.uuid,
1314 libcfs_nid2str(req->rq_peer.nid),
1316 export->exp_connection->c_peer.nid));
1318 class_export_put(export);
1320 } else if (OBD_FAIL_PRECHECK(OBD_FAIL_TGT_RECOVERY_CONNECT) &&
1322 spin_unlock(&export->exp_lock);
1325 export->exp_connecting = 1;
1326 spin_unlock(&export->exp_lock);
1327 LASSERT(export->exp_obd == target);
1329 rc = target_handle_reconnect(&conn, export, &cluuid);
1332 /* If we found an export, we already unlocked. */
1335 OBD_FAIL_TIMEOUT(OBD_FAIL_TGT_DELAY_CONNECT, 2 * obd_timeout);
1336 } else if (req->rq_export == NULL &&
1337 atomic_read(&export->exp_rpc_count) > 0) {
1338 LCONSOLE_WARN("%s: Client %s (at %s) refused connection, still busy with %d references\n",
1339 target->obd_name, cluuid.uuid,
1340 libcfs_nid2str(req->rq_peer.nid),
1341 refcount_read(&export->exp_handle.h_ref));
1342 GOTO(out, rc = -EBUSY);
1343 } else if (lustre_msg_get_conn_cnt(req->rq_reqmsg) == 1 &&
1345 if (!strstr(cluuid.uuid, "mdt"))
1346 LCONSOLE_WARN("%s: Rejecting reconnect from the known client %s (at %s) because it is indicating it is a new client\n",
1347 target->obd_name, cluuid.uuid,
1348 libcfs_nid2str(req->rq_peer.nid));
1349 GOTO(out, rc = -EALREADY);
1351 OBD_FAIL_TIMEOUT(OBD_FAIL_TGT_DELAY_RECONNECT, 2 * obd_timeout);
1357 CDEBUG(D_HA, "%s: connection from %s@%s %st%llu exp %p cur %lld last %lld\n",
1358 target->obd_name, cluuid.uuid, libcfs_nid2str(req->rq_peer.nid),
1359 target->obd_recovering ? "recovering/" : "", data->ocd_transno,
1360 export, ktime_get_seconds(),
1361 export ? export->exp_last_request_time : 0);
1364 * If this is the first time a client connects, reset the recovery
1365 * timer. Discard lightweight connections which might be local.
1367 if (!lw_client && rc == 0 && target->obd_recovering)
1368 check_and_start_recovery_timer(target, req, export == NULL);
1371 * We want to handle EALREADY but *not* -EALREADY from
1372 * target_handle_reconnect(), return reconnection state in a flag.
1374 if (rc == EALREADY) {
1375 lustre_msg_add_op_flags(req->rq_repmsg, MSG_CONNECT_RECONNECT);
1381 /* Tell the client if we support replayable requests. */
1382 if (target->obd_replayable)
1383 lustre_msg_add_op_flags(req->rq_repmsg, MSG_CONNECT_REPLAYABLE);
1384 client_nid = &req->rq_peer.nid;
1386 if (export == NULL) {
1387 /* allow lightweight connections during recovery */
1389 * allow "new" MDT to be connected during recovery, since we
1390 * need retrieve recovery update records from it
1392 if (target->obd_recovering && !lw_client && !mds_mds_conn) {
1393 struct hrtimer *timer = &target->obd_recovery_timer;
1402 connected = atomic_read(&target->obd_connected_clients);
1403 in_progress = atomic_read(&target->obd_lock_replay_clients);
1405 atomic_read(&target->obd_max_recoverable_clients);
1406 stale = target->obd_stale_clients;
1407 remaining = hrtimer_get_remaining(timer);
1408 left = ktime_divns(remaining, NSEC_PER_SEC);
1410 if (ktime_to_ns(remaining) > 0) {
1411 msg = "to recover in";
1414 msg = "already passed deadline";
1417 target_check_recovery_timer(target);
1420 LCONSOLE_WARN("%s: Denying connection for new client %s (at %s), waiting for %d known clients (%d recovered, %d in progress, and %d evicted) %s %lld:%.02lld\n",
1421 target->obd_name, cluuid.uuid,
1422 libcfs_nid2str(req->rq_peer.nid), known,
1423 connected - in_progress, in_progress,
1424 stale, msg, timeout / 60, timeout % 60);
1428 rc = obd_connect(req->rq_svc_thread->t_env,
1429 &export, target, &cluuid, data,
1431 if (mds_conn && OBD_FAIL_CHECK(OBD_FAIL_TGT_RCVG_FLAG))
1432 lustre_msg_add_op_flags(req->rq_repmsg,
1433 MSG_CONNECT_RECOVERING);
1435 conn.cookie = export->exp_handle.h_cookie;
1436 rc = rev_import_init(export);
1440 new_mds_mds_conn = true;
1443 rc = obd_reconnect(req->rq_svc_thread->t_env,
1444 export, target, &cluuid, data, client_nid);
1449 LASSERT(target->u.obt.obt_magic == OBT_MAGIC);
1450 data->ocd_instance = target->u.obt.obt_instance;
1453 * Return only the parts of obd_connect_data that we understand, so the
1454 * client knows that we don't understand the rest.
1457 tmpsize = req_capsule_get_size(&req->rq_pill, &RMF_CONNECT_DATA,
1459 tmpdata = req_capsule_server_get(&req->rq_pill,
1462 * Don't use struct assignment here, because the client reply
1463 * buffer may be smaller/larger than the local struct
1466 memcpy(tmpdata, data, min(tmpsize, size));
1470 * If the client and the server are the same node, we will already
1471 * have an export that really points to the client's DLM export,
1472 * because we have a shared handles table.
1474 * XXX this will go away when shaver stops sending the "connect" handle
1475 * in the real "remote handle" field of the request --phik 24 Apr 2003
1477 ptlrpc_request_change_export(req, export);
1479 spin_lock(&export->exp_lock);
1480 if (export->exp_conn_cnt >= lustre_msg_get_conn_cnt(req->rq_reqmsg)) {
1481 spin_unlock(&export->exp_lock);
1483 "%s: %s already connected at greater or equal conn_cnt: %d >= %d\n",
1484 cluuid.uuid, libcfs_nid2str(req->rq_peer.nid),
1485 export->exp_conn_cnt,
1486 lustre_msg_get_conn_cnt(req->rq_reqmsg));
1488 GOTO(out, rc = -EALREADY);
1490 LASSERT(lustre_msg_get_conn_cnt(req->rq_reqmsg) > 0);
1491 export->exp_conn_cnt = lustre_msg_get_conn_cnt(req->rq_reqmsg);
1492 spin_unlock(&export->exp_lock);
1494 if (export->exp_connection != NULL) {
1495 /* Check to see if connection came from another NID. */
1496 if (export->exp_connection->c_peer.nid != req->rq_peer.nid)
1497 obd_nid_del(export->exp_obd, export);
1499 ptlrpc_connection_put(export->exp_connection);
1502 export->exp_connection = ptlrpc_connection_get(req->rq_peer,
1505 if (!export->exp_connection)
1506 GOTO(out, rc = -ENOTCONN);
1508 obd_nid_add(export->exp_obd, export);
1510 lustre_msg_set_handle(req->rq_repmsg, &conn);
1512 rc = rev_import_reconnect(export, req);
1516 if (target->obd_recovering && !export->exp_in_recovery && !lw_client) {
1518 __u64 transno = data->ocd_transno;
1520 spin_lock(&export->exp_lock);
1522 * possible race with class_disconnect_stale_exports,
1523 * export may be already in the eviction process
1525 if (export->exp_failed) {
1526 spin_unlock(&export->exp_lock);
1527 GOTO(out, rc = -ENODEV);
1529 export->exp_in_recovery = 1;
1530 export->exp_req_replay_needed = 1;
1531 export->exp_lock_replay_needed = 1;
1532 spin_unlock(&export->exp_lock);
1534 has_transno = !!(lustre_msg_get_op_flags(req->rq_reqmsg) &
1535 MSG_CONNECT_TRANSNO);
1536 if (has_transno && transno == 0)
1537 CWARN("Connect with zero transno!\n");
1539 if (has_transno && transno > 0 &&
1540 transno < target->obd_next_recovery_transno &&
1541 transno > target->obd_last_committed) {
1542 /* Another way is to use cmpxchg() to be lock-free. */
1543 spin_lock(&target->obd_recovery_task_lock);
1544 if (transno < target->obd_next_recovery_transno)
1545 target->obd_next_recovery_transno = transno;
1546 spin_unlock(&target->obd_recovery_task_lock);
1549 atomic_inc(&target->obd_req_replay_clients);
1550 atomic_inc(&target->obd_lock_replay_clients);
1552 * Note: MDS-MDS connection is allowed to be connected during
1553 * recovery, no matter if the exports needs to be recoveried.
1554 * Because we need retrieve updates logs from all other MDTs.
1555 * So if the MDS-MDS export is new, obd_max_recoverable_clients
1556 * also needs to be increased to match other recovery checking
1559 if (new_mds_mds_conn)
1560 atomic_inc(&target->obd_max_recoverable_clients);
1562 if (atomic_inc_return(&target->obd_connected_clients) ==
1563 atomic_read(&target->obd_max_recoverable_clients))
1564 wake_up(&target->obd_next_transno_waitq);
1567 /* Tell the client we're in recovery, when client is involved in it. */
1568 if (target->obd_recovering && !lw_client)
1569 lustre_msg_add_op_flags(req->rq_repmsg, MSG_CONNECT_RECOVERING);
1573 spin_lock(&export->exp_lock);
1574 export->exp_connecting = 0;
1575 spin_unlock(&export->exp_lock);
1577 class_export_put(export);
1579 if (target != NULL) {
1580 spin_lock(&target->obd_dev_lock);
1581 target->obd_conn_inprogress--;
1582 spin_unlock(&target->obd_dev_lock);
1583 class_decref(target, "find", current);
1585 req->rq_status = rc;
1589 int target_handle_disconnect(struct ptlrpc_request *req)
1595 rc = req_capsule_server_pack(&req->rq_pill);
1599 /* In case of target disconnect, updating sec ctx immediately is
1600 * required in order to record latest sequence number used.
1601 * Sequence is normally updated on export destroy, but this event
1602 * can occur too late, ie after a new target connect request has
1604 * Maintaining correct sequence when client connection becomes idle
1605 * ensures that GSS does not erroneously consider requests as replays.
1607 rc = sptlrpc_export_update_ctx(req->rq_export);
1611 /* Keep the rq_export around so we can send the reply. */
1612 req->rq_status = obd_disconnect(class_export_get(req->rq_export));
1617 void target_destroy_export(struct obd_export *exp)
1619 struct obd_import *imp = NULL;
1621 * exports created from last_rcvd data, and "fake"
1622 * exports created by lctl don't have an import
1624 spin_lock(&exp->exp_lock);
1625 if (exp->exp_imp_reverse != NULL) {
1626 imp = exp->exp_imp_reverse;
1627 exp->exp_imp_reverse = NULL;
1629 spin_unlock(&exp->exp_lock);
1631 client_destroy_import(imp);
1633 LASSERT_ATOMIC_ZERO(&exp->exp_locks_count);
1634 LASSERT_ATOMIC_ZERO(&exp->exp_rpc_count);
1635 LASSERT_ATOMIC_ZERO(&exp->exp_cb_count);
1636 LASSERT_ATOMIC_ZERO(&exp->exp_replay_count);
1638 EXPORT_SYMBOL(target_destroy_export);
1641 * Recovery functions
1643 static void target_request_copy_get(struct ptlrpc_request *req)
1645 class_export_rpc_inc(req->rq_export);
1646 LASSERT(list_empty(&req->rq_list));
1647 INIT_LIST_HEAD(&req->rq_replay_list);
1649 /* Increase refcount to keep request in queue. */
1650 atomic_inc(&req->rq_refcount);
1651 /* Let export know it has replays to be handled. */
1652 atomic_inc(&req->rq_export->exp_replay_count);
1655 static void target_request_copy_put(struct ptlrpc_request *req)
1657 LASSERT(list_empty(&req->rq_replay_list));
1658 LASSERT_ATOMIC_POS(&req->rq_export->exp_replay_count);
1660 atomic_dec(&req->rq_export->exp_replay_count);
1661 class_export_rpc_dec(req->rq_export);
1662 ptlrpc_server_drop_request(req);
1665 static int target_exp_enqueue_req_replay(struct ptlrpc_request *req)
1667 __u64 transno = lustre_msg_get_transno(req->rq_reqmsg);
1668 struct obd_export *exp = req->rq_export;
1669 struct ptlrpc_request *reqiter;
1670 struct ptlrpc_request *dup_req = NULL;
1675 spin_lock(&exp->exp_lock);
1676 list_for_each_entry(reqiter, &exp->exp_req_replay_queue,
1678 if (lustre_msg_get_transno(reqiter->rq_reqmsg) == transno) {
1686 /* We expect it with RESENT and REPLAY flags. */
1687 if ((lustre_msg_get_flags(req->rq_reqmsg) &
1688 (MSG_RESENT | MSG_REPLAY)) != (MSG_RESENT | MSG_REPLAY))
1689 CERROR("invalid flags %x of resent replay\n",
1690 lustre_msg_get_flags(req->rq_reqmsg));
1692 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY) {
1695 new_conn = lustre_msg_get_conn_cnt(req->rq_reqmsg);
1697 lustre_msg_get_conn_cnt(dup_req->rq_reqmsg))
1698 lustre_msg_set_conn_cnt(dup_req->rq_reqmsg,
1702 list_add_tail(&req->rq_replay_list,
1703 &exp->exp_req_replay_queue);
1706 spin_unlock(&exp->exp_lock);
1710 static void target_exp_dequeue_req_replay(struct ptlrpc_request *req)
1712 LASSERT(!list_empty(&req->rq_replay_list));
1713 LASSERT(req->rq_export);
1715 spin_lock(&req->rq_export->exp_lock);
1716 list_del_init(&req->rq_replay_list);
1717 spin_unlock(&req->rq_export->exp_lock);
1720 static void target_finish_recovery(struct lu_target *lut)
1722 struct obd_device *obd = lut->lut_obd;
1726 /* Only log a recovery message when recovery has occurred. */
1727 if (obd->obd_recovery_start) {
1728 time64_t now = ktime_get_seconds();
1729 time64_t elapsed_time;
1731 elapsed_time = max_t(time64_t, now - obd->obd_recovery_start,
1733 LCONSOLE_INFO("%s: Recovery over after %lld:%.02lld, of %d clients %d recovered and %d %s evicted.\n",
1734 obd->obd_name, elapsed_time / 60,
1736 atomic_read(&obd->obd_max_recoverable_clients),
1737 atomic_read(&obd->obd_connected_clients),
1738 obd->obd_stale_clients,
1739 obd->obd_stale_clients == 1 ? "was" : "were");
1742 ldlm_reprocess_recovery_done(obd->obd_namespace);
1743 spin_lock(&obd->obd_recovery_task_lock);
1744 if (!list_empty(&obd->obd_req_replay_queue) ||
1745 !list_empty(&obd->obd_lock_replay_queue) ||
1746 !list_empty(&obd->obd_final_req_queue)) {
1747 CERROR("%s: Recovery queues ( %s%s%s) are not empty\n",
1749 list_empty(&obd->obd_req_replay_queue) ? "" : "req ",
1750 list_empty(&obd->obd_lock_replay_queue) ? \
1752 list_empty(&obd->obd_final_req_queue) ? \
1754 spin_unlock(&obd->obd_recovery_task_lock);
1757 spin_unlock(&obd->obd_recovery_task_lock);
1759 obd->obd_recovery_end = ktime_get_seconds();
1761 /* When recovery finished, cleanup orphans on MDS and OST. */
1762 if (obd->obd_type && OBP(obd, postrecov)) {
1763 int rc = OBP(obd, postrecov)(obd);
1766 LCONSOLE_WARN("%s: Post recovery failed, rc %d\n",
1772 static void abort_req_replay_queue(struct obd_device *obd)
1774 struct ptlrpc_request *req, *n;
1775 LIST_HEAD(abort_list);
1777 spin_lock(&obd->obd_recovery_task_lock);
1778 list_splice_init(&obd->obd_req_replay_queue, &abort_list);
1779 spin_unlock(&obd->obd_recovery_task_lock);
1780 list_for_each_entry_safe(req, n, &abort_list, rq_list) {
1781 DEBUG_REQ(D_WARNING, req, "aborted:");
1782 req->rq_status = -ENOTCONN;
1783 if (ptlrpc_error(req)) {
1784 DEBUG_REQ(D_ERROR, req,
1785 "failed abort_req_reply; skipping");
1787 target_exp_dequeue_req_replay(req);
1788 target_request_copy_put(req);
1792 static void abort_lock_replay_queue(struct obd_device *obd)
1794 struct ptlrpc_request *req, *n;
1795 LIST_HEAD(abort_list);
1797 spin_lock(&obd->obd_recovery_task_lock);
1798 list_splice_init(&obd->obd_lock_replay_queue, &abort_list);
1799 spin_unlock(&obd->obd_recovery_task_lock);
1800 list_for_each_entry_safe(req, n, &abort_list, rq_list) {
1801 DEBUG_REQ(D_ERROR, req, "aborted:");
1802 req->rq_status = -ENOTCONN;
1803 if (ptlrpc_error(req)) {
1804 DEBUG_REQ(D_ERROR, req,
1805 "failed abort_lock_reply; skipping");
1807 target_request_copy_put(req);
1812 * Called from a cleanup function if the device is being cleaned up
1813 * forcefully. The exports should all have been disconnected already,
1814 * the only thing left to do is
1815 * - clear the recovery flags
1816 * - cancel the timer
1817 * - free queued requests and replies, but don't send replies
1818 * Because the obd_stopping flag is set, no new requests should be received.
1820 void target_cleanup_recovery(struct obd_device *obd)
1822 struct ptlrpc_request *req, *n;
1823 LIST_HEAD(clean_list);
1825 spin_lock(&obd->obd_dev_lock);
1826 if (!obd->obd_recovering) {
1827 spin_unlock(&obd->obd_dev_lock);
1831 obd->obd_recovering = obd->obd_abort_recovery = 0;
1832 spin_unlock(&obd->obd_dev_lock);
1834 spin_lock(&obd->obd_recovery_task_lock);
1835 target_cancel_recovery_timer(obd);
1836 list_splice_init(&obd->obd_req_replay_queue, &clean_list);
1837 spin_unlock(&obd->obd_recovery_task_lock);
1839 list_for_each_entry_safe(req, n, &clean_list, rq_list) {
1840 LASSERT(req->rq_reply_state == NULL);
1841 target_exp_dequeue_req_replay(req);
1842 target_request_copy_put(req);
1845 spin_lock(&obd->obd_recovery_task_lock);
1846 list_splice_init(&obd->obd_lock_replay_queue, &clean_list);
1847 list_splice_init(&obd->obd_final_req_queue, &clean_list);
1848 spin_unlock(&obd->obd_recovery_task_lock);
1850 list_for_each_entry_safe(req, n, &clean_list, rq_list) {
1851 LASSERT(req->rq_reply_state == NULL);
1852 target_request_copy_put(req);
1857 EXPORT_SYMBOL(target_cleanup_recovery);
1859 /* obd_recovery_task_lock should be held */
1860 void target_cancel_recovery_timer(struct obd_device *obd)
1862 CDEBUG(D_HA, "%s: cancel recovery timer\n", obd->obd_name);
1863 hrtimer_cancel(&obd->obd_recovery_timer);
1866 static void target_start_recovery_timer(struct obd_device *obd)
1870 if (obd->obd_recovery_start != 0)
1873 spin_lock(&obd->obd_dev_lock);
1874 if (!obd->obd_recovering || obd->obd_abort_recovery) {
1875 spin_unlock(&obd->obd_dev_lock);
1879 LASSERT(obd->obd_recovery_timeout != 0);
1881 if (obd->obd_recovery_start != 0) {
1882 spin_unlock(&obd->obd_dev_lock);
1886 obd->obd_recovery_start = ktime_get_seconds();
1887 delay = ktime_set(obd->obd_recovery_start +
1888 obd->obd_recovery_timeout, 0);
1889 hrtimer_start(&obd->obd_recovery_timer, delay, HRTIMER_MODE_ABS);
1890 spin_unlock(&obd->obd_dev_lock);
1892 LCONSOLE_WARN("%s: Will be in recovery for at least %u:%02u, or until %d client%s reconnect%s\n",
1894 obd->obd_recovery_timeout / 60,
1895 obd->obd_recovery_timeout % 60,
1896 atomic_read(&obd->obd_max_recoverable_clients),
1897 (atomic_read(&obd->obd_max_recoverable_clients) == 1) ?
1899 (atomic_read(&obd->obd_max_recoverable_clients) == 1) ?
1904 * extend recovery window.
1906 * if @extend is true, extend recovery window to have @dr_timeout remaining
1907 * at least; otherwise, make sure the recovery timeout value is not less
1910 static void extend_recovery_timer(struct obd_device *obd, timeout_t dr_timeout,
1917 spin_lock(&obd->obd_dev_lock);
1918 if (!obd->obd_recovering || obd->obd_abort_recovery ||
1919 obd->obd_stopping) {
1920 spin_unlock(&obd->obd_dev_lock);
1923 LASSERT(obd->obd_recovery_start != 0);
1925 left_ns = hrtimer_get_remaining(&obd->obd_recovery_timer);
1926 left = ktime_divns(left_ns, NSEC_PER_SEC);
1929 timeout = obd->obd_recovery_timeout;
1930 /* dr_timeout will happen after the hrtimer has expired.
1931 * Add the excess time to the soft recovery timeout without
1932 * exceeding the hard recovery timeout.
1934 if (dr_timeout > left) {
1935 timeout += dr_timeout - left;
1936 timeout = min_t(timeout_t, obd->obd_recovery_time_hard,
1940 timeout = clamp_t(timeout_t, dr_timeout,
1941 obd->obd_recovery_timeout,
1942 obd->obd_recovery_time_hard);
1945 if (timeout == obd->obd_recovery_time_hard)
1946 CWARN("%s: extended recovery timer reached hard limit: %d, extend: %d\n",
1947 obd->obd_name, timeout, extend);
1949 if (obd->obd_recovery_timeout < timeout) {
1952 obd->obd_recovery_timeout = timeout;
1953 end = ktime_set(obd->obd_recovery_start + timeout, 0);
1954 now = ktime_set(ktime_get_seconds(), 0);
1955 left_ns = ktime_sub(end, now);
1956 hrtimer_start(&obd->obd_recovery_timer, end, HRTIMER_MODE_ABS);
1957 left = ktime_divns(left_ns, NSEC_PER_SEC);
1959 spin_unlock(&obd->obd_dev_lock);
1961 CDEBUG(D_HA, "%s: recovery timer will expire in %d seconds\n",
1962 obd->obd_name, left);
1965 /* Reset the timer with each new client connection */
1967 * This timer is actually reconnect_timer, which is for making sure
1968 * the total recovery window is at least as big as my reconnect
1969 * attempt timing. So the initial recovery time_out will be set to
1970 * OBD_RECOVERY_FACTOR * obd_timeout. If the timeout coming
1971 * from client is bigger than this, then the recovery time_out will
1972 * be extended to make sure the client could be reconnected, in the
1973 * process, the timeout from the new client should be ignored.
1976 check_and_start_recovery_timer(struct obd_device *obd,
1977 struct ptlrpc_request *req,
1980 timeout_t service_timeout = lustre_msg_get_service_timeout(req->rq_reqmsg);
1981 struct obd_device_target *obt = &obd->u.obt;
1983 if (!new_client && service_timeout)
1985 * Teach server about old server's estimates, as first guess
1986 * at how long new requests will take.
1988 at_measured(&req->rq_rqbd->rqbd_svcpt->scp_at_estimate,
1991 target_start_recovery_timer(obd);
1994 * Convert the service time to RPC timeout,
1995 * and reuse service_timeout to limit stack usage.
1997 service_timeout = at_est2timeout(service_timeout);
1999 if (OBD_FAIL_CHECK(OBD_FAIL_TGT_SLUGGISH_NET) &&
2000 service_timeout < at_extra)
2001 service_timeout = at_extra;
2004 * We expect other clients to timeout within service_timeout, then try
2005 * to reconnect, then try the failover server. The max delay between
2006 * connect attempts is SWITCH_MAX + SWITCH_INC + INITIAL.
2008 service_timeout += 2 * INITIAL_CONNECT_TIMEOUT;
2010 LASSERT(obt->obt_magic == OBT_MAGIC);
2011 service_timeout += 2 * (CONNECTION_SWITCH_MAX + CONNECTION_SWITCH_INC);
2012 if (service_timeout > obd->obd_recovery_timeout && !new_client)
2013 extend_recovery_timer(obd, service_timeout, false);
2016 /** Health checking routines */
2017 static inline int exp_connect_healthy(struct obd_export *exp)
2019 return exp->exp_in_recovery;
2022 /** if export done req_replay or has replay in queue */
2023 static inline int exp_req_replay_healthy(struct obd_export *exp)
2025 return (!exp->exp_req_replay_needed ||
2026 atomic_read(&exp->exp_replay_count) > 0);
2030 static inline int exp_req_replay_healthy_or_from_mdt(struct obd_export *exp)
2032 return (exp_connect_flags(exp) & OBD_CONNECT_MDS_MDS) ||
2033 exp_req_replay_healthy(exp);
2036 /** if export done lock_replay or has replay in queue */
2037 static inline int exp_lock_replay_healthy(struct obd_export *exp)
2039 return (!exp->exp_lock_replay_needed ||
2040 atomic_read(&exp->exp_replay_count) > 0);
2043 static inline int exp_vbr_healthy(struct obd_export *exp)
2045 return !exp->exp_vbr_failed;
2048 static inline int exp_finished(struct obd_export *exp)
2050 return exp->exp_in_recovery && !exp->exp_lock_replay_needed;
2053 static inline int exp_finished_or_from_mdt(struct obd_export *exp)
2055 return (exp_connect_flags(exp) & OBD_CONNECT_MDS_MDS) ||
2059 static int check_for_next_transno(struct lu_target *lut)
2061 struct ptlrpc_request *req = NULL;
2062 struct obd_device *obd = lut->lut_obd;
2063 struct target_distribute_txn_data *tdtd = lut->lut_tdtd;
2064 int wake_up = 0, connected, completed, queue_len;
2065 __u64 req_transno = 0;
2066 __u64 update_transno = 0;
2067 __u64 next_transno = 0;
2071 spin_lock(&obd->obd_recovery_task_lock);
2072 if (!list_empty(&obd->obd_req_replay_queue)) {
2073 req = list_entry(obd->obd_req_replay_queue.next,
2074 struct ptlrpc_request, rq_list);
2075 req_transno = lustre_msg_get_transno(req->rq_reqmsg);
2078 if (!obd->obd_abort_recov_mdt && tdtd)
2079 update_transno = distribute_txn_get_next_transno(tdtd);
2081 connected = atomic_read(&obd->obd_connected_clients);
2082 completed = connected - atomic_read(&obd->obd_req_replay_clients);
2083 queue_len = obd->obd_requests_queued_for_recovery;
2084 next_transno = obd->obd_next_recovery_transno;
2087 "max: %d, connected: %d, completed: %d, queue_len: %d, req_transno: %llu, next_transno: %llu\n",
2088 atomic_read(&obd->obd_max_recoverable_clients),
2089 connected, completed,
2090 queue_len, req_transno, next_transno);
2092 if (obd->obd_abort_recovery) {
2093 CDEBUG(D_HA, "waking for aborted recovery\n");
2095 } else if (obd->obd_recovery_expired) {
2096 CDEBUG(D_HA, "waking for expired recovery\n");
2098 } else if (!obd->obd_abort_recov_mdt && tdtd && req &&
2099 is_req_replayed_by_update(req)) {
2100 LASSERTF(req_transno < next_transno,
2101 "req_transno %llu next_transno%llu\n", req_transno,
2103 CDEBUG(D_HA, "waking for duplicate req (%llu)\n",
2106 } else if (req_transno == next_transno ||
2107 (update_transno != 0 && update_transno <= next_transno)) {
2108 CDEBUG(D_HA, "waking for next (%lld)\n", next_transno);
2110 } else if (queue_len > 0 &&
2111 queue_len == atomic_read(&obd->obd_req_replay_clients)) {
2112 /** handle gaps occured due to lost reply or VBR */
2113 LASSERTF(req_transno >= next_transno,
2114 "req_transno: %llu, next_transno: %llu\n",
2115 req_transno, next_transno);
2117 "%s: waking for gap in transno, VBR is %s (skip: %lld, ql: %d, comp: %d, conn: %d, next: %lld, next_update %lld last_committed: %lld)\n",
2118 obd->obd_name, obd->obd_version_recov ? "ON" : "OFF",
2119 next_transno, queue_len, completed, connected,
2120 req_transno, update_transno, obd->obd_last_committed);
2121 obd->obd_next_recovery_transno = req_transno;
2123 } else if (atomic_read(&obd->obd_req_replay_clients) == 0) {
2124 CDEBUG(D_HA, "waking for completed recovery\n");
2126 } else if (OBD_FAIL_CHECK(OBD_FAIL_MDS_RECOVERY_ACCEPTS_GAPS)) {
2128 "accepting transno gaps is explicitly allowed by fail_lock, waking up (%lld)\n",
2130 obd->obd_next_recovery_transno = req_transno;
2133 spin_unlock(&obd->obd_recovery_task_lock);
2137 static int check_for_next_lock(struct lu_target *lut)
2139 struct obd_device *obd = lut->lut_obd;
2142 spin_lock(&obd->obd_recovery_task_lock);
2143 if (!list_empty(&obd->obd_lock_replay_queue)) {
2144 CDEBUG(D_HA, "waking for next lock\n");
2146 } else if (atomic_read(&obd->obd_lock_replay_clients) == 0) {
2147 CDEBUG(D_HA, "waking for completed lock replay\n");
2149 } else if (obd->obd_abort_recovery) {
2150 CDEBUG(D_HA, "waking for aborted recovery\n");
2152 } else if (obd->obd_recovery_expired) {
2153 CDEBUG(D_HA, "waking for expired recovery\n");
2156 spin_unlock(&obd->obd_recovery_task_lock);
2161 static int check_update_llog(struct lu_target *lut)
2163 struct obd_device *obd = lut->lut_obd;
2164 struct target_distribute_txn_data *tdtd = lut->lut_tdtd;
2166 if (obd->obd_abort_recovery) {
2167 CDEBUG(D_HA, "waking for aborted recovery\n");
2171 if (atomic_read(&tdtd->tdtd_recovery_threads_count) == 0) {
2172 CDEBUG(D_HA, "waking for completion of reading update log\n");
2180 * wait for recovery events,
2181 * check its status with help of check_routine
2182 * evict dead clients via health_check
2184 static int target_recovery_overseer(struct lu_target *lut,
2185 int (*check_routine)(struct lu_target *),
2186 int (*health_check)(struct obd_export *))
2188 struct obd_device *obd = lut->lut_obd;
2189 struct target_distribute_txn_data *tdtd;
2193 if (obd->obd_recovering && obd->obd_recovery_start == 0) {
2194 now = ktime_get_seconds();
2195 if (now - last > 600) {
2196 LCONSOLE_INFO("%s: in recovery but waiting for the first client to connect\n",
2201 if (obd->obd_recovery_start != 0 && ktime_get_seconds() >=
2202 (obd->obd_recovery_start + obd->obd_recovery_time_hard)) {
2203 __u64 next_update_transno = 0;
2206 * Only abort the recovery if there are no update recovery
2209 spin_lock(&obd->obd_recovery_task_lock);
2210 if (!obd->obd_abort_recov_mdt && lut->lut_tdtd) {
2211 next_update_transno =
2212 distribute_txn_get_next_transno(lut->lut_tdtd);
2214 tdtd = lut->lut_tdtd;
2216 * If next_update_transno == 0, it probably because
2217 * updatelog retrieve threads did not get any records
2218 * yet, let's wait those threads stopped
2220 if (next_update_transno == 0) {
2221 spin_unlock(&obd->obd_recovery_task_lock);
2223 while (wait_event_timeout(
2224 tdtd->tdtd_recovery_threads_waitq,
2225 check_update_llog(lut),
2226 cfs_time_seconds(60)) == 0);
2228 spin_lock(&obd->obd_recovery_task_lock);
2229 next_update_transno =
2230 distribute_txn_get_next_transno(tdtd);
2234 if (next_update_transno != 0 && !obd->obd_abort_recovery) {
2235 obd->obd_next_recovery_transno = next_update_transno;
2236 spin_unlock(&obd->obd_recovery_task_lock);
2238 * Disconnect unfinished exports from clients, and
2239 * keep connection from MDT to make sure the update
2240 * recovery will still keep trying until some one
2241 * manually abort the recovery
2243 class_disconnect_stale_exports(obd,
2244 exp_finished_or_from_mdt);
2245 /* Abort all of replay & replay lock req from clients */
2246 abort_req_replay_queue(obd);
2247 abort_lock_replay_queue(obd);
2249 "%s: there are still update replay (%#llx)in the queue.\n",
2250 obd->obd_name, next_update_transno);
2252 obd->obd_abort_recovery = 1;
2253 spin_unlock(&obd->obd_recovery_task_lock);
2254 CWARN("%s recovery is aborted by hard timeout\n",
2259 while (wait_event_timeout(obd->obd_next_transno_waitq,
2261 cfs_time_seconds(60)) == 0)
2262 ; /* wait indefinitely for event, but don't trigger watchdog */
2264 if (obd->obd_abort_recovery) {
2265 CWARN("recovery is aborted, evict exports in recovery\n");
2266 if (lut->lut_tdtd != NULL) {
2267 tdtd = lut->lut_tdtd;
2269 * Let's wait all of the update log recovery thread
2273 tdtd->tdtd_recovery_threads_waitq,
2274 atomic_read(&tdtd->tdtd_recovery_threads_count)
2276 /* Then abort the update recovery list */
2277 dtrq_list_destroy(lut->lut_tdtd);
2280 /** evict exports which didn't finish recovery yet */
2281 class_disconnect_stale_exports(obd, exp_finished);
2283 } else if (obd->obd_recovery_expired) {
2284 obd->obd_recovery_expired = 0;
2286 /** If some clients died being recovered, evict them */
2287 LCONSOLE_WARN("%s: recovery is timed out, evict stale exports\n",
2289 /** evict cexports with no replay in queue, they are stalled */
2290 class_disconnect_stale_exports(obd, health_check);
2292 /** continue with VBR */
2293 spin_lock(&obd->obd_dev_lock);
2294 obd->obd_version_recov = 1;
2295 spin_unlock(&obd->obd_dev_lock);
2297 * reset timer, recovery will proceed with versions now,
2298 * timeout is set just to handle reconnection delays
2300 extend_recovery_timer(obd, RECONNECT_DELAY_MAX, true);
2302 * Wait for recovery events again, after evicting bad clients
2309 static struct ptlrpc_request *target_next_replay_lock(struct lu_target *lut)
2311 struct obd_device *obd = lut->lut_obd;
2312 struct ptlrpc_request *req = NULL;
2314 CDEBUG(D_HA, "Waiting for lock\n");
2315 if (target_recovery_overseer(lut, check_for_next_lock,
2316 exp_lock_replay_healthy))
2317 abort_lock_replay_queue(obd);
2319 spin_lock(&obd->obd_recovery_task_lock);
2320 if (!list_empty(&obd->obd_lock_replay_queue)) {
2321 req = list_entry(obd->obd_lock_replay_queue.next,
2322 struct ptlrpc_request, rq_list);
2323 list_del_init(&req->rq_list);
2324 spin_unlock(&obd->obd_recovery_task_lock);
2326 spin_unlock(&obd->obd_recovery_task_lock);
2327 LASSERT(list_empty(&obd->obd_lock_replay_queue));
2328 LASSERT(atomic_read(&obd->obd_lock_replay_clients) == 0);
2329 /** evict exports failed VBR */
2330 class_disconnect_stale_exports(obd, exp_vbr_healthy);
2335 static struct ptlrpc_request *target_next_final_ping(struct obd_device *obd)
2337 struct ptlrpc_request *req = NULL;
2339 spin_lock(&obd->obd_recovery_task_lock);
2340 if (!list_empty(&obd->obd_final_req_queue)) {
2341 req = list_entry(obd->obd_final_req_queue.next,
2342 struct ptlrpc_request, rq_list);
2343 list_del_init(&req->rq_list);
2344 spin_unlock(&obd->obd_recovery_task_lock);
2345 if (req->rq_export->exp_in_recovery) {
2346 spin_lock(&req->rq_export->exp_lock);
2347 req->rq_export->exp_in_recovery = 0;
2348 spin_unlock(&req->rq_export->exp_lock);
2351 spin_unlock(&obd->obd_recovery_task_lock);
2356 static void handle_recovery_req(struct ptlrpc_thread *thread,
2357 struct ptlrpc_request *req,
2358 svc_handler_t handler)
2363 * export can be evicted during recovery, no need to handle replays for
2364 * it after that, discard such request silently
2366 if (req->rq_export->exp_disconnected)
2369 req->rq_session.lc_thread = thread;
2370 req->rq_svc_thread = thread;
2371 req->rq_svc_thread->t_env->le_ses = &req->rq_session;
2373 /* thread context */
2374 lu_context_enter(&thread->t_env->le_ctx);
2376 lu_context_exit(&thread->t_env->le_ctx);
2378 req->rq_svc_thread->t_env->le_ses = NULL;
2380 /* don't reset timer for final stage */
2381 if (!exp_finished(req->rq_export)) {
2382 timeout_t timeout = obd_timeout;
2385 * Add request @timeout to the recovery time so next request from
2386 * this client may come in recovery time
2389 struct ptlrpc_service_part *svcpt;
2390 timeout_t est_timeout;
2392 svcpt = req->rq_rqbd->rqbd_svcpt;
2394 * If the server sent early reply for this request,
2395 * the client will recalculate the timeout according to
2396 * current server estimate service time, so we will
2397 * use the maxium timeout here for waiting the client
2398 * sending the next req
2400 est_timeout = at_get(&svcpt->scp_at_estimate);
2401 timeout = max_t(timeout_t, at_est2timeout(est_timeout),
2402 lustre_msg_get_timeout(req->rq_reqmsg));
2404 * Add 2 net_latency, one for balance rq_deadline
2405 * (see ptl_send_rpc), one for resend the req to server,
2406 * Note: client will pack net_latency in replay req
2407 * (see ptlrpc_replay_req)
2409 timeout += 2 * lustre_msg_get_service_timeout(req->rq_reqmsg);
2411 extend_recovery_timer(class_exp2obd(req->rq_export), timeout,
2417 /** Checking routines for recovery */
2418 static int check_for_recovery_ready(struct lu_target *lut)
2420 struct obd_device *obd = lut->lut_obd;
2421 unsigned int clnts = atomic_read(&obd->obd_connected_clients);
2424 "connected %d stale %d max_recoverable_clients %d abort %d expired %d\n",
2425 clnts, obd->obd_stale_clients,
2426 atomic_read(&obd->obd_max_recoverable_clients),
2427 obd->obd_abort_recovery, obd->obd_recovery_expired);
2429 if (!obd->obd_abort_recovery && !obd->obd_recovery_expired) {
2431 atomic_read(&obd->obd_max_recoverable_clients));
2432 if (clnts + obd->obd_stale_clients <
2433 atomic_read(&obd->obd_max_recoverable_clients))
2437 if (!obd->obd_abort_recov_mdt && lut->lut_tdtd != NULL) {
2438 if (!lut->lut_tdtd->tdtd_replay_ready &&
2439 !obd->obd_abort_recovery && !obd->obd_stopping) {
2441 * Let's extend recovery timer, in case the recovery
2442 * timer expired, and some clients got evicted
2444 extend_recovery_timer(obd, obd->obd_recovery_timeout,
2447 "%s update recovery is not ready, extend recovery %d\n",
2448 obd->obd_name, obd->obd_recovery_timeout);
2457 REQUEST_RECOVERY = 1,
2458 UPDATE_RECOVERY = 2,
2461 static __u64 get_next_replay_req_transno(struct obd_device *obd)
2465 if (!list_empty(&obd->obd_req_replay_queue)) {
2466 struct ptlrpc_request *req;
2468 req = list_entry(obd->obd_req_replay_queue.next,
2469 struct ptlrpc_request, rq_list);
2470 transno = lustre_msg_get_transno(req->rq_reqmsg);
2476 static __u64 get_next_transno(struct lu_target *lut, int *type)
2478 struct obd_device *obd = lut->lut_obd;
2479 struct target_distribute_txn_data *tdtd = lut->lut_tdtd;
2481 __u64 update_transno;
2485 transno = get_next_replay_req_transno(obd);
2487 *type = REQUEST_RECOVERY;
2489 if (!tdtd || obd->obd_abort_recov_mdt)
2492 update_transno = distribute_txn_get_next_transno(tdtd);
2493 if (transno == 0 || (transno >= update_transno &&
2494 update_transno != 0)) {
2495 transno = update_transno;
2497 *type = UPDATE_RECOVERY;
2504 * drop duplicate replay request
2506 * Because the operation has been replayed by update recovery, the request
2507 * with the same transno will be dropped and also notify the client to send
2508 * next replay request.
2510 * \param[in] env execution environment
2511 * \param[in] obd failover obd device
2512 * \param[in] req request to be dropped
2514 static void drop_duplicate_replay_req(struct lu_env *env,
2515 struct obd_device *obd,
2516 struct ptlrpc_request *req)
2518 DEBUG_REQ(D_HA, req,
2519 "remove t%lld from %s because duplicate update records found",
2520 lustre_msg_get_transno(req->rq_reqmsg),
2521 libcfs_nid2str(req->rq_peer.nid));
2524 * Right now, only for MDS reint operation update replay and
2525 * normal request replay can have the same transno
2527 if (lustre_msg_get_opc(req->rq_reqmsg) == MDS_REINT) {
2528 req_capsule_set(&req->rq_pill, &RQF_MDS_REINT);
2529 req->rq_status = req_capsule_server_pack(&req->rq_pill);
2530 if (likely(req->rq_export))
2531 target_committed_to_req(req);
2532 lustre_msg_set_transno(req->rq_repmsg, req->rq_transno);
2533 target_send_reply(req, req->rq_status, 0);
2535 DEBUG_REQ(D_ERROR, req, "wrong opc from %s",
2536 libcfs_nid2str(req->rq_peer.nid));
2538 target_exp_dequeue_req_replay(req);
2539 target_request_copy_put(req);
2540 obd->obd_replayed_requests++;
2543 #define WATCHDOG_TIMEOUT (obd_timeout * 10)
2545 static void replay_request_or_update(struct lu_env *env,
2546 struct lu_target *lut,
2547 struct target_recovery_data *trd,
2548 struct ptlrpc_thread *thread)
2550 struct obd_device *obd = lut->lut_obd;
2551 struct ptlrpc_request *req = NULL;
2557 CDEBUG(D_HA, "Waiting for transno %lld\n",
2558 obd->obd_next_recovery_transno);
2560 /* Replay all of request and update by transno */
2562 struct target_distribute_txn_data *tdtd = lut->lut_tdtd;
2564 CFS_FAIL_TIMEOUT(OBD_FAIL_TGT_REPLAY_DELAY2, cfs_fail_val);
2567 * It is needed to extend recovery window above
2568 * recovery_time_soft. Extending is possible only in the
2569 * end of recovery window (see more details in
2570 * handle_recovery_req()).
2572 CFS_FAIL_TIMEOUT_MS(OBD_FAIL_TGT_REPLAY_DELAY, 300);
2574 if (target_recovery_overseer(lut, check_for_next_transno,
2575 exp_req_replay_healthy_or_from_mdt)) {
2576 abort_req_replay_queue(obd);
2577 abort_lock_replay_queue(obd);
2581 spin_lock(&obd->obd_recovery_task_lock);
2582 transno = get_next_transno(lut, &type);
2583 if (type == REQUEST_RECOVERY && transno != 0) {
2585 * Drop replay request from client side, if the
2586 * replay has been executed by update with the
2589 req = list_entry(obd->obd_req_replay_queue.next,
2590 struct ptlrpc_request, rq_list);
2592 list_del_init(&req->rq_list);
2593 obd->obd_requests_queued_for_recovery--;
2594 spin_unlock(&obd->obd_recovery_task_lock);
2597 * Let's check if the request has been redone by
2600 if (is_req_replayed_by_update(req)) {
2601 struct distribute_txn_replay_req *dtrq;
2603 dtrq = distribute_txn_lookup_finish_list(tdtd,
2605 LASSERT(dtrq != NULL);
2606 spin_lock(&tdtd->tdtd_replay_list_lock);
2607 list_del_init(&dtrq->dtrq_list);
2608 spin_unlock(&tdtd->tdtd_replay_list_lock);
2611 drop_duplicate_replay_req(env, obd, req);
2616 LASSERT(trd->trd_processing_task == current->pid);
2617 DEBUG_REQ(D_HA, req, "processing x%llu t%lld from %s",
2619 lustre_msg_get_transno(req->rq_reqmsg),
2620 libcfs_nid2str(req->rq_peer.nid));
2622 ptlrpc_watchdog_init(&thread->t_watchdog,
2624 handle_recovery_req(thread, req,
2625 trd->trd_recovery_handler);
2626 ptlrpc_watchdog_disable(&thread->t_watchdog);
2629 * bz18031: increase next_recovery_transno before
2630 * target_request_copy_put() will drop exp_rpc reference
2632 spin_lock(&obd->obd_recovery_task_lock);
2633 obd->obd_next_recovery_transno++;
2634 spin_unlock(&obd->obd_recovery_task_lock);
2635 target_exp_dequeue_req_replay(req);
2636 target_request_copy_put(req);
2637 obd->obd_replayed_requests++;
2638 } else if (type == UPDATE_RECOVERY && transno != 0) {
2639 struct distribute_txn_replay_req *dtrq;
2642 spin_unlock(&obd->obd_recovery_task_lock);
2644 LASSERT(tdtd != NULL);
2645 dtrq = distribute_txn_get_next_req(tdtd);
2646 lu_context_enter(&thread->t_env->le_ctx);
2647 ptlrpc_watchdog_init(&thread->t_watchdog,
2649 rc = tdtd->tdtd_replay_handler(env, tdtd, dtrq);
2650 ptlrpc_watchdog_disable(&thread->t_watchdog);
2651 lu_context_exit(&thread->t_env->le_ctx);
2652 extend_recovery_timer(obd, obd_timeout, true);
2654 if (rc == 0 && dtrq->dtrq_xid != 0) {
2656 "Move x%llu t%llu to finish list\n",
2658 dtrq->dtrq_master_transno);
2660 /* Add it to the replay finish list */
2661 spin_lock(&tdtd->tdtd_replay_list_lock);
2662 list_add(&dtrq->dtrq_list,
2663 &tdtd->tdtd_replay_finish_list);
2664 spin_unlock(&tdtd->tdtd_replay_list_lock);
2666 spin_lock(&obd->obd_recovery_task_lock);
2667 if (transno == obd->obd_next_recovery_transno)
2668 obd->obd_next_recovery_transno++;
2670 obd->obd_next_recovery_transno)
2671 obd->obd_next_recovery_transno =
2673 spin_unlock(&obd->obd_recovery_task_lock);
2678 spin_unlock(&obd->obd_recovery_task_lock);
2680 LASSERT(list_empty(&obd->obd_req_replay_queue));
2681 LASSERT(atomic_read(&obd->obd_req_replay_clients) == 0);
2682 /** evict exports failed VBR */
2683 class_disconnect_stale_exports(obd, exp_vbr_healthy);
2689 static int target_recovery_thread(void *arg)
2691 struct lu_target *lut = arg;
2692 struct obd_device *obd = lut->lut_obd;
2693 struct ptlrpc_request *req;
2694 struct target_recovery_data *trd = &obd->obd_recovery_data;
2695 unsigned long delta;
2697 struct ptlrpc_thread *thread = NULL;
2701 OBD_ALLOC_PTR(thread);
2707 GOTO(out_thread, rc = -ENOMEM);
2708 rc = lu_env_add(env);
2712 rc = lu_context_init(&env->le_ctx, LCT_MD_THREAD | LCT_DT_THREAD);
2714 GOTO(out_env_remove, rc);
2716 thread->t_env = env;
2717 thread->t_id = -1; /* force filter_iobuf_get/put to use local buffers */
2718 env->le_ctx.lc_thread = thread;
2719 tgt_io_thread_init(thread); /* init thread_big_cache for IO requests */
2721 CDEBUG(D_HA, "%s: started recovery thread pid %d\n", obd->obd_name,
2723 trd->trd_processing_task = current->pid;
2725 spin_lock(&obd->obd_dev_lock);
2726 obd->obd_recovering = 1;
2727 spin_unlock(&obd->obd_dev_lock);
2728 complete(&trd->trd_starting);
2730 /* first of all, we have to know the first transno to replay */
2731 if (target_recovery_overseer(lut, check_for_recovery_ready,
2732 exp_connect_healthy)) {
2733 abort_req_replay_queue(obd);
2734 abort_lock_replay_queue(obd);
2735 if (lut->lut_tdtd != NULL)
2736 dtrq_list_destroy(lut->lut_tdtd);
2739 /* next stage: replay requests or update */
2741 CDEBUG(D_INFO, "1: request replay stage - %d clients from t%llu\n",
2742 atomic_read(&obd->obd_req_replay_clients),
2743 obd->obd_next_recovery_transno);
2744 replay_request_or_update(env, lut, trd, thread);
2747 * The second stage: replay locks
2749 CDEBUG(D_INFO, "2: lock replay stage - %d clients\n",
2750 atomic_read(&obd->obd_lock_replay_clients));
2751 while ((req = target_next_replay_lock(lut))) {
2752 LASSERT(trd->trd_processing_task == current->pid);
2753 DEBUG_REQ(D_HA, req, "processing lock from %s:",
2754 libcfs_nid2str(req->rq_peer.nid));
2755 if (OBD_FAIL_CHECK(OBD_FAIL_LDLM_LOCK_REPLAY)) {
2756 req->rq_status = -ENODEV;
2757 target_request_copy_put(req);
2760 handle_recovery_req(thread, req,
2761 trd->trd_recovery_handler);
2762 target_request_copy_put(req);
2763 obd->obd_replayed_locks++;
2767 * The third stage: reply on final pings, at this moment all clients
2768 * must have request in final queue
2770 CFS_FAIL_TIMEOUT(OBD_FAIL_TGT_REPLAY_RECONNECT, cfs_fail_val);
2771 CDEBUG(D_INFO, "3: final stage - process recovery completion pings\n");
2772 /** Update server last boot epoch */
2773 tgt_boot_epoch_update(lut);
2775 * We drop recoverying flag to forward all new requests
2776 * to regular mds_handle() since now
2778 spin_lock(&obd->obd_dev_lock);
2779 obd->obd_recovering = obd->obd_abort_recovery = 0;
2780 spin_unlock(&obd->obd_dev_lock);
2781 spin_lock(&obd->obd_recovery_task_lock);
2782 target_cancel_recovery_timer(obd);
2783 spin_unlock(&obd->obd_recovery_task_lock);
2784 while ((req = target_next_final_ping(obd))) {
2785 LASSERT(trd->trd_processing_task == current->pid);
2786 DEBUG_REQ(D_HA, req, "processing final ping from %s:",
2787 libcfs_nid2str(req->rq_peer.nid));
2788 handle_recovery_req(thread, req,
2789 trd->trd_recovery_handler);
2791 * Because the waiting client can not send ping to server,
2792 * so we need refresh the last_request_time, to avoid the
2793 * export is being evicted
2795 ptlrpc_update_export_timer(req->rq_export, 0);
2796 target_request_copy_put(req);
2799 delta = jiffies_to_msecs(jiffies - delta) / MSEC_PER_SEC;
2800 CDEBUG(D_INFO, "4: recovery completed in %lus - %d/%d reqs/locks\n",
2801 delta, obd->obd_replayed_requests, obd->obd_replayed_locks);
2802 if (delta > OBD_RECOVERY_TIME_SOFT) {
2803 CWARN("too long recovery - read logs\n");
2804 libcfs_debug_dumplog();
2807 target_finish_recovery(lut);
2808 lu_context_fini(&env->le_ctx);
2809 trd->trd_processing_task = 0;
2810 complete_all(&trd->trd_finishing);
2811 tgt_io_thread_done(thread);
2817 OBD_FREE_PTR(thread);
2821 static int target_start_recovery_thread(struct lu_target *lut,
2822 svc_handler_t handler)
2824 struct obd_device *obd = lut->lut_obd;
2826 struct target_recovery_data *trd = &obd->obd_recovery_data;
2829 memset(trd, 0, sizeof(*trd));
2830 init_completion(&trd->trd_starting);
2831 init_completion(&trd->trd_finishing);
2832 trd->trd_recovery_handler = handler;
2834 rc = server_name2index(obd->obd_name, &index, NULL);
2838 if (!IS_ERR(kthread_run(target_recovery_thread,
2839 lut, "tgt_recover_%d", index))) {
2840 wait_for_completion(&trd->trd_starting);
2841 LASSERT(obd->obd_recovering != 0);
2849 void target_stop_recovery_thread(struct obd_device *obd)
2851 if (obd->obd_recovery_data.trd_processing_task > 0) {
2852 struct target_recovery_data *trd = &obd->obd_recovery_data;
2853 /** recovery can be done but postrecovery is not yet */
2854 spin_lock(&obd->obd_dev_lock);
2855 if (obd->obd_recovering) {
2856 CERROR("%s: Aborting recovery\n", obd->obd_name);
2857 obd->obd_abort_recovery = 1;
2858 wake_up(&obd->obd_next_transno_waitq);
2860 spin_unlock(&obd->obd_dev_lock);
2861 wait_for_completion(&trd->trd_finishing);
2864 EXPORT_SYMBOL(target_stop_recovery_thread);
2866 void target_recovery_fini(struct obd_device *obd)
2868 class_disconnect_exports(obd);
2869 target_stop_recovery_thread(obd);
2870 target_cleanup_recovery(obd);
2872 EXPORT_SYMBOL(target_recovery_fini);
2874 static enum hrtimer_restart target_recovery_expired(struct hrtimer *timer)
2876 struct obd_device *obd = container_of(timer, struct obd_device,
2877 obd_recovery_timer);
2880 "%s: recovery timed out; %d clients are still in recovery after %llu seconds (%d clients connected)\n",
2881 obd->obd_name, atomic_read(&obd->obd_lock_replay_clients),
2882 ktime_get_seconds() - obd->obd_recovery_start,
2883 atomic_read(&obd->obd_connected_clients));
2885 obd->obd_recovery_expired = 1;
2886 wake_up(&obd->obd_next_transno_waitq);
2887 return HRTIMER_NORESTART;
2890 void target_recovery_init(struct lu_target *lut, svc_handler_t handler)
2892 struct obd_device *obd = lut->lut_obd;
2894 if (lut->lut_bottom->dd_rdonly)
2897 if (atomic_read(&obd->obd_max_recoverable_clients) == 0) {
2898 /** Update server last boot epoch */
2899 tgt_boot_epoch_update(lut);
2903 CDEBUG(D_HA, "RECOVERY: service %s, %d recoverable clients, "
2904 "last_transno %llu\n", obd->obd_name,
2905 atomic_read(&obd->obd_max_recoverable_clients),
2906 obd->obd_last_committed);
2907 LASSERT(obd->obd_stopping == 0);
2908 obd->obd_next_recovery_transno = obd->obd_last_committed + 1;
2909 obd->obd_recovery_start = 0;
2910 obd->obd_recovery_end = 0;
2912 hrtimer_init(&obd->obd_recovery_timer, CLOCK_MONOTONIC,
2914 obd->obd_recovery_timer.function = &target_recovery_expired;
2915 target_start_recovery_thread(lut, handler);
2917 EXPORT_SYMBOL(target_recovery_init);
2919 static int target_process_req_flags(struct obd_device *obd,
2920 struct ptlrpc_request *req)
2922 struct obd_export *exp = req->rq_export;
2924 LASSERT(exp != NULL);
2925 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REQ_REPLAY_DONE) {
2926 /* client declares he's ready to replay locks */
2927 spin_lock(&exp->exp_lock);
2928 if (exp->exp_req_replay_needed) {
2929 exp->exp_req_replay_needed = 0;
2930 spin_unlock(&exp->exp_lock);
2932 LASSERT_ATOMIC_POS(&obd->obd_req_replay_clients);
2933 atomic_dec(&obd->obd_req_replay_clients);
2935 spin_unlock(&exp->exp_lock);
2938 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_LOCK_REPLAY_DONE) {
2940 * client declares he's ready to complete recovery
2941 * so, we put the request on th final queue
2943 spin_lock(&exp->exp_lock);
2944 if (exp->exp_lock_replay_needed) {
2945 exp->exp_lock_replay_needed = 0;
2946 spin_unlock(&exp->exp_lock);
2948 LASSERT_ATOMIC_POS(&obd->obd_lock_replay_clients);
2949 atomic_dec(&obd->obd_lock_replay_clients);
2951 spin_unlock(&exp->exp_lock);
2957 int target_queue_recovery_request(struct ptlrpc_request *req,
2958 struct obd_device *obd)
2960 __u64 transno = lustre_msg_get_transno(req->rq_reqmsg);
2961 struct ptlrpc_request *reqiter;
2966 if (obd->obd_recovery_data.trd_processing_task == current->pid) {
2967 /* Processing the queue right now, don't re-add. */
2971 target_process_req_flags(obd, req);
2973 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_LOCK_REPLAY_DONE) {
2974 if (unlikely(OBD_FAIL_CHECK(OBD_FAIL_TGT_RECOVERY_REQ_RACE))) {
2975 if (cfs_fail_val == 1) {
2978 wake_up(&cfs_race_waitq);
2980 schedule_timeout_interruptible(
2981 cfs_time_seconds(1));
2986 * client declares he's ready to complete recovery
2987 * so, we put the request on th final queue
2989 target_request_copy_get(req);
2990 DEBUG_REQ(D_HA, req, "queue final req");
2991 wake_up(&obd->obd_next_transno_waitq);
2992 spin_lock(&obd->obd_recovery_task_lock);
2993 if (obd->obd_recovering) {
2994 struct ptlrpc_request *tmp;
2995 struct ptlrpc_request *duplicate = NULL;
2997 if (likely(!req->rq_export->exp_replay_done)) {
2998 req->rq_export->exp_replay_done = 1;
2999 list_add_tail(&req->rq_list,
3000 &obd->obd_final_req_queue);
3001 spin_unlock(&obd->obd_recovery_task_lock);
3006 * XXX O(n), but only happens if final ping is
3007 * timed out, probably reorganize the list as
3010 list_for_each_entry_safe(reqiter, tmp,
3011 &obd->obd_final_req_queue,
3013 if (reqiter->rq_export == req->rq_export) {
3014 list_del_init(&reqiter->rq_list);
3015 duplicate = reqiter;
3020 list_add_tail(&req->rq_list,
3021 &obd->obd_final_req_queue);
3022 req->rq_export->exp_replay_done = 1;
3023 spin_unlock(&obd->obd_recovery_task_lock);
3025 if (duplicate != NULL) {
3026 DEBUG_REQ(D_HA, duplicate,
3027 "put prev final req");
3028 target_request_copy_put(duplicate);
3032 spin_unlock(&obd->obd_recovery_task_lock);
3033 target_request_copy_put(req);
3034 RETURN(obd->obd_stopping ? -ENOTCONN : 1);
3037 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REQ_REPLAY_DONE) {
3038 /* client declares he's ready to replay locks */
3039 target_request_copy_get(req);
3040 DEBUG_REQ(D_HA, req, "queue lock replay req");
3041 wake_up(&obd->obd_next_transno_waitq);
3042 spin_lock(&obd->obd_recovery_task_lock);
3043 LASSERT(obd->obd_recovering);
3044 /* usually due to recovery abort */
3045 if (!req->rq_export->exp_in_recovery) {
3046 spin_unlock(&obd->obd_recovery_task_lock);
3047 target_request_copy_put(req);
3050 LASSERT(req->rq_export->exp_lock_replay_needed);
3051 list_add_tail(&req->rq_list, &obd->obd_lock_replay_queue);
3052 spin_unlock(&obd->obd_recovery_task_lock);
3057 * CAVEAT EMPTOR: The incoming request message has been swabbed
3058 * (i.e. buflens etc are in my own byte order), but type-dependent
3059 * buffers (eg mdt_body, ost_body etc) have NOT been swabbed.
3063 INIT_LIST_HEAD(&req->rq_list);
3064 DEBUG_REQ(D_HA, req, "not queueing");
3069 * If we're processing the queue, we want don't want to queue this
3072 * Also, if this request has a transno less than the one we're waiting
3073 * for, we should process it now. It could (and currently always will)
3074 * be an open request for a descriptor that was opened some time ago.
3076 * Also, a resent, replayed request that has already been
3077 * handled will pass through here and be processed immediately.
3080 "Next recovery transno: %llu, current: %llu, replaying\n",
3081 obd->obd_next_recovery_transno, transno);
3084 * If the request has been replayed by update replay, then sends this
3085 * request to the recovery thread (replay_request_or_update()), where
3086 * it will be handled
3088 spin_lock(&obd->obd_recovery_task_lock);
3089 if (transno < obd->obd_next_recovery_transno &&
3090 !is_req_replayed_by_update(req)) {
3091 /* Processing the queue right now, don't re-add. */
3092 LASSERT(list_empty(&req->rq_list));
3093 spin_unlock(&obd->obd_recovery_task_lock);
3096 spin_unlock(&obd->obd_recovery_task_lock);
3098 if (OBD_FAIL_CHECK(OBD_FAIL_TGT_REPLAY_DROP))
3101 target_request_copy_get(req);
3102 if (!req->rq_export->exp_in_recovery) {
3103 target_request_copy_put(req);
3106 LASSERT(req->rq_export->exp_req_replay_needed);
3108 if (target_exp_enqueue_req_replay(req)) {
3109 DEBUG_REQ(D_ERROR, req, "dropping resent queued req");
3110 target_request_copy_put(req);
3115 spin_lock(&obd->obd_recovery_task_lock);
3116 LASSERT(obd->obd_recovering);
3117 list_for_each_entry(reqiter, &obd->obd_req_replay_queue, rq_list) {
3118 if (lustre_msg_get_transno(reqiter->rq_reqmsg) > transno) {
3119 list_add_tail(&req->rq_list, &reqiter->rq_list);
3124 if (unlikely(lustre_msg_get_transno(reqiter->rq_reqmsg) ==
3126 DEBUG_REQ(D_ERROR, req,
3127 "dropping replay: transno has been claimed by another client");
3128 spin_unlock(&obd->obd_recovery_task_lock);
3129 target_exp_dequeue_req_replay(req);
3130 target_request_copy_put(req);
3136 list_add_tail(&req->rq_list, &obd->obd_req_replay_queue);
3138 obd->obd_requests_queued_for_recovery++;
3139 spin_unlock(&obd->obd_recovery_task_lock);
3140 wake_up(&obd->obd_next_transno_waitq);
3144 void target_committed_to_req(struct ptlrpc_request *req)
3146 struct obd_export *exp = req->rq_export;
3148 if (!exp->exp_obd->obd_no_transno && req->rq_repmsg != NULL)
3149 lustre_msg_set_last_committed(req->rq_repmsg,
3150 exp->exp_last_committed);
3152 DEBUG_REQ(D_IOCTL, req,
3153 "not sending last_committed update (%d/%d)",
3154 exp->exp_obd->obd_no_transno,
3155 req->rq_repmsg == NULL);
3157 CDEBUG(D_INFO, "last_committed %llu, transno %llu, xid %llu\n",
3158 exp->exp_last_committed, req->rq_transno, req->rq_xid);
3161 #endif /* HAVE_SERVER_SUPPORT */
3164 * Packs current SLV and Limit into \a req.
3166 int target_pack_pool_reply(struct ptlrpc_request *req)
3168 struct obd_device *obd;
3173 * Check that we still have all structures alive as this may
3174 * be some late RPC at shutdown time.
3176 if (unlikely(!req->rq_export || !req->rq_export->exp_obd ||
3177 !exp_connect_lru_resize(req->rq_export))) {
3178 lustre_msg_set_slv(req->rq_repmsg, 0);
3179 lustre_msg_set_limit(req->rq_repmsg, 0);
3183 /* OBD is alive here as export is alive, which we checked above. */
3184 obd = req->rq_export->exp_obd;
3186 read_lock(&obd->obd_pool_lock);
3187 lustre_msg_set_slv(req->rq_repmsg, obd->obd_pool_slv);
3188 lustre_msg_set_limit(req->rq_repmsg, obd->obd_pool_limit);
3189 read_unlock(&obd->obd_pool_lock);
3194 static int target_send_reply_msg(struct ptlrpc_request *req,
3195 int rc, int fail_id)
3197 if (OBD_FAIL_CHECK_ORSET(fail_id & ~OBD_FAIL_ONCE, OBD_FAIL_ONCE)) {
3198 DEBUG_REQ(D_ERROR, req, "dropping reply");
3202 * We can have a null rq_reqmsg in the event of bad signature or
3203 * no context when unwrapping
3205 if (req->rq_reqmsg &&
3206 unlikely(lustre_msg_get_opc(req->rq_reqmsg) == MDS_REINT &&
3207 OBD_FAIL_CHECK(OBD_FAIL_MDS_REINT_MULTI_NET_REP)))
3211 DEBUG_REQ(D_NET, req, "processing error (%d)", rc);
3212 req->rq_status = rc;
3213 return ptlrpc_send_error(req, 1);
3215 DEBUG_REQ(D_NET, req, "sending reply");
3217 return ptlrpc_send_reply(req, PTLRPC_REPLY_MAYBE_DIFFICULT);
3220 void target_send_reply(struct ptlrpc_request *req, int rc, int fail_id)
3222 struct ptlrpc_service_part *svcpt;
3224 struct ptlrpc_reply_state *rs;
3225 struct obd_export *exp;
3229 if (req->rq_no_reply) {
3234 svcpt = req->rq_rqbd->rqbd_svcpt;
3235 rs = req->rq_reply_state;
3236 if (rs == NULL || !rs->rs_difficult) {
3238 target_send_reply_msg(req, rc, fail_id);
3243 /* must be an export if locks saved */
3244 LASSERT(req->rq_export != NULL);
3245 /* req/reply consistent */
3246 LASSERT(rs->rs_svcpt == svcpt);
3249 LASSERT(!rs->rs_scheduled);
3250 LASSERT(!rs->rs_scheduled_ever);
3251 LASSERT(!rs->rs_handled);
3252 LASSERT(!rs->rs_on_net);
3253 LASSERT(rs->rs_export == NULL);
3254 LASSERT(list_empty(&rs->rs_obd_list));
3255 LASSERT(list_empty(&rs->rs_exp_list));
3257 exp = class_export_get(req->rq_export);
3259 /* disable reply scheduling while I'm setting up */
3260 rs->rs_scheduled = 1;
3262 rs->rs_xid = req->rq_xid;
3263 rs->rs_transno = req->rq_transno;
3264 rs->rs_export = exp;
3265 rs->rs_opc = lustre_msg_get_opc(req->rq_reqmsg);
3267 spin_lock(&exp->exp_uncommitted_replies_lock);
3268 CDEBUG(D_NET, "rs transno = %llu, last committed = %llu\n",
3269 rs->rs_transno, exp->exp_last_committed);
3270 if (rs->rs_transno > exp->exp_last_committed) {
3271 /* not committed already */
3272 list_add_tail(&rs->rs_obd_list,
3273 &exp->exp_uncommitted_replies);
3275 spin_unlock(&exp->exp_uncommitted_replies_lock);
3277 spin_lock(&exp->exp_lock);
3278 list_add_tail(&rs->rs_exp_list, &exp->exp_outstanding_replies);
3279 spin_unlock(&exp->exp_lock);
3281 netrc = target_send_reply_msg(req, rc, fail_id);
3283 spin_lock(&svcpt->scp_rep_lock);
3285 atomic_inc(&svcpt->scp_nreps_difficult);
3289 * error sending: reply is off the net. Also we need +1
3290 * reply ref until ptlrpc_handle_rs() is done
3291 * with the reply state (if the send was successful, there
3292 * would have been +1 ref for the net, which
3293 * reply_out_callback leaves alone)
3296 ptlrpc_rs_addref(rs);
3299 spin_lock(&rs->rs_lock);
3300 if (rs->rs_transno <= exp->exp_last_committed ||
3301 (!rs->rs_on_net && !rs->rs_no_ack) ||
3302 list_empty(&rs->rs_exp_list) || /* completed already */
3303 list_empty(&rs->rs_obd_list)) {
3304 CDEBUG(D_HA, "Schedule reply immediately\n");
3305 ptlrpc_dispatch_difficult_reply(rs);
3307 list_add(&rs->rs_list, &svcpt->scp_rep_active);
3308 rs->rs_scheduled = 0; /* allow notifier to schedule */
3310 spin_unlock(&rs->rs_lock);
3311 spin_unlock(&svcpt->scp_rep_lock);
3315 enum ldlm_mode lck_compat_array[] = {
3316 [LCK_EX] = LCK_COMPAT_EX,
3317 [LCK_PW] = LCK_COMPAT_PW,
3318 [LCK_PR] = LCK_COMPAT_PR,
3319 [LCK_CW] = LCK_COMPAT_CW,
3320 [LCK_CR] = LCK_COMPAT_CR,
3321 [LCK_NL] = LCK_COMPAT_NL,
3322 [LCK_GROUP] = LCK_COMPAT_GROUP,
3323 [LCK_COS] = LCK_COMPAT_COS,
3327 * Rather arbitrary mapping from LDLM error codes to errno values. This should
3328 * not escape to the user level.
3330 int ldlm_error2errno(enum ldlm_error error)
3336 case ELDLM_LOCK_MATCHED:
3339 case ELDLM_LOCK_CHANGED:
3342 case ELDLM_LOCK_ABORTED:
3345 case ELDLM_LOCK_REPLACED:
3348 case ELDLM_NO_LOCK_DATA:
3351 case ELDLM_NAMESPACE_EXISTS:
3354 case ELDLM_BAD_NAMESPACE:
3358 if (((int)error) < 0) { /* cast to signed type */
3359 result = error; /* as ldlm_error can be unsigned */
3361 CERROR("Invalid DLM result code: %d\n", error);
3367 EXPORT_SYMBOL(ldlm_error2errno);
3370 * Dual to ldlm_error2errno(): maps errno values back to enum ldlm_error.
3372 enum ldlm_error ldlm_errno2error(int err_no)
3381 error = ELDLM_LOCK_CHANGED;
3384 error = ELDLM_LOCK_ABORTED;
3387 error = ELDLM_LOCK_REPLACED;
3390 error = ELDLM_NO_LOCK_DATA;
3393 error = ELDLM_NAMESPACE_EXISTS;
3396 error = ELDLM_BAD_NAMESPACE;
3404 #if LUSTRE_TRACKS_LOCK_EXP_REFS
3405 void ldlm_dump_export_locks(struct obd_export *exp)
3407 spin_lock(&exp->exp_locks_list_guard);
3408 if (!list_empty(&exp->exp_locks_list)) {
3409 struct ldlm_lock *lock;
3411 CERROR("dumping locks for export %p, ignore if the unmount doesn't hang\n",
3413 list_for_each_entry(lock, &exp->exp_locks_list,
3415 LDLM_ERROR(lock, "lock:");
3417 spin_unlock(&exp->exp_locks_list_guard);
3421 #ifdef HAVE_SERVER_SUPPORT
3422 static inline const char *bulk2type(struct ptlrpc_request *req)
3424 if (req->rq_bulk_read)
3426 if (req->rq_bulk_write)
3431 int target_bulk_io(struct obd_export *exp, struct ptlrpc_bulk_desc *desc)
3433 struct ptlrpc_request *req = desc->bd_req;
3434 time64_t start = ktime_get_seconds();
3440 /* If there is eviction in progress, wait for it to finish. */
3442 exp->exp_obd->obd_evict_inprogress_waitq,
3443 !atomic_read(&exp->exp_obd->obd_evict_inprogress));
3445 /* Check if client was evicted or reconnected already. */
3446 if (exp->exp_failed ||
3447 exp->exp_conn_cnt > lustre_msg_get_conn_cnt(req->rq_reqmsg)) {
3450 if (req->rq_bulk_read)
3451 rc = sptlrpc_svc_wrap_bulk(req, desc);
3453 if (OCD_HAS_FLAG(&exp->exp_connect_data, BULK_MBITS))
3454 req->rq_mbits = lustre_msg_get_mbits(req->rq_reqmsg);
3455 else /* old version, bulk matchbits is rq_xid */
3456 req->rq_mbits = req->rq_xid;
3459 rc = ptlrpc_start_bulk_transfer(desc);
3463 DEBUG_REQ(D_ERROR, req, "bulk %s failed: rc = %d",
3464 bulk2type(req), rc);
3468 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_SENDPAGE)) {
3469 ptlrpc_abort_bulk(desc);
3473 /* limit actual bulk transfer to bulk_timeout seconds */
3474 deadline = start + bulk_timeout;
3475 if (deadline > req->rq_deadline)
3476 deadline = req->rq_deadline;
3479 time64_t timeoutl = deadline - ktime_get_seconds();
3480 time64_t rq_deadline;
3482 while (timeoutl >= 0 &&
3483 wait_event_idle_timeout(
3485 !ptlrpc_server_bulk_active(desc) ||
3488 lustre_msg_get_conn_cnt(req->rq_reqmsg),
3489 timeoutl ? cfs_time_seconds(1) : 1) == 0)
3491 rc = timeoutl < 0 ? -ETIMEDOUT : 0;
3493 /* Wait again if we changed rq_deadline. */
3494 rq_deadline = READ_ONCE(req->rq_deadline);
3495 deadline = start + bulk_timeout;
3496 if (deadline > rq_deadline)
3497 deadline = rq_deadline;
3498 } while (rc == -ETIMEDOUT &&
3499 deadline > ktime_get_seconds());
3501 if (rc == -ETIMEDOUT) {
3502 DEBUG_REQ(D_ERROR, req, "timeout on bulk %s after %lld%+llds",
3503 bulk2type(req), deadline - start,
3504 ktime_get_real_seconds() - deadline);
3505 ptlrpc_abort_bulk(desc);
3506 } else if (exp->exp_failed) {
3507 DEBUG_REQ(D_ERROR, req, "Eviction on bulk %s",
3510 ptlrpc_abort_bulk(desc);
3511 } else if (exp->exp_conn_cnt >
3512 lustre_msg_get_conn_cnt(req->rq_reqmsg)) {
3513 DEBUG_REQ(D_ERROR, req, "Reconnect on bulk %s",
3515 /* We don't reply anyway. */
3517 ptlrpc_abort_bulk(desc);
3518 } else if (desc->bd_failure) {
3519 DEBUG_REQ(D_ERROR, req, "network error on bulk %s",
3521 /* XXX should this be a different errno? */
3524 if (req->rq_bulk_write)
3525 rc = sptlrpc_svc_unwrap_bulk(req, desc);
3526 if (rc == 0 && desc->bd_nob_transferred != desc->bd_nob) {
3527 DEBUG_REQ(D_ERROR, req, "truncated bulk %s %d(%d)",
3528 bulk2type(req), desc->bd_nob_transferred,
3530 /* XXX should this be a different errno? */
3537 EXPORT_SYMBOL(target_bulk_io);
3539 #endif /* HAVE_SERVER_SUPPORT */