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/
33 * This file deals with various client/target related logic including recovery.
35 * TODO: This code more logically belongs in the ptlrpc module than in ldlm and
39 #define DEBUG_SUBSYSTEM S_LDLM
41 #include <cl_object.h>
42 #include <linux/jiffies.h>
43 #include <linux/kernel.h>
44 #include <linux/kthread.h>
45 #include <libcfs/libcfs.h>
47 #include <obd_class.h>
48 #include <lustre_dlm.h>
49 #include <lustre_net.h>
50 #include <lustre_sec.h>
51 #include "ldlm_internal.h"
54 * @priority: If non-zero, move the selected connection to the list head.
55 * @create: If zero, only search in existing connections.
57 static int import_set_conn(struct obd_import *imp, struct obd_uuid *uuid,
58 int priority, int create)
60 struct ptlrpc_connection *ptlrpc_conn;
61 struct obd_import_conn *imp_conn = NULL, *item;
62 lnet_nid_t nid4refnet = LNET_NID_ANY;
67 if (!create && !priority) {
68 CDEBUG(D_HA, "Nothing to do\n");
72 if (imp->imp_connection &&
73 imp->imp_connection->c_remote_uuid.uuid[0] == 0)
74 /* nid4refnet is used to restrict network connections */
75 nid4refnet = imp->imp_connection->c_self;
76 ptlrpc_conn = ptlrpc_uuid_to_connection(uuid, nid4refnet);
78 CDEBUG(D_HA, "can't find connection %s\n", uuid->uuid);
83 OBD_ALLOC(imp_conn, sizeof(*imp_conn));
85 GOTO(out_put, rc = -ENOMEM);
88 spin_lock(&imp->imp_lock);
89 list_for_each_entry(item, &imp->imp_conn_list, oic_item) {
90 if (obd_uuid_equals(uuid, &item->oic_uuid)) {
92 list_move(&item->oic_item,
94 item->oic_last_attempt = 0;
96 CDEBUG(D_HA, "imp %p@%s: found existing conn %s%s\n",
97 imp, imp->imp_obd->obd_name, uuid->uuid,
98 (priority ? ", moved to head" : ""));
99 spin_unlock(&imp->imp_lock);
100 GOTO(out_free, rc = 0);
103 /* No existing import connection found for \a uuid. */
105 imp_conn->oic_conn = ptlrpc_conn;
106 imp_conn->oic_uuid = *uuid;
107 imp_conn->oic_last_attempt = 0;
109 list_add(&imp_conn->oic_item, &imp->imp_conn_list);
111 list_add_tail(&imp_conn->oic_item,
112 &imp->imp_conn_list);
113 CDEBUG(D_HA, "imp %p@%s: add connection %s at %s\n",
114 imp, imp->imp_obd->obd_name, uuid->uuid,
115 (priority ? "head" : "tail"));
117 spin_unlock(&imp->imp_lock);
118 GOTO(out_free, rc = -ENOENT);
121 spin_unlock(&imp->imp_lock);
125 OBD_FREE(imp_conn, sizeof(*imp_conn));
127 ptlrpc_connection_put(ptlrpc_conn);
131 int import_set_conn_priority(struct obd_import *imp, struct obd_uuid *uuid)
133 return import_set_conn(imp, uuid, 1, 0);
136 int client_import_add_conn(struct obd_import *imp, struct obd_uuid *uuid,
139 return import_set_conn(imp, uuid, priority, 1);
141 EXPORT_SYMBOL(client_import_add_conn);
143 int client_import_dyn_add_conn(struct obd_import *imp, struct obd_uuid *uuid,
144 lnet_nid_t prim_nid, int priority)
146 struct ptlrpc_connection *ptlrpc_conn;
149 ptlrpc_conn = ptlrpc_uuid_to_connection(uuid, prim_nid);
151 const char *str_uuid = obd_uuid2str(uuid);
153 rc = class_add_uuid(str_uuid, prim_nid);
155 CERROR("%s: failed to add UUID '%s': rc = %d\n",
156 imp->imp_obd->obd_name, str_uuid, rc);
160 return import_set_conn(imp, uuid, priority, 1);
162 EXPORT_SYMBOL(client_import_dyn_add_conn);
164 int client_import_add_nids_to_conn(struct obd_import *imp, lnet_nid_t *nids,
165 int nid_count, struct obd_uuid *uuid)
167 struct obd_import_conn *conn;
171 if (nid_count <= 0 || !nids)
174 spin_lock(&imp->imp_lock);
175 list_for_each_entry(conn, &imp->imp_conn_list, oic_item) {
176 if (class_check_uuid(&conn->oic_uuid, nids[0])) {
177 *uuid = conn->oic_uuid;
178 spin_unlock(&imp->imp_lock);
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 revimp->imp_connect_data = *data;
987 rc = sptlrpc_import_sec_adapt(revimp, req->rq_svc_ctx, &req->rq_flvr);
989 CERROR("%s: cannot get reverse import %s security: rc = %d\n",
990 revimp->imp_client->cli_name,
991 libcfs_id2str(req->rq_peer), rc);
999 * Allocate a new reverse import for an export.
1001 * \retval -errno in case error hit
1002 * \retval 0 if reverse import correctly init
1004 int rev_import_init(struct obd_export *export)
1006 struct obd_device *obd = export->exp_obd;
1007 struct obd_import *revimp;
1009 LASSERT(export->exp_imp_reverse == NULL);
1011 revimp = class_new_import(obd);
1015 revimp->imp_remote_handle.cookie = 0ULL;
1016 revimp->imp_client = &obd->obd_ldlm_client;
1017 revimp->imp_dlm_fake = 1;
1019 /* it is safe to connect import in new state as no sends possible */
1020 spin_lock(&export->exp_lock);
1021 export->exp_imp_reverse = revimp;
1022 spin_unlock(&export->exp_lock);
1023 class_import_put(revimp);
1027 EXPORT_SYMBOL(rev_import_init);
1030 * Handle reconnect for an export.
1032 * \param exp export to handle reconnect process
1033 * \param req client reconnect request
1035 * \retval -rc in case securitfy flavor can't be changed
1036 * \retval 0 in case none problems
1038 static int rev_import_reconnect(struct obd_export *exp,
1039 struct ptlrpc_request *req)
1041 struct obd_import *revimp = exp->exp_imp_reverse;
1042 struct lustre_handle *lh;
1045 /* avoid sending a request until import flags are changed */
1046 ptlrpc_import_enter_resend(revimp);
1048 ptlrpc_connection_put(revimp->imp_connection);
1051 * client from recovery don't have a handle so we need to take from
1052 * request. it may produce situation when wrong client connected
1053 * to recovery as we trust a client uuid
1055 lh = req_capsule_client_get(&req->rq_pill, &RMF_CONN);
1056 revimp->imp_remote_handle = *lh;
1059 * unknown versions will be caught in
1060 * ptlrpc_handle_server_req_in->lustre_unpack_msg()
1062 revimp->imp_msg_magic = req->rq_reqmsg->lm_magic;
1064 revimp->imp_connection = ptlrpc_connection_addref(exp->exp_connection);
1066 rc = rev_import_flags_update(revimp, req);
1069 * it is safe to still be in RECOVERY phase as we are not able
1070 * to setup correct security flavor so requests are not able to
1071 * be delivered correctly
1076 /* resend all rpc's via new connection */
1077 return ptlrpc_import_recovery_state_machine(revimp);
1080 int target_handle_connect(struct ptlrpc_request *req)
1082 struct obd_device *target = NULL;
1083 struct obd_export *export = NULL;
1085 * connect handle - filled from target_handle_reconnect in
1088 struct lustre_handle conn;
1089 struct lustre_handle *tmp;
1090 struct obd_uuid cluuid;
1095 bool mds_conn = false, lw_client = false, initial_conn = false;
1096 bool mds_mds_conn = false;
1097 bool new_mds_mds_conn = false;
1098 struct obd_connect_data *data, *tmpdata;
1100 lnet_nid_t *client_nid = NULL;
1101 struct ptlrpc_connection *pcon = NULL;
1105 OBD_RACE(OBD_FAIL_TGT_CONN_RACE);
1107 str = req_capsule_client_get(&req->rq_pill, &RMF_TGTUUID);
1109 DEBUG_REQ(D_ERROR, req, "bad target UUID for connect");
1110 GOTO(out, rc = -EINVAL);
1113 target = class_dev_by_str(str);
1115 deuuidify(str, NULL, &target_start, &target_len);
1116 LCONSOLE_ERROR_MSG(0x137,
1117 "%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",
1118 str, libcfs_nid2str(req->rq_peer.nid));
1119 GOTO(out, rc = -ENODEV);
1122 spin_lock(&target->obd_dev_lock);
1124 target->obd_conn_inprogress++;
1126 if (target->obd_stopping || !target->obd_set_up) {
1127 spin_unlock(&target->obd_dev_lock);
1129 deuuidify(str, NULL, &target_start, &target_len);
1130 LCONSOLE_INFO("%.*s: Not available for connect from %s (%s)\n",
1131 target_len, target_start,
1132 libcfs_nid2str(req->rq_peer.nid),
1133 (target->obd_stopping ?
1134 "stopping" : "not set up"));
1135 GOTO(out, rc = -ENODEV);
1138 if (target->obd_no_conn) {
1139 spin_unlock(&target->obd_dev_lock);
1142 "%s: Temporarily refusing client connection from %s\n",
1143 target->obd_name, libcfs_nid2str(req->rq_peer.nid));
1144 GOTO(out, rc = -EAGAIN);
1147 spin_unlock(&target->obd_dev_lock);
1149 str = req_capsule_client_get(&req->rq_pill, &RMF_CLUUID);
1151 DEBUG_REQ(D_ERROR, req, "bad client UUID for connect");
1152 GOTO(out, rc = -EINVAL);
1155 obd_str2uuid(&cluuid, str);
1157 tmp = req_capsule_client_get(&req->rq_pill, &RMF_CONN);
1159 GOTO(out, rc = -EPROTO);
1163 size = req_capsule_get_size(&req->rq_pill, &RMF_CONNECT_DATA,
1165 if (size < 0 || size > 8 * sizeof(struct obd_connect_data))
1166 GOTO(out, rc = -EPROTO);
1167 data = req_capsule_client_get(&req->rq_pill, &RMF_CONNECT_DATA);
1169 GOTO(out, rc = -EPROTO);
1171 rc = req_capsule_server_pack(&req->rq_pill);
1175 #if LUSTRE_VERSION_CODE < OBD_OCD_VERSION(3, 0, 53, 0)
1177 * Don't allow clients to connect that are using old 1.8 format
1178 * protocol conventions (LUSTRE_MSG_MAGIC_v1, !MSGHDR_CKSUM_INCOMPAT18,
1179 * ldlm_flock_policy_wire format, MDT_ATTR_xTIME_SET, etc). The
1180 * FULL20 flag should be set on all connections since 2.0, but no
1181 * longer affects behaviour.
1183 * Later this check will be disabled and the flag can be retired
1184 * completely once interop with 3.0 is no longer needed.
1186 if (!(data->ocd_connect_flags & OBD_CONNECT_FULL20))
1187 GOTO(out, rc = -EPROTO);
1190 * Don't allow liblustre clients to connect.
1191 * - testing was disabled in v2_2_50_0-61-g6a75d65
1192 * - building was disabled in v2_5_58_0-28-g7277179
1193 * - client code was deleted in v2_6_50_0-101-gcdfbc72,
1194 * - clients were refused connect for version difference > 0.0.1.32
1196 if (lustre_msg_get_op_flags(req->rq_reqmsg) & MSG_CONNECT_LIBCLIENT) {
1197 DEBUG_REQ(D_WARNING, req, "Refusing libclient connection");
1198 GOTO(out, rc = -EPROTO);
1203 * Note: lw_client is needed in MDS-MDS failover during update log
1204 * processing, so we needs to allow lw_client to be connected at
1205 * anytime, instead of only the initial connection
1207 lw_client = OCD_HAS_FLAG(data, LIGHTWEIGHT);
1209 if (lustre_msg_get_op_flags(req->rq_reqmsg) & MSG_CONNECT_INITIAL) {
1210 initial_conn = true;
1211 mds_conn = OCD_HAS_FLAG(data, MDS);
1212 mds_mds_conn = OCD_HAS_FLAG(data, MDS_MDS);
1215 * OBD_CONNECT_MNE_SWAB is removed at 2.14
1216 * Checking OBD_CONNECT_FID can be removed in the future.
1218 * Via check OBD_CONNECT_FID, we can distinguish whether
1219 * the OBD_CONNECT_MDS_MDS/OBD_CONNECT_MNE_SWAB is from
1220 * MGC or MDT, since MGC does not use OBD_CONNECT_FID.
1223 (data->ocd_connect_flags & OBD_CONNECT_MDS_MDS) &&
1224 (data->ocd_connect_flags & OBD_CONNECT_FID) &&
1225 (data->ocd_connect_flags & OBD_CONNECT_VERSION)) {
1226 __u32 major = OBD_OCD_VERSION_MAJOR(data->ocd_version);
1227 __u32 minor = OBD_OCD_VERSION_MINOR(data->ocd_version);
1228 __u32 patch = OBD_OCD_VERSION_PATCH(data->ocd_version);
1231 * We do not support the MDT-MDT interoperations with
1232 * different version MDT because of protocol changes.
1234 if (unlikely(major != LUSTRE_MAJOR ||
1235 minor != LUSTRE_MINOR ||
1236 abs(patch - LUSTRE_PATCH) > 3)) {
1237 LCONSOLE_WARN("%s (%u.%u.%u.%u) refused the connection from different version MDT (%d.%d.%d.%d) %s %s\n",
1238 target->obd_name, LUSTRE_MAJOR,
1239 LUSTRE_MINOR, LUSTRE_PATCH,
1240 LUSTRE_FIX, major, minor, patch,
1241 OBD_OCD_VERSION_FIX(data->ocd_version),
1242 libcfs_nid2str(req->rq_peer.nid),
1244 GOTO(out, rc = -EPROTO);
1249 /* lctl gets a backstage, all-access pass. */
1250 if (obd_uuid_equals(&cluuid, &target->obd_uuid))
1251 goto dont_check_exports;
1253 export = obd_uuid_lookup(target, &cluuid);
1257 /* We've found an export in the hash. */
1259 spin_lock(&export->exp_lock);
1261 if (export->exp_connecting) { /* b=9635, et. al. */
1262 spin_unlock(&export->exp_lock);
1263 LCONSOLE_WARN("%s: Export %p already connecting from %s\n",
1264 export->exp_obd->obd_name, export,
1265 libcfs_nid2str(req->rq_peer.nid));
1266 class_export_put(export);
1269 } else if ((mds_conn || (lw_client && initial_conn) ||
1270 OCD_HAS_FLAG(data, MDS_MDS)) && export->exp_connection) {
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 lw_client ? "LWP" : "MDS",
1277 libcfs_nid2str(req->rq_peer.nid),
1278 libcfs_nid2str(export->exp_connection->c_peer.nid));
1280 /* New connection from the same NID. */
1281 LCONSOLE_WARN("%s: Received new %s connection from %s, %s former export from same NID\n",
1283 lw_client ? "LWP" : "MDS",
1284 libcfs_nid2str(req->rq_peer.nid),
1285 OCD_HAS_FLAG(data, MDS_MDS) ?
1289 if (req->rq_peer.nid == export->exp_connection->c_peer.nid &&
1290 OCD_HAS_FLAG(data, MDS_MDS)) {
1292 * Because exports between MDTs will always be
1293 * kept, let's do not fail such export if they
1294 * come from the same NID, otherwise it might
1295 * cause eviction between MDTs, which might
1296 * cause namespace inconsistency
1298 spin_lock(&export->exp_lock);
1299 export->exp_connecting = 1;
1300 export->exp_conn_cnt = 0;
1301 spin_unlock(&export->exp_lock);
1302 conn.cookie = export->exp_handle.h_cookie;
1305 class_fail_export(export);
1306 class_export_put(export);
1310 } else if (export->exp_connection != NULL && initial_conn &&
1311 req->rq_peer.nid != export->exp_connection->c_peer.nid) {
1312 spin_unlock(&export->exp_lock);
1313 /* In MDS failover we have static UUID but NID can change. */
1314 LCONSOLE_WARN("%s: Client %s seen on new nid %s when existing nid %s is already connected\n",
1315 target->obd_name, cluuid.uuid,
1316 libcfs_nid2str(req->rq_peer.nid),
1318 export->exp_connection->c_peer.nid));
1320 class_export_put(export);
1322 } else if (OBD_FAIL_PRECHECK(OBD_FAIL_TGT_RECOVERY_CONNECT) &&
1324 spin_unlock(&export->exp_lock);
1327 export->exp_connecting = 1;
1328 spin_unlock(&export->exp_lock);
1329 LASSERT(export->exp_obd == target);
1331 rc = target_handle_reconnect(&conn, export, &cluuid);
1334 /* If we found an export, we already unlocked. */
1337 OBD_FAIL_TIMEOUT(OBD_FAIL_TGT_DELAY_CONNECT, 2 * obd_timeout);
1338 } else if (req->rq_export == NULL &&
1339 atomic_read(&export->exp_rpc_count) > 0) {
1340 LCONSOLE_WARN("%s: Client %s (at %s) refused connection, still busy with %d references\n",
1341 target->obd_name, cluuid.uuid,
1342 libcfs_nid2str(req->rq_peer.nid),
1343 refcount_read(&export->exp_handle.h_ref));
1344 GOTO(out, rc = -EBUSY);
1345 } else if (lustre_msg_get_conn_cnt(req->rq_reqmsg) == 1 &&
1347 if (!strstr(cluuid.uuid, "mdt"))
1348 LCONSOLE_WARN("%s: Rejecting reconnect from the known client %s (at %s) because it is indicating it is a new client\n",
1349 target->obd_name, cluuid.uuid,
1350 libcfs_nid2str(req->rq_peer.nid));
1351 GOTO(out, rc = -EALREADY);
1353 OBD_FAIL_TIMEOUT(OBD_FAIL_TGT_DELAY_RECONNECT, 2 * obd_timeout);
1359 CDEBUG(D_HA, "%s: connection from %s@%s %st%llu exp %p cur %lld last %lld\n",
1360 target->obd_name, cluuid.uuid, libcfs_nid2str(req->rq_peer.nid),
1361 target->obd_recovering ? "recovering/" : "", data->ocd_transno,
1362 export, ktime_get_seconds(),
1363 export ? export->exp_last_request_time : 0);
1366 * If this is the first time a client connects, reset the recovery
1367 * timer. Discard lightweight connections which might be local.
1369 if (!lw_client && rc == 0 && target->obd_recovering)
1370 check_and_start_recovery_timer(target, req, export == NULL);
1373 * We want to handle EALREADY but *not* -EALREADY from
1374 * target_handle_reconnect(), return reconnection state in a flag.
1376 if (rc == EALREADY) {
1377 lustre_msg_add_op_flags(req->rq_repmsg, MSG_CONNECT_RECONNECT);
1383 /* Tell the client if we support replayable requests. */
1384 if (target->obd_replayable)
1385 lustre_msg_add_op_flags(req->rq_repmsg, MSG_CONNECT_REPLAYABLE);
1386 client_nid = &req->rq_peer.nid;
1388 if (export == NULL) {
1389 /* allow lightweight connections during recovery */
1391 * allow "new" MDT to be connected during recovery, since we
1392 * need retrieve recovery update records from it
1394 if (target->obd_recovering && !lw_client && !mds_mds_conn) {
1395 struct hrtimer *timer = &target->obd_recovery_timer;
1404 connected = atomic_read(&target->obd_connected_clients);
1405 in_progress = atomic_read(&target->obd_lock_replay_clients);
1407 atomic_read(&target->obd_max_recoverable_clients);
1408 stale = target->obd_stale_clients;
1409 remaining = hrtimer_get_remaining(timer);
1410 left = ktime_divns(remaining, NSEC_PER_SEC);
1412 if (ktime_to_ns(remaining) > 0) {
1413 msg = "to recover in";
1416 msg = "already passed deadline";
1419 target_check_recovery_timer(target);
1422 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",
1423 target->obd_name, cluuid.uuid,
1424 libcfs_nid2str(req->rq_peer.nid), known,
1425 connected - in_progress, in_progress,
1426 stale, msg, timeout / 60, timeout % 60);
1430 rc = obd_connect(req->rq_svc_thread->t_env,
1431 &export, target, &cluuid, data,
1433 if (mds_conn && OBD_FAIL_CHECK(OBD_FAIL_TGT_RCVG_FLAG))
1434 lustre_msg_add_op_flags(req->rq_repmsg,
1435 MSG_CONNECT_RECOVERING);
1437 conn.cookie = export->exp_handle.h_cookie;
1438 rc = rev_import_init(export);
1442 new_mds_mds_conn = true;
1445 rc = obd_reconnect(req->rq_svc_thread->t_env,
1446 export, target, &cluuid, data, client_nid);
1451 LASSERT(target->u.obt.obt_magic == OBT_MAGIC);
1452 data->ocd_instance = target->u.obt.obt_instance;
1455 * Return only the parts of obd_connect_data that we understand, so the
1456 * client knows that we don't understand the rest.
1459 tmpsize = req_capsule_get_size(&req->rq_pill, &RMF_CONNECT_DATA,
1461 tmpdata = req_capsule_server_get(&req->rq_pill,
1464 * Don't use struct assignment here, because the client reply
1465 * buffer may be smaller/larger than the local struct
1468 memcpy(tmpdata, data, min(tmpsize, size));
1472 * If the client and the server are the same node, we will already
1473 * have an export that really points to the client's DLM export,
1474 * because we have a shared handles table.
1476 * XXX this will go away when shaver stops sending the "connect" handle
1477 * in the real "remote handle" field of the request --phik 24 Apr 2003
1479 ptlrpc_request_change_export(req, export);
1481 pcon = ptlrpc_connection_get(req->rq_peer, req->rq_self, &cluuid);
1483 GOTO(out, rc = -ENOTCONN);
1485 spin_lock(&export->exp_lock);
1487 if (export->exp_disconnected) {
1488 spin_unlock(&export->exp_lock);
1489 GOTO(out, rc = -ENODEV);
1491 if (export->exp_conn_cnt >= lustre_msg_get_conn_cnt(req->rq_reqmsg)) {
1492 spin_unlock(&export->exp_lock);
1494 "%s: %s already connected at greater or equal conn_cnt: %d >= %d\n",
1495 cluuid.uuid, libcfs_nid2str(req->rq_peer.nid),
1496 export->exp_conn_cnt,
1497 lustre_msg_get_conn_cnt(req->rq_reqmsg));
1499 GOTO(out, rc = -EALREADY);
1501 LASSERT(lustre_msg_get_conn_cnt(req->rq_reqmsg) > 0);
1502 export->exp_conn_cnt = lustre_msg_get_conn_cnt(req->rq_reqmsg);
1504 /* Check to see if connection came from another NID. */
1505 if (export->exp_connection != NULL &&
1506 export->exp_connection->c_peer.nid != req->rq_peer.nid) {
1507 obd_nid_del(export->exp_obd, export);
1508 ptlrpc_connection_put(export->exp_connection);
1509 export->exp_connection = NULL;
1512 if (export->exp_connection == NULL) {
1513 export->exp_connection = pcon;
1516 obd_nid_add(export->exp_obd, export);
1518 spin_unlock(&export->exp_lock);
1520 lustre_msg_set_handle(req->rq_repmsg, &conn);
1522 rc = rev_import_reconnect(export, req);
1526 if (target->obd_recovering && !export->exp_in_recovery && !lw_client) {
1528 __u64 transno = data->ocd_transno;
1530 spin_lock(&export->exp_lock);
1532 * possible race with class_disconnect_stale_exports,
1533 * export may be already in the eviction process
1535 if (export->exp_failed) {
1536 spin_unlock(&export->exp_lock);
1537 GOTO(out, rc = -ENODEV);
1539 export->exp_in_recovery = 1;
1540 export->exp_req_replay_needed = 1;
1541 export->exp_lock_replay_needed = 1;
1542 spin_unlock(&export->exp_lock);
1544 has_transno = !!(lustre_msg_get_op_flags(req->rq_reqmsg) &
1545 MSG_CONNECT_TRANSNO);
1546 if (has_transno && transno == 0)
1547 CWARN("Connect with zero transno!\n");
1549 if (has_transno && transno > 0 &&
1550 transno < target->obd_next_recovery_transno &&
1551 transno > target->obd_last_committed) {
1552 /* Another way is to use cmpxchg() to be lock-free. */
1553 spin_lock(&target->obd_recovery_task_lock);
1554 if (transno < target->obd_next_recovery_transno)
1555 target->obd_next_recovery_transno = transno;
1556 spin_unlock(&target->obd_recovery_task_lock);
1559 atomic_inc(&target->obd_req_replay_clients);
1560 atomic_inc(&target->obd_lock_replay_clients);
1562 * Note: MDS-MDS connection is allowed to be connected during
1563 * recovery, no matter if the exports needs to be recoveried.
1564 * Because we need retrieve updates logs from all other MDTs.
1565 * So if the MDS-MDS export is new, obd_max_recoverable_clients
1566 * also needs to be increased to match other recovery checking
1569 if (new_mds_mds_conn)
1570 atomic_inc(&target->obd_max_recoverable_clients);
1572 if (atomic_inc_return(&target->obd_connected_clients) ==
1573 atomic_read(&target->obd_max_recoverable_clients))
1574 wake_up(&target->obd_next_transno_waitq);
1577 /* Tell the client we're in recovery, when client is involved in it. */
1578 if (target->obd_recovering && !lw_client)
1579 lustre_msg_add_op_flags(req->rq_repmsg, MSG_CONNECT_RECOVERING);
1583 spin_lock(&export->exp_lock);
1584 export->exp_connecting = 0;
1585 spin_unlock(&export->exp_lock);
1587 class_export_put(export);
1589 if (target != NULL) {
1590 spin_lock(&target->obd_dev_lock);
1591 target->obd_conn_inprogress--;
1592 spin_unlock(&target->obd_dev_lock);
1593 class_decref(target, "find", current);
1596 ptlrpc_connection_put(pcon);
1597 req->rq_status = rc;
1601 int target_handle_disconnect(struct ptlrpc_request *req)
1607 rc = req_capsule_server_pack(&req->rq_pill);
1611 /* In case of target disconnect, updating sec ctx immediately is
1612 * required in order to record latest sequence number used.
1613 * Sequence is normally updated on export destroy, but this event
1614 * can occur too late, ie after a new target connect request has
1616 * Maintaining correct sequence when client connection becomes idle
1617 * ensures that GSS does not erroneously consider requests as replays.
1619 rc = sptlrpc_export_update_ctx(req->rq_export);
1623 /* Keep the rq_export around so we can send the reply. */
1624 req->rq_status = obd_disconnect(class_export_get(req->rq_export));
1629 void target_destroy_export(struct obd_export *exp)
1631 struct obd_import *imp = NULL;
1633 * exports created from last_rcvd data, and "fake"
1634 * exports created by lctl don't have an import
1636 spin_lock(&exp->exp_lock);
1637 if (exp->exp_imp_reverse != NULL) {
1638 imp = exp->exp_imp_reverse;
1639 exp->exp_imp_reverse = NULL;
1641 spin_unlock(&exp->exp_lock);
1643 client_destroy_import(imp);
1645 LASSERT_ATOMIC_ZERO(&exp->exp_locks_count);
1646 LASSERT_ATOMIC_ZERO(&exp->exp_rpc_count);
1647 LASSERT_ATOMIC_ZERO(&exp->exp_cb_count);
1648 LASSERT_ATOMIC_ZERO(&exp->exp_replay_count);
1650 EXPORT_SYMBOL(target_destroy_export);
1653 * Recovery functions
1655 static void target_request_copy_get(struct ptlrpc_request *req)
1657 class_export_rpc_inc(req->rq_export);
1658 LASSERT(list_empty(&req->rq_list));
1659 INIT_LIST_HEAD(&req->rq_replay_list);
1661 /* Increase refcount to keep request in queue. */
1662 atomic_inc(&req->rq_refcount);
1663 /* Let export know it has replays to be handled. */
1664 atomic_inc(&req->rq_export->exp_replay_count);
1667 static void target_request_copy_put(struct ptlrpc_request *req)
1669 LASSERT(list_empty(&req->rq_replay_list));
1670 LASSERT_ATOMIC_POS(&req->rq_export->exp_replay_count);
1672 atomic_dec(&req->rq_export->exp_replay_count);
1673 class_export_rpc_dec(req->rq_export);
1674 ptlrpc_server_drop_request(req);
1677 static int target_exp_enqueue_req_replay(struct ptlrpc_request *req)
1679 __u64 transno = lustre_msg_get_transno(req->rq_reqmsg);
1680 struct obd_export *exp = req->rq_export;
1681 struct ptlrpc_request *reqiter;
1682 struct ptlrpc_request *dup_req = NULL;
1687 spin_lock(&exp->exp_lock);
1688 list_for_each_entry(reqiter, &exp->exp_req_replay_queue,
1690 if (lustre_msg_get_transno(reqiter->rq_reqmsg) == transno) {
1698 /* We expect it with RESENT and REPLAY flags. */
1699 if ((lustre_msg_get_flags(req->rq_reqmsg) &
1700 (MSG_RESENT | MSG_REPLAY)) != (MSG_RESENT | MSG_REPLAY))
1701 CERROR("invalid flags %x of resent replay\n",
1702 lustre_msg_get_flags(req->rq_reqmsg));
1704 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY) {
1707 new_conn = lustre_msg_get_conn_cnt(req->rq_reqmsg);
1709 lustre_msg_get_conn_cnt(dup_req->rq_reqmsg))
1710 lustre_msg_set_conn_cnt(dup_req->rq_reqmsg,
1714 list_add_tail(&req->rq_replay_list,
1715 &exp->exp_req_replay_queue);
1718 spin_unlock(&exp->exp_lock);
1722 static void target_exp_dequeue_req_replay(struct ptlrpc_request *req)
1724 LASSERT(!list_empty(&req->rq_replay_list));
1725 LASSERT(req->rq_export);
1727 spin_lock(&req->rq_export->exp_lock);
1728 list_del_init(&req->rq_replay_list);
1729 spin_unlock(&req->rq_export->exp_lock);
1732 static void target_finish_recovery(struct lu_target *lut)
1734 struct obd_device *obd = lut->lut_obd;
1738 /* Only log a recovery message when recovery has occurred. */
1739 if (obd->obd_recovery_start) {
1740 time64_t now = ktime_get_seconds();
1741 time64_t elapsed_time;
1743 elapsed_time = max_t(time64_t, now - obd->obd_recovery_start,
1745 LCONSOLE_INFO("%s: Recovery over after %lld:%.02lld, of %d clients %d recovered and %d %s evicted.\n",
1746 obd->obd_name, elapsed_time / 60,
1748 atomic_read(&obd->obd_max_recoverable_clients),
1749 atomic_read(&obd->obd_connected_clients),
1750 obd->obd_stale_clients,
1751 obd->obd_stale_clients == 1 ? "was" : "were");
1752 if (obd->obd_stale_clients && do_dump_on_eviction(obd))
1753 libcfs_debug_dumplog();
1756 ldlm_reprocess_recovery_done(obd->obd_namespace);
1757 spin_lock(&obd->obd_recovery_task_lock);
1758 if (!list_empty(&obd->obd_req_replay_queue) ||
1759 !list_empty(&obd->obd_lock_replay_queue) ||
1760 !list_empty(&obd->obd_final_req_queue)) {
1761 CERROR("%s: Recovery queues ( %s%s%s) are not empty\n",
1763 list_empty(&obd->obd_req_replay_queue) ? "" : "req ",
1764 list_empty(&obd->obd_lock_replay_queue) ? \
1766 list_empty(&obd->obd_final_req_queue) ? \
1768 spin_unlock(&obd->obd_recovery_task_lock);
1771 spin_unlock(&obd->obd_recovery_task_lock);
1773 obd->obd_recovery_end = ktime_get_seconds();
1775 /* When recovery finished, cleanup orphans on MDS and OST. */
1776 if (obd->obd_type && OBP(obd, postrecov)) {
1777 int rc = OBP(obd, postrecov)(obd);
1780 LCONSOLE_WARN("%s: Post recovery failed, rc %d\n",
1786 static void abort_req_replay_queue(struct obd_device *obd)
1788 struct ptlrpc_request *req, *n;
1789 LIST_HEAD(abort_list);
1791 spin_lock(&obd->obd_recovery_task_lock);
1792 list_splice_init(&obd->obd_req_replay_queue, &abort_list);
1793 spin_unlock(&obd->obd_recovery_task_lock);
1794 list_for_each_entry_safe(req, n, &abort_list, rq_list) {
1795 DEBUG_REQ(D_WARNING, req, "aborted:");
1796 req->rq_status = -ENOTCONN;
1797 if (ptlrpc_error(req)) {
1798 DEBUG_REQ(D_ERROR, req,
1799 "failed abort_req_reply; skipping");
1801 target_exp_dequeue_req_replay(req);
1802 target_request_copy_put(req);
1806 static void abort_lock_replay_queue(struct obd_device *obd)
1808 struct ptlrpc_request *req, *n;
1809 LIST_HEAD(abort_list);
1811 spin_lock(&obd->obd_recovery_task_lock);
1812 list_splice_init(&obd->obd_lock_replay_queue, &abort_list);
1813 spin_unlock(&obd->obd_recovery_task_lock);
1814 list_for_each_entry_safe(req, n, &abort_list, rq_list) {
1815 DEBUG_REQ(D_ERROR, req, "aborted:");
1816 req->rq_status = -ENOTCONN;
1817 if (ptlrpc_error(req)) {
1818 DEBUG_REQ(D_ERROR, req,
1819 "failed abort_lock_reply; skipping");
1821 target_request_copy_put(req);
1826 * Called from a cleanup function if the device is being cleaned up
1827 * forcefully. The exports should all have been disconnected already,
1828 * the only thing left to do is
1829 * - clear the recovery flags
1830 * - cancel the timer
1831 * - free queued requests and replies, but don't send replies
1832 * Because the obd_stopping flag is set, no new requests should be received.
1834 void target_cleanup_recovery(struct obd_device *obd)
1836 struct ptlrpc_request *req, *n;
1837 LIST_HEAD(clean_list);
1839 spin_lock(&obd->obd_dev_lock);
1840 if (!obd->obd_recovering) {
1841 spin_unlock(&obd->obd_dev_lock);
1845 obd->obd_recovering = obd->obd_abort_recovery = 0;
1846 obd->obd_abort_recov_mdt = 0;
1847 spin_unlock(&obd->obd_dev_lock);
1849 spin_lock(&obd->obd_recovery_task_lock);
1850 target_cancel_recovery_timer(obd);
1851 list_splice_init(&obd->obd_req_replay_queue, &clean_list);
1852 spin_unlock(&obd->obd_recovery_task_lock);
1854 list_for_each_entry_safe(req, n, &clean_list, rq_list) {
1855 LASSERT(req->rq_reply_state == NULL);
1856 target_exp_dequeue_req_replay(req);
1857 target_request_copy_put(req);
1860 spin_lock(&obd->obd_recovery_task_lock);
1861 list_splice_init(&obd->obd_lock_replay_queue, &clean_list);
1862 list_splice_init(&obd->obd_final_req_queue, &clean_list);
1863 spin_unlock(&obd->obd_recovery_task_lock);
1865 list_for_each_entry_safe(req, n, &clean_list, rq_list) {
1866 LASSERT(req->rq_reply_state == NULL);
1867 target_request_copy_put(req);
1872 EXPORT_SYMBOL(target_cleanup_recovery);
1874 /* obd_recovery_task_lock should be held */
1875 void target_cancel_recovery_timer(struct obd_device *obd)
1877 CDEBUG(D_HA, "%s: cancel recovery timer\n", obd->obd_name);
1878 hrtimer_cancel(&obd->obd_recovery_timer);
1881 static void target_start_recovery_timer(struct obd_device *obd)
1885 if (obd->obd_recovery_start != 0)
1888 spin_lock(&obd->obd_dev_lock);
1889 if (!obd->obd_recovering || obd->obd_abort_recovery) {
1890 spin_unlock(&obd->obd_dev_lock);
1894 LASSERT(obd->obd_recovery_timeout != 0);
1896 if (obd->obd_recovery_start != 0) {
1897 spin_unlock(&obd->obd_dev_lock);
1901 obd->obd_recovery_start = ktime_get_seconds();
1902 delay = ktime_set(obd->obd_recovery_start +
1903 obd->obd_recovery_timeout, 0);
1904 hrtimer_start(&obd->obd_recovery_timer, delay, HRTIMER_MODE_ABS);
1905 spin_unlock(&obd->obd_dev_lock);
1907 LCONSOLE_WARN("%s: Will be in recovery for at least %u:%02u, or until %d client%s reconnect%s\n",
1909 obd->obd_recovery_timeout / 60,
1910 obd->obd_recovery_timeout % 60,
1911 atomic_read(&obd->obd_max_recoverable_clients),
1912 (atomic_read(&obd->obd_max_recoverable_clients) == 1) ?
1914 (atomic_read(&obd->obd_max_recoverable_clients) == 1) ?
1919 * extend recovery window.
1921 * if @extend is true, extend recovery window to have @dr_timeout remaining
1922 * at least; otherwise, make sure the recovery timeout value is not less
1925 static void extend_recovery_timer(struct obd_device *obd, timeout_t dr_timeout,
1932 spin_lock(&obd->obd_dev_lock);
1933 if (!obd->obd_recovering || obd->obd_abort_recovery ||
1934 obd->obd_stopping) {
1935 spin_unlock(&obd->obd_dev_lock);
1938 LASSERT(obd->obd_recovery_start != 0);
1940 left_ns = hrtimer_get_remaining(&obd->obd_recovery_timer);
1941 left = ktime_divns(left_ns, NSEC_PER_SEC);
1944 timeout = obd->obd_recovery_timeout;
1945 /* dr_timeout will happen after the hrtimer has expired.
1946 * Add the excess time to the soft recovery timeout without
1947 * exceeding the hard recovery timeout.
1949 if (dr_timeout > left) {
1950 timeout += dr_timeout - left;
1951 timeout = min_t(timeout_t, obd->obd_recovery_time_hard,
1955 timeout = clamp_t(timeout_t, dr_timeout,
1956 obd->obd_recovery_timeout,
1957 obd->obd_recovery_time_hard);
1960 if (timeout == obd->obd_recovery_time_hard)
1961 CWARN("%s: extended recovery timer reached hard limit: %d, extend: %d\n",
1962 obd->obd_name, timeout, extend);
1964 if (obd->obd_recovery_timeout < timeout) {
1967 obd->obd_recovery_timeout = timeout;
1968 end = ktime_set(obd->obd_recovery_start + timeout, 0);
1969 now = ktime_set(ktime_get_seconds(), 0);
1970 left_ns = ktime_sub(end, now);
1971 hrtimer_start(&obd->obd_recovery_timer, end, HRTIMER_MODE_ABS);
1972 left = ktime_divns(left_ns, NSEC_PER_SEC);
1974 spin_unlock(&obd->obd_dev_lock);
1976 CDEBUG(D_HA, "%s: recovery timer will expire in %d seconds\n",
1977 obd->obd_name, left);
1980 /* Reset the timer with each new client connection */
1982 * This timer is actually reconnect_timer, which is for making sure
1983 * the total recovery window is at least as big as my reconnect
1984 * attempt timing. So the initial recovery time_out will be set to
1985 * OBD_RECOVERY_FACTOR * obd_timeout. If the timeout coming
1986 * from client is bigger than this, then the recovery time_out will
1987 * be extended to make sure the client could be reconnected, in the
1988 * process, the timeout from the new client should be ignored.
1991 check_and_start_recovery_timer(struct obd_device *obd,
1992 struct ptlrpc_request *req,
1995 timeout_t service_timeout = lustre_msg_get_service_timeout(req->rq_reqmsg);
1996 struct obd_device_target *obt = &obd->u.obt;
1998 if (!new_client && service_timeout)
2000 * Teach server about old server's estimates, as first guess
2001 * at how long new requests will take.
2003 at_measured(&req->rq_rqbd->rqbd_svcpt->scp_at_estimate,
2006 target_start_recovery_timer(obd);
2009 * Convert the service time to RPC timeout,
2010 * and reuse service_timeout to limit stack usage.
2012 service_timeout = at_est2timeout(service_timeout);
2014 if (OBD_FAIL_CHECK(OBD_FAIL_TGT_SLUGGISH_NET) &&
2015 service_timeout < at_extra)
2016 service_timeout = at_extra;
2019 * We expect other clients to timeout within service_timeout, then try
2020 * to reconnect, then try the failover server. The max delay between
2021 * connect attempts is SWITCH_MAX + SWITCH_INC + INITIAL.
2023 service_timeout += 2 * INITIAL_CONNECT_TIMEOUT;
2025 LASSERT(obt->obt_magic == OBT_MAGIC);
2026 service_timeout += 2 * (CONNECTION_SWITCH_MAX + CONNECTION_SWITCH_INC);
2027 if (service_timeout > obd->obd_recovery_timeout && !new_client)
2028 extend_recovery_timer(obd, service_timeout, false);
2031 /** Health checking routines */
2032 static inline int exp_connect_healthy(struct obd_export *exp)
2034 return exp->exp_in_recovery;
2037 /** if export done req_replay or has replay in queue */
2038 static inline int exp_req_replay_healthy(struct obd_export *exp)
2040 return (!exp->exp_req_replay_needed ||
2041 atomic_read(&exp->exp_replay_count) > 0);
2045 static inline int exp_req_replay_healthy_or_from_mdt(struct obd_export *exp)
2047 return (exp_connect_flags(exp) & OBD_CONNECT_MDS_MDS) ||
2048 exp_req_replay_healthy(exp);
2051 /** if export done lock_replay or has replay in queue */
2052 static inline int exp_lock_replay_healthy(struct obd_export *exp)
2054 return (!exp->exp_lock_replay_needed ||
2055 atomic_read(&exp->exp_replay_count) > 0);
2058 static inline int exp_vbr_healthy(struct obd_export *exp)
2060 return !exp->exp_vbr_failed;
2063 static inline int exp_finished(struct obd_export *exp)
2065 return exp->exp_in_recovery && !exp->exp_lock_replay_needed;
2068 static inline int exp_finished_or_from_mdt(struct obd_export *exp)
2070 return (exp_connect_flags(exp) & OBD_CONNECT_MDS_MDS) ||
2074 static int check_for_next_transno(struct lu_target *lut)
2076 struct ptlrpc_request *req = NULL;
2077 struct obd_device *obd = lut->lut_obd;
2078 struct target_distribute_txn_data *tdtd = lut->lut_tdtd;
2079 int wake_up = 0, connected, completed, queue_len;
2080 __u64 req_transno = 0;
2081 __u64 update_transno = 0;
2082 __u64 next_transno = 0;
2086 spin_lock(&obd->obd_recovery_task_lock);
2087 if (!list_empty(&obd->obd_req_replay_queue)) {
2088 req = list_entry(obd->obd_req_replay_queue.next,
2089 struct ptlrpc_request, rq_list);
2090 req_transno = lustre_msg_get_transno(req->rq_reqmsg);
2093 if (!obd->obd_abort_recov_mdt && tdtd)
2094 update_transno = distribute_txn_get_next_transno(tdtd);
2096 connected = atomic_read(&obd->obd_connected_clients);
2097 completed = connected - atomic_read(&obd->obd_req_replay_clients);
2098 queue_len = obd->obd_requests_queued_for_recovery;
2099 next_transno = obd->obd_next_recovery_transno;
2102 "max: %d, connected: %d, completed: %d, queue_len: %d, req_transno: %llu, next_transno: %llu\n",
2103 atomic_read(&obd->obd_max_recoverable_clients),
2104 connected, completed,
2105 queue_len, req_transno, next_transno);
2107 if (obd->obd_abort_recovery) {
2108 CDEBUG(D_HA, "waking for aborted recovery\n");
2110 } else if (obd->obd_recovery_expired) {
2111 CDEBUG(D_HA, "waking for expired recovery\n");
2113 } else if (!obd->obd_abort_recov_mdt && tdtd && req &&
2114 is_req_replayed_by_update(req)) {
2115 LASSERTF(req_transno < next_transno,
2116 "req_transno %llu next_transno%llu\n", req_transno,
2118 CDEBUG(D_HA, "waking for duplicate req (%llu)\n",
2121 } else if (req_transno == next_transno ||
2122 (update_transno != 0 && update_transno <= next_transno)) {
2123 CDEBUG(D_HA, "waking for next (%lld)\n", next_transno);
2125 } else if (queue_len > 0 &&
2126 queue_len == atomic_read(&obd->obd_req_replay_clients)) {
2127 /** handle gaps occured due to lost reply or VBR */
2128 LASSERTF(req_transno >= next_transno,
2129 "req_transno: %llu, next_transno: %llu\n",
2130 req_transno, next_transno);
2132 "%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",
2133 obd->obd_name, obd->obd_version_recov ? "ON" : "OFF",
2134 next_transno, queue_len, completed, connected,
2135 req_transno, update_transno, obd->obd_last_committed);
2136 obd->obd_next_recovery_transno = req_transno;
2138 } else if (atomic_read(&obd->obd_req_replay_clients) == 0) {
2139 CDEBUG(D_HA, "waking for completed recovery\n");
2141 } else if (OBD_FAIL_CHECK(OBD_FAIL_MDS_RECOVERY_ACCEPTS_GAPS)) {
2143 "accepting transno gaps is explicitly allowed by fail_lock, waking up (%lld)\n",
2145 obd->obd_next_recovery_transno = req_transno;
2148 spin_unlock(&obd->obd_recovery_task_lock);
2152 static int check_for_next_lock(struct lu_target *lut)
2154 struct obd_device *obd = lut->lut_obd;
2157 spin_lock(&obd->obd_recovery_task_lock);
2158 if (!list_empty(&obd->obd_lock_replay_queue)) {
2159 CDEBUG(D_HA, "waking for next lock\n");
2161 } else if (atomic_read(&obd->obd_lock_replay_clients) == 0) {
2162 CDEBUG(D_HA, "waking for completed lock replay\n");
2164 } else if (obd->obd_abort_recovery) {
2165 CDEBUG(D_HA, "waking for aborted recovery\n");
2167 } else if (obd->obd_recovery_expired) {
2168 CDEBUG(D_HA, "waking for expired recovery\n");
2171 spin_unlock(&obd->obd_recovery_task_lock);
2176 static int check_update_llog(struct lu_target *lut)
2178 struct obd_device *obd = lut->lut_obd;
2179 struct target_distribute_txn_data *tdtd = lut->lut_tdtd;
2181 if (obd->obd_abort_recovery) {
2182 CDEBUG(D_HA, "waking for aborted recovery\n");
2186 if (atomic_read(&tdtd->tdtd_recovery_threads_count) == 0) {
2187 CDEBUG(D_HA, "waking for completion of reading update log\n");
2195 * wait for recovery events,
2196 * check its status with help of check_routine
2197 * evict dead clients via health_check
2199 static int target_recovery_overseer(struct lu_target *lut,
2200 int (*check_routine)(struct lu_target *),
2201 int (*health_check)(struct obd_export *))
2203 struct obd_device *obd = lut->lut_obd;
2204 struct target_distribute_txn_data *tdtd;
2208 if (obd->obd_recovering && obd->obd_recovery_start == 0) {
2209 now = ktime_get_seconds();
2210 if (now - last > 600) {
2211 LCONSOLE_INFO("%s: in recovery but waiting for the first client to connect\n",
2216 if (obd->obd_recovery_start != 0 && ktime_get_seconds() >=
2217 (obd->obd_recovery_start + obd->obd_recovery_time_hard)) {
2218 __u64 next_update_transno = 0;
2221 * Only abort the recovery if there are no update recovery
2224 spin_lock(&obd->obd_recovery_task_lock);
2225 if (!obd->obd_abort_recov_mdt && lut->lut_tdtd) {
2226 next_update_transno =
2227 distribute_txn_get_next_transno(lut->lut_tdtd);
2229 tdtd = lut->lut_tdtd;
2231 * If next_update_transno == 0, it probably because
2232 * updatelog retrieve threads did not get any records
2233 * yet, let's wait those threads stopped
2235 if (next_update_transno == 0) {
2236 spin_unlock(&obd->obd_recovery_task_lock);
2238 while (wait_event_timeout(
2239 tdtd->tdtd_recovery_threads_waitq,
2240 check_update_llog(lut),
2241 cfs_time_seconds(60)) == 0);
2243 spin_lock(&obd->obd_recovery_task_lock);
2244 next_update_transno =
2245 distribute_txn_get_next_transno(tdtd);
2249 if (next_update_transno != 0 && !obd->obd_abort_recovery) {
2250 obd->obd_next_recovery_transno = next_update_transno;
2251 spin_unlock(&obd->obd_recovery_task_lock);
2253 * Disconnect unfinished exports from clients, and
2254 * keep connection from MDT to make sure the update
2255 * recovery will still keep trying until some one
2256 * manually abort the recovery
2258 class_disconnect_stale_exports(obd,
2259 exp_finished_or_from_mdt);
2260 /* Abort all of replay & replay lock req from clients */
2261 abort_req_replay_queue(obd);
2262 abort_lock_replay_queue(obd);
2264 "%s: there are still update replay (%#llx)in the queue.\n",
2265 obd->obd_name, next_update_transno);
2267 obd->obd_abort_recovery = 1;
2268 spin_unlock(&obd->obd_recovery_task_lock);
2269 CWARN("%s recovery is aborted by hard timeout\n",
2274 while (wait_event_timeout(obd->obd_next_transno_waitq,
2276 cfs_time_seconds(60)) == 0)
2277 ; /* wait indefinitely for event, but don't trigger watchdog */
2279 if (obd->obd_abort_recovery) {
2280 CWARN("recovery is aborted, evict exports in recovery\n");
2281 if (lut->lut_tdtd != NULL) {
2282 tdtd = lut->lut_tdtd;
2284 * Let's wait all of the update log recovery thread
2288 tdtd->tdtd_recovery_threads_waitq,
2289 atomic_read(&tdtd->tdtd_recovery_threads_count)
2291 /* Then abort the update recovery list */
2292 dtrq_list_destroy(lut->lut_tdtd);
2295 /** evict exports which didn't finish recovery yet */
2296 class_disconnect_stale_exports(obd, exp_finished);
2298 } else if (obd->obd_recovery_expired) {
2299 obd->obd_recovery_expired = 0;
2301 /** If some clients died being recovered, evict them */
2302 LCONSOLE_WARN("%s: recovery is timed out, evict stale exports\n",
2304 /** evict cexports with no replay in queue, they are stalled */
2305 class_disconnect_stale_exports(obd, health_check);
2307 /** continue with VBR */
2308 spin_lock(&obd->obd_dev_lock);
2309 obd->obd_version_recov = 1;
2310 spin_unlock(&obd->obd_dev_lock);
2312 * reset timer, recovery will proceed with versions now,
2313 * timeout is set just to handle reconnection delays
2315 extend_recovery_timer(obd, RECONNECT_DELAY_MAX, true);
2317 * Wait for recovery events again, after evicting bad clients
2324 static struct ptlrpc_request *target_next_replay_lock(struct lu_target *lut)
2326 struct obd_device *obd = lut->lut_obd;
2327 struct ptlrpc_request *req = NULL;
2329 CDEBUG(D_HA, "Waiting for lock\n");
2330 if (target_recovery_overseer(lut, check_for_next_lock,
2331 exp_lock_replay_healthy))
2332 abort_lock_replay_queue(obd);
2334 spin_lock(&obd->obd_recovery_task_lock);
2335 if (!list_empty(&obd->obd_lock_replay_queue)) {
2336 req = list_entry(obd->obd_lock_replay_queue.next,
2337 struct ptlrpc_request, rq_list);
2338 list_del_init(&req->rq_list);
2339 spin_unlock(&obd->obd_recovery_task_lock);
2341 spin_unlock(&obd->obd_recovery_task_lock);
2342 LASSERT(list_empty(&obd->obd_lock_replay_queue));
2343 LASSERT(atomic_read(&obd->obd_lock_replay_clients) == 0);
2344 /** evict exports failed VBR */
2345 class_disconnect_stale_exports(obd, exp_vbr_healthy);
2350 static struct ptlrpc_request *target_next_final_ping(struct obd_device *obd)
2352 struct ptlrpc_request *req = NULL;
2354 spin_lock(&obd->obd_recovery_task_lock);
2355 if (!list_empty(&obd->obd_final_req_queue)) {
2356 req = list_entry(obd->obd_final_req_queue.next,
2357 struct ptlrpc_request, rq_list);
2358 list_del_init(&req->rq_list);
2359 spin_unlock(&obd->obd_recovery_task_lock);
2360 if (req->rq_export->exp_in_recovery) {
2361 spin_lock(&req->rq_export->exp_lock);
2362 req->rq_export->exp_in_recovery = 0;
2363 spin_unlock(&req->rq_export->exp_lock);
2366 spin_unlock(&obd->obd_recovery_task_lock);
2371 static void handle_recovery_req(struct ptlrpc_thread *thread,
2372 struct ptlrpc_request *req,
2373 svc_handler_t handler)
2378 * export can be evicted during recovery, no need to handle replays for
2379 * it after that, discard such request silently
2381 if (req->rq_export->exp_disconnected)
2384 req->rq_session.lc_thread = thread;
2385 req->rq_svc_thread = thread;
2386 req->rq_svc_thread->t_env->le_ses = &req->rq_session;
2388 /* thread context */
2389 lu_context_enter(&thread->t_env->le_ctx);
2391 lu_context_exit(&thread->t_env->le_ctx);
2393 req->rq_svc_thread->t_env->le_ses = NULL;
2395 /* don't reset timer for final stage */
2396 if (!exp_finished(req->rq_export)) {
2397 timeout_t timeout = obd_timeout;
2400 * Add request @timeout to the recovery time so next request from
2401 * this client may come in recovery time
2404 struct ptlrpc_service_part *svcpt;
2405 timeout_t est_timeout;
2407 svcpt = req->rq_rqbd->rqbd_svcpt;
2409 * If the server sent early reply for this request,
2410 * the client will recalculate the timeout according to
2411 * current server estimate service time, so we will
2412 * use the maxium timeout here for waiting the client
2413 * sending the next req
2415 est_timeout = at_get(&svcpt->scp_at_estimate);
2416 timeout = max_t(timeout_t, at_est2timeout(est_timeout),
2417 lustre_msg_get_timeout(req->rq_reqmsg));
2419 * Add 2 net_latency, one for balance rq_deadline
2420 * (see ptl_send_rpc), one for resend the req to server,
2421 * Note: client will pack net_latency in replay req
2422 * (see ptlrpc_replay_req)
2424 timeout += 2 * lustre_msg_get_service_timeout(req->rq_reqmsg);
2426 extend_recovery_timer(class_exp2obd(req->rq_export), timeout,
2432 /** Checking routines for recovery */
2433 static int check_for_recovery_ready(struct lu_target *lut)
2435 struct obd_device *obd = lut->lut_obd;
2436 unsigned int clnts = atomic_read(&obd->obd_connected_clients);
2439 "connected %d stale %d max_recoverable_clients %d abort %d expired %d\n",
2440 clnts, obd->obd_stale_clients,
2441 atomic_read(&obd->obd_max_recoverable_clients),
2442 obd->obd_abort_recovery, obd->obd_recovery_expired);
2444 if (!obd->obd_abort_recovery && !obd->obd_recovery_expired) {
2446 atomic_read(&obd->obd_max_recoverable_clients));
2447 if (clnts + obd->obd_stale_clients <
2448 atomic_read(&obd->obd_max_recoverable_clients))
2452 if (!obd->obd_abort_recov_mdt && lut->lut_tdtd != NULL) {
2453 if (!lut->lut_tdtd->tdtd_replay_ready &&
2454 !obd->obd_abort_recovery && !obd->obd_stopping) {
2456 * Let's extend recovery timer, in case the recovery
2457 * timer expired, and some clients got evicted
2459 extend_recovery_timer(obd, obd->obd_recovery_timeout,
2462 "%s update recovery is not ready, extend recovery %d\n",
2463 obd->obd_name, obd->obd_recovery_timeout);
2472 REQUEST_RECOVERY = 1,
2473 UPDATE_RECOVERY = 2,
2476 static __u64 get_next_replay_req_transno(struct obd_device *obd)
2480 if (!list_empty(&obd->obd_req_replay_queue)) {
2481 struct ptlrpc_request *req;
2483 req = list_entry(obd->obd_req_replay_queue.next,
2484 struct ptlrpc_request, rq_list);
2485 transno = lustre_msg_get_transno(req->rq_reqmsg);
2491 static __u64 get_next_transno(struct lu_target *lut, int *type)
2493 struct obd_device *obd = lut->lut_obd;
2494 struct target_distribute_txn_data *tdtd = lut->lut_tdtd;
2496 __u64 update_transno;
2500 transno = get_next_replay_req_transno(obd);
2502 *type = REQUEST_RECOVERY;
2504 if (!tdtd || obd->obd_abort_recov_mdt)
2507 update_transno = distribute_txn_get_next_transno(tdtd);
2508 if (transno == 0 || (transno >= update_transno &&
2509 update_transno != 0)) {
2510 transno = update_transno;
2512 *type = UPDATE_RECOVERY;
2519 * drop duplicate replay request
2521 * Because the operation has been replayed by update recovery, the request
2522 * with the same transno will be dropped and also notify the client to send
2523 * next replay request.
2525 * \param[in] env execution environment
2526 * \param[in] obd failover obd device
2527 * \param[in] req request to be dropped
2529 static void drop_duplicate_replay_req(struct lu_env *env,
2530 struct obd_device *obd,
2531 struct ptlrpc_request *req)
2533 DEBUG_REQ(D_HA, req,
2534 "remove t%lld from %s because duplicate update records found",
2535 lustre_msg_get_transno(req->rq_reqmsg),
2536 libcfs_nid2str(req->rq_peer.nid));
2539 * Right now, only for MDS reint operation update replay and
2540 * normal request replay can have the same transno
2542 if (lustre_msg_get_opc(req->rq_reqmsg) == MDS_REINT) {
2543 req_capsule_set(&req->rq_pill, &RQF_MDS_REINT);
2544 req->rq_status = req_capsule_server_pack(&req->rq_pill);
2545 if (likely(req->rq_export))
2546 target_committed_to_req(req);
2547 lustre_msg_set_transno(req->rq_repmsg, req->rq_transno);
2548 target_send_reply(req, req->rq_status, 0);
2550 DEBUG_REQ(D_ERROR, req, "wrong opc from %s",
2551 libcfs_nid2str(req->rq_peer.nid));
2553 target_exp_dequeue_req_replay(req);
2554 target_request_copy_put(req);
2555 obd->obd_replayed_requests++;
2558 #define WATCHDOG_TIMEOUT (obd_timeout * 10)
2560 static void replay_request_or_update(struct lu_env *env,
2561 struct lu_target *lut,
2562 struct target_recovery_data *trd,
2563 struct ptlrpc_thread *thread)
2565 struct obd_device *obd = lut->lut_obd;
2566 struct ptlrpc_request *req = NULL;
2572 CDEBUG(D_HA, "Waiting for transno %lld\n",
2573 obd->obd_next_recovery_transno);
2575 /* Replay all of request and update by transno */
2577 struct target_distribute_txn_data *tdtd = lut->lut_tdtd;
2579 CFS_FAIL_TIMEOUT(OBD_FAIL_TGT_REPLAY_DELAY2, cfs_fail_val);
2582 * It is needed to extend recovery window above
2583 * recovery_time_soft. Extending is possible only in the
2584 * end of recovery window (see more details in
2585 * handle_recovery_req()).
2587 CFS_FAIL_TIMEOUT_MS(OBD_FAIL_TGT_REPLAY_DELAY, 300);
2589 if (target_recovery_overseer(lut, check_for_next_transno,
2590 exp_req_replay_healthy_or_from_mdt)) {
2591 abort_req_replay_queue(obd);
2592 abort_lock_replay_queue(obd);
2596 spin_lock(&obd->obd_recovery_task_lock);
2597 transno = get_next_transno(lut, &type);
2598 if (type == REQUEST_RECOVERY && transno != 0) {
2600 * Drop replay request from client side, if the
2601 * replay has been executed by update with the
2604 req = list_entry(obd->obd_req_replay_queue.next,
2605 struct ptlrpc_request, rq_list);
2607 list_del_init(&req->rq_list);
2608 obd->obd_requests_queued_for_recovery--;
2609 spin_unlock(&obd->obd_recovery_task_lock);
2612 * Let's check if the request has been redone by
2615 if (is_req_replayed_by_update(req)) {
2616 struct distribute_txn_replay_req *dtrq;
2618 dtrq = distribute_txn_lookup_finish_list(tdtd,
2620 LASSERT(dtrq != NULL);
2621 spin_lock(&tdtd->tdtd_replay_list_lock);
2622 list_del_init(&dtrq->dtrq_list);
2623 spin_unlock(&tdtd->tdtd_replay_list_lock);
2626 drop_duplicate_replay_req(env, obd, req);
2631 LASSERT(trd->trd_processing_task == current->pid);
2632 DEBUG_REQ(D_HA, req, "processing x%llu t%lld from %s",
2634 lustre_msg_get_transno(req->rq_reqmsg),
2635 libcfs_nid2str(req->rq_peer.nid));
2637 ptlrpc_watchdog_init(&thread->t_watchdog,
2639 handle_recovery_req(thread, req,
2640 trd->trd_recovery_handler);
2641 ptlrpc_watchdog_disable(&thread->t_watchdog);
2644 * bz18031: increase next_recovery_transno before
2645 * target_request_copy_put() will drop exp_rpc reference
2647 spin_lock(&obd->obd_recovery_task_lock);
2648 obd->obd_next_recovery_transno++;
2649 spin_unlock(&obd->obd_recovery_task_lock);
2650 target_exp_dequeue_req_replay(req);
2651 target_request_copy_put(req);
2652 obd->obd_replayed_requests++;
2653 } else if (type == UPDATE_RECOVERY && transno != 0) {
2654 struct distribute_txn_replay_req *dtrq;
2657 spin_unlock(&obd->obd_recovery_task_lock);
2659 LASSERT(tdtd != NULL);
2660 dtrq = distribute_txn_get_next_req(tdtd);
2661 lu_context_enter(&thread->t_env->le_ctx);
2662 ptlrpc_watchdog_init(&thread->t_watchdog,
2664 rc = tdtd->tdtd_replay_handler(env, tdtd, dtrq);
2665 ptlrpc_watchdog_disable(&thread->t_watchdog);
2666 lu_context_exit(&thread->t_env->le_ctx);
2667 extend_recovery_timer(obd, obd_timeout, true);
2669 if (rc == 0 && dtrq->dtrq_xid != 0) {
2671 "Move x%llu t%llu to finish list\n",
2673 dtrq->dtrq_master_transno);
2675 /* Add it to the replay finish list */
2676 spin_lock(&tdtd->tdtd_replay_list_lock);
2677 list_add(&dtrq->dtrq_list,
2678 &tdtd->tdtd_replay_finish_list);
2679 spin_unlock(&tdtd->tdtd_replay_list_lock);
2681 spin_lock(&obd->obd_recovery_task_lock);
2682 if (transno == obd->obd_next_recovery_transno)
2683 obd->obd_next_recovery_transno++;
2685 obd->obd_next_recovery_transno)
2686 obd->obd_next_recovery_transno =
2688 spin_unlock(&obd->obd_recovery_task_lock);
2693 spin_unlock(&obd->obd_recovery_task_lock);
2695 LASSERT(list_empty(&obd->obd_req_replay_queue));
2696 LASSERT(atomic_read(&obd->obd_req_replay_clients) == 0);
2697 /** evict exports failed VBR */
2698 class_disconnect_stale_exports(obd, exp_vbr_healthy);
2704 static int target_recovery_thread(void *arg)
2706 struct lu_target *lut = arg;
2707 struct obd_device *obd = lut->lut_obd;
2708 struct ptlrpc_request *req;
2709 struct target_recovery_data *trd = &obd->obd_recovery_data;
2710 unsigned long delta;
2712 struct ptlrpc_thread *thread = NULL;
2716 OBD_ALLOC_PTR(thread);
2722 GOTO(out_thread, rc = -ENOMEM);
2723 rc = lu_env_add(env);
2727 rc = lu_context_init(&env->le_ctx, LCT_MD_THREAD | LCT_DT_THREAD);
2729 GOTO(out_env_remove, rc);
2731 thread->t_env = env;
2732 thread->t_id = -1; /* force filter_iobuf_get/put to use local buffers */
2733 thread->t_task = current;
2734 env->le_ctx.lc_thread = thread;
2735 tgt_io_thread_init(thread); /* init thread_big_cache for IO requests */
2737 CDEBUG(D_HA, "%s: started recovery thread pid %d\n", obd->obd_name,
2739 trd->trd_processing_task = current->pid;
2741 spin_lock(&obd->obd_dev_lock);
2742 obd->obd_recovering = 1;
2743 spin_unlock(&obd->obd_dev_lock);
2744 complete(&trd->trd_starting);
2746 /* first of all, we have to know the first transno to replay */
2747 if (target_recovery_overseer(lut, check_for_recovery_ready,
2748 exp_connect_healthy)) {
2749 abort_req_replay_queue(obd);
2750 abort_lock_replay_queue(obd);
2751 if (lut->lut_tdtd != NULL)
2752 dtrq_list_destroy(lut->lut_tdtd);
2755 /* next stage: replay requests or update */
2757 CDEBUG(D_INFO, "1: request replay stage - %d clients from t%llu\n",
2758 atomic_read(&obd->obd_req_replay_clients),
2759 obd->obd_next_recovery_transno);
2760 replay_request_or_update(env, lut, trd, thread);
2763 * The second stage: replay locks
2765 CDEBUG(D_INFO, "2: lock replay stage - %d clients\n",
2766 atomic_read(&obd->obd_lock_replay_clients));
2767 while ((req = target_next_replay_lock(lut))) {
2768 LASSERT(trd->trd_processing_task == current->pid);
2769 DEBUG_REQ(D_HA, req, "processing lock from %s:",
2770 libcfs_nid2str(req->rq_peer.nid));
2771 if (OBD_FAIL_CHECK(OBD_FAIL_LDLM_LOCK_REPLAY)) {
2772 req->rq_status = -ENODEV;
2773 target_request_copy_put(req);
2776 handle_recovery_req(thread, req,
2777 trd->trd_recovery_handler);
2778 target_request_copy_put(req);
2779 obd->obd_replayed_locks++;
2783 * The third stage: reply on final pings, at this moment all clients
2784 * must have request in final queue
2786 CFS_FAIL_TIMEOUT(OBD_FAIL_TGT_REPLAY_RECONNECT, cfs_fail_val);
2787 CDEBUG(D_INFO, "3: final stage - process recovery completion pings\n");
2788 /** Update server last boot epoch */
2789 tgt_boot_epoch_update(lut);
2791 * We drop recoverying flag to forward all new requests
2792 * to regular mds_handle() since now
2794 spin_lock(&obd->obd_dev_lock);
2795 obd->obd_recovering = obd->obd_abort_recovery = 0;
2796 obd->obd_abort_recov_mdt = 0;
2797 spin_unlock(&obd->obd_dev_lock);
2798 spin_lock(&obd->obd_recovery_task_lock);
2799 target_cancel_recovery_timer(obd);
2800 spin_unlock(&obd->obd_recovery_task_lock);
2801 while ((req = target_next_final_ping(obd))) {
2802 LASSERT(trd->trd_processing_task == current->pid);
2803 DEBUG_REQ(D_HA, req, "processing final ping from %s:",
2804 libcfs_nid2str(req->rq_peer.nid));
2805 handle_recovery_req(thread, req,
2806 trd->trd_recovery_handler);
2808 * Because the waiting client can not send ping to server,
2809 * so we need refresh the last_request_time, to avoid the
2810 * export is being evicted
2812 ptlrpc_update_export_timer(req->rq_export, 0);
2813 target_request_copy_put(req);
2816 delta = jiffies_to_msecs(jiffies - delta) / MSEC_PER_SEC;
2817 CDEBUG(D_INFO, "4: recovery completed in %lus - %d/%d reqs/locks\n",
2818 delta, obd->obd_replayed_requests, obd->obd_replayed_locks);
2819 if (delta > OBD_RECOVERY_TIME_SOFT) {
2820 CWARN("too long recovery - read logs\n");
2821 libcfs_debug_dumplog();
2824 target_finish_recovery(lut);
2825 lu_context_fini(&env->le_ctx);
2826 trd->trd_processing_task = 0;
2827 complete_all(&trd->trd_finishing);
2828 tgt_io_thread_done(thread);
2834 OBD_FREE_PTR(thread);
2838 static int target_start_recovery_thread(struct lu_target *lut,
2839 svc_handler_t handler)
2841 struct obd_device *obd = lut->lut_obd;
2843 struct target_recovery_data *trd = &obd->obd_recovery_data;
2846 memset(trd, 0, sizeof(*trd));
2847 init_completion(&trd->trd_starting);
2848 init_completion(&trd->trd_finishing);
2849 trd->trd_recovery_handler = handler;
2851 rc = server_name2index(obd->obd_name, &index, NULL);
2855 if (!IS_ERR(kthread_run(target_recovery_thread,
2856 lut, "tgt_recover_%d", index))) {
2857 wait_for_completion(&trd->trd_starting);
2858 LASSERT(obd->obd_recovering != 0);
2866 void target_stop_recovery_thread(struct obd_device *obd)
2868 if (obd->obd_recovery_data.trd_processing_task > 0) {
2869 struct target_recovery_data *trd = &obd->obd_recovery_data;
2870 /** recovery can be done but postrecovery is not yet */
2871 spin_lock(&obd->obd_dev_lock);
2872 if (obd->obd_recovering) {
2873 CERROR("%s: Aborting recovery\n", obd->obd_name);
2874 obd->obd_abort_recovery = 1;
2875 wake_up(&obd->obd_next_transno_waitq);
2877 spin_unlock(&obd->obd_dev_lock);
2878 wait_for_completion(&trd->trd_finishing);
2881 EXPORT_SYMBOL(target_stop_recovery_thread);
2883 void target_recovery_fini(struct obd_device *obd)
2885 class_disconnect_exports(obd);
2886 target_stop_recovery_thread(obd);
2887 target_cleanup_recovery(obd);
2889 EXPORT_SYMBOL(target_recovery_fini);
2891 static enum hrtimer_restart target_recovery_expired(struct hrtimer *timer)
2893 struct obd_device *obd = container_of(timer, struct obd_device,
2894 obd_recovery_timer);
2897 "%s: recovery timed out; %d clients are still in recovery after %llu seconds (%d clients connected)\n",
2898 obd->obd_name, atomic_read(&obd->obd_lock_replay_clients),
2899 ktime_get_seconds() - obd->obd_recovery_start,
2900 atomic_read(&obd->obd_connected_clients));
2902 obd->obd_recovery_expired = 1;
2903 wake_up(&obd->obd_next_transno_waitq);
2904 return HRTIMER_NORESTART;
2907 void target_recovery_init(struct lu_target *lut, svc_handler_t handler)
2909 struct obd_device *obd = lut->lut_obd;
2911 if (lut->lut_bottom->dd_rdonly)
2914 if (atomic_read(&obd->obd_max_recoverable_clients) == 0) {
2915 /** Update server last boot epoch */
2916 tgt_boot_epoch_update(lut);
2920 CDEBUG(D_HA, "RECOVERY: service %s, %d recoverable clients, "
2921 "last_transno %llu\n", obd->obd_name,
2922 atomic_read(&obd->obd_max_recoverable_clients),
2923 obd->obd_last_committed);
2924 LASSERT(obd->obd_stopping == 0);
2925 obd->obd_next_recovery_transno = obd->obd_last_committed + 1;
2926 obd->obd_recovery_start = 0;
2927 obd->obd_recovery_end = 0;
2929 hrtimer_init(&obd->obd_recovery_timer, CLOCK_MONOTONIC,
2931 obd->obd_recovery_timer.function = &target_recovery_expired;
2932 target_start_recovery_thread(lut, handler);
2934 EXPORT_SYMBOL(target_recovery_init);
2936 static int target_process_req_flags(struct obd_device *obd,
2937 struct ptlrpc_request *req)
2939 struct obd_export *exp = req->rq_export;
2941 LASSERT(exp != NULL);
2942 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REQ_REPLAY_DONE) {
2943 /* client declares he's ready to replay locks */
2944 spin_lock(&exp->exp_lock);
2945 if (exp->exp_req_replay_needed) {
2946 exp->exp_req_replay_needed = 0;
2947 spin_unlock(&exp->exp_lock);
2949 LASSERT_ATOMIC_POS(&obd->obd_req_replay_clients);
2950 atomic_dec(&obd->obd_req_replay_clients);
2952 spin_unlock(&exp->exp_lock);
2955 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_LOCK_REPLAY_DONE) {
2957 * client declares he's ready to complete recovery
2958 * so, we put the request on th final queue
2960 spin_lock(&exp->exp_lock);
2961 if (exp->exp_lock_replay_needed) {
2962 exp->exp_lock_replay_needed = 0;
2963 spin_unlock(&exp->exp_lock);
2965 LASSERT_ATOMIC_POS(&obd->obd_lock_replay_clients);
2966 atomic_dec(&obd->obd_lock_replay_clients);
2968 spin_unlock(&exp->exp_lock);
2974 int target_queue_recovery_request(struct ptlrpc_request *req,
2975 struct obd_device *obd)
2977 __u64 transno = lustre_msg_get_transno(req->rq_reqmsg);
2978 struct ptlrpc_request *reqiter;
2983 if (obd->obd_recovery_data.trd_processing_task == current->pid) {
2984 /* Processing the queue right now, don't re-add. */
2988 target_process_req_flags(obd, req);
2990 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_LOCK_REPLAY_DONE) {
2991 if (unlikely(OBD_FAIL_CHECK(OBD_FAIL_TGT_RECOVERY_REQ_RACE))) {
2992 if (cfs_fail_val == 1) {
2995 wake_up(&cfs_race_waitq);
2997 schedule_timeout_interruptible(
2998 cfs_time_seconds(1));
3003 * client declares he's ready to complete recovery
3004 * so, we put the request on th final queue
3006 target_request_copy_get(req);
3007 DEBUG_REQ(D_HA, req, "queue final req");
3008 wake_up(&obd->obd_next_transno_waitq);
3009 spin_lock(&obd->obd_recovery_task_lock);
3010 if (obd->obd_recovering) {
3011 struct ptlrpc_request *tmp;
3012 struct ptlrpc_request *duplicate = NULL;
3014 if (likely(!req->rq_export->exp_replay_done)) {
3015 req->rq_export->exp_replay_done = 1;
3016 list_add_tail(&req->rq_list,
3017 &obd->obd_final_req_queue);
3018 spin_unlock(&obd->obd_recovery_task_lock);
3023 * XXX O(n), but only happens if final ping is
3024 * timed out, probably reorganize the list as
3027 list_for_each_entry_safe(reqiter, tmp,
3028 &obd->obd_final_req_queue,
3030 if (reqiter->rq_export == req->rq_export) {
3031 list_del_init(&reqiter->rq_list);
3032 duplicate = reqiter;
3037 list_add_tail(&req->rq_list,
3038 &obd->obd_final_req_queue);
3039 req->rq_export->exp_replay_done = 1;
3040 spin_unlock(&obd->obd_recovery_task_lock);
3042 if (duplicate != NULL) {
3043 DEBUG_REQ(D_HA, duplicate,
3044 "put prev final req");
3045 target_request_copy_put(duplicate);
3049 spin_unlock(&obd->obd_recovery_task_lock);
3050 target_request_copy_put(req);
3051 RETURN(obd->obd_stopping ? -ENOTCONN : 1);
3054 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REQ_REPLAY_DONE) {
3055 /* client declares he's ready to replay locks */
3056 target_request_copy_get(req);
3057 DEBUG_REQ(D_HA, req, "queue lock replay req");
3058 wake_up(&obd->obd_next_transno_waitq);
3059 spin_lock(&obd->obd_recovery_task_lock);
3060 LASSERT(obd->obd_recovering);
3061 /* usually due to recovery abort */
3062 if (!req->rq_export->exp_in_recovery) {
3063 spin_unlock(&obd->obd_recovery_task_lock);
3064 target_request_copy_put(req);
3067 LASSERT(req->rq_export->exp_lock_replay_needed);
3068 list_add_tail(&req->rq_list, &obd->obd_lock_replay_queue);
3069 spin_unlock(&obd->obd_recovery_task_lock);
3074 * CAVEAT EMPTOR: The incoming request message has been swabbed
3075 * (i.e. buflens etc are in my own byte order), but type-dependent
3076 * buffers (eg mdt_body, ost_body etc) have NOT been swabbed.
3080 INIT_LIST_HEAD(&req->rq_list);
3081 DEBUG_REQ(D_HA, req, "not queueing");
3086 * If we're processing the queue, we want don't want to queue this
3089 * Also, if this request has a transno less than the one we're waiting
3090 * for, we should process it now. It could (and currently always will)
3091 * be an open request for a descriptor that was opened some time ago.
3093 * Also, a resent, replayed request that has already been
3094 * handled will pass through here and be processed immediately.
3097 "Next recovery transno: %llu, current: %llu, replaying\n",
3098 obd->obd_next_recovery_transno, transno);
3101 * If the request has been replayed by update replay, then sends this
3102 * request to the recovery thread (replay_request_or_update()), where
3103 * it will be handled
3105 spin_lock(&obd->obd_recovery_task_lock);
3106 if (transno < obd->obd_next_recovery_transno &&
3107 !is_req_replayed_by_update(req)) {
3108 /* Processing the queue right now, don't re-add. */
3109 LASSERT(list_empty(&req->rq_list));
3110 spin_unlock(&obd->obd_recovery_task_lock);
3113 spin_unlock(&obd->obd_recovery_task_lock);
3115 if (OBD_FAIL_CHECK(OBD_FAIL_TGT_REPLAY_DROP))
3118 target_request_copy_get(req);
3119 if (!req->rq_export->exp_in_recovery) {
3120 target_request_copy_put(req);
3123 LASSERT(req->rq_export->exp_req_replay_needed);
3125 if (target_exp_enqueue_req_replay(req)) {
3126 DEBUG_REQ(D_ERROR, req, "dropping resent queued req");
3127 target_request_copy_put(req);
3132 spin_lock(&obd->obd_recovery_task_lock);
3133 LASSERT(obd->obd_recovering);
3134 list_for_each_entry(reqiter, &obd->obd_req_replay_queue, rq_list) {
3135 if (lustre_msg_get_transno(reqiter->rq_reqmsg) > transno) {
3136 list_add_tail(&req->rq_list, &reqiter->rq_list);
3141 if (unlikely(lustre_msg_get_transno(reqiter->rq_reqmsg) ==
3143 DEBUG_REQ(D_ERROR, req,
3144 "dropping replay: transno has been claimed by another client");
3145 spin_unlock(&obd->obd_recovery_task_lock);
3146 target_exp_dequeue_req_replay(req);
3147 target_request_copy_put(req);
3153 list_add_tail(&req->rq_list, &obd->obd_req_replay_queue);
3155 obd->obd_requests_queued_for_recovery++;
3156 spin_unlock(&obd->obd_recovery_task_lock);
3157 wake_up(&obd->obd_next_transno_waitq);
3161 void target_committed_to_req(struct ptlrpc_request *req)
3163 struct obd_export *exp = req->rq_export;
3165 if (!exp->exp_obd->obd_no_transno && req->rq_repmsg != NULL)
3166 lustre_msg_set_last_committed(req->rq_repmsg,
3167 exp->exp_last_committed);
3169 DEBUG_REQ(D_IOCTL, req,
3170 "not sending last_committed update (%d/%d)",
3171 exp->exp_obd->obd_no_transno,
3172 req->rq_repmsg == NULL);
3174 CDEBUG(D_INFO, "last_committed %llu, transno %llu, xid %llu\n",
3175 exp->exp_last_committed, req->rq_transno, req->rq_xid);
3178 #endif /* HAVE_SERVER_SUPPORT */
3181 * Packs current SLV and Limit into \a req.
3183 int target_pack_pool_reply(struct ptlrpc_request *req)
3185 struct obd_device *obd;
3190 * Check that we still have all structures alive as this may
3191 * be some late RPC at shutdown time.
3193 if (unlikely(!req->rq_export || !req->rq_export->exp_obd ||
3194 !exp_connect_lru_resize(req->rq_export))) {
3195 lustre_msg_set_slv(req->rq_repmsg, 0);
3196 lustre_msg_set_limit(req->rq_repmsg, 0);
3200 /* OBD is alive here as export is alive, which we checked above. */
3201 obd = req->rq_export->exp_obd;
3203 read_lock(&obd->obd_pool_lock);
3204 lustre_msg_set_slv(req->rq_repmsg, obd->obd_pool_slv);
3205 lustre_msg_set_limit(req->rq_repmsg, obd->obd_pool_limit);
3206 read_unlock(&obd->obd_pool_lock);
3211 static int target_send_reply_msg(struct ptlrpc_request *req,
3212 int rc, int fail_id)
3214 if (OBD_FAIL_CHECK_ORSET(fail_id & ~OBD_FAIL_ONCE, OBD_FAIL_ONCE)) {
3215 DEBUG_REQ(D_ERROR, req, "dropping reply");
3219 * We can have a null rq_reqmsg in the event of bad signature or
3220 * no context when unwrapping
3222 if (req->rq_reqmsg &&
3223 unlikely(lustre_msg_get_opc(req->rq_reqmsg) == MDS_REINT &&
3224 OBD_FAIL_CHECK(OBD_FAIL_MDS_REINT_MULTI_NET_REP)))
3228 DEBUG_REQ(D_NET, req, "processing error (%d)", rc);
3229 req->rq_status = rc;
3230 return ptlrpc_send_error(req, 1);
3232 DEBUG_REQ(D_NET, req, "sending reply");
3234 return ptlrpc_send_reply(req, PTLRPC_REPLY_MAYBE_DIFFICULT);
3237 void target_send_reply(struct ptlrpc_request *req, int rc, int fail_id)
3239 struct ptlrpc_service_part *svcpt;
3241 struct ptlrpc_reply_state *rs;
3242 struct obd_export *exp;
3246 if (req->rq_no_reply) {
3251 svcpt = req->rq_rqbd->rqbd_svcpt;
3252 rs = req->rq_reply_state;
3253 if (rs == NULL || !rs->rs_difficult) {
3255 target_send_reply_msg(req, rc, fail_id);
3260 /* must be an export if locks saved */
3261 LASSERT(req->rq_export != NULL);
3262 /* req/reply consistent */
3263 LASSERT(rs->rs_svcpt == svcpt);
3266 LASSERT(!rs->rs_scheduled);
3267 LASSERT(!rs->rs_scheduled_ever);
3268 LASSERT(!rs->rs_handled);
3269 LASSERT(!rs->rs_on_net);
3270 LASSERT(rs->rs_export == NULL);
3271 LASSERT(list_empty(&rs->rs_obd_list));
3272 LASSERT(list_empty(&rs->rs_exp_list));
3274 exp = class_export_get(req->rq_export);
3276 /* disable reply scheduling while I'm setting up */
3277 rs->rs_scheduled = 1;
3279 rs->rs_xid = req->rq_xid;
3280 rs->rs_transno = req->rq_transno;
3281 rs->rs_export = exp;
3282 rs->rs_opc = lustre_msg_get_opc(req->rq_reqmsg);
3284 spin_lock(&exp->exp_uncommitted_replies_lock);
3285 CDEBUG(D_NET, "rs transno = %llu, last committed = %llu\n",
3286 rs->rs_transno, exp->exp_last_committed);
3287 if (rs->rs_transno > exp->exp_last_committed) {
3288 /* not committed already */
3289 list_add_tail(&rs->rs_obd_list,
3290 &exp->exp_uncommitted_replies);
3292 spin_unlock(&exp->exp_uncommitted_replies_lock);
3294 spin_lock(&exp->exp_lock);
3295 list_add_tail(&rs->rs_exp_list, &exp->exp_outstanding_replies);
3296 spin_unlock(&exp->exp_lock);
3298 netrc = target_send_reply_msg(req, rc, fail_id);
3300 spin_lock(&svcpt->scp_rep_lock);
3302 atomic_inc(&svcpt->scp_nreps_difficult);
3306 * error sending: reply is off the net. Also we need +1
3307 * reply ref until ptlrpc_handle_rs() is done
3308 * with the reply state (if the send was successful, there
3309 * would have been +1 ref for the net, which
3310 * reply_out_callback leaves alone)
3313 ptlrpc_rs_addref(rs);
3316 spin_lock(&rs->rs_lock);
3317 if (rs->rs_transno <= exp->exp_last_committed ||
3318 (!rs->rs_on_net && !rs->rs_no_ack) ||
3319 list_empty(&rs->rs_exp_list) || /* completed already */
3320 list_empty(&rs->rs_obd_list)) {
3321 CDEBUG(D_HA, "Schedule reply immediately\n");
3322 ptlrpc_dispatch_difficult_reply(rs);
3324 list_add(&rs->rs_list, &svcpt->scp_rep_active);
3325 rs->rs_scheduled = 0; /* allow notifier to schedule */
3327 spin_unlock(&rs->rs_lock);
3328 spin_unlock(&svcpt->scp_rep_lock);
3332 enum ldlm_mode lck_compat_array[] = {
3333 [LCK_EX] = LCK_COMPAT_EX,
3334 [LCK_PW] = LCK_COMPAT_PW,
3335 [LCK_PR] = LCK_COMPAT_PR,
3336 [LCK_CW] = LCK_COMPAT_CW,
3337 [LCK_CR] = LCK_COMPAT_CR,
3338 [LCK_NL] = LCK_COMPAT_NL,
3339 [LCK_GROUP] = LCK_COMPAT_GROUP,
3340 [LCK_COS] = LCK_COMPAT_COS,
3344 * Rather arbitrary mapping from LDLM error codes to errno values. This should
3345 * not escape to the user level.
3347 int ldlm_error2errno(enum ldlm_error error)
3353 case ELDLM_LOCK_MATCHED:
3356 case ELDLM_LOCK_CHANGED:
3359 case ELDLM_LOCK_ABORTED:
3362 case ELDLM_LOCK_REPLACED:
3365 case ELDLM_NO_LOCK_DATA:
3368 case ELDLM_NAMESPACE_EXISTS:
3371 case ELDLM_BAD_NAMESPACE:
3375 if (((int)error) < 0) { /* cast to signed type */
3376 result = error; /* as ldlm_error can be unsigned */
3378 CERROR("Invalid DLM result code: %d\n", error);
3384 EXPORT_SYMBOL(ldlm_error2errno);
3387 * Dual to ldlm_error2errno(): maps errno values back to enum ldlm_error.
3389 enum ldlm_error ldlm_errno2error(int err_no)
3398 error = ELDLM_LOCK_CHANGED;
3401 error = ELDLM_LOCK_ABORTED;
3404 error = ELDLM_LOCK_REPLACED;
3407 error = ELDLM_NO_LOCK_DATA;
3410 error = ELDLM_NAMESPACE_EXISTS;
3413 error = ELDLM_BAD_NAMESPACE;
3421 #if LUSTRE_TRACKS_LOCK_EXP_REFS
3422 void ldlm_dump_export_locks(struct obd_export *exp)
3424 spin_lock(&exp->exp_locks_list_guard);
3425 if (!list_empty(&exp->exp_locks_list)) {
3426 struct ldlm_lock *lock;
3428 CERROR("dumping locks for export %p, ignore if the unmount doesn't hang\n",
3430 list_for_each_entry(lock, &exp->exp_locks_list,
3432 LDLM_ERROR(lock, "lock:");
3434 spin_unlock(&exp->exp_locks_list_guard);
3438 #ifdef HAVE_SERVER_SUPPORT
3439 static inline const char *bulk2type(struct ptlrpc_request *req)
3441 if (req->rq_bulk_read)
3443 if (req->rq_bulk_write)
3448 int target_bulk_io(struct obd_export *exp, struct ptlrpc_bulk_desc *desc)
3450 struct ptlrpc_request *req = desc->bd_req;
3451 time64_t start = ktime_get_seconds();
3457 /* If there is eviction in progress, wait for it to finish. */
3459 exp->exp_obd->obd_evict_inprogress_waitq,
3460 !atomic_read(&exp->exp_obd->obd_evict_inprogress));
3462 /* Check if client was evicted or reconnected already. */
3463 if (exp->exp_failed ||
3464 exp->exp_conn_cnt > lustre_msg_get_conn_cnt(req->rq_reqmsg)) {
3467 if (req->rq_bulk_read)
3468 rc = sptlrpc_svc_wrap_bulk(req, desc);
3470 if (OCD_HAS_FLAG(&exp->exp_connect_data, BULK_MBITS))
3471 req->rq_mbits = lustre_msg_get_mbits(req->rq_reqmsg);
3472 else /* old version, bulk matchbits is rq_xid */
3473 req->rq_mbits = req->rq_xid;
3476 rc = ptlrpc_start_bulk_transfer(desc);
3480 DEBUG_REQ(D_ERROR, req, "bulk %s failed: rc = %d",
3481 bulk2type(req), rc);
3485 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_SENDPAGE)) {
3486 ptlrpc_abort_bulk(desc);
3490 /* limit actual bulk transfer to bulk_timeout seconds */
3491 deadline = start + bulk_timeout;
3492 if (deadline > req->rq_deadline)
3493 deadline = req->rq_deadline;
3496 time64_t timeoutl = deadline - ktime_get_seconds();
3497 time64_t rq_deadline;
3499 while (timeoutl >= 0 &&
3500 wait_event_idle_timeout(
3502 !ptlrpc_server_bulk_active(desc) ||
3505 lustre_msg_get_conn_cnt(req->rq_reqmsg),
3506 timeoutl ? cfs_time_seconds(1) : 1) == 0)
3508 rc = timeoutl < 0 ? -ETIMEDOUT : 0;
3510 /* Wait again if we changed rq_deadline. */
3511 rq_deadline = READ_ONCE(req->rq_deadline);
3512 deadline = start + bulk_timeout;
3513 if (deadline > rq_deadline)
3514 deadline = rq_deadline;
3515 } while (rc == -ETIMEDOUT &&
3516 deadline > ktime_get_seconds());
3518 if (rc == -ETIMEDOUT) {
3519 DEBUG_REQ(D_ERROR, req, "timeout on bulk %s after %lld%+llds",
3520 bulk2type(req), deadline - start,
3521 ktime_get_real_seconds() - deadline);
3522 ptlrpc_abort_bulk(desc);
3523 } else if (exp->exp_failed) {
3524 DEBUG_REQ(D_ERROR, req, "Eviction on bulk %s",
3527 ptlrpc_abort_bulk(desc);
3528 } else if (exp->exp_conn_cnt >
3529 lustre_msg_get_conn_cnt(req->rq_reqmsg)) {
3530 DEBUG_REQ(D_ERROR, req, "Reconnect on bulk %s",
3532 /* We don't reply anyway. */
3534 ptlrpc_abort_bulk(desc);
3535 } else if (desc->bd_failure) {
3536 DEBUG_REQ(D_ERROR, req, "network error on bulk %s",
3538 /* XXX should this be a different errno? */
3541 if (req->rq_bulk_write)
3542 rc = sptlrpc_svc_unwrap_bulk(req, desc);
3543 if (rc == 0 && desc->bd_nob_transferred != desc->bd_nob) {
3544 DEBUG_REQ(D_ERROR, req, "truncated bulk %s %d(%d)",
3545 bulk2type(req), desc->bd_nob_transferred,
3547 /* XXX should this be a different errno? */
3554 EXPORT_SYMBOL(target_bulk_io);
3556 #endif /* HAVE_SERVER_SUPPORT */