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/fs_struct.h>
43 #include <linux/jiffies.h>
44 #include <linux/kernel.h>
45 #include <linux/kthread.h>
46 #include <libcfs/libcfs.h>
48 #include <obd_class.h>
49 #include <lustre_dlm.h>
50 #include <lustre_net.h>
51 #include <lustre_sec.h>
52 #include "ldlm_internal.h"
55 * @priority: If non-zero, move the selected connection to the list head.
56 * @create: If zero, only search in existing connections.
58 static int import_set_conn(struct obd_import *imp, struct obd_uuid *uuid,
59 int priority, int create)
61 struct ptlrpc_connection *ptlrpc_conn;
62 struct obd_import_conn *imp_conn = NULL, *item;
63 u32 refnet = LNET_NET_ANY;
68 if (!create && !priority) {
69 CDEBUG(D_HA, "Nothing to do\n");
73 if (imp->imp_connection &&
74 imp->imp_connection->c_remote_uuid.uuid[0] == 0)
75 /* refnet is used to restrict network connections */
76 refnet = LNET_NID_NET(&imp->imp_connection->c_self);
77 ptlrpc_conn = ptlrpc_uuid_to_connection(uuid, refnet);
79 CDEBUG(D_HA, "can't find connection %s\n", uuid->uuid);
84 OBD_ALLOC(imp_conn, sizeof(*imp_conn));
86 GOTO(out_put, rc = -ENOMEM);
89 spin_lock(&imp->imp_lock);
90 list_for_each_entry(item, &imp->imp_conn_list, oic_item) {
91 if (obd_uuid_equals(uuid, &item->oic_uuid)) {
93 list_move(&item->oic_item,
95 item->oic_last_attempt = 0;
97 CDEBUG(D_HA, "imp %p@%s: found existing conn %s%s\n",
98 imp, imp->imp_obd->obd_name, uuid->uuid,
99 (priority ? ", moved to head" : ""));
100 spin_unlock(&imp->imp_lock);
101 GOTO(out_free, rc = 0);
104 /* No existing import connection found for \a uuid. */
106 imp_conn->oic_conn = ptlrpc_conn;
107 imp_conn->oic_uuid = *uuid;
108 imp_conn->oic_last_attempt = 0;
110 list_add(&imp_conn->oic_item, &imp->imp_conn_list);
112 list_add_tail(&imp_conn->oic_item,
113 &imp->imp_conn_list);
114 CDEBUG(D_HA, "imp %p@%s: add connection %s at %s\n",
115 imp, imp->imp_obd->obd_name, uuid->uuid,
116 (priority ? "head" : "tail"));
118 spin_unlock(&imp->imp_lock);
119 GOTO(out_free, rc = -ENOENT);
122 spin_unlock(&imp->imp_lock);
126 OBD_FREE(imp_conn, sizeof(*imp_conn));
128 ptlrpc_connection_put(ptlrpc_conn);
132 int import_set_conn_priority(struct obd_import *imp, struct obd_uuid *uuid)
134 return import_set_conn(imp, uuid, 1, 0);
137 int client_import_add_conn(struct obd_import *imp, struct obd_uuid *uuid,
140 return import_set_conn(imp, uuid, priority, 1);
142 EXPORT_SYMBOL(client_import_add_conn);
144 int client_import_dyn_add_conn(struct obd_import *imp, struct obd_uuid *uuid,
145 lnet_nid_t prim_nid, int priority)
147 struct ptlrpc_connection *ptlrpc_conn;
150 ptlrpc_conn = ptlrpc_uuid_to_connection(uuid, LNET_NIDNET(prim_nid));
152 const char *str_uuid = obd_uuid2str(uuid);
154 rc = class_add_uuid(str_uuid, prim_nid);
156 CERROR("%s: failed to add UUID '%s': rc = %d\n",
157 imp->imp_obd->obd_name, str_uuid, rc);
161 return import_set_conn(imp, uuid, priority, 1);
163 EXPORT_SYMBOL(client_import_dyn_add_conn);
165 int client_import_add_nids_to_conn(struct obd_import *imp, lnet_nid_t *nids,
166 int nid_count, struct obd_uuid *uuid)
168 struct obd_import_conn *conn;
172 if (nid_count <= 0 || !nids)
175 spin_lock(&imp->imp_lock);
176 list_for_each_entry(conn, &imp->imp_conn_list, oic_item) {
177 if (class_check_uuid(&conn->oic_uuid, nids[0])) {
178 *uuid = conn->oic_uuid;
179 spin_unlock(&imp->imp_lock);
180 rc = class_add_nids_to_uuid(&conn->oic_uuid, nids,
185 spin_unlock(&imp->imp_lock);
188 EXPORT_SYMBOL(client_import_add_nids_to_conn);
190 int client_import_del_conn(struct obd_import *imp, struct obd_uuid *uuid)
192 struct obd_import_conn *imp_conn;
193 struct obd_export *dlmexp;
198 spin_lock(&imp->imp_lock);
199 if (list_empty(&imp->imp_conn_list)) {
200 LASSERT(!imp->imp_connection);
204 list_for_each_entry(imp_conn, &imp->imp_conn_list, oic_item) {
205 if (!obd_uuid_equals(uuid, &imp_conn->oic_uuid))
207 LASSERT(imp_conn->oic_conn);
209 if (imp_conn == imp->imp_conn_current) {
210 LASSERT(imp_conn->oic_conn == imp->imp_connection);
212 if (imp->imp_state != LUSTRE_IMP_CLOSED &&
213 imp->imp_state != LUSTRE_IMP_DISCON) {
214 CERROR("can't remove current connection\n");
215 GOTO(out, rc = -EBUSY);
218 ptlrpc_connection_put(imp->imp_connection);
219 imp->imp_connection = NULL;
221 dlmexp = class_conn2export(&imp->imp_dlm_handle);
222 if (dlmexp && dlmexp->exp_connection) {
223 LASSERT(dlmexp->exp_connection ==
225 ptlrpc_connection_put(dlmexp->exp_connection);
226 dlmexp->exp_connection = NULL;
230 class_export_put(dlmexp);
233 list_del(&imp_conn->oic_item);
234 ptlrpc_connection_put(imp_conn->oic_conn);
235 OBD_FREE(imp_conn, sizeof(*imp_conn));
236 CDEBUG(D_HA, "imp %p@%s: remove connection %s\n",
237 imp, imp->imp_obd->obd_name, uuid->uuid);
242 spin_unlock(&imp->imp_lock);
244 CERROR("connection %s not found\n", uuid->uuid);
247 EXPORT_SYMBOL(client_import_del_conn);
250 * Find conn UUID by peer NID. \a peer is a server NID. This function is used
251 * to find a conn uuid of \a imp which can reach \a peer.
253 int client_import_find_conn(struct obd_import *imp, lnet_nid_t peer,
254 struct obd_uuid *uuid)
256 struct obd_import_conn *conn;
261 spin_lock(&imp->imp_lock);
262 list_for_each_entry(conn, &imp->imp_conn_list, oic_item) {
263 /* Check if conn UUID does have this peer NID. */
264 if (class_check_uuid(&conn->oic_uuid, peer)) {
265 *uuid = conn->oic_uuid;
270 spin_unlock(&imp->imp_lock);
273 EXPORT_SYMBOL(client_import_find_conn);
275 void client_destroy_import(struct obd_import *imp)
278 * Drop security policy instance after all RPCs have finished/aborted
279 * to let all busy contexts be released.
281 class_import_get(imp);
282 class_destroy_import(imp);
283 sptlrpc_import_sec_put(imp);
284 class_import_put(imp);
286 EXPORT_SYMBOL(client_destroy_import);
289 * Check whether or not the OSC is on MDT.
292 * setup 0:{fsname}-OSTxxxx-osc[-MDTxxxx] 1:lustre-OST0000_UUID 2:NID
294 * setup 0:{fsname}-OSTxxxx-osc 1:lustre-OST0000_UUID 2:NID
297 static int osc_on_mdt(char *obdname)
301 ptr = strrchr(obdname, '-');
305 if (strncmp(ptr + 1, "MDT", 3) == 0)
312 * Configure an RPC client OBD device.
317 * 3 - inactive-on-startup
318 * 4 - restrictive net
320 int client_obd_setup(struct obd_device *obd, struct lustre_cfg *lcfg)
322 struct client_obd *cli = &obd->u.cli;
323 struct obd_import *imp;
324 struct obd_uuid server_uuid;
325 int rq_portal, rp_portal, connect_op;
326 const char *name = obd->obd_type->typ_name;
327 enum ldlm_ns_type ns_type = LDLM_NS_TYPE_UNKNOWN;
328 char *cli_name = lustre_cfg_buf(lcfg, 0);
329 struct ptlrpc_connection fake_conn = { .c_self = {},
330 .c_remote_uuid.uuid[0] = 0 };
336 * In a more perfect world, we would hang a ptlrpc_client off of
337 * obd_type and just use the values from there.
339 if (!strcmp(name, LUSTRE_OSC_NAME)) {
340 rq_portal = OST_REQUEST_PORTAL;
341 rp_portal = OSC_REPLY_PORTAL;
342 connect_op = OST_CONNECT;
343 cli->cl_sp_me = LUSTRE_SP_CLI;
344 cli->cl_sp_to = LUSTRE_SP_OST;
345 ns_type = LDLM_NS_TYPE_OSC;
346 } else if (!strcmp(name, LUSTRE_MDC_NAME) ||
347 !strcmp(name, LUSTRE_LWP_NAME)) {
348 rq_portal = MDS_REQUEST_PORTAL;
349 rp_portal = MDC_REPLY_PORTAL;
350 connect_op = MDS_CONNECT;
351 if (is_lwp_on_ost(cli_name))
352 cli->cl_sp_me = LUSTRE_SP_OST;
353 else if (is_lwp_on_mdt(cli_name))
354 cli->cl_sp_me = LUSTRE_SP_MDT;
356 cli->cl_sp_me = LUSTRE_SP_CLI;
357 cli->cl_sp_to = LUSTRE_SP_MDT;
358 ns_type = LDLM_NS_TYPE_MDC;
359 } else if (!strcmp(name, LUSTRE_OSP_NAME)) {
360 if (strstr(lustre_cfg_buf(lcfg, 1), "OST") == NULL) {
361 /* OSP_on_MDT for other MDTs */
362 connect_op = MDS_CONNECT;
363 cli->cl_sp_to = LUSTRE_SP_MDT;
364 ns_type = LDLM_NS_TYPE_MDC;
365 rq_portal = OUT_PORTAL;
367 /* OSP on MDT for OST */
368 connect_op = OST_CONNECT;
369 cli->cl_sp_to = LUSTRE_SP_OST;
370 ns_type = LDLM_NS_TYPE_OSC;
371 rq_portal = OST_REQUEST_PORTAL;
373 rp_portal = OSC_REPLY_PORTAL;
374 cli->cl_sp_me = LUSTRE_SP_MDT;
375 } else if (!strcmp(name, LUSTRE_MGC_NAME)) {
376 rq_portal = MGS_REQUEST_PORTAL;
377 rp_portal = MGC_REPLY_PORTAL;
378 connect_op = MGS_CONNECT;
379 cli->cl_sp_me = LUSTRE_SP_MGC;
380 cli->cl_sp_to = LUSTRE_SP_MGS;
381 cli->cl_flvr_mgc.sf_rpc = SPTLRPC_FLVR_INVALID;
382 ns_type = LDLM_NS_TYPE_MGC;
384 CERROR("unknown client OBD type \"%s\", can't setup\n",
389 if (LUSTRE_CFG_BUFLEN(lcfg, 1) < 1) {
390 CERROR("requires a TARGET UUID\n");
394 if (LUSTRE_CFG_BUFLEN(lcfg, 1) > 37) {
395 CERROR("client UUID must be less than 38 characters\n");
399 if (LUSTRE_CFG_BUFLEN(lcfg, 2) < 1) {
400 CERROR("setup requires a SERVER UUID\n");
404 if (LUSTRE_CFG_BUFLEN(lcfg, 2) > 37) {
405 CERROR("target UUID must be less than 38 characters\n");
409 init_rwsem(&cli->cl_sem);
410 mutex_init(&cli->cl_mgc_mutex);
412 init_rwsem(&cli->cl_seq_rwsem);
413 cli->cl_conn_count = 0;
414 memcpy(server_uuid.uuid, lustre_cfg_buf(lcfg, 2),
415 min_t(unsigned int, LUSTRE_CFG_BUFLEN(lcfg, 2),
416 sizeof(server_uuid)));
418 cli->cl_dirty_pages = 0;
419 cli->cl_dirty_max_pages = 0;
420 cli->cl_avail_grant = 0;
421 /* FIXME: Should limit this for the sum of all cl_dirty_max_pages. */
423 * cl_dirty_max_pages may be changed at connect time in
424 * ptlrpc_connect_interpret().
426 client_adjust_max_dirty(cli);
427 init_waitqueue_head(&cli->cl_cache_waiters);
428 INIT_LIST_HEAD(&cli->cl_loi_ready_list);
429 INIT_LIST_HEAD(&cli->cl_loi_hp_ready_list);
430 INIT_LIST_HEAD(&cli->cl_loi_write_list);
431 INIT_LIST_HEAD(&cli->cl_loi_read_list);
432 spin_lock_init(&cli->cl_loi_list_lock);
433 atomic_set(&cli->cl_pending_w_pages, 0);
434 atomic_set(&cli->cl_pending_r_pages, 0);
435 cli->cl_r_in_flight = 0;
436 cli->cl_w_in_flight = 0;
438 spin_lock_init(&cli->cl_read_rpc_hist.oh_lock);
439 spin_lock_init(&cli->cl_write_rpc_hist.oh_lock);
440 spin_lock_init(&cli->cl_read_page_hist.oh_lock);
441 spin_lock_init(&cli->cl_write_page_hist.oh_lock);
442 spin_lock_init(&cli->cl_read_offset_hist.oh_lock);
443 spin_lock_init(&cli->cl_write_offset_hist.oh_lock);
446 INIT_LIST_HEAD(&cli->cl_lru_osc);
447 atomic_set(&cli->cl_lru_shrinkers, 0);
448 atomic_long_set(&cli->cl_lru_busy, 0);
449 atomic_long_set(&cli->cl_lru_in_list, 0);
450 INIT_LIST_HEAD(&cli->cl_lru_list);
451 spin_lock_init(&cli->cl_lru_list_lock);
452 atomic_long_set(&cli->cl_unstable_count, 0);
453 INIT_LIST_HEAD(&cli->cl_shrink_list);
454 INIT_LIST_HEAD(&cli->cl_grant_chain);
456 INIT_LIST_HEAD(&cli->cl_flight_waiters);
457 cli->cl_rpcs_in_flight = 0;
459 init_waitqueue_head(&cli->cl_destroy_waitq);
460 atomic_set(&cli->cl_destroy_in_flight, 0);
463 cli->cl_supp_cksum_types = OBD_CKSUM_CRC32;
464 cli->cl_preferred_cksum_type = 0;
465 #ifdef ENABLE_CHECKSUM
466 /* Turn on checksumming by default. */
467 cli->cl_checksum = 1;
469 * The supported checksum types will be worked out at connect time
470 * Set cl_chksum* to CRC32 for now to avoid returning screwed info
473 cli->cl_cksum_type = cli->cl_supp_cksum_types;
475 atomic_set(&cli->cl_resends, OSC_DEFAULT_RESENDS);
478 * Set it to possible maximum size. It may be reduced by ocd_brw_size
479 * from OFD after connecting.
481 cli->cl_max_pages_per_rpc = PTLRPC_MAX_BRW_PAGES;
483 cli->cl_max_short_io_bytes = OBD_DEF_SHORT_IO_BYTES;
486 * set cl_chunkbits default value to PAGE_SHIFT,
487 * it will be updated at OSC connection time.
489 cli->cl_chunkbits = PAGE_SHIFT;
491 if (!strcmp(name, LUSTRE_MDC_NAME)) {
492 cli->cl_max_rpcs_in_flight = OBD_MAX_RIF_DEFAULT;
493 } else if (cfs_totalram_pages() >> (20 - PAGE_SHIFT) <= 128 /* MB */) {
494 cli->cl_max_rpcs_in_flight = 2;
495 } else if (cfs_totalram_pages() >> (20 - PAGE_SHIFT) <= 256 /* MB */) {
496 cli->cl_max_rpcs_in_flight = 3;
497 } else if (cfs_totalram_pages() >> (20 - PAGE_SHIFT) <= 512 /* MB */) {
498 cli->cl_max_rpcs_in_flight = 4;
500 if (osc_on_mdt(obd->obd_name))
501 cli->cl_max_rpcs_in_flight = OBD_MAX_RIF_MAX;
503 cli->cl_max_rpcs_in_flight = OBD_MAX_RIF_DEFAULT;
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 lnet_nid4_to_nid(LNET_MKNID(refnet, 0), &fake_conn.c_self);
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 spin_lock(&imp->imp_lock);
657 if (imp->imp_state == LUSTRE_IMP_CLOSED && imp->imp_deactive) {
658 /* need to reactivate import if trying to connect
659 * to a previously disconnected
661 imp->imp_deactive = 0;
662 imp->imp_invalid = 0;
664 spin_unlock(&imp->imp_lock);
666 imp->imp_dlm_handle = conn;
667 rc = ptlrpc_init_import(imp);
671 ocd = &imp->imp_connect_data;
674 imp->imp_connect_flags_orig = data->ocd_connect_flags;
675 imp->imp_connect_flags2_orig = data->ocd_connect_flags2;
678 rc = ptlrpc_connect_import(imp);
680 LASSERT(imp->imp_state == LUSTRE_IMP_DISCON);
683 LASSERT(*exp != NULL && (*exp)->exp_connection);
686 LASSERTF((ocd->ocd_connect_flags & data->ocd_connect_flags) ==
687 ocd->ocd_connect_flags, "old %#llx, new %#llx\n",
688 data->ocd_connect_flags, ocd->ocd_connect_flags);
689 data->ocd_connect_flags = ocd->ocd_connect_flags;
690 data->ocd_connect_flags2 = ocd->ocd_connect_flags2;
693 ptlrpc_pinger_add_import(imp);
699 cli->cl_conn_count--;
700 class_disconnect(*exp);
704 up_write(&cli->cl_sem);
706 if (!rc && localdata && !cli->cl_cache) {
707 cli->cl_cache = (struct cl_client_cache *)localdata;
708 cl_cache_incref(cli->cl_cache);
709 cli->cl_lru_left = &cli->cl_cache->ccc_lru_left;
711 /* add this osc into entity list */
712 LASSERT(list_empty(&cli->cl_lru_osc));
713 spin_lock(&cli->cl_cache->ccc_lru_lock);
714 list_add(&cli->cl_lru_osc, &cli->cl_cache->ccc_lru);
715 spin_unlock(&cli->cl_cache->ccc_lru_lock);
720 EXPORT_SYMBOL(client_connect_import);
722 int client_disconnect_export(struct obd_export *exp)
724 struct obd_device *obd = class_exp2obd(exp);
725 struct client_obd *cli;
726 struct obd_import *imp;
732 CERROR("invalid export for disconnect: exp %p cookie %#llx\n",
733 exp, exp ? exp->exp_handle.h_cookie : -1);
738 imp = cli->cl_import;
740 down_write(&cli->cl_sem);
741 CDEBUG(D_INFO, "disconnect %s - %zu\n", obd->obd_name,
744 if (cli->cl_conn_count == 0) {
745 CERROR("disconnecting disconnected device (%s)\n",
747 GOTO(out_disconnect, rc = -EINVAL);
750 cli->cl_conn_count--;
751 if (cli->cl_conn_count != 0)
752 GOTO(out_disconnect, rc = 0);
755 * Mark import deactivated now, so we don't try to reconnect if any
756 * of the cleanup RPCs fails (e.g. LDLM cancel, etc). We don't
757 * fully deactivate the import, or that would drop all requests.
759 spin_lock(&imp->imp_lock);
760 imp->imp_deactive = 1;
761 spin_unlock(&imp->imp_lock);
764 * Some non-replayable imports (MDS's OSCs) are pinged, so just
765 * delete it regardless. (It's safe to delete an import that was
768 (void)ptlrpc_pinger_del_import(imp);
770 if (obd->obd_namespace != NULL) {
771 /* obd_force == local only */
772 ldlm_cli_cancel_unused(obd->obd_namespace, NULL,
773 obd->obd_force ? LCF_LOCAL : 0, NULL);
774 ldlm_namespace_free_prior(obd->obd_namespace, imp,
779 * There's no need to hold sem while disconnecting an import,
780 * and it may actually cause deadlock in GSS.
782 up_write(&cli->cl_sem);
783 rc = ptlrpc_disconnect_import(imp, 0);
784 down_write(&cli->cl_sem);
786 ptlrpc_invalidate_import(imp);
792 * Use server style - class_disconnect should be always called for
795 err = class_disconnect(exp);
799 up_write(&cli->cl_sem);
803 EXPORT_SYMBOL(client_disconnect_export);
805 #ifdef HAVE_SERVER_SUPPORT
806 int server_disconnect_export(struct obd_export *exp)
812 /* Disconnect early so that clients can't keep using export. */
813 rc = class_disconnect(exp);
814 /* Close import to avoid sending any requests. */
815 if (exp->exp_imp_reverse)
816 ptlrpc_cleanup_imp(exp->exp_imp_reverse);
818 ldlm_bl_thread_wakeup();
820 /* complete all outstanding replies */
821 spin_lock(&exp->exp_lock);
822 while (!list_empty(&exp->exp_outstanding_replies)) {
823 struct ptlrpc_reply_state *rs =
824 list_entry(exp->exp_outstanding_replies.next,
825 struct ptlrpc_reply_state, rs_exp_list);
826 struct ptlrpc_service_part *svcpt = rs->rs_svcpt;
828 spin_lock(&svcpt->scp_rep_lock);
830 list_del_init(&rs->rs_exp_list);
832 spin_lock(&rs->rs_lock);
833 /* clear rs_convert_lock to make sure rs is handled and put */
834 rs->rs_convert_lock = 0;
835 ptlrpc_schedule_difficult_reply(rs);
836 spin_unlock(&rs->rs_lock);
838 spin_unlock(&svcpt->scp_rep_lock);
840 spin_unlock(&exp->exp_lock);
844 EXPORT_SYMBOL(server_disconnect_export);
846 static inline int target_check_recovery_timer(struct obd_device *target)
851 if (!target->obd_recovering || target->obd_recovery_start == 0)
854 remaining = hrtimer_get_remaining(&target->obd_recovery_timer);
855 timeout = ktime_divns(remaining, NSEC_PER_SEC);
859 /* the recovery timer should expire, but it isn't triggered,
860 * it's better to abort the recovery of this target to speed up
861 * the recovery of the whole cluster. */
862 spin_lock(&target->obd_dev_lock);
863 if (target->obd_recovering) {
864 CERROR("%s: Aborting recovery\n", target->obd_name);
865 target->obd_abort_recovery = 1;
866 wake_up(&target->obd_next_transno_waitq);
868 spin_unlock(&target->obd_dev_lock);
873 * --------------------------------------------------------------------------
874 * from old lib/target.c
875 * --------------------------------------------------------------------------
877 static int target_handle_reconnect(struct lustre_handle *conn,
878 struct obd_export *exp,
879 struct obd_uuid *cluuid)
881 struct obd_device *target;
882 struct lustre_handle *hdl;
888 hdl = &exp->exp_imp_reverse->imp_remote_handle;
889 if (!exp->exp_connection || !lustre_handle_is_used(hdl)) {
890 conn->cookie = exp->exp_handle.h_cookie;
892 "connect export for UUID '%s' at %p, cookie %#llx\n",
893 cluuid->uuid, exp, conn->cookie);
897 target = exp->exp_obd;
899 /* Might be a re-connect after a partition. */
900 if (memcmp(&conn->cookie, &hdl->cookie, sizeof(conn->cookie))) {
901 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",
903 obd_uuid2str(&exp->exp_client_uuid),
904 obd_export_nid2str(exp),
905 hdl->cookie, conn->cookie);
906 memset(conn, 0, sizeof(*conn));
908 * target_handle_connect() treats EALREADY and
909 * -EALREADY differently. -EALREADY is an error
910 * (same UUID, different handle).
915 if (!target->obd_recovering) {
916 LCONSOLE_WARN("%s: Client %s (at %s) reconnecting\n",
917 target->obd_name, obd_uuid2str(&exp->exp_client_uuid),
918 obd_export_nid2str(exp));
919 GOTO(out_already, rc);
922 remaining = hrtimer_get_remaining(&target->obd_recovery_timer);
923 timeout = ktime_divns(remaining, NSEC_PER_SEC);
925 LCONSOLE_WARN("%s: Client %s (at %s) reconnected, waiting for %d clients in recovery for %lld:%.02lld\n",
927 obd_uuid2str(&exp->exp_client_uuid),
928 obd_export_nid2str(exp),
929 atomic_read(&target->obd_max_recoverable_clients),
930 timeout / 60, timeout % 60);
932 struct target_distribute_txn_data *tdtd;
937 target_check_recovery_timer(target);
939 tdtd = class_exp2tgt(exp)->lut_tdtd;
940 if (tdtd && tdtd->tdtd_show_update_logs_retrievers)
941 buf = tdtd->tdtd_show_update_logs_retrievers(
942 tdtd->tdtd_show_retrievers_cbdata,
946 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",
948 obd_uuid2str(&exp->exp_client_uuid),
949 obd_export_nid2str(exp), count,
950 buf ? buf : "unknown (not enough RAM)",
951 (abs(timeout) + target->obd_recovery_timeout) / 60,
952 (abs(timeout) + target->obd_recovery_timeout) % 60,
955 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",
956 target->obd_name, abs(timeout) / 60,
957 abs(timeout) % 60, target->obd_name);
964 conn->cookie = exp->exp_handle.h_cookie;
966 * target_handle_connect() treats EALREADY and
967 * -EALREADY differently. EALREADY means we are
968 * doing a valid reconnect from the same client.
974 check_and_start_recovery_timer(struct obd_device *obd,
975 struct ptlrpc_request *req, int new_client);
978 * update flags for import during reconnect process
980 static int rev_import_flags_update(struct obd_import *revimp,
981 struct ptlrpc_request *req)
984 struct obd_connect_data *data;
986 data = req_capsule_client_get(&req->rq_pill, &RMF_CONNECT_DATA);
988 if (data->ocd_connect_flags & OBD_CONNECT_AT)
989 revimp->imp_msghdr_flags |= MSGHDR_AT_SUPPORT;
991 revimp->imp_msghdr_flags &= ~MSGHDR_AT_SUPPORT;
993 revimp->imp_msghdr_flags |= MSGHDR_CKSUM_INCOMPAT18;
995 revimp->imp_connect_data = *data;
996 rc = sptlrpc_import_sec_adapt(revimp, req->rq_svc_ctx, &req->rq_flvr);
998 CERROR("%s: cannot get reverse import %s security: rc = %d\n",
999 revimp->imp_client->cli_name,
1000 libcfs_idstr(&req->rq_peer), rc);
1008 * Allocate a new reverse import for an export.
1010 * \retval -errno in case error hit
1011 * \retval 0 if reverse import correctly init
1013 int rev_import_init(struct obd_export *export)
1015 struct obd_device *obd = export->exp_obd;
1016 struct obd_import *revimp;
1018 LASSERT(export->exp_imp_reverse == NULL);
1020 revimp = class_new_import(obd);
1024 revimp->imp_remote_handle.cookie = 0ULL;
1025 revimp->imp_client = &obd->obd_ldlm_client;
1026 revimp->imp_dlm_fake = 1;
1028 /* it is safe to connect import in new state as no sends possible */
1029 spin_lock(&export->exp_lock);
1030 export->exp_imp_reverse = revimp;
1031 spin_unlock(&export->exp_lock);
1032 class_import_put(revimp);
1036 EXPORT_SYMBOL(rev_import_init);
1039 * Handle reconnect for an export.
1041 * \param exp export to handle reconnect process
1042 * \param req client reconnect request
1044 * \retval -rc in case securitfy flavor can't be changed
1045 * \retval 0 in case none problems
1047 static int rev_import_reconnect(struct obd_export *exp,
1048 struct ptlrpc_request *req)
1050 struct obd_import *revimp = exp->exp_imp_reverse;
1051 struct lustre_handle *lh;
1054 /* avoid sending a request until import flags are changed */
1055 ptlrpc_import_enter_resend(revimp);
1057 ptlrpc_connection_put(revimp->imp_connection);
1060 * client from recovery don't have a handle so we need to take from
1061 * request. it may produce situation when wrong client connected
1062 * to recovery as we trust a client uuid
1064 lh = req_capsule_client_get(&req->rq_pill, &RMF_CONN);
1065 revimp->imp_remote_handle = *lh;
1068 * unknown versions will be caught in
1069 * ptlrpc_handle_server_req_in->lustre_unpack_msg()
1071 revimp->imp_msg_magic = req->rq_reqmsg->lm_magic;
1073 revimp->imp_connection = ptlrpc_connection_addref(exp->exp_connection);
1075 rc = rev_import_flags_update(revimp, req);
1078 * it is safe to still be in RECOVERY phase as we are not able
1079 * to setup correct security flavor so requests are not able to
1080 * be delivered correctly
1085 /* resend all rpc's via new connection */
1086 return ptlrpc_import_recovery_state_machine(revimp);
1089 int target_handle_connect(struct ptlrpc_request *req)
1091 struct obd_device *target = NULL;
1092 struct obd_export *export = NULL;
1094 * connect handle - filled from target_handle_reconnect in
1097 struct lustre_handle conn;
1098 struct lustre_handle *tmp;
1099 struct obd_uuid cluuid;
1104 bool mds_conn = false, lw_client = false, initial_conn = false;
1105 bool mds_mds_conn = false;
1106 bool new_mds_mds_conn = false;
1107 struct obd_connect_data *data, *tmpdata;
1109 lnet_nid_t client_nid;
1110 #if LUSTRE_VERSION_CODE < OBD_OCD_VERSION(3, 0, 53, 0)
1111 int tmp_exp_old_falloc;
1113 struct ptlrpc_connection *pcon = NULL;
1117 OBD_RACE(OBD_FAIL_TGT_CONN_RACE);
1119 str = req_capsule_client_get(&req->rq_pill, &RMF_TGTUUID);
1121 DEBUG_REQ(D_ERROR, req, "bad target UUID for connect");
1122 GOTO(out, rc = -EINVAL);
1125 target = class_dev_by_str(str);
1127 deuuidify(str, NULL, &target_start, &target_len);
1128 LCONSOLE_ERROR_MSG(0x137,
1129 "%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",
1130 str, libcfs_nidstr(&req->rq_peer.nid));
1131 GOTO(out, rc = -ENODEV);
1134 spin_lock(&target->obd_dev_lock);
1136 target->obd_conn_inprogress++;
1138 if (target->obd_stopping || !target->obd_set_up) {
1139 spin_unlock(&target->obd_dev_lock);
1141 deuuidify(str, NULL, &target_start, &target_len);
1142 LCONSOLE_INFO("%.*s: Not available for connect from %s (%s)\n",
1143 target_len, target_start,
1144 libcfs_nidstr(&req->rq_peer.nid),
1145 (target->obd_stopping ?
1146 "stopping" : "not set up"));
1147 GOTO(out, rc = -ENODEV);
1150 if (target->obd_no_conn) {
1151 spin_unlock(&target->obd_dev_lock);
1154 "%s: Temporarily refusing client connection from %s\n",
1155 target->obd_name, libcfs_nidstr(&req->rq_peer.nid));
1156 GOTO(out, rc = -EAGAIN);
1159 spin_unlock(&target->obd_dev_lock);
1161 str = req_capsule_client_get(&req->rq_pill, &RMF_CLUUID);
1163 DEBUG_REQ(D_ERROR, req, "bad client UUID for connect");
1164 GOTO(out, rc = -EINVAL);
1167 obd_str2uuid(&cluuid, str);
1169 tmp = req_capsule_client_get(&req->rq_pill, &RMF_CONN);
1171 GOTO(out, rc = -EPROTO);
1175 size = req_capsule_get_size(&req->rq_pill, &RMF_CONNECT_DATA,
1177 if (size < 0 || size > 8 * sizeof(struct obd_connect_data))
1178 GOTO(out, rc = -EPROTO);
1179 data = req_capsule_client_get(&req->rq_pill, &RMF_CONNECT_DATA);
1181 GOTO(out, rc = -EPROTO);
1183 rc = req_capsule_server_pack(&req->rq_pill);
1187 #if LUSTRE_VERSION_CODE < OBD_OCD_VERSION(3, 0, 53, 0)
1189 * Don't allow clients to connect that are using old 1.8 format
1190 * protocol conventions (LUSTRE_MSG_MAGIC_v1, !MSGHDR_CKSUM_INCOMPAT18,
1191 * ldlm_flock_policy_wire format, MDT_ATTR_xTIME_SET, etc). The
1192 * FULL20 flag should be set on all connections since 2.0, but no
1193 * longer affects behaviour.
1195 * Later this check will be disabled and the flag can be retired
1196 * completely once interop with 3.0 is no longer needed.
1198 if (!(data->ocd_connect_flags & OBD_CONNECT_FULL20))
1199 GOTO(out, rc = -EPROTO);
1201 /* Old clients will have 'tmp_exp_old_falloc' as 1.
1202 * Newer clients (2.15) and beyond will have it set as 0
1204 tmp_exp_old_falloc =
1205 !!(data->ocd_connect_flags & OBD_CONNECT_OLD_FALLOC);
1207 CDEBUG(D_INFO, "%s: ocd_connect_flags: %#llx tmp_exp_old_falloc: %d\n",
1208 target->obd_name, data->ocd_connect_flags, tmp_exp_old_falloc);
1211 * Don't allow liblustre clients to connect.
1212 * - testing was disabled in v2_2_50_0-61-g6a75d65
1213 * - building was disabled in v2_5_58_0-28-g7277179
1214 * - client code was deleted in v2_6_50_0-101-gcdfbc72,
1215 * - clients were refused connect for version difference > 0.0.1.32
1217 if (lustre_msg_get_op_flags(req->rq_reqmsg) & MSG_CONNECT_LIBCLIENT) {
1218 DEBUG_REQ(D_WARNING, req, "Refusing libclient connection");
1219 GOTO(out, rc = -EPROTO);
1224 * Note: lw_client is needed in MDS-MDS failover during update log
1225 * processing, so we needs to allow lw_client to be connected at
1226 * anytime, instead of only the initial connection
1228 lw_client = OCD_HAS_FLAG(data, LIGHTWEIGHT);
1230 if (lustre_msg_get_op_flags(req->rq_reqmsg) & MSG_CONNECT_INITIAL) {
1231 initial_conn = true;
1232 mds_conn = OCD_HAS_FLAG(data, MDS);
1233 mds_mds_conn = OCD_HAS_FLAG(data, MDS_MDS);
1236 * OBD_CONNECT_MNE_SWAB is removed at 2.14
1237 * Checking OBD_CONNECT_FID can be removed in the future.
1239 * Via check OBD_CONNECT_FID, we can distinguish whether
1240 * the OBD_CONNECT_MDS_MDS/OBD_CONNECT_MNE_SWAB is from
1241 * MGC or MDT, since MGC does not use OBD_CONNECT_FID.
1244 (data->ocd_connect_flags & OBD_CONNECT_MDS_MDS) &&
1245 (data->ocd_connect_flags & OBD_CONNECT_FID) &&
1246 (data->ocd_connect_flags & OBD_CONNECT_VERSION)) {
1247 __u32 major = OBD_OCD_VERSION_MAJOR(data->ocd_version);
1248 __u32 minor = OBD_OCD_VERSION_MINOR(data->ocd_version);
1249 __u32 patch = OBD_OCD_VERSION_PATCH(data->ocd_version);
1252 * We do not support the MDT-MDT interoperations with
1253 * different version MDT because of protocol changes.
1255 if (unlikely(major != LUSTRE_MAJOR ||
1256 minor != LUSTRE_MINOR ||
1257 abs(patch - LUSTRE_PATCH) > 3)) {
1258 LCONSOLE_WARN("%s (%u.%u.%u.%u) refused the connection from different version MDT (%d.%d.%d.%d) %s %s\n",
1259 target->obd_name, LUSTRE_MAJOR,
1260 LUSTRE_MINOR, LUSTRE_PATCH,
1261 LUSTRE_FIX, major, minor, patch,
1262 OBD_OCD_VERSION_FIX(data->ocd_version),
1263 libcfs_nidstr(&req->rq_peer.nid),
1265 GOTO(out, rc = -EPROTO);
1270 /* lctl gets a backstage, all-access pass. */
1271 if (obd_uuid_equals(&cluuid, &target->obd_uuid))
1272 goto dont_check_exports;
1274 export = obd_uuid_lookup(target, &cluuid);
1278 /* We've found an export in the hash. */
1280 spin_lock(&export->exp_lock);
1282 if (export->exp_connecting) { /* b=9635, et. al. */
1283 spin_unlock(&export->exp_lock);
1284 LCONSOLE_WARN("%s: Export %p already connecting from %s\n",
1285 export->exp_obd->obd_name, export,
1286 libcfs_nidstr(&req->rq_peer.nid));
1287 class_export_put(export);
1290 } else if ((mds_conn || (lw_client && initial_conn) ||
1291 OCD_HAS_FLAG(data, MDS_MDS)) && export->exp_connection) {
1292 spin_unlock(&export->exp_lock);
1293 if (!nid_same(&req->rq_peer.nid,
1294 &export->exp_connection->c_peer.nid)) {
1295 /* MDS or LWP reconnected after failover. */
1296 LCONSOLE_WARN("%s: Received %s connection from %s, removing former export from %s\n",
1298 lw_client ? "LWP" : "MDS",
1299 libcfs_nidstr(&req->rq_peer.nid),
1300 libcfs_nidstr(&export->exp_connection->c_peer.nid));
1302 /* New connection from the same NID. */
1303 LCONSOLE_WARN("%s: Received new %s connection from %s, %s former export from same NID\n",
1305 lw_client ? "LWP" : "MDS",
1306 libcfs_nidstr(&req->rq_peer.nid),
1307 OCD_HAS_FLAG(data, MDS_MDS) ?
1311 if (nid_same(&req->rq_peer.nid,
1312 &export->exp_connection->c_peer.nid) &&
1313 OCD_HAS_FLAG(data, MDS_MDS)) {
1315 * Because exports between MDTs will always be
1316 * kept, let's do not fail such export if they
1317 * come from the same NID, otherwise it might
1318 * cause eviction between MDTs, which might
1319 * cause namespace inconsistency
1321 spin_lock(&export->exp_lock);
1322 export->exp_connecting = 1;
1323 export->exp_conn_cnt = 0;
1324 spin_unlock(&export->exp_lock);
1325 conn.cookie = export->exp_handle.h_cookie;
1328 class_fail_export(export);
1329 class_export_put(export);
1333 } else if (export->exp_connection != NULL && initial_conn &&
1334 !nid_same(&req->rq_peer.nid,
1335 &export->exp_connection->c_peer.nid)) {
1336 spin_unlock(&export->exp_lock);
1337 /* In MDS failover we have static UUID but NID can change. */
1338 LCONSOLE_WARN("%s: Client %s seen on new nid %s when existing nid %s is already connected\n",
1339 target->obd_name, cluuid.uuid,
1340 libcfs_nidstr(&req->rq_peer.nid),
1342 &export->exp_connection->c_peer.nid));
1344 class_export_put(export);
1346 } else if (OBD_FAIL_PRECHECK(OBD_FAIL_TGT_RECOVERY_CONNECT) &&
1348 spin_unlock(&export->exp_lock);
1351 export->exp_connecting = 1;
1352 spin_unlock(&export->exp_lock);
1353 LASSERT(export->exp_obd == target);
1355 rc = target_handle_reconnect(&conn, export, &cluuid);
1358 /* If we found an export, we already unlocked. */
1361 OBD_FAIL_TIMEOUT(OBD_FAIL_TGT_DELAY_CONNECT, 2 * obd_timeout);
1362 } else if (req->rq_export == NULL &&
1363 atomic_read(&export->exp_rpc_count) > 0) {
1364 LCONSOLE_WARN("%s: Client %s (at %s) refused connection, still busy with %d references\n",
1365 target->obd_name, cluuid.uuid,
1366 libcfs_nidstr(&req->rq_peer.nid),
1367 refcount_read(&export->exp_handle.h_ref));
1368 GOTO(out, rc = -EBUSY);
1369 } else if (lustre_msg_get_conn_cnt(req->rq_reqmsg) == 1 &&
1371 if (!strstr(cluuid.uuid, "mdt"))
1372 LCONSOLE_WARN("%s: Rejecting reconnect from the known client %s (at %s) because it is indicating it is a new client\n",
1373 target->obd_name, cluuid.uuid,
1374 libcfs_nidstr(&req->rq_peer.nid));
1375 GOTO(out, rc = -EALREADY);
1377 OBD_FAIL_TIMEOUT(OBD_FAIL_TGT_DELAY_RECONNECT, 2 * obd_timeout);
1383 CDEBUG(D_HA, "%s: connection from %s@%s %st%llu exp %p cur %lld last %lld\n",
1384 target->obd_name, cluuid.uuid, libcfs_nidstr(&req->rq_peer.nid),
1385 target->obd_recovering ? "recovering/" : "", data->ocd_transno,
1386 export, ktime_get_seconds(),
1387 export ? export->exp_last_request_time : 0);
1390 * If this is the first time a client connects, reset the recovery
1391 * timer. Discard lightweight connections which might be local.
1393 if (!lw_client && rc == 0 && target->obd_recovering)
1394 check_and_start_recovery_timer(target, req, export == NULL);
1397 * We want to handle EALREADY but *not* -EALREADY from
1398 * target_handle_reconnect(), return reconnection state in a flag.
1400 if (rc == EALREADY) {
1401 lustre_msg_add_op_flags(req->rq_repmsg, MSG_CONNECT_RECONNECT);
1407 /* Tell the client if we support replayable requests. */
1408 if (target->obd_replayable)
1409 lustre_msg_add_op_flags(req->rq_repmsg, MSG_CONNECT_REPLAYABLE);
1410 client_nid = lnet_nid_to_nid4(&req->rq_peer.nid);
1412 if (export == NULL) {
1413 /* allow lightweight connections during recovery */
1415 * allow "new" MDT to be connected during recovery, since we
1416 * need retrieve recovery update records from it
1418 if (target->obd_recovering && !lw_client && !mds_mds_conn) {
1419 struct hrtimer *timer = &target->obd_recovery_timer;
1428 connected = atomic_read(&target->obd_connected_clients);
1429 in_progress = atomic_read(&target->obd_lock_replay_clients);
1431 atomic_read(&target->obd_max_recoverable_clients);
1432 stale = target->obd_stale_clients;
1433 remaining = hrtimer_get_remaining(timer);
1434 left = ktime_divns(remaining, NSEC_PER_SEC);
1436 if (ktime_to_ns(remaining) > 0) {
1437 msg = "to recover in";
1440 msg = "already passed deadline";
1443 target_check_recovery_timer(target);
1446 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",
1447 target->obd_name, cluuid.uuid,
1448 libcfs_nidstr(&req->rq_peer.nid), known,
1449 connected - in_progress, in_progress,
1450 stale, msg, timeout / 60, timeout % 60);
1454 rc = obd_connect(req->rq_svc_thread->t_env,
1455 &export, target, &cluuid, data,
1457 if (mds_conn && OBD_FAIL_CHECK(OBD_FAIL_TGT_RCVG_FLAG))
1458 lustre_msg_add_op_flags(req->rq_repmsg,
1459 MSG_CONNECT_RECOVERING);
1461 conn.cookie = export->exp_handle.h_cookie;
1462 rc = rev_import_init(export);
1466 new_mds_mds_conn = true;
1469 rc = obd_reconnect(req->rq_svc_thread->t_env,
1470 export, target, &cluuid, data, &client_nid);
1475 data->ocd_instance = obd2obt(target)->obt_instance;
1478 * Return only the parts of obd_connect_data that we understand, so the
1479 * client knows that we don't understand the rest.
1482 tmpsize = req_capsule_get_size(&req->rq_pill, &RMF_CONNECT_DATA,
1484 tmpdata = req_capsule_server_get(&req->rq_pill,
1487 * Don't use struct assignment here, because the client reply
1488 * buffer may be smaller/larger than the local struct
1491 memcpy(tmpdata, data, min(tmpsize, size));
1495 * If the client and the server are the same node, we will already
1496 * have an export that really points to the client's DLM export,
1497 * because we have a shared handles table.
1499 * XXX this will go away when shaver stops sending the "connect" handle
1500 * in the real "remote handle" field of the request --phik 24 Apr 2003
1502 ptlrpc_request_change_export(req, export);
1504 pcon = ptlrpc_connection_get(&req->rq_peer,
1505 &req->rq_self, &cluuid);
1507 GOTO(out, rc = -ENOTCONN);
1509 spin_lock(&export->exp_lock);
1511 if (export->exp_disconnected) {
1512 spin_unlock(&export->exp_lock);
1513 GOTO(out, rc = -ENODEV);
1515 if (export->exp_conn_cnt >= lustre_msg_get_conn_cnt(req->rq_reqmsg)) {
1516 spin_unlock(&export->exp_lock);
1518 "%s: %s already connected at greater or equal conn_cnt: %d >= %d\n",
1519 cluuid.uuid, libcfs_nidstr(&req->rq_peer.nid),
1520 export->exp_conn_cnt,
1521 lustre_msg_get_conn_cnt(req->rq_reqmsg));
1523 GOTO(out, rc = -EALREADY);
1525 LASSERT(lustre_msg_get_conn_cnt(req->rq_reqmsg) > 0);
1526 export->exp_conn_cnt = lustre_msg_get_conn_cnt(req->rq_reqmsg);
1528 #if LUSTRE_VERSION_CODE < OBD_OCD_VERSION(3, 0, 53, 0)
1529 /* make 'tmp_exp_old_falloc' persistent by saving it into
1530 * server side export object(obd_export)
1532 export->exp_old_falloc = tmp_exp_old_falloc;
1535 /* Check to see if connection came from another NID. */
1536 if (export->exp_connection != NULL &&
1537 !nid_same(&export->exp_connection->c_peer.nid,
1538 &req->rq_peer.nid)) {
1539 obd_nid_del(export->exp_obd, export);
1540 ptlrpc_connection_put(export->exp_connection);
1541 export->exp_connection = NULL;
1544 if (export->exp_connection == NULL) {
1545 export->exp_connection = pcon;
1548 obd_nid_add(export->exp_obd, export);
1550 spin_unlock(&export->exp_lock);
1552 lustre_msg_set_handle(req->rq_repmsg, &conn);
1554 rc = rev_import_reconnect(export, req);
1558 if (target->obd_recovering && !export->exp_in_recovery && !lw_client) {
1560 __u64 transno = data->ocd_transno;
1562 spin_lock(&export->exp_lock);
1564 * possible race with class_disconnect_stale_exports,
1565 * export may be already in the eviction process
1567 if (export->exp_failed) {
1568 spin_unlock(&export->exp_lock);
1569 GOTO(out, rc = -ENODEV);
1571 export->exp_in_recovery = 1;
1572 export->exp_req_replay_needed = 1;
1573 export->exp_lock_replay_needed = 1;
1574 spin_unlock(&export->exp_lock);
1576 has_transno = !!(lustre_msg_get_op_flags(req->rq_reqmsg) &
1577 MSG_CONNECT_TRANSNO);
1578 if (has_transno && transno == 0)
1579 CWARN("Connect with zero transno!\n");
1581 if (has_transno && transno > 0 &&
1582 transno < target->obd_next_recovery_transno &&
1583 transno > target->obd_last_committed) {
1584 /* Another way is to use cmpxchg() to be lock-free. */
1585 spin_lock(&target->obd_recovery_task_lock);
1586 if (transno < target->obd_next_recovery_transno)
1587 target->obd_next_recovery_transno = transno;
1588 spin_unlock(&target->obd_recovery_task_lock);
1591 atomic_inc(&target->obd_req_replay_clients);
1592 atomic_inc(&target->obd_lock_replay_clients);
1594 * Note: MDS-MDS connection is allowed to be connected during
1595 * recovery, no matter if the exports needs to be recoveried.
1596 * Because we need retrieve updates logs from all other MDTs.
1597 * So if the MDS-MDS export is new, obd_max_recoverable_clients
1598 * also needs to be increased to match other recovery checking
1601 if (new_mds_mds_conn)
1602 atomic_inc(&target->obd_max_recoverable_clients);
1604 if (atomic_inc_return(&target->obd_connected_clients) ==
1605 atomic_read(&target->obd_max_recoverable_clients))
1606 wake_up(&target->obd_next_transno_waitq);
1609 /* Tell the client we're in recovery, when client is involved in it. */
1610 if (target->obd_recovering && !lw_client)
1611 lustre_msg_add_op_flags(req->rq_repmsg, MSG_CONNECT_RECOVERING);
1615 spin_lock(&export->exp_lock);
1616 export->exp_connecting = 0;
1617 spin_unlock(&export->exp_lock);
1619 class_export_put(export);
1621 if (target != NULL) {
1622 spin_lock(&target->obd_dev_lock);
1623 target->obd_conn_inprogress--;
1624 spin_unlock(&target->obd_dev_lock);
1625 class_decref(target, "find", current);
1628 ptlrpc_connection_put(pcon);
1629 req->rq_status = rc;
1633 int target_handle_disconnect(struct ptlrpc_request *req)
1639 rc = req_capsule_server_pack(&req->rq_pill);
1643 /* In case of target disconnect, updating sec ctx immediately is
1644 * required in order to record latest sequence number used.
1645 * Sequence is normally updated on export destroy, but this event
1646 * can occur too late, ie after a new target connect request has
1648 * Maintaining correct sequence when client connection becomes idle
1649 * ensures that GSS does not erroneously consider requests as replays.
1651 rc = sptlrpc_export_update_ctx(req->rq_export);
1655 /* Keep the rq_export around so we can send the reply. */
1656 req->rq_status = obd_disconnect(class_export_get(req->rq_export));
1661 void target_destroy_export(struct obd_export *exp)
1663 struct obd_import *imp = NULL;
1665 * exports created from last_rcvd data, and "fake"
1666 * exports created by lctl don't have an import
1668 spin_lock(&exp->exp_lock);
1669 if (exp->exp_imp_reverse != NULL) {
1670 imp = exp->exp_imp_reverse;
1671 exp->exp_imp_reverse = NULL;
1673 spin_unlock(&exp->exp_lock);
1675 client_destroy_import(imp);
1677 LASSERT_ATOMIC_ZERO(&exp->exp_locks_count);
1678 LASSERT_ATOMIC_ZERO(&exp->exp_rpc_count);
1679 LASSERT_ATOMIC_ZERO(&exp->exp_cb_count);
1680 LASSERT_ATOMIC_ZERO(&exp->exp_replay_count);
1682 EXPORT_SYMBOL(target_destroy_export);
1685 * Recovery functions
1687 static void target_request_copy_get(struct ptlrpc_request *req)
1689 class_export_rpc_inc(req->rq_export);
1690 LASSERT(list_empty(&req->rq_list));
1691 INIT_LIST_HEAD(&req->rq_replay_list);
1693 /* Increase refcount to keep request in queue. */
1694 atomic_inc(&req->rq_refcount);
1695 /* Let export know it has replays to be handled. */
1696 atomic_inc(&req->rq_export->exp_replay_count);
1699 static void target_request_copy_put(struct ptlrpc_request *req)
1701 LASSERT(list_empty(&req->rq_replay_list));
1702 LASSERT_ATOMIC_POS(&req->rq_export->exp_replay_count);
1704 atomic_dec(&req->rq_export->exp_replay_count);
1705 class_export_rpc_dec(req->rq_export);
1706 ptlrpc_server_drop_request(req);
1709 static int target_exp_enqueue_req_replay(struct ptlrpc_request *req)
1711 __u64 transno = lustre_msg_get_transno(req->rq_reqmsg);
1712 struct obd_export *exp = req->rq_export;
1713 struct ptlrpc_request *reqiter;
1714 struct ptlrpc_request *dup_req = NULL;
1719 spin_lock(&exp->exp_lock);
1720 list_for_each_entry(reqiter, &exp->exp_req_replay_queue,
1722 if (lustre_msg_get_transno(reqiter->rq_reqmsg) == transno) {
1730 /* We expect it with RESENT and REPLAY flags. */
1731 if ((lustre_msg_get_flags(req->rq_reqmsg) &
1732 (MSG_RESENT | MSG_REPLAY)) != (MSG_RESENT | MSG_REPLAY))
1733 CERROR("invalid flags %x of resent replay\n",
1734 lustre_msg_get_flags(req->rq_reqmsg));
1736 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY) {
1739 new_conn = lustre_msg_get_conn_cnt(req->rq_reqmsg);
1741 lustre_msg_get_conn_cnt(dup_req->rq_reqmsg))
1742 lustre_msg_set_conn_cnt(dup_req->rq_reqmsg,
1746 list_add_tail(&req->rq_replay_list,
1747 &exp->exp_req_replay_queue);
1750 spin_unlock(&exp->exp_lock);
1754 static void target_exp_dequeue_req_replay(struct ptlrpc_request *req)
1756 LASSERT(!list_empty(&req->rq_replay_list));
1757 LASSERT(req->rq_export);
1759 spin_lock(&req->rq_export->exp_lock);
1760 list_del_init(&req->rq_replay_list);
1761 spin_unlock(&req->rq_export->exp_lock);
1764 static void target_finish_recovery(struct lu_target *lut)
1766 struct obd_device *obd = lut->lut_obd;
1770 /* Only log a recovery message when recovery has occurred. */
1771 if (obd->obd_recovery_start) {
1772 time64_t now = ktime_get_seconds();
1773 time64_t elapsed_time;
1775 elapsed_time = max_t(time64_t, now - obd->obd_recovery_start,
1777 LCONSOLE_INFO("%s: Recovery over after %lld:%.02lld, of %d clients %d recovered and %d %s evicted.\n",
1778 obd->obd_name, elapsed_time / 60,
1780 atomic_read(&obd->obd_max_recoverable_clients),
1781 atomic_read(&obd->obd_connected_clients),
1782 obd->obd_stale_clients,
1783 obd->obd_stale_clients == 1 ? "was" : "were");
1784 if (obd->obd_stale_clients && do_dump_on_eviction(obd))
1785 libcfs_debug_dumplog();
1788 ldlm_reprocess_recovery_done(obd->obd_namespace);
1789 spin_lock(&obd->obd_recovery_task_lock);
1790 if (!list_empty(&obd->obd_req_replay_queue) ||
1791 !list_empty(&obd->obd_lock_replay_queue) ||
1792 !list_empty(&obd->obd_final_req_queue)) {
1793 CERROR("%s: Recovery queues ( %s%s%s) are not empty\n",
1795 list_empty(&obd->obd_req_replay_queue) ? "" : "req ",
1796 list_empty(&obd->obd_lock_replay_queue) ? \
1798 list_empty(&obd->obd_final_req_queue) ? \
1800 spin_unlock(&obd->obd_recovery_task_lock);
1803 spin_unlock(&obd->obd_recovery_task_lock);
1805 obd->obd_recovery_end = ktime_get_seconds();
1807 /* When recovery finished, cleanup orphans on MDS and OST. */
1808 if (obd->obd_type && OBP(obd, postrecov)) {
1809 int rc = OBP(obd, postrecov)(obd);
1812 LCONSOLE_WARN("%s: Post recovery failed, rc %d\n",
1818 static void abort_req_replay_queue(struct obd_device *obd)
1820 struct ptlrpc_request *req, *n;
1821 LIST_HEAD(abort_list);
1823 spin_lock(&obd->obd_recovery_task_lock);
1824 list_splice_init(&obd->obd_req_replay_queue, &abort_list);
1825 spin_unlock(&obd->obd_recovery_task_lock);
1826 list_for_each_entry_safe(req, n, &abort_list, rq_list) {
1827 DEBUG_REQ(D_WARNING, req, "aborted:");
1828 req->rq_status = -ENOTCONN;
1829 if (ptlrpc_error(req)) {
1830 DEBUG_REQ(D_ERROR, req,
1831 "failed abort_req_reply; skipping");
1833 target_exp_dequeue_req_replay(req);
1834 target_request_copy_put(req);
1838 static void abort_lock_replay_queue(struct obd_device *obd)
1840 struct ptlrpc_request *req, *n;
1841 LIST_HEAD(abort_list);
1843 spin_lock(&obd->obd_recovery_task_lock);
1844 list_splice_init(&obd->obd_lock_replay_queue, &abort_list);
1845 spin_unlock(&obd->obd_recovery_task_lock);
1846 list_for_each_entry_safe(req, n, &abort_list, rq_list) {
1847 DEBUG_REQ(D_ERROR, req, "aborted:");
1848 req->rq_status = -ENOTCONN;
1849 if (ptlrpc_error(req)) {
1850 DEBUG_REQ(D_ERROR, req,
1851 "failed abort_lock_reply; skipping");
1853 target_request_copy_put(req);
1858 * Called from a cleanup function if the device is being cleaned up
1859 * forcefully. The exports should all have been disconnected already,
1860 * the only thing left to do is
1861 * - clear the recovery flags
1862 * - cancel the timer
1863 * - free queued requests and replies, but don't send replies
1864 * Because the obd_stopping flag is set, no new requests should be received.
1866 void target_cleanup_recovery(struct obd_device *obd)
1868 struct ptlrpc_request *req, *n;
1869 LIST_HEAD(clean_list);
1871 spin_lock(&obd->obd_dev_lock);
1872 if (!obd->obd_recovering) {
1873 spin_unlock(&obd->obd_dev_lock);
1877 obd->obd_recovering = obd->obd_abort_recovery = 0;
1878 obd->obd_abort_recov_mdt = 0;
1879 spin_unlock(&obd->obd_dev_lock);
1881 spin_lock(&obd->obd_recovery_task_lock);
1882 target_cancel_recovery_timer(obd);
1883 list_splice_init(&obd->obd_req_replay_queue, &clean_list);
1884 spin_unlock(&obd->obd_recovery_task_lock);
1886 list_for_each_entry_safe(req, n, &clean_list, rq_list) {
1887 LASSERT(req->rq_reply_state == NULL);
1888 target_exp_dequeue_req_replay(req);
1889 target_request_copy_put(req);
1892 spin_lock(&obd->obd_recovery_task_lock);
1893 list_splice_init(&obd->obd_lock_replay_queue, &clean_list);
1894 list_splice_init(&obd->obd_final_req_queue, &clean_list);
1895 spin_unlock(&obd->obd_recovery_task_lock);
1897 list_for_each_entry_safe(req, n, &clean_list, rq_list) {
1898 LASSERT(req->rq_reply_state == NULL);
1899 target_request_copy_put(req);
1904 EXPORT_SYMBOL(target_cleanup_recovery);
1906 /* obd_recovery_task_lock should be held */
1907 void target_cancel_recovery_timer(struct obd_device *obd)
1909 CDEBUG(D_HA, "%s: cancel recovery timer\n", obd->obd_name);
1910 hrtimer_cancel(&obd->obd_recovery_timer);
1913 static void target_start_recovery_timer(struct obd_device *obd)
1917 if (obd->obd_recovery_start != 0)
1920 spin_lock(&obd->obd_dev_lock);
1921 if (!obd->obd_recovering || obd->obd_abort_recovery) {
1922 spin_unlock(&obd->obd_dev_lock);
1926 LASSERT(obd->obd_recovery_timeout != 0);
1928 if (obd->obd_recovery_start != 0) {
1929 spin_unlock(&obd->obd_dev_lock);
1933 obd->obd_recovery_start = ktime_get_seconds();
1934 delay = ktime_set(obd->obd_recovery_start +
1935 obd->obd_recovery_timeout, 0);
1936 hrtimer_start(&obd->obd_recovery_timer, delay, HRTIMER_MODE_ABS);
1937 spin_unlock(&obd->obd_dev_lock);
1939 LCONSOLE_WARN("%s: Will be in recovery for at least %u:%02u, or until %d client%s reconnect%s\n",
1941 obd->obd_recovery_timeout / 60,
1942 obd->obd_recovery_timeout % 60,
1943 atomic_read(&obd->obd_max_recoverable_clients),
1944 (atomic_read(&obd->obd_max_recoverable_clients) == 1) ?
1946 (atomic_read(&obd->obd_max_recoverable_clients) == 1) ?
1951 * extend recovery window.
1953 * if @extend is true, extend recovery window to have @dr_timeout remaining
1954 * at least; otherwise, make sure the recovery timeout value is not less
1957 static void extend_recovery_timer(struct obd_device *obd, timeout_t dr_timeout,
1964 spin_lock(&obd->obd_dev_lock);
1965 if (!obd->obd_recovering || obd->obd_abort_recovery ||
1966 obd->obd_stopping) {
1967 spin_unlock(&obd->obd_dev_lock);
1970 LASSERT(obd->obd_recovery_start != 0);
1972 left_ns = hrtimer_get_remaining(&obd->obd_recovery_timer);
1973 left = ktime_divns(left_ns, NSEC_PER_SEC);
1976 timeout = obd->obd_recovery_timeout;
1977 /* dr_timeout will happen after the hrtimer has expired.
1978 * Add the excess time to the soft recovery timeout without
1979 * exceeding the hard recovery timeout.
1981 if (dr_timeout > left) {
1982 timeout += dr_timeout - left;
1983 timeout = min_t(timeout_t, obd->obd_recovery_time_hard,
1987 timeout = clamp_t(timeout_t, dr_timeout,
1988 obd->obd_recovery_timeout,
1989 obd->obd_recovery_time_hard);
1992 if (timeout == obd->obd_recovery_time_hard)
1993 CWARN("%s: extended recovery timer reached hard limit: %d, extend: %d\n",
1994 obd->obd_name, timeout, extend);
1996 if (obd->obd_recovery_timeout < timeout) {
1999 obd->obd_recovery_timeout = timeout;
2000 end = ktime_set(obd->obd_recovery_start + timeout, 0);
2001 now = ktime_set(ktime_get_seconds(), 0);
2002 left_ns = ktime_sub(end, now);
2003 hrtimer_start(&obd->obd_recovery_timer, end, HRTIMER_MODE_ABS);
2004 left = ktime_divns(left_ns, NSEC_PER_SEC);
2006 spin_unlock(&obd->obd_dev_lock);
2008 CDEBUG(D_HA, "%s: recovery timer will expire in %d seconds\n",
2009 obd->obd_name, left);
2012 /* Reset the timer with each new client connection */
2014 * This timer is actually reconnect_timer, which is for making sure
2015 * the total recovery window is at least as big as my reconnect
2016 * attempt timing. So the initial recovery time_out will be set to
2017 * OBD_RECOVERY_FACTOR * obd_timeout. If the timeout coming
2018 * from client is bigger than this, then the recovery time_out will
2019 * be extended to make sure the client could be reconnected, in the
2020 * process, the timeout from the new client should be ignored.
2023 check_and_start_recovery_timer(struct obd_device *obd,
2024 struct ptlrpc_request *req,
2027 timeout_t service_timeout = lustre_msg_get_service_timeout(req->rq_reqmsg);
2028 struct obd_device_target *obt = obd2obt(obd);
2030 if (!new_client && service_timeout)
2032 * Teach server about old server's estimates, as first guess
2033 * at how long new requests will take.
2035 at_measured(&req->rq_rqbd->rqbd_svcpt->scp_at_estimate,
2038 target_start_recovery_timer(obd);
2041 * Convert the service time to RPC timeout,
2042 * and reuse service_timeout to limit stack usage.
2044 service_timeout = at_est2timeout(service_timeout);
2046 if (OBD_FAIL_CHECK(OBD_FAIL_TGT_SLUGGISH_NET) &&
2047 service_timeout < at_extra)
2048 service_timeout = at_extra;
2051 * We expect other clients to timeout within service_timeout, then try
2052 * to reconnect, then try the failover server. The max delay between
2053 * connect attempts is SWITCH_MAX + SWITCH_INC + INITIAL.
2055 service_timeout += 2 * INITIAL_CONNECT_TIMEOUT;
2057 LASSERT(obt->obt_magic == OBT_MAGIC);
2058 service_timeout += 2 * (CONNECTION_SWITCH_MAX + CONNECTION_SWITCH_INC);
2059 if (service_timeout > obd->obd_recovery_timeout && !new_client)
2060 extend_recovery_timer(obd, service_timeout, false);
2063 /** Health checking routines */
2064 static inline int exp_connect_healthy(struct obd_export *exp)
2066 return exp->exp_in_recovery;
2069 /** if export done req_replay or has replay in queue */
2070 static inline int exp_req_replay_healthy(struct obd_export *exp)
2072 return (!exp->exp_req_replay_needed ||
2073 atomic_read(&exp->exp_replay_count) > 0);
2077 static inline int exp_req_replay_healthy_or_from_mdt(struct obd_export *exp)
2079 return (exp_connect_flags(exp) & OBD_CONNECT_MDS_MDS) ||
2080 exp_req_replay_healthy(exp);
2083 /** if export done lock_replay or has replay in queue */
2084 static inline int exp_lock_replay_healthy(struct obd_export *exp)
2086 return (!exp->exp_lock_replay_needed ||
2087 atomic_read(&exp->exp_replay_count) > 0);
2090 static inline int exp_vbr_healthy(struct obd_export *exp)
2092 return !exp->exp_vbr_failed;
2095 static inline int exp_finished(struct obd_export *exp)
2097 return exp->exp_in_recovery && !exp->exp_lock_replay_needed;
2100 static inline int exp_finished_or_from_mdt(struct obd_export *exp)
2102 return (exp_connect_flags(exp) & OBD_CONNECT_MDS_MDS) ||
2106 static int check_for_next_transno(struct lu_target *lut)
2108 struct ptlrpc_request *req = NULL;
2109 struct obd_device *obd = lut->lut_obd;
2110 struct target_distribute_txn_data *tdtd = lut->lut_tdtd;
2111 int wake_up = 0, connected, completed, queue_len;
2112 __u64 req_transno = 0;
2113 __u64 update_transno = 0;
2114 __u64 next_transno = 0;
2118 spin_lock(&obd->obd_recovery_task_lock);
2119 if (!list_empty(&obd->obd_req_replay_queue)) {
2120 req = list_entry(obd->obd_req_replay_queue.next,
2121 struct ptlrpc_request, rq_list);
2122 req_transno = lustre_msg_get_transno(req->rq_reqmsg);
2125 if (!obd->obd_abort_recov_mdt && tdtd)
2126 update_transno = distribute_txn_get_next_transno(tdtd);
2128 connected = atomic_read(&obd->obd_connected_clients);
2129 completed = connected - atomic_read(&obd->obd_req_replay_clients);
2130 queue_len = obd->obd_requests_queued_for_recovery;
2131 next_transno = obd->obd_next_recovery_transno;
2134 "max: %d, connected: %d, completed: %d, queue_len: %d, req_transno: %llu, next_transno: %llu\n",
2135 atomic_read(&obd->obd_max_recoverable_clients),
2136 connected, completed,
2137 queue_len, req_transno, next_transno);
2139 if (obd->obd_abort_recovery) {
2140 CDEBUG(D_HA, "waking for aborted recovery\n");
2142 } else if (obd->obd_recovery_expired) {
2143 CDEBUG(D_HA, "waking for expired recovery\n");
2145 } else if (!obd->obd_abort_recov_mdt && tdtd && req &&
2146 is_req_replayed_by_update(req)) {
2147 LASSERTF(req_transno < next_transno,
2148 "req_transno %llu next_transno%llu\n", req_transno,
2150 CDEBUG(D_HA, "waking for duplicate req (%llu)\n",
2153 } else if (req_transno == next_transno ||
2154 (update_transno != 0 && update_transno <= next_transno)) {
2155 CDEBUG(D_HA, "waking for next (%lld)\n", next_transno);
2157 } else if (queue_len > 0 &&
2158 queue_len == atomic_read(&obd->obd_req_replay_clients)) {
2159 /** handle gaps occured due to lost reply or VBR */
2160 LASSERTF(req_transno >= next_transno,
2161 "req_transno: %llu, next_transno: %llu\n",
2162 req_transno, next_transno);
2164 "%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",
2165 obd->obd_name, obd->obd_version_recov ? "ON" : "OFF",
2166 next_transno, queue_len, completed, connected,
2167 req_transno, update_transno, obd->obd_last_committed);
2168 obd->obd_next_recovery_transno = req_transno;
2170 } else if (atomic_read(&obd->obd_req_replay_clients) == 0) {
2171 CDEBUG(D_HA, "waking for completed recovery\n");
2173 } else if (OBD_FAIL_CHECK(OBD_FAIL_MDS_RECOVERY_ACCEPTS_GAPS)) {
2175 "accepting transno gaps is explicitly allowed by fail_lock, waking up (%lld)\n",
2177 obd->obd_next_recovery_transno = req_transno;
2180 spin_unlock(&obd->obd_recovery_task_lock);
2184 static int check_for_next_lock(struct lu_target *lut)
2186 struct obd_device *obd = lut->lut_obd;
2189 spin_lock(&obd->obd_recovery_task_lock);
2190 if (!list_empty(&obd->obd_lock_replay_queue)) {
2191 CDEBUG(D_HA, "waking for next lock\n");
2193 } else if (atomic_read(&obd->obd_lock_replay_clients) == 0) {
2194 CDEBUG(D_HA, "waking for completed lock replay\n");
2196 } else if (obd->obd_abort_recovery) {
2197 CDEBUG(D_HA, "waking for aborted recovery\n");
2199 } else if (obd->obd_recovery_expired) {
2200 CDEBUG(D_HA, "waking for expired recovery\n");
2203 spin_unlock(&obd->obd_recovery_task_lock);
2208 static int check_update_llog(struct lu_target *lut)
2210 struct obd_device *obd = lut->lut_obd;
2211 struct target_distribute_txn_data *tdtd = lut->lut_tdtd;
2213 if (obd->obd_abort_recovery) {
2214 CDEBUG(D_HA, "waking for aborted recovery\n");
2218 if (atomic_read(&tdtd->tdtd_recovery_threads_count) == 0) {
2219 CDEBUG(D_HA, "waking for completion of reading update log\n");
2227 * wait for recovery events,
2228 * check its status with help of check_routine
2229 * evict dead clients via health_check
2231 static int target_recovery_overseer(struct lu_target *lut,
2232 int (*check_routine)(struct lu_target *),
2233 int (*health_check)(struct obd_export *))
2235 struct obd_device *obd = lut->lut_obd;
2236 struct target_distribute_txn_data *tdtd;
2240 if (obd->obd_recovering && obd->obd_recovery_start == 0) {
2241 now = ktime_get_seconds();
2242 if (now - last > 600) {
2243 LCONSOLE_INFO("%s: in recovery but waiting for the first client to connect\n",
2248 if (obd->obd_recovery_start != 0 && ktime_get_seconds() >=
2249 (obd->obd_recovery_start + obd->obd_recovery_time_hard)) {
2250 __u64 next_update_transno = 0;
2253 * Only abort the recovery if there are no update recovery
2256 spin_lock(&obd->obd_recovery_task_lock);
2257 if (!obd->obd_abort_recov_mdt && lut->lut_tdtd) {
2258 next_update_transno =
2259 distribute_txn_get_next_transno(lut->lut_tdtd);
2261 tdtd = lut->lut_tdtd;
2263 * If next_update_transno == 0, it probably because
2264 * updatelog retrieve threads did not get any records
2265 * yet, let's wait those threads stopped
2267 if (next_update_transno == 0) {
2268 spin_unlock(&obd->obd_recovery_task_lock);
2270 while (wait_event_timeout(
2271 tdtd->tdtd_recovery_threads_waitq,
2272 check_update_llog(lut),
2273 cfs_time_seconds(60)) == 0);
2275 spin_lock(&obd->obd_recovery_task_lock);
2276 next_update_transno =
2277 distribute_txn_get_next_transno(tdtd);
2281 if (next_update_transno != 0 && !obd->obd_abort_recovery) {
2282 obd->obd_next_recovery_transno = next_update_transno;
2283 spin_unlock(&obd->obd_recovery_task_lock);
2285 * Disconnect unfinished exports from clients, and
2286 * keep connection from MDT to make sure the update
2287 * recovery will still keep trying until some one
2288 * manually abort the recovery
2290 class_disconnect_stale_exports(obd,
2291 exp_finished_or_from_mdt);
2292 /* Abort all of replay & replay lock req from clients */
2293 abort_req_replay_queue(obd);
2294 abort_lock_replay_queue(obd);
2296 "%s: there are still update replay (%#llx)in the queue.\n",
2297 obd->obd_name, next_update_transno);
2299 obd->obd_abort_recovery = 1;
2300 spin_unlock(&obd->obd_recovery_task_lock);
2301 CWARN("%s recovery is aborted by hard timeout\n",
2306 while (wait_event_timeout(obd->obd_next_transno_waitq,
2308 cfs_time_seconds(60)) == 0)
2309 ; /* wait indefinitely for event, but don't trigger watchdog */
2311 if (obd->obd_abort_recovery) {
2312 CWARN("recovery is aborted, evict exports in recovery\n");
2313 if (lut->lut_tdtd != NULL) {
2314 tdtd = lut->lut_tdtd;
2316 * Let's wait all of the update log recovery thread
2320 tdtd->tdtd_recovery_threads_waitq,
2321 atomic_read(&tdtd->tdtd_recovery_threads_count)
2323 /* Then abort the update recovery list */
2324 dtrq_list_destroy(lut->lut_tdtd);
2327 /** evict exports which didn't finish recovery yet */
2328 class_disconnect_stale_exports(obd, exp_finished);
2330 } else if (obd->obd_recovery_expired) {
2331 obd->obd_recovery_expired = 0;
2333 /** If some clients died being recovered, evict them */
2334 LCONSOLE_WARN("%s: recovery is timed out, evict stale exports\n",
2336 /** evict cexports with no replay in queue, they are stalled */
2337 class_disconnect_stale_exports(obd, health_check);
2339 /** continue with VBR */
2340 spin_lock(&obd->obd_dev_lock);
2341 obd->obd_version_recov = 1;
2342 spin_unlock(&obd->obd_dev_lock);
2344 * reset timer, recovery will proceed with versions now,
2345 * timeout is set just to handle reconnection delays
2347 extend_recovery_timer(obd, RECONNECT_DELAY_MAX, true);
2349 * Wait for recovery events again, after evicting bad clients
2356 static struct ptlrpc_request *target_next_replay_lock(struct lu_target *lut)
2358 struct obd_device *obd = lut->lut_obd;
2359 struct ptlrpc_request *req = NULL;
2361 CDEBUG(D_HA, "Waiting for lock\n");
2362 if (target_recovery_overseer(lut, check_for_next_lock,
2363 exp_lock_replay_healthy))
2364 abort_lock_replay_queue(obd);
2366 spin_lock(&obd->obd_recovery_task_lock);
2367 if (!list_empty(&obd->obd_lock_replay_queue)) {
2368 req = list_entry(obd->obd_lock_replay_queue.next,
2369 struct ptlrpc_request, rq_list);
2370 list_del_init(&req->rq_list);
2371 spin_unlock(&obd->obd_recovery_task_lock);
2373 spin_unlock(&obd->obd_recovery_task_lock);
2374 LASSERT(list_empty(&obd->obd_lock_replay_queue));
2375 LASSERT(atomic_read(&obd->obd_lock_replay_clients) == 0);
2376 /** evict exports failed VBR */
2377 class_disconnect_stale_exports(obd, exp_vbr_healthy);
2382 static struct ptlrpc_request *target_next_final_ping(struct obd_device *obd)
2384 struct ptlrpc_request *req = NULL;
2386 spin_lock(&obd->obd_recovery_task_lock);
2387 if (!list_empty(&obd->obd_final_req_queue)) {
2388 req = list_entry(obd->obd_final_req_queue.next,
2389 struct ptlrpc_request, rq_list);
2390 list_del_init(&req->rq_list);
2391 spin_unlock(&obd->obd_recovery_task_lock);
2392 if (req->rq_export->exp_in_recovery) {
2393 spin_lock(&req->rq_export->exp_lock);
2394 req->rq_export->exp_in_recovery = 0;
2395 spin_unlock(&req->rq_export->exp_lock);
2398 spin_unlock(&obd->obd_recovery_task_lock);
2403 static void handle_recovery_req(struct ptlrpc_thread *thread,
2404 struct ptlrpc_request *req,
2405 svc_handler_t handler)
2410 * export can be evicted during recovery, no need to handle replays for
2411 * it after that, discard such request silently
2413 if (req->rq_export->exp_disconnected)
2416 req->rq_session.lc_thread = thread;
2417 req->rq_svc_thread = thread;
2418 req->rq_svc_thread->t_env->le_ses = &req->rq_session;
2420 /* thread context */
2421 lu_context_enter(&thread->t_env->le_ctx);
2423 lu_context_exit(&thread->t_env->le_ctx);
2425 req->rq_svc_thread->t_env->le_ses = NULL;
2427 /* don't reset timer for final stage */
2428 if (!exp_finished(req->rq_export)) {
2429 timeout_t timeout = obd_timeout;
2432 * Add request @timeout to the recovery time so next request from
2433 * this client may come in recovery time
2436 struct ptlrpc_service_part *svcpt;
2437 timeout_t est_timeout;
2439 svcpt = req->rq_rqbd->rqbd_svcpt;
2441 * If the server sent early reply for this request,
2442 * the client will recalculate the timeout according to
2443 * current server estimate service time, so we will
2444 * use the maxium timeout here for waiting the client
2445 * sending the next req
2447 est_timeout = at_get(&svcpt->scp_at_estimate);
2448 timeout = max_t(timeout_t, at_est2timeout(est_timeout),
2449 lustre_msg_get_timeout(req->rq_reqmsg));
2451 * Add 2 net_latency, one for balance rq_deadline
2452 * (see ptl_send_rpc), one for resend the req to server,
2453 * Note: client will pack net_latency in replay req
2454 * (see ptlrpc_replay_req)
2456 timeout += 2 * lustre_msg_get_service_timeout(req->rq_reqmsg);
2458 extend_recovery_timer(class_exp2obd(req->rq_export), timeout,
2464 /** Checking routines for recovery */
2465 static int check_for_recovery_ready(struct lu_target *lut)
2467 struct obd_device *obd = lut->lut_obd;
2468 unsigned int clnts = atomic_read(&obd->obd_connected_clients);
2471 "connected %d stale %d max_recoverable_clients %d abort %d expired %d\n",
2472 clnts, obd->obd_stale_clients,
2473 atomic_read(&obd->obd_max_recoverable_clients),
2474 obd->obd_abort_recovery, obd->obd_recovery_expired);
2476 if (!obd->obd_abort_recovery && !obd->obd_recovery_expired) {
2478 atomic_read(&obd->obd_max_recoverable_clients));
2479 if (clnts + obd->obd_stale_clients <
2480 atomic_read(&obd->obd_max_recoverable_clients))
2484 if (!obd->obd_abort_recov_mdt && lut->lut_tdtd != NULL) {
2485 if (!lut->lut_tdtd->tdtd_replay_ready &&
2486 !obd->obd_abort_recovery && !obd->obd_stopping) {
2488 * Let's extend recovery timer, in case the recovery
2489 * timer expired, and some clients got evicted
2491 extend_recovery_timer(obd, obd->obd_recovery_timeout,
2494 "%s update recovery is not ready, extend recovery %d\n",
2495 obd->obd_name, obd->obd_recovery_timeout);
2504 REQUEST_RECOVERY = 1,
2505 UPDATE_RECOVERY = 2,
2508 static __u64 get_next_replay_req_transno(struct obd_device *obd)
2512 if (!list_empty(&obd->obd_req_replay_queue)) {
2513 struct ptlrpc_request *req;
2515 req = list_entry(obd->obd_req_replay_queue.next,
2516 struct ptlrpc_request, rq_list);
2517 transno = lustre_msg_get_transno(req->rq_reqmsg);
2523 static __u64 get_next_transno(struct lu_target *lut, int *type)
2525 struct obd_device *obd = lut->lut_obd;
2526 struct target_distribute_txn_data *tdtd = lut->lut_tdtd;
2528 __u64 update_transno;
2532 transno = get_next_replay_req_transno(obd);
2534 *type = REQUEST_RECOVERY;
2536 if (!tdtd || obd->obd_abort_recov_mdt)
2539 update_transno = distribute_txn_get_next_transno(tdtd);
2540 if (transno == 0 || (transno >= update_transno &&
2541 update_transno != 0)) {
2542 transno = update_transno;
2544 *type = UPDATE_RECOVERY;
2551 * drop duplicate replay request
2553 * Because the operation has been replayed by update recovery, the request
2554 * with the same transno will be dropped and also notify the client to send
2555 * next replay request.
2557 * \param[in] env execution environment
2558 * \param[in] obd failover obd device
2559 * \param[in] req request to be dropped
2561 static void drop_duplicate_replay_req(struct lu_env *env,
2562 struct obd_device *obd,
2563 struct ptlrpc_request *req)
2565 DEBUG_REQ(D_HA, req,
2566 "remove t%lld from %s because duplicate update records found",
2567 lustre_msg_get_transno(req->rq_reqmsg),
2568 libcfs_nidstr(&req->rq_peer.nid));
2571 * Right now, only for MDS reint operation update replay and
2572 * normal request replay can have the same transno
2574 if (lustre_msg_get_opc(req->rq_reqmsg) == MDS_REINT) {
2575 req_capsule_set(&req->rq_pill, &RQF_MDS_REINT);
2576 req->rq_status = req_capsule_server_pack(&req->rq_pill);
2577 if (likely(req->rq_export))
2578 target_committed_to_req(req);
2579 lustre_msg_set_transno(req->rq_repmsg, req->rq_transno);
2580 target_send_reply(req, req->rq_status, 0);
2582 DEBUG_REQ(D_ERROR, req, "wrong opc from %s",
2583 libcfs_nidstr(&req->rq_peer.nid));
2585 target_exp_dequeue_req_replay(req);
2586 target_request_copy_put(req);
2587 obd->obd_replayed_requests++;
2590 #define WATCHDOG_TIMEOUT (obd_timeout * 10)
2592 static void replay_request_or_update(struct lu_env *env,
2593 struct lu_target *lut,
2594 struct target_recovery_data *trd,
2595 struct ptlrpc_thread *thread)
2597 struct obd_device *obd = lut->lut_obd;
2598 struct ptlrpc_request *req = NULL;
2604 CDEBUG(D_HA, "Waiting for transno %lld\n",
2605 obd->obd_next_recovery_transno);
2607 /* Replay all of request and update by transno */
2609 struct target_distribute_txn_data *tdtd = lut->lut_tdtd;
2611 CFS_FAIL_TIMEOUT(OBD_FAIL_TGT_REPLAY_DELAY2, cfs_fail_val);
2614 * It is needed to extend recovery window above
2615 * recovery_time_soft. Extending is possible only in the
2616 * end of recovery window (see more details in
2617 * handle_recovery_req()).
2619 CFS_FAIL_TIMEOUT_MS(OBD_FAIL_TGT_REPLAY_DELAY, 300);
2621 if (target_recovery_overseer(lut, check_for_next_transno,
2622 exp_req_replay_healthy_or_from_mdt)) {
2623 abort_req_replay_queue(obd);
2624 abort_lock_replay_queue(obd);
2628 spin_lock(&obd->obd_recovery_task_lock);
2629 transno = get_next_transno(lut, &type);
2630 if (type == REQUEST_RECOVERY && transno != 0) {
2632 * Drop replay request from client side, if the
2633 * replay has been executed by update with the
2636 req = list_entry(obd->obd_req_replay_queue.next,
2637 struct ptlrpc_request, rq_list);
2639 list_del_init(&req->rq_list);
2640 obd->obd_requests_queued_for_recovery--;
2641 spin_unlock(&obd->obd_recovery_task_lock);
2644 * Let's check if the request has been redone by
2647 if (is_req_replayed_by_update(req)) {
2648 struct distribute_txn_replay_req *dtrq;
2650 dtrq = distribute_txn_lookup_finish_list(tdtd,
2652 LASSERT(dtrq != NULL);
2653 spin_lock(&tdtd->tdtd_replay_list_lock);
2654 list_del_init(&dtrq->dtrq_list);
2655 spin_unlock(&tdtd->tdtd_replay_list_lock);
2658 drop_duplicate_replay_req(env, obd, req);
2663 LASSERT(trd->trd_processing_task == current->pid);
2664 DEBUG_REQ(D_HA, req, "processing x%llu t%lld from %s",
2666 lustre_msg_get_transno(req->rq_reqmsg),
2667 libcfs_nidstr(&req->rq_peer.nid));
2669 ptlrpc_watchdog_init(&thread->t_watchdog,
2671 handle_recovery_req(thread, req,
2672 trd->trd_recovery_handler);
2673 ptlrpc_watchdog_disable(&thread->t_watchdog);
2676 * bz18031: increase next_recovery_transno before
2677 * target_request_copy_put() will drop exp_rpc reference
2679 spin_lock(&obd->obd_recovery_task_lock);
2680 obd->obd_next_recovery_transno++;
2681 spin_unlock(&obd->obd_recovery_task_lock);
2682 target_exp_dequeue_req_replay(req);
2683 target_request_copy_put(req);
2684 obd->obd_replayed_requests++;
2685 } else if (type == UPDATE_RECOVERY && transno != 0) {
2686 struct distribute_txn_replay_req *dtrq;
2689 spin_unlock(&obd->obd_recovery_task_lock);
2691 LASSERT(tdtd != NULL);
2692 dtrq = distribute_txn_get_next_req(tdtd);
2693 lu_context_enter(&thread->t_env->le_ctx);
2694 ptlrpc_watchdog_init(&thread->t_watchdog,
2696 rc = tdtd->tdtd_replay_handler(env, tdtd, dtrq);
2697 ptlrpc_watchdog_disable(&thread->t_watchdog);
2698 lu_context_exit(&thread->t_env->le_ctx);
2699 extend_recovery_timer(obd, obd_timeout, true);
2701 if (rc == 0 && dtrq->dtrq_xid != 0) {
2703 "Move x%llu t%llu to finish list\n",
2705 dtrq->dtrq_master_transno);
2707 /* Add it to the replay finish list */
2708 spin_lock(&tdtd->tdtd_replay_list_lock);
2709 list_add(&dtrq->dtrq_list,
2710 &tdtd->tdtd_replay_finish_list);
2711 spin_unlock(&tdtd->tdtd_replay_list_lock);
2713 spin_lock(&obd->obd_recovery_task_lock);
2714 if (transno == obd->obd_next_recovery_transno)
2715 obd->obd_next_recovery_transno++;
2717 obd->obd_next_recovery_transno)
2718 obd->obd_next_recovery_transno =
2720 spin_unlock(&obd->obd_recovery_task_lock);
2725 spin_unlock(&obd->obd_recovery_task_lock);
2727 LASSERT(list_empty(&obd->obd_req_replay_queue));
2728 LASSERT(atomic_read(&obd->obd_req_replay_clients) == 0);
2729 /** evict exports failed VBR */
2730 class_disconnect_stale_exports(obd, exp_vbr_healthy);
2736 static int target_recovery_thread(void *arg)
2738 struct lu_target *lut = arg;
2739 struct obd_device *obd = lut->lut_obd;
2740 struct ptlrpc_request *req;
2741 struct target_recovery_data *trd = &obd->obd_recovery_data;
2742 unsigned long delta;
2744 struct ptlrpc_thread *thread = NULL;
2748 unshare_fs_struct();
2749 OBD_ALLOC_PTR(thread);
2755 GOTO(out_thread, rc = -ENOMEM);
2756 rc = lu_env_add(env);
2760 rc = lu_context_init(&env->le_ctx, LCT_MD_THREAD | LCT_DT_THREAD);
2762 GOTO(out_env_remove, rc);
2764 thread->t_env = env;
2765 thread->t_id = -1; /* force filter_iobuf_get/put to use local buffers */
2766 thread->t_task = current;
2767 env->le_ctx.lc_thread = thread;
2768 tgt_io_thread_init(thread); /* init thread_big_cache for IO requests */
2770 CDEBUG(D_HA, "%s: started recovery thread pid %d\n", obd->obd_name,
2772 trd->trd_processing_task = current->pid;
2774 spin_lock(&obd->obd_dev_lock);
2775 obd->obd_recovering = 1;
2776 spin_unlock(&obd->obd_dev_lock);
2777 complete(&trd->trd_starting);
2779 /* first of all, we have to know the first transno to replay */
2780 if (target_recovery_overseer(lut, check_for_recovery_ready,
2781 exp_connect_healthy)) {
2782 abort_req_replay_queue(obd);
2783 abort_lock_replay_queue(obd);
2784 if (lut->lut_tdtd != NULL)
2785 dtrq_list_destroy(lut->lut_tdtd);
2788 /* next stage: replay requests or update */
2790 CDEBUG(D_INFO, "1: request replay stage - %d clients from t%llu\n",
2791 atomic_read(&obd->obd_req_replay_clients),
2792 obd->obd_next_recovery_transno);
2793 replay_request_or_update(env, lut, trd, thread);
2796 * The second stage: replay locks
2798 CDEBUG(D_INFO, "2: lock replay stage - %d clients\n",
2799 atomic_read(&obd->obd_lock_replay_clients));
2800 while ((req = target_next_replay_lock(lut))) {
2801 LASSERT(trd->trd_processing_task == current->pid);
2802 DEBUG_REQ(D_HA, req, "processing lock from %s:",
2803 libcfs_nidstr(&req->rq_peer.nid));
2804 if (OBD_FAIL_CHECK(OBD_FAIL_LDLM_LOCK_REPLAY)) {
2805 req->rq_status = -ENODEV;
2806 target_request_copy_put(req);
2809 handle_recovery_req(thread, req,
2810 trd->trd_recovery_handler);
2811 target_request_copy_put(req);
2812 obd->obd_replayed_locks++;
2816 * The third stage: reply on final pings, at this moment all clients
2817 * must have request in final queue
2819 CFS_FAIL_TIMEOUT(OBD_FAIL_TGT_REPLAY_RECONNECT, cfs_fail_val);
2820 CDEBUG(D_INFO, "3: final stage - process recovery completion pings\n");
2821 /** Update server last boot epoch */
2822 tgt_boot_epoch_update(lut);
2824 * We drop recoverying flag to forward all new requests
2825 * to regular mds_handle() since now
2827 spin_lock(&obd->obd_dev_lock);
2828 obd->obd_recovering = obd->obd_abort_recovery = 0;
2829 obd->obd_abort_recov_mdt = 0;
2830 spin_unlock(&obd->obd_dev_lock);
2831 spin_lock(&obd->obd_recovery_task_lock);
2832 target_cancel_recovery_timer(obd);
2833 spin_unlock(&obd->obd_recovery_task_lock);
2834 while ((req = target_next_final_ping(obd))) {
2835 LASSERT(trd->trd_processing_task == current->pid);
2836 DEBUG_REQ(D_HA, req, "processing final ping from %s:",
2837 libcfs_nidstr(&req->rq_peer.nid));
2838 handle_recovery_req(thread, req,
2839 trd->trd_recovery_handler);
2841 * Because the waiting client can not send ping to server,
2842 * so we need refresh the last_request_time, to avoid the
2843 * export is being evicted
2845 ptlrpc_update_export_timer(req->rq_export, 0);
2846 target_request_copy_put(req);
2849 delta = jiffies_to_msecs(jiffies - delta) / MSEC_PER_SEC;
2850 CDEBUG(D_INFO, "4: recovery completed in %lus - %d/%d reqs/locks\n",
2851 delta, obd->obd_replayed_requests, obd->obd_replayed_locks);
2852 if (delta > OBD_RECOVERY_TIME_SOFT) {
2853 CWARN("too long recovery - read logs\n");
2854 libcfs_debug_dumplog();
2857 target_finish_recovery(lut);
2858 lu_context_fini(&env->le_ctx);
2859 trd->trd_processing_task = 0;
2860 complete_all(&trd->trd_finishing);
2861 tgt_io_thread_done(thread);
2867 OBD_FREE_PTR(thread);
2871 static int target_start_recovery_thread(struct lu_target *lut,
2872 svc_handler_t handler)
2874 struct obd_device *obd = lut->lut_obd;
2876 struct target_recovery_data *trd = &obd->obd_recovery_data;
2879 memset(trd, 0, sizeof(*trd));
2880 init_completion(&trd->trd_starting);
2881 init_completion(&trd->trd_finishing);
2882 trd->trd_recovery_handler = handler;
2884 rc = server_name2index(obd->obd_name, &index, NULL);
2888 if (!IS_ERR(kthread_run(target_recovery_thread,
2889 lut, "tgt_recover_%d", index))) {
2890 wait_for_completion(&trd->trd_starting);
2891 LASSERT(obd->obd_recovering != 0);
2899 void target_stop_recovery_thread(struct obd_device *obd)
2901 if (obd->obd_recovery_data.trd_processing_task > 0) {
2902 struct target_recovery_data *trd = &obd->obd_recovery_data;
2903 /** recovery can be done but postrecovery is not yet */
2904 spin_lock(&obd->obd_dev_lock);
2905 if (obd->obd_recovering) {
2906 CERROR("%s: Aborting recovery\n", obd->obd_name);
2907 obd->obd_abort_recovery = 1;
2908 wake_up(&obd->obd_next_transno_waitq);
2910 spin_unlock(&obd->obd_dev_lock);
2911 wait_for_completion(&trd->trd_finishing);
2914 EXPORT_SYMBOL(target_stop_recovery_thread);
2916 void target_recovery_fini(struct obd_device *obd)
2918 class_disconnect_exports(obd);
2919 target_stop_recovery_thread(obd);
2920 target_cleanup_recovery(obd);
2922 EXPORT_SYMBOL(target_recovery_fini);
2924 static enum hrtimer_restart target_recovery_expired(struct hrtimer *timer)
2926 struct obd_device *obd = container_of(timer, struct obd_device,
2927 obd_recovery_timer);
2930 "%s: recovery timed out; %d clients are still in recovery after %llu seconds (%d clients connected)\n",
2931 obd->obd_name, atomic_read(&obd->obd_lock_replay_clients),
2932 ktime_get_seconds() - obd->obd_recovery_start,
2933 atomic_read(&obd->obd_connected_clients));
2935 obd->obd_recovery_expired = 1;
2936 wake_up(&obd->obd_next_transno_waitq);
2937 return HRTIMER_NORESTART;
2940 void target_recovery_init(struct lu_target *lut, svc_handler_t handler)
2942 struct obd_device *obd = lut->lut_obd;
2944 if (lut->lut_bottom->dd_rdonly)
2947 if (atomic_read(&obd->obd_max_recoverable_clients) == 0) {
2948 /** Update server last boot epoch */
2949 tgt_boot_epoch_update(lut);
2953 CDEBUG(D_HA, "RECOVERY: service %s, %d recoverable clients, "
2954 "last_transno %llu\n", obd->obd_name,
2955 atomic_read(&obd->obd_max_recoverable_clients),
2956 obd->obd_last_committed);
2957 LASSERT(obd->obd_stopping == 0);
2958 obd->obd_next_recovery_transno = obd->obd_last_committed + 1;
2959 obd->obd_recovery_start = 0;
2960 obd->obd_recovery_end = 0;
2962 hrtimer_init(&obd->obd_recovery_timer, CLOCK_MONOTONIC,
2964 obd->obd_recovery_timer.function = &target_recovery_expired;
2965 target_start_recovery_thread(lut, handler);
2967 EXPORT_SYMBOL(target_recovery_init);
2969 static int target_process_req_flags(struct obd_device *obd,
2970 struct ptlrpc_request *req)
2972 struct obd_export *exp = req->rq_export;
2974 LASSERT(exp != NULL);
2975 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REQ_REPLAY_DONE) {
2976 /* client declares he's ready to replay locks */
2977 spin_lock(&exp->exp_lock);
2978 if (exp->exp_req_replay_needed) {
2979 exp->exp_req_replay_needed = 0;
2980 spin_unlock(&exp->exp_lock);
2982 LASSERT_ATOMIC_POS(&obd->obd_req_replay_clients);
2983 atomic_dec(&obd->obd_req_replay_clients);
2985 spin_unlock(&exp->exp_lock);
2988 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_LOCK_REPLAY_DONE) {
2990 * client declares he's ready to complete recovery
2991 * so, we put the request on th final queue
2993 spin_lock(&exp->exp_lock);
2994 if (exp->exp_lock_replay_needed) {
2995 exp->exp_lock_replay_needed = 0;
2996 spin_unlock(&exp->exp_lock);
2998 LASSERT_ATOMIC_POS(&obd->obd_lock_replay_clients);
2999 atomic_dec(&obd->obd_lock_replay_clients);
3001 spin_unlock(&exp->exp_lock);
3007 int target_queue_recovery_request(struct ptlrpc_request *req,
3008 struct obd_device *obd)
3010 __u64 transno = lustre_msg_get_transno(req->rq_reqmsg);
3011 struct ptlrpc_request *reqiter;
3016 if (obd->obd_recovery_data.trd_processing_task == current->pid) {
3017 /* Processing the queue right now, don't re-add. */
3021 target_process_req_flags(obd, req);
3023 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_LOCK_REPLAY_DONE) {
3024 if (unlikely(OBD_FAIL_CHECK(OBD_FAIL_TGT_RECOVERY_REQ_RACE))) {
3025 if (cfs_fail_val == 1) {
3028 wake_up(&cfs_race_waitq);
3030 schedule_timeout_interruptible(
3031 cfs_time_seconds(1));
3036 * client declares he's ready to complete recovery
3037 * so, we put the request on th final queue
3039 target_request_copy_get(req);
3040 DEBUG_REQ(D_HA, req, "queue final req");
3041 wake_up(&obd->obd_next_transno_waitq);
3042 spin_lock(&obd->obd_recovery_task_lock);
3043 if (obd->obd_recovering) {
3044 struct ptlrpc_request *tmp;
3045 struct ptlrpc_request *duplicate = NULL;
3047 if (likely(!req->rq_export->exp_replay_done)) {
3048 req->rq_export->exp_replay_done = 1;
3049 list_add_tail(&req->rq_list,
3050 &obd->obd_final_req_queue);
3051 spin_unlock(&obd->obd_recovery_task_lock);
3056 * XXX O(n), but only happens if final ping is
3057 * timed out, probably reorganize the list as
3060 list_for_each_entry_safe(reqiter, tmp,
3061 &obd->obd_final_req_queue,
3063 if (reqiter->rq_export == req->rq_export) {
3064 list_del_init(&reqiter->rq_list);
3065 duplicate = reqiter;
3070 list_add_tail(&req->rq_list,
3071 &obd->obd_final_req_queue);
3072 req->rq_export->exp_replay_done = 1;
3073 spin_unlock(&obd->obd_recovery_task_lock);
3075 if (duplicate != NULL) {
3076 DEBUG_REQ(D_HA, duplicate,
3077 "put prev final req");
3078 target_request_copy_put(duplicate);
3082 spin_unlock(&obd->obd_recovery_task_lock);
3083 target_request_copy_put(req);
3084 RETURN(obd->obd_stopping ? -ENOTCONN : 1);
3087 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REQ_REPLAY_DONE) {
3088 /* client declares he's ready to replay locks */
3089 target_request_copy_get(req);
3090 DEBUG_REQ(D_HA, req, "queue lock replay req");
3091 wake_up(&obd->obd_next_transno_waitq);
3092 spin_lock(&obd->obd_recovery_task_lock);
3093 LASSERT(obd->obd_recovering);
3094 /* usually due to recovery abort */
3095 if (!req->rq_export->exp_in_recovery) {
3096 spin_unlock(&obd->obd_recovery_task_lock);
3097 target_request_copy_put(req);
3100 LASSERT(req->rq_export->exp_lock_replay_needed);
3101 list_add_tail(&req->rq_list, &obd->obd_lock_replay_queue);
3102 spin_unlock(&obd->obd_recovery_task_lock);
3107 * CAVEAT EMPTOR: The incoming request message has been swabbed
3108 * (i.e. buflens etc are in my own byte order), but type-dependent
3109 * buffers (eg mdt_body, ost_body etc) have NOT been swabbed.
3113 INIT_LIST_HEAD(&req->rq_list);
3114 DEBUG_REQ(D_HA, req, "not queueing");
3119 * If we're processing the queue, we want don't want to queue this
3122 * Also, if this request has a transno less than the one we're waiting
3123 * for, we should process it now. It could (and currently always will)
3124 * be an open request for a descriptor that was opened some time ago.
3126 * Also, a resent, replayed request that has already been
3127 * handled will pass through here and be processed immediately.
3130 "Next recovery transno: %llu, current: %llu, replaying\n",
3131 obd->obd_next_recovery_transno, transno);
3134 * If the request has been replayed by update replay, then sends this
3135 * request to the recovery thread (replay_request_or_update()), where
3136 * it will be handled
3138 spin_lock(&obd->obd_recovery_task_lock);
3139 if (transno < obd->obd_next_recovery_transno &&
3140 !is_req_replayed_by_update(req)) {
3141 /* Processing the queue right now, don't re-add. */
3142 LASSERT(list_empty(&req->rq_list));
3143 spin_unlock(&obd->obd_recovery_task_lock);
3146 spin_unlock(&obd->obd_recovery_task_lock);
3148 if (OBD_FAIL_CHECK(OBD_FAIL_TGT_REPLAY_DROP))
3151 target_request_copy_get(req);
3152 if (!req->rq_export->exp_in_recovery) {
3153 target_request_copy_put(req);
3156 LASSERT(req->rq_export->exp_req_replay_needed);
3158 if (target_exp_enqueue_req_replay(req)) {
3159 DEBUG_REQ(D_ERROR, req, "dropping resent queued req");
3160 target_request_copy_put(req);
3165 spin_lock(&obd->obd_recovery_task_lock);
3166 LASSERT(obd->obd_recovering);
3167 list_for_each_entry(reqiter, &obd->obd_req_replay_queue, rq_list) {
3168 if (lustre_msg_get_transno(reqiter->rq_reqmsg) > transno) {
3169 list_add_tail(&req->rq_list, &reqiter->rq_list);
3174 if (unlikely(lustre_msg_get_transno(reqiter->rq_reqmsg) ==
3176 DEBUG_REQ(D_ERROR, req,
3177 "dropping replay: transno has been claimed by another client");
3178 spin_unlock(&obd->obd_recovery_task_lock);
3179 target_exp_dequeue_req_replay(req);
3180 target_request_copy_put(req);
3186 list_add_tail(&req->rq_list, &obd->obd_req_replay_queue);
3188 obd->obd_requests_queued_for_recovery++;
3189 spin_unlock(&obd->obd_recovery_task_lock);
3190 wake_up(&obd->obd_next_transno_waitq);
3194 void target_committed_to_req(struct ptlrpc_request *req)
3196 struct obd_export *exp = req->rq_export;
3198 if (!exp->exp_obd->obd_no_transno && req->rq_repmsg != NULL)
3199 lustre_msg_set_last_committed(req->rq_repmsg,
3200 exp->exp_last_committed);
3202 DEBUG_REQ(D_IOCTL, req,
3203 "not sending last_committed update (%d/%d)",
3204 exp->exp_obd->obd_no_transno,
3205 req->rq_repmsg == NULL);
3207 CDEBUG(D_INFO, "last_committed %llu, transno %llu, xid %llu\n",
3208 exp->exp_last_committed, req->rq_transno, req->rq_xid);
3211 #endif /* HAVE_SERVER_SUPPORT */
3214 * Packs current SLV and Limit into \a req.
3216 int target_pack_pool_reply(struct ptlrpc_request *req)
3218 struct obd_device *obd;
3223 * Check that we still have all structures alive as this may
3224 * be some late RPC at shutdown time.
3226 if (unlikely(!req->rq_export || !req->rq_export->exp_obd ||
3227 !exp_connect_lru_resize(req->rq_export))) {
3228 lustre_msg_set_slv(req->rq_repmsg, 0);
3229 lustre_msg_set_limit(req->rq_repmsg, 0);
3233 /* OBD is alive here as export is alive, which we checked above. */
3234 obd = req->rq_export->exp_obd;
3236 read_lock(&obd->obd_pool_lock);
3237 lustre_msg_set_slv(req->rq_repmsg, obd->obd_pool_slv);
3238 lustre_msg_set_limit(req->rq_repmsg, obd->obd_pool_limit);
3239 read_unlock(&obd->obd_pool_lock);
3244 static int target_send_reply_msg(struct ptlrpc_request *req,
3245 int rc, int fail_id)
3247 if (OBD_FAIL_CHECK_ORSET(fail_id & ~OBD_FAIL_ONCE, OBD_FAIL_ONCE)) {
3248 DEBUG_REQ(D_ERROR, req, "dropping reply");
3252 * We can have a null rq_reqmsg in the event of bad signature or
3253 * no context when unwrapping
3255 if (req->rq_reqmsg &&
3256 unlikely(lustre_msg_get_opc(req->rq_reqmsg) == MDS_REINT &&
3257 OBD_FAIL_CHECK(OBD_FAIL_MDS_REINT_MULTI_NET_REP)))
3261 DEBUG_REQ(D_NET, req, "processing error (%d)", rc);
3262 req->rq_status = rc;
3263 return ptlrpc_send_error(req, 1);
3265 DEBUG_REQ(D_NET, req, "sending reply");
3267 return ptlrpc_send_reply(req, PTLRPC_REPLY_MAYBE_DIFFICULT);
3270 void target_send_reply(struct ptlrpc_request *req, int rc, int fail_id)
3272 struct ptlrpc_service_part *svcpt;
3274 struct ptlrpc_reply_state *rs;
3275 struct obd_export *exp;
3279 if (req->rq_no_reply) {
3284 svcpt = req->rq_rqbd->rqbd_svcpt;
3285 rs = req->rq_reply_state;
3286 if (rs == NULL || !rs->rs_difficult) {
3288 target_send_reply_msg(req, rc, fail_id);
3293 /* must be an export if locks saved */
3294 LASSERT(req->rq_export != NULL);
3295 /* req/reply consistent */
3296 LASSERT(rs->rs_svcpt == svcpt);
3299 LASSERT(!rs->rs_scheduled);
3300 LASSERT(!rs->rs_scheduled_ever);
3301 LASSERT(!rs->rs_handled);
3302 LASSERT(!rs->rs_sent);
3303 LASSERT(!rs->rs_unlinked);
3304 LASSERT(rs->rs_export == NULL);
3305 LASSERT(list_empty(&rs->rs_obd_list));
3306 LASSERT(list_empty(&rs->rs_exp_list));
3308 exp = class_export_get(req->rq_export);
3310 /* disable reply scheduling while I'm setting up */
3311 rs->rs_scheduled = 1;
3313 rs->rs_unlinked = 0;
3314 rs->rs_xid = req->rq_xid;
3315 rs->rs_transno = req->rq_transno;
3316 rs->rs_export = exp;
3317 rs->rs_opc = lustre_msg_get_opc(req->rq_reqmsg);
3319 spin_lock(&exp->exp_uncommitted_replies_lock);
3320 CDEBUG(D_NET, "rs transno = %llu, last committed = %llu\n",
3321 rs->rs_transno, exp->exp_last_committed);
3322 if (rs->rs_transno > exp->exp_last_committed) {
3323 /* not committed already */
3324 list_add_tail(&rs->rs_obd_list,
3325 &exp->exp_uncommitted_replies);
3327 spin_unlock(&exp->exp_uncommitted_replies_lock);
3329 spin_lock(&exp->exp_lock);
3330 list_add_tail(&rs->rs_exp_list, &exp->exp_outstanding_replies);
3331 spin_unlock(&exp->exp_lock);
3333 netrc = target_send_reply_msg(req, rc, fail_id);
3335 spin_lock(&svcpt->scp_rep_lock);
3337 atomic_inc(&svcpt->scp_nreps_difficult);
3341 * error sending: reply is off the net. Also we need +1
3342 * reply ref until ptlrpc_handle_rs() is done
3343 * with the reply state (if the send was successful, there
3344 * would have been +1 ref for the net, which
3345 * reply_out_callback leaves alone)
3348 rs->rs_unlinked = 1;
3349 ptlrpc_rs_addref(rs);
3352 spin_lock(&rs->rs_lock);
3353 if (rs->rs_transno <= exp->exp_last_committed ||
3354 (rs->rs_unlinked && !rs->rs_no_ack) ||
3355 list_empty(&rs->rs_exp_list) || /* completed already */
3356 list_empty(&rs->rs_obd_list)) {
3357 CDEBUG(D_HA, "Schedule reply immediately\n");
3358 ptlrpc_dispatch_difficult_reply(rs);
3360 list_add(&rs->rs_list, &svcpt->scp_rep_active);
3361 rs->rs_scheduled = 0; /* allow notifier to schedule */
3363 spin_unlock(&rs->rs_lock);
3364 spin_unlock(&svcpt->scp_rep_lock);
3368 enum ldlm_mode lck_compat_array[] = {
3369 [LCK_EX] = LCK_COMPAT_EX,
3370 [LCK_PW] = LCK_COMPAT_PW,
3371 [LCK_PR] = LCK_COMPAT_PR,
3372 [LCK_CW] = LCK_COMPAT_CW,
3373 [LCK_CR] = LCK_COMPAT_CR,
3374 [LCK_NL] = LCK_COMPAT_NL,
3375 [LCK_GROUP] = LCK_COMPAT_GROUP,
3376 [LCK_COS] = LCK_COMPAT_COS,
3380 * Rather arbitrary mapping from LDLM error codes to errno values. This should
3381 * not escape to the user level.
3383 int ldlm_error2errno(enum ldlm_error error)
3389 case ELDLM_LOCK_MATCHED:
3392 case ELDLM_LOCK_CHANGED:
3395 case ELDLM_LOCK_ABORTED:
3398 case ELDLM_LOCK_REPLACED:
3401 case ELDLM_NO_LOCK_DATA:
3404 case ELDLM_NAMESPACE_EXISTS:
3407 case ELDLM_BAD_NAMESPACE:
3411 if (((int)error) < 0) { /* cast to signed type */
3412 result = error; /* as ldlm_error can be unsigned */
3414 CERROR("Invalid DLM result code: %d\n", error);
3420 EXPORT_SYMBOL(ldlm_error2errno);
3423 * Dual to ldlm_error2errno(): maps errno values back to enum ldlm_error.
3425 enum ldlm_error ldlm_errno2error(int err_no)
3434 error = ELDLM_LOCK_CHANGED;
3437 error = ELDLM_LOCK_ABORTED;
3440 error = ELDLM_LOCK_REPLACED;
3443 error = ELDLM_NO_LOCK_DATA;
3446 error = ELDLM_NAMESPACE_EXISTS;
3449 error = ELDLM_BAD_NAMESPACE;
3457 #if LUSTRE_TRACKS_LOCK_EXP_REFS
3458 void ldlm_dump_export_locks(struct obd_export *exp)
3460 spin_lock(&exp->exp_locks_list_guard);
3461 if (!list_empty(&exp->exp_locks_list)) {
3462 struct ldlm_lock *lock;
3464 CERROR("dumping locks for export %p, ignore if the unmount doesn't hang\n",
3466 list_for_each_entry(lock, &exp->exp_locks_list,
3468 LDLM_ERROR(lock, "lock:");
3470 spin_unlock(&exp->exp_locks_list_guard);
3474 #ifdef HAVE_SERVER_SUPPORT
3475 static inline const char *bulk2type(struct ptlrpc_request *req)
3477 if (req->rq_bulk_read)
3479 if (req->rq_bulk_write)
3484 int target_bulk_io(struct obd_export *exp, struct ptlrpc_bulk_desc *desc)
3486 struct ptlrpc_request *req = desc->bd_req;
3487 time64_t start = ktime_get_seconds();
3493 /* If there is eviction in progress, wait for it to finish. */
3495 exp->exp_obd->obd_evict_inprogress_waitq,
3496 !atomic_read(&exp->exp_obd->obd_evict_inprogress));
3498 /* Check if client was evicted or reconnected already. */
3499 if (exp->exp_failed ||
3500 exp->exp_conn_cnt > lustre_msg_get_conn_cnt(req->rq_reqmsg)) {
3503 if (req->rq_bulk_read)
3504 rc = sptlrpc_svc_wrap_bulk(req, desc);
3506 if (OCD_HAS_FLAG(&exp->exp_connect_data, BULK_MBITS))
3507 req->rq_mbits = lustre_msg_get_mbits(req->rq_reqmsg);
3508 else /* old version, bulk matchbits is rq_xid */
3509 req->rq_mbits = req->rq_xid;
3512 rc = ptlrpc_start_bulk_transfer(desc);
3516 DEBUG_REQ(D_ERROR, req, "bulk %s failed: rc = %d",
3517 bulk2type(req), rc);
3521 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_SENDPAGE)) {
3522 ptlrpc_abort_bulk(desc);
3526 /* limit actual bulk transfer to bulk_timeout seconds */
3527 deadline = start + bulk_timeout;
3528 if (deadline > req->rq_deadline)
3529 deadline = req->rq_deadline;
3532 time64_t timeoutl = deadline - ktime_get_seconds();
3533 time64_t rq_deadline;
3535 while (timeoutl >= 0 &&
3536 wait_event_idle_timeout(
3538 !ptlrpc_server_bulk_active(desc) ||
3541 lustre_msg_get_conn_cnt(req->rq_reqmsg),
3542 timeoutl ? cfs_time_seconds(1) : 1) == 0)
3544 rc = timeoutl < 0 ? -ETIMEDOUT : 0;
3546 /* Wait again if we changed rq_deadline. */
3547 rq_deadline = READ_ONCE(req->rq_deadline);
3548 deadline = start + bulk_timeout;
3549 if (deadline > rq_deadline)
3550 deadline = rq_deadline;
3551 } while (rc == -ETIMEDOUT &&
3552 deadline > ktime_get_seconds());
3554 if (rc == -ETIMEDOUT) {
3555 DEBUG_REQ(D_ERROR, req, "timeout on bulk %s after %lld%+llds",
3556 bulk2type(req), deadline - start,
3557 ktime_get_real_seconds() - deadline);
3558 ptlrpc_abort_bulk(desc);
3559 } else if (exp->exp_failed) {
3560 DEBUG_REQ(D_ERROR, req, "Eviction on bulk %s",
3563 ptlrpc_abort_bulk(desc);
3564 } else if (exp->exp_conn_cnt >
3565 lustre_msg_get_conn_cnt(req->rq_reqmsg)) {
3566 DEBUG_REQ(D_ERROR, req, "Reconnect on bulk %s",
3568 /* We don't reply anyway. */
3570 ptlrpc_abort_bulk(desc);
3571 } else if (desc->bd_failure) {
3572 DEBUG_REQ(D_ERROR, req, "network error on bulk %s",
3574 /* XXX should this be a different errno? */
3577 if (req->rq_bulk_write)
3578 rc = sptlrpc_svc_unwrap_bulk(req, desc);
3579 if (rc == 0 && desc->bd_nob_transferred != desc->bd_nob) {
3580 DEBUG_REQ(D_ERROR, req, "truncated bulk %s %d(%d)",
3581 bulk2type(req), desc->bd_nob_transferred,
3583 /* XXX should this be a different errno? */
3590 EXPORT_SYMBOL(target_bulk_io);
3592 #endif /* HAVE_SERVER_SUPPORT */