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 <uapi/linux/lustre/lustre_ioctl.h>
53 #include "ldlm_internal.h"
56 * @priority: If non-zero, move the selected connection to the list head.
57 * @create: If zero, only search in existing connections.
59 static int import_set_conn(struct obd_import *imp, struct obd_uuid *uuid,
60 int priority, int create)
62 struct ptlrpc_connection *ptlrpc_conn;
63 struct obd_import_conn *imp_conn = NULL, *item;
64 u32 refnet = imp->imp_conn_restricted_net;
69 if (!create && !priority) {
70 CDEBUG(D_HA, "Nothing to do\n");
74 /* refnet is used to restrict network connections */
75 if (refnet != LNET_NET_ANY)
76 CDEBUG(D_HA, "imp %s: restrict %s to %s net\n",
77 imp->imp_obd->obd_name, uuid->uuid,
78 libcfs_net2str(refnet));
80 ptlrpc_conn = ptlrpc_uuid_to_connection(uuid, refnet);
82 CDEBUG(D_HA, "can't find connection %s\n", uuid->uuid);
87 OBD_ALLOC(imp_conn, sizeof(*imp_conn));
89 GOTO(out_put, rc = -ENOMEM);
92 spin_lock(&imp->imp_lock);
93 list_for_each_entry(item, &imp->imp_conn_list, oic_item) {
94 if (obd_uuid_equals(uuid, &item->oic_uuid)) {
96 list_move(&item->oic_item,
98 item->oic_last_attempt = 0;
100 CDEBUG(D_HA, "imp %p@%s: found existing conn %s%s\n",
101 imp, imp->imp_obd->obd_name, uuid->uuid,
102 (priority ? ", moved to head" : ""));
103 spin_unlock(&imp->imp_lock);
104 GOTO(out_free, rc = 0);
107 /* No existing import connection found for \a uuid. */
109 imp_conn->oic_conn = ptlrpc_conn;
110 imp_conn->oic_uuid = *uuid;
111 imp_conn->oic_last_attempt = 0;
113 list_add(&imp_conn->oic_item, &imp->imp_conn_list);
115 list_add_tail(&imp_conn->oic_item,
116 &imp->imp_conn_list);
117 CDEBUG(D_HA, "imp %p@%s: add connection %s at %s\n",
118 imp, imp->imp_obd->obd_name, uuid->uuid,
119 (priority ? "head" : "tail"));
121 spin_unlock(&imp->imp_lock);
122 GOTO(out_free, rc = -ENOENT);
125 spin_unlock(&imp->imp_lock);
129 OBD_FREE(imp_conn, sizeof(*imp_conn));
131 ptlrpc_connection_put(ptlrpc_conn);
135 int import_set_conn_priority(struct obd_import *imp, struct obd_uuid *uuid)
137 return import_set_conn(imp, uuid, 1, 0);
140 int client_import_add_conn(struct obd_import *imp, struct obd_uuid *uuid,
143 return import_set_conn(imp, uuid, priority, 1);
145 EXPORT_SYMBOL(client_import_add_conn);
147 int client_import_dyn_add_conn(struct obd_import *imp, struct obd_uuid *uuid,
148 lnet_nid_t prim_nid4, int priority)
150 struct ptlrpc_connection *ptlrpc_conn;
151 struct lnet_nid prim_nid;
154 lnet_nid4_to_nid(prim_nid4, &prim_nid);
155 ptlrpc_conn = ptlrpc_uuid_to_connection(uuid, LNET_NID_NET(&prim_nid));
157 const char *str_uuid = obd_uuid2str(uuid);
159 rc = class_add_uuid(str_uuid, &prim_nid);
161 CERROR("%s: failed to add UUID '%s': rc = %d\n",
162 imp->imp_obd->obd_name, str_uuid, rc);
166 return import_set_conn(imp, uuid, priority, 1);
168 EXPORT_SYMBOL(client_import_dyn_add_conn);
170 int client_import_add_nids_to_conn(struct obd_import *imp, lnet_nid_t *nids,
171 int nid_count, struct obd_uuid *uuid)
173 struct obd_import_conn *conn;
177 if (nid_count <= 0 || !nids)
180 spin_lock(&imp->imp_lock);
181 list_for_each_entry(conn, &imp->imp_conn_list, oic_item) {
184 lnet_nid4_to_nid(nids[0], &nid);
185 if (class_check_uuid(&conn->oic_uuid, &nid)) {
186 *uuid = conn->oic_uuid;
187 spin_unlock(&imp->imp_lock);
188 rc = class_add_nids_to_uuid(&conn->oic_uuid, nids,
193 spin_unlock(&imp->imp_lock);
196 EXPORT_SYMBOL(client_import_add_nids_to_conn);
198 int client_import_del_conn(struct obd_import *imp, struct obd_uuid *uuid)
200 struct obd_import_conn *imp_conn;
201 struct obd_export *dlmexp;
206 spin_lock(&imp->imp_lock);
207 if (list_empty(&imp->imp_conn_list)) {
208 LASSERT(!imp->imp_connection);
212 list_for_each_entry(imp_conn, &imp->imp_conn_list, oic_item) {
213 if (!obd_uuid_equals(uuid, &imp_conn->oic_uuid))
215 LASSERT(imp_conn->oic_conn);
217 if (imp_conn == imp->imp_conn_current) {
218 LASSERT(imp_conn->oic_conn == imp->imp_connection);
220 if (imp->imp_state != LUSTRE_IMP_CLOSED &&
221 imp->imp_state != LUSTRE_IMP_DISCON) {
222 CERROR("can't remove current connection\n");
223 GOTO(out, rc = -EBUSY);
226 ptlrpc_connection_put(imp->imp_connection);
227 imp->imp_connection = NULL;
229 dlmexp = class_conn2export(&imp->imp_dlm_handle);
230 if (dlmexp && dlmexp->exp_connection) {
231 LASSERT(dlmexp->exp_connection ==
233 ptlrpc_connection_put(dlmexp->exp_connection);
234 dlmexp->exp_connection = NULL;
238 class_export_put(dlmexp);
241 list_del(&imp_conn->oic_item);
242 ptlrpc_connection_put(imp_conn->oic_conn);
243 OBD_FREE(imp_conn, sizeof(*imp_conn));
244 CDEBUG(D_HA, "imp %p@%s: remove connection %s\n",
245 imp, imp->imp_obd->obd_name, uuid->uuid);
250 spin_unlock(&imp->imp_lock);
252 CERROR("connection %s not found\n", uuid->uuid);
255 EXPORT_SYMBOL(client_import_del_conn);
257 void client_destroy_import(struct obd_import *imp)
260 * Drop security policy instance after all RPCs have finished/aborted
261 * to let all busy contexts be released.
263 class_import_get(imp);
264 class_destroy_import(imp);
265 sptlrpc_import_sec_put(imp);
266 class_import_put(imp);
268 EXPORT_SYMBOL(client_destroy_import);
271 * Check whether or not the OSC is on MDT.
274 * setup 0:{fsname}-OSTxxxx-osc[-MDTxxxx] 1:lustre-OST0000_UUID 2:NID
276 * setup 0:{fsname}-OSTxxxx-osc 1:lustre-OST0000_UUID 2:NID
279 static int osc_on_mdt(char *obdname)
283 ptr = strrchr(obdname, '-');
287 if (strncmp(ptr + 1, "MDT", 3) == 0)
294 * Configure an RPC client OBD device.
299 * 3 - inactive-on-startup
300 * 4 - restrictive net
302 int client_obd_setup(struct obd_device *obd, struct lustre_cfg *lcfg)
304 struct client_obd *cli = &obd->u.cli;
305 struct obd_import *imp;
306 struct obd_uuid server_uuid;
307 int rq_portal, rp_portal, connect_op;
308 const char *name = obd->obd_type->typ_name;
309 enum ldlm_ns_type ns_type = LDLM_NS_TYPE_UNKNOWN;
310 char *cli_name = lustre_cfg_buf(lcfg, 0);
316 * In a more perfect world, we would hang a ptlrpc_client off of
317 * obd_type and just use the values from there.
319 if (!strcmp(name, LUSTRE_OSC_NAME)) {
320 rq_portal = OST_REQUEST_PORTAL;
321 rp_portal = OSC_REPLY_PORTAL;
322 connect_op = OST_CONNECT;
323 cli->cl_sp_me = LUSTRE_SP_CLI;
324 cli->cl_sp_to = LUSTRE_SP_OST;
325 ns_type = LDLM_NS_TYPE_OSC;
326 } else if (!strcmp(name, LUSTRE_MDC_NAME) ||
327 !strcmp(name, LUSTRE_LWP_NAME)) {
328 rq_portal = MDS_REQUEST_PORTAL;
329 rp_portal = MDC_REPLY_PORTAL;
330 connect_op = MDS_CONNECT;
331 if (is_lwp_on_ost(cli_name))
332 cli->cl_sp_me = LUSTRE_SP_OST;
333 else if (is_lwp_on_mdt(cli_name))
334 cli->cl_sp_me = LUSTRE_SP_MDT;
336 cli->cl_sp_me = LUSTRE_SP_CLI;
337 cli->cl_sp_to = LUSTRE_SP_MDT;
338 ns_type = LDLM_NS_TYPE_MDC;
339 } else if (!strcmp(name, LUSTRE_OSP_NAME)) {
340 if (strstr(lustre_cfg_buf(lcfg, 1), "OST") == NULL) {
341 /* OSP_on_MDT for other MDTs */
342 connect_op = MDS_CONNECT;
343 cli->cl_sp_to = LUSTRE_SP_MDT;
344 ns_type = LDLM_NS_TYPE_MDC;
345 rq_portal = OUT_PORTAL;
347 /* OSP on MDT for OST */
348 connect_op = OST_CONNECT;
349 cli->cl_sp_to = LUSTRE_SP_OST;
350 ns_type = LDLM_NS_TYPE_OSC;
351 rq_portal = OST_REQUEST_PORTAL;
353 rp_portal = OSC_REPLY_PORTAL;
354 cli->cl_sp_me = LUSTRE_SP_MDT;
355 } else if (!strcmp(name, LUSTRE_MGC_NAME)) {
356 rq_portal = MGS_REQUEST_PORTAL;
357 rp_portal = MGC_REPLY_PORTAL;
358 connect_op = MGS_CONNECT;
359 cli->cl_sp_me = LUSTRE_SP_MGC;
360 cli->cl_sp_to = LUSTRE_SP_MGS;
361 cli->cl_flvr_mgc.sf_rpc = SPTLRPC_FLVR_INVALID;
362 ns_type = LDLM_NS_TYPE_MGC;
364 CERROR("unknown client OBD type \"%s\", can't setup\n",
369 if (LUSTRE_CFG_BUFLEN(lcfg, 1) < 1) {
370 CERROR("requires a TARGET UUID\n");
374 if (LUSTRE_CFG_BUFLEN(lcfg, 1) > 37) {
375 CERROR("client UUID must be less than 38 characters\n");
379 if (LUSTRE_CFG_BUFLEN(lcfg, 2) < 1) {
380 CERROR("setup requires a SERVER UUID\n");
384 if (LUSTRE_CFG_BUFLEN(lcfg, 2) > 37) {
385 CERROR("target UUID must be less than 38 characters\n");
389 init_rwsem(&cli->cl_sem);
390 mutex_init(&cli->cl_mgc_mutex);
392 init_rwsem(&cli->cl_seq_rwsem);
393 cli->cl_conn_count = 0;
394 memcpy(server_uuid.uuid, lustre_cfg_buf(lcfg, 2),
395 min_t(unsigned int, LUSTRE_CFG_BUFLEN(lcfg, 2),
396 sizeof(server_uuid)));
398 cli->cl_dirty_pages = 0;
399 cli->cl_dirty_max_pages = 0;
400 cli->cl_avail_grant = 0;
401 /* FIXME: Should limit this for the sum of all cl_dirty_max_pages. */
403 * cl_dirty_max_pages may be changed at connect time in
404 * ptlrpc_connect_interpret().
406 client_adjust_max_dirty(cli);
407 init_waitqueue_head(&cli->cl_cache_waiters);
408 INIT_LIST_HEAD(&cli->cl_loi_ready_list);
409 INIT_LIST_HEAD(&cli->cl_loi_hp_ready_list);
410 INIT_LIST_HEAD(&cli->cl_loi_write_list);
411 INIT_LIST_HEAD(&cli->cl_loi_read_list);
412 spin_lock_init(&cli->cl_loi_list_lock);
413 atomic_set(&cli->cl_pending_w_pages, 0);
414 atomic_set(&cli->cl_pending_r_pages, 0);
415 cli->cl_r_in_flight = 0;
416 cli->cl_w_in_flight = 0;
418 cli->cl_stats_init = ktime_get_real();
419 spin_lock_init(&cli->cl_read_rpc_hist.oh_lock);
420 spin_lock_init(&cli->cl_write_rpc_hist.oh_lock);
421 spin_lock_init(&cli->cl_read_page_hist.oh_lock);
422 spin_lock_init(&cli->cl_write_page_hist.oh_lock);
423 spin_lock_init(&cli->cl_read_offset_hist.oh_lock);
424 spin_lock_init(&cli->cl_write_offset_hist.oh_lock);
425 spin_lock_init(&cli->cl_batch_rpc_hist.oh_lock);
428 INIT_LIST_HEAD(&cli->cl_lru_osc);
429 atomic_set(&cli->cl_lru_shrinkers, 0);
430 atomic_long_set(&cli->cl_lru_busy, 0);
431 atomic_long_set(&cli->cl_lru_in_list, 0);
432 INIT_LIST_HEAD(&cli->cl_lru_list);
433 spin_lock_init(&cli->cl_lru_list_lock);
434 atomic_long_set(&cli->cl_unstable_count, 0);
435 INIT_LIST_HEAD(&cli->cl_shrink_list);
436 INIT_LIST_HEAD(&cli->cl_grant_chain);
438 INIT_LIST_HEAD(&cli->cl_flight_waiters);
439 cli->cl_rpcs_in_flight = 0;
441 init_waitqueue_head(&cli->cl_destroy_waitq);
442 atomic_set(&cli->cl_destroy_in_flight, 0);
445 cli->cl_supp_cksum_types = OBD_CKSUM_CRC32;
446 cli->cl_preferred_cksum_type = 0;
447 #ifdef CONFIG_ENABLE_CHECKSUM
448 /* Turn on checksumming by default. */
449 cli->cl_checksum = 1;
451 * The supported checksum types will be worked out at connect time
452 * Set cl_chksum* to CRC32 for now to avoid returning screwed info
455 cli->cl_cksum_type = cli->cl_supp_cksum_types;
457 atomic_set(&cli->cl_resends, OSC_DEFAULT_RESENDS);
460 * Set it to possible maximum size. It may be reduced by ocd_brw_size
461 * from OFD after connecting.
463 cli->cl_max_pages_per_rpc = PTLRPC_MAX_BRW_PAGES;
465 cli->cl_max_short_io_bytes = OBD_DEF_SHORT_IO_BYTES;
468 * set cl_chunkbits default value to PAGE_SHIFT,
469 * it will be updated at OSC connection time.
471 cli->cl_chunkbits = PAGE_SHIFT;
473 if (!strcmp(name, LUSTRE_MDC_NAME)) {
474 cli->cl_max_rpcs_in_flight = OBD_MAX_RIF_DEFAULT;
475 } else if (cfs_totalram_pages() >> (20 - PAGE_SHIFT) <= 128 /* MB */) {
476 cli->cl_max_rpcs_in_flight = 2;
477 } else if (cfs_totalram_pages() >> (20 - PAGE_SHIFT) <= 256 /* MB */) {
478 cli->cl_max_rpcs_in_flight = 3;
479 } else if (cfs_totalram_pages() >> (20 - PAGE_SHIFT) <= 512 /* MB */) {
480 cli->cl_max_rpcs_in_flight = 4;
482 if (osc_on_mdt(obd->obd_name))
483 cli->cl_max_rpcs_in_flight = OBD_MAX_RIF_MAX;
485 cli->cl_max_rpcs_in_flight = OBD_MAX_RIF_DEFAULT;
488 spin_lock_init(&cli->cl_mod_rpcs_hist.oh_lock);
489 cli->cl_max_mod_rpcs_in_flight = 0;
490 cli->cl_mod_rpcs_in_flight = 0;
491 cli->cl_close_rpcs_in_flight = 0;
492 init_waitqueue_head(&cli->cl_mod_rpcs_waitq);
493 cli->cl_mod_rpcs_init = ktime_get_real();
494 cli->cl_mod_tag_bitmap = NULL;
496 INIT_LIST_HEAD(&cli->cl_chg_dev_linkage);
498 if (connect_op == MDS_CONNECT) {
499 cli->cl_max_mod_rpcs_in_flight = cli->cl_max_rpcs_in_flight - 1;
500 OBD_ALLOC(cli->cl_mod_tag_bitmap,
501 BITS_TO_LONGS(OBD_MAX_RIF_MAX) * sizeof(long));
502 if (cli->cl_mod_tag_bitmap == NULL)
503 GOTO(err, rc = -ENOMEM);
508 CERROR("ldlm_get_ref failed: %d\n", rc);
512 ptlrpc_init_client(rq_portal, rp_portal, name,
513 &obd->obd_ldlm_client);
515 imp = class_new_import(obd);
517 GOTO(err_ldlm, rc = -ENOENT);
518 imp->imp_client = &obd->obd_ldlm_client;
519 imp->imp_connect_op = connect_op;
520 memcpy(cli->cl_target_uuid.uuid, lustre_cfg_buf(lcfg, 1),
521 LUSTRE_CFG_BUFLEN(lcfg, 1));
522 class_import_put(imp);
524 if (lustre_cfg_buf(lcfg, 4)) {
525 __u32 refnet = libcfs_str2net(lustre_cfg_string(lcfg, 4));
527 if (refnet == LNET_NET_ANY) {
529 CERROR("%s: bad mount option 'network=%s': rc = %d\n",
530 obd->obd_name, lustre_cfg_string(lcfg, 4),
532 GOTO(err_import, rc);
534 imp->imp_conn_restricted_net = refnet;
536 imp->imp_conn_restricted_net = LNET_NET_ANY;
539 rc = client_import_add_conn(imp, &server_uuid, 1);
541 CERROR("can't add initial connection\n");
542 GOTO(err_import, rc);
544 imp->imp_connection = NULL;
546 cli->cl_import = imp;
547 /* cli->cl_max_mds_easize updated by mdc_init_ea_size() */
548 cli->cl_max_mds_easize = sizeof(struct lov_mds_md_v3);
550 if (LUSTRE_CFG_BUFLEN(lcfg, 3) > 0) {
551 if (!strcmp(lustre_cfg_string(lcfg, 3), "inactive")) {
552 CDEBUG(D_HA, "marking %s %s->%s as inactive\n",
554 cli->cl_target_uuid.uuid);
555 spin_lock(&imp->imp_lock);
556 imp->imp_deactive = 1;
557 spin_unlock(&imp->imp_lock);
561 obd->obd_namespace = ldlm_namespace_new(obd, obd->obd_name,
562 LDLM_NAMESPACE_CLIENT,
563 LDLM_NAMESPACE_GREEDY,
565 if (IS_ERR(obd->obd_namespace)) {
566 rc = PTR_ERR(obd->obd_namespace);
567 CERROR("%s: unable to create client namespace: rc = %d\n",
569 obd->obd_namespace = NULL;
570 GOTO(err_import, rc);
576 class_destroy_import(imp);
580 if (cli->cl_mod_tag_bitmap != NULL)
581 OBD_FREE(cli->cl_mod_tag_bitmap,
582 BITS_TO_LONGS(OBD_MAX_RIF_MAX) * sizeof(long));
583 cli->cl_mod_tag_bitmap = NULL;
587 EXPORT_SYMBOL(client_obd_setup);
589 int client_obd_cleanup(struct obd_device *obd)
591 struct client_obd *cli = &obd->u.cli;
595 ldlm_namespace_free_post(obd->obd_namespace);
596 obd->obd_namespace = NULL;
598 obd_cleanup_client_import(obd);
599 LASSERT(obd->u.cli.cl_import == NULL);
603 if (cli->cl_mod_tag_bitmap != NULL)
604 OBD_FREE(cli->cl_mod_tag_bitmap,
605 BITS_TO_LONGS(OBD_MAX_RIF_MAX) * sizeof(long));
606 cli->cl_mod_tag_bitmap = NULL;
610 EXPORT_SYMBOL(client_obd_cleanup);
612 /* ->o_connect() method for client side (OSC and MDC and MGC) */
613 int client_connect_import(const struct lu_env *env,
614 struct obd_export **exp,
615 struct obd_device *obd, struct obd_uuid *cluuid,
616 struct obd_connect_data *data, void *localdata)
618 struct client_obd *cli = &obd->u.cli;
619 struct obd_import *imp = cli->cl_import;
620 struct obd_connect_data *ocd;
621 struct lustre_handle conn = { 0 };
627 down_write(&cli->cl_sem);
628 if (cli->cl_conn_count > 0)
629 GOTO(out_sem, rc = -EALREADY);
631 rc = class_connect(&conn, obd, cluuid);
635 cli->cl_conn_count++;
636 *exp = class_conn2export(&conn);
638 LASSERT(obd->obd_namespace);
640 spin_lock(&imp->imp_lock);
641 if (imp->imp_state == LUSTRE_IMP_CLOSED && imp->imp_deactive) {
642 /* need to reactivate import if trying to connect
643 * to a previously disconnected
645 imp->imp_deactive = 0;
646 imp->imp_invalid = 0;
648 spin_unlock(&imp->imp_lock);
650 imp->imp_dlm_handle = conn;
651 rc = ptlrpc_init_import(imp);
655 ocd = &imp->imp_connect_data;
658 imp->imp_connect_flags_orig = data->ocd_connect_flags;
659 imp->imp_connect_flags2_orig = data->ocd_connect_flags2;
662 rc = ptlrpc_connect_import(imp);
664 LASSERT(imp->imp_state == LUSTRE_IMP_DISCON);
667 LASSERT(*exp != NULL && (*exp)->exp_connection);
670 LASSERTF((ocd->ocd_connect_flags & data->ocd_connect_flags) ==
671 ocd->ocd_connect_flags, "old %#llx, new %#llx\n",
672 data->ocd_connect_flags, ocd->ocd_connect_flags);
673 data->ocd_connect_flags = ocd->ocd_connect_flags;
674 data->ocd_connect_flags2 = ocd->ocd_connect_flags2;
677 ptlrpc_pinger_add_import(imp);
683 cli->cl_conn_count--;
684 class_disconnect(*exp);
688 up_write(&cli->cl_sem);
690 if (!rc && localdata && !cli->cl_cache) {
691 cli->cl_cache = (struct cl_client_cache *)localdata;
692 cl_cache_incref(cli->cl_cache);
693 cli->cl_lru_left = &cli->cl_cache->ccc_lru_left;
695 /* add this osc into entity list */
696 LASSERT(list_empty(&cli->cl_lru_osc));
697 spin_lock(&cli->cl_cache->ccc_lru_lock);
698 list_add(&cli->cl_lru_osc, &cli->cl_cache->ccc_lru);
699 spin_unlock(&cli->cl_cache->ccc_lru_lock);
704 EXPORT_SYMBOL(client_connect_import);
706 int client_disconnect_export(struct obd_export *exp)
708 struct obd_device *obd = class_exp2obd(exp);
709 struct client_obd *cli;
710 struct obd_import *imp;
716 CERROR("invalid export for disconnect: exp %p cookie %#llx\n",
717 exp, exp ? exp->exp_handle.h_cookie : -1);
722 imp = cli->cl_import;
724 down_write(&cli->cl_sem);
725 CDEBUG(D_INFO, "disconnect %s - %zu\n", obd->obd_name,
728 if (cli->cl_conn_count == 0) {
729 CERROR("disconnecting disconnected device (%s)\n",
731 GOTO(out_disconnect, rc = -EINVAL);
734 cli->cl_conn_count--;
735 if (cli->cl_conn_count != 0)
736 GOTO(out_disconnect, rc = 0);
739 * Mark import deactivated now, so we don't try to reconnect if any
740 * of the cleanup RPCs fails (e.g. LDLM cancel, etc). We don't
741 * fully deactivate the import, or that would drop all requests.
743 spin_lock(&imp->imp_lock);
744 imp->imp_deactive = 1;
745 spin_unlock(&imp->imp_lock);
748 * Some non-replayable imports (MDS's OSCs) are pinged, so just
749 * delete it regardless. (It's safe to delete an import that was
752 (void)ptlrpc_pinger_del_import(imp);
754 if (obd->obd_namespace != NULL) {
755 /* obd_force == local only */
756 ldlm_cli_cancel_unused(obd->obd_namespace, NULL,
757 obd->obd_force ? LCF_LOCAL : 0, NULL);
758 ldlm_namespace_free_prior(obd->obd_namespace, imp,
763 * There's no need to hold sem while disconnecting an import,
764 * and it may actually cause deadlock in GSS.
766 up_write(&cli->cl_sem);
767 rc = ptlrpc_disconnect_import(imp, 0);
768 down_write(&cli->cl_sem);
770 ptlrpc_invalidate_import(imp);
776 * Use server style - class_disconnect should be always called for
779 err = class_disconnect(exp);
783 up_write(&cli->cl_sem);
787 EXPORT_SYMBOL(client_disconnect_export);
789 #ifdef HAVE_SERVER_SUPPORT
790 int server_disconnect_export(struct obd_export *exp)
796 /* Disconnect early so that clients can't keep using export. */
797 rc = class_disconnect(exp);
798 /* Close import to avoid sending any requests. */
799 if (exp->exp_imp_reverse)
800 ptlrpc_cleanup_imp(exp->exp_imp_reverse);
802 ldlm_bl_thread_wakeup();
804 /* complete all outstanding replies */
805 spin_lock(&exp->exp_lock);
806 while (!list_empty(&exp->exp_outstanding_replies)) {
807 struct ptlrpc_reply_state *rs =
808 list_first_entry(&exp->exp_outstanding_replies,
809 struct ptlrpc_reply_state,
811 struct ptlrpc_service_part *svcpt = rs->rs_svcpt;
813 spin_lock(&svcpt->scp_rep_lock);
815 list_del_init(&rs->rs_exp_list);
817 spin_lock(&rs->rs_lock);
818 /* clear rs_convert_lock to make sure rs is handled and put */
819 rs->rs_convert_lock = 0;
820 ptlrpc_schedule_difficult_reply(rs);
821 spin_unlock(&rs->rs_lock);
823 spin_unlock(&svcpt->scp_rep_lock);
825 spin_unlock(&exp->exp_lock);
829 EXPORT_SYMBOL(server_disconnect_export);
831 static inline int target_check_recovery_timer(struct obd_device *target)
836 if (!target->obd_recovering || target->obd_recovery_start == 0)
839 remaining = hrtimer_get_remaining(&target->obd_recovery_timer);
840 timeout = ktime_divns(remaining, NSEC_PER_SEC);
844 /* the recovery timer should expire, but it isn't triggered,
845 * it's better to abort the recovery of this target to speed up
846 * the recovery of the whole cluster. */
847 spin_lock(&target->obd_dev_lock);
848 if (target->obd_recovering) {
849 CERROR("%s: Aborting recovery\n", target->obd_name);
850 target->obd_abort_recovery = 1;
851 wake_up(&target->obd_next_transno_waitq);
853 spin_unlock(&target->obd_dev_lock);
858 * --------------------------------------------------------------------------
859 * from old lib/target.c
860 * --------------------------------------------------------------------------
862 static int target_handle_reconnect(struct lustre_handle *conn,
863 struct obd_export *exp,
864 struct obd_uuid *cluuid)
866 struct obd_device *target;
867 struct lustre_handle *hdl;
873 hdl = &exp->exp_imp_reverse->imp_remote_handle;
874 if (!exp->exp_connection || !lustre_handle_is_used(hdl)) {
875 conn->cookie = exp->exp_handle.h_cookie;
877 "connect export for UUID '%s' at %p, cookie %#llx\n",
878 cluuid->uuid, exp, conn->cookie);
882 target = exp->exp_obd;
884 /* Might be a re-connect after a partition. */
885 if (memcmp(&conn->cookie, &hdl->cookie, sizeof(conn->cookie))) {
886 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",
888 obd_uuid2str(&exp->exp_client_uuid),
889 obd_export_nid2str(exp),
890 hdl->cookie, conn->cookie);
891 memset(conn, 0, sizeof(*conn));
893 * target_handle_connect() treats EALREADY and
894 * -EALREADY differently. -EALREADY is an error
895 * (same UUID, different handle).
900 if (!target->obd_recovering) {
901 LCONSOLE_WARN("%s: Client %s (at %s) reconnecting\n",
902 target->obd_name, obd_uuid2str(&exp->exp_client_uuid),
903 obd_export_nid2str(exp));
904 GOTO(out_already, rc);
907 remaining = hrtimer_get_remaining(&target->obd_recovery_timer);
908 timeout = ktime_divns(remaining, NSEC_PER_SEC);
910 LCONSOLE_WARN("%s: Client %s (at %s) reconnected, waiting for %d clients in recovery for %lld:%.02lld\n",
912 obd_uuid2str(&exp->exp_client_uuid),
913 obd_export_nid2str(exp),
914 atomic_read(&target->obd_max_recoverable_clients),
915 timeout / 60, timeout % 60);
917 struct target_distribute_txn_data *tdtd;
922 target_check_recovery_timer(target);
924 tdtd = class_exp2tgt(exp)->lut_tdtd;
925 if (tdtd && tdtd->tdtd_show_update_logs_retrievers)
926 buf = tdtd->tdtd_show_update_logs_retrievers(
927 tdtd->tdtd_show_retrievers_cbdata,
931 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",
933 obd_uuid2str(&exp->exp_client_uuid),
934 obd_export_nid2str(exp), count,
935 buf ? buf : "unknown (not enough RAM)",
936 (abs(timeout) + target->obd_recovery_timeout) / 60,
937 (abs(timeout) + target->obd_recovery_timeout) % 60,
940 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",
941 target->obd_name, abs(timeout) / 60,
942 abs(timeout) % 60, target->obd_name);
949 conn->cookie = exp->exp_handle.h_cookie;
951 * target_handle_connect() treats EALREADY and
952 * -EALREADY differently. EALREADY means we are
953 * doing a valid reconnect from the same client.
959 check_and_start_recovery_timer(struct obd_device *obd,
960 struct ptlrpc_request *req, int new_client);
963 * update flags for import during reconnect process
965 static int rev_import_flags_update(struct obd_import *revimp,
966 struct ptlrpc_request *req)
969 struct obd_connect_data *data;
971 data = req_capsule_client_get(&req->rq_pill, &RMF_CONNECT_DATA);
973 if (data->ocd_connect_flags & OBD_CONNECT_AT)
974 revimp->imp_msghdr_flags |= MSGHDR_AT_SUPPORT;
976 revimp->imp_msghdr_flags &= ~MSGHDR_AT_SUPPORT;
978 revimp->imp_msghdr_flags |= MSGHDR_CKSUM_INCOMPAT18;
980 revimp->imp_connect_data = *data;
981 rc = sptlrpc_import_sec_adapt(revimp, req->rq_svc_ctx, &req->rq_flvr);
983 CERROR("%s: cannot get reverse import %s security: rc = %d\n",
984 revimp->imp_client->cli_name,
985 libcfs_idstr(&req->rq_peer), rc);
993 * Allocate a new reverse import for an export.
995 * \retval -errno in case error hit
996 * \retval 0 if reverse import correctly init
998 int rev_import_init(struct obd_export *export)
1000 struct obd_device *obd = export->exp_obd;
1001 struct obd_import *revimp;
1003 LASSERT(export->exp_imp_reverse == NULL);
1005 revimp = class_new_import(obd);
1009 revimp->imp_remote_handle.cookie = 0ULL;
1010 revimp->imp_client = &obd->obd_ldlm_client;
1011 revimp->imp_dlm_fake = 1;
1013 /* it is safe to connect import in new state as no sends possible */
1014 spin_lock(&export->exp_lock);
1015 export->exp_imp_reverse = revimp;
1016 spin_unlock(&export->exp_lock);
1017 class_import_put(revimp);
1021 EXPORT_SYMBOL(rev_import_init);
1024 * Handle reconnect for an export.
1026 * \param exp export to handle reconnect process
1027 * \param req client reconnect request
1029 * \retval -rc in case securitfy flavor can't be changed
1030 * \retval 0 in case none problems
1032 static int rev_import_reconnect(struct obd_export *exp,
1033 struct ptlrpc_request *req)
1035 struct obd_import *revimp = exp->exp_imp_reverse;
1036 struct lustre_handle *lh;
1039 /* avoid sending a request until import flags are changed */
1040 ptlrpc_import_enter_resend(revimp);
1042 ptlrpc_connection_put(revimp->imp_connection);
1045 * client from recovery don't have a handle so we need to take from
1046 * request. it may produce situation when wrong client connected
1047 * to recovery as we trust a client uuid
1049 lh = req_capsule_client_get(&req->rq_pill, &RMF_CONN);
1050 revimp->imp_remote_handle = *lh;
1053 * unknown versions will be caught in
1054 * ptlrpc_handle_server_req_in->lustre_unpack_msg()
1056 revimp->imp_msg_magic = req->rq_reqmsg->lm_magic;
1058 revimp->imp_connection = ptlrpc_connection_addref(exp->exp_connection);
1060 rc = rev_import_flags_update(revimp, req);
1063 * it is safe to still be in RECOVERY phase as we are not able
1064 * to setup correct security flavor so requests are not able to
1065 * be delivered correctly
1070 /* resend all rpc's via new connection */
1071 return ptlrpc_import_recovery_state_machine(revimp);
1074 int target_handle_connect(struct ptlrpc_request *req)
1076 struct obd_device *target = NULL;
1077 struct obd_export *export = NULL;
1079 * connect handle - filled from target_handle_reconnect in
1082 struct lustre_handle conn;
1083 struct lustre_handle *tmp;
1084 struct obd_uuid cluuid;
1089 bool mds_conn = false, lw_client = false, initial_conn = false;
1090 bool mds_mds_conn = false;
1091 bool new_mds_mds_conn = false;
1092 struct obd_connect_data *data, *tmpdata;
1094 #if LUSTRE_VERSION_CODE < OBD_OCD_VERSION(3, 0, 53, 0)
1095 int tmp_exp_old_falloc;
1097 struct ptlrpc_connection *pcon = NULL;
1098 bool reconnected = false;
1102 CFS_RACE(OBD_FAIL_TGT_CONN_RACE);
1104 str = req_capsule_client_get(&req->rq_pill, &RMF_TGTUUID);
1106 DEBUG_REQ(D_ERROR, req, "bad target UUID for connect");
1107 GOTO(out, rc = -EINVAL);
1110 target = class_dev_by_str(str);
1112 deuuidify(str, NULL, &target_start, &target_len);
1113 LCONSOLE_ERROR_MSG(0x137,
1114 "%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",
1115 str, libcfs_nidstr(&req->rq_peer.nid));
1116 GOTO(out, rc = -ENODEV);
1119 spin_lock(&target->obd_dev_lock);
1121 target->obd_conn_inprogress++;
1123 if (target->obd_stopping || !target->obd_set_up) {
1124 spin_unlock(&target->obd_dev_lock);
1126 deuuidify(str, NULL, &target_start, &target_len);
1127 LCONSOLE_INFO("%.*s: Not available for connect from %s (%s)\n",
1128 target_len, target_start,
1129 libcfs_nidstr(&req->rq_peer.nid),
1130 (target->obd_stopping ?
1131 "stopping" : "not set up"));
1132 GOTO(out, rc = -ENODEV);
1135 if (target->obd_no_conn) {
1136 spin_unlock(&target->obd_dev_lock);
1139 "%s: Temporarily refusing client connection from %s\n",
1140 target->obd_name, libcfs_nidstr(&req->rq_peer.nid));
1141 GOTO(out, rc = -EAGAIN);
1144 spin_unlock(&target->obd_dev_lock);
1146 str = req_capsule_client_get(&req->rq_pill, &RMF_CLUUID);
1148 DEBUG_REQ(D_ERROR, req, "bad client UUID for connect");
1149 GOTO(out, rc = -EINVAL);
1152 obd_str2uuid(&cluuid, str);
1154 tmp = req_capsule_client_get(&req->rq_pill, &RMF_CONN);
1156 GOTO(out, rc = -EPROTO);
1160 size = req_capsule_get_size(&req->rq_pill, &RMF_CONNECT_DATA,
1162 if (size < 0 || size > 8 * sizeof(struct obd_connect_data))
1163 GOTO(out, rc = -EPROTO);
1164 data = req_capsule_client_get(&req->rq_pill, &RMF_CONNECT_DATA);
1166 GOTO(out, rc = -EPROTO);
1168 rc = req_capsule_server_pack(&req->rq_pill);
1172 #if LUSTRE_VERSION_CODE < OBD_OCD_VERSION(3, 0, 53, 0)
1174 * Don't allow clients to connect that are using old 1.8 format
1175 * protocol conventions (LUSTRE_MSG_MAGIC_v1, !MSGHDR_CKSUM_INCOMPAT18,
1176 * ldlm_flock_policy_wire format, MDT_ATTR_xTIME_SET, etc). The
1177 * FULL20 flag should be set on all connections since 2.0, but no
1178 * longer affects behaviour.
1180 * Later this check will be disabled and the flag can be retired
1181 * completely once interop with 3.0 is no longer needed.
1183 if (!(data->ocd_connect_flags & OBD_CONNECT_FULL20))
1184 GOTO(out, rc = -EPROTO);
1186 /* Old clients will have 'tmp_exp_old_falloc' as 1.
1187 * Newer clients (2.15) and beyond will have it set as 0
1189 tmp_exp_old_falloc =
1190 !!(data->ocd_connect_flags & OBD_CONNECT_OLD_FALLOC);
1192 CDEBUG(D_INFO, "%s: ocd_connect_flags: %#llx tmp_exp_old_falloc: %d\n",
1193 target->obd_name, data->ocd_connect_flags, tmp_exp_old_falloc);
1196 * Don't allow liblustre clients to connect.
1197 * - testing was disabled in v2_2_50_0-61-g6a75d65
1198 * - building was disabled in v2_5_58_0-28-g7277179
1199 * - client code was deleted in v2_6_50_0-101-gcdfbc72,
1200 * - clients were refused connect for version difference > 0.0.1.32
1202 if (lustre_msg_get_op_flags(req->rq_reqmsg) & MSG_CONNECT_LIBCLIENT) {
1203 DEBUG_REQ(D_WARNING, req, "Refusing libclient connection");
1204 GOTO(out, rc = -EPROTO);
1209 * Note: lw_client is needed in MDS-MDS failover during update log
1210 * processing, so we needs to allow lw_client to be connected at
1211 * anytime, instead of only the initial connection
1213 lw_client = OCD_HAS_FLAG(data, LIGHTWEIGHT);
1215 if (lustre_msg_get_op_flags(req->rq_reqmsg) & MSG_CONNECT_INITIAL) {
1216 initial_conn = true;
1217 mds_conn = OCD_HAS_FLAG(data, MDS);
1218 mds_mds_conn = OCD_HAS_FLAG(data, MDS_MDS);
1221 * OBD_CONNECT_MNE_SWAB is removed at 2.14
1222 * Checking OBD_CONNECT_FID can be removed in the future.
1224 * Via check OBD_CONNECT_FID, we can distinguish whether
1225 * the OBD_CONNECT_MDS_MDS/OBD_CONNECT_MNE_SWAB is from
1226 * MGC or MDT, since MGC does not use OBD_CONNECT_FID.
1229 (data->ocd_connect_flags & OBD_CONNECT_MDS_MDS) &&
1230 (data->ocd_connect_flags & OBD_CONNECT_FID) &&
1231 (data->ocd_connect_flags & OBD_CONNECT_VERSION)) {
1232 __u32 major = OBD_OCD_VERSION_MAJOR(data->ocd_version);
1233 __u32 minor = OBD_OCD_VERSION_MINOR(data->ocd_version);
1234 __u32 patch = OBD_OCD_VERSION_PATCH(data->ocd_version);
1237 * We do not support the MDT-MDT interoperations with
1238 * different version MDT because of protocol changes.
1240 if (unlikely(major != LUSTRE_MAJOR ||
1241 minor != LUSTRE_MINOR ||
1242 abs(patch - LUSTRE_PATCH) > 3)) {
1243 LCONSOLE_WARN("%s (%u.%u.%u.%u) refused the connection from different version MDT (%d.%d.%d.%d) %s %s\n",
1244 target->obd_name, LUSTRE_MAJOR,
1245 LUSTRE_MINOR, LUSTRE_PATCH,
1246 LUSTRE_FIX, major, minor, patch,
1247 OBD_OCD_VERSION_FIX(data->ocd_version),
1248 libcfs_nidstr(&req->rq_peer.nid),
1250 GOTO(out, rc = -EPROTO);
1255 /* lctl gets a backstage, all-access pass. */
1256 if (obd_uuid_equals(&cluuid, &target->obd_uuid))
1257 goto dont_check_exports;
1259 export = obd_uuid_lookup(target, &cluuid);
1263 /* We've found an export in the hash. */
1265 spin_lock(&export->exp_lock);
1267 if (export->exp_connecting) { /* b=9635, et. al. */
1268 spin_unlock(&export->exp_lock);
1269 LCONSOLE_WARN("%s: Export %p already connecting from %s\n",
1270 export->exp_obd->obd_name, export,
1271 libcfs_nidstr(&req->rq_peer.nid));
1272 class_export_put(export);
1275 } else if ((mds_conn || (lw_client && initial_conn) ||
1276 OCD_HAS_FLAG(data, MDS_MDS)) && export->exp_connection) {
1277 spin_unlock(&export->exp_lock);
1278 if (!nid_same(&req->rq_peer.nid,
1279 &export->exp_connection->c_peer.nid)) {
1280 /* MDS or LWP reconnected after failover. */
1281 LCONSOLE_WARN("%s: Received %s connection from %s, removing former export from %s\n",
1283 lw_client ? "LWP" : "MDS",
1284 libcfs_nidstr(&req->rq_peer.nid),
1285 libcfs_nidstr(&export->exp_connection->c_peer.nid));
1287 /* New connection from the same NID. */
1288 LCONSOLE_WARN("%s: Received new %s connection from %s, %s former export from same NID\n",
1290 lw_client ? "LWP" : "MDS",
1291 libcfs_nidstr(&req->rq_peer.nid),
1292 OCD_HAS_FLAG(data, MDS_MDS) ?
1296 if (nid_same(&req->rq_peer.nid,
1297 &export->exp_connection->c_peer.nid) &&
1298 OCD_HAS_FLAG(data, MDS_MDS)) {
1300 * Because exports between MDTs will always be
1301 * kept, let's do not fail such export if they
1302 * come from the same NID, otherwise it might
1303 * cause eviction between MDTs, which might
1304 * cause namespace inconsistency
1306 spin_lock(&export->exp_lock);
1307 export->exp_connecting = 1;
1308 export->exp_conn_cnt = 0;
1309 spin_unlock(&export->exp_lock);
1310 conn.cookie = export->exp_handle.h_cookie;
1313 class_fail_export(export);
1314 class_export_put(export);
1318 } else if (export->exp_connection != NULL && initial_conn &&
1319 !nid_same(&req->rq_peer.nid,
1320 &export->exp_connection->c_peer.nid)) {
1321 spin_unlock(&export->exp_lock);
1322 /* In MDS failover we have static UUID but NID can change. */
1323 LCONSOLE_WARN("%s: Client %s seen on new nid %s when existing nid %s is already connected\n",
1324 target->obd_name, cluuid.uuid,
1325 libcfs_nidstr(&req->rq_peer.nid),
1327 &export->exp_connection->c_peer.nid));
1329 class_export_put(export);
1331 } else if (CFS_FAIL_PRECHECK(OBD_FAIL_TGT_RECOVERY_CONNECT) &&
1333 spin_unlock(&export->exp_lock);
1336 export->exp_connecting = 1;
1337 spin_unlock(&export->exp_lock);
1338 LASSERT(export->exp_obd == target);
1340 rc = target_handle_reconnect(&conn, export, &cluuid);
1343 /* If we found an export, we already unlocked. */
1346 CFS_FAIL_TIMEOUT(OBD_FAIL_TGT_DELAY_CONNECT, 2 * obd_timeout);
1347 } else if (req->rq_export == NULL &&
1348 atomic_read(&export->exp_rpc_count) > 0) {
1349 LCONSOLE_WARN("%s: Client %s (at %s) refused connection, still busy with %d references\n",
1350 target->obd_name, cluuid.uuid,
1351 libcfs_nidstr(&req->rq_peer.nid),
1352 refcount_read(&export->exp_handle.h_ref));
1353 GOTO(out, rc = -EBUSY);
1354 } else if (lustre_msg_get_conn_cnt(req->rq_reqmsg) == 1 &&
1356 if (!strstr(cluuid.uuid, "mdt"))
1357 LCONSOLE_WARN("%s: Rejecting reconnect from the known client %s (at %s) because it is indicating it is a new client\n",
1358 target->obd_name, cluuid.uuid,
1359 libcfs_nidstr(&req->rq_peer.nid));
1360 GOTO(out, rc = -EALREADY);
1362 CFS_FAIL_TIMEOUT(OBD_FAIL_TGT_DELAY_RECONNECT, 2 * obd_timeout);
1368 CDEBUG(D_HA, "%s: connection from %s@%s %st%llu exp %p cur %lld last %lld\n",
1369 target->obd_name, cluuid.uuid, libcfs_nidstr(&req->rq_peer.nid),
1370 target->obd_recovering ? "recovering/" : "", data->ocd_transno,
1371 export, ktime_get_seconds(),
1372 export ? export->exp_last_request_time : 0);
1375 * If this is the first time a client connects, reset the recovery
1376 * timer. Discard lightweight connections which might be local.
1378 if (!lw_client && rc == 0 && target->obd_recovering)
1379 check_and_start_recovery_timer(target, req, export == NULL);
1382 * We want to handle EALREADY but *not* -EALREADY from
1383 * target_handle_reconnect(), return reconnection state in a flag.
1385 if (rc == EALREADY) {
1386 lustre_msg_add_op_flags(req->rq_repmsg, MSG_CONNECT_RECONNECT);
1392 /* Tell the client if we support replayable requests. */
1393 if (target->obd_replayable)
1394 lustre_msg_add_op_flags(req->rq_repmsg, MSG_CONNECT_REPLAYABLE);
1396 if (export == NULL) {
1397 /* allow lightweight connections during recovery */
1399 * allow "new" MDT to be connected during recovery, since we
1400 * need retrieve recovery update records from it
1402 if (target->obd_recovering && !lw_client && !mds_mds_conn) {
1403 struct hrtimer *timer = &target->obd_recovery_timer;
1412 connected = atomic_read(&target->obd_connected_clients);
1413 in_progress = atomic_read(&target->obd_lock_replay_clients);
1415 atomic_read(&target->obd_max_recoverable_clients);
1416 stale = target->obd_stale_clients;
1417 remaining = hrtimer_get_remaining(timer);
1418 left = ktime_divns(remaining, NSEC_PER_SEC);
1420 if (ktime_to_ns(remaining) > 0) {
1421 msg = "to recover in";
1424 msg = "already passed deadline";
1427 target_check_recovery_timer(target);
1430 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",
1431 target->obd_name, cluuid.uuid,
1432 libcfs_nidstr(&req->rq_peer.nid), known,
1433 connected - in_progress, in_progress,
1434 stale, msg, timeout / 60, timeout % 60);
1438 rc = obd_connect(req->rq_svc_thread->t_env,
1439 &export, target, &cluuid, data,
1441 if (mds_conn && CFS_FAIL_CHECK(OBD_FAIL_TGT_RCVG_FLAG))
1442 lustre_msg_add_op_flags(req->rq_repmsg,
1443 MSG_CONNECT_RECOVERING);
1445 conn.cookie = export->exp_handle.h_cookie;
1446 rc = rev_import_init(export);
1450 new_mds_mds_conn = true;
1453 if (CFS_FAIL_CHECK(OBD_FAIL_MDS_CONNECT_VS_EVICT)) {
1454 class_export_get(export);
1455 class_fail_export(export);
1456 class_export_put(export);
1458 rc = obd_reconnect(req->rq_svc_thread->t_env,
1459 export, target, &cluuid, data,
1467 data->ocd_instance = obd2obt(target)->obt_instance;
1470 * Return only the parts of obd_connect_data that we understand, so the
1471 * client knows that we don't understand the rest.
1474 tmpsize = req_capsule_get_size(&req->rq_pill, &RMF_CONNECT_DATA,
1476 tmpdata = req_capsule_server_get(&req->rq_pill,
1479 * Don't use struct assignment here, because the client reply
1480 * buffer may be smaller/larger than the local struct
1483 memcpy(tmpdata, data, min(tmpsize, size));
1487 * If the client and the server are the same node, we will already
1488 * have an export that really points to the client's DLM export,
1489 * because we have a shared handles table.
1491 * XXX this will go away when shaver stops sending the "connect" handle
1492 * in the real "remote handle" field of the request --phik 24 Apr 2003
1494 ptlrpc_request_change_export(req, export);
1496 pcon = ptlrpc_connection_get(&req->rq_peer,
1497 &req->rq_self, &cluuid);
1499 GOTO(out, rc = -ENOTCONN);
1501 spin_lock(&export->exp_lock);
1503 if (export->exp_disconnected) {
1504 spin_unlock(&export->exp_lock);
1507 * for each connect called disconnect
1508 * should be called to cleanup stuff
1510 class_export_get(export);
1511 obd_disconnect(export);
1514 GOTO(out, rc = -ENODEV);
1516 if (export->exp_conn_cnt >= lustre_msg_get_conn_cnt(req->rq_reqmsg)) {
1517 spin_unlock(&export->exp_lock);
1519 "%s: %s already connected at greater or equal conn_cnt: %d >= %d\n",
1520 cluuid.uuid, libcfs_nidstr(&req->rq_peer.nid),
1521 export->exp_conn_cnt,
1522 lustre_msg_get_conn_cnt(req->rq_reqmsg));
1524 GOTO(out, rc = -EALREADY);
1526 LASSERT(lustre_msg_get_conn_cnt(req->rq_reqmsg) > 0);
1527 export->exp_conn_cnt = lustre_msg_get_conn_cnt(req->rq_reqmsg);
1529 #if LUSTRE_VERSION_CODE < OBD_OCD_VERSION(3, 0, 53, 0)
1530 /* make 'tmp_exp_old_falloc' persistent by saving it into
1531 * server side export object(obd_export)
1533 export->exp_old_falloc = tmp_exp_old_falloc;
1536 /* Check to see if connection came from another NID. */
1537 if (export->exp_connection != NULL &&
1538 !nid_same(&export->exp_connection->c_peer.nid,
1539 &req->rq_peer.nid)) {
1540 obd_nid_del(export->exp_obd, export);
1541 ptlrpc_connection_put(export->exp_connection);
1542 export->exp_connection = NULL;
1545 if (export->exp_connection == NULL) {
1546 export->exp_connection = pcon;
1549 obd_nid_add(export->exp_obd, export);
1551 spin_unlock(&export->exp_lock);
1553 lustre_msg_set_handle(req->rq_repmsg, &conn);
1555 rc = rev_import_reconnect(export, req);
1559 if (target->obd_recovering && !export->exp_in_recovery && !lw_client) {
1561 __u64 transno = data->ocd_transno;
1563 spin_lock(&export->exp_lock);
1565 * possible race with class_disconnect_stale_exports,
1566 * export may be already in the eviction process
1568 if (export->exp_failed) {
1569 spin_unlock(&export->exp_lock);
1570 GOTO(out, rc = -ENODEV);
1572 export->exp_in_recovery = 1;
1573 export->exp_req_replay_needed = 1;
1574 export->exp_lock_replay_needed = 1;
1575 spin_unlock(&export->exp_lock);
1577 has_transno = !!(lustre_msg_get_op_flags(req->rq_reqmsg) &
1578 MSG_CONNECT_TRANSNO);
1579 if (has_transno && transno == 0)
1580 CWARN("Connect with zero transno!\n");
1582 if (has_transno && transno > 0 &&
1583 transno < target->obd_next_recovery_transno &&
1584 transno > target->obd_last_committed) {
1585 /* Another way is to use cmpxchg() to be lock-free. */
1586 spin_lock(&target->obd_recovery_task_lock);
1587 if (transno < target->obd_next_recovery_transno)
1588 target->obd_next_recovery_transno = transno;
1589 spin_unlock(&target->obd_recovery_task_lock);
1592 atomic_inc(&target->obd_req_replay_clients);
1593 atomic_inc(&target->obd_lock_replay_clients);
1595 * Note: MDS-MDS connection is allowed to be connected during
1596 * recovery, no matter if the exports needs to be recoveried.
1597 * Because we need retrieve updates logs from all other MDTs.
1598 * So if the MDS-MDS export is new, obd_max_recoverable_clients
1599 * also needs to be increased to match other recovery checking
1602 if (new_mds_mds_conn)
1603 atomic_inc(&target->obd_max_recoverable_clients);
1605 if (atomic_inc_return(&target->obd_connected_clients) ==
1606 atomic_read(&target->obd_max_recoverable_clients))
1607 wake_up(&target->obd_next_transno_waitq);
1610 /* Tell the client we're in recovery, when client is involved in it. */
1611 if (target->obd_recovering && !lw_client)
1612 lustre_msg_add_op_flags(req->rq_repmsg, MSG_CONNECT_RECOVERING);
1616 spin_lock(&export->exp_lock);
1617 export->exp_connecting = 0;
1618 spin_unlock(&export->exp_lock);
1620 class_export_put(export);
1622 if (target != NULL) {
1623 spin_lock(&target->obd_dev_lock);
1624 target->obd_conn_inprogress--;
1625 spin_unlock(&target->obd_dev_lock);
1626 class_decref(target, "find", current);
1629 ptlrpc_connection_put(pcon);
1630 req->rq_status = rc;
1634 int target_handle_disconnect(struct ptlrpc_request *req)
1640 rc = req_capsule_server_pack(&req->rq_pill);
1644 /* In case of target disconnect, updating sec ctx immediately is
1645 * required in order to record latest sequence number used.
1646 * Sequence is normally updated on export destroy, but this event
1647 * can occur too late, ie after a new target connect request has
1649 * Maintaining correct sequence when client connection becomes idle
1650 * ensures that GSS does not erroneously consider requests as replays.
1652 rc = sptlrpc_export_update_ctx(req->rq_export);
1656 /* Keep the rq_export around so we can send the reply. */
1657 req->rq_status = obd_disconnect(class_export_get(req->rq_export));
1662 void target_destroy_export(struct obd_export *exp)
1664 struct obd_import *imp = NULL;
1666 * exports created from last_rcvd data, and "fake"
1667 * exports created by lctl don't have an import
1669 spin_lock(&exp->exp_lock);
1670 if (exp->exp_imp_reverse != NULL) {
1671 imp = exp->exp_imp_reverse;
1672 exp->exp_imp_reverse = NULL;
1674 spin_unlock(&exp->exp_lock);
1676 client_destroy_import(imp);
1678 LASSERT(atomic_read(&exp->exp_locks_count) == 0);
1679 LASSERT(atomic_read(&exp->exp_rpc_count) == 0);
1680 LASSERT(atomic_read(&exp->exp_cb_count) == 0);
1681 LASSERT(atomic_read(&exp->exp_replay_count) == 0);
1683 EXPORT_SYMBOL(target_destroy_export);
1686 * Recovery functions
1688 static void target_request_copy_get(struct ptlrpc_request *req)
1690 class_export_rpc_inc(req->rq_export);
1691 LASSERT(list_empty(&req->rq_list));
1692 INIT_LIST_HEAD(&req->rq_replay_list);
1694 /* Increase refcount to keep request in queue. */
1695 atomic_inc(&req->rq_refcount);
1696 /* Let export know it has replays to be handled. */
1697 atomic_inc(&req->rq_export->exp_replay_count);
1700 static void target_request_copy_put(struct ptlrpc_request *req)
1702 LASSERT(list_empty(&req->rq_replay_list));
1703 LASSERT(atomic_read(&(req)->rq_export->exp_replay_count) > 0);
1705 atomic_dec(&req->rq_export->exp_replay_count);
1706 class_export_rpc_dec(req->rq_export);
1707 ptlrpc_server_drop_request(req);
1710 static int target_exp_enqueue_req_replay(struct ptlrpc_request *req)
1712 __u64 transno = lustre_msg_get_transno(req->rq_reqmsg);
1713 struct obd_export *exp = req->rq_export;
1714 struct ptlrpc_request *reqiter;
1715 struct ptlrpc_request *dup_req = NULL;
1720 spin_lock(&exp->exp_lock);
1721 list_for_each_entry(reqiter, &exp->exp_req_replay_queue,
1723 if (lustre_msg_get_transno(reqiter->rq_reqmsg) == transno) {
1731 /* We expect it with RESENT and REPLAY flags. */
1732 if ((lustre_msg_get_flags(req->rq_reqmsg) &
1733 (MSG_RESENT | MSG_REPLAY)) != (MSG_RESENT | MSG_REPLAY))
1734 CERROR("invalid flags %x of resent replay\n",
1735 lustre_msg_get_flags(req->rq_reqmsg));
1737 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY) {
1740 new_conn = lustre_msg_get_conn_cnt(req->rq_reqmsg);
1742 lustre_msg_get_conn_cnt(dup_req->rq_reqmsg))
1743 lustre_msg_set_conn_cnt(dup_req->rq_reqmsg,
1747 list_add_tail(&req->rq_replay_list,
1748 &exp->exp_req_replay_queue);
1751 spin_unlock(&exp->exp_lock);
1755 static void target_exp_dequeue_req_replay(struct ptlrpc_request *req)
1757 LASSERT(!list_empty(&req->rq_replay_list));
1758 LASSERT(req->rq_export);
1760 spin_lock(&req->rq_export->exp_lock);
1761 list_del_init(&req->rq_replay_list);
1762 spin_unlock(&req->rq_export->exp_lock);
1765 static void target_finish_recovery(struct lu_target *lut)
1767 struct obd_device *obd = lut->lut_obd;
1771 /* Only log a recovery message when recovery has occurred. */
1772 if (obd->obd_recovery_start) {
1773 time64_t now = ktime_get_seconds();
1774 time64_t elapsed_time;
1776 elapsed_time = max_t(time64_t, now - obd->obd_recovery_start,
1778 LCONSOLE_INFO("%s: Recovery over after %lld:%.02lld, of %d clients %d recovered and %d %s evicted.\n",
1779 obd->obd_name, elapsed_time / 60,
1781 atomic_read(&obd->obd_max_recoverable_clients),
1782 atomic_read(&obd->obd_connected_clients),
1783 obd->obd_stale_clients,
1784 obd->obd_stale_clients == 1 ? "was" : "were");
1785 if (obd->obd_stale_clients && do_dump_on_eviction(obd))
1786 libcfs_debug_dumplog();
1789 ldlm_reprocess_recovery_done(obd->obd_namespace);
1790 spin_lock(&obd->obd_recovery_task_lock);
1791 if (!list_empty(&obd->obd_req_replay_queue) ||
1792 !list_empty(&obd->obd_lock_replay_queue) ||
1793 !list_empty(&obd->obd_final_req_queue)) {
1794 CERROR("%s: Recovery queues ( %s%s%s) are not empty\n",
1796 list_empty(&obd->obd_req_replay_queue) ? "" : "req ",
1797 list_empty(&obd->obd_lock_replay_queue) ? \
1799 list_empty(&obd->obd_final_req_queue) ? \
1801 spin_unlock(&obd->obd_recovery_task_lock);
1804 spin_unlock(&obd->obd_recovery_task_lock);
1806 obd->obd_recovery_end = ktime_get_seconds();
1808 /* When recovery finished, cleanup orphans on MDS and OST. */
1809 if (obd->obd_type && obd->obd_type->typ_dt_ops->o_postrecov) {
1810 int rc = obd->obd_type->typ_dt_ops->o_postrecov(obd);
1813 LCONSOLE_WARN("%s: Post recovery failed, rc %d\n",
1819 static void abort_req_replay_queue(struct obd_device *obd)
1821 struct ptlrpc_request *req, *n;
1822 LIST_HEAD(abort_list);
1824 spin_lock(&obd->obd_recovery_task_lock);
1825 list_splice_init(&obd->obd_req_replay_queue, &abort_list);
1826 spin_unlock(&obd->obd_recovery_task_lock);
1827 list_for_each_entry_safe(req, n, &abort_list, rq_list) {
1828 DEBUG_REQ(D_WARNING, req, "aborted:");
1829 req->rq_status = -ENOTCONN;
1830 if (ptlrpc_error(req)) {
1831 DEBUG_REQ(D_ERROR, req,
1832 "failed abort_req_reply; skipping");
1834 target_exp_dequeue_req_replay(req);
1835 target_request_copy_put(req);
1839 static void abort_lock_replay_queue(struct obd_device *obd)
1841 struct ptlrpc_request *req, *n;
1842 LIST_HEAD(abort_list);
1844 spin_lock(&obd->obd_recovery_task_lock);
1845 list_splice_init(&obd->obd_lock_replay_queue, &abort_list);
1846 spin_unlock(&obd->obd_recovery_task_lock);
1847 list_for_each_entry_safe(req, n, &abort_list, rq_list) {
1848 DEBUG_REQ(D_ERROR, req, "aborted:");
1849 req->rq_status = -ENOTCONN;
1850 if (ptlrpc_error(req)) {
1851 DEBUG_REQ(D_ERROR, req,
1852 "failed abort_lock_reply; skipping");
1854 target_request_copy_put(req);
1859 * Called from a cleanup function if the device is being cleaned up
1860 * forcefully. The exports should all have been disconnected already,
1861 * the only thing left to do is
1862 * - clear the recovery flags
1863 * - cancel the timer
1864 * - free queued requests and replies, but don't send replies
1865 * Because the obd_stopping flag is set, no new requests should be received.
1867 void target_cleanup_recovery(struct obd_device *obd)
1869 struct ptlrpc_request *req, *n;
1870 LIST_HEAD(clean_list);
1872 spin_lock(&obd->obd_dev_lock);
1873 if (!obd->obd_recovering) {
1874 spin_unlock(&obd->obd_dev_lock);
1878 obd->obd_recovering = 0;
1879 obd->obd_abort_recovery = 0;
1880 obd->obd_abort_mdt_recovery = 0;
1881 spin_unlock(&obd->obd_dev_lock);
1883 spin_lock(&obd->obd_recovery_task_lock);
1884 target_cancel_recovery_timer(obd);
1885 list_splice_init(&obd->obd_req_replay_queue, &clean_list);
1886 spin_unlock(&obd->obd_recovery_task_lock);
1888 list_for_each_entry_safe(req, n, &clean_list, rq_list) {
1889 LASSERT(req->rq_reply_state == NULL);
1890 target_exp_dequeue_req_replay(req);
1891 target_request_copy_put(req);
1894 spin_lock(&obd->obd_recovery_task_lock);
1895 list_splice_init(&obd->obd_lock_replay_queue, &clean_list);
1896 list_splice_init(&obd->obd_final_req_queue, &clean_list);
1897 spin_unlock(&obd->obd_recovery_task_lock);
1899 list_for_each_entry_safe(req, n, &clean_list, rq_list) {
1900 LASSERT(req->rq_reply_state == NULL);
1901 target_request_copy_put(req);
1906 EXPORT_SYMBOL(target_cleanup_recovery);
1908 /* obd_recovery_task_lock should be held */
1909 void target_cancel_recovery_timer(struct obd_device *obd)
1911 CDEBUG(D_HA, "%s: cancel recovery timer\n", obd->obd_name);
1912 hrtimer_cancel(&obd->obd_recovery_timer);
1915 static void target_start_recovery_timer(struct obd_device *obd)
1919 if (obd->obd_recovery_start != 0)
1922 spin_lock(&obd->obd_dev_lock);
1923 if (!obd->obd_recovering || obd->obd_abort_recovery) {
1924 spin_unlock(&obd->obd_dev_lock);
1928 LASSERT(obd->obd_recovery_timeout != 0);
1930 if (obd->obd_recovery_start != 0) {
1931 spin_unlock(&obd->obd_dev_lock);
1935 obd->obd_recovery_start = ktime_get_seconds();
1936 delay = ktime_set(obd->obd_recovery_start +
1937 obd->obd_recovery_timeout, 0);
1938 hrtimer_start(&obd->obd_recovery_timer, delay, HRTIMER_MODE_ABS);
1939 spin_unlock(&obd->obd_dev_lock);
1941 LCONSOLE_WARN("%s: Will be in recovery for at least %u:%02u, or until %d client%s reconnect%s\n",
1943 obd->obd_recovery_timeout / 60,
1944 obd->obd_recovery_timeout % 60,
1945 atomic_read(&obd->obd_max_recoverable_clients),
1946 (atomic_read(&obd->obd_max_recoverable_clients) == 1) ?
1948 (atomic_read(&obd->obd_max_recoverable_clients) == 1) ?
1953 * extend recovery window.
1955 * if @extend is true, extend recovery window to have @dr_timeout remaining
1956 * at least; otherwise, make sure the recovery timeout value is not less
1959 static void extend_recovery_timer(struct obd_device *obd, timeout_t dr_timeout,
1966 spin_lock(&obd->obd_dev_lock);
1967 if (!obd->obd_recovering || obd->obd_abort_recovery ||
1968 obd->obd_stopping) {
1969 spin_unlock(&obd->obd_dev_lock);
1972 LASSERT(obd->obd_recovery_start != 0);
1974 left_ns = hrtimer_get_remaining(&obd->obd_recovery_timer);
1975 left = ktime_divns(left_ns, NSEC_PER_SEC);
1978 timeout = obd->obd_recovery_timeout;
1979 /* dr_timeout will happen after the hrtimer has expired.
1980 * Add the excess time to the soft recovery timeout without
1981 * exceeding the hard recovery timeout.
1983 if (dr_timeout > left) {
1984 timeout += dr_timeout - left;
1985 timeout = min_t(timeout_t, obd->obd_recovery_time_hard,
1989 timeout = clamp_t(timeout_t, dr_timeout,
1990 obd->obd_recovery_timeout,
1991 obd->obd_recovery_time_hard);
1994 if (timeout == obd->obd_recovery_time_hard)
1995 CWARN("%s: extended recovery timer reached hard limit: %d, extend: %d\n",
1996 obd->obd_name, timeout, extend);
1998 if (obd->obd_recovery_timeout < timeout) {
2001 obd->obd_recovery_timeout = timeout;
2002 end = ktime_set(obd->obd_recovery_start + timeout, 0);
2003 now = ktime_set(ktime_get_seconds(), 0);
2004 left_ns = ktime_sub(end, now);
2005 hrtimer_start(&obd->obd_recovery_timer, end, HRTIMER_MODE_ABS);
2006 left = ktime_divns(left_ns, NSEC_PER_SEC);
2008 spin_unlock(&obd->obd_dev_lock);
2010 CDEBUG(D_HA, "%s: recovery timer will expire in %d seconds\n",
2011 obd->obd_name, left);
2014 /* Reset the timer with each new client connection */
2016 * This timer is actually reconnect_timer, which is for making sure
2017 * the total recovery window is at least as big as my reconnect
2018 * attempt timing. So the initial recovery time_out will be set to
2019 * OBD_RECOVERY_FACTOR * obd_timeout. If the timeout coming
2020 * from client is bigger than this, then the recovery time_out will
2021 * be extended to make sure the client could be reconnected, in the
2022 * process, the timeout from the new client should be ignored.
2025 check_and_start_recovery_timer(struct obd_device *obd,
2026 struct ptlrpc_request *req,
2029 timeout_t service_timeout = lustre_msg_get_service_timeout(req->rq_reqmsg);
2030 struct obd_device_target *obt = obd2obt(obd);
2032 if (!new_client && service_timeout)
2034 * Teach server about old server's estimates, as first guess
2035 * at how long new requests will take.
2037 at_measured(&req->rq_rqbd->rqbd_svcpt->scp_at_estimate,
2040 target_start_recovery_timer(obd);
2043 * Convert the service time to RPC timeout,
2044 * and reuse service_timeout to limit stack usage.
2046 service_timeout = at_est2timeout(service_timeout);
2048 if (CFS_FAIL_CHECK(OBD_FAIL_TGT_SLUGGISH_NET) &&
2049 service_timeout < at_extra)
2050 service_timeout = at_extra;
2053 * We expect other clients to timeout within service_timeout, then try
2054 * to reconnect, then try the failover server. The max delay between
2055 * connect attempts is SWITCH_MAX + SWITCH_INC + INITIAL.
2057 service_timeout += 2 * INITIAL_CONNECT_TIMEOUT;
2059 LASSERT(obt->obt_magic == OBT_MAGIC);
2060 service_timeout += 2 * (CONNECTION_SWITCH_MAX + CONNECTION_SWITCH_INC);
2061 if (service_timeout > obd->obd_recovery_timeout && !new_client)
2062 extend_recovery_timer(obd, service_timeout, false);
2065 /** Health checking routines */
2066 static inline int exp_connect_healthy(struct obd_export *exp)
2068 return exp->exp_in_recovery;
2071 /** if export done req_replay or has replay in queue */
2072 static inline int exp_req_replay_healthy(struct obd_export *exp)
2074 return (!exp->exp_req_replay_needed ||
2075 atomic_read(&exp->exp_replay_count) > 0);
2079 static inline int exp_req_replay_healthy_or_from_mdt(struct obd_export *exp)
2081 return (exp_connect_flags(exp) & OBD_CONNECT_MDS_MDS) ||
2082 exp_req_replay_healthy(exp);
2085 /** if export done lock_replay or has replay in queue */
2086 static inline int exp_lock_replay_healthy(struct obd_export *exp)
2088 return (!exp->exp_lock_replay_needed ||
2089 atomic_read(&exp->exp_replay_count) > 0);
2092 static inline int exp_vbr_healthy(struct obd_export *exp)
2094 return !exp->exp_vbr_failed;
2097 static inline int exp_finished(struct obd_export *exp)
2099 return exp->exp_in_recovery && !exp->exp_lock_replay_needed;
2102 static inline int exp_finished_or_from_mdt(struct obd_export *exp)
2104 return (exp_connect_flags(exp) & OBD_CONNECT_MDS_MDS) ||
2108 static int check_for_next_transno(struct lu_target *lut)
2110 struct ptlrpc_request *req = NULL;
2111 struct obd_device *obd = lut->lut_obd;
2112 struct target_distribute_txn_data *tdtd = lut->lut_tdtd;
2113 int wake_up = 0, connected, completed, queue_len;
2114 __u64 req_transno = 0;
2115 __u64 update_transno = 0;
2116 __u64 next_transno = 0;
2120 spin_lock(&obd->obd_recovery_task_lock);
2121 if (!list_empty(&obd->obd_req_replay_queue)) {
2122 req = list_first_entry(&obd->obd_req_replay_queue,
2123 struct ptlrpc_request, rq_list);
2124 req_transno = lustre_msg_get_transno(req->rq_reqmsg);
2127 if (!obd_mdt_recovery_abort(obd) && tdtd)
2128 update_transno = distribute_txn_get_next_transno(tdtd);
2130 connected = atomic_read(&obd->obd_connected_clients);
2131 completed = connected - atomic_read(&obd->obd_req_replay_clients);
2132 queue_len = obd->obd_requests_queued_for_recovery;
2133 next_transno = obd->obd_next_recovery_transno;
2136 "max: %d, connected: %d, completed: %d, queue_len: %d, req_transno: %llu, next_transno: %llu\n",
2137 atomic_read(&obd->obd_max_recoverable_clients),
2138 connected, completed,
2139 queue_len, req_transno, next_transno);
2141 if (obd_recovery_abort(obd)) {
2142 CDEBUG(D_HA, "waking for aborted recovery\n");
2144 } else if (obd->obd_recovery_expired) {
2145 CDEBUG(D_HA, "waking for expired recovery\n");
2147 } else if (!obd_mdt_recovery_abort(obd) && tdtd && req &&
2148 is_req_replayed_by_update(req)) {
2149 LASSERTF(req_transno < next_transno,
2150 "req_transno %llu next_transno%llu\n", req_transno,
2152 CDEBUG(D_HA, "waking for duplicate req (%llu)\n",
2155 } else if (req_transno == next_transno ||
2156 (update_transno != 0 && update_transno <= next_transno)) {
2157 CDEBUG(D_HA, "waking for next (%lld)\n", next_transno);
2159 } else if (queue_len > 0 &&
2160 queue_len == atomic_read(&obd->obd_req_replay_clients)) {
2161 /** handle gaps occured due to lost reply or VBR */
2162 LASSERTF(req_transno >= next_transno,
2163 "req_transno: %llu, next_transno: %llu\n",
2164 req_transno, next_transno);
2166 "%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",
2167 obd->obd_name, obd->obd_version_recov ? "ON" : "OFF",
2168 next_transno, queue_len, completed, connected,
2169 req_transno, update_transno, obd->obd_last_committed);
2170 obd->obd_next_recovery_transno = req_transno;
2172 } else if (atomic_read(&obd->obd_req_replay_clients) == 0) {
2173 CDEBUG(D_HA, "waking for completed recovery\n");
2175 } else if (CFS_FAIL_CHECK(OBD_FAIL_MDS_RECOVERY_ACCEPTS_GAPS)) {
2177 "accepting transno gaps is explicitly allowed by fail_lock, waking up (%lld)\n",
2179 obd->obd_next_recovery_transno = req_transno;
2182 spin_unlock(&obd->obd_recovery_task_lock);
2186 static int check_for_next_lock(struct lu_target *lut)
2188 struct obd_device *obd = lut->lut_obd;
2191 spin_lock(&obd->obd_recovery_task_lock);
2192 if (!list_empty(&obd->obd_lock_replay_queue)) {
2193 CDEBUG(D_HA, "waking for next lock\n");
2195 } else if (atomic_read(&obd->obd_lock_replay_clients) == 0) {
2196 CDEBUG(D_HA, "waking for completed lock replay\n");
2198 } else if (obd->obd_abort_recovery) {
2199 CDEBUG(D_HA, "waking for aborted recovery\n");
2201 } else if (obd->obd_recovery_expired) {
2202 CDEBUG(D_HA, "waking for expired recovery\n");
2205 spin_unlock(&obd->obd_recovery_task_lock);
2210 static int check_update_llog(struct lu_target *lut)
2212 struct obd_device *obd = lut->lut_obd;
2213 struct target_distribute_txn_data *tdtd = lut->lut_tdtd;
2215 if (obd_mdt_recovery_abort(obd)) {
2216 CDEBUG(D_HA, "waking for aborted recovery\n");
2220 if (atomic_read(&tdtd->tdtd_recovery_threads_count) == 0) {
2221 CDEBUG(D_HA, "waking for completion of reading update log\n");
2229 * wait for recovery events,
2230 * check its status with help of check_routine
2231 * evict dead clients via health_check
2233 static int target_recovery_overseer(struct lu_target *lut,
2234 int (*check_routine)(struct lu_target *),
2235 int (*health_check)(struct obd_export *))
2237 struct obd_device *obd = lut->lut_obd;
2238 struct target_distribute_txn_data *tdtd;
2242 if (obd->obd_recovering && obd->obd_recovery_start == 0) {
2243 now = ktime_get_seconds();
2244 if (now - last > 600) {
2245 LCONSOLE_INFO("%s: in recovery but waiting for the first client to connect\n",
2250 if (obd->obd_recovery_start != 0 && ktime_get_seconds() >=
2251 (obd->obd_recovery_start + obd->obd_recovery_time_hard)) {
2252 __u64 next_update_transno = 0;
2255 * Only abort the recovery if there are no update recovery
2258 spin_lock(&obd->obd_recovery_task_lock);
2259 if (!obd_mdt_recovery_abort(obd) && lut->lut_tdtd) {
2260 next_update_transno =
2261 distribute_txn_get_next_transno(lut->lut_tdtd);
2263 tdtd = lut->lut_tdtd;
2265 * If next_update_transno == 0, it probably because
2266 * updatelog retrieve threads did not get any records
2267 * yet, let's wait those threads stopped
2269 if (next_update_transno == 0) {
2270 spin_unlock(&obd->obd_recovery_task_lock);
2272 while (wait_event_timeout(
2273 tdtd->tdtd_recovery_threads_waitq,
2274 check_update_llog(lut),
2275 cfs_time_seconds(60)) == 0);
2277 spin_lock(&obd->obd_recovery_task_lock);
2278 next_update_transno =
2279 distribute_txn_get_next_transno(tdtd);
2283 if (next_update_transno != 0 && !obd_recovery_abort(obd)) {
2284 obd->obd_next_recovery_transno = next_update_transno;
2285 spin_unlock(&obd->obd_recovery_task_lock);
2287 * Disconnect unfinished exports from clients, and
2288 * keep connection from MDT to make sure the update
2289 * recovery will still keep trying until some one
2290 * manually abort the recovery
2292 class_disconnect_stale_exports(obd,
2293 exp_finished_or_from_mdt);
2294 /* Abort all of replay & replay lock req from clients */
2295 abort_req_replay_queue(obd);
2296 abort_lock_replay_queue(obd);
2298 "%s: there are still update replay (%#llx)in the queue.\n",
2299 obd->obd_name, next_update_transno);
2301 obd->obd_abort_recovery = 1;
2302 spin_unlock(&obd->obd_recovery_task_lock);
2303 CWARN("%s recovery is aborted by hard timeout\n",
2308 while (wait_event_timeout(obd->obd_next_transno_waitq,
2310 cfs_time_seconds(60)) == 0)
2311 ; /* wait indefinitely for event, but don't trigger watchdog */
2313 if (obd_recovery_abort(obd)) {
2314 CWARN("recovery is aborted, evict exports in recovery\n");
2315 if (lut->lut_tdtd != NULL) {
2316 tdtd = lut->lut_tdtd;
2318 * Let's wait all of the update log recovery thread
2322 tdtd->tdtd_recovery_threads_waitq,
2323 atomic_read(&tdtd->tdtd_recovery_threads_count)
2325 /* Then abort the update recovery list */
2326 dtrq_list_destroy(lut->lut_tdtd);
2329 /** evict exports which didn't finish recovery yet */
2330 class_disconnect_stale_exports(obd, exp_finished);
2332 } else if (obd->obd_recovery_expired) {
2333 obd->obd_recovery_expired = 0;
2335 /** If some clients died being recovered, evict them */
2336 LCONSOLE_WARN("%s: recovery is timed out, evict stale exports\n",
2338 /** evict cexports with no replay in queue, they are stalled */
2339 class_disconnect_stale_exports(obd, health_check);
2341 /** continue with VBR */
2342 spin_lock(&obd->obd_dev_lock);
2343 obd->obd_version_recov = 1;
2344 spin_unlock(&obd->obd_dev_lock);
2346 * reset timer, recovery will proceed with versions now,
2347 * timeout is set just to handle reconnection delays
2349 extend_recovery_timer(obd, RECONNECT_DELAY_MAX, true);
2351 * Wait for recovery events again, after evicting bad clients
2358 static struct ptlrpc_request *target_next_replay_lock(struct lu_target *lut)
2360 struct obd_device *obd = lut->lut_obd;
2361 struct ptlrpc_request *req = NULL;
2363 CDEBUG(D_HA, "Waiting for lock\n");
2364 if (target_recovery_overseer(lut, check_for_next_lock,
2365 exp_lock_replay_healthy))
2366 abort_lock_replay_queue(obd);
2368 spin_lock(&obd->obd_recovery_task_lock);
2369 if (!list_empty(&obd->obd_lock_replay_queue)) {
2370 req = list_first_entry(&obd->obd_lock_replay_queue,
2371 struct ptlrpc_request, rq_list);
2372 list_del_init(&req->rq_list);
2373 spin_unlock(&obd->obd_recovery_task_lock);
2375 spin_unlock(&obd->obd_recovery_task_lock);
2376 LASSERT(list_empty(&obd->obd_lock_replay_queue));
2377 LASSERT(atomic_read(&obd->obd_lock_replay_clients) == 0);
2378 /** evict exports failed VBR */
2379 class_disconnect_stale_exports(obd, exp_vbr_healthy);
2384 static struct ptlrpc_request *target_next_final_ping(struct obd_device *obd)
2386 struct ptlrpc_request *req = NULL;
2388 spin_lock(&obd->obd_recovery_task_lock);
2389 if (!list_empty(&obd->obd_final_req_queue)) {
2390 req = list_first_entry(&obd->obd_final_req_queue,
2391 struct ptlrpc_request, rq_list);
2392 list_del_init(&req->rq_list);
2393 spin_unlock(&obd->obd_recovery_task_lock);
2394 if (req->rq_export->exp_in_recovery) {
2395 spin_lock(&req->rq_export->exp_lock);
2396 req->rq_export->exp_in_recovery = 0;
2397 spin_unlock(&req->rq_export->exp_lock);
2400 spin_unlock(&obd->obd_recovery_task_lock);
2405 static void handle_recovery_req(struct ptlrpc_thread *thread,
2406 struct ptlrpc_request *req,
2407 svc_handler_t handler)
2412 * export can be evicted during recovery, no need to handle replays for
2413 * it after that, discard such request silently
2415 if (req->rq_export->exp_disconnected)
2418 req->rq_session.lc_thread = thread;
2419 req->rq_svc_thread = thread;
2420 req->rq_svc_thread->t_env->le_ses = &req->rq_session;
2422 /* thread context */
2423 lu_context_enter(&thread->t_env->le_ctx);
2425 lu_context_exit(&thread->t_env->le_ctx);
2427 req->rq_svc_thread->t_env->le_ses = NULL;
2429 /* don't reset timer for final stage */
2430 if (!exp_finished(req->rq_export)) {
2431 timeout_t timeout = obd_timeout;
2434 * Add request @timeout to the recovery time so next request from
2435 * this client may come in recovery time
2438 struct ptlrpc_service_part *svcpt;
2439 timeout_t est_timeout;
2441 svcpt = req->rq_rqbd->rqbd_svcpt;
2443 * If the server sent early reply for this request,
2444 * the client will recalculate the timeout according to
2445 * current server estimate service time, so we will
2446 * use the maxium timeout here for waiting the client
2447 * sending the next req
2449 est_timeout = at_get(&svcpt->scp_at_estimate);
2450 timeout = max_t(timeout_t, at_est2timeout(est_timeout),
2451 lustre_msg_get_timeout(req->rq_reqmsg));
2453 * Add 2 net_latency, one for balance rq_deadline
2454 * (see ptl_send_rpc), one for resend the req to server,
2455 * Note: client will pack net_latency in replay req
2456 * (see ptlrpc_replay_req)
2458 timeout += 2 * lustre_msg_get_service_timeout(req->rq_reqmsg);
2460 extend_recovery_timer(class_exp2obd(req->rq_export), timeout,
2466 /** Checking routines for recovery */
2467 static int check_for_recovery_ready(struct lu_target *lut)
2469 struct obd_device *obd = lut->lut_obd;
2470 unsigned int clnts = atomic_read(&obd->obd_connected_clients);
2473 "connected %d stale %d max_recoverable_clients %d abort %d expired %d\n",
2474 clnts, obd->obd_stale_clients,
2475 atomic_read(&obd->obd_max_recoverable_clients),
2476 obd->obd_abort_recovery, obd->obd_recovery_expired);
2478 if (!obd_recovery_abort(obd) && !obd->obd_recovery_expired) {
2480 atomic_read(&obd->obd_max_recoverable_clients));
2481 if (clnts + obd->obd_stale_clients <
2482 atomic_read(&obd->obd_max_recoverable_clients))
2486 if (!obd_mdt_recovery_abort(obd) && lut->lut_tdtd &&
2487 !lut->lut_tdtd->tdtd_replay_ready) {
2488 /* Let's extend recovery timer, in case the recovery timer
2489 * expired, and some clients got evicted
2491 extend_recovery_timer(obd, obd->obd_recovery_timeout, true);
2493 "%s update recovery is not ready, extend recovery %d\n",
2494 obd->obd_name, obd->obd_recovery_timeout);
2502 REQUEST_RECOVERY = 1,
2503 UPDATE_RECOVERY = 2,
2506 static __u64 get_next_replay_req_transno(struct obd_device *obd)
2510 if (!list_empty(&obd->obd_req_replay_queue)) {
2511 struct ptlrpc_request *req;
2513 req = list_first_entry(&obd->obd_req_replay_queue,
2514 struct ptlrpc_request, rq_list);
2515 transno = lustre_msg_get_transno(req->rq_reqmsg);
2521 static __u64 get_next_transno(struct lu_target *lut, int *type)
2523 struct obd_device *obd = lut->lut_obd;
2524 struct target_distribute_txn_data *tdtd = lut->lut_tdtd;
2526 __u64 update_transno;
2530 transno = get_next_replay_req_transno(obd);
2532 *type = REQUEST_RECOVERY;
2534 if (!tdtd || obd_mdt_recovery_abort(obd))
2537 update_transno = distribute_txn_get_next_transno(tdtd);
2538 if (transno == 0 || (transno >= update_transno &&
2539 update_transno != 0)) {
2540 transno = update_transno;
2542 *type = UPDATE_RECOVERY;
2549 * drop duplicate replay request
2551 * Because the operation has been replayed by update recovery, the request
2552 * with the same transno will be dropped and also notify the client to send
2553 * next replay request.
2555 * \param[in] env execution environment
2556 * \param[in] obd failover obd device
2557 * \param[in] req request to be dropped
2559 * \retval true duplicate replay update is dropped
2560 * \retval false duplicate replay update is not dropped
2562 static bool drop_duplicate_replay_req(struct lu_env *env,
2563 struct obd_device *obd,
2564 struct ptlrpc_request *req)
2566 __u32 opc = lustre_msg_get_opc(req->rq_reqmsg);
2568 DEBUG_REQ(D_HA, req,
2569 "remove t%lld from %s because duplicate update records found",
2570 lustre_msg_get_transno(req->rq_reqmsg),
2571 libcfs_nidstr(&req->rq_peer.nid));
2574 * Right now, only for MDS reint operation update replay and
2575 * normal request replay can have the same transno
2577 if (opc == MDS_REINT) {
2578 req_capsule_set(&req->rq_pill, &RQF_MDS_REINT);
2579 req->rq_status = req_capsule_server_pack(&req->rq_pill);
2580 if (likely(req->rq_export))
2581 target_committed_to_req(req);
2582 lustre_msg_set_transno(req->rq_repmsg, req->rq_transno);
2583 target_send_reply(req, req->rq_status, 0);
2584 } else if (opc == MDS_CLOSE) {
2585 DEBUG_REQ(D_HA, req, "duplicate close replay from %s\n",
2586 libcfs_nidstr(&req->rq_peer.nid));
2589 DEBUG_REQ(D_ERROR, req, "wrong opc %d from %s\n", opc,
2590 libcfs_nidstr(&req->rq_peer.nid));
2592 target_exp_dequeue_req_replay(req);
2593 target_request_copy_put(req);
2594 obd->obd_replayed_requests++;
2599 #define WATCHDOG_TIMEOUT (obd_timeout * 10)
2601 static void replay_request_or_update(struct lu_env *env,
2602 struct lu_target *lut,
2603 struct target_recovery_data *trd,
2604 struct ptlrpc_thread *thread)
2606 struct obd_device *obd = lut->lut_obd;
2607 struct ptlrpc_request *req = NULL;
2613 CDEBUG(D_HA, "Waiting for transno %lld\n",
2614 obd->obd_next_recovery_transno);
2616 /* Replay all of request and update by transno */
2618 struct target_distribute_txn_data *tdtd = lut->lut_tdtd;
2620 CFS_FAIL_TIMEOUT(OBD_FAIL_TGT_REPLAY_DELAY2, cfs_fail_val);
2623 * It is needed to extend recovery window above
2624 * recovery_time_soft. Extending is possible only in the
2625 * end of recovery window (see more details in
2626 * handle_recovery_req()).
2628 CFS_FAIL_TIMEOUT_MS(OBD_FAIL_TGT_REPLAY_DELAY, 300);
2630 if (target_recovery_overseer(lut, check_for_next_transno,
2631 exp_req_replay_healthy_or_from_mdt)) {
2632 abort_req_replay_queue(obd);
2633 abort_lock_replay_queue(obd);
2637 spin_lock(&obd->obd_recovery_task_lock);
2638 transno = get_next_transno(lut, &type);
2639 if (type == REQUEST_RECOVERY && transno != 0) {
2640 bool update = false;
2642 * Drop replay request from client side, if the
2643 * replay has been executed by update with the
2646 req = list_first_entry(&obd->obd_req_replay_queue,
2647 struct ptlrpc_request, rq_list);
2649 list_del_init(&req->rq_list);
2650 obd->obd_requests_queued_for_recovery--;
2651 spin_unlock(&obd->obd_recovery_task_lock);
2654 * Let's check if the request has been redone by
2657 if (is_req_replayed_by_update(req)) {
2658 struct distribute_txn_replay_req *dtrq;
2660 dtrq = distribute_txn_lookup_finish_list(tdtd,
2662 LASSERT(dtrq != NULL);
2663 spin_lock(&tdtd->tdtd_replay_list_lock);
2664 list_del_init(&dtrq->dtrq_list);
2665 spin_unlock(&tdtd->tdtd_replay_list_lock);
2668 if (drop_duplicate_replay_req(env, obd, req))
2670 /* not dropped yet */
2674 LASSERT(trd->trd_processing_task == current->pid);
2675 DEBUG_REQ(D_HA, req, "processing x%llu t%lld from %s",
2677 lustre_msg_get_transno(req->rq_reqmsg),
2678 libcfs_nidstr(&req->rq_peer.nid));
2680 ptlrpc_watchdog_init(&thread->t_watchdog,
2682 handle_recovery_req(thread, req,
2683 trd->trd_recovery_handler);
2684 ptlrpc_watchdog_disable(&thread->t_watchdog);
2687 * bz18031: increase next_recovery_transno before
2688 * target_request_copy_put() will drop exp_rpc reference
2691 spin_lock(&obd->obd_recovery_task_lock);
2692 obd->obd_next_recovery_transno++;
2693 spin_unlock(&obd->obd_recovery_task_lock);
2695 target_exp_dequeue_req_replay(req);
2696 target_request_copy_put(req);
2697 obd->obd_replayed_requests++;
2698 } else if (type == UPDATE_RECOVERY && transno != 0) {
2699 struct distribute_txn_replay_req *dtrq;
2702 spin_unlock(&obd->obd_recovery_task_lock);
2704 LASSERT(tdtd != NULL);
2705 dtrq = distribute_txn_get_next_req(tdtd);
2706 lu_context_enter(&thread->t_env->le_ctx);
2707 ptlrpc_watchdog_init(&thread->t_watchdog,
2709 rc = tdtd->tdtd_replay_handler(env, tdtd, dtrq);
2710 ptlrpc_watchdog_disable(&thread->t_watchdog);
2711 lu_context_exit(&thread->t_env->le_ctx);
2712 extend_recovery_timer(obd, obd_timeout, true);
2714 if (rc == 0 && dtrq->dtrq_xid != 0) {
2716 "Move x%llu t%llu to finish list\n",
2718 dtrq->dtrq_master_transno);
2720 /* Add it to the replay finish list */
2721 spin_lock(&tdtd->tdtd_replay_list_lock);
2722 list_add(&dtrq->dtrq_list,
2723 &tdtd->tdtd_replay_finish_list);
2724 spin_unlock(&tdtd->tdtd_replay_list_lock);
2726 spin_lock(&obd->obd_recovery_task_lock);
2727 if (transno == obd->obd_next_recovery_transno)
2728 obd->obd_next_recovery_transno++;
2730 obd->obd_next_recovery_transno)
2731 obd->obd_next_recovery_transno =
2733 spin_unlock(&obd->obd_recovery_task_lock);
2738 spin_unlock(&obd->obd_recovery_task_lock);
2740 LASSERT(list_empty(&obd->obd_req_replay_queue));
2741 LASSERT(atomic_read(&obd->obd_req_replay_clients) == 0);
2742 /** evict exports failed VBR */
2743 class_disconnect_stale_exports(obd, exp_vbr_healthy);
2749 static int target_recovery_thread(void *arg)
2751 struct lu_target *lut = arg;
2752 struct obd_device *obd = lut->lut_obd;
2753 struct ptlrpc_request *req;
2754 struct target_recovery_data *trd = &obd->obd_recovery_data;
2755 unsigned long delta;
2757 struct ptlrpc_thread *thread = NULL;
2761 unshare_fs_struct();
2762 OBD_ALLOC_PTR(thread);
2768 GOTO(out_thread, rc = -ENOMEM);
2769 rc = lu_env_add(env);
2773 rc = lu_context_init(&env->le_ctx, LCT_MD_THREAD | LCT_DT_THREAD);
2775 GOTO(out_env_remove, rc);
2777 thread->t_env = env;
2778 thread->t_id = -1; /* force filter_iobuf_get/put to use local buffers */
2779 thread->t_task = current;
2780 env->le_ctx.lc_thread = thread;
2781 tgt_io_thread_init(thread); /* init thread_big_cache for IO requests */
2783 CDEBUG(D_HA, "%s: started recovery thread pid %d\n", obd->obd_name,
2785 trd->trd_processing_task = current->pid;
2787 spin_lock(&obd->obd_dev_lock);
2788 obd->obd_recovering = 1;
2789 spin_unlock(&obd->obd_dev_lock);
2790 complete(&trd->trd_starting);
2792 /* first of all, we have to know the first transno to replay */
2793 if (target_recovery_overseer(lut, check_for_recovery_ready,
2794 exp_connect_healthy)) {
2795 abort_req_replay_queue(obd);
2796 abort_lock_replay_queue(obd);
2797 if (lut->lut_tdtd != NULL)
2798 dtrq_list_destroy(lut->lut_tdtd);
2801 /* next stage: replay requests or update */
2803 CDEBUG(D_INFO, "1: request replay stage - %d clients from t%llu\n",
2804 atomic_read(&obd->obd_req_replay_clients),
2805 obd->obd_next_recovery_transno);
2806 replay_request_or_update(env, lut, trd, thread);
2809 * The second stage: replay locks
2811 CDEBUG(D_INFO, "2: lock replay stage - %d clients\n",
2812 atomic_read(&obd->obd_lock_replay_clients));
2813 while ((req = target_next_replay_lock(lut))) {
2814 LASSERT(trd->trd_processing_task == current->pid);
2815 DEBUG_REQ(D_HA, req, "processing lock from %s:",
2816 libcfs_nidstr(&req->rq_peer.nid));
2817 if (CFS_FAIL_CHECK(OBD_FAIL_LDLM_LOCK_REPLAY)) {
2818 req->rq_status = -ENODEV;
2819 target_request_copy_put(req);
2822 handle_recovery_req(thread, req,
2823 trd->trd_recovery_handler);
2824 target_request_copy_put(req);
2825 obd->obd_replayed_locks++;
2829 * The third stage: reply on final pings, at this moment all clients
2830 * must have request in final queue
2832 CFS_FAIL_TIMEOUT(OBD_FAIL_TGT_REPLAY_RECONNECT, cfs_fail_val);
2833 CDEBUG(D_INFO, "3: final stage - process recovery completion pings\n");
2834 /** Update server last boot epoch */
2835 tgt_boot_epoch_update(lut);
2837 /* cancel update llogs upon recovery abort */
2838 if (obd->obd_abort_recovery || obd->obd_abort_mdt_recovery)
2839 obd->obd_type->typ_dt_ops->o_iocontrol(OBD_IOC_LLOG_CANCEL,
2840 obd->obd_self_export,
2843 list_for_each_entry(req, &obd->obd_final_req_queue, rq_list) {
2845 * Because the waiting client can not send ping to server,
2846 * so we need refresh the last_request_time, to avoid the
2847 * export is being evicted
2849 ptlrpc_update_export_timer(req->rq_export, 0);
2853 * We drop recoverying flag to forward all new requests
2854 * to regular mds_handle() since now
2856 spin_lock(&obd->obd_dev_lock);
2857 obd->obd_recovering = 0;
2858 obd->obd_abort_recovery = 0;
2859 obd->obd_abort_mdt_recovery = 0;
2860 spin_unlock(&obd->obd_dev_lock);
2861 spin_lock(&obd->obd_recovery_task_lock);
2862 target_cancel_recovery_timer(obd);
2863 spin_unlock(&obd->obd_recovery_task_lock);
2864 while ((req = target_next_final_ping(obd))) {
2865 LASSERT(trd->trd_processing_task == current->pid);
2866 DEBUG_REQ(D_HA, req, "processing final ping from %s:",
2867 libcfs_nidstr(&req->rq_peer.nid));
2868 handle_recovery_req(thread, req,
2869 trd->trd_recovery_handler);
2870 target_request_copy_put(req);
2873 delta = jiffies_to_msecs(jiffies - delta) / MSEC_PER_SEC;
2874 CDEBUG(D_INFO, "4: recovery completed in %lus - %d/%d reqs/locks\n",
2875 delta, obd->obd_replayed_requests, obd->obd_replayed_locks);
2876 if (delta > OBD_RECOVERY_TIME_SOFT) {
2877 CWARN("too long recovery - read logs\n");
2878 libcfs_debug_dumplog();
2881 target_finish_recovery(lut);
2882 lu_context_fini(&env->le_ctx);
2883 trd->trd_processing_task = 0;
2884 complete_all(&trd->trd_finishing);
2885 tgt_io_thread_done(thread);
2891 OBD_FREE_PTR(thread);
2895 static int target_start_recovery_thread(struct lu_target *lut,
2896 svc_handler_t handler)
2898 struct obd_device *obd = lut->lut_obd;
2900 struct target_recovery_data *trd = &obd->obd_recovery_data;
2903 memset(trd, 0, sizeof(*trd));
2904 init_completion(&trd->trd_starting);
2905 init_completion(&trd->trd_finishing);
2906 trd->trd_recovery_handler = handler;
2908 rc = server_name2index(obd->obd_name, &index, NULL);
2912 if (!IS_ERR(kthread_run(target_recovery_thread,
2913 lut, "tgt_recover_%d", index))) {
2914 wait_for_completion(&trd->trd_starting);
2915 LASSERT(obd->obd_recovering != 0);
2923 void target_stop_recovery_thread(struct obd_device *obd)
2925 if (obd->obd_recovery_data.trd_processing_task > 0) {
2926 struct target_recovery_data *trd = &obd->obd_recovery_data;
2927 /** recovery can be done but postrecovery is not yet */
2928 spin_lock(&obd->obd_dev_lock);
2929 if (obd->obd_recovering) {
2930 CERROR("%s: Aborting recovery\n", obd->obd_name);
2931 obd->obd_abort_recovery = 1;
2932 wake_up(&obd->obd_next_transno_waitq);
2934 spin_unlock(&obd->obd_dev_lock);
2935 wait_for_completion(&trd->trd_finishing);
2938 EXPORT_SYMBOL(target_stop_recovery_thread);
2940 void target_recovery_fini(struct obd_device *obd)
2942 class_disconnect_exports(obd);
2943 target_stop_recovery_thread(obd);
2944 target_cleanup_recovery(obd);
2946 EXPORT_SYMBOL(target_recovery_fini);
2948 static enum hrtimer_restart target_recovery_expired(struct hrtimer *timer)
2950 struct obd_device *obd = container_of(timer, struct obd_device,
2951 obd_recovery_timer);
2954 "%s: recovery timed out; %d clients are still in recovery after %llu seconds (%d clients connected)\n",
2955 obd->obd_name, atomic_read(&obd->obd_lock_replay_clients),
2956 ktime_get_seconds() - obd->obd_recovery_start,
2957 atomic_read(&obd->obd_connected_clients));
2959 obd->obd_recovery_expired = 1;
2960 wake_up(&obd->obd_next_transno_waitq);
2961 return HRTIMER_NORESTART;
2964 void target_recovery_init(struct lu_target *lut, svc_handler_t handler)
2966 struct obd_device *obd = lut->lut_obd;
2968 if (lut->lut_bottom->dd_rdonly)
2971 if (atomic_read(&obd->obd_max_recoverable_clients) == 0) {
2972 /** Update server last boot epoch */
2973 tgt_boot_epoch_update(lut);
2977 CDEBUG(D_HA, "RECOVERY: service %s, %d recoverable clients, "
2978 "last_transno %llu\n", obd->obd_name,
2979 atomic_read(&obd->obd_max_recoverable_clients),
2980 obd->obd_last_committed);
2981 LASSERT(obd->obd_stopping == 0);
2982 obd->obd_next_recovery_transno = obd->obd_last_committed + 1;
2983 obd->obd_recovery_start = 0;
2984 obd->obd_recovery_end = 0;
2986 hrtimer_init(&obd->obd_recovery_timer, CLOCK_MONOTONIC,
2988 obd->obd_recovery_timer.function = &target_recovery_expired;
2989 target_start_recovery_thread(lut, handler);
2991 EXPORT_SYMBOL(target_recovery_init);
2993 static int target_process_req_flags(struct obd_device *obd,
2994 struct ptlrpc_request *req)
2996 struct obd_export *exp = req->rq_export;
2998 LASSERT(exp != NULL);
2999 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REQ_REPLAY_DONE) {
3000 /* client declares he's ready to replay locks */
3001 spin_lock(&exp->exp_lock);
3002 if (exp->exp_req_replay_needed) {
3003 exp->exp_req_replay_needed = 0;
3004 spin_unlock(&exp->exp_lock);
3006 LASSERT(atomic_read(&(obd)->obd_req_replay_clients) >
3008 atomic_dec(&obd->obd_req_replay_clients);
3010 spin_unlock(&exp->exp_lock);
3013 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_LOCK_REPLAY_DONE) {
3015 * client declares he's ready to complete recovery
3016 * so, we put the request on th final queue
3018 spin_lock(&exp->exp_lock);
3019 if (exp->exp_lock_replay_needed) {
3020 exp->exp_lock_replay_needed = 0;
3021 spin_unlock(&exp->exp_lock);
3023 LASSERT(atomic_read(&(obd)->obd_lock_replay_clients) >
3025 atomic_dec(&obd->obd_lock_replay_clients);
3027 spin_unlock(&exp->exp_lock);
3033 int target_queue_recovery_request(struct ptlrpc_request *req,
3034 struct obd_device *obd)
3036 __u64 transno = lustre_msg_get_transno(req->rq_reqmsg);
3037 struct ptlrpc_request *reqiter;
3042 if (obd->obd_recovery_data.trd_processing_task == current->pid) {
3043 /* Processing the queue right now, don't re-add. */
3047 target_process_req_flags(obd, req);
3049 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_LOCK_REPLAY_DONE) {
3050 if (unlikely(CFS_FAIL_CHECK(OBD_FAIL_TGT_RECOVERY_REQ_RACE))) {
3051 if (cfs_fail_val == 1) {
3054 wake_up(&cfs_race_waitq);
3056 schedule_timeout_interruptible(
3057 cfs_time_seconds(1));
3062 * client declares he's ready to complete recovery
3063 * so, we put the request on th final queue
3065 target_request_copy_get(req);
3066 DEBUG_REQ(D_HA, req, "queue final req");
3067 wake_up(&obd->obd_next_transno_waitq);
3068 spin_lock(&obd->obd_recovery_task_lock);
3069 if (obd->obd_recovering) {
3070 struct ptlrpc_request *tmp;
3071 struct ptlrpc_request *duplicate = NULL;
3073 if (likely(!req->rq_export->exp_replay_done)) {
3074 req->rq_export->exp_replay_done = 1;
3075 list_add_tail(&req->rq_list,
3076 &obd->obd_final_req_queue);
3077 spin_unlock(&obd->obd_recovery_task_lock);
3082 * XXX O(n), but only happens if final ping is
3083 * timed out, probably reorganize the list as
3086 list_for_each_entry_safe(reqiter, tmp,
3087 &obd->obd_final_req_queue,
3089 if (reqiter->rq_export == req->rq_export) {
3090 list_del_init(&reqiter->rq_list);
3091 duplicate = reqiter;
3096 list_add_tail(&req->rq_list,
3097 &obd->obd_final_req_queue);
3098 req->rq_export->exp_replay_done = 1;
3099 spin_unlock(&obd->obd_recovery_task_lock);
3101 if (duplicate != NULL) {
3102 DEBUG_REQ(D_HA, duplicate,
3103 "put prev final req");
3104 target_request_copy_put(duplicate);
3108 spin_unlock(&obd->obd_recovery_task_lock);
3109 target_request_copy_put(req);
3110 RETURN(obd->obd_stopping ? -ENOTCONN : 1);
3113 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REQ_REPLAY_DONE) {
3114 /* client declares he's ready to replay locks */
3115 target_request_copy_get(req);
3116 DEBUG_REQ(D_HA, req, "queue lock replay req");
3117 wake_up(&obd->obd_next_transno_waitq);
3118 spin_lock(&obd->obd_recovery_task_lock);
3119 LASSERT(obd->obd_recovering);
3120 /* usually due to recovery abort */
3121 if (!req->rq_export->exp_in_recovery) {
3122 spin_unlock(&obd->obd_recovery_task_lock);
3123 target_request_copy_put(req);
3126 LASSERT(req->rq_export->exp_lock_replay_needed);
3127 list_add_tail(&req->rq_list, &obd->obd_lock_replay_queue);
3128 spin_unlock(&obd->obd_recovery_task_lock);
3133 * CAVEAT EMPTOR: The incoming request message has been swabbed
3134 * (i.e. buflens etc are in my own byte order), but type-dependent
3135 * buffers (eg mdt_body, ost_body etc) have NOT been swabbed.
3139 INIT_LIST_HEAD(&req->rq_list);
3140 DEBUG_REQ(D_HA, req, "not queueing");
3145 * If we're processing the queue, we want don't want to queue this
3148 * Also, if this request has a transno less than the one we're waiting
3149 * for, we should process it now. It could (and currently always will)
3150 * be an open request for a descriptor that was opened some time ago.
3152 * Also, a resent, replayed request that has already been
3153 * handled will pass through here and be processed immediately.
3156 "Next recovery transno: %llu, current: %llu, replaying\n",
3157 obd->obd_next_recovery_transno, transno);
3160 * If the request has been replayed by update replay, then sends this
3161 * request to the recovery thread (replay_request_or_update()), where
3162 * it will be handled
3164 spin_lock(&obd->obd_recovery_task_lock);
3165 if (transno < obd->obd_next_recovery_transno &&
3166 !is_req_replayed_by_update(req)) {
3167 /* Processing the queue right now, don't re-add. */
3168 LASSERT(list_empty(&req->rq_list));
3169 spin_unlock(&obd->obd_recovery_task_lock);
3172 spin_unlock(&obd->obd_recovery_task_lock);
3174 if (CFS_FAIL_CHECK(OBD_FAIL_TGT_REPLAY_DROP))
3177 target_request_copy_get(req);
3178 if (!req->rq_export->exp_in_recovery) {
3179 target_request_copy_put(req);
3182 LASSERT(req->rq_export->exp_req_replay_needed);
3184 if (target_exp_enqueue_req_replay(req)) {
3185 DEBUG_REQ(D_ERROR, req, "dropping resent queued req");
3186 target_request_copy_put(req);
3191 spin_lock(&obd->obd_recovery_task_lock);
3192 LASSERT(obd->obd_recovering);
3193 list_for_each_entry(reqiter, &obd->obd_req_replay_queue, rq_list) {
3194 if (lustre_msg_get_transno(reqiter->rq_reqmsg) > transno) {
3195 list_add_tail(&req->rq_list, &reqiter->rq_list);
3200 if (unlikely(lustre_msg_get_transno(reqiter->rq_reqmsg) ==
3202 DEBUG_REQ(D_ERROR, req,
3203 "dropping replay: transno has been claimed by another client");
3204 spin_unlock(&obd->obd_recovery_task_lock);
3205 target_exp_dequeue_req_replay(req);
3206 target_request_copy_put(req);
3212 list_add_tail(&req->rq_list, &obd->obd_req_replay_queue);
3214 obd->obd_requests_queued_for_recovery++;
3215 spin_unlock(&obd->obd_recovery_task_lock);
3216 wake_up(&obd->obd_next_transno_waitq);
3220 void target_committed_to_req(struct ptlrpc_request *req)
3222 struct obd_export *exp = req->rq_export;
3224 if (!exp->exp_obd->obd_no_transno && req->rq_repmsg != NULL)
3225 lustre_msg_set_last_committed(req->rq_repmsg,
3226 exp->exp_last_committed);
3228 DEBUG_REQ(D_IOCTL, req,
3229 "not sending last_committed update (%d/%d)",
3230 exp->exp_obd->obd_no_transno,
3231 req->rq_repmsg == NULL);
3233 CDEBUG(D_INFO, "last_committed %llu, transno %llu, xid %llu\n",
3234 exp->exp_last_committed, req->rq_transno, req->rq_xid);
3237 #endif /* HAVE_SERVER_SUPPORT */
3240 * Packs current SLV and Limit into \a req.
3242 int target_pack_pool_reply(struct ptlrpc_request *req)
3244 struct obd_device *obd;
3249 * Check that we still have all structures alive as this may
3250 * be some late RPC at shutdown time.
3252 if (unlikely(!req->rq_export || !req->rq_export->exp_obd ||
3253 !exp_connect_lru_resize(req->rq_export))) {
3254 lustre_msg_set_slv(req->rq_repmsg, 0);
3255 lustre_msg_set_limit(req->rq_repmsg, 0);
3259 /* OBD is alive here as export is alive, which we checked above. */
3260 obd = req->rq_export->exp_obd;
3262 read_lock(&obd->obd_pool_lock);
3263 lustre_msg_set_slv(req->rq_repmsg, obd->obd_pool_slv);
3264 lustre_msg_set_limit(req->rq_repmsg, obd->obd_pool_limit);
3265 read_unlock(&obd->obd_pool_lock);
3270 static int target_send_reply_msg(struct ptlrpc_request *req,
3271 int rc, int fail_id)
3273 if (CFS_FAIL_CHECK_ORSET(fail_id & ~CFS_FAIL_ONCE, CFS_FAIL_ONCE)) {
3274 DEBUG_REQ(D_ERROR, req, "dropping reply");
3278 * We can have a null rq_reqmsg in the event of bad signature or
3279 * no context when unwrapping
3281 if (req->rq_reqmsg &&
3282 unlikely(lustre_msg_get_opc(req->rq_reqmsg) == MDS_REINT &&
3283 CFS_FAIL_CHECK(OBD_FAIL_MDS_REINT_MULTI_NET_REP)))
3287 DEBUG_REQ(D_NET, req, "processing error (%d)", rc);
3288 req->rq_status = rc;
3289 return ptlrpc_send_error(req, 1);
3291 DEBUG_REQ(D_NET, req, "sending reply");
3293 return ptlrpc_send_reply(req, PTLRPC_REPLY_MAYBE_DIFFICULT);
3296 void target_send_reply(struct ptlrpc_request *req, int rc, int fail_id)
3298 struct ptlrpc_service_part *svcpt;
3300 struct ptlrpc_reply_state *rs;
3301 struct obd_export *exp;
3305 if (req->rq_no_reply) {
3310 svcpt = req->rq_rqbd->rqbd_svcpt;
3311 rs = req->rq_reply_state;
3312 if (rs == NULL || !rs->rs_difficult) {
3314 target_send_reply_msg(req, rc, fail_id);
3319 /* must be an export if locks saved */
3320 LASSERT(req->rq_export != NULL);
3321 /* req/reply consistent */
3322 LASSERT(rs->rs_svcpt == svcpt);
3325 LASSERT(!rs->rs_scheduled);
3326 LASSERT(!rs->rs_scheduled_ever);
3327 LASSERT(!rs->rs_handled);
3328 LASSERT(!rs->rs_sent);
3329 LASSERT(!rs->rs_unlinked);
3330 LASSERT(rs->rs_export == NULL);
3331 LASSERT(list_empty(&rs->rs_obd_list));
3332 LASSERT(list_empty(&rs->rs_exp_list));
3334 exp = class_export_get(req->rq_export);
3336 /* disable reply scheduling while I'm setting up */
3337 rs->rs_scheduled = 1;
3339 rs->rs_unlinked = 0;
3340 rs->rs_xid = req->rq_xid;
3341 rs->rs_transno = req->rq_transno;
3342 rs->rs_export = exp;
3343 rs->rs_opc = lustre_msg_get_opc(req->rq_reqmsg);
3345 spin_lock(&exp->exp_uncommitted_replies_lock);
3346 CDEBUG(D_NET, "rs transno = %llu, last committed = %llu\n",
3347 rs->rs_transno, exp->exp_last_committed);
3348 if (rs->rs_transno > exp->exp_last_committed) {
3349 /* not committed already */
3350 list_add_tail(&rs->rs_obd_list,
3351 &exp->exp_uncommitted_replies);
3353 spin_unlock(&exp->exp_uncommitted_replies_lock);
3355 spin_lock(&exp->exp_lock);
3356 list_add_tail(&rs->rs_exp_list, &exp->exp_outstanding_replies);
3357 spin_unlock(&exp->exp_lock);
3359 netrc = target_send_reply_msg(req, rc, fail_id);
3361 spin_lock(&svcpt->scp_rep_lock);
3363 atomic_inc(&svcpt->scp_nreps_difficult);
3367 * error sending: reply is off the net. Also we need +1
3368 * reply ref until ptlrpc_handle_rs() is done
3369 * with the reply state (if the send was successful, there
3370 * would have been +1 ref for the net, which
3371 * reply_out_callback leaves alone)
3374 rs->rs_unlinked = 1;
3375 ptlrpc_rs_addref(rs);
3378 spin_lock(&rs->rs_lock);
3379 if (rs->rs_transno <= exp->exp_last_committed ||
3380 (rs->rs_unlinked && !rs->rs_no_ack) ||
3381 list_empty(&rs->rs_exp_list) || /* completed already */
3382 list_empty(&rs->rs_obd_list)) {
3383 CDEBUG(D_HA, "Schedule reply immediately\n");
3384 ptlrpc_dispatch_difficult_reply(rs);
3386 list_add(&rs->rs_list, &svcpt->scp_rep_active);
3387 rs->rs_scheduled = 0; /* allow notifier to schedule */
3389 spin_unlock(&rs->rs_lock);
3390 spin_unlock(&svcpt->scp_rep_lock);
3394 enum ldlm_mode lck_compat_array[] = {
3395 [LCK_EX] = LCK_COMPAT_EX,
3396 [LCK_PW] = LCK_COMPAT_PW,
3397 [LCK_PR] = LCK_COMPAT_PR,
3398 [LCK_CW] = LCK_COMPAT_CW,
3399 [LCK_CR] = LCK_COMPAT_CR,
3400 [LCK_NL] = LCK_COMPAT_NL,
3401 [LCK_GROUP] = LCK_COMPAT_GROUP,
3402 [LCK_COS] = LCK_COMPAT_COS,
3403 [LCK_TXN] = LCK_COMPAT_TXN,
3407 * Rather arbitrary mapping from LDLM error codes to errno values. This should
3408 * not escape to the user level.
3410 int ldlm_error2errno(enum ldlm_error error)
3416 case ELDLM_LOCK_MATCHED:
3419 case ELDLM_LOCK_CHANGED:
3422 case ELDLM_LOCK_ABORTED:
3425 case ELDLM_LOCK_REPLACED:
3428 case ELDLM_NO_LOCK_DATA:
3431 case ELDLM_NAMESPACE_EXISTS:
3434 case ELDLM_BAD_NAMESPACE:
3438 if (((int)error) < 0) { /* cast to signed type */
3439 result = error; /* as ldlm_error can be unsigned */
3441 CERROR("Invalid DLM result code: %d\n", error);
3447 EXPORT_SYMBOL(ldlm_error2errno);
3450 * Dual to ldlm_error2errno(): maps errno values back to enum ldlm_error.
3452 enum ldlm_error ldlm_errno2error(int err_no)
3461 error = ELDLM_LOCK_CHANGED;
3464 error = ELDLM_LOCK_ABORTED;
3467 error = ELDLM_LOCK_REPLACED;
3470 error = ELDLM_NO_LOCK_DATA;
3473 error = ELDLM_NAMESPACE_EXISTS;
3476 error = ELDLM_BAD_NAMESPACE;
3484 #if LUSTRE_TRACKS_LOCK_EXP_REFS
3485 void ldlm_dump_export_locks(struct obd_export *exp)
3487 spin_lock(&exp->exp_locks_list_guard);
3488 if (!list_empty(&exp->exp_locks_list)) {
3489 struct ldlm_lock *lock;
3491 CERROR("dumping locks for export %p, ignore if the unmount doesn't hang\n",
3493 list_for_each_entry(lock, &exp->exp_locks_list,
3495 LDLM_ERROR(lock, "lock:");
3497 spin_unlock(&exp->exp_locks_list_guard);
3501 #ifdef HAVE_SERVER_SUPPORT
3502 static inline const char *bulk2type(struct ptlrpc_request *req)
3504 if (req->rq_bulk_read)
3506 if (req->rq_bulk_write)
3511 int target_bulk_io(struct obd_export *exp, struct ptlrpc_bulk_desc *desc)
3513 struct ptlrpc_request *req = desc->bd_req;
3514 time64_t start = ktime_get_seconds();
3520 /* Check if client was evicted or reconnected already. */
3521 if (exp->exp_failed ||
3522 exp->exp_conn_cnt > lustre_msg_get_conn_cnt(req->rq_reqmsg)) {
3525 if (req->rq_bulk_read)
3526 rc = sptlrpc_svc_wrap_bulk(req, desc);
3528 if (OCD_HAS_FLAG(&exp->exp_connect_data, BULK_MBITS))
3529 req->rq_mbits = lustre_msg_get_mbits(req->rq_reqmsg);
3530 else /* old version, bulk matchbits is rq_xid */
3531 req->rq_mbits = req->rq_xid;
3534 rc = ptlrpc_start_bulk_transfer(desc);
3538 DEBUG_REQ(D_ERROR, req, "bulk %s failed: rc = %d",
3539 bulk2type(req), rc);
3543 if (CFS_FAIL_CHECK(OBD_FAIL_MDS_SENDPAGE)) {
3544 ptlrpc_abort_bulk(desc);
3548 /* limit actual bulk transfer to bulk_timeout seconds */
3549 deadline = start + bulk_timeout;
3550 if (deadline > req->rq_deadline)
3551 deadline = req->rq_deadline;
3554 time64_t timeoutl = deadline - ktime_get_seconds();
3555 time64_t rq_deadline;
3557 while (timeoutl >= 0 &&
3558 wait_event_idle_timeout(
3560 !ptlrpc_server_bulk_active(desc) ||
3563 lustre_msg_get_conn_cnt(req->rq_reqmsg),
3564 timeoutl ? cfs_time_seconds(1) : 1) == 0)
3566 rc = timeoutl < 0 ? -ETIMEDOUT : 0;
3568 /* Wait again if we changed rq_deadline. */
3569 rq_deadline = READ_ONCE(req->rq_deadline);
3570 deadline = start + bulk_timeout;
3571 if (deadline > rq_deadline)
3572 deadline = rq_deadline;
3573 } while (rc == -ETIMEDOUT &&
3574 deadline > ktime_get_seconds());
3576 if (rc == -ETIMEDOUT) {
3577 DEBUG_REQ(D_ERROR, req, "timeout on bulk %s after %lld%+llds",
3578 bulk2type(req), deadline - start,
3579 ktime_get_real_seconds() - deadline);
3580 ptlrpc_abort_bulk(desc);
3581 } else if (exp->exp_failed) {
3582 DEBUG_REQ(D_ERROR, req, "Eviction on bulk %s",
3585 ptlrpc_abort_bulk(desc);
3586 } else if (exp->exp_conn_cnt >
3587 lustre_msg_get_conn_cnt(req->rq_reqmsg)) {
3588 DEBUG_REQ(D_ERROR, req, "Reconnect on bulk %s",
3590 /* We don't reply anyway. */
3592 ptlrpc_abort_bulk(desc);
3593 } else if (desc->bd_failure) {
3594 DEBUG_REQ(D_ERROR, req, "network error on bulk %s",
3596 /* XXX should this be a different errno? */
3599 if (req->rq_bulk_write)
3600 rc = sptlrpc_svc_unwrap_bulk(req, desc);
3601 if (rc == 0 && desc->bd_nob_transferred != desc->bd_nob) {
3602 DEBUG_REQ(D_ERROR, req, "truncated bulk %s %d(%d)",
3603 bulk2type(req), desc->bd_nob_transferred,
3605 /* XXX should this be a different errno? */
3612 EXPORT_SYMBOL(target_bulk_io);
3614 #endif /* HAVE_SERVER_SUPPORT */