4 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
6 * This program is free software; you can redistribute it and/or modify
7 * it under the terms of the GNU General Public License version 2 only,
8 * as published by the Free Software Foundation.
10 * This program is distributed in the hope that it will be useful, but
11 * WITHOUT ANY WARRANTY; without even the implied warranty of
12 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
13 * General Public License version 2 for more details (a copy is included
14 * in the LICENSE file that accompanied this code).
16 * You should have received a copy of the GNU General Public License
17 * version 2 along with this program; If not, see
18 * http://www.sun.com/software/products/lustre/docs/GPLv2.pdf
20 * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
21 * CA 95054 USA or visit www.sun.com if you need additional information or
27 * Copyright (c) 2003, 2010, Oracle and/or its affiliates. All rights reserved.
28 * Use is subject to license terms.
30 * Copyright (c) 2011, 2015, Intel Corporation.
33 * This file is part of Lustre, http://www.lustre.org/
34 * Lustre is a trademark of Sun Microsystems, Inc.
36 * lustre/ptlrpc/import.c
38 * Author: Mike Shaver <shaver@clusterfs.com>
41 #define DEBUG_SUBSYSTEM S_RPC
43 #include "../include/obd_support.h"
44 #include "../include/lustre_ha.h"
45 #include "../include/lustre_net.h"
46 #include "../include/lustre_import.h"
47 #include "../include/lustre_export.h"
48 #include "../include/obd.h"
49 #include "../include/obd_cksum.h"
50 #include "../include/obd_class.h"
52 #include "ptlrpc_internal.h"
54 struct ptlrpc_connect_async_args
{
55 __u64 pcaa_peer_committed
;
56 int pcaa_initial_connect
;
60 * Updates import \a imp current state to provided \a state value
61 * Helper function. Must be called under imp_lock.
63 static void __import_set_state(struct obd_import
*imp
,
64 enum lustre_imp_state state
)
67 case LUSTRE_IMP_CLOSED
:
69 case LUSTRE_IMP_DISCON
:
70 case LUSTRE_IMP_CONNECTING
:
72 case LUSTRE_IMP_REPLAY_WAIT
:
73 imp
->imp_replay_state
= LUSTRE_IMP_REPLAY_LOCKS
;
76 imp
->imp_replay_state
= LUSTRE_IMP_REPLAY
;
79 imp
->imp_state
= state
;
80 imp
->imp_state_hist
[imp
->imp_state_hist_idx
].ish_state
= state
;
81 imp
->imp_state_hist
[imp
->imp_state_hist_idx
].ish_time
=
82 ktime_get_real_seconds();
83 imp
->imp_state_hist_idx
= (imp
->imp_state_hist_idx
+ 1) %
87 /* A CLOSED import should remain so. */
88 #define IMPORT_SET_STATE_NOLOCK(imp, state) \
90 if (imp->imp_state != LUSTRE_IMP_CLOSED) { \
91 CDEBUG(D_HA, "%p %s: changing import state from %s to %s\n", \
92 imp, obd2cli_tgt(imp->imp_obd), \
93 ptlrpc_import_state_name(imp->imp_state), \
94 ptlrpc_import_state_name(state)); \
95 __import_set_state(imp, state); \
99 #define IMPORT_SET_STATE(imp, state) \
101 spin_lock(&imp->imp_lock); \
102 IMPORT_SET_STATE_NOLOCK(imp, state); \
103 spin_unlock(&imp->imp_lock); \
106 static int ptlrpc_connect_interpret(const struct lu_env
*env
,
107 struct ptlrpc_request
*request
,
109 int ptlrpc_import_recovery_state_machine(struct obd_import
*imp
);
111 /* Only this function is allowed to change the import state when it is
112 * CLOSED. I would rather refcount the import and free it after
113 * disconnection like we do with exports. To do that, the client_obd
114 * will need to save the peer info somewhere other than in the import,
117 int ptlrpc_init_import(struct obd_import
*imp
)
119 spin_lock(&imp
->imp_lock
);
121 imp
->imp_generation
++;
122 imp
->imp_state
= LUSTRE_IMP_NEW
;
124 spin_unlock(&imp
->imp_lock
);
128 EXPORT_SYMBOL(ptlrpc_init_import
);
130 #define UUID_STR "_UUID"
131 static void deuuidify(char *uuid
, const char *prefix
, char **uuid_start
,
134 *uuid_start
= !prefix
|| strncmp(uuid
, prefix
, strlen(prefix
))
135 ? uuid
: uuid
+ strlen(prefix
);
137 *uuid_len
= strlen(*uuid_start
);
139 if (*uuid_len
< strlen(UUID_STR
))
142 if (!strncmp(*uuid_start
+ *uuid_len
- strlen(UUID_STR
),
143 UUID_STR
, strlen(UUID_STR
)))
144 *uuid_len
-= strlen(UUID_STR
);
148 * Returns true if import was FULL, false if import was already not
150 * @imp - import to be disconnected
151 * @conn_cnt - connection count (epoch) of the request that timed out
152 * and caused the disconnection. In some cases, multiple
153 * inflight requests can fail to a single target (e.g. OST
154 * bulk requests) and if one has already caused a reconnection
155 * (increasing the import->conn_cnt) the older failure should
156 * not also cause a reconnection. If zero it forces a reconnect.
158 int ptlrpc_set_import_discon(struct obd_import
*imp
, __u32 conn_cnt
)
162 spin_lock(&imp
->imp_lock
);
164 if (imp
->imp_state
== LUSTRE_IMP_FULL
&&
165 (conn_cnt
== 0 || conn_cnt
== imp
->imp_conn_cnt
)) {
169 deuuidify(obd2cli_tgt(imp
->imp_obd
), NULL
,
170 &target_start
, &target_len
);
172 if (imp
->imp_replayable
) {
173 LCONSOLE_WARN("%s: Connection to %.*s (at %s) was lost; in progress operations using this service will wait for recovery to complete\n",
174 imp
->imp_obd
->obd_name
, target_len
, target_start
,
175 libcfs_nid2str(imp
->imp_connection
->c_peer
.nid
));
177 LCONSOLE_ERROR_MSG(0x166, "%s: Connection to %.*s (at %s) was lost; in progress operations using this service will fail\n",
178 imp
->imp_obd
->obd_name
,
179 target_len
, target_start
,
180 libcfs_nid2str(imp
->imp_connection
->c_peer
.nid
));
182 IMPORT_SET_STATE_NOLOCK(imp
, LUSTRE_IMP_DISCON
);
183 spin_unlock(&imp
->imp_lock
);
185 if (obd_dump_on_timeout
)
186 libcfs_debug_dumplog();
188 obd_import_event(imp
->imp_obd
, imp
, IMP_EVENT_DISCON
);
191 spin_unlock(&imp
->imp_lock
);
192 CDEBUG(D_HA
, "%s: import %p already %s (conn %u, was %u): %s\n",
193 imp
->imp_client
->cli_name
, imp
,
194 (imp
->imp_state
== LUSTRE_IMP_FULL
&&
195 imp
->imp_conn_cnt
> conn_cnt
) ?
196 "reconnected" : "not connected", imp
->imp_conn_cnt
,
197 conn_cnt
, ptlrpc_import_state_name(imp
->imp_state
));
204 * This acts as a barrier; all existing requests are rejected, and
205 * no new requests will be accepted until the import is valid again.
207 void ptlrpc_deactivate_import(struct obd_import
*imp
)
209 CDEBUG(D_HA
, "setting import %s INVALID\n", obd2cli_tgt(imp
->imp_obd
));
211 spin_lock(&imp
->imp_lock
);
212 imp
->imp_invalid
= 1;
213 imp
->imp_generation
++;
214 spin_unlock(&imp
->imp_lock
);
216 ptlrpc_abort_inflight(imp
);
217 obd_import_event(imp
->imp_obd
, imp
, IMP_EVENT_INACTIVE
);
219 EXPORT_SYMBOL(ptlrpc_deactivate_import
);
222 ptlrpc_inflight_deadline(struct ptlrpc_request
*req
, time64_t now
)
226 if (!(((req
->rq_phase
== RQ_PHASE_RPC
) && !req
->rq_waiting
) ||
227 (req
->rq_phase
== RQ_PHASE_BULK
) ||
228 (req
->rq_phase
== RQ_PHASE_NEW
)))
231 if (req
->rq_timedout
)
234 if (req
->rq_phase
== RQ_PHASE_NEW
)
237 dl
= req
->rq_deadline
;
245 static unsigned int ptlrpc_inflight_timeout(struct obd_import
*imp
)
247 time64_t now
= ktime_get_real_seconds();
248 struct list_head
*tmp
, *n
;
249 struct ptlrpc_request
*req
;
250 unsigned int timeout
= 0;
252 spin_lock(&imp
->imp_lock
);
253 list_for_each_safe(tmp
, n
, &imp
->imp_sending_list
) {
254 req
= list_entry(tmp
, struct ptlrpc_request
, rq_list
);
255 timeout
= max(ptlrpc_inflight_deadline(req
, now
), timeout
);
257 spin_unlock(&imp
->imp_lock
);
262 * This function will invalidate the import, if necessary, then block
263 * for all the RPC completions, and finally notify the obd to
264 * invalidate its state (ie cancel locks, clear pending requests,
267 void ptlrpc_invalidate_import(struct obd_import
*imp
)
269 struct list_head
*tmp
, *n
;
270 struct ptlrpc_request
*req
;
271 struct l_wait_info lwi
;
272 unsigned int timeout
;
275 atomic_inc(&imp
->imp_inval_count
);
277 if (!imp
->imp_invalid
|| imp
->imp_obd
->obd_no_recov
)
278 ptlrpc_deactivate_import(imp
);
280 CFS_FAIL_TIMEOUT(OBD_FAIL_MGS_CONNECT_NET
, 3 * cfs_fail_val
/ 2);
281 LASSERT(imp
->imp_invalid
);
283 /* Wait forever until inflight == 0. We really can't do it another
284 * way because in some cases we need to wait for very long reply
285 * unlink. We can't do anything before that because there is really
286 * no guarantee that some rdma transfer is not in progress right now.
289 /* Calculate max timeout for waiting on rpcs to error
290 * out. Use obd_timeout if calculated value is smaller
293 if (!OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_REPL_UNLINK
)) {
294 timeout
= ptlrpc_inflight_timeout(imp
);
295 timeout
+= timeout
/ 3;
298 timeout
= obd_timeout
;
300 /* decrease the interval to increase race condition */
305 "Sleeping %d sec for inflight to error out\n",
308 /* Wait for all requests to error out and call completion
309 * callbacks. Cap it at obd_timeout -- these should all
310 * have been locally cancelled by ptlrpc_abort_inflight.
312 lwi
= LWI_TIMEOUT_INTERVAL(
313 cfs_timeout_cap(cfs_time_seconds(timeout
)),
314 (timeout
> 1)?cfs_time_seconds(1):cfs_time_seconds(1)/2,
316 rc
= l_wait_event(imp
->imp_recovery_waitq
,
317 (atomic_read(&imp
->imp_inflight
) == 0),
320 const char *cli_tgt
= obd2cli_tgt(imp
->imp_obd
);
322 CERROR("%s: rc = %d waiting for callback (%d != 0)\n",
324 atomic_read(&imp
->imp_inflight
));
326 spin_lock(&imp
->imp_lock
);
327 if (atomic_read(&imp
->imp_inflight
) == 0) {
328 int count
= atomic_read(&imp
->imp_unregistering
);
330 /* We know that "unregistering" rpcs only can
331 * survive in sending or delaying lists (they
332 * maybe waiting for long reply unlink in
333 * sluggish nets). Let's check this. If there
334 * is no inflight and unregistering != 0, this
337 LASSERTF(count
== 0, "Some RPCs are still unregistering: %d\n",
340 /* Let's save one loop as soon as inflight have
341 * dropped to zero. No new inflights possible at
346 list_for_each_safe(tmp
, n
,
347 &imp
->imp_sending_list
) {
348 req
= list_entry(tmp
,
349 struct ptlrpc_request
,
351 DEBUG_REQ(D_ERROR
, req
,
352 "still on sending list");
354 list_for_each_safe(tmp
, n
,
355 &imp
->imp_delayed_list
) {
356 req
= list_entry(tmp
,
357 struct ptlrpc_request
,
359 DEBUG_REQ(D_ERROR
, req
,
360 "still on delayed list");
363 CERROR("%s: RPCs in \"%s\" phase found (%d). Network is sluggish? Waiting them to error out.\n",
365 ptlrpc_phase2str(RQ_PHASE_UNREGISTERING
),
369 spin_unlock(&imp
->imp_lock
);
374 * Let's additionally check that no new rpcs added to import in
375 * "invalidate" state.
377 LASSERT(atomic_read(&imp
->imp_inflight
) == 0);
378 obd_import_event(imp
->imp_obd
, imp
, IMP_EVENT_INVALIDATE
);
379 sptlrpc_import_flush_all_ctx(imp
);
381 atomic_dec(&imp
->imp_inval_count
);
382 wake_up_all(&imp
->imp_recovery_waitq
);
384 EXPORT_SYMBOL(ptlrpc_invalidate_import
);
386 /* unset imp_invalid */
387 void ptlrpc_activate_import(struct obd_import
*imp
)
389 struct obd_device
*obd
= imp
->imp_obd
;
391 spin_lock(&imp
->imp_lock
);
392 if (imp
->imp_deactive
!= 0) {
393 spin_unlock(&imp
->imp_lock
);
397 imp
->imp_invalid
= 0;
398 spin_unlock(&imp
->imp_lock
);
399 obd_import_event(obd
, imp
, IMP_EVENT_ACTIVE
);
401 EXPORT_SYMBOL(ptlrpc_activate_import
);
403 static void ptlrpc_pinger_force(struct obd_import
*imp
)
405 CDEBUG(D_HA
, "%s: waking up pinger s:%s\n", obd2cli_tgt(imp
->imp_obd
),
406 ptlrpc_import_state_name(imp
->imp_state
));
408 spin_lock(&imp
->imp_lock
);
409 imp
->imp_force_verify
= 1;
410 spin_unlock(&imp
->imp_lock
);
412 if (imp
->imp_state
!= LUSTRE_IMP_CONNECTING
)
413 ptlrpc_pinger_wake_up();
416 void ptlrpc_fail_import(struct obd_import
*imp
, __u32 conn_cnt
)
418 LASSERT(!imp
->imp_dlm_fake
);
420 if (ptlrpc_set_import_discon(imp
, conn_cnt
)) {
421 if (!imp
->imp_replayable
) {
422 CDEBUG(D_HA
, "import %s@%s for %s not replayable, auto-deactivating\n",
423 obd2cli_tgt(imp
->imp_obd
),
424 imp
->imp_connection
->c_remote_uuid
.uuid
,
425 imp
->imp_obd
->obd_name
);
426 ptlrpc_deactivate_import(imp
);
429 ptlrpc_pinger_force(imp
);
432 EXPORT_SYMBOL(ptlrpc_fail_import
);
434 int ptlrpc_reconnect_import(struct obd_import
*imp
)
436 struct l_wait_info lwi
;
437 int secs
= cfs_time_seconds(obd_timeout
);
440 ptlrpc_pinger_force(imp
);
442 CDEBUG(D_HA
, "%s: recovery started, waiting %u seconds\n",
443 obd2cli_tgt(imp
->imp_obd
), secs
);
445 lwi
= LWI_TIMEOUT(secs
, NULL
, NULL
);
446 rc
= l_wait_event(imp
->imp_recovery_waitq
,
447 !ptlrpc_import_in_recovery(imp
), &lwi
);
448 CDEBUG(D_HA
, "%s: recovery finished s:%s\n", obd2cli_tgt(imp
->imp_obd
),
449 ptlrpc_import_state_name(imp
->imp_state
));
452 EXPORT_SYMBOL(ptlrpc_reconnect_import
);
455 * Connection on import \a imp is changed to another one (if more than one is
456 * present). We typically chose connection that we have not tried to connect to
459 static int import_select_connection(struct obd_import
*imp
)
461 struct obd_import_conn
*imp_conn
= NULL
, *conn
;
462 struct obd_export
*dlmexp
;
464 int target_len
, tried_all
= 1;
466 spin_lock(&imp
->imp_lock
);
468 if (list_empty(&imp
->imp_conn_list
)) {
469 CERROR("%s: no connections available\n",
470 imp
->imp_obd
->obd_name
);
471 spin_unlock(&imp
->imp_lock
);
475 list_for_each_entry(conn
, &imp
->imp_conn_list
, oic_item
) {
476 CDEBUG(D_HA
, "%s: connect to NID %s last attempt %llu\n",
477 imp
->imp_obd
->obd_name
,
478 libcfs_nid2str(conn
->oic_conn
->c_peer
.nid
),
479 conn
->oic_last_attempt
);
481 /* If we have not tried this connection since
482 * the last successful attempt, go with this one
484 if ((conn
->oic_last_attempt
== 0) ||
485 cfs_time_beforeq_64(conn
->oic_last_attempt
,
486 imp
->imp_last_success_conn
)) {
492 /* If all of the connections have already been tried
493 * since the last successful connection; just choose the
494 * least recently used
498 else if (cfs_time_before_64(conn
->oic_last_attempt
,
499 imp_conn
->oic_last_attempt
))
503 /* if not found, simply choose the current one */
504 if (!imp_conn
|| imp
->imp_force_reconnect
) {
505 LASSERT(imp
->imp_conn_current
);
506 imp_conn
= imp
->imp_conn_current
;
509 LASSERT(imp_conn
->oic_conn
);
511 /* If we've tried everything, and we're back to the beginning of the
512 * list, increase our timeout and try again. It will be reset when
513 * we do finally connect. (FIXME: really we should wait for all network
514 * state associated with the last connection attempt to drain before
515 * trying to reconnect on it.)
517 if (tried_all
&& (imp
->imp_conn_list
.next
== &imp_conn
->oic_item
)) {
518 struct adaptive_timeout
*at
= &imp
->imp_at
.iat_net_latency
;
520 if (at_get(at
) < CONNECTION_SWITCH_MAX
) {
521 at_measured(at
, at_get(at
) + CONNECTION_SWITCH_INC
);
522 if (at_get(at
) > CONNECTION_SWITCH_MAX
)
523 at_reset(at
, CONNECTION_SWITCH_MAX
);
525 LASSERT(imp_conn
->oic_last_attempt
);
526 CDEBUG(D_HA
, "%s: tried all connections, increasing latency to %ds\n",
527 imp
->imp_obd
->obd_name
, at_get(at
));
530 imp_conn
->oic_last_attempt
= cfs_time_current_64();
532 /* switch connection, don't mind if it's same as the current one */
533 ptlrpc_connection_put(imp
->imp_connection
);
534 imp
->imp_connection
= ptlrpc_connection_addref(imp_conn
->oic_conn
);
536 dlmexp
= class_conn2export(&imp
->imp_dlm_handle
);
537 ptlrpc_connection_put(dlmexp
->exp_connection
);
538 dlmexp
->exp_connection
= ptlrpc_connection_addref(imp_conn
->oic_conn
);
539 class_export_put(dlmexp
);
541 if (imp
->imp_conn_current
!= imp_conn
) {
542 if (imp
->imp_conn_current
) {
543 deuuidify(obd2cli_tgt(imp
->imp_obd
), NULL
,
544 &target_start
, &target_len
);
546 CDEBUG(D_HA
, "%s: Connection changing to %.*s (at %s)\n",
547 imp
->imp_obd
->obd_name
,
548 target_len
, target_start
,
549 libcfs_nid2str(imp_conn
->oic_conn
->c_peer
.nid
));
552 imp
->imp_conn_current
= imp_conn
;
555 CDEBUG(D_HA
, "%s: import %p using connection %s/%s\n",
556 imp
->imp_obd
->obd_name
, imp
, imp_conn
->oic_uuid
.uuid
,
557 libcfs_nid2str(imp_conn
->oic_conn
->c_peer
.nid
));
559 spin_unlock(&imp
->imp_lock
);
565 * must be called under imp_lock
567 static int ptlrpc_first_transno(struct obd_import
*imp
, __u64
*transno
)
569 struct ptlrpc_request
*req
;
570 struct list_head
*tmp
;
572 /* The requests in committed_list always have smaller transnos than
573 * the requests in replay_list
575 if (!list_empty(&imp
->imp_committed_list
)) {
576 tmp
= imp
->imp_committed_list
.next
;
577 req
= list_entry(tmp
, struct ptlrpc_request
, rq_replay_list
);
578 *transno
= req
->rq_transno
;
579 if (req
->rq_transno
== 0) {
580 DEBUG_REQ(D_ERROR
, req
,
581 "zero transno in committed_list");
586 if (!list_empty(&imp
->imp_replay_list
)) {
587 tmp
= imp
->imp_replay_list
.next
;
588 req
= list_entry(tmp
, struct ptlrpc_request
, rq_replay_list
);
589 *transno
= req
->rq_transno
;
590 if (req
->rq_transno
== 0) {
591 DEBUG_REQ(D_ERROR
, req
, "zero transno in replay_list");
600 * Attempt to (re)connect import \a imp. This includes all preparations,
601 * initializing CONNECT RPC request and passing it to ptlrpcd for
603 * Returns 0 on success or error code.
605 int ptlrpc_connect_import(struct obd_import
*imp
)
607 struct obd_device
*obd
= imp
->imp_obd
;
608 int initial_connect
= 0;
610 __u64 committed_before_reconnect
= 0;
611 struct ptlrpc_request
*request
;
612 char *bufs
[] = { NULL
,
613 obd2cli_tgt(imp
->imp_obd
),
615 (char *)&imp
->imp_dlm_handle
,
616 (char *)&imp
->imp_connect_data
};
617 struct ptlrpc_connect_async_args
*aa
;
620 spin_lock(&imp
->imp_lock
);
621 if (imp
->imp_state
== LUSTRE_IMP_CLOSED
) {
622 spin_unlock(&imp
->imp_lock
);
623 CERROR("can't connect to a closed import\n");
625 } else if (imp
->imp_state
== LUSTRE_IMP_FULL
) {
626 spin_unlock(&imp
->imp_lock
);
627 CERROR("already connected\n");
629 } else if (imp
->imp_state
== LUSTRE_IMP_CONNECTING
) {
630 spin_unlock(&imp
->imp_lock
);
631 CERROR("already connecting\n");
635 IMPORT_SET_STATE_NOLOCK(imp
, LUSTRE_IMP_CONNECTING
);
638 imp
->imp_resend_replay
= 0;
640 if (!lustre_handle_is_used(&imp
->imp_remote_handle
))
643 committed_before_reconnect
= imp
->imp_peer_committed_transno
;
645 set_transno
= ptlrpc_first_transno(imp
,
646 &imp
->imp_connect_data
.ocd_transno
);
647 spin_unlock(&imp
->imp_lock
);
649 rc
= import_select_connection(imp
);
653 rc
= sptlrpc_import_sec_adapt(imp
, NULL
, NULL
);
657 /* Reset connect flags to the originally requested flags, in case
658 * the server is updated on-the-fly we will get the new features.
660 imp
->imp_connect_data
.ocd_connect_flags
= imp
->imp_connect_flags_orig
;
661 /* Reset ocd_version each time so the server knows the exact versions */
662 imp
->imp_connect_data
.ocd_version
= LUSTRE_VERSION_CODE
;
663 imp
->imp_msghdr_flags
&= ~MSGHDR_AT_SUPPORT
;
664 imp
->imp_msghdr_flags
&= ~MSGHDR_CKSUM_INCOMPAT18
;
666 rc
= obd_reconnect(NULL
, imp
->imp_obd
->obd_self_export
, obd
,
667 &obd
->obd_uuid
, &imp
->imp_connect_data
, NULL
);
671 request
= ptlrpc_request_alloc(imp
, &RQF_MDS_CONNECT
);
677 rc
= ptlrpc_request_bufs_pack(request
, LUSTRE_OBD_VERSION
,
678 imp
->imp_connect_op
, bufs
, NULL
);
680 ptlrpc_request_free(request
);
684 /* Report the rpc service time to the server so that it knows how long
685 * to wait for clients to join recovery
687 lustre_msg_set_service_time(request
->rq_reqmsg
,
688 at_timeout2est(request
->rq_timeout
));
690 /* The amount of time we give the server to process the connect req.
691 * import_select_connection will increase the net latency on
692 * repeated reconnect attempts to cover slow networks.
693 * We override/ignore the server rpc completion estimate here,
694 * which may be large if this is a reconnect attempt
696 request
->rq_timeout
= INITIAL_CONNECT_TIMEOUT
;
697 lustre_msg_set_timeout(request
->rq_reqmsg
, request
->rq_timeout
);
699 lustre_msg_add_op_flags(request
->rq_reqmsg
, MSG_CONNECT_NEXT_VER
);
701 request
->rq_no_resend
= request
->rq_no_delay
= 1;
702 request
->rq_send_state
= LUSTRE_IMP_CONNECTING
;
703 /* Allow a slightly larger reply for future growth compatibility */
704 req_capsule_set_size(&request
->rq_pill
, &RMF_CONNECT_DATA
, RCL_SERVER
,
705 sizeof(struct obd_connect_data
)+16*sizeof(__u64
));
706 ptlrpc_request_set_replen(request
);
707 request
->rq_interpret_reply
= ptlrpc_connect_interpret
;
709 CLASSERT(sizeof(*aa
) <= sizeof(request
->rq_async_args
));
710 aa
= ptlrpc_req_async_args(request
);
711 memset(aa
, 0, sizeof(*aa
));
713 aa
->pcaa_peer_committed
= committed_before_reconnect
;
714 aa
->pcaa_initial_connect
= initial_connect
;
716 if (aa
->pcaa_initial_connect
) {
717 spin_lock(&imp
->imp_lock
);
718 imp
->imp_replayable
= 1;
719 spin_unlock(&imp
->imp_lock
);
720 lustre_msg_add_op_flags(request
->rq_reqmsg
,
721 MSG_CONNECT_INITIAL
);
725 lustre_msg_add_op_flags(request
->rq_reqmsg
,
726 MSG_CONNECT_TRANSNO
);
728 DEBUG_REQ(D_RPCTRACE
, request
, "(re)connect request (timeout %d)",
729 request
->rq_timeout
);
730 ptlrpcd_add_req(request
);
734 IMPORT_SET_STATE(imp
, LUSTRE_IMP_DISCON
);
738 EXPORT_SYMBOL(ptlrpc_connect_import
);
740 static void ptlrpc_maybe_ping_import_soon(struct obd_import
*imp
)
744 spin_lock(&imp
->imp_lock
);
745 force_verify
= imp
->imp_force_verify
!= 0;
746 spin_unlock(&imp
->imp_lock
);
749 ptlrpc_pinger_wake_up();
752 static int ptlrpc_busy_reconnect(int rc
)
754 return (rc
== -EBUSY
) || (rc
== -EAGAIN
);
758 * interpret_reply callback for connect RPCs.
759 * Looks into returned status of connect operation and decides
760 * what to do with the import - i.e enter recovery, promote it to
761 * full state for normal operations of disconnect it due to an error.
763 static int ptlrpc_connect_interpret(const struct lu_env
*env
,
764 struct ptlrpc_request
*request
,
767 struct ptlrpc_connect_async_args
*aa
= data
;
768 struct obd_import
*imp
= request
->rq_import
;
769 struct client_obd
*cli
= &imp
->imp_obd
->u
.cli
;
770 struct lustre_handle old_hdl
;
771 __u64 old_connect_flags
;
773 struct obd_connect_data
*ocd
;
774 struct obd_export
*exp
;
777 spin_lock(&imp
->imp_lock
);
778 if (imp
->imp_state
== LUSTRE_IMP_CLOSED
) {
779 imp
->imp_connect_tried
= 1;
780 spin_unlock(&imp
->imp_lock
);
785 /* if this reconnect to busy export - not need select new target
788 imp
->imp_force_reconnect
= ptlrpc_busy_reconnect(rc
);
789 spin_unlock(&imp
->imp_lock
);
790 ptlrpc_maybe_ping_import_soon(imp
);
793 spin_unlock(&imp
->imp_lock
);
795 LASSERT(imp
->imp_conn_current
);
797 msg_flags
= lustre_msg_get_op_flags(request
->rq_repmsg
);
799 ret
= req_capsule_get_size(&request
->rq_pill
, &RMF_CONNECT_DATA
,
801 /* server replied obd_connect_data is always bigger */
802 ocd
= req_capsule_server_sized_get(&request
->rq_pill
,
803 &RMF_CONNECT_DATA
, ret
);
806 CERROR("%s: no connect data from server\n",
807 imp
->imp_obd
->obd_name
);
812 spin_lock(&imp
->imp_lock
);
814 /* All imports are pingable */
815 imp
->imp_pingable
= 1;
816 imp
->imp_force_reconnect
= 0;
817 imp
->imp_force_verify
= 0;
819 imp
->imp_connect_data
= *ocd
;
821 CDEBUG(D_HA
, "%s: connect to target with instance %u\n",
822 imp
->imp_obd
->obd_name
, ocd
->ocd_instance
);
823 exp
= class_conn2export(&imp
->imp_dlm_handle
);
825 spin_unlock(&imp
->imp_lock
);
827 /* check that server granted subset of flags we asked for. */
828 if ((ocd
->ocd_connect_flags
& imp
->imp_connect_flags_orig
) !=
829 ocd
->ocd_connect_flags
) {
830 CERROR("%s: Server didn't granted asked subset of flags: asked=%#llx grranted=%#llx\n",
831 imp
->imp_obd
->obd_name
, imp
->imp_connect_flags_orig
,
832 ocd
->ocd_connect_flags
);
838 /* This could happen if export is cleaned during the
841 CERROR("%s: missing export after connect\n",
842 imp
->imp_obd
->obd_name
);
846 old_connect_flags
= exp_connect_flags(exp
);
847 exp
->exp_connect_data
= *ocd
;
848 imp
->imp_obd
->obd_self_export
->exp_connect_data
= *ocd
;
849 class_export_put(exp
);
851 obd_import_event(imp
->imp_obd
, imp
, IMP_EVENT_OCD
);
853 if (aa
->pcaa_initial_connect
) {
854 spin_lock(&imp
->imp_lock
);
855 if (msg_flags
& MSG_CONNECT_REPLAYABLE
) {
856 imp
->imp_replayable
= 1;
857 spin_unlock(&imp
->imp_lock
);
858 CDEBUG(D_HA
, "connected to replayable target: %s\n",
859 obd2cli_tgt(imp
->imp_obd
));
861 imp
->imp_replayable
= 0;
862 spin_unlock(&imp
->imp_lock
);
865 /* if applies, adjust the imp->imp_msg_magic here
866 * according to reply flags
869 imp
->imp_remote_handle
=
870 *lustre_msg_get_handle(request
->rq_repmsg
);
872 /* Initial connects are allowed for clients with non-random
873 * uuids when servers are in recovery. Simply signal the
874 * servers replay is complete and wait in REPLAY_WAIT.
876 if (msg_flags
& MSG_CONNECT_RECOVERING
) {
877 CDEBUG(D_HA
, "connect to %s during recovery\n",
878 obd2cli_tgt(imp
->imp_obd
));
879 IMPORT_SET_STATE(imp
, LUSTRE_IMP_REPLAY_LOCKS
);
881 IMPORT_SET_STATE(imp
, LUSTRE_IMP_FULL
);
882 ptlrpc_activate_import(imp
);
889 /* Determine what recovery state to move the import to. */
890 if (msg_flags
& MSG_CONNECT_RECONNECT
) {
891 memset(&old_hdl
, 0, sizeof(old_hdl
));
892 if (!memcmp(&old_hdl
, lustre_msg_get_handle(request
->rq_repmsg
),
894 LCONSOLE_WARN("Reconnect to %s (at @%s) failed due bad handle %#llx\n",
895 obd2cli_tgt(imp
->imp_obd
),
896 imp
->imp_connection
->c_remote_uuid
.uuid
,
897 imp
->imp_dlm_handle
.cookie
);
902 if (memcmp(&imp
->imp_remote_handle
,
903 lustre_msg_get_handle(request
->rq_repmsg
),
904 sizeof(imp
->imp_remote_handle
))) {
905 int level
= msg_flags
& MSG_CONNECT_RECOVERING
?
908 /* Bug 16611/14775: if server handle have changed,
909 * that means some sort of disconnection happened.
910 * If the server is not in recovery, that also means it
911 * already erased all of our state because of previous
912 * eviction. If it is in recovery - we are safe to
913 * participate since we can reestablish all of our state
916 if ((msg_flags
& MSG_CONNECT_RECOVERING
)) {
917 CDEBUG(level
, "%s@%s changed server handle from %#llx to %#llx but is still in recovery\n",
918 obd2cli_tgt(imp
->imp_obd
),
919 imp
->imp_connection
->c_remote_uuid
.uuid
,
920 imp
->imp_remote_handle
.cookie
,
921 lustre_msg_get_handle(
922 request
->rq_repmsg
)->cookie
);
924 LCONSOLE_WARN("Evicted from %s (at %s) after server handle changed from %#llx to %#llx\n",
925 obd2cli_tgt(imp
->imp_obd
),
926 imp
->imp_connection
-> \
928 imp
->imp_remote_handle
.cookie
,
929 lustre_msg_get_handle(
930 request
->rq_repmsg
)->cookie
);
933 imp
->imp_remote_handle
=
934 *lustre_msg_get_handle(request
->rq_repmsg
);
936 if (!(msg_flags
& MSG_CONNECT_RECOVERING
)) {
937 IMPORT_SET_STATE(imp
, LUSTRE_IMP_EVICTED
);
943 CDEBUG(D_HA
, "reconnected to %s@%s after partition\n",
944 obd2cli_tgt(imp
->imp_obd
),
945 imp
->imp_connection
->c_remote_uuid
.uuid
);
948 if (imp
->imp_invalid
) {
949 CDEBUG(D_HA
, "%s: reconnected but import is invalid; marking evicted\n",
950 imp
->imp_obd
->obd_name
);
951 IMPORT_SET_STATE(imp
, LUSTRE_IMP_EVICTED
);
952 } else if (msg_flags
& MSG_CONNECT_RECOVERING
) {
953 CDEBUG(D_HA
, "%s: reconnected to %s during replay\n",
954 imp
->imp_obd
->obd_name
,
955 obd2cli_tgt(imp
->imp_obd
));
957 spin_lock(&imp
->imp_lock
);
958 imp
->imp_resend_replay
= 1;
959 spin_unlock(&imp
->imp_lock
);
961 IMPORT_SET_STATE(imp
, imp
->imp_replay_state
);
963 IMPORT_SET_STATE(imp
, LUSTRE_IMP_RECOVER
);
965 } else if ((msg_flags
& MSG_CONNECT_RECOVERING
) && !imp
->imp_invalid
) {
966 LASSERT(imp
->imp_replayable
);
967 imp
->imp_remote_handle
=
968 *lustre_msg_get_handle(request
->rq_repmsg
);
969 imp
->imp_last_replay_transno
= 0;
970 IMPORT_SET_STATE(imp
, LUSTRE_IMP_REPLAY
);
972 DEBUG_REQ(D_HA
, request
, "%s: evicting (reconnect/recover flags not set: %x)",
973 imp
->imp_obd
->obd_name
, msg_flags
);
974 imp
->imp_remote_handle
=
975 *lustre_msg_get_handle(request
->rq_repmsg
);
976 IMPORT_SET_STATE(imp
, LUSTRE_IMP_EVICTED
);
979 /* Sanity checks for a reconnected import. */
980 if (!(imp
->imp_replayable
) != !(msg_flags
& MSG_CONNECT_REPLAYABLE
)) {
981 CERROR("imp_replayable flag does not match server after reconnect. We should LBUG right here.\n");
984 if (lustre_msg_get_last_committed(request
->rq_repmsg
) > 0 &&
985 lustre_msg_get_last_committed(request
->rq_repmsg
) <
986 aa
->pcaa_peer_committed
) {
987 CERROR("%s went back in time (transno %lld was previously committed, server now claims %lld)! See https://bugzilla.lustre.org/show_bug.cgi?id=9646\n",
988 obd2cli_tgt(imp
->imp_obd
), aa
->pcaa_peer_committed
,
989 lustre_msg_get_last_committed(request
->rq_repmsg
));
993 rc
= ptlrpc_import_recovery_state_machine(imp
);
995 if (rc
== -ENOTCONN
) {
996 CDEBUG(D_HA
, "evicted/aborted by %s@%s during recovery; invalidating and reconnecting\n",
997 obd2cli_tgt(imp
->imp_obd
),
998 imp
->imp_connection
->c_remote_uuid
.uuid
);
999 ptlrpc_connect_import(imp
);
1000 imp
->imp_connect_tried
= 1;
1005 spin_lock(&imp
->imp_lock
);
1006 list_del(&imp
->imp_conn_current
->oic_item
);
1007 list_add(&imp
->imp_conn_current
->oic_item
, &imp
->imp_conn_list
);
1008 imp
->imp_last_success_conn
=
1009 imp
->imp_conn_current
->oic_last_attempt
;
1011 spin_unlock(&imp
->imp_lock
);
1013 if ((imp
->imp_connect_flags_orig
& OBD_CONNECT_IBITS
) &&
1014 !(ocd
->ocd_connect_flags
& OBD_CONNECT_IBITS
)) {
1015 LCONSOLE_WARN("%s: MDS %s does not support ibits lock, either very old or invalid: requested %llx, replied %llx\n",
1016 imp
->imp_obd
->obd_name
,
1017 imp
->imp_connection
->c_remote_uuid
.uuid
,
1018 imp
->imp_connect_flags_orig
,
1019 ocd
->ocd_connect_flags
);
1024 if ((ocd
->ocd_connect_flags
& OBD_CONNECT_VERSION
) &&
1025 (ocd
->ocd_version
> LUSTRE_VERSION_CODE
+
1026 LUSTRE_VERSION_OFFSET_WARN
||
1027 ocd
->ocd_version
< LUSTRE_VERSION_CODE
-
1028 LUSTRE_VERSION_OFFSET_WARN
)) {
1029 /* Sigh, some compilers do not like #ifdef in the middle
1030 * of macro arguments
1032 const char *older
= "older. Consider upgrading server or downgrading client"
1034 const char *newer
= "newer than client version. Consider upgrading client"
1037 LCONSOLE_WARN("Server %s version (%d.%d.%d.%d) is much %s (%s)\n",
1038 obd2cli_tgt(imp
->imp_obd
),
1039 OBD_OCD_VERSION_MAJOR(ocd
->ocd_version
),
1040 OBD_OCD_VERSION_MINOR(ocd
->ocd_version
),
1041 OBD_OCD_VERSION_PATCH(ocd
->ocd_version
),
1042 OBD_OCD_VERSION_FIX(ocd
->ocd_version
),
1043 ocd
->ocd_version
> LUSTRE_VERSION_CODE
?
1044 newer
: older
, LUSTRE_VERSION_STRING
);
1047 #if LUSTRE_VERSION_CODE < OBD_OCD_VERSION(3, 2, 50, 0)
1048 /* Check if server has LU-1252 fix applied to not always swab
1049 * the IR MNE entries. Do this only once per connection. This
1050 * fixup is version-limited, because we don't want to carry the
1051 * OBD_CONNECT_MNE_SWAB flag around forever, just so long as we
1052 * need interop with unpatched 2.2 servers. For newer servers,
1053 * the client will do MNE swabbing only as needed. LU-1644
1055 if (unlikely((ocd
->ocd_connect_flags
& OBD_CONNECT_VERSION
) &&
1056 !(ocd
->ocd_connect_flags
& OBD_CONNECT_MNE_SWAB
) &&
1057 OBD_OCD_VERSION_MAJOR(ocd
->ocd_version
) == 2 &&
1058 OBD_OCD_VERSION_MINOR(ocd
->ocd_version
) == 2 &&
1059 OBD_OCD_VERSION_PATCH(ocd
->ocd_version
) < 55 &&
1060 strcmp(imp
->imp_obd
->obd_type
->typ_name
,
1061 LUSTRE_MGC_NAME
) == 0))
1062 imp
->imp_need_mne_swab
= 1;
1063 else /* clear if server was upgraded since last connect */
1064 imp
->imp_need_mne_swab
= 0;
1066 #warning "LU-1644: Remove old OBD_CONNECT_MNE_SWAB fixup and imp_need_mne_swab"
1069 if (ocd
->ocd_connect_flags
& OBD_CONNECT_CKSUM
) {
1070 /* We sent to the server ocd_cksum_types with bits set
1071 * for algorithms we understand. The server masked off
1072 * the checksum types it doesn't support
1074 if ((ocd
->ocd_cksum_types
&
1075 cksum_types_supported_client()) == 0) {
1076 LCONSOLE_WARN("The negotiation of the checksum algorithm to use with server %s failed (%x/%x), disabling checksums\n",
1077 obd2cli_tgt(imp
->imp_obd
),
1078 ocd
->ocd_cksum_types
,
1079 cksum_types_supported_client());
1080 cli
->cl_checksum
= 0;
1081 cli
->cl_supp_cksum_types
= OBD_CKSUM_ADLER
;
1083 cli
->cl_supp_cksum_types
= ocd
->ocd_cksum_types
;
1086 /* The server does not support OBD_CONNECT_CKSUM.
1087 * Enforce ADLER for backward compatibility
1089 cli
->cl_supp_cksum_types
= OBD_CKSUM_ADLER
;
1091 cli
->cl_cksum_type
= cksum_type_select(cli
->cl_supp_cksum_types
);
1093 if (ocd
->ocd_connect_flags
& OBD_CONNECT_BRW_SIZE
)
1094 cli
->cl_max_pages_per_rpc
=
1095 min(ocd
->ocd_brw_size
>> PAGE_CACHE_SHIFT
,
1096 cli
->cl_max_pages_per_rpc
);
1097 else if (imp
->imp_connect_op
== MDS_CONNECT
||
1098 imp
->imp_connect_op
== MGS_CONNECT
)
1099 cli
->cl_max_pages_per_rpc
= 1;
1101 /* Reset ns_connect_flags only for initial connect. It might be
1102 * changed in while using FS and if we reset it in reconnect
1103 * this leads to losing user settings done before such as
1104 * disable lru_resize, etc.
1106 if (old_connect_flags
!= exp_connect_flags(exp
) ||
1107 aa
->pcaa_initial_connect
) {
1108 CDEBUG(D_HA
, "%s: Resetting ns_connect_flags to server flags: %#llx\n",
1109 imp
->imp_obd
->obd_name
, ocd
->ocd_connect_flags
);
1110 imp
->imp_obd
->obd_namespace
->ns_connect_flags
=
1111 ocd
->ocd_connect_flags
;
1112 imp
->imp_obd
->obd_namespace
->ns_orig_connect_flags
=
1113 ocd
->ocd_connect_flags
;
1116 if ((ocd
->ocd_connect_flags
& OBD_CONNECT_AT
) &&
1117 (imp
->imp_msg_magic
== LUSTRE_MSG_MAGIC_V2
))
1118 /* We need a per-message support flag, because
1119 * a. we don't know if the incoming connect reply
1120 * supports AT or not (in reply_in_callback)
1121 * until we unpack it.
1122 * b. failovered server means export and flags are gone
1123 * (in ptlrpc_send_reply).
1124 * Can only be set when we know AT is supported at
1127 imp
->imp_msghdr_flags
|= MSGHDR_AT_SUPPORT
;
1129 imp
->imp_msghdr_flags
&= ~MSGHDR_AT_SUPPORT
;
1131 if ((ocd
->ocd_connect_flags
& OBD_CONNECT_FULL20
) &&
1132 (imp
->imp_msg_magic
== LUSTRE_MSG_MAGIC_V2
))
1133 imp
->imp_msghdr_flags
|= MSGHDR_CKSUM_INCOMPAT18
;
1135 imp
->imp_msghdr_flags
&= ~MSGHDR_CKSUM_INCOMPAT18
;
1137 LASSERT((cli
->cl_max_pages_per_rpc
<= PTLRPC_MAX_BRW_PAGES
) &&
1138 (cli
->cl_max_pages_per_rpc
> 0));
1142 imp
->imp_connect_tried
= 1;
1145 IMPORT_SET_STATE(imp
, LUSTRE_IMP_DISCON
);
1146 if (rc
== -EACCES
) {
1148 * Give up trying to reconnect
1149 * EACCES means client has no permission for connection
1151 imp
->imp_obd
->obd_no_recov
= 1;
1152 ptlrpc_deactivate_import(imp
);
1155 if (rc
== -EPROTO
) {
1156 struct obd_connect_data
*ocd
;
1158 /* reply message might not be ready */
1159 if (!request
->rq_repmsg
)
1162 ocd
= req_capsule_server_get(&request
->rq_pill
,
1165 (ocd
->ocd_connect_flags
& OBD_CONNECT_VERSION
) &&
1166 (ocd
->ocd_version
!= LUSTRE_VERSION_CODE
)) {
1168 * Actually servers are only supposed to refuse
1169 * connection from liblustre clients, so we
1170 * should never see this from VFS context
1172 LCONSOLE_ERROR_MSG(0x16a, "Server %s version (%d.%d.%d.%d) refused connection from this client with an incompatible version (%s). Client must be recompiled\n",
1173 obd2cli_tgt(imp
->imp_obd
),
1174 OBD_OCD_VERSION_MAJOR(ocd
->ocd_version
),
1175 OBD_OCD_VERSION_MINOR(ocd
->ocd_version
),
1176 OBD_OCD_VERSION_PATCH(ocd
->ocd_version
),
1177 OBD_OCD_VERSION_FIX(ocd
->ocd_version
),
1178 LUSTRE_VERSION_STRING
);
1179 ptlrpc_deactivate_import(imp
);
1180 IMPORT_SET_STATE(imp
, LUSTRE_IMP_CLOSED
);
1185 ptlrpc_maybe_ping_import_soon(imp
);
1187 CDEBUG(D_HA
, "recovery of %s on %s failed (%d)\n",
1188 obd2cli_tgt(imp
->imp_obd
),
1189 (char *)imp
->imp_connection
->c_remote_uuid
.uuid
, rc
);
1192 wake_up_all(&imp
->imp_recovery_waitq
);
1197 * interpret callback for "completed replay" RPCs.
1198 * \see signal_completed_replay
1200 static int completed_replay_interpret(const struct lu_env
*env
,
1201 struct ptlrpc_request
*req
,
1204 atomic_dec(&req
->rq_import
->imp_replay_inflight
);
1205 if (req
->rq_status
== 0 &&
1206 !req
->rq_import
->imp_vbr_failed
) {
1207 ptlrpc_import_recovery_state_machine(req
->rq_import
);
1209 if (req
->rq_import
->imp_vbr_failed
) {
1211 "%s: version recovery fails, reconnecting\n",
1212 req
->rq_import
->imp_obd
->obd_name
);
1214 CDEBUG(D_HA
, "%s: LAST_REPLAY message error: %d, reconnecting\n",
1215 req
->rq_import
->imp_obd
->obd_name
,
1218 ptlrpc_connect_import(req
->rq_import
);
1225 * Let server know that we have no requests to replay anymore.
1226 * Achieved by just sending a PING request
1228 static int signal_completed_replay(struct obd_import
*imp
)
1230 struct ptlrpc_request
*req
;
1232 if (unlikely(OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_FINISH_REPLAY
)))
1235 LASSERT(atomic_read(&imp
->imp_replay_inflight
) == 0);
1236 atomic_inc(&imp
->imp_replay_inflight
);
1238 req
= ptlrpc_request_alloc_pack(imp
, &RQF_OBD_PING
, LUSTRE_OBD_VERSION
,
1241 atomic_dec(&imp
->imp_replay_inflight
);
1245 ptlrpc_request_set_replen(req
);
1246 req
->rq_send_state
= LUSTRE_IMP_REPLAY_WAIT
;
1247 lustre_msg_add_flags(req
->rq_reqmsg
,
1248 MSG_LOCK_REPLAY_DONE
| MSG_REQ_REPLAY_DONE
);
1250 req
->rq_timeout
*= 3;
1251 req
->rq_interpret_reply
= completed_replay_interpret
;
1253 ptlrpcd_add_req(req
);
1258 * In kernel code all import invalidation happens in its own
1259 * separate thread, so that whatever application happened to encounter
1260 * a problem could still be killed or otherwise continue
1262 static int ptlrpc_invalidate_import_thread(void *data
)
1264 struct obd_import
*imp
= data
;
1266 unshare_fs_struct();
1268 CDEBUG(D_HA
, "thread invalidate import %s to %s@%s\n",
1269 imp
->imp_obd
->obd_name
, obd2cli_tgt(imp
->imp_obd
),
1270 imp
->imp_connection
->c_remote_uuid
.uuid
);
1272 ptlrpc_invalidate_import(imp
);
1274 if (obd_dump_on_eviction
) {
1275 CERROR("dump the log upon eviction\n");
1276 libcfs_debug_dumplog();
1279 IMPORT_SET_STATE(imp
, LUSTRE_IMP_RECOVER
);
1280 ptlrpc_import_recovery_state_machine(imp
);
1282 class_import_put(imp
);
1287 * This is the state machine for client-side recovery on import.
1289 * Typically we have two possibly paths. If we came to server and it is not
1290 * in recovery, we just enter IMP_EVICTED state, invalidate our import
1291 * state and reconnect from scratch.
1292 * If we came to server that is in recovery, we enter IMP_REPLAY import state.
1293 * We go through our list of requests to replay and send them to server one by
1295 * After sending all request from the list we change import state to
1296 * IMP_REPLAY_LOCKS and re-request all the locks we believe we have from server
1297 * and also all the locks we don't yet have and wait for server to grant us.
1298 * After that we send a special "replay completed" request and change import
1299 * state to IMP_REPLAY_WAIT.
1300 * Upon receiving reply to that "replay completed" RPC we enter IMP_RECOVER
1301 * state and resend all requests from sending list.
1302 * After that we promote import to FULL state and send all delayed requests
1303 * and import is fully operational after that.
1306 int ptlrpc_import_recovery_state_machine(struct obd_import
*imp
)
1313 if (imp
->imp_state
== LUSTRE_IMP_EVICTED
) {
1314 deuuidify(obd2cli_tgt(imp
->imp_obd
), NULL
,
1315 &target_start
, &target_len
);
1316 /* Don't care about MGC eviction */
1317 if (strcmp(imp
->imp_obd
->obd_type
->typ_name
,
1318 LUSTRE_MGC_NAME
) != 0) {
1319 LCONSOLE_ERROR_MSG(0x167, "%s: This client was evicted by %.*s; in progress operations using this service will fail.\n",
1320 imp
->imp_obd
->obd_name
, target_len
,
1323 CDEBUG(D_HA
, "evicted from %s@%s; invalidating\n",
1324 obd2cli_tgt(imp
->imp_obd
),
1325 imp
->imp_connection
->c_remote_uuid
.uuid
);
1326 /* reset vbr_failed flag upon eviction */
1327 spin_lock(&imp
->imp_lock
);
1328 imp
->imp_vbr_failed
= 0;
1329 spin_unlock(&imp
->imp_lock
);
1332 struct task_struct
*task
;
1333 /* bug 17802: XXX client_disconnect_export vs connect request
1334 * race. if client is evicted at this time, we start
1335 * invalidate thread without reference to import and import can
1336 * be freed at same time.
1338 class_import_get(imp
);
1339 task
= kthread_run(ptlrpc_invalidate_import_thread
, imp
,
1342 class_import_put(imp
);
1343 CERROR("error starting invalidate thread: %d\n", rc
);
1352 if (imp
->imp_state
== LUSTRE_IMP_REPLAY
) {
1353 CDEBUG(D_HA
, "replay requested by %s\n",
1354 obd2cli_tgt(imp
->imp_obd
));
1355 rc
= ptlrpc_replay_next(imp
, &inflight
);
1356 if (inflight
== 0 &&
1357 atomic_read(&imp
->imp_replay_inflight
) == 0) {
1358 IMPORT_SET_STATE(imp
, LUSTRE_IMP_REPLAY_LOCKS
);
1359 rc
= ldlm_replay_locks(imp
);
1366 if (imp
->imp_state
== LUSTRE_IMP_REPLAY_LOCKS
) {
1367 if (atomic_read(&imp
->imp_replay_inflight
) == 0) {
1368 IMPORT_SET_STATE(imp
, LUSTRE_IMP_REPLAY_WAIT
);
1369 rc
= signal_completed_replay(imp
);
1376 if (imp
->imp_state
== LUSTRE_IMP_REPLAY_WAIT
) {
1377 if (atomic_read(&imp
->imp_replay_inflight
) == 0) {
1378 IMPORT_SET_STATE(imp
, LUSTRE_IMP_RECOVER
);
1382 if (imp
->imp_state
== LUSTRE_IMP_RECOVER
) {
1383 CDEBUG(D_HA
, "reconnected to %s@%s\n",
1384 obd2cli_tgt(imp
->imp_obd
),
1385 imp
->imp_connection
->c_remote_uuid
.uuid
);
1387 rc
= ptlrpc_resend(imp
);
1390 IMPORT_SET_STATE(imp
, LUSTRE_IMP_FULL
);
1391 ptlrpc_activate_import(imp
);
1393 deuuidify(obd2cli_tgt(imp
->imp_obd
), NULL
,
1394 &target_start
, &target_len
);
1395 LCONSOLE_INFO("%s: Connection restored to %.*s (at %s)\n",
1396 imp
->imp_obd
->obd_name
,
1397 target_len
, target_start
,
1398 libcfs_nid2str(imp
->imp_connection
->c_peer
.nid
));
1401 if (imp
->imp_state
== LUSTRE_IMP_FULL
) {
1402 wake_up_all(&imp
->imp_recovery_waitq
);
1403 ptlrpc_wake_delayed(imp
);
1410 int ptlrpc_disconnect_import(struct obd_import
*imp
, int noclose
)
1412 struct ptlrpc_request
*req
;
1415 if (imp
->imp_obd
->obd_force
)
1418 switch (imp
->imp_connect_op
) {
1420 rq_opc
= OST_DISCONNECT
;
1423 rq_opc
= MDS_DISCONNECT
;
1426 rq_opc
= MGS_DISCONNECT
;
1430 CERROR("%s: don't know how to disconnect from %s (connect_op %d): rc = %d\n",
1431 imp
->imp_obd
->obd_name
, obd2cli_tgt(imp
->imp_obd
),
1432 imp
->imp_connect_op
, rc
);
1436 if (ptlrpc_import_in_recovery(imp
)) {
1437 struct l_wait_info lwi
;
1441 if (imp
->imp_server_timeout
)
1442 timeout
= cfs_time_seconds(obd_timeout
/ 2);
1444 timeout
= cfs_time_seconds(obd_timeout
);
1446 int idx
= import_at_get_index(imp
,
1447 imp
->imp_client
->cli_request_portal
);
1448 timeout
= cfs_time_seconds(
1449 at_get(&imp
->imp_at
.iat_service_estimate
[idx
]));
1452 lwi
= LWI_TIMEOUT_INTR(cfs_timeout_cap(timeout
),
1453 back_to_sleep
, LWI_ON_SIGNAL_NOOP
, NULL
);
1454 rc
= l_wait_event(imp
->imp_recovery_waitq
,
1455 !ptlrpc_import_in_recovery(imp
), &lwi
);
1459 spin_lock(&imp
->imp_lock
);
1460 if (imp
->imp_state
!= LUSTRE_IMP_FULL
)
1462 spin_unlock(&imp
->imp_lock
);
1464 req
= ptlrpc_request_alloc_pack(imp
, &RQF_MDS_DISCONNECT
,
1465 LUSTRE_OBD_VERSION
, rq_opc
);
1467 /* We are disconnecting, do not retry a failed DISCONNECT rpc if
1468 * it fails. We can get through the above with a down server
1469 * if the client doesn't know the server is gone yet.
1471 req
->rq_no_resend
= 1;
1473 /* We want client umounts to happen quickly, no matter the
1476 req
->rq_timeout
= min_t(int, req
->rq_timeout
,
1477 INITIAL_CONNECT_TIMEOUT
);
1479 IMPORT_SET_STATE(imp
, LUSTRE_IMP_CONNECTING
);
1480 req
->rq_send_state
= LUSTRE_IMP_CONNECTING
;
1481 ptlrpc_request_set_replen(req
);
1482 rc
= ptlrpc_queue_wait(req
);
1483 ptlrpc_req_finished(req
);
1487 spin_lock(&imp
->imp_lock
);
1490 IMPORT_SET_STATE_NOLOCK(imp
, LUSTRE_IMP_DISCON
);
1492 IMPORT_SET_STATE_NOLOCK(imp
, LUSTRE_IMP_CLOSED
);
1493 memset(&imp
->imp_remote_handle
, 0, sizeof(imp
->imp_remote_handle
));
1494 spin_unlock(&imp
->imp_lock
);
1496 if (rc
== -ETIMEDOUT
|| rc
== -ENOTCONN
|| rc
== -ESHUTDOWN
)
1501 EXPORT_SYMBOL(ptlrpc_disconnect_import
);
1503 /* Adaptive Timeout utils */
1504 extern unsigned int at_min
, at_max
, at_history
;
1506 /* Bin into timeslices using AT_BINS bins.
1507 * This gives us a max of the last binlimit*AT_BINS secs without the storage,
1508 * but still smoothing out a return to normalcy from a slow response.
1509 * (E.g. remember the maximum latency in each minute of the last 4 minutes.)
1511 int at_measured(struct adaptive_timeout
*at
, unsigned int val
)
1513 unsigned int old
= at
->at_current
;
1514 time64_t now
= ktime_get_real_seconds();
1515 long binlimit
= max_t(long, at_history
/ AT_BINS
, 1);
1518 CDEBUG(D_OTHER
, "add %u to %p time=%lu v=%u (%u %u %u %u)\n",
1519 val
, at
, (long)(now
- at
->at_binstart
), at
->at_current
,
1520 at
->at_hist
[0], at
->at_hist
[1], at
->at_hist
[2], at
->at_hist
[3]);
1523 /* 0's don't count, because we never want our timeout to
1524 * drop to 0, and because 0 could mean an error
1528 spin_lock(&at
->at_lock
);
1530 if (unlikely(at
->at_binstart
== 0)) {
1531 /* Special case to remove default from history */
1532 at
->at_current
= val
;
1533 at
->at_worst_ever
= val
;
1534 at
->at_worst_time
= now
;
1535 at
->at_hist
[0] = val
;
1536 at
->at_binstart
= now
;
1537 } else if (now
- at
->at_binstart
< binlimit
) {
1539 at
->at_hist
[0] = max(val
, at
->at_hist
[0]);
1540 at
->at_current
= max(val
, at
->at_current
);
1543 unsigned int maxv
= val
;
1544 /* move bins over */
1545 shift
= (u32
)(now
- at
->at_binstart
) / binlimit
;
1547 for (i
= AT_BINS
- 1; i
>= 0; i
--) {
1549 at
->at_hist
[i
] = at
->at_hist
[i
- shift
];
1550 maxv
= max(maxv
, at
->at_hist
[i
]);
1555 at
->at_hist
[0] = val
;
1556 at
->at_current
= maxv
;
1557 at
->at_binstart
+= shift
* binlimit
;
1560 if (at
->at_current
> at
->at_worst_ever
) {
1561 at
->at_worst_ever
= at
->at_current
;
1562 at
->at_worst_time
= now
;
1565 if (at
->at_flags
& AT_FLG_NOHIST
)
1566 /* Only keep last reported val; keeping the rest of the history
1569 at
->at_current
= val
;
1572 at
->at_current
= min(at
->at_current
, at_max
);
1573 at
->at_current
= max(at
->at_current
, at_min
);
1575 if (at
->at_current
!= old
)
1576 CDEBUG(D_OTHER
, "AT %p change: old=%u new=%u delta=%d (val=%u) hist %u %u %u %u\n",
1578 old
, at
->at_current
, at
->at_current
- old
, val
,
1579 at
->at_hist
[0], at
->at_hist
[1], at
->at_hist
[2],
1582 /* if we changed, report the old value */
1583 old
= (at
->at_current
!= old
) ? old
: 0;
1585 spin_unlock(&at
->at_lock
);
1589 /* Find the imp_at index for a given portal; assign if space available */
1590 int import_at_get_index(struct obd_import
*imp
, int portal
)
1592 struct imp_at
*at
= &imp
->imp_at
;
1595 for (i
= 0; i
< IMP_AT_MAX_PORTALS
; i
++) {
1596 if (at
->iat_portal
[i
] == portal
)
1598 if (at
->iat_portal
[i
] == 0)
1603 /* Not found in list, add it under a lock */
1604 spin_lock(&imp
->imp_lock
);
1606 /* Check unused under lock */
1607 for (; i
< IMP_AT_MAX_PORTALS
; i
++) {
1608 if (at
->iat_portal
[i
] == portal
)
1610 if (at
->iat_portal
[i
] == 0)
1615 /* Not enough portals? */
1616 LASSERT(i
< IMP_AT_MAX_PORTALS
);
1618 at
->iat_portal
[i
] = portal
;
1620 spin_unlock(&imp
->imp_lock
);