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, 2012, 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
=
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); \
107 static int ptlrpc_connect_interpret(const struct lu_env
*env
,
108 struct ptlrpc_request
*request
,
110 int ptlrpc_import_recovery_state_machine(struct obd_import
*imp
);
112 /* Only this function is allowed to change the import state when it is
113 * CLOSED. I would rather refcount the import and free it after
114 * disconnection like we do with exports. To do that, the client_obd
115 * 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 void deuuidify(char *uuid
, const char *prefix
, char **uuid_start
, int *uuid_len
)
133 *uuid_start
= !prefix
|| strncmp(uuid
, prefix
, strlen(prefix
))
134 ? uuid
: uuid
+ strlen(prefix
);
136 *uuid_len
= strlen(*uuid_start
);
138 if (*uuid_len
< strlen(UUID_STR
))
141 if (!strncmp(*uuid_start
+ *uuid_len
- strlen(UUID_STR
),
142 UUID_STR
, strlen(UUID_STR
)))
143 *uuid_len
-= strlen(UUID_STR
);
145 EXPORT_SYMBOL(deuuidify
);
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
, time_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 time_t now
= get_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. */
288 /* Calculate max timeout for waiting on rpcs to error
289 * out. Use obd_timeout if calculated value is smaller
291 if (!OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_REPL_UNLINK
)) {
292 timeout
= ptlrpc_inflight_timeout(imp
);
293 timeout
+= timeout
/ 3;
296 timeout
= obd_timeout
;
298 /* decrease the interval to increase race condition */
303 "Sleeping %d sec for inflight to error out\n",
306 /* Wait for all requests to error out and call completion
307 * callbacks. Cap it at obd_timeout -- these should all
308 * have been locally cancelled by ptlrpc_abort_inflight. */
309 lwi
= LWI_TIMEOUT_INTERVAL(
310 cfs_timeout_cap(cfs_time_seconds(timeout
)),
311 (timeout
> 1)?cfs_time_seconds(1):cfs_time_seconds(1)/2,
313 rc
= l_wait_event(imp
->imp_recovery_waitq
,
314 (atomic_read(&imp
->imp_inflight
) == 0),
317 const char *cli_tgt
= obd2cli_tgt(imp
->imp_obd
);
319 CERROR("%s: rc = %d waiting for callback (%d != 0)\n",
321 atomic_read(&imp
->imp_inflight
));
323 spin_lock(&imp
->imp_lock
);
324 if (atomic_read(&imp
->imp_inflight
) == 0) {
325 int count
= atomic_read(&imp
->imp_unregistering
);
327 /* We know that "unregistering" rpcs only can
328 * survive in sending or delaying lists (they
329 * maybe waiting for long reply unlink in
330 * sluggish nets). Let's check this. If there
331 * is no inflight and unregistering != 0, this
333 LASSERTF(count
== 0, "Some RPCs are still unregistering: %d\n",
336 /* Let's save one loop as soon as inflight have
337 * dropped to zero. No new inflights possible at
341 list_for_each_safe(tmp
, n
,
342 &imp
->imp_sending_list
) {
343 req
= list_entry(tmp
,
344 struct ptlrpc_request
,
346 DEBUG_REQ(D_ERROR
, req
,
347 "still on sending list");
349 list_for_each_safe(tmp
, n
,
350 &imp
->imp_delayed_list
) {
351 req
= list_entry(tmp
,
352 struct ptlrpc_request
,
354 DEBUG_REQ(D_ERROR
, req
,
355 "still on delayed list");
358 CERROR("%s: RPCs in \"%s\" phase found (%d). Network is sluggish? Waiting them to error out.\n",
360 ptlrpc_phase2str(RQ_PHASE_UNREGISTERING
),
364 spin_unlock(&imp
->imp_lock
);
369 * Let's additionally check that no new rpcs added to import in
370 * "invalidate" state.
372 LASSERT(atomic_read(&imp
->imp_inflight
) == 0);
373 obd_import_event(imp
->imp_obd
, imp
, IMP_EVENT_INVALIDATE
);
374 sptlrpc_import_flush_all_ctx(imp
);
376 atomic_dec(&imp
->imp_inval_count
);
377 wake_up_all(&imp
->imp_recovery_waitq
);
379 EXPORT_SYMBOL(ptlrpc_invalidate_import
);
381 /* unset imp_invalid */
382 void ptlrpc_activate_import(struct obd_import
*imp
)
384 struct obd_device
*obd
= imp
->imp_obd
;
386 spin_lock(&imp
->imp_lock
);
387 if (imp
->imp_deactive
!= 0) {
388 spin_unlock(&imp
->imp_lock
);
392 imp
->imp_invalid
= 0;
393 spin_unlock(&imp
->imp_lock
);
394 obd_import_event(obd
, imp
, IMP_EVENT_ACTIVE
);
396 EXPORT_SYMBOL(ptlrpc_activate_import
);
398 static void ptlrpc_pinger_force(struct obd_import
*imp
)
400 CDEBUG(D_HA
, "%s: waking up pinger s:%s\n", obd2cli_tgt(imp
->imp_obd
),
401 ptlrpc_import_state_name(imp
->imp_state
));
403 spin_lock(&imp
->imp_lock
);
404 imp
->imp_force_verify
= 1;
405 spin_unlock(&imp
->imp_lock
);
407 if (imp
->imp_state
!= LUSTRE_IMP_CONNECTING
)
408 ptlrpc_pinger_wake_up();
411 void ptlrpc_fail_import(struct obd_import
*imp
, __u32 conn_cnt
)
413 LASSERT(!imp
->imp_dlm_fake
);
415 if (ptlrpc_set_import_discon(imp
, conn_cnt
)) {
416 if (!imp
->imp_replayable
) {
417 CDEBUG(D_HA
, "import %s@%s for %s not replayable, auto-deactivating\n",
418 obd2cli_tgt(imp
->imp_obd
),
419 imp
->imp_connection
->c_remote_uuid
.uuid
,
420 imp
->imp_obd
->obd_name
);
421 ptlrpc_deactivate_import(imp
);
424 ptlrpc_pinger_force(imp
);
427 EXPORT_SYMBOL(ptlrpc_fail_import
);
429 int ptlrpc_reconnect_import(struct obd_import
*imp
)
432 struct l_wait_info lwi
;
433 int secs
= cfs_time_seconds(obd_timeout
);
436 ptlrpc_pinger_force(imp
);
438 CDEBUG(D_HA
, "%s: recovery started, waiting %u seconds\n",
439 obd2cli_tgt(imp
->imp_obd
), secs
);
441 lwi
= LWI_TIMEOUT(secs
, NULL
, NULL
);
442 rc
= l_wait_event(imp
->imp_recovery_waitq
,
443 !ptlrpc_import_in_recovery(imp
), &lwi
);
444 CDEBUG(D_HA
, "%s: recovery finished s:%s\n", obd2cli_tgt(imp
->imp_obd
),
445 ptlrpc_import_state_name(imp
->imp_state
));
448 ptlrpc_set_import_discon(imp
, 0);
449 /* Force a new connect attempt */
450 ptlrpc_invalidate_import(imp
);
451 /* Do a fresh connect next time by zeroing the handle */
452 ptlrpc_disconnect_import(imp
, 1);
453 /* Wait for all invalidate calls to finish */
454 if (atomic_read(&imp
->imp_inval_count
) > 0) {
456 struct l_wait_info lwi
= LWI_INTR(LWI_ON_SIGNAL_NOOP
, NULL
);
457 rc
= l_wait_event(imp
->imp_recovery_waitq
,
458 (atomic_read(&imp
->imp_inval_count
) == 0),
461 CERROR("Interrupted, inval=%d\n",
462 atomic_read(&imp
->imp_inval_count
));
465 /* Allow reconnect attempts */
466 imp
->imp_obd
->obd_no_recov
= 0;
467 /* Remove 'invalid' flag */
468 ptlrpc_activate_import(imp
);
469 /* Attempt a new connect */
470 ptlrpc_recover_import(imp
, NULL
, 0);
474 EXPORT_SYMBOL(ptlrpc_reconnect_import
);
477 * Connection on import \a imp is changed to another one (if more than one is
478 * present). We typically chose connection that we have not tried to connect to
481 static int import_select_connection(struct obd_import
*imp
)
483 struct obd_import_conn
*imp_conn
= NULL
, *conn
;
484 struct obd_export
*dlmexp
;
486 int target_len
, tried_all
= 1;
488 spin_lock(&imp
->imp_lock
);
490 if (list_empty(&imp
->imp_conn_list
)) {
491 CERROR("%s: no connections available\n",
492 imp
->imp_obd
->obd_name
);
493 spin_unlock(&imp
->imp_lock
);
497 list_for_each_entry(conn
, &imp
->imp_conn_list
, oic_item
) {
498 CDEBUG(D_HA
, "%s: connect to NID %s last attempt %llu\n",
499 imp
->imp_obd
->obd_name
,
500 libcfs_nid2str(conn
->oic_conn
->c_peer
.nid
),
501 conn
->oic_last_attempt
);
503 /* If we have not tried this connection since
504 the last successful attempt, go with this one */
505 if ((conn
->oic_last_attempt
== 0) ||
506 cfs_time_beforeq_64(conn
->oic_last_attempt
,
507 imp
->imp_last_success_conn
)) {
513 /* If all of the connections have already been tried
514 since the last successful connection; just choose the
515 least recently used */
518 else if (cfs_time_before_64(conn
->oic_last_attempt
,
519 imp_conn
->oic_last_attempt
))
523 /* if not found, simply choose the current one */
524 if (!imp_conn
|| imp
->imp_force_reconnect
) {
525 LASSERT(imp
->imp_conn_current
);
526 imp_conn
= imp
->imp_conn_current
;
529 LASSERT(imp_conn
->oic_conn
);
531 /* If we've tried everything, and we're back to the beginning of the
532 list, increase our timeout and try again. It will be reset when
533 we do finally connect. (FIXME: really we should wait for all network
534 state associated with the last connection attempt to drain before
535 trying to reconnect on it.) */
536 if (tried_all
&& (imp
->imp_conn_list
.next
== &imp_conn
->oic_item
)) {
537 struct adaptive_timeout
*at
= &imp
->imp_at
.iat_net_latency
;
538 if (at_get(at
) < CONNECTION_SWITCH_MAX
) {
539 at_measured(at
, at_get(at
) + CONNECTION_SWITCH_INC
);
540 if (at_get(at
) > CONNECTION_SWITCH_MAX
)
541 at_reset(at
, CONNECTION_SWITCH_MAX
);
543 LASSERT(imp_conn
->oic_last_attempt
);
544 CDEBUG(D_HA
, "%s: tried all connections, increasing latency to %ds\n",
545 imp
->imp_obd
->obd_name
, at_get(at
));
548 imp_conn
->oic_last_attempt
= cfs_time_current_64();
550 /* switch connection, don't mind if it's same as the current one */
551 ptlrpc_connection_put(imp
->imp_connection
);
552 imp
->imp_connection
= ptlrpc_connection_addref(imp_conn
->oic_conn
);
554 dlmexp
= class_conn2export(&imp
->imp_dlm_handle
);
555 LASSERT(dlmexp
!= NULL
);
556 ptlrpc_connection_put(dlmexp
->exp_connection
);
557 dlmexp
->exp_connection
= ptlrpc_connection_addref(imp_conn
->oic_conn
);
558 class_export_put(dlmexp
);
560 if (imp
->imp_conn_current
!= imp_conn
) {
561 if (imp
->imp_conn_current
) {
562 deuuidify(obd2cli_tgt(imp
->imp_obd
), NULL
,
563 &target_start
, &target_len
);
565 CDEBUG(D_HA
, "%s: Connection changing to %.*s (at %s)\n",
566 imp
->imp_obd
->obd_name
,
567 target_len
, target_start
,
568 libcfs_nid2str(imp_conn
->oic_conn
->c_peer
.nid
));
571 imp
->imp_conn_current
= imp_conn
;
574 CDEBUG(D_HA
, "%s: import %p using connection %s/%s\n",
575 imp
->imp_obd
->obd_name
, imp
, imp_conn
->oic_uuid
.uuid
,
576 libcfs_nid2str(imp_conn
->oic_conn
->c_peer
.nid
));
578 spin_unlock(&imp
->imp_lock
);
584 * must be called under imp_lock
586 static int ptlrpc_first_transno(struct obd_import
*imp
, __u64
*transno
)
588 struct ptlrpc_request
*req
;
589 struct list_head
*tmp
;
591 /* The requests in committed_list always have smaller transnos than
592 * the requests in replay_list */
593 if (!list_empty(&imp
->imp_committed_list
)) {
594 tmp
= imp
->imp_committed_list
.next
;
595 req
= list_entry(tmp
, struct ptlrpc_request
, rq_replay_list
);
596 *transno
= req
->rq_transno
;
597 if (req
->rq_transno
== 0) {
598 DEBUG_REQ(D_ERROR
, req
,
599 "zero transno in committed_list");
604 if (!list_empty(&imp
->imp_replay_list
)) {
605 tmp
= imp
->imp_replay_list
.next
;
606 req
= list_entry(tmp
, struct ptlrpc_request
, rq_replay_list
);
607 *transno
= req
->rq_transno
;
608 if (req
->rq_transno
== 0) {
609 DEBUG_REQ(D_ERROR
, req
, "zero transno in replay_list");
618 * Attempt to (re)connect import \a imp. This includes all preparations,
619 * initializing CONNECT RPC request and passing it to ptlrpcd for
621 * Returns 0 on success or error code.
623 int ptlrpc_connect_import(struct obd_import
*imp
)
625 struct obd_device
*obd
= imp
->imp_obd
;
626 int initial_connect
= 0;
628 __u64 committed_before_reconnect
= 0;
629 struct ptlrpc_request
*request
;
630 char *bufs
[] = { NULL
,
631 obd2cli_tgt(imp
->imp_obd
),
633 (char *)&imp
->imp_dlm_handle
,
634 (char *)&imp
->imp_connect_data
};
635 struct ptlrpc_connect_async_args
*aa
;
638 spin_lock(&imp
->imp_lock
);
639 if (imp
->imp_state
== LUSTRE_IMP_CLOSED
) {
640 spin_unlock(&imp
->imp_lock
);
641 CERROR("can't connect to a closed import\n");
643 } else if (imp
->imp_state
== LUSTRE_IMP_FULL
) {
644 spin_unlock(&imp
->imp_lock
);
645 CERROR("already connected\n");
647 } else if (imp
->imp_state
== LUSTRE_IMP_CONNECTING
) {
648 spin_unlock(&imp
->imp_lock
);
649 CERROR("already connecting\n");
653 IMPORT_SET_STATE_NOLOCK(imp
, LUSTRE_IMP_CONNECTING
);
656 imp
->imp_resend_replay
= 0;
658 if (!lustre_handle_is_used(&imp
->imp_remote_handle
))
661 committed_before_reconnect
= imp
->imp_peer_committed_transno
;
663 set_transno
= ptlrpc_first_transno(imp
,
664 &imp
->imp_connect_data
.ocd_transno
);
665 spin_unlock(&imp
->imp_lock
);
667 rc
= import_select_connection(imp
);
671 rc
= sptlrpc_import_sec_adapt(imp
, NULL
, NULL
);
675 /* Reset connect flags to the originally requested flags, in case
676 * the server is updated on-the-fly we will get the new features. */
677 imp
->imp_connect_data
.ocd_connect_flags
= imp
->imp_connect_flags_orig
;
678 /* Reset ocd_version each time so the server knows the exact versions */
679 imp
->imp_connect_data
.ocd_version
= LUSTRE_VERSION_CODE
;
680 imp
->imp_msghdr_flags
&= ~MSGHDR_AT_SUPPORT
;
681 imp
->imp_msghdr_flags
&= ~MSGHDR_CKSUM_INCOMPAT18
;
683 rc
= obd_reconnect(NULL
, imp
->imp_obd
->obd_self_export
, obd
,
684 &obd
->obd_uuid
, &imp
->imp_connect_data
, NULL
);
688 request
= ptlrpc_request_alloc(imp
, &RQF_MDS_CONNECT
);
689 if (request
== NULL
) {
694 rc
= ptlrpc_request_bufs_pack(request
, LUSTRE_OBD_VERSION
,
695 imp
->imp_connect_op
, bufs
, NULL
);
697 ptlrpc_request_free(request
);
701 /* Report the rpc service time to the server so that it knows how long
702 * to wait for clients to join recovery */
703 lustre_msg_set_service_time(request
->rq_reqmsg
,
704 at_timeout2est(request
->rq_timeout
));
706 /* The amount of time we give the server to process the connect req.
707 * import_select_connection will increase the net latency on
708 * repeated reconnect attempts to cover slow networks.
709 * We override/ignore the server rpc completion estimate here,
710 * which may be large if this is a reconnect attempt */
711 request
->rq_timeout
= INITIAL_CONNECT_TIMEOUT
;
712 lustre_msg_set_timeout(request
->rq_reqmsg
, request
->rq_timeout
);
714 lustre_msg_add_op_flags(request
->rq_reqmsg
, MSG_CONNECT_NEXT_VER
);
716 request
->rq_no_resend
= request
->rq_no_delay
= 1;
717 request
->rq_send_state
= LUSTRE_IMP_CONNECTING
;
718 /* Allow a slightly larger reply for future growth compatibility */
719 req_capsule_set_size(&request
->rq_pill
, &RMF_CONNECT_DATA
, RCL_SERVER
,
720 sizeof(struct obd_connect_data
)+16*sizeof(__u64
));
721 ptlrpc_request_set_replen(request
);
722 request
->rq_interpret_reply
= ptlrpc_connect_interpret
;
724 CLASSERT(sizeof(*aa
) <= sizeof(request
->rq_async_args
));
725 aa
= ptlrpc_req_async_args(request
);
726 memset(aa
, 0, sizeof(*aa
));
728 aa
->pcaa_peer_committed
= committed_before_reconnect
;
729 aa
->pcaa_initial_connect
= initial_connect
;
731 if (aa
->pcaa_initial_connect
) {
732 spin_lock(&imp
->imp_lock
);
733 imp
->imp_replayable
= 1;
734 spin_unlock(&imp
->imp_lock
);
735 lustre_msg_add_op_flags(request
->rq_reqmsg
,
736 MSG_CONNECT_INITIAL
);
740 lustre_msg_add_op_flags(request
->rq_reqmsg
,
741 MSG_CONNECT_TRANSNO
);
743 DEBUG_REQ(D_RPCTRACE
, request
, "(re)connect request (timeout %d)",
744 request
->rq_timeout
);
745 ptlrpcd_add_req(request
);
749 IMPORT_SET_STATE(imp
, LUSTRE_IMP_DISCON
);
753 EXPORT_SYMBOL(ptlrpc_connect_import
);
755 static void ptlrpc_maybe_ping_import_soon(struct obd_import
*imp
)
759 spin_lock(&imp
->imp_lock
);
760 force_verify
= imp
->imp_force_verify
!= 0;
761 spin_unlock(&imp
->imp_lock
);
764 ptlrpc_pinger_wake_up();
767 static int ptlrpc_busy_reconnect(int rc
)
769 return (rc
== -EBUSY
) || (rc
== -EAGAIN
);
773 * interpret_reply callback for connect RPCs.
774 * Looks into returned status of connect operation and decides
775 * what to do with the import - i.e enter recovery, promote it to
776 * full state for normal operations of disconnect it due to an error.
778 static int ptlrpc_connect_interpret(const struct lu_env
*env
,
779 struct ptlrpc_request
*request
,
782 struct ptlrpc_connect_async_args
*aa
= data
;
783 struct obd_import
*imp
= request
->rq_import
;
784 struct client_obd
*cli
= &imp
->imp_obd
->u
.cli
;
785 struct lustre_handle old_hdl
;
786 __u64 old_connect_flags
;
788 struct obd_connect_data
*ocd
;
789 struct obd_export
*exp
;
792 spin_lock(&imp
->imp_lock
);
793 if (imp
->imp_state
== LUSTRE_IMP_CLOSED
) {
794 imp
->imp_connect_tried
= 1;
795 spin_unlock(&imp
->imp_lock
);
800 /* if this reconnect to busy export - not need select new target
802 imp
->imp_force_reconnect
= ptlrpc_busy_reconnect(rc
);
803 spin_unlock(&imp
->imp_lock
);
804 ptlrpc_maybe_ping_import_soon(imp
);
807 spin_unlock(&imp
->imp_lock
);
809 LASSERT(imp
->imp_conn_current
);
811 msg_flags
= lustre_msg_get_op_flags(request
->rq_repmsg
);
813 ret
= req_capsule_get_size(&request
->rq_pill
, &RMF_CONNECT_DATA
,
815 /* server replied obd_connect_data is always bigger */
816 ocd
= req_capsule_server_sized_get(&request
->rq_pill
,
817 &RMF_CONNECT_DATA
, ret
);
820 CERROR("%s: no connect data from server\n",
821 imp
->imp_obd
->obd_name
);
826 spin_lock(&imp
->imp_lock
);
828 /* All imports are pingable */
829 imp
->imp_pingable
= 1;
830 imp
->imp_force_reconnect
= 0;
831 imp
->imp_force_verify
= 0;
833 imp
->imp_connect_data
= *ocd
;
835 CDEBUG(D_HA
, "%s: connect to target with instance %u\n",
836 imp
->imp_obd
->obd_name
, ocd
->ocd_instance
);
837 exp
= class_conn2export(&imp
->imp_dlm_handle
);
839 spin_unlock(&imp
->imp_lock
);
841 /* check that server granted subset of flags we asked for. */
842 if ((ocd
->ocd_connect_flags
& imp
->imp_connect_flags_orig
) !=
843 ocd
->ocd_connect_flags
) {
844 CERROR("%s: Server didn't granted asked subset of flags: asked=%#llx grranted=%#llx\n",
845 imp
->imp_obd
->obd_name
, imp
->imp_connect_flags_orig
,
846 ocd
->ocd_connect_flags
);
852 /* This could happen if export is cleaned during the
854 CERROR("%s: missing export after connect\n",
855 imp
->imp_obd
->obd_name
);
859 old_connect_flags
= exp_connect_flags(exp
);
860 exp
->exp_connect_data
= *ocd
;
861 imp
->imp_obd
->obd_self_export
->exp_connect_data
= *ocd
;
862 class_export_put(exp
);
864 obd_import_event(imp
->imp_obd
, imp
, IMP_EVENT_OCD
);
866 if (aa
->pcaa_initial_connect
) {
867 spin_lock(&imp
->imp_lock
);
868 if (msg_flags
& MSG_CONNECT_REPLAYABLE
) {
869 imp
->imp_replayable
= 1;
870 spin_unlock(&imp
->imp_lock
);
871 CDEBUG(D_HA
, "connected to replayable target: %s\n",
872 obd2cli_tgt(imp
->imp_obd
));
874 imp
->imp_replayable
= 0;
875 spin_unlock(&imp
->imp_lock
);
878 /* if applies, adjust the imp->imp_msg_magic here
879 * according to reply flags */
881 imp
->imp_remote_handle
=
882 *lustre_msg_get_handle(request
->rq_repmsg
);
884 /* Initial connects are allowed for clients with non-random
885 * uuids when servers are in recovery. Simply signal the
886 * servers replay is complete and wait in REPLAY_WAIT. */
887 if (msg_flags
& MSG_CONNECT_RECOVERING
) {
888 CDEBUG(D_HA
, "connect to %s during recovery\n",
889 obd2cli_tgt(imp
->imp_obd
));
890 IMPORT_SET_STATE(imp
, LUSTRE_IMP_REPLAY_LOCKS
);
892 IMPORT_SET_STATE(imp
, LUSTRE_IMP_FULL
);
893 ptlrpc_activate_import(imp
);
900 /* Determine what recovery state to move the import to. */
901 if (msg_flags
& MSG_CONNECT_RECONNECT
) {
902 memset(&old_hdl
, 0, sizeof(old_hdl
));
903 if (!memcmp(&old_hdl
, lustre_msg_get_handle(request
->rq_repmsg
),
905 LCONSOLE_WARN("Reconnect to %s (at @%s) failed due bad handle %#llx\n",
906 obd2cli_tgt(imp
->imp_obd
),
907 imp
->imp_connection
->c_remote_uuid
.uuid
,
908 imp
->imp_dlm_handle
.cookie
);
913 if (memcmp(&imp
->imp_remote_handle
,
914 lustre_msg_get_handle(request
->rq_repmsg
),
915 sizeof(imp
->imp_remote_handle
))) {
916 int level
= msg_flags
& MSG_CONNECT_RECOVERING
?
919 /* Bug 16611/14775: if server handle have changed,
920 * that means some sort of disconnection happened.
921 * If the server is not in recovery, that also means it
922 * already erased all of our state because of previous
923 * eviction. If it is in recovery - we are safe to
924 * participate since we can reestablish all of our state
925 * with server again */
926 if ((msg_flags
& MSG_CONNECT_RECOVERING
)) {
927 CDEBUG(level
, "%s@%s changed server handle from %#llx to %#llx but is still in recovery\n",
928 obd2cli_tgt(imp
->imp_obd
),
929 imp
->imp_connection
->c_remote_uuid
.uuid
,
930 imp
->imp_remote_handle
.cookie
,
931 lustre_msg_get_handle(
932 request
->rq_repmsg
)->cookie
);
934 LCONSOLE_WARN("Evicted from %s (at %s) after server handle changed from %#llx to %#llx\n",
935 obd2cli_tgt(imp
->imp_obd
),
936 imp
->imp_connection
-> \
938 imp
->imp_remote_handle
.cookie
,
939 lustre_msg_get_handle(
940 request
->rq_repmsg
)->cookie
);
944 imp
->imp_remote_handle
=
945 *lustre_msg_get_handle(request
->rq_repmsg
);
947 if (!(msg_flags
& MSG_CONNECT_RECOVERING
)) {
948 IMPORT_SET_STATE(imp
, LUSTRE_IMP_EVICTED
);
954 CDEBUG(D_HA
, "reconnected to %s@%s after partition\n",
955 obd2cli_tgt(imp
->imp_obd
),
956 imp
->imp_connection
->c_remote_uuid
.uuid
);
959 if (imp
->imp_invalid
) {
960 CDEBUG(D_HA
, "%s: reconnected but import is invalid; marking evicted\n",
961 imp
->imp_obd
->obd_name
);
962 IMPORT_SET_STATE(imp
, LUSTRE_IMP_EVICTED
);
963 } else if (msg_flags
& MSG_CONNECT_RECOVERING
) {
964 CDEBUG(D_HA
, "%s: reconnected to %s during replay\n",
965 imp
->imp_obd
->obd_name
,
966 obd2cli_tgt(imp
->imp_obd
));
968 spin_lock(&imp
->imp_lock
);
969 imp
->imp_resend_replay
= 1;
970 spin_unlock(&imp
->imp_lock
);
972 IMPORT_SET_STATE(imp
, imp
->imp_replay_state
);
974 IMPORT_SET_STATE(imp
, LUSTRE_IMP_RECOVER
);
976 } else if ((msg_flags
& MSG_CONNECT_RECOVERING
) && !imp
->imp_invalid
) {
977 LASSERT(imp
->imp_replayable
);
978 imp
->imp_remote_handle
=
979 *lustre_msg_get_handle(request
->rq_repmsg
);
980 imp
->imp_last_replay_transno
= 0;
981 IMPORT_SET_STATE(imp
, LUSTRE_IMP_REPLAY
);
983 DEBUG_REQ(D_HA
, request
, "%s: evicting (reconnect/recover flags not set: %x)",
984 imp
->imp_obd
->obd_name
, msg_flags
);
985 imp
->imp_remote_handle
=
986 *lustre_msg_get_handle(request
->rq_repmsg
);
987 IMPORT_SET_STATE(imp
, LUSTRE_IMP_EVICTED
);
990 /* Sanity checks for a reconnected import. */
991 if (!(imp
->imp_replayable
) != !(msg_flags
& MSG_CONNECT_REPLAYABLE
)) {
992 CERROR("imp_replayable flag does not match server after reconnect. We should LBUG right here.\n");
995 if (lustre_msg_get_last_committed(request
->rq_repmsg
) > 0 &&
996 lustre_msg_get_last_committed(request
->rq_repmsg
) <
997 aa
->pcaa_peer_committed
) {
998 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",
999 obd2cli_tgt(imp
->imp_obd
), aa
->pcaa_peer_committed
,
1000 lustre_msg_get_last_committed(request
->rq_repmsg
));
1004 rc
= ptlrpc_import_recovery_state_machine(imp
);
1006 if (rc
== -ENOTCONN
) {
1007 CDEBUG(D_HA
, "evicted/aborted by %s@%s during recovery; invalidating and reconnecting\n",
1008 obd2cli_tgt(imp
->imp_obd
),
1009 imp
->imp_connection
->c_remote_uuid
.uuid
);
1010 ptlrpc_connect_import(imp
);
1011 imp
->imp_connect_tried
= 1;
1016 spin_lock(&imp
->imp_lock
);
1017 list_del(&imp
->imp_conn_current
->oic_item
);
1018 list_add(&imp
->imp_conn_current
->oic_item
,
1019 &imp
->imp_conn_list
);
1020 imp
->imp_last_success_conn
=
1021 imp
->imp_conn_current
->oic_last_attempt
;
1023 spin_unlock(&imp
->imp_lock
);
1025 if ((imp
->imp_connect_flags_orig
& OBD_CONNECT_IBITS
) &&
1026 !(ocd
->ocd_connect_flags
& OBD_CONNECT_IBITS
)) {
1027 LCONSOLE_WARN("%s: MDS %s does not support ibits lock, either very old or invalid: requested %llx, replied %llx\n",
1028 imp
->imp_obd
->obd_name
,
1029 imp
->imp_connection
->c_remote_uuid
.uuid
,
1030 imp
->imp_connect_flags_orig
,
1031 ocd
->ocd_connect_flags
);
1036 if ((ocd
->ocd_connect_flags
& OBD_CONNECT_VERSION
) &&
1037 (ocd
->ocd_version
> LUSTRE_VERSION_CODE
+
1038 LUSTRE_VERSION_OFFSET_WARN
||
1039 ocd
->ocd_version
< LUSTRE_VERSION_CODE
-
1040 LUSTRE_VERSION_OFFSET_WARN
)) {
1041 /* Sigh, some compilers do not like #ifdef in the middle
1042 of macro arguments */
1043 const char *older
= "older. Consider upgrading server or downgrading client"
1045 const char *newer
= "newer than client version. Consider upgrading client"
1048 LCONSOLE_WARN("Server %s version (%d.%d.%d.%d) is much %s (%s)\n",
1049 obd2cli_tgt(imp
->imp_obd
),
1050 OBD_OCD_VERSION_MAJOR(ocd
->ocd_version
),
1051 OBD_OCD_VERSION_MINOR(ocd
->ocd_version
),
1052 OBD_OCD_VERSION_PATCH(ocd
->ocd_version
),
1053 OBD_OCD_VERSION_FIX(ocd
->ocd_version
),
1054 ocd
->ocd_version
> LUSTRE_VERSION_CODE
?
1055 newer
: older
, LUSTRE_VERSION_STRING
);
1058 #if LUSTRE_VERSION_CODE < OBD_OCD_VERSION(3, 2, 50, 0)
1059 /* Check if server has LU-1252 fix applied to not always swab
1060 * the IR MNE entries. Do this only once per connection. This
1061 * fixup is version-limited, because we don't want to carry the
1062 * OBD_CONNECT_MNE_SWAB flag around forever, just so long as we
1063 * need interop with unpatched 2.2 servers. For newer servers,
1064 * the client will do MNE swabbing only as needed. LU-1644 */
1065 if (unlikely((ocd
->ocd_connect_flags
& OBD_CONNECT_VERSION
) &&
1066 !(ocd
->ocd_connect_flags
& OBD_CONNECT_MNE_SWAB
) &&
1067 OBD_OCD_VERSION_MAJOR(ocd
->ocd_version
) == 2 &&
1068 OBD_OCD_VERSION_MINOR(ocd
->ocd_version
) == 2 &&
1069 OBD_OCD_VERSION_PATCH(ocd
->ocd_version
) < 55 &&
1070 strcmp(imp
->imp_obd
->obd_type
->typ_name
,
1071 LUSTRE_MGC_NAME
) == 0))
1072 imp
->imp_need_mne_swab
= 1;
1073 else /* clear if server was upgraded since last connect */
1074 imp
->imp_need_mne_swab
= 0;
1076 #warning "LU-1644: Remove old OBD_CONNECT_MNE_SWAB fixup and imp_need_mne_swab"
1079 if (ocd
->ocd_connect_flags
& OBD_CONNECT_CKSUM
) {
1080 /* We sent to the server ocd_cksum_types with bits set
1081 * for algorithms we understand. The server masked off
1082 * the checksum types it doesn't support */
1083 if ((ocd
->ocd_cksum_types
&
1084 cksum_types_supported_client()) == 0) {
1085 LCONSOLE_WARN("The negotiation of the checksum algorithm to use with server %s failed (%x/%x), disabling checksums\n",
1086 obd2cli_tgt(imp
->imp_obd
),
1087 ocd
->ocd_cksum_types
,
1088 cksum_types_supported_client());
1089 cli
->cl_checksum
= 0;
1090 cli
->cl_supp_cksum_types
= OBD_CKSUM_ADLER
;
1092 cli
->cl_supp_cksum_types
= ocd
->ocd_cksum_types
;
1095 /* The server does not support OBD_CONNECT_CKSUM.
1096 * Enforce ADLER for backward compatibility*/
1097 cli
->cl_supp_cksum_types
= OBD_CKSUM_ADLER
;
1099 cli
->cl_cksum_type
= cksum_type_select(cli
->cl_supp_cksum_types
);
1101 if (ocd
->ocd_connect_flags
& OBD_CONNECT_BRW_SIZE
)
1102 cli
->cl_max_pages_per_rpc
=
1103 min(ocd
->ocd_brw_size
>> PAGE_CACHE_SHIFT
,
1104 cli
->cl_max_pages_per_rpc
);
1105 else if (imp
->imp_connect_op
== MDS_CONNECT
||
1106 imp
->imp_connect_op
== MGS_CONNECT
)
1107 cli
->cl_max_pages_per_rpc
= 1;
1109 /* Reset ns_connect_flags only for initial connect. It might be
1110 * changed in while using FS and if we reset it in reconnect
1111 * this leads to losing user settings done before such as
1112 * disable lru_resize, etc. */
1113 if (old_connect_flags
!= exp_connect_flags(exp
) ||
1114 aa
->pcaa_initial_connect
) {
1115 CDEBUG(D_HA
, "%s: Resetting ns_connect_flags to server flags: %#llx\n",
1116 imp
->imp_obd
->obd_name
, ocd
->ocd_connect_flags
);
1117 imp
->imp_obd
->obd_namespace
->ns_connect_flags
=
1118 ocd
->ocd_connect_flags
;
1119 imp
->imp_obd
->obd_namespace
->ns_orig_connect_flags
=
1120 ocd
->ocd_connect_flags
;
1123 if ((ocd
->ocd_connect_flags
& OBD_CONNECT_AT
) &&
1124 (imp
->imp_msg_magic
== LUSTRE_MSG_MAGIC_V2
))
1125 /* We need a per-message support flag, because
1126 a. we don't know if the incoming connect reply
1127 supports AT or not (in reply_in_callback)
1129 b. failovered server means export and flags are gone
1130 (in ptlrpc_send_reply).
1131 Can only be set when we know AT is supported at
1133 imp
->imp_msghdr_flags
|= MSGHDR_AT_SUPPORT
;
1135 imp
->imp_msghdr_flags
&= ~MSGHDR_AT_SUPPORT
;
1137 if ((ocd
->ocd_connect_flags
& OBD_CONNECT_FULL20
) &&
1138 (imp
->imp_msg_magic
== LUSTRE_MSG_MAGIC_V2
))
1139 imp
->imp_msghdr_flags
|= MSGHDR_CKSUM_INCOMPAT18
;
1141 imp
->imp_msghdr_flags
&= ~MSGHDR_CKSUM_INCOMPAT18
;
1143 LASSERT((cli
->cl_max_pages_per_rpc
<= PTLRPC_MAX_BRW_PAGES
) &&
1144 (cli
->cl_max_pages_per_rpc
> 0));
1148 imp
->imp_connect_tried
= 1;
1151 IMPORT_SET_STATE(imp
, LUSTRE_IMP_DISCON
);
1152 if (rc
== -EACCES
) {
1154 * Give up trying to reconnect
1155 * EACCES means client has no permission for connection
1157 imp
->imp_obd
->obd_no_recov
= 1;
1158 ptlrpc_deactivate_import(imp
);
1161 if (rc
== -EPROTO
) {
1162 struct obd_connect_data
*ocd
;
1164 /* reply message might not be ready */
1165 if (request
->rq_repmsg
== NULL
)
1168 ocd
= req_capsule_server_get(&request
->rq_pill
,
1171 (ocd
->ocd_connect_flags
& OBD_CONNECT_VERSION
) &&
1172 (ocd
->ocd_version
!= LUSTRE_VERSION_CODE
)) {
1174 * Actually servers are only supposed to refuse
1175 * connection from liblustre clients, so we
1176 * should never see this from VFS context
1178 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",
1179 obd2cli_tgt(imp
->imp_obd
),
1180 OBD_OCD_VERSION_MAJOR(ocd
->ocd_version
),
1181 OBD_OCD_VERSION_MINOR(ocd
->ocd_version
),
1182 OBD_OCD_VERSION_PATCH(ocd
->ocd_version
),
1183 OBD_OCD_VERSION_FIX(ocd
->ocd_version
),
1184 LUSTRE_VERSION_STRING
);
1185 ptlrpc_deactivate_import(imp
);
1186 IMPORT_SET_STATE(imp
, LUSTRE_IMP_CLOSED
);
1191 ptlrpc_maybe_ping_import_soon(imp
);
1193 CDEBUG(D_HA
, "recovery of %s on %s failed (%d)\n",
1194 obd2cli_tgt(imp
->imp_obd
),
1195 (char *)imp
->imp_connection
->c_remote_uuid
.uuid
, rc
);
1198 wake_up_all(&imp
->imp_recovery_waitq
);
1203 * interpret callback for "completed replay" RPCs.
1204 * \see signal_completed_replay
1206 static int completed_replay_interpret(const struct lu_env
*env
,
1207 struct ptlrpc_request
*req
,
1210 atomic_dec(&req
->rq_import
->imp_replay_inflight
);
1211 if (req
->rq_status
== 0 &&
1212 !req
->rq_import
->imp_vbr_failed
) {
1213 ptlrpc_import_recovery_state_machine(req
->rq_import
);
1215 if (req
->rq_import
->imp_vbr_failed
) {
1217 "%s: version recovery fails, reconnecting\n",
1218 req
->rq_import
->imp_obd
->obd_name
);
1220 CDEBUG(D_HA
, "%s: LAST_REPLAY message error: %d, reconnecting\n",
1221 req
->rq_import
->imp_obd
->obd_name
,
1224 ptlrpc_connect_import(req
->rq_import
);
1231 * Let server know that we have no requests to replay anymore.
1232 * Achieved by just sending a PING request
1234 static int signal_completed_replay(struct obd_import
*imp
)
1236 struct ptlrpc_request
*req
;
1238 if (unlikely(OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_FINISH_REPLAY
)))
1241 LASSERT(atomic_read(&imp
->imp_replay_inflight
) == 0);
1242 atomic_inc(&imp
->imp_replay_inflight
);
1244 req
= ptlrpc_request_alloc_pack(imp
, &RQF_OBD_PING
, LUSTRE_OBD_VERSION
,
1247 atomic_dec(&imp
->imp_replay_inflight
);
1251 ptlrpc_request_set_replen(req
);
1252 req
->rq_send_state
= LUSTRE_IMP_REPLAY_WAIT
;
1253 lustre_msg_add_flags(req
->rq_reqmsg
,
1254 MSG_LOCK_REPLAY_DONE
| MSG_REQ_REPLAY_DONE
);
1256 req
->rq_timeout
*= 3;
1257 req
->rq_interpret_reply
= completed_replay_interpret
;
1259 ptlrpcd_add_req(req
);
1264 * In kernel code all import invalidation happens in its own
1265 * separate thread, so that whatever application happened to encounter
1266 * a problem could still be killed or otherwise continue
1268 static int ptlrpc_invalidate_import_thread(void *data
)
1270 struct obd_import
*imp
= data
;
1272 unshare_fs_struct();
1274 CDEBUG(D_HA
, "thread invalidate import %s to %s@%s\n",
1275 imp
->imp_obd
->obd_name
, obd2cli_tgt(imp
->imp_obd
),
1276 imp
->imp_connection
->c_remote_uuid
.uuid
);
1278 ptlrpc_invalidate_import(imp
);
1280 if (obd_dump_on_eviction
) {
1281 CERROR("dump the log upon eviction\n");
1282 libcfs_debug_dumplog();
1285 IMPORT_SET_STATE(imp
, LUSTRE_IMP_RECOVER
);
1286 ptlrpc_import_recovery_state_machine(imp
);
1288 class_import_put(imp
);
1293 * This is the state machine for client-side recovery on import.
1295 * Typically we have two possibly paths. If we came to server and it is not
1296 * in recovery, we just enter IMP_EVICTED state, invalidate our import
1297 * state and reconnect from scratch.
1298 * If we came to server that is in recovery, we enter IMP_REPLAY import state.
1299 * We go through our list of requests to replay and send them to server one by
1301 * After sending all request from the list we change import state to
1302 * IMP_REPLAY_LOCKS and re-request all the locks we believe we have from server
1303 * and also all the locks we don't yet have and wait for server to grant us.
1304 * After that we send a special "replay completed" request and change import
1305 * state to IMP_REPLAY_WAIT.
1306 * Upon receiving reply to that "replay completed" RPC we enter IMP_RECOVER
1307 * state and resend all requests from sending list.
1308 * After that we promote import to FULL state and send all delayed requests
1309 * and import is fully operational after that.
1312 int ptlrpc_import_recovery_state_machine(struct obd_import
*imp
)
1319 if (imp
->imp_state
== LUSTRE_IMP_EVICTED
) {
1320 deuuidify(obd2cli_tgt(imp
->imp_obd
), NULL
,
1321 &target_start
, &target_len
);
1322 /* Don't care about MGC eviction */
1323 if (strcmp(imp
->imp_obd
->obd_type
->typ_name
,
1324 LUSTRE_MGC_NAME
) != 0) {
1325 LCONSOLE_ERROR_MSG(0x167, "%s: This client was evicted by %.*s; in progress operations using this service will fail.\n",
1326 imp
->imp_obd
->obd_name
, target_len
,
1329 CDEBUG(D_HA
, "evicted from %s@%s; invalidating\n",
1330 obd2cli_tgt(imp
->imp_obd
),
1331 imp
->imp_connection
->c_remote_uuid
.uuid
);
1332 /* reset vbr_failed flag upon eviction */
1333 spin_lock(&imp
->imp_lock
);
1334 imp
->imp_vbr_failed
= 0;
1335 spin_unlock(&imp
->imp_lock
);
1338 struct task_struct
*task
;
1339 /* bug 17802: XXX client_disconnect_export vs connect request
1340 * race. if client will evicted at this time, we start
1341 * invalidate thread without reference to import and import can
1342 * be freed at same time. */
1343 class_import_get(imp
);
1344 task
= kthread_run(ptlrpc_invalidate_import_thread
, imp
,
1347 class_import_put(imp
);
1348 CERROR("error starting invalidate thread: %d\n", rc
);
1357 if (imp
->imp_state
== LUSTRE_IMP_REPLAY
) {
1358 CDEBUG(D_HA
, "replay requested by %s\n",
1359 obd2cli_tgt(imp
->imp_obd
));
1360 rc
= ptlrpc_replay_next(imp
, &inflight
);
1361 if (inflight
== 0 &&
1362 atomic_read(&imp
->imp_replay_inflight
) == 0) {
1363 IMPORT_SET_STATE(imp
, LUSTRE_IMP_REPLAY_LOCKS
);
1364 rc
= ldlm_replay_locks(imp
);
1371 if (imp
->imp_state
== LUSTRE_IMP_REPLAY_LOCKS
) {
1372 if (atomic_read(&imp
->imp_replay_inflight
) == 0) {
1373 IMPORT_SET_STATE(imp
, LUSTRE_IMP_REPLAY_WAIT
);
1374 rc
= signal_completed_replay(imp
);
1381 if (imp
->imp_state
== LUSTRE_IMP_REPLAY_WAIT
) {
1382 if (atomic_read(&imp
->imp_replay_inflight
) == 0) {
1383 IMPORT_SET_STATE(imp
, LUSTRE_IMP_RECOVER
);
1387 if (imp
->imp_state
== LUSTRE_IMP_RECOVER
) {
1388 CDEBUG(D_HA
, "reconnected to %s@%s\n",
1389 obd2cli_tgt(imp
->imp_obd
),
1390 imp
->imp_connection
->c_remote_uuid
.uuid
);
1392 rc
= ptlrpc_resend(imp
);
1395 IMPORT_SET_STATE(imp
, LUSTRE_IMP_FULL
);
1396 ptlrpc_activate_import(imp
);
1398 deuuidify(obd2cli_tgt(imp
->imp_obd
), NULL
,
1399 &target_start
, &target_len
);
1400 LCONSOLE_INFO("%s: Connection restored to %.*s (at %s)\n",
1401 imp
->imp_obd
->obd_name
,
1402 target_len
, target_start
,
1403 libcfs_nid2str(imp
->imp_connection
->c_peer
.nid
));
1406 if (imp
->imp_state
== LUSTRE_IMP_FULL
) {
1407 wake_up_all(&imp
->imp_recovery_waitq
);
1408 ptlrpc_wake_delayed(imp
);
1415 int ptlrpc_disconnect_import(struct obd_import
*imp
, int noclose
)
1417 struct ptlrpc_request
*req
;
1420 if (imp
->imp_obd
->obd_force
)
1423 switch (imp
->imp_connect_op
) {
1425 rq_opc
= OST_DISCONNECT
;
1428 rq_opc
= MDS_DISCONNECT
;
1431 rq_opc
= MGS_DISCONNECT
;
1435 CERROR("%s: don't know how to disconnect from %s (connect_op %d): rc = %d\n",
1436 imp
->imp_obd
->obd_name
, obd2cli_tgt(imp
->imp_obd
),
1437 imp
->imp_connect_op
, rc
);
1441 if (ptlrpc_import_in_recovery(imp
)) {
1442 struct l_wait_info lwi
;
1446 if (imp
->imp_server_timeout
)
1447 timeout
= cfs_time_seconds(obd_timeout
/ 2);
1449 timeout
= cfs_time_seconds(obd_timeout
);
1451 int idx
= import_at_get_index(imp
,
1452 imp
->imp_client
->cli_request_portal
);
1453 timeout
= cfs_time_seconds(
1454 at_get(&imp
->imp_at
.iat_service_estimate
[idx
]));
1457 lwi
= LWI_TIMEOUT_INTR(cfs_timeout_cap(timeout
),
1458 back_to_sleep
, LWI_ON_SIGNAL_NOOP
, NULL
);
1459 rc
= l_wait_event(imp
->imp_recovery_waitq
,
1460 !ptlrpc_import_in_recovery(imp
), &lwi
);
1464 spin_lock(&imp
->imp_lock
);
1465 if (imp
->imp_state
!= LUSTRE_IMP_FULL
)
1467 spin_unlock(&imp
->imp_lock
);
1469 req
= ptlrpc_request_alloc_pack(imp
, &RQF_MDS_DISCONNECT
,
1470 LUSTRE_OBD_VERSION
, rq_opc
);
1472 /* We are disconnecting, do not retry a failed DISCONNECT rpc if
1473 * it fails. We can get through the above with a down server
1474 * if the client doesn't know the server is gone yet. */
1475 req
->rq_no_resend
= 1;
1477 /* We want client umounts to happen quickly, no matter the
1479 req
->rq_timeout
= min_t(int, req
->rq_timeout
,
1480 INITIAL_CONNECT_TIMEOUT
);
1482 IMPORT_SET_STATE(imp
, LUSTRE_IMP_CONNECTING
);
1483 req
->rq_send_state
= LUSTRE_IMP_CONNECTING
;
1484 ptlrpc_request_set_replen(req
);
1485 rc
= ptlrpc_queue_wait(req
);
1486 ptlrpc_req_finished(req
);
1490 spin_lock(&imp
->imp_lock
);
1493 IMPORT_SET_STATE_NOLOCK(imp
, LUSTRE_IMP_DISCON
);
1495 IMPORT_SET_STATE_NOLOCK(imp
, LUSTRE_IMP_CLOSED
);
1496 memset(&imp
->imp_remote_handle
, 0, sizeof(imp
->imp_remote_handle
));
1497 spin_unlock(&imp
->imp_lock
);
1499 if (rc
== -ETIMEDOUT
|| rc
== -ENOTCONN
|| rc
== -ESHUTDOWN
)
1504 EXPORT_SYMBOL(ptlrpc_disconnect_import
);
1506 void ptlrpc_cleanup_imp(struct obd_import
*imp
)
1508 spin_lock(&imp
->imp_lock
);
1509 IMPORT_SET_STATE_NOLOCK(imp
, LUSTRE_IMP_CLOSED
);
1510 imp
->imp_generation
++;
1511 spin_unlock(&imp
->imp_lock
);
1512 ptlrpc_abort_inflight(imp
);
1514 EXPORT_SYMBOL(ptlrpc_cleanup_imp
);
1516 /* Adaptive Timeout utils */
1517 extern unsigned int at_min
, at_max
, at_history
;
1519 /* Bin into timeslices using AT_BINS bins.
1520 This gives us a max of the last binlimit*AT_BINS secs without the storage,
1521 but still smoothing out a return to normalcy from a slow response.
1522 (E.g. remember the maximum latency in each minute of the last 4 minutes.) */
1523 int at_measured(struct adaptive_timeout
*at
, unsigned int val
)
1525 unsigned int old
= at
->at_current
;
1526 time_t now
= get_seconds();
1527 time_t binlimit
= max_t(time_t, at_history
/ AT_BINS
, 1);
1530 CDEBUG(D_OTHER
, "add %u to %p time=%lu v=%u (%u %u %u %u)\n",
1531 val
, at
, now
- at
->at_binstart
, at
->at_current
,
1532 at
->at_hist
[0], at
->at_hist
[1], at
->at_hist
[2], at
->at_hist
[3]);
1535 /* 0's don't count, because we never want our timeout to
1536 drop to 0, and because 0 could mean an error */
1539 spin_lock(&at
->at_lock
);
1541 if (unlikely(at
->at_binstart
== 0)) {
1542 /* Special case to remove default from history */
1543 at
->at_current
= val
;
1544 at
->at_worst_ever
= val
;
1545 at
->at_worst_time
= now
;
1546 at
->at_hist
[0] = val
;
1547 at
->at_binstart
= now
;
1548 } else if (now
- at
->at_binstart
< binlimit
) {
1550 at
->at_hist
[0] = max(val
, at
->at_hist
[0]);
1551 at
->at_current
= max(val
, at
->at_current
);
1554 unsigned int maxv
= val
;
1555 /* move bins over */
1556 shift
= (now
- at
->at_binstart
) / binlimit
;
1558 for (i
= AT_BINS
- 1; i
>= 0; i
--) {
1560 at
->at_hist
[i
] = at
->at_hist
[i
- shift
];
1561 maxv
= max(maxv
, at
->at_hist
[i
]);
1566 at
->at_hist
[0] = val
;
1567 at
->at_current
= maxv
;
1568 at
->at_binstart
+= shift
* binlimit
;
1571 if (at
->at_current
> at
->at_worst_ever
) {
1572 at
->at_worst_ever
= at
->at_current
;
1573 at
->at_worst_time
= now
;
1576 if (at
->at_flags
& AT_FLG_NOHIST
)
1577 /* Only keep last reported val; keeping the rest of the history
1579 at
->at_current
= val
;
1582 at
->at_current
= min(at
->at_current
, at_max
);
1583 at
->at_current
= max(at
->at_current
, at_min
);
1585 if (at
->at_current
!= old
)
1586 CDEBUG(D_OTHER
, "AT %p change: old=%u new=%u delta=%d (val=%u) hist %u %u %u %u\n",
1588 old
, at
->at_current
, at
->at_current
- old
, val
,
1589 at
->at_hist
[0], at
->at_hist
[1], at
->at_hist
[2],
1592 /* if we changed, report the old value */
1593 old
= (at
->at_current
!= old
) ? old
: 0;
1595 spin_unlock(&at
->at_lock
);
1599 /* Find the imp_at index for a given portal; assign if space available */
1600 int import_at_get_index(struct obd_import
*imp
, int portal
)
1602 struct imp_at
*at
= &imp
->imp_at
;
1605 for (i
= 0; i
< IMP_AT_MAX_PORTALS
; i
++) {
1606 if (at
->iat_portal
[i
] == portal
)
1608 if (at
->iat_portal
[i
] == 0)
1613 /* Not found in list, add it under a lock */
1614 spin_lock(&imp
->imp_lock
);
1616 /* Check unused under lock */
1617 for (; i
< IMP_AT_MAX_PORTALS
; i
++) {
1618 if (at
->iat_portal
[i
] == portal
)
1620 if (at
->iat_portal
[i
] == 0)
1625 /* Not enough portals? */
1626 LASSERT(i
< IMP_AT_MAX_PORTALS
);
1628 at
->iat_portal
[i
] = portal
;
1630 spin_unlock(&imp
->imp_lock
);