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
)
66 imp
->imp_state
= state
;
67 imp
->imp_state_hist
[imp
->imp_state_hist_idx
].ish_state
= state
;
68 imp
->imp_state_hist
[imp
->imp_state_hist_idx
].ish_time
=
69 cfs_time_current_sec();
70 imp
->imp_state_hist_idx
= (imp
->imp_state_hist_idx
+ 1) %
74 /* A CLOSED import should remain so. */
75 #define IMPORT_SET_STATE_NOLOCK(imp, state) \
77 if (imp->imp_state != LUSTRE_IMP_CLOSED) { \
78 CDEBUG(D_HA, "%p %s: changing import state from %s to %s\n", \
79 imp, obd2cli_tgt(imp->imp_obd), \
80 ptlrpc_import_state_name(imp->imp_state), \
81 ptlrpc_import_state_name(state)); \
82 __import_set_state(imp, state); \
86 #define IMPORT_SET_STATE(imp, state) \
88 spin_lock(&imp->imp_lock); \
89 IMPORT_SET_STATE_NOLOCK(imp, state); \
90 spin_unlock(&imp->imp_lock); \
94 static int ptlrpc_connect_interpret(const struct lu_env
*env
,
95 struct ptlrpc_request
*request
,
97 int ptlrpc_import_recovery_state_machine(struct obd_import
*imp
);
99 /* Only this function is allowed to change the import state when it is
100 * CLOSED. I would rather refcount the import and free it after
101 * disconnection like we do with exports. To do that, the client_obd
102 * will need to save the peer info somewhere other than in the import,
104 int ptlrpc_init_import(struct obd_import
*imp
)
106 spin_lock(&imp
->imp_lock
);
108 imp
->imp_generation
++;
109 imp
->imp_state
= LUSTRE_IMP_NEW
;
111 spin_unlock(&imp
->imp_lock
);
115 EXPORT_SYMBOL(ptlrpc_init_import
);
117 #define UUID_STR "_UUID"
118 void deuuidify(char *uuid
, const char *prefix
, char **uuid_start
, int *uuid_len
)
120 *uuid_start
= !prefix
|| strncmp(uuid
, prefix
, strlen(prefix
))
121 ? uuid
: uuid
+ strlen(prefix
);
123 *uuid_len
= strlen(*uuid_start
);
125 if (*uuid_len
< strlen(UUID_STR
))
128 if (!strncmp(*uuid_start
+ *uuid_len
- strlen(UUID_STR
),
129 UUID_STR
, strlen(UUID_STR
)))
130 *uuid_len
-= strlen(UUID_STR
);
132 EXPORT_SYMBOL(deuuidify
);
135 * Returns true if import was FULL, false if import was already not
137 * @imp - import to be disconnected
138 * @conn_cnt - connection count (epoch) of the request that timed out
139 * and caused the disconnection. In some cases, multiple
140 * inflight requests can fail to a single target (e.g. OST
141 * bulk requests) and if one has already caused a reconnection
142 * (increasing the import->conn_cnt) the older failure should
143 * not also cause a reconnection. If zero it forces a reconnect.
145 int ptlrpc_set_import_discon(struct obd_import
*imp
, __u32 conn_cnt
)
149 spin_lock(&imp
->imp_lock
);
151 if (imp
->imp_state
== LUSTRE_IMP_FULL
&&
152 (conn_cnt
== 0 || conn_cnt
== imp
->imp_conn_cnt
)) {
156 deuuidify(obd2cli_tgt(imp
->imp_obd
), NULL
,
157 &target_start
, &target_len
);
159 if (imp
->imp_replayable
) {
160 LCONSOLE_WARN("%s: Connection to %.*s (at %s) was "
161 "lost; in progress operations using this "
162 "service will wait for recovery to complete\n",
163 imp
->imp_obd
->obd_name
, target_len
, target_start
,
164 libcfs_nid2str(imp
->imp_connection
->c_peer
.nid
));
166 LCONSOLE_ERROR_MSG(0x166, "%s: Connection to "
167 "%.*s (at %s) was lost; in progress "
168 "operations using this service will fail\n",
169 imp
->imp_obd
->obd_name
,
170 target_len
, target_start
,
171 libcfs_nid2str(imp
->imp_connection
->c_peer
.nid
));
173 IMPORT_SET_STATE_NOLOCK(imp
, LUSTRE_IMP_DISCON
);
174 spin_unlock(&imp
->imp_lock
);
176 if (obd_dump_on_timeout
)
177 libcfs_debug_dumplog();
179 obd_import_event(imp
->imp_obd
, imp
, IMP_EVENT_DISCON
);
182 spin_unlock(&imp
->imp_lock
);
183 CDEBUG(D_HA
, "%s: import %p already %s (conn %u, was %u): %s\n",
184 imp
->imp_client
->cli_name
, imp
,
185 (imp
->imp_state
== LUSTRE_IMP_FULL
&&
186 imp
->imp_conn_cnt
> conn_cnt
) ?
187 "reconnected" : "not connected", imp
->imp_conn_cnt
,
188 conn_cnt
, ptlrpc_import_state_name(imp
->imp_state
));
194 /* Must be called with imp_lock held! */
195 static void ptlrpc_deactivate_and_unlock_import(struct obd_import
*imp
)
197 assert_spin_locked(&imp
->imp_lock
);
199 CDEBUG(D_HA
, "setting import %s INVALID\n", obd2cli_tgt(imp
->imp_obd
));
200 imp
->imp_invalid
= 1;
201 imp
->imp_generation
++;
202 spin_unlock(&imp
->imp_lock
);
204 ptlrpc_abort_inflight(imp
);
205 obd_import_event(imp
->imp_obd
, imp
, IMP_EVENT_INACTIVE
);
209 * This acts as a barrier; all existing requests are rejected, and
210 * no new requests will be accepted until the import is valid again.
212 void ptlrpc_deactivate_import(struct obd_import
*imp
)
214 spin_lock(&imp
->imp_lock
);
215 ptlrpc_deactivate_and_unlock_import(imp
);
217 EXPORT_SYMBOL(ptlrpc_deactivate_import
);
220 ptlrpc_inflight_deadline(struct ptlrpc_request
*req
, time_t now
)
224 if (!(((req
->rq_phase
== RQ_PHASE_RPC
) && !req
->rq_waiting
) ||
225 (req
->rq_phase
== RQ_PHASE_BULK
) ||
226 (req
->rq_phase
== RQ_PHASE_NEW
)))
229 if (req
->rq_timedout
)
232 if (req
->rq_phase
== RQ_PHASE_NEW
)
235 dl
= req
->rq_deadline
;
243 static unsigned int ptlrpc_inflight_timeout(struct obd_import
*imp
)
245 time_t now
= cfs_time_current_sec();
246 struct list_head
*tmp
, *n
;
247 struct ptlrpc_request
*req
;
248 unsigned int timeout
= 0;
250 spin_lock(&imp
->imp_lock
);
251 list_for_each_safe(tmp
, n
, &imp
->imp_sending_list
) {
252 req
= list_entry(tmp
, struct ptlrpc_request
, rq_list
);
253 timeout
= max(ptlrpc_inflight_deadline(req
, now
), timeout
);
255 spin_unlock(&imp
->imp_lock
);
260 * This function will invalidate the import, if necessary, then block
261 * for all the RPC completions, and finally notify the obd to
262 * invalidate its state (ie cancel locks, clear pending requests,
265 void ptlrpc_invalidate_import(struct obd_import
*imp
)
267 struct list_head
*tmp
, *n
;
268 struct ptlrpc_request
*req
;
269 struct l_wait_info lwi
;
270 unsigned int timeout
;
273 atomic_inc(&imp
->imp_inval_count
);
275 if (!imp
->imp_invalid
|| imp
->imp_obd
->obd_no_recov
)
276 ptlrpc_deactivate_import(imp
);
278 CFS_FAIL_TIMEOUT(OBD_FAIL_MGS_CONNECT_NET
, 3 * cfs_fail_val
/ 2);
279 LASSERT(imp
->imp_invalid
);
281 /* Wait forever until inflight == 0. We really can't do it another
282 * way because in some cases we need to wait for very long reply
283 * unlink. We can't do anything before that because there is really
284 * no guarantee that some rdma transfer is not in progress right now. */
286 /* Calculate max timeout for waiting on rpcs to error
287 * out. Use obd_timeout if calculated value is smaller
289 if (!OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_REPL_UNLINK
)) {
290 timeout
= ptlrpc_inflight_timeout(imp
);
291 timeout
+= timeout
/ 3;
294 timeout
= obd_timeout
;
296 /* decrease the interval to increase race condition */
300 CDEBUG(D_RPCTRACE
,"Sleeping %d sec for inflight to error out\n",
303 /* Wait for all requests to error out and call completion
304 * callbacks. Cap it at obd_timeout -- these should all
305 * have been locally cancelled by ptlrpc_abort_inflight. */
306 lwi
= LWI_TIMEOUT_INTERVAL(
307 cfs_timeout_cap(cfs_time_seconds(timeout
)),
308 (timeout
> 1)?cfs_time_seconds(1):cfs_time_seconds(1)/2,
310 rc
= l_wait_event(imp
->imp_recovery_waitq
,
311 (atomic_read(&imp
->imp_inflight
) == 0),
314 const char *cli_tgt
= obd2cli_tgt(imp
->imp_obd
);
316 CERROR("%s: rc = %d waiting for callback (%d != 0)\n",
318 atomic_read(&imp
->imp_inflight
));
320 spin_lock(&imp
->imp_lock
);
321 if (atomic_read(&imp
->imp_inflight
) == 0) {
322 int count
= atomic_read(&imp
->imp_unregistering
);
324 /* We know that "unregistering" rpcs only can
325 * survive in sending or delaying lists (they
326 * maybe waiting for long reply unlink in
327 * sluggish nets). Let's check this. If there
328 * is no inflight and unregistering != 0, this
330 LASSERTF(count
== 0, "Some RPCs are still "
331 "unregistering: %d\n", count
);
333 /* Let's save one loop as soon as inflight have
334 * dropped to zero. No new inflights possible at
338 list_for_each_safe(tmp
, n
,
339 &imp
->imp_sending_list
) {
340 req
= list_entry(tmp
,
341 struct ptlrpc_request
,
343 DEBUG_REQ(D_ERROR
, req
,
344 "still on sending list");
346 list_for_each_safe(tmp
, n
,
347 &imp
->imp_delayed_list
) {
348 req
= list_entry(tmp
,
349 struct ptlrpc_request
,
351 DEBUG_REQ(D_ERROR
, req
,
352 "still on delayed list");
355 CERROR("%s: RPCs in \"%s\" phase found (%d). "
356 "Network is sluggish? Waiting them "
357 "to error out.\n", cli_tgt
,
358 ptlrpc_phase2str(RQ_PHASE_UNREGISTERING
),
362 spin_unlock(&imp
->imp_lock
);
367 * Let's additionally check that no new rpcs added to import in
368 * "invalidate" state.
370 LASSERT(atomic_read(&imp
->imp_inflight
) == 0);
371 obd_import_event(imp
->imp_obd
, imp
, IMP_EVENT_INVALIDATE
);
372 sptlrpc_import_flush_all_ctx(imp
);
374 atomic_dec(&imp
->imp_inval_count
);
375 wake_up_all(&imp
->imp_recovery_waitq
);
377 EXPORT_SYMBOL(ptlrpc_invalidate_import
);
379 /* unset imp_invalid */
380 void ptlrpc_activate_import(struct obd_import
*imp
)
382 struct obd_device
*obd
= imp
->imp_obd
;
384 spin_lock(&imp
->imp_lock
);
385 if (imp
->imp_deactive
!= 0) {
386 spin_unlock(&imp
->imp_lock
);
390 imp
->imp_invalid
= 0;
391 spin_unlock(&imp
->imp_lock
);
392 obd_import_event(obd
, imp
, IMP_EVENT_ACTIVE
);
394 EXPORT_SYMBOL(ptlrpc_activate_import
);
396 static void ptlrpc_pinger_force(struct obd_import
*imp
)
398 CDEBUG(D_HA
, "%s: waking up pinger s:%s\n", obd2cli_tgt(imp
->imp_obd
),
399 ptlrpc_import_state_name(imp
->imp_state
));
401 spin_lock(&imp
->imp_lock
);
402 imp
->imp_force_verify
= 1;
403 spin_unlock(&imp
->imp_lock
);
405 if (imp
->imp_state
!= LUSTRE_IMP_CONNECTING
)
406 ptlrpc_pinger_wake_up();
409 void ptlrpc_fail_import(struct obd_import
*imp
, __u32 conn_cnt
)
411 LASSERT(!imp
->imp_dlm_fake
);
413 if (ptlrpc_set_import_discon(imp
, conn_cnt
)) {
414 if (!imp
->imp_replayable
) {
415 CDEBUG(D_HA
, "import %s@%s for %s not replayable, "
416 "auto-deactivating\n",
417 obd2cli_tgt(imp
->imp_obd
),
418 imp
->imp_connection
->c_remote_uuid
.uuid
,
419 imp
->imp_obd
->obd_name
);
420 ptlrpc_deactivate_import(imp
);
423 ptlrpc_pinger_force(imp
);
426 EXPORT_SYMBOL(ptlrpc_fail_import
);
428 int ptlrpc_reconnect_import(struct obd_import
*imp
)
431 struct l_wait_info lwi
;
432 int secs
= cfs_time_seconds(obd_timeout
);
435 ptlrpc_pinger_force(imp
);
437 CDEBUG(D_HA
, "%s: recovery started, waiting %u seconds\n",
438 obd2cli_tgt(imp
->imp_obd
), secs
);
440 lwi
= LWI_TIMEOUT(secs
, NULL
, NULL
);
441 rc
= l_wait_event(imp
->imp_recovery_waitq
,
442 !ptlrpc_import_in_recovery(imp
), &lwi
);
443 CDEBUG(D_HA
, "%s: recovery finished s:%s\n", obd2cli_tgt(imp
->imp_obd
),
444 ptlrpc_import_state_name(imp
->imp_state
));
447 ptlrpc_set_import_discon(imp
, 0);
448 /* Force a new connect attempt */
449 ptlrpc_invalidate_import(imp
);
450 /* Do a fresh connect next time by zeroing the handle */
451 ptlrpc_disconnect_import(imp
, 1);
452 /* Wait for all invalidate calls to finish */
453 if (atomic_read(&imp
->imp_inval_count
) > 0) {
455 struct l_wait_info lwi
= LWI_INTR(LWI_ON_SIGNAL_NOOP
, NULL
);
456 rc
= l_wait_event(imp
->imp_recovery_waitq
,
457 (atomic_read(&imp
->imp_inval_count
) == 0),
460 CERROR("Interrupted, inval=%d\n",
461 atomic_read(&imp
->imp_inval_count
));
464 /* Allow reconnect attempts */
465 imp
->imp_obd
->obd_no_recov
= 0;
466 /* Remove 'invalid' flag */
467 ptlrpc_activate_import(imp
);
468 /* Attempt a new connect */
469 ptlrpc_recover_import(imp
, NULL
, 0);
473 EXPORT_SYMBOL(ptlrpc_reconnect_import
);
476 * Connection on import \a imp is changed to another one (if more than one is
477 * present). We typically chose connection that we have not tried to connect to
480 static int import_select_connection(struct obd_import
*imp
)
482 struct obd_import_conn
*imp_conn
= NULL
, *conn
;
483 struct obd_export
*dlmexp
;
485 int target_len
, tried_all
= 1;
487 spin_lock(&imp
->imp_lock
);
489 if (list_empty(&imp
->imp_conn_list
)) {
490 CERROR("%s: no connections available\n",
491 imp
->imp_obd
->obd_name
);
492 spin_unlock(&imp
->imp_lock
);
496 list_for_each_entry(conn
, &imp
->imp_conn_list
, oic_item
) {
497 CDEBUG(D_HA
, "%s: connect to NID %s last attempt "LPU64
"\n",
498 imp
->imp_obd
->obd_name
,
499 libcfs_nid2str(conn
->oic_conn
->c_peer
.nid
),
500 conn
->oic_last_attempt
);
502 /* If we have not tried this connection since
503 the last successful attempt, go with this one */
504 if ((conn
->oic_last_attempt
== 0) ||
505 cfs_time_beforeq_64(conn
->oic_last_attempt
,
506 imp
->imp_last_success_conn
)) {
512 /* If all of the connections have already been tried
513 since the last successful connection; just choose the
514 least recently used */
517 else if (cfs_time_before_64(conn
->oic_last_attempt
,
518 imp_conn
->oic_last_attempt
))
522 /* if not found, simply choose the current one */
523 if (!imp_conn
|| imp
->imp_force_reconnect
) {
524 LASSERT(imp
->imp_conn_current
);
525 imp_conn
= imp
->imp_conn_current
;
528 LASSERT(imp_conn
->oic_conn
);
530 /* If we've tried everything, and we're back to the beginning of the
531 list, increase our timeout and try again. It will be reset when
532 we do finally connect. (FIXME: really we should wait for all network
533 state associated with the last connection attempt to drain before
534 trying to reconnect on it.) */
535 if (tried_all
&& (imp
->imp_conn_list
.next
== &imp_conn
->oic_item
)) {
536 struct adaptive_timeout
*at
= &imp
->imp_at
.iat_net_latency
;
537 if (at_get(at
) < CONNECTION_SWITCH_MAX
) {
538 at_measured(at
, at_get(at
) + CONNECTION_SWITCH_INC
);
539 if (at_get(at
) > CONNECTION_SWITCH_MAX
)
540 at_reset(at
, CONNECTION_SWITCH_MAX
);
542 LASSERT(imp_conn
->oic_last_attempt
);
543 CDEBUG(D_HA
, "%s: tried all connections, increasing latency "
544 "to %ds\n", imp
->imp_obd
->obd_name
, at_get(at
));
547 imp_conn
->oic_last_attempt
= cfs_time_current_64();
549 /* switch connection, don't mind if it's same as the current one */
550 if (imp
->imp_connection
)
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 if (dlmexp
->exp_connection
)
557 ptlrpc_connection_put(dlmexp
->exp_connection
);
558 dlmexp
->exp_connection
= ptlrpc_connection_addref(imp_conn
->oic_conn
);
559 class_export_put(dlmexp
);
561 if (imp
->imp_conn_current
!= imp_conn
) {
562 if (imp
->imp_conn_current
) {
563 deuuidify(obd2cli_tgt(imp
->imp_obd
), NULL
,
564 &target_start
, &target_len
);
566 CDEBUG(D_HA
, "%s: Connection changing to"
568 imp
->imp_obd
->obd_name
,
569 target_len
, target_start
,
570 libcfs_nid2str(imp_conn
->oic_conn
->c_peer
.nid
));
573 imp
->imp_conn_current
= imp_conn
;
576 CDEBUG(D_HA
, "%s: import %p using connection %s/%s\n",
577 imp
->imp_obd
->obd_name
, imp
, imp_conn
->oic_uuid
.uuid
,
578 libcfs_nid2str(imp_conn
->oic_conn
->c_peer
.nid
));
580 spin_unlock(&imp
->imp_lock
);
586 * must be called under imp_lock
588 static int ptlrpc_first_transno(struct obd_import
*imp
, __u64
*transno
)
590 struct ptlrpc_request
*req
;
591 struct list_head
*tmp
;
593 /* The requests in committed_list always have smaller transnos than
594 * the requests in replay_list */
595 if (!list_empty(&imp
->imp_committed_list
)) {
596 tmp
= imp
->imp_committed_list
.next
;
597 req
= list_entry(tmp
, struct ptlrpc_request
, rq_replay_list
);
598 *transno
= req
->rq_transno
;
599 if (req
->rq_transno
== 0) {
600 DEBUG_REQ(D_ERROR
, req
,
601 "zero transno in committed_list");
606 if (!list_empty(&imp
->imp_replay_list
)) {
607 tmp
= imp
->imp_replay_list
.next
;
608 req
= list_entry(tmp
, struct ptlrpc_request
, rq_replay_list
);
609 *transno
= req
->rq_transno
;
610 if (req
->rq_transno
== 0) {
611 DEBUG_REQ(D_ERROR
, req
, "zero transno in replay_list");
620 * Attempt to (re)connect import \a imp. This includes all preparations,
621 * initializing CONNECT RPC request and passing it to ptlrpcd for
623 * Returns 0 on success or error code.
625 int ptlrpc_connect_import(struct obd_import
*imp
)
627 struct obd_device
*obd
= imp
->imp_obd
;
628 int initial_connect
= 0;
630 __u64 committed_before_reconnect
= 0;
631 struct ptlrpc_request
*request
;
632 char *bufs
[] = { NULL
,
633 obd2cli_tgt(imp
->imp_obd
),
635 (char *)&imp
->imp_dlm_handle
,
636 (char *)&imp
->imp_connect_data
};
637 struct ptlrpc_connect_async_args
*aa
;
640 spin_lock(&imp
->imp_lock
);
641 if (imp
->imp_state
== LUSTRE_IMP_CLOSED
) {
642 spin_unlock(&imp
->imp_lock
);
643 CERROR("can't connect to a closed import\n");
645 } else if (imp
->imp_state
== LUSTRE_IMP_FULL
) {
646 spin_unlock(&imp
->imp_lock
);
647 CERROR("already connected\n");
649 } else if (imp
->imp_state
== LUSTRE_IMP_CONNECTING
) {
650 spin_unlock(&imp
->imp_lock
);
651 CERROR("already connecting\n");
655 IMPORT_SET_STATE_NOLOCK(imp
, LUSTRE_IMP_CONNECTING
);
658 imp
->imp_resend_replay
= 0;
660 if (!lustre_handle_is_used(&imp
->imp_remote_handle
))
663 committed_before_reconnect
= imp
->imp_peer_committed_transno
;
665 set_transno
= ptlrpc_first_transno(imp
,
666 &imp
->imp_connect_data
.ocd_transno
);
667 spin_unlock(&imp
->imp_lock
);
669 rc
= import_select_connection(imp
);
673 rc
= sptlrpc_import_sec_adapt(imp
, NULL
, NULL
);
677 /* Reset connect flags to the originally requested flags, in case
678 * the server is updated on-the-fly we will get the new features. */
679 imp
->imp_connect_data
.ocd_connect_flags
= imp
->imp_connect_flags_orig
;
680 /* Reset ocd_version each time so the server knows the exact versions */
681 imp
->imp_connect_data
.ocd_version
= LUSTRE_VERSION_CODE
;
682 imp
->imp_msghdr_flags
&= ~MSGHDR_AT_SUPPORT
;
683 imp
->imp_msghdr_flags
&= ~MSGHDR_CKSUM_INCOMPAT18
;
685 rc
= obd_reconnect(NULL
, imp
->imp_obd
->obd_self_export
, obd
,
686 &obd
->obd_uuid
, &imp
->imp_connect_data
, NULL
);
690 request
= ptlrpc_request_alloc(imp
, &RQF_MDS_CONNECT
);
692 GOTO(out
, rc
= -ENOMEM
);
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
, PDL_POLICY_ROUND
, -1);
749 IMPORT_SET_STATE(imp
, LUSTRE_IMP_DISCON
);
754 EXPORT_SYMBOL(ptlrpc_connect_import
);
756 static void ptlrpc_maybe_ping_import_soon(struct obd_import
*imp
)
760 spin_lock(&imp
->imp_lock
);
761 force_verify
= imp
->imp_force_verify
!= 0;
762 spin_unlock(&imp
->imp_lock
);
765 ptlrpc_pinger_wake_up();
768 static int ptlrpc_busy_reconnect(int rc
)
770 return (rc
== -EBUSY
) || (rc
== -EAGAIN
);
774 * interpret_reply callback for connect RPCs.
775 * Looks into returned status of connect operation and decides
776 * what to do with the import - i.e enter recovery, promote it to
777 * full state for normal operations of disconnect it due to an error.
779 static int ptlrpc_connect_interpret(const struct lu_env
*env
,
780 struct ptlrpc_request
*request
,
783 struct ptlrpc_connect_async_args
*aa
= data
;
784 struct obd_import
*imp
= request
->rq_import
;
785 struct client_obd
*cli
= &imp
->imp_obd
->u
.cli
;
786 struct lustre_handle old_hdl
;
787 __u64 old_connect_flags
;
789 struct obd_connect_data
*ocd
;
790 struct obd_export
*exp
;
793 spin_lock(&imp
->imp_lock
);
794 if (imp
->imp_state
== LUSTRE_IMP_CLOSED
) {
795 imp
->imp_connect_tried
= 1;
796 spin_unlock(&imp
->imp_lock
);
801 /* if this reconnect to busy export - not need select new target
803 imp
->imp_force_reconnect
= ptlrpc_busy_reconnect(rc
);
804 spin_unlock(&imp
->imp_lock
);
805 ptlrpc_maybe_ping_import_soon(imp
);
808 spin_unlock(&imp
->imp_lock
);
810 LASSERT(imp
->imp_conn_current
);
812 msg_flags
= lustre_msg_get_op_flags(request
->rq_repmsg
);
814 ret
= req_capsule_get_size(&request
->rq_pill
, &RMF_CONNECT_DATA
,
816 /* server replied obd_connect_data is always bigger */
817 ocd
= req_capsule_server_sized_get(&request
->rq_pill
,
818 &RMF_CONNECT_DATA
, ret
);
821 CERROR("%s: no connect data from server\n",
822 imp
->imp_obd
->obd_name
);
827 spin_lock(&imp
->imp_lock
);
829 /* All imports are pingable */
830 imp
->imp_pingable
= 1;
831 imp
->imp_force_reconnect
= 0;
832 imp
->imp_force_verify
= 0;
834 imp
->imp_connect_data
= *ocd
;
836 CDEBUG(D_HA
, "%s: connect to target with instance %u\n",
837 imp
->imp_obd
->obd_name
, ocd
->ocd_instance
);
838 exp
= class_conn2export(&imp
->imp_dlm_handle
);
840 spin_unlock(&imp
->imp_lock
);
842 /* check that server granted subset of flags we asked for. */
843 if ((ocd
->ocd_connect_flags
& imp
->imp_connect_flags_orig
) !=
844 ocd
->ocd_connect_flags
) {
845 CERROR("%s: Server didn't granted asked subset of flags: "
846 "asked="LPX64
" grranted="LPX64
"\n",
847 imp
->imp_obd
->obd_name
,imp
->imp_connect_flags_orig
,
848 ocd
->ocd_connect_flags
);
849 GOTO(out
, rc
= -EPROTO
);
853 /* This could happen if export is cleaned during the
855 CERROR("%s: missing export after connect\n",
856 imp
->imp_obd
->obd_name
);
857 GOTO(out
, rc
= -ENODEV
);
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
);
896 GOTO(finish
, rc
= 0);
899 /* Determine what recovery state to move the import to. */
900 if (MSG_CONNECT_RECONNECT
& msg_flags
) {
901 memset(&old_hdl
, 0, sizeof(old_hdl
));
902 if (!memcmp(&old_hdl
, lustre_msg_get_handle(request
->rq_repmsg
),
904 LCONSOLE_WARN("Reconnect to %s (at @%s) failed due "
905 "bad handle "LPX64
"\n",
906 obd2cli_tgt(imp
->imp_obd
),
907 imp
->imp_connection
->c_remote_uuid
.uuid
,
908 imp
->imp_dlm_handle
.cookie
);
909 GOTO(out
, rc
= -ENOTCONN
);
912 if (memcmp(&imp
->imp_remote_handle
,
913 lustre_msg_get_handle(request
->rq_repmsg
),
914 sizeof(imp
->imp_remote_handle
))) {
915 int level
= msg_flags
& MSG_CONNECT_RECOVERING
?
918 /* Bug 16611/14775: if server handle have changed,
919 * that means some sort of disconnection happened.
920 * If the server is not in recovery, that also means it
921 * already erased all of our state because of previous
922 * eviction. If it is in recovery - we are safe to
923 * participate since we can reestablish all of our state
924 * with server again */
925 if ((MSG_CONNECT_RECOVERING
& msg_flags
)) {
926 CDEBUG(level
,"%s@%s changed server handle from "
928 " but is still in recovery\n",
929 obd2cli_tgt(imp
->imp_obd
),
930 imp
->imp_connection
->c_remote_uuid
.uuid
,
931 imp
->imp_remote_handle
.cookie
,
932 lustre_msg_get_handle(
933 request
->rq_repmsg
)->cookie
);
935 LCONSOLE_WARN("Evicted from %s (at %s) "
936 "after server handle changed from "
937 LPX64
" to "LPX64
"\n",
938 obd2cli_tgt(imp
->imp_obd
),
939 imp
->imp_connection
-> \
941 imp
->imp_remote_handle
.cookie
,
942 lustre_msg_get_handle(
943 request
->rq_repmsg
)->cookie
);
947 imp
->imp_remote_handle
=
948 *lustre_msg_get_handle(request
->rq_repmsg
);
950 if (!(MSG_CONNECT_RECOVERING
& msg_flags
)) {
951 IMPORT_SET_STATE(imp
, LUSTRE_IMP_EVICTED
);
952 GOTO(finish
, rc
= 0);
956 CDEBUG(D_HA
, "reconnected to %s@%s after partition\n",
957 obd2cli_tgt(imp
->imp_obd
),
958 imp
->imp_connection
->c_remote_uuid
.uuid
);
961 if (imp
->imp_invalid
) {
962 CDEBUG(D_HA
, "%s: reconnected but import is invalid; "
963 "marking evicted\n", imp
->imp_obd
->obd_name
);
964 IMPORT_SET_STATE(imp
, LUSTRE_IMP_EVICTED
);
965 } else if (MSG_CONNECT_RECOVERING
& msg_flags
) {
966 CDEBUG(D_HA
, "%s: reconnected to %s during replay\n",
967 imp
->imp_obd
->obd_name
,
968 obd2cli_tgt(imp
->imp_obd
));
970 spin_lock(&imp
->imp_lock
);
971 imp
->imp_resend_replay
= 1;
972 spin_unlock(&imp
->imp_lock
);
974 IMPORT_SET_STATE(imp
, LUSTRE_IMP_REPLAY
);
976 IMPORT_SET_STATE(imp
, LUSTRE_IMP_RECOVER
);
978 } else if ((MSG_CONNECT_RECOVERING
& msg_flags
) && !imp
->imp_invalid
) {
979 LASSERT(imp
->imp_replayable
);
980 imp
->imp_remote_handle
=
981 *lustre_msg_get_handle(request
->rq_repmsg
);
982 imp
->imp_last_replay_transno
= 0;
983 IMPORT_SET_STATE(imp
, LUSTRE_IMP_REPLAY
);
985 DEBUG_REQ(D_HA
, request
, "%s: evicting (reconnect/recover flags"
986 " not set: %x)", imp
->imp_obd
->obd_name
, msg_flags
);
987 imp
->imp_remote_handle
=
988 *lustre_msg_get_handle(request
->rq_repmsg
);
989 IMPORT_SET_STATE(imp
, LUSTRE_IMP_EVICTED
);
992 /* Sanity checks for a reconnected import. */
993 if (!(imp
->imp_replayable
) != !(msg_flags
& MSG_CONNECT_REPLAYABLE
)) {
994 CERROR("imp_replayable flag does not match server "
995 "after reconnect. We should LBUG right here.\n");
998 if (lustre_msg_get_last_committed(request
->rq_repmsg
) > 0 &&
999 lustre_msg_get_last_committed(request
->rq_repmsg
) <
1000 aa
->pcaa_peer_committed
) {
1001 CERROR("%s went back in time (transno "LPD64
1002 " was previously committed, server now claims "LPD64
1003 ")! See https://bugzilla.lustre.org/show_bug.cgi?"
1005 obd2cli_tgt(imp
->imp_obd
), aa
->pcaa_peer_committed
,
1006 lustre_msg_get_last_committed(request
->rq_repmsg
));
1010 rc
= ptlrpc_import_recovery_state_machine(imp
);
1012 if (rc
== -ENOTCONN
) {
1013 CDEBUG(D_HA
, "evicted/aborted by %s@%s during recovery;"
1014 "invalidating and reconnecting\n",
1015 obd2cli_tgt(imp
->imp_obd
),
1016 imp
->imp_connection
->c_remote_uuid
.uuid
);
1017 ptlrpc_connect_import(imp
);
1018 imp
->imp_connect_tried
= 1;
1023 spin_lock(&imp
->imp_lock
);
1024 list_del(&imp
->imp_conn_current
->oic_item
);
1025 list_add(&imp
->imp_conn_current
->oic_item
,
1026 &imp
->imp_conn_list
);
1027 imp
->imp_last_success_conn
=
1028 imp
->imp_conn_current
->oic_last_attempt
;
1030 spin_unlock(&imp
->imp_lock
);
1032 if (!ocd
->ocd_ibits_known
&&
1033 ocd
->ocd_connect_flags
& OBD_CONNECT_IBITS
)
1034 CERROR("Inodebits aware server returned zero compatible"
1037 if ((ocd
->ocd_connect_flags
& OBD_CONNECT_VERSION
) &&
1038 (ocd
->ocd_version
> LUSTRE_VERSION_CODE
+
1039 LUSTRE_VERSION_OFFSET_WARN
||
1040 ocd
->ocd_version
< LUSTRE_VERSION_CODE
-
1041 LUSTRE_VERSION_OFFSET_WARN
)) {
1042 /* Sigh, some compilers do not like #ifdef in the middle
1043 of macro arguments */
1044 const char *older
= "older. Consider upgrading server "
1045 "or downgrading client";
1046 const char *newer
= "newer than client version. "
1047 "Consider upgrading client";
1049 LCONSOLE_WARN("Server %s version (%d.%d.%d.%d) "
1050 "is much %s (%s)\n",
1051 obd2cli_tgt(imp
->imp_obd
),
1052 OBD_OCD_VERSION_MAJOR(ocd
->ocd_version
),
1053 OBD_OCD_VERSION_MINOR(ocd
->ocd_version
),
1054 OBD_OCD_VERSION_PATCH(ocd
->ocd_version
),
1055 OBD_OCD_VERSION_FIX(ocd
->ocd_version
),
1056 ocd
->ocd_version
> LUSTRE_VERSION_CODE
?
1057 newer
: older
, LUSTRE_VERSION_STRING
);
1060 #if LUSTRE_VERSION_CODE < OBD_OCD_VERSION(3, 2, 50, 0)
1061 /* Check if server has LU-1252 fix applied to not always swab
1062 * the IR MNE entries. Do this only once per connection. This
1063 * fixup is version-limited, because we don't want to carry the
1064 * OBD_CONNECT_MNE_SWAB flag around forever, just so long as we
1065 * need interop with unpatched 2.2 servers. For newer servers,
1066 * the client will do MNE swabbing only as needed. LU-1644 */
1067 if (unlikely((ocd
->ocd_connect_flags
& OBD_CONNECT_VERSION
) &&
1068 !(ocd
->ocd_connect_flags
& OBD_CONNECT_MNE_SWAB
) &&
1069 OBD_OCD_VERSION_MAJOR(ocd
->ocd_version
) == 2 &&
1070 OBD_OCD_VERSION_MINOR(ocd
->ocd_version
) == 2 &&
1071 OBD_OCD_VERSION_PATCH(ocd
->ocd_version
) < 55 &&
1072 strcmp(imp
->imp_obd
->obd_type
->typ_name
,
1073 LUSTRE_MGC_NAME
) == 0))
1074 imp
->imp_need_mne_swab
= 1;
1075 else /* clear if server was upgraded since last connect */
1076 imp
->imp_need_mne_swab
= 0;
1078 #warning "LU-1644: Remove old OBD_CONNECT_MNE_SWAB fixup and imp_need_mne_swab"
1081 if (ocd
->ocd_connect_flags
& OBD_CONNECT_CKSUM
) {
1082 /* We sent to the server ocd_cksum_types with bits set
1083 * for algorithms we understand. The server masked off
1084 * the checksum types it doesn't support */
1085 if ((ocd
->ocd_cksum_types
&
1086 cksum_types_supported_client()) == 0) {
1087 LCONSOLE_WARN("The negotiation of the checksum "
1088 "algorithm to use with server %s "
1089 "failed (%x/%x), disabling "
1091 obd2cli_tgt(imp
->imp_obd
),
1092 ocd
->ocd_cksum_types
,
1093 cksum_types_supported_client());
1094 cli
->cl_checksum
= 0;
1095 cli
->cl_supp_cksum_types
= OBD_CKSUM_ADLER
;
1097 cli
->cl_supp_cksum_types
= ocd
->ocd_cksum_types
;
1100 /* The server does not support OBD_CONNECT_CKSUM.
1101 * Enforce ADLER for backward compatibility*/
1102 cli
->cl_supp_cksum_types
= OBD_CKSUM_ADLER
;
1104 cli
->cl_cksum_type
=cksum_type_select(cli
->cl_supp_cksum_types
);
1106 if (ocd
->ocd_connect_flags
& OBD_CONNECT_BRW_SIZE
)
1107 cli
->cl_max_pages_per_rpc
=
1108 min(ocd
->ocd_brw_size
>> PAGE_CACHE_SHIFT
,
1109 cli
->cl_max_pages_per_rpc
);
1110 else if (imp
->imp_connect_op
== MDS_CONNECT
||
1111 imp
->imp_connect_op
== MGS_CONNECT
)
1112 cli
->cl_max_pages_per_rpc
= 1;
1114 /* Reset ns_connect_flags only for initial connect. It might be
1115 * changed in while using FS and if we reset it in reconnect
1116 * this leads to losing user settings done before such as
1117 * disable lru_resize, etc. */
1118 if (old_connect_flags
!= exp_connect_flags(exp
) ||
1119 aa
->pcaa_initial_connect
) {
1120 CDEBUG(D_HA
, "%s: Resetting ns_connect_flags to server "
1121 "flags: "LPX64
"\n", imp
->imp_obd
->obd_name
,
1122 ocd
->ocd_connect_flags
);
1123 imp
->imp_obd
->obd_namespace
->ns_connect_flags
=
1124 ocd
->ocd_connect_flags
;
1125 imp
->imp_obd
->obd_namespace
->ns_orig_connect_flags
=
1126 ocd
->ocd_connect_flags
;
1129 if ((ocd
->ocd_connect_flags
& OBD_CONNECT_AT
) &&
1130 (imp
->imp_msg_magic
== LUSTRE_MSG_MAGIC_V2
))
1131 /* We need a per-message support flag, because
1132 a. we don't know if the incoming connect reply
1133 supports AT or not (in reply_in_callback)
1135 b. failovered server means export and flags are gone
1136 (in ptlrpc_send_reply).
1137 Can only be set when we know AT is supported at
1139 imp
->imp_msghdr_flags
|= MSGHDR_AT_SUPPORT
;
1141 imp
->imp_msghdr_flags
&= ~MSGHDR_AT_SUPPORT
;
1143 if ((ocd
->ocd_connect_flags
& OBD_CONNECT_FULL20
) &&
1144 (imp
->imp_msg_magic
== LUSTRE_MSG_MAGIC_V2
))
1145 imp
->imp_msghdr_flags
|= MSGHDR_CKSUM_INCOMPAT18
;
1147 imp
->imp_msghdr_flags
&= ~MSGHDR_CKSUM_INCOMPAT18
;
1149 LASSERT((cli
->cl_max_pages_per_rpc
<= PTLRPC_MAX_BRW_PAGES
) &&
1150 (cli
->cl_max_pages_per_rpc
> 0));
1154 imp
->imp_connect_tried
= 1;
1157 IMPORT_SET_STATE(imp
, LUSTRE_IMP_DISCON
);
1158 if (rc
== -EACCES
) {
1160 * Give up trying to reconnect
1161 * EACCES means client has no permission for connection
1163 imp
->imp_obd
->obd_no_recov
= 1;
1164 ptlrpc_deactivate_import(imp
);
1167 if (rc
== -EPROTO
) {
1168 struct obd_connect_data
*ocd
;
1170 /* reply message might not be ready */
1171 if (request
->rq_repmsg
== NULL
)
1174 ocd
= req_capsule_server_get(&request
->rq_pill
,
1177 (ocd
->ocd_connect_flags
& OBD_CONNECT_VERSION
) &&
1178 (ocd
->ocd_version
!= LUSTRE_VERSION_CODE
)) {
1180 * Actually servers are only supposed to refuse
1181 * connection from liblustre clients, so we
1182 * should never see this from VFS context
1184 LCONSOLE_ERROR_MSG(0x16a, "Server %s version "
1186 " refused connection from this client "
1187 "with an incompatible version (%s). "
1188 "Client must be recompiled\n",
1189 obd2cli_tgt(imp
->imp_obd
),
1190 OBD_OCD_VERSION_MAJOR(ocd
->ocd_version
),
1191 OBD_OCD_VERSION_MINOR(ocd
->ocd_version
),
1192 OBD_OCD_VERSION_PATCH(ocd
->ocd_version
),
1193 OBD_OCD_VERSION_FIX(ocd
->ocd_version
),
1194 LUSTRE_VERSION_STRING
);
1195 ptlrpc_deactivate_import(imp
);
1196 IMPORT_SET_STATE(imp
, LUSTRE_IMP_CLOSED
);
1201 ptlrpc_maybe_ping_import_soon(imp
);
1203 CDEBUG(D_HA
, "recovery of %s on %s failed (%d)\n",
1204 obd2cli_tgt(imp
->imp_obd
),
1205 (char *)imp
->imp_connection
->c_remote_uuid
.uuid
, rc
);
1208 wake_up_all(&imp
->imp_recovery_waitq
);
1213 * interpret callback for "completed replay" RPCs.
1214 * \see signal_completed_replay
1216 static int completed_replay_interpret(const struct lu_env
*env
,
1217 struct ptlrpc_request
*req
,
1218 void * data
, int rc
)
1220 atomic_dec(&req
->rq_import
->imp_replay_inflight
);
1221 if (req
->rq_status
== 0 &&
1222 !req
->rq_import
->imp_vbr_failed
) {
1223 ptlrpc_import_recovery_state_machine(req
->rq_import
);
1225 if (req
->rq_import
->imp_vbr_failed
) {
1227 "%s: version recovery fails, reconnecting\n",
1228 req
->rq_import
->imp_obd
->obd_name
);
1230 CDEBUG(D_HA
, "%s: LAST_REPLAY message error: %d, "
1232 req
->rq_import
->imp_obd
->obd_name
,
1235 ptlrpc_connect_import(req
->rq_import
);
1242 * Let server know that we have no requests to replay anymore.
1243 * Achieved by just sending a PING request
1245 static int signal_completed_replay(struct obd_import
*imp
)
1247 struct ptlrpc_request
*req
;
1249 if (unlikely(OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_FINISH_REPLAY
)))
1252 LASSERT(atomic_read(&imp
->imp_replay_inflight
) == 0);
1253 atomic_inc(&imp
->imp_replay_inflight
);
1255 req
= ptlrpc_request_alloc_pack(imp
, &RQF_OBD_PING
, LUSTRE_OBD_VERSION
,
1258 atomic_dec(&imp
->imp_replay_inflight
);
1262 ptlrpc_request_set_replen(req
);
1263 req
->rq_send_state
= LUSTRE_IMP_REPLAY_WAIT
;
1264 lustre_msg_add_flags(req
->rq_reqmsg
,
1265 MSG_LOCK_REPLAY_DONE
| MSG_REQ_REPLAY_DONE
);
1267 req
->rq_timeout
*= 3;
1268 req
->rq_interpret_reply
= completed_replay_interpret
;
1270 ptlrpcd_add_req(req
, PDL_POLICY_ROUND
, -1);
1275 * In kernel code all import invalidation happens in its own
1276 * separate thread, so that whatever application happened to encounter
1277 * a problem could still be killed or otherwise continue
1279 static int ptlrpc_invalidate_import_thread(void *data
)
1281 struct obd_import
*imp
= data
;
1283 unshare_fs_struct();
1285 CDEBUG(D_HA
, "thread invalidate import %s to %s@%s\n",
1286 imp
->imp_obd
->obd_name
, obd2cli_tgt(imp
->imp_obd
),
1287 imp
->imp_connection
->c_remote_uuid
.uuid
);
1289 ptlrpc_invalidate_import(imp
);
1291 if (obd_dump_on_eviction
) {
1292 CERROR("dump the log upon eviction\n");
1293 libcfs_debug_dumplog();
1296 IMPORT_SET_STATE(imp
, LUSTRE_IMP_RECOVER
);
1297 ptlrpc_import_recovery_state_machine(imp
);
1299 class_import_put(imp
);
1304 * This is the state machine for client-side recovery on import.
1306 * Typically we have two possibly paths. If we came to server and it is not
1307 * in recovery, we just enter IMP_EVICTED state, invalidate our import
1308 * state and reconnect from scratch.
1309 * If we came to server that is in recovery, we enter IMP_REPLAY import state.
1310 * We go through our list of requests to replay and send them to server one by
1312 * After sending all request from the list we change import state to
1313 * IMP_REPLAY_LOCKS and re-request all the locks we believe we have from server
1314 * and also all the locks we don't yet have and wait for server to grant us.
1315 * After that we send a special "replay completed" request and change import
1316 * state to IMP_REPLAY_WAIT.
1317 * Upon receiving reply to that "replay completed" RPC we enter IMP_RECOVER
1318 * state and resend all requests from sending list.
1319 * After that we promote import to FULL state and send all delayed requests
1320 * and import is fully operational after that.
1323 int ptlrpc_import_recovery_state_machine(struct obd_import
*imp
)
1330 if (imp
->imp_state
== LUSTRE_IMP_EVICTED
) {
1331 deuuidify(obd2cli_tgt(imp
->imp_obd
), NULL
,
1332 &target_start
, &target_len
);
1333 /* Don't care about MGC eviction */
1334 if (strcmp(imp
->imp_obd
->obd_type
->typ_name
,
1335 LUSTRE_MGC_NAME
) != 0) {
1336 LCONSOLE_ERROR_MSG(0x167, "%s: This client was evicted "
1337 "by %.*s; in progress operations "
1338 "using this service will fail.\n",
1339 imp
->imp_obd
->obd_name
, target_len
,
1342 CDEBUG(D_HA
, "evicted from %s@%s; invalidating\n",
1343 obd2cli_tgt(imp
->imp_obd
),
1344 imp
->imp_connection
->c_remote_uuid
.uuid
);
1345 /* reset vbr_failed flag upon eviction */
1346 spin_lock(&imp
->imp_lock
);
1347 imp
->imp_vbr_failed
= 0;
1348 spin_unlock(&imp
->imp_lock
);
1351 struct task_struct
*task
;
1352 /* bug 17802: XXX client_disconnect_export vs connect request
1353 * race. if client will evicted at this time, we start
1354 * invalidate thread without reference to import and import can
1355 * be freed at same time. */
1356 class_import_get(imp
);
1357 task
= kthread_run(ptlrpc_invalidate_import_thread
, imp
,
1360 class_import_put(imp
);
1361 CERROR("error starting invalidate thread: %d\n", rc
);
1370 if (imp
->imp_state
== LUSTRE_IMP_REPLAY
) {
1371 CDEBUG(D_HA
, "replay requested by %s\n",
1372 obd2cli_tgt(imp
->imp_obd
));
1373 rc
= ptlrpc_replay_next(imp
, &inflight
);
1374 if (inflight
== 0 &&
1375 atomic_read(&imp
->imp_replay_inflight
) == 0) {
1376 IMPORT_SET_STATE(imp
, LUSTRE_IMP_REPLAY_LOCKS
);
1377 rc
= ldlm_replay_locks(imp
);
1384 if (imp
->imp_state
== LUSTRE_IMP_REPLAY_LOCKS
) {
1385 if (atomic_read(&imp
->imp_replay_inflight
) == 0) {
1386 IMPORT_SET_STATE(imp
, LUSTRE_IMP_REPLAY_WAIT
);
1387 rc
= signal_completed_replay(imp
);
1394 if (imp
->imp_state
== LUSTRE_IMP_REPLAY_WAIT
) {
1395 if (atomic_read(&imp
->imp_replay_inflight
) == 0) {
1396 IMPORT_SET_STATE(imp
, LUSTRE_IMP_RECOVER
);
1400 if (imp
->imp_state
== LUSTRE_IMP_RECOVER
) {
1401 CDEBUG(D_HA
, "reconnected to %s@%s\n",
1402 obd2cli_tgt(imp
->imp_obd
),
1403 imp
->imp_connection
->c_remote_uuid
.uuid
);
1405 rc
= ptlrpc_resend(imp
);
1408 IMPORT_SET_STATE(imp
, LUSTRE_IMP_FULL
);
1409 ptlrpc_activate_import(imp
);
1411 deuuidify(obd2cli_tgt(imp
->imp_obd
), NULL
,
1412 &target_start
, &target_len
);
1413 LCONSOLE_INFO("%s: Connection restored to %.*s (at %s)\n",
1414 imp
->imp_obd
->obd_name
,
1415 target_len
, target_start
,
1416 libcfs_nid2str(imp
->imp_connection
->c_peer
.nid
));
1419 if (imp
->imp_state
== LUSTRE_IMP_FULL
) {
1420 wake_up_all(&imp
->imp_recovery_waitq
);
1421 ptlrpc_wake_delayed(imp
);
1428 int ptlrpc_disconnect_import(struct obd_import
*imp
, int noclose
)
1430 struct ptlrpc_request
*req
;
1433 if (imp
->imp_obd
->obd_force
)
1434 GOTO(set_state
, rc
);
1436 switch (imp
->imp_connect_op
) {
1438 rq_opc
= OST_DISCONNECT
;
1441 rq_opc
= MDS_DISCONNECT
;
1444 rq_opc
= MGS_DISCONNECT
;
1448 CERROR("%s: don't know how to disconnect from %s "
1449 "(connect_op %d): rc = %d\n",
1450 imp
->imp_obd
->obd_name
, obd2cli_tgt(imp
->imp_obd
),
1451 imp
->imp_connect_op
, rc
);
1455 if (ptlrpc_import_in_recovery(imp
)) {
1456 struct l_wait_info lwi
;
1457 cfs_duration_t timeout
;
1460 if (imp
->imp_server_timeout
)
1461 timeout
= cfs_time_seconds(obd_timeout
/ 2);
1463 timeout
= cfs_time_seconds(obd_timeout
);
1465 int idx
= import_at_get_index(imp
,
1466 imp
->imp_client
->cli_request_portal
);
1467 timeout
= cfs_time_seconds(
1468 at_get(&imp
->imp_at
.iat_service_estimate
[idx
]));
1471 lwi
= LWI_TIMEOUT_INTR(cfs_timeout_cap(timeout
),
1472 back_to_sleep
, LWI_ON_SIGNAL_NOOP
, NULL
);
1473 rc
= l_wait_event(imp
->imp_recovery_waitq
,
1474 !ptlrpc_import_in_recovery(imp
), &lwi
);
1478 spin_lock(&imp
->imp_lock
);
1479 if (imp
->imp_state
!= LUSTRE_IMP_FULL
)
1481 spin_unlock(&imp
->imp_lock
);
1483 req
= ptlrpc_request_alloc_pack(imp
, &RQF_MDS_DISCONNECT
,
1484 LUSTRE_OBD_VERSION
, rq_opc
);
1486 /* We are disconnecting, do not retry a failed DISCONNECT rpc if
1487 * it fails. We can get through the above with a down server
1488 * if the client doesn't know the server is gone yet. */
1489 req
->rq_no_resend
= 1;
1491 /* We want client umounts to happen quickly, no matter the
1493 req
->rq_timeout
= min_t(int, req
->rq_timeout
,
1494 INITIAL_CONNECT_TIMEOUT
);
1496 IMPORT_SET_STATE(imp
, LUSTRE_IMP_CONNECTING
);
1497 req
->rq_send_state
= LUSTRE_IMP_CONNECTING
;
1498 ptlrpc_request_set_replen(req
);
1499 rc
= ptlrpc_queue_wait(req
);
1500 ptlrpc_req_finished(req
);
1504 spin_lock(&imp
->imp_lock
);
1507 IMPORT_SET_STATE_NOLOCK(imp
, LUSTRE_IMP_DISCON
);
1509 IMPORT_SET_STATE_NOLOCK(imp
, LUSTRE_IMP_CLOSED
);
1510 memset(&imp
->imp_remote_handle
, 0, sizeof(imp
->imp_remote_handle
));
1511 spin_unlock(&imp
->imp_lock
);
1513 if (rc
== -ETIMEDOUT
|| rc
== -ENOTCONN
|| rc
== -ESHUTDOWN
)
1518 EXPORT_SYMBOL(ptlrpc_disconnect_import
);
1520 void ptlrpc_cleanup_imp(struct obd_import
*imp
)
1522 spin_lock(&imp
->imp_lock
);
1523 IMPORT_SET_STATE_NOLOCK(imp
, LUSTRE_IMP_CLOSED
);
1524 imp
->imp_generation
++;
1525 spin_unlock(&imp
->imp_lock
);
1526 ptlrpc_abort_inflight(imp
);
1528 EXPORT_SYMBOL(ptlrpc_cleanup_imp
);
1530 /* Adaptive Timeout utils */
1531 extern unsigned int at_min
, at_max
, at_history
;
1533 /* Bin into timeslices using AT_BINS bins.
1534 This gives us a max of the last binlimit*AT_BINS secs without the storage,
1535 but still smoothing out a return to normalcy from a slow response.
1536 (E.g. remember the maximum latency in each minute of the last 4 minutes.) */
1537 int at_measured(struct adaptive_timeout
*at
, unsigned int val
)
1539 unsigned int old
= at
->at_current
;
1540 time_t now
= cfs_time_current_sec();
1541 time_t binlimit
= max_t(time_t, at_history
/ AT_BINS
, 1);
1544 CDEBUG(D_OTHER
, "add %u to %p time=%lu v=%u (%u %u %u %u)\n",
1545 val
, at
, now
- at
->at_binstart
, at
->at_current
,
1546 at
->at_hist
[0], at
->at_hist
[1], at
->at_hist
[2], at
->at_hist
[3]);
1549 /* 0's don't count, because we never want our timeout to
1550 drop to 0, and because 0 could mean an error */
1553 spin_lock(&at
->at_lock
);
1555 if (unlikely(at
->at_binstart
== 0)) {
1556 /* Special case to remove default from history */
1557 at
->at_current
= val
;
1558 at
->at_worst_ever
= val
;
1559 at
->at_worst_time
= now
;
1560 at
->at_hist
[0] = val
;
1561 at
->at_binstart
= now
;
1562 } else if (now
- at
->at_binstart
< binlimit
) {
1564 at
->at_hist
[0] = max(val
, at
->at_hist
[0]);
1565 at
->at_current
= max(val
, at
->at_current
);
1568 unsigned int maxv
= val
;
1569 /* move bins over */
1570 shift
= (now
- at
->at_binstart
) / binlimit
;
1572 for (i
= AT_BINS
- 1; i
>= 0; i
--) {
1574 at
->at_hist
[i
] = at
->at_hist
[i
- shift
];
1575 maxv
= max(maxv
, at
->at_hist
[i
]);
1580 at
->at_hist
[0] = val
;
1581 at
->at_current
= maxv
;
1582 at
->at_binstart
+= shift
* binlimit
;
1585 if (at
->at_current
> at
->at_worst_ever
) {
1586 at
->at_worst_ever
= at
->at_current
;
1587 at
->at_worst_time
= now
;
1590 if (at
->at_flags
& AT_FLG_NOHIST
)
1591 /* Only keep last reported val; keeping the rest of the history
1593 at
->at_current
= val
;
1596 at
->at_current
= min(at
->at_current
, at_max
);
1597 at
->at_current
= max(at
->at_current
, at_min
);
1599 if (at
->at_current
!= old
)
1600 CDEBUG(D_OTHER
, "AT %p change: old=%u new=%u delta=%d "
1601 "(val=%u) hist %u %u %u %u\n", at
,
1602 old
, at
->at_current
, at
->at_current
- old
, val
,
1603 at
->at_hist
[0], at
->at_hist
[1], at
->at_hist
[2],
1606 /* if we changed, report the old value */
1607 old
= (at
->at_current
!= old
) ? old
: 0;
1609 spin_unlock(&at
->at_lock
);
1613 /* Find the imp_at index for a given portal; assign if space available */
1614 int import_at_get_index(struct obd_import
*imp
, int portal
)
1616 struct imp_at
*at
= &imp
->imp_at
;
1619 for (i
= 0; i
< IMP_AT_MAX_PORTALS
; i
++) {
1620 if (at
->iat_portal
[i
] == portal
)
1622 if (at
->iat_portal
[i
] == 0)
1627 /* Not found in list, add it under a lock */
1628 spin_lock(&imp
->imp_lock
);
1630 /* Check unused under lock */
1631 for (; i
< IMP_AT_MAX_PORTALS
; i
++) {
1632 if (at
->iat_portal
[i
] == portal
)
1634 if (at
->iat_portal
[i
] == 0)
1639 /* Not enough portals? */
1640 LASSERT(i
< IMP_AT_MAX_PORTALS
);
1642 at
->iat_portal
[i
] = portal
;
1644 spin_unlock(&imp
->imp_lock
);