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 <obd_support.h>
44 #include <lustre_ha.h>
45 #include <lustre_net.h>
46 #include <lustre_import.h>
47 #include <lustre_export.h>
49 #include <obd_cksum.h>
50 #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 LASSERT(imp
->imp_invalid
);
280 /* Wait forever until inflight == 0. We really can't do it another
281 * way because in some cases we need to wait for very long reply
282 * unlink. We can't do anything before that because there is really
283 * no guarantee that some rdma transfer is not in progress right now. */
285 /* Calculate max timeout for waiting on rpcs to error
286 * out. Use obd_timeout if calculated value is smaller
288 if (!OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_REPL_UNLINK
)) {
289 timeout
= ptlrpc_inflight_timeout(imp
);
290 timeout
+= timeout
/ 3;
293 timeout
= obd_timeout
;
295 /* decrease the interval to increase race condition */
299 CDEBUG(D_RPCTRACE
,"Sleeping %d sec for inflight to error out\n",
302 /* Wait for all requests to error out and call completion
303 * callbacks. Cap it at obd_timeout -- these should all
304 * have been locally cancelled by ptlrpc_abort_inflight. */
305 lwi
= LWI_TIMEOUT_INTERVAL(
306 cfs_timeout_cap(cfs_time_seconds(timeout
)),
307 (timeout
> 1)?cfs_time_seconds(1):cfs_time_seconds(1)/2,
309 rc
= l_wait_event(imp
->imp_recovery_waitq
,
310 (atomic_read(&imp
->imp_inflight
) == 0),
313 const char *cli_tgt
= obd2cli_tgt(imp
->imp_obd
);
315 CERROR("%s: rc = %d waiting for callback (%d != 0)\n",
317 atomic_read(&imp
->imp_inflight
));
319 spin_lock(&imp
->imp_lock
);
320 if (atomic_read(&imp
->imp_inflight
) == 0) {
321 int count
= atomic_read(&imp
->imp_unregistering
);
323 /* We know that "unregistering" rpcs only can
324 * survive in sending or delaying lists (they
325 * maybe waiting for long reply unlink in
326 * sluggish nets). Let's check this. If there
327 * is no inflight and unregistering != 0, this
329 LASSERTF(count
== 0, "Some RPCs are still "
330 "unregistering: %d\n", count
);
332 /* Let's save one loop as soon as inflight have
333 * dropped to zero. No new inflights possible at
337 list_for_each_safe(tmp
, n
,
338 &imp
->imp_sending_list
) {
339 req
= list_entry(tmp
,
340 struct ptlrpc_request
,
342 DEBUG_REQ(D_ERROR
, req
,
343 "still on sending list");
345 list_for_each_safe(tmp
, n
,
346 &imp
->imp_delayed_list
) {
347 req
= list_entry(tmp
,
348 struct ptlrpc_request
,
350 DEBUG_REQ(D_ERROR
, req
,
351 "still on delayed list");
354 CERROR("%s: RPCs in \"%s\" phase found (%d). "
355 "Network is sluggish? Waiting them "
356 "to error out.\n", cli_tgt
,
357 ptlrpc_phase2str(RQ_PHASE_UNREGISTERING
),
361 spin_unlock(&imp
->imp_lock
);
366 * Let's additionally check that no new rpcs added to import in
367 * "invalidate" state.
369 LASSERT(atomic_read(&imp
->imp_inflight
) == 0);
370 obd_import_event(imp
->imp_obd
, imp
, IMP_EVENT_INVALIDATE
);
371 sptlrpc_import_flush_all_ctx(imp
);
373 atomic_dec(&imp
->imp_inval_count
);
374 wake_up_all(&imp
->imp_recovery_waitq
);
376 EXPORT_SYMBOL(ptlrpc_invalidate_import
);
378 /* unset imp_invalid */
379 void ptlrpc_activate_import(struct obd_import
*imp
)
381 struct obd_device
*obd
= imp
->imp_obd
;
383 spin_lock(&imp
->imp_lock
);
384 if (imp
->imp_deactive
!= 0) {
385 spin_unlock(&imp
->imp_lock
);
389 imp
->imp_invalid
= 0;
390 spin_unlock(&imp
->imp_lock
);
391 obd_import_event(obd
, imp
, IMP_EVENT_ACTIVE
);
393 EXPORT_SYMBOL(ptlrpc_activate_import
);
395 void ptlrpc_fail_import(struct obd_import
*imp
, __u32 conn_cnt
)
397 LASSERT(!imp
->imp_dlm_fake
);
399 if (ptlrpc_set_import_discon(imp
, conn_cnt
)) {
400 if (!imp
->imp_replayable
) {
401 CDEBUG(D_HA
, "import %s@%s for %s not replayable, "
402 "auto-deactivating\n",
403 obd2cli_tgt(imp
->imp_obd
),
404 imp
->imp_connection
->c_remote_uuid
.uuid
,
405 imp
->imp_obd
->obd_name
);
406 ptlrpc_deactivate_import(imp
);
409 CDEBUG(D_HA
, "%s: waking up pinger\n",
410 obd2cli_tgt(imp
->imp_obd
));
412 spin_lock(&imp
->imp_lock
);
413 imp
->imp_force_verify
= 1;
414 spin_unlock(&imp
->imp_lock
);
416 ptlrpc_pinger_wake_up();
419 EXPORT_SYMBOL(ptlrpc_fail_import
);
421 int ptlrpc_reconnect_import(struct obd_import
*imp
)
423 ptlrpc_set_import_discon(imp
, 0);
424 /* Force a new connect attempt */
425 ptlrpc_invalidate_import(imp
);
426 /* Do a fresh connect next time by zeroing the handle */
427 ptlrpc_disconnect_import(imp
, 1);
428 /* Wait for all invalidate calls to finish */
429 if (atomic_read(&imp
->imp_inval_count
) > 0) {
431 struct l_wait_info lwi
= LWI_INTR(LWI_ON_SIGNAL_NOOP
, NULL
);
432 rc
= l_wait_event(imp
->imp_recovery_waitq
,
433 (atomic_read(&imp
->imp_inval_count
) == 0),
436 CERROR("Interrupted, inval=%d\n",
437 atomic_read(&imp
->imp_inval_count
));
440 /* Allow reconnect attempts */
441 imp
->imp_obd
->obd_no_recov
= 0;
442 /* Remove 'invalid' flag */
443 ptlrpc_activate_import(imp
);
444 /* Attempt a new connect */
445 ptlrpc_recover_import(imp
, NULL
, 0);
448 EXPORT_SYMBOL(ptlrpc_reconnect_import
);
451 * Connection on import \a imp is changed to another one (if more than one is
452 * present). We typically chose connection that we have not tried to connect to
455 static int import_select_connection(struct obd_import
*imp
)
457 struct obd_import_conn
*imp_conn
= NULL
, *conn
;
458 struct obd_export
*dlmexp
;
460 int target_len
, tried_all
= 1;
462 spin_lock(&imp
->imp_lock
);
464 if (list_empty(&imp
->imp_conn_list
)) {
465 CERROR("%s: no connections available\n",
466 imp
->imp_obd
->obd_name
);
467 spin_unlock(&imp
->imp_lock
);
471 list_for_each_entry(conn
, &imp
->imp_conn_list
, oic_item
) {
472 CDEBUG(D_HA
, "%s: connect to NID %s last attempt "LPU64
"\n",
473 imp
->imp_obd
->obd_name
,
474 libcfs_nid2str(conn
->oic_conn
->c_peer
.nid
),
475 conn
->oic_last_attempt
);
477 /* If we have not tried this connection since
478 the last successful attempt, go with this one */
479 if ((conn
->oic_last_attempt
== 0) ||
480 cfs_time_beforeq_64(conn
->oic_last_attempt
,
481 imp
->imp_last_success_conn
)) {
487 /* If all of the connections have already been tried
488 since the last successful connection; just choose the
489 least recently used */
492 else if (cfs_time_before_64(conn
->oic_last_attempt
,
493 imp_conn
->oic_last_attempt
))
497 /* if not found, simply choose the current one */
498 if (!imp_conn
|| imp
->imp_force_reconnect
) {
499 LASSERT(imp
->imp_conn_current
);
500 imp_conn
= imp
->imp_conn_current
;
503 LASSERT(imp_conn
->oic_conn
);
505 /* If we've tried everything, and we're back to the beginning of the
506 list, increase our timeout and try again. It will be reset when
507 we do finally connect. (FIXME: really we should wait for all network
508 state associated with the last connection attempt to drain before
509 trying to reconnect on it.) */
510 if (tried_all
&& (imp
->imp_conn_list
.next
== &imp_conn
->oic_item
)) {
511 struct adaptive_timeout
*at
= &imp
->imp_at
.iat_net_latency
;
512 if (at_get(at
) < CONNECTION_SWITCH_MAX
) {
513 at_measured(at
, at_get(at
) + CONNECTION_SWITCH_INC
);
514 if (at_get(at
) > CONNECTION_SWITCH_MAX
)
515 at_reset(at
, CONNECTION_SWITCH_MAX
);
517 LASSERT(imp_conn
->oic_last_attempt
);
518 CDEBUG(D_HA
, "%s: tried all connections, increasing latency "
519 "to %ds\n", imp
->imp_obd
->obd_name
, at_get(at
));
522 imp_conn
->oic_last_attempt
= cfs_time_current_64();
524 /* switch connection, don't mind if it's same as the current one */
525 if (imp
->imp_connection
)
526 ptlrpc_connection_put(imp
->imp_connection
);
527 imp
->imp_connection
= ptlrpc_connection_addref(imp_conn
->oic_conn
);
529 dlmexp
= class_conn2export(&imp
->imp_dlm_handle
);
530 LASSERT(dlmexp
!= NULL
);
531 if (dlmexp
->exp_connection
)
532 ptlrpc_connection_put(dlmexp
->exp_connection
);
533 dlmexp
->exp_connection
= ptlrpc_connection_addref(imp_conn
->oic_conn
);
534 class_export_put(dlmexp
);
536 if (imp
->imp_conn_current
!= imp_conn
) {
537 if (imp
->imp_conn_current
) {
538 deuuidify(obd2cli_tgt(imp
->imp_obd
), NULL
,
539 &target_start
, &target_len
);
541 CDEBUG(D_HA
, "%s: Connection changing to"
543 imp
->imp_obd
->obd_name
,
544 target_len
, target_start
,
545 libcfs_nid2str(imp_conn
->oic_conn
->c_peer
.nid
));
548 imp
->imp_conn_current
= imp_conn
;
551 CDEBUG(D_HA
, "%s: import %p using connection %s/%s\n",
552 imp
->imp_obd
->obd_name
, imp
, imp_conn
->oic_uuid
.uuid
,
553 libcfs_nid2str(imp_conn
->oic_conn
->c_peer
.nid
));
555 spin_unlock(&imp
->imp_lock
);
561 * must be called under imp_lock
563 static int ptlrpc_first_transno(struct obd_import
*imp
, __u64
*transno
)
565 struct ptlrpc_request
*req
;
566 struct list_head
*tmp
;
568 /* The requests in committed_list always have smaller transnos than
569 * the requests in replay_list */
570 if (!list_empty(&imp
->imp_committed_list
)) {
571 tmp
= imp
->imp_committed_list
.next
;
572 req
= list_entry(tmp
, struct ptlrpc_request
, rq_replay_list
);
573 *transno
= req
->rq_transno
;
574 if (req
->rq_transno
== 0) {
575 DEBUG_REQ(D_ERROR
, req
,
576 "zero transno in committed_list");
581 if (!list_empty(&imp
->imp_replay_list
)) {
582 tmp
= imp
->imp_replay_list
.next
;
583 req
= list_entry(tmp
, struct ptlrpc_request
, rq_replay_list
);
584 *transno
= req
->rq_transno
;
585 if (req
->rq_transno
== 0) {
586 DEBUG_REQ(D_ERROR
, req
, "zero transno in replay_list");
595 * Attempt to (re)connect import \a imp. This includes all preparations,
596 * initializing CONNECT RPC request and passing it to ptlrpcd for
598 * Returns 0 on success or error code.
600 int ptlrpc_connect_import(struct obd_import
*imp
)
602 struct obd_device
*obd
= imp
->imp_obd
;
603 int initial_connect
= 0;
605 __u64 committed_before_reconnect
= 0;
606 struct ptlrpc_request
*request
;
607 char *bufs
[] = { NULL
,
608 obd2cli_tgt(imp
->imp_obd
),
610 (char *)&imp
->imp_dlm_handle
,
611 (char *)&imp
->imp_connect_data
};
612 struct ptlrpc_connect_async_args
*aa
;
615 spin_lock(&imp
->imp_lock
);
616 if (imp
->imp_state
== LUSTRE_IMP_CLOSED
) {
617 spin_unlock(&imp
->imp_lock
);
618 CERROR("can't connect to a closed import\n");
620 } else if (imp
->imp_state
== LUSTRE_IMP_FULL
) {
621 spin_unlock(&imp
->imp_lock
);
622 CERROR("already connected\n");
624 } else if (imp
->imp_state
== LUSTRE_IMP_CONNECTING
) {
625 spin_unlock(&imp
->imp_lock
);
626 CERROR("already connecting\n");
630 IMPORT_SET_STATE_NOLOCK(imp
, LUSTRE_IMP_CONNECTING
);
633 imp
->imp_resend_replay
= 0;
635 if (!lustre_handle_is_used(&imp
->imp_remote_handle
))
638 committed_before_reconnect
= imp
->imp_peer_committed_transno
;
640 set_transno
= ptlrpc_first_transno(imp
,
641 &imp
->imp_connect_data
.ocd_transno
);
642 spin_unlock(&imp
->imp_lock
);
644 rc
= import_select_connection(imp
);
648 rc
= sptlrpc_import_sec_adapt(imp
, NULL
, NULL
);
652 /* Reset connect flags to the originally requested flags, in case
653 * the server is updated on-the-fly we will get the new features. */
654 imp
->imp_connect_data
.ocd_connect_flags
= imp
->imp_connect_flags_orig
;
655 /* Reset ocd_version each time so the server knows the exact versions */
656 imp
->imp_connect_data
.ocd_version
= LUSTRE_VERSION_CODE
;
657 imp
->imp_msghdr_flags
&= ~MSGHDR_AT_SUPPORT
;
658 imp
->imp_msghdr_flags
&= ~MSGHDR_CKSUM_INCOMPAT18
;
660 rc
= obd_reconnect(NULL
, imp
->imp_obd
->obd_self_export
, obd
,
661 &obd
->obd_uuid
, &imp
->imp_connect_data
, NULL
);
665 request
= ptlrpc_request_alloc(imp
, &RQF_MDS_CONNECT
);
667 GOTO(out
, rc
= -ENOMEM
);
669 rc
= ptlrpc_request_bufs_pack(request
, LUSTRE_OBD_VERSION
,
670 imp
->imp_connect_op
, bufs
, NULL
);
672 ptlrpc_request_free(request
);
676 /* Report the rpc service time to the server so that it knows how long
677 * to wait for clients to join recovery */
678 lustre_msg_set_service_time(request
->rq_reqmsg
,
679 at_timeout2est(request
->rq_timeout
));
681 /* The amount of time we give the server to process the connect req.
682 * import_select_connection will increase the net latency on
683 * repeated reconnect attempts to cover slow networks.
684 * We override/ignore the server rpc completion estimate here,
685 * which may be large if this is a reconnect attempt */
686 request
->rq_timeout
= INITIAL_CONNECT_TIMEOUT
;
687 lustre_msg_set_timeout(request
->rq_reqmsg
, request
->rq_timeout
);
689 lustre_msg_add_op_flags(request
->rq_reqmsg
, MSG_CONNECT_NEXT_VER
);
691 request
->rq_no_resend
= request
->rq_no_delay
= 1;
692 request
->rq_send_state
= LUSTRE_IMP_CONNECTING
;
693 /* Allow a slightly larger reply for future growth compatibility */
694 req_capsule_set_size(&request
->rq_pill
, &RMF_CONNECT_DATA
, RCL_SERVER
,
695 sizeof(struct obd_connect_data
)+16*sizeof(__u64
));
696 ptlrpc_request_set_replen(request
);
697 request
->rq_interpret_reply
= ptlrpc_connect_interpret
;
699 CLASSERT(sizeof(*aa
) <= sizeof(request
->rq_async_args
));
700 aa
= ptlrpc_req_async_args(request
);
701 memset(aa
, 0, sizeof(*aa
));
703 aa
->pcaa_peer_committed
= committed_before_reconnect
;
704 aa
->pcaa_initial_connect
= initial_connect
;
706 if (aa
->pcaa_initial_connect
) {
707 spin_lock(&imp
->imp_lock
);
708 imp
->imp_replayable
= 1;
709 spin_unlock(&imp
->imp_lock
);
710 lustre_msg_add_op_flags(request
->rq_reqmsg
,
711 MSG_CONNECT_INITIAL
);
715 lustre_msg_add_op_flags(request
->rq_reqmsg
,
716 MSG_CONNECT_TRANSNO
);
718 DEBUG_REQ(D_RPCTRACE
, request
, "(re)connect request (timeout %d)",
719 request
->rq_timeout
);
720 ptlrpcd_add_req(request
, PDL_POLICY_ROUND
, -1);
724 IMPORT_SET_STATE(imp
, LUSTRE_IMP_DISCON
);
729 EXPORT_SYMBOL(ptlrpc_connect_import
);
731 static void ptlrpc_maybe_ping_import_soon(struct obd_import
*imp
)
735 spin_lock(&imp
->imp_lock
);
736 force_verify
= imp
->imp_force_verify
!= 0;
737 spin_unlock(&imp
->imp_lock
);
740 ptlrpc_pinger_wake_up();
743 static int ptlrpc_busy_reconnect(int rc
)
745 return (rc
== -EBUSY
) || (rc
== -EAGAIN
);
749 * interpret_reply callback for connect RPCs.
750 * Looks into returned status of connect operation and decides
751 * what to do with the import - i.e enter recovery, promote it to
752 * full state for normal operations of disconnect it due to an error.
754 static int ptlrpc_connect_interpret(const struct lu_env
*env
,
755 struct ptlrpc_request
*request
,
758 struct ptlrpc_connect_async_args
*aa
= data
;
759 struct obd_import
*imp
= request
->rq_import
;
760 struct client_obd
*cli
= &imp
->imp_obd
->u
.cli
;
761 struct lustre_handle old_hdl
;
762 __u64 old_connect_flags
;
764 struct obd_connect_data
*ocd
;
765 struct obd_export
*exp
;
768 spin_lock(&imp
->imp_lock
);
769 if (imp
->imp_state
== LUSTRE_IMP_CLOSED
) {
770 imp
->imp_connect_tried
= 1;
771 spin_unlock(&imp
->imp_lock
);
776 /* if this reconnect to busy export - not need select new target
778 imp
->imp_force_reconnect
= ptlrpc_busy_reconnect(rc
);
779 spin_unlock(&imp
->imp_lock
);
780 ptlrpc_maybe_ping_import_soon(imp
);
783 spin_unlock(&imp
->imp_lock
);
785 LASSERT(imp
->imp_conn_current
);
787 msg_flags
= lustre_msg_get_op_flags(request
->rq_repmsg
);
789 ret
= req_capsule_get_size(&request
->rq_pill
, &RMF_CONNECT_DATA
,
791 /* server replied obd_connect_data is always bigger */
792 ocd
= req_capsule_server_sized_get(&request
->rq_pill
,
793 &RMF_CONNECT_DATA
, ret
);
796 CERROR("%s: no connect data from server\n",
797 imp
->imp_obd
->obd_name
);
802 spin_lock(&imp
->imp_lock
);
804 /* All imports are pingable */
805 imp
->imp_pingable
= 1;
806 imp
->imp_force_reconnect
= 0;
807 imp
->imp_force_verify
= 0;
809 imp
->imp_connect_data
= *ocd
;
811 CDEBUG(D_HA
, "%s: connect to target with instance %u\n",
812 imp
->imp_obd
->obd_name
, ocd
->ocd_instance
);
813 exp
= class_conn2export(&imp
->imp_dlm_handle
);
815 spin_unlock(&imp
->imp_lock
);
817 /* check that server granted subset of flags we asked for. */
818 if ((ocd
->ocd_connect_flags
& imp
->imp_connect_flags_orig
) !=
819 ocd
->ocd_connect_flags
) {
820 CERROR("%s: Server didn't granted asked subset of flags: "
821 "asked="LPX64
" grranted="LPX64
"\n",
822 imp
->imp_obd
->obd_name
,imp
->imp_connect_flags_orig
,
823 ocd
->ocd_connect_flags
);
824 GOTO(out
, rc
= -EPROTO
);
828 /* This could happen if export is cleaned during the
830 CERROR("%s: missing export after connect\n",
831 imp
->imp_obd
->obd_name
);
832 GOTO(out
, rc
= -ENODEV
);
834 old_connect_flags
= exp_connect_flags(exp
);
835 exp
->exp_connect_data
= *ocd
;
836 imp
->imp_obd
->obd_self_export
->exp_connect_data
= *ocd
;
837 class_export_put(exp
);
839 obd_import_event(imp
->imp_obd
, imp
, IMP_EVENT_OCD
);
841 if (aa
->pcaa_initial_connect
) {
842 spin_lock(&imp
->imp_lock
);
843 if (msg_flags
& MSG_CONNECT_REPLAYABLE
) {
844 imp
->imp_replayable
= 1;
845 spin_unlock(&imp
->imp_lock
);
846 CDEBUG(D_HA
, "connected to replayable target: %s\n",
847 obd2cli_tgt(imp
->imp_obd
));
849 imp
->imp_replayable
= 0;
850 spin_unlock(&imp
->imp_lock
);
853 /* if applies, adjust the imp->imp_msg_magic here
854 * according to reply flags */
856 imp
->imp_remote_handle
=
857 *lustre_msg_get_handle(request
->rq_repmsg
);
859 /* Initial connects are allowed for clients with non-random
860 * uuids when servers are in recovery. Simply signal the
861 * servers replay is complete and wait in REPLAY_WAIT. */
862 if (msg_flags
& MSG_CONNECT_RECOVERING
) {
863 CDEBUG(D_HA
, "connect to %s during recovery\n",
864 obd2cli_tgt(imp
->imp_obd
));
865 IMPORT_SET_STATE(imp
, LUSTRE_IMP_REPLAY_LOCKS
);
867 IMPORT_SET_STATE(imp
, LUSTRE_IMP_FULL
);
868 ptlrpc_activate_import(imp
);
871 GOTO(finish
, rc
= 0);
874 /* Determine what recovery state to move the import to. */
875 if (MSG_CONNECT_RECONNECT
& msg_flags
) {
876 memset(&old_hdl
, 0, sizeof(old_hdl
));
877 if (!memcmp(&old_hdl
, lustre_msg_get_handle(request
->rq_repmsg
),
879 LCONSOLE_WARN("Reconnect to %s (at @%s) failed due "
880 "bad handle "LPX64
"\n",
881 obd2cli_tgt(imp
->imp_obd
),
882 imp
->imp_connection
->c_remote_uuid
.uuid
,
883 imp
->imp_dlm_handle
.cookie
);
884 GOTO(out
, rc
= -ENOTCONN
);
887 if (memcmp(&imp
->imp_remote_handle
,
888 lustre_msg_get_handle(request
->rq_repmsg
),
889 sizeof(imp
->imp_remote_handle
))) {
890 int level
= msg_flags
& MSG_CONNECT_RECOVERING
?
893 /* Bug 16611/14775: if server handle have changed,
894 * that means some sort of disconnection happened.
895 * If the server is not in recovery, that also means it
896 * already erased all of our state because of previous
897 * eviction. If it is in recovery - we are safe to
898 * participate since we can reestablish all of our state
899 * with server again */
900 if ((MSG_CONNECT_RECOVERING
& msg_flags
)) {
901 CDEBUG(level
,"%s@%s changed server handle from "
903 " but is still in recovery\n",
904 obd2cli_tgt(imp
->imp_obd
),
905 imp
->imp_connection
->c_remote_uuid
.uuid
,
906 imp
->imp_remote_handle
.cookie
,
907 lustre_msg_get_handle(
908 request
->rq_repmsg
)->cookie
);
910 LCONSOLE_WARN("Evicted from %s (at %s) "
911 "after server handle changed from "
912 LPX64
" to "LPX64
"\n",
913 obd2cli_tgt(imp
->imp_obd
),
914 imp
->imp_connection
-> \
916 imp
->imp_remote_handle
.cookie
,
917 lustre_msg_get_handle(
918 request
->rq_repmsg
)->cookie
);
922 imp
->imp_remote_handle
=
923 *lustre_msg_get_handle(request
->rq_repmsg
);
925 if (!(MSG_CONNECT_RECOVERING
& msg_flags
)) {
926 IMPORT_SET_STATE(imp
, LUSTRE_IMP_EVICTED
);
927 GOTO(finish
, rc
= 0);
931 CDEBUG(D_HA
, "reconnected to %s@%s after partition\n",
932 obd2cli_tgt(imp
->imp_obd
),
933 imp
->imp_connection
->c_remote_uuid
.uuid
);
936 if (imp
->imp_invalid
) {
937 CDEBUG(D_HA
, "%s: reconnected but import is invalid; "
938 "marking evicted\n", imp
->imp_obd
->obd_name
);
939 IMPORT_SET_STATE(imp
, LUSTRE_IMP_EVICTED
);
940 } else if (MSG_CONNECT_RECOVERING
& msg_flags
) {
941 CDEBUG(D_HA
, "%s: reconnected to %s during replay\n",
942 imp
->imp_obd
->obd_name
,
943 obd2cli_tgt(imp
->imp_obd
));
945 spin_lock(&imp
->imp_lock
);
946 imp
->imp_resend_replay
= 1;
947 spin_unlock(&imp
->imp_lock
);
949 IMPORT_SET_STATE(imp
, LUSTRE_IMP_REPLAY
);
951 IMPORT_SET_STATE(imp
, LUSTRE_IMP_RECOVER
);
953 } else if ((MSG_CONNECT_RECOVERING
& msg_flags
) && !imp
->imp_invalid
) {
954 LASSERT(imp
->imp_replayable
);
955 imp
->imp_remote_handle
=
956 *lustre_msg_get_handle(request
->rq_repmsg
);
957 imp
->imp_last_replay_transno
= 0;
958 IMPORT_SET_STATE(imp
, LUSTRE_IMP_REPLAY
);
960 DEBUG_REQ(D_HA
, request
, "%s: evicting (reconnect/recover flags"
961 " not set: %x)", imp
->imp_obd
->obd_name
, msg_flags
);
962 imp
->imp_remote_handle
=
963 *lustre_msg_get_handle(request
->rq_repmsg
);
964 IMPORT_SET_STATE(imp
, LUSTRE_IMP_EVICTED
);
967 /* Sanity checks for a reconnected import. */
968 if (!(imp
->imp_replayable
) != !(msg_flags
& MSG_CONNECT_REPLAYABLE
)) {
969 CERROR("imp_replayable flag does not match server "
970 "after reconnect. We should LBUG right here.\n");
973 if (lustre_msg_get_last_committed(request
->rq_repmsg
) > 0 &&
974 lustre_msg_get_last_committed(request
->rq_repmsg
) <
975 aa
->pcaa_peer_committed
) {
976 CERROR("%s went back in time (transno "LPD64
977 " was previously committed, server now claims "LPD64
978 ")! See https://bugzilla.lustre.org/show_bug.cgi?"
980 obd2cli_tgt(imp
->imp_obd
), aa
->pcaa_peer_committed
,
981 lustre_msg_get_last_committed(request
->rq_repmsg
));
985 rc
= ptlrpc_import_recovery_state_machine(imp
);
987 if (rc
== -ENOTCONN
) {
988 CDEBUG(D_HA
, "evicted/aborted by %s@%s during recovery;"
989 "invalidating and reconnecting\n",
990 obd2cli_tgt(imp
->imp_obd
),
991 imp
->imp_connection
->c_remote_uuid
.uuid
);
992 ptlrpc_connect_import(imp
);
993 imp
->imp_connect_tried
= 1;
998 spin_lock(&imp
->imp_lock
);
999 list_del(&imp
->imp_conn_current
->oic_item
);
1000 list_add(&imp
->imp_conn_current
->oic_item
,
1001 &imp
->imp_conn_list
);
1002 imp
->imp_last_success_conn
=
1003 imp
->imp_conn_current
->oic_last_attempt
;
1005 spin_unlock(&imp
->imp_lock
);
1007 if (!ocd
->ocd_ibits_known
&&
1008 ocd
->ocd_connect_flags
& OBD_CONNECT_IBITS
)
1009 CERROR("Inodebits aware server returned zero compatible"
1012 if ((ocd
->ocd_connect_flags
& OBD_CONNECT_VERSION
) &&
1013 (ocd
->ocd_version
> LUSTRE_VERSION_CODE
+
1014 LUSTRE_VERSION_OFFSET_WARN
||
1015 ocd
->ocd_version
< LUSTRE_VERSION_CODE
-
1016 LUSTRE_VERSION_OFFSET_WARN
)) {
1017 /* Sigh, some compilers do not like #ifdef in the middle
1018 of macro arguments */
1019 const char *older
= "older. Consider upgrading server "
1020 "or downgrading client";
1021 const char *newer
= "newer than client version. "
1022 "Consider upgrading client";
1024 LCONSOLE_WARN("Server %s version (%d.%d.%d.%d) "
1025 "is much %s (%s)\n",
1026 obd2cli_tgt(imp
->imp_obd
),
1027 OBD_OCD_VERSION_MAJOR(ocd
->ocd_version
),
1028 OBD_OCD_VERSION_MINOR(ocd
->ocd_version
),
1029 OBD_OCD_VERSION_PATCH(ocd
->ocd_version
),
1030 OBD_OCD_VERSION_FIX(ocd
->ocd_version
),
1031 ocd
->ocd_version
> LUSTRE_VERSION_CODE
?
1032 newer
: older
, LUSTRE_VERSION_STRING
);
1035 #if LUSTRE_VERSION_CODE < OBD_OCD_VERSION(3, 2, 50, 0)
1036 /* Check if server has LU-1252 fix applied to not always swab
1037 * the IR MNE entries. Do this only once per connection. This
1038 * fixup is version-limited, because we don't want to carry the
1039 * OBD_CONNECT_MNE_SWAB flag around forever, just so long as we
1040 * need interop with unpatched 2.2 servers. For newer servers,
1041 * the client will do MNE swabbing only as needed. LU-1644 */
1042 if (unlikely((ocd
->ocd_connect_flags
& OBD_CONNECT_VERSION
) &&
1043 !(ocd
->ocd_connect_flags
& OBD_CONNECT_MNE_SWAB
) &&
1044 OBD_OCD_VERSION_MAJOR(ocd
->ocd_version
) == 2 &&
1045 OBD_OCD_VERSION_MINOR(ocd
->ocd_version
) == 2 &&
1046 OBD_OCD_VERSION_PATCH(ocd
->ocd_version
) < 55 &&
1047 strcmp(imp
->imp_obd
->obd_type
->typ_name
,
1048 LUSTRE_MGC_NAME
) == 0))
1049 imp
->imp_need_mne_swab
= 1;
1050 else /* clear if server was upgraded since last connect */
1051 imp
->imp_need_mne_swab
= 0;
1053 #warning "LU-1644: Remove old OBD_CONNECT_MNE_SWAB fixup and imp_need_mne_swab"
1056 if (ocd
->ocd_connect_flags
& OBD_CONNECT_CKSUM
) {
1057 /* We sent to the server ocd_cksum_types with bits set
1058 * for algorithms we understand. The server masked off
1059 * the checksum types it doesn't support */
1060 if ((ocd
->ocd_cksum_types
&
1061 cksum_types_supported_client()) == 0) {
1062 LCONSOLE_WARN("The negotiation of the checksum "
1063 "algorithm to use with server %s "
1064 "failed (%x/%x), disabling "
1066 obd2cli_tgt(imp
->imp_obd
),
1067 ocd
->ocd_cksum_types
,
1068 cksum_types_supported_client());
1069 cli
->cl_checksum
= 0;
1070 cli
->cl_supp_cksum_types
= OBD_CKSUM_ADLER
;
1072 cli
->cl_supp_cksum_types
= ocd
->ocd_cksum_types
;
1075 /* The server does not support OBD_CONNECT_CKSUM.
1076 * Enforce ADLER for backward compatibility*/
1077 cli
->cl_supp_cksum_types
= OBD_CKSUM_ADLER
;
1079 cli
->cl_cksum_type
=cksum_type_select(cli
->cl_supp_cksum_types
);
1081 if (ocd
->ocd_connect_flags
& OBD_CONNECT_BRW_SIZE
)
1082 cli
->cl_max_pages_per_rpc
=
1083 min(ocd
->ocd_brw_size
>> PAGE_CACHE_SHIFT
,
1084 cli
->cl_max_pages_per_rpc
);
1085 else if (imp
->imp_connect_op
== MDS_CONNECT
||
1086 imp
->imp_connect_op
== MGS_CONNECT
)
1087 cli
->cl_max_pages_per_rpc
= 1;
1089 /* Reset ns_connect_flags only for initial connect. It might be
1090 * changed in while using FS and if we reset it in reconnect
1091 * this leads to losing user settings done before such as
1092 * disable lru_resize, etc. */
1093 if (old_connect_flags
!= exp_connect_flags(exp
) ||
1094 aa
->pcaa_initial_connect
) {
1095 CDEBUG(D_HA
, "%s: Resetting ns_connect_flags to server "
1096 "flags: "LPX64
"\n", imp
->imp_obd
->obd_name
,
1097 ocd
->ocd_connect_flags
);
1098 imp
->imp_obd
->obd_namespace
->ns_connect_flags
=
1099 ocd
->ocd_connect_flags
;
1100 imp
->imp_obd
->obd_namespace
->ns_orig_connect_flags
=
1101 ocd
->ocd_connect_flags
;
1104 if ((ocd
->ocd_connect_flags
& OBD_CONNECT_AT
) &&
1105 (imp
->imp_msg_magic
== LUSTRE_MSG_MAGIC_V2
))
1106 /* We need a per-message support flag, because
1107 a. we don't know if the incoming connect reply
1108 supports AT or not (in reply_in_callback)
1110 b. failovered server means export and flags are gone
1111 (in ptlrpc_send_reply).
1112 Can only be set when we know AT is supported at
1114 imp
->imp_msghdr_flags
|= MSGHDR_AT_SUPPORT
;
1116 imp
->imp_msghdr_flags
&= ~MSGHDR_AT_SUPPORT
;
1118 if ((ocd
->ocd_connect_flags
& OBD_CONNECT_FULL20
) &&
1119 (imp
->imp_msg_magic
== LUSTRE_MSG_MAGIC_V2
))
1120 imp
->imp_msghdr_flags
|= MSGHDR_CKSUM_INCOMPAT18
;
1122 imp
->imp_msghdr_flags
&= ~MSGHDR_CKSUM_INCOMPAT18
;
1124 LASSERT((cli
->cl_max_pages_per_rpc
<= PTLRPC_MAX_BRW_PAGES
) &&
1125 (cli
->cl_max_pages_per_rpc
> 0));
1129 imp
->imp_connect_tried
= 1;
1132 IMPORT_SET_STATE(imp
, LUSTRE_IMP_DISCON
);
1133 if (rc
== -EACCES
) {
1135 * Give up trying to reconnect
1136 * EACCES means client has no permission for connection
1138 imp
->imp_obd
->obd_no_recov
= 1;
1139 ptlrpc_deactivate_import(imp
);
1142 if (rc
== -EPROTO
) {
1143 struct obd_connect_data
*ocd
;
1145 /* reply message might not be ready */
1146 if (request
->rq_repmsg
== NULL
)
1149 ocd
= req_capsule_server_get(&request
->rq_pill
,
1152 (ocd
->ocd_connect_flags
& OBD_CONNECT_VERSION
) &&
1153 (ocd
->ocd_version
!= LUSTRE_VERSION_CODE
)) {
1155 * Actually servers are only supposed to refuse
1156 * connection from liblustre clients, so we
1157 * should never see this from VFS context
1159 LCONSOLE_ERROR_MSG(0x16a, "Server %s version "
1161 " refused connection from this client "
1162 "with an incompatible version (%s). "
1163 "Client must be recompiled\n",
1164 obd2cli_tgt(imp
->imp_obd
),
1165 OBD_OCD_VERSION_MAJOR(ocd
->ocd_version
),
1166 OBD_OCD_VERSION_MINOR(ocd
->ocd_version
),
1167 OBD_OCD_VERSION_PATCH(ocd
->ocd_version
),
1168 OBD_OCD_VERSION_FIX(ocd
->ocd_version
),
1169 LUSTRE_VERSION_STRING
);
1170 ptlrpc_deactivate_import(imp
);
1171 IMPORT_SET_STATE(imp
, LUSTRE_IMP_CLOSED
);
1176 ptlrpc_maybe_ping_import_soon(imp
);
1178 CDEBUG(D_HA
, "recovery of %s on %s failed (%d)\n",
1179 obd2cli_tgt(imp
->imp_obd
),
1180 (char *)imp
->imp_connection
->c_remote_uuid
.uuid
, rc
);
1183 wake_up_all(&imp
->imp_recovery_waitq
);
1188 * interpret callback for "completed replay" RPCs.
1189 * \see signal_completed_replay
1191 static int completed_replay_interpret(const struct lu_env
*env
,
1192 struct ptlrpc_request
*req
,
1193 void * data
, int rc
)
1195 atomic_dec(&req
->rq_import
->imp_replay_inflight
);
1196 if (req
->rq_status
== 0 &&
1197 !req
->rq_import
->imp_vbr_failed
) {
1198 ptlrpc_import_recovery_state_machine(req
->rq_import
);
1200 if (req
->rq_import
->imp_vbr_failed
) {
1202 "%s: version recovery fails, reconnecting\n",
1203 req
->rq_import
->imp_obd
->obd_name
);
1205 CDEBUG(D_HA
, "%s: LAST_REPLAY message error: %d, "
1207 req
->rq_import
->imp_obd
->obd_name
,
1210 ptlrpc_connect_import(req
->rq_import
);
1217 * Let server know that we have no requests to replay anymore.
1218 * Achieved by just sending a PING request
1220 static int signal_completed_replay(struct obd_import
*imp
)
1222 struct ptlrpc_request
*req
;
1224 if (unlikely(OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_FINISH_REPLAY
)))
1227 LASSERT(atomic_read(&imp
->imp_replay_inflight
) == 0);
1228 atomic_inc(&imp
->imp_replay_inflight
);
1230 req
= ptlrpc_request_alloc_pack(imp
, &RQF_OBD_PING
, LUSTRE_OBD_VERSION
,
1233 atomic_dec(&imp
->imp_replay_inflight
);
1237 ptlrpc_request_set_replen(req
);
1238 req
->rq_send_state
= LUSTRE_IMP_REPLAY_WAIT
;
1239 lustre_msg_add_flags(req
->rq_reqmsg
,
1240 MSG_LOCK_REPLAY_DONE
| MSG_REQ_REPLAY_DONE
);
1242 req
->rq_timeout
*= 3;
1243 req
->rq_interpret_reply
= completed_replay_interpret
;
1245 ptlrpcd_add_req(req
, PDL_POLICY_ROUND
, -1);
1250 * In kernel code all import invalidation happens in its own
1251 * separate thread, so that whatever application happened to encounter
1252 * a problem could still be killed or otherwise continue
1254 static int ptlrpc_invalidate_import_thread(void *data
)
1256 struct obd_import
*imp
= data
;
1258 unshare_fs_struct();
1260 CDEBUG(D_HA
, "thread invalidate import %s to %s@%s\n",
1261 imp
->imp_obd
->obd_name
, obd2cli_tgt(imp
->imp_obd
),
1262 imp
->imp_connection
->c_remote_uuid
.uuid
);
1264 ptlrpc_invalidate_import(imp
);
1266 if (obd_dump_on_eviction
) {
1267 CERROR("dump the log upon eviction\n");
1268 libcfs_debug_dumplog();
1271 IMPORT_SET_STATE(imp
, LUSTRE_IMP_RECOVER
);
1272 ptlrpc_import_recovery_state_machine(imp
);
1274 class_import_put(imp
);
1279 * This is the state machine for client-side recovery on import.
1281 * Typically we have two possibly paths. If we came to server and it is not
1282 * in recovery, we just enter IMP_EVICTED state, invalidate our import
1283 * state and reconnect from scratch.
1284 * If we came to server that is in recovery, we enter IMP_REPLAY import state.
1285 * We go through our list of requests to replay and send them to server one by
1287 * After sending all request from the list we change import state to
1288 * IMP_REPLAY_LOCKS and re-request all the locks we believe we have from server
1289 * and also all the locks we don't yet have and wait for server to grant us.
1290 * After that we send a special "replay completed" request and change import
1291 * state to IMP_REPLAY_WAIT.
1292 * Upon receiving reply to that "replay completed" RPC we enter IMP_RECOVER
1293 * state and resend all requests from sending list.
1294 * After that we promote import to FULL state and send all delayed requests
1295 * and import is fully operational after that.
1298 int ptlrpc_import_recovery_state_machine(struct obd_import
*imp
)
1305 if (imp
->imp_state
== LUSTRE_IMP_EVICTED
) {
1306 deuuidify(obd2cli_tgt(imp
->imp_obd
), NULL
,
1307 &target_start
, &target_len
);
1308 /* Don't care about MGC eviction */
1309 if (strcmp(imp
->imp_obd
->obd_type
->typ_name
,
1310 LUSTRE_MGC_NAME
) != 0) {
1311 LCONSOLE_ERROR_MSG(0x167, "%s: This client was evicted "
1312 "by %.*s; in progress operations "
1313 "using this service will fail.\n",
1314 imp
->imp_obd
->obd_name
, target_len
,
1317 CDEBUG(D_HA
, "evicted from %s@%s; invalidating\n",
1318 obd2cli_tgt(imp
->imp_obd
),
1319 imp
->imp_connection
->c_remote_uuid
.uuid
);
1320 /* reset vbr_failed flag upon eviction */
1321 spin_lock(&imp
->imp_lock
);
1322 imp
->imp_vbr_failed
= 0;
1323 spin_unlock(&imp
->imp_lock
);
1326 struct task_struct
*task
;
1327 /* bug 17802: XXX client_disconnect_export vs connect request
1328 * race. if client will evicted at this time, we start
1329 * invalidate thread without reference to import and import can
1330 * be freed at same time. */
1331 class_import_get(imp
);
1332 task
= kthread_run(ptlrpc_invalidate_import_thread
, imp
,
1335 class_import_put(imp
);
1336 CERROR("error starting invalidate thread: %d\n", rc
);
1345 if (imp
->imp_state
== LUSTRE_IMP_REPLAY
) {
1346 CDEBUG(D_HA
, "replay requested by %s\n",
1347 obd2cli_tgt(imp
->imp_obd
));
1348 rc
= ptlrpc_replay_next(imp
, &inflight
);
1349 if (inflight
== 0 &&
1350 atomic_read(&imp
->imp_replay_inflight
) == 0) {
1351 IMPORT_SET_STATE(imp
, LUSTRE_IMP_REPLAY_LOCKS
);
1352 rc
= ldlm_replay_locks(imp
);
1359 if (imp
->imp_state
== LUSTRE_IMP_REPLAY_LOCKS
) {
1360 if (atomic_read(&imp
->imp_replay_inflight
) == 0) {
1361 IMPORT_SET_STATE(imp
, LUSTRE_IMP_REPLAY_WAIT
);
1362 rc
= signal_completed_replay(imp
);
1369 if (imp
->imp_state
== LUSTRE_IMP_REPLAY_WAIT
) {
1370 if (atomic_read(&imp
->imp_replay_inflight
) == 0) {
1371 IMPORT_SET_STATE(imp
, LUSTRE_IMP_RECOVER
);
1375 if (imp
->imp_state
== LUSTRE_IMP_RECOVER
) {
1376 CDEBUG(D_HA
, "reconnected to %s@%s\n",
1377 obd2cli_tgt(imp
->imp_obd
),
1378 imp
->imp_connection
->c_remote_uuid
.uuid
);
1380 rc
= ptlrpc_resend(imp
);
1383 IMPORT_SET_STATE(imp
, LUSTRE_IMP_FULL
);
1384 ptlrpc_activate_import(imp
);
1386 deuuidify(obd2cli_tgt(imp
->imp_obd
), NULL
,
1387 &target_start
, &target_len
);
1388 LCONSOLE_INFO("%s: Connection restored to %.*s (at %s)\n",
1389 imp
->imp_obd
->obd_name
,
1390 target_len
, target_start
,
1391 libcfs_nid2str(imp
->imp_connection
->c_peer
.nid
));
1394 if (imp
->imp_state
== LUSTRE_IMP_FULL
) {
1395 wake_up_all(&imp
->imp_recovery_waitq
);
1396 ptlrpc_wake_delayed(imp
);
1403 int ptlrpc_disconnect_import(struct obd_import
*imp
, int noclose
)
1405 struct ptlrpc_request
*req
;
1408 if (imp
->imp_obd
->obd_force
)
1409 GOTO(set_state
, rc
);
1411 switch (imp
->imp_connect_op
) {
1413 rq_opc
= OST_DISCONNECT
;
1416 rq_opc
= MDS_DISCONNECT
;
1419 rq_opc
= MGS_DISCONNECT
;
1423 CERROR("%s: don't know how to disconnect from %s "
1424 "(connect_op %d): rc = %d\n",
1425 imp
->imp_obd
->obd_name
, obd2cli_tgt(imp
->imp_obd
),
1426 imp
->imp_connect_op
, rc
);
1430 if (ptlrpc_import_in_recovery(imp
)) {
1431 struct l_wait_info lwi
;
1432 cfs_duration_t timeout
;
1435 if (imp
->imp_server_timeout
)
1436 timeout
= cfs_time_seconds(obd_timeout
/ 2);
1438 timeout
= cfs_time_seconds(obd_timeout
);
1440 int idx
= import_at_get_index(imp
,
1441 imp
->imp_client
->cli_request_portal
);
1442 timeout
= cfs_time_seconds(
1443 at_get(&imp
->imp_at
.iat_service_estimate
[idx
]));
1446 lwi
= LWI_TIMEOUT_INTR(cfs_timeout_cap(timeout
),
1447 back_to_sleep
, LWI_ON_SIGNAL_NOOP
, NULL
);
1448 rc
= l_wait_event(imp
->imp_recovery_waitq
,
1449 !ptlrpc_import_in_recovery(imp
), &lwi
);
1453 spin_lock(&imp
->imp_lock
);
1454 if (imp
->imp_state
!= LUSTRE_IMP_FULL
)
1456 spin_unlock(&imp
->imp_lock
);
1458 req
= ptlrpc_request_alloc_pack(imp
, &RQF_MDS_DISCONNECT
,
1459 LUSTRE_OBD_VERSION
, rq_opc
);
1461 /* We are disconnecting, do not retry a failed DISCONNECT rpc if
1462 * it fails. We can get through the above with a down server
1463 * if the client doesn't know the server is gone yet. */
1464 req
->rq_no_resend
= 1;
1466 /* We want client umounts to happen quickly, no matter the
1468 req
->rq_timeout
= min_t(int, req
->rq_timeout
,
1469 INITIAL_CONNECT_TIMEOUT
);
1471 IMPORT_SET_STATE(imp
, LUSTRE_IMP_CONNECTING
);
1472 req
->rq_send_state
= LUSTRE_IMP_CONNECTING
;
1473 ptlrpc_request_set_replen(req
);
1474 rc
= ptlrpc_queue_wait(req
);
1475 ptlrpc_req_finished(req
);
1479 spin_lock(&imp
->imp_lock
);
1482 IMPORT_SET_STATE_NOLOCK(imp
, LUSTRE_IMP_DISCON
);
1484 IMPORT_SET_STATE_NOLOCK(imp
, LUSTRE_IMP_CLOSED
);
1485 memset(&imp
->imp_remote_handle
, 0, sizeof(imp
->imp_remote_handle
));
1486 spin_unlock(&imp
->imp_lock
);
1488 if (rc
== -ETIMEDOUT
|| rc
== -ENOTCONN
|| rc
== -ESHUTDOWN
)
1493 EXPORT_SYMBOL(ptlrpc_disconnect_import
);
1495 void ptlrpc_cleanup_imp(struct obd_import
*imp
)
1497 spin_lock(&imp
->imp_lock
);
1498 IMPORT_SET_STATE_NOLOCK(imp
, LUSTRE_IMP_CLOSED
);
1499 imp
->imp_generation
++;
1500 spin_unlock(&imp
->imp_lock
);
1501 ptlrpc_abort_inflight(imp
);
1503 EXPORT_SYMBOL(ptlrpc_cleanup_imp
);
1505 /* Adaptive Timeout utils */
1506 extern unsigned int at_min
, at_max
, at_history
;
1508 /* Bin into timeslices using AT_BINS bins.
1509 This gives us a max of the last binlimit*AT_BINS secs without the storage,
1510 but still smoothing out a return to normalcy from a slow response.
1511 (E.g. remember the maximum latency in each minute of the last 4 minutes.) */
1512 int at_measured(struct adaptive_timeout
*at
, unsigned int val
)
1514 unsigned int old
= at
->at_current
;
1515 time_t now
= cfs_time_current_sec();
1516 time_t binlimit
= max_t(time_t, at_history
/ AT_BINS
, 1);
1519 CDEBUG(D_OTHER
, "add %u to %p time=%lu v=%u (%u %u %u %u)\n",
1520 val
, at
, now
- at
->at_binstart
, at
->at_current
,
1521 at
->at_hist
[0], at
->at_hist
[1], at
->at_hist
[2], at
->at_hist
[3]);
1524 /* 0's don't count, because we never want our timeout to
1525 drop to 0, and because 0 could mean an error */
1528 spin_lock(&at
->at_lock
);
1530 if (unlikely(at
->at_binstart
== 0)) {
1531 /* Special case to remove default from history */
1532 at
->at_current
= val
;
1533 at
->at_worst_ever
= val
;
1534 at
->at_worst_time
= now
;
1535 at
->at_hist
[0] = val
;
1536 at
->at_binstart
= now
;
1537 } else if (now
- at
->at_binstart
< binlimit
) {
1539 at
->at_hist
[0] = max(val
, at
->at_hist
[0]);
1540 at
->at_current
= max(val
, at
->at_current
);
1543 unsigned int maxv
= val
;
1544 /* move bins over */
1545 shift
= (now
- at
->at_binstart
) / binlimit
;
1547 for (i
= AT_BINS
- 1; i
>= 0; i
--) {
1549 at
->at_hist
[i
] = at
->at_hist
[i
- shift
];
1550 maxv
= max(maxv
, at
->at_hist
[i
]);
1555 at
->at_hist
[0] = val
;
1556 at
->at_current
= maxv
;
1557 at
->at_binstart
+= shift
* binlimit
;
1560 if (at
->at_current
> at
->at_worst_ever
) {
1561 at
->at_worst_ever
= at
->at_current
;
1562 at
->at_worst_time
= now
;
1565 if (at
->at_flags
& AT_FLG_NOHIST
)
1566 /* Only keep last reported val; keeping the rest of the history
1568 at
->at_current
= val
;
1571 at
->at_current
= min(at
->at_current
, at_max
);
1572 at
->at_current
= max(at
->at_current
, at_min
);
1574 if (at
->at_current
!= old
)
1575 CDEBUG(D_OTHER
, "AT %p change: old=%u new=%u delta=%d "
1576 "(val=%u) hist %u %u %u %u\n", at
,
1577 old
, at
->at_current
, at
->at_current
- old
, val
,
1578 at
->at_hist
[0], at
->at_hist
[1], at
->at_hist
[2],
1581 /* if we changed, report the old value */
1582 old
= (at
->at_current
!= old
) ? old
: 0;
1584 spin_unlock(&at
->at_lock
);
1588 /* Find the imp_at index for a given portal; assign if space available */
1589 int import_at_get_index(struct obd_import
*imp
, int portal
)
1591 struct imp_at
*at
= &imp
->imp_at
;
1594 for (i
= 0; i
< IMP_AT_MAX_PORTALS
; i
++) {
1595 if (at
->iat_portal
[i
] == portal
)
1597 if (at
->iat_portal
[i
] == 0)
1602 /* Not found in list, add it under a lock */
1603 spin_lock(&imp
->imp_lock
);
1605 /* Check unused under lock */
1606 for (; i
< IMP_AT_MAX_PORTALS
; i
++) {
1607 if (at
->iat_portal
[i
] == portal
)
1609 if (at
->iat_portal
[i
] == 0)
1614 /* Not enough portals? */
1615 LASSERT(i
< IMP_AT_MAX_PORTALS
);
1617 at
->iat_portal
[i
] = portal
;
1619 spin_unlock(&imp
->imp_lock
);