4 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
6 * This program is free software; you can redistribute it and/or modify
7 * it under the terms of the GNU General Public License version 2 only,
8 * as published by the Free Software Foundation.
10 * This program is distributed in the hope that it will be useful, but
11 * WITHOUT ANY WARRANTY; without even the implied warranty of
12 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
13 * General Public License version 2 for more details (a copy is included
14 * in the LICENSE file that accompanied this code).
16 * You should have received a copy of the GNU General Public License
17 * version 2 along with this program; If not, see
18 * http://www.gnu.org/licenses/gpl-2.0.html
23 * Copyright (c) 2003, 2010, Oracle and/or its affiliates. All rights reserved.
24 * Use is subject to license terms.
26 * Copyright (c) 2011, 2016, Intel Corporation.
29 * This file is part of Lustre, http://www.lustre.org/
30 * Lustre is a trademark of Sun Microsystems, Inc.
32 * lustre/ptlrpc/import.c
34 * Author: Mike Shaver <shaver@clusterfs.com>
37 #define DEBUG_SUBSYSTEM S_RPC
39 #include <linux/kthread.h>
40 #include <linux/delay.h>
41 #include <obd_support.h>
42 #include <lustre_ha.h>
43 #include <lustre_net.h>
44 #include <lustre_import.h>
45 #include <lustre_export.h>
47 #include <obd_cksum.h>
48 #include <obd_class.h>
50 #include "ptlrpc_internal.h"
52 struct ptlrpc_connect_async_args {
53 __u64 pcaa_peer_committed;
54 int pcaa_initial_connect;
58 * Updates import \a imp current state to provided \a state value
59 * Helper function. Must be called under imp_lock.
61 static void __import_set_state(struct obd_import *imp,
62 enum lustre_imp_state state)
65 case LUSTRE_IMP_CLOSED:
67 case LUSTRE_IMP_DISCON:
68 case LUSTRE_IMP_CONNECTING:
70 case LUSTRE_IMP_REPLAY_WAIT:
71 imp->imp_replay_state = LUSTRE_IMP_REPLAY_LOCKS;
74 imp->imp_replay_state = LUSTRE_IMP_REPLAY;
76 imp->imp_state = state;
77 imp->imp_state_hist[imp->imp_state_hist_idx].ish_state = state;
78 imp->imp_state_hist[imp->imp_state_hist_idx].ish_time =
79 ktime_get_real_seconds();
80 imp->imp_state_hist_idx = (imp->imp_state_hist_idx + 1) %
84 /* A CLOSED import should remain so. */
85 #define IMPORT_SET_STATE_NOLOCK(imp, state) \
87 if (imp->imp_state != LUSTRE_IMP_CLOSED) { \
88 CDEBUG(D_HA, "%p %s: changing import state from %s to %s\n", \
89 imp, obd2cli_tgt(imp->imp_obd), \
90 ptlrpc_import_state_name(imp->imp_state), \
91 ptlrpc_import_state_name(state)); \
92 __import_set_state(imp, state); \
96 #define IMPORT_SET_STATE(imp, state) \
98 spin_lock(&imp->imp_lock); \
99 IMPORT_SET_STATE_NOLOCK(imp, state); \
100 spin_unlock(&imp->imp_lock); \
103 void ptlrpc_import_enter_resend(struct obd_import *imp)
105 IMPORT_SET_STATE(imp, LUSTRE_IMP_RECOVER);
107 EXPORT_SYMBOL(ptlrpc_import_enter_resend);
110 static int ptlrpc_connect_interpret(const struct lu_env *env,
111 struct ptlrpc_request *request,
112 void * data, int rc);
113 int ptlrpc_import_recovery_state_machine(struct obd_import *imp);
115 /* Only this function is allowed to change the import state when it is
116 * CLOSED. I would rather refcount the import and free it after
117 * disconnection like we do with exports. To do that, the client_obd
118 * will need to save the peer info somewhere other than in the import,
120 int ptlrpc_init_import(struct obd_import *imp)
122 spin_lock(&imp->imp_lock);
124 imp->imp_generation++;
125 imp->imp_state = LUSTRE_IMP_NEW;
127 spin_unlock(&imp->imp_lock);
131 EXPORT_SYMBOL(ptlrpc_init_import);
133 #define UUID_STR "_UUID"
134 void deuuidify(char *uuid, const char *prefix, char **uuid_start, int *uuid_len)
136 *uuid_start = !prefix || strncmp(uuid, prefix, strlen(prefix))
137 ? uuid : uuid + strlen(prefix);
139 *uuid_len = strlen(*uuid_start);
141 if (*uuid_len < strlen(UUID_STR))
144 if (!strncmp(*uuid_start + *uuid_len - strlen(UUID_STR),
145 UUID_STR, strlen(UUID_STR)))
146 *uuid_len -= strlen(UUID_STR);
150 * Returns true if import was FULL, false if import was already not
152 * @imp - import to be disconnected
153 * @conn_cnt - connection count (epoch) of the request that timed out
154 * and caused the disconnection. In some cases, multiple
155 * inflight requests can fail to a single target (e.g. OST
156 * bulk requests) and if one has already caused a reconnection
157 * (increasing the import->conn_cnt) the older failure should
158 * not also cause a reconnection. If zero it forces a reconnect.
160 int ptlrpc_set_import_discon(struct obd_import *imp, __u32 conn_cnt)
164 spin_lock(&imp->imp_lock);
166 if (imp->imp_state == LUSTRE_IMP_FULL &&
167 (conn_cnt == 0 || conn_cnt == imp->imp_conn_cnt)) {
171 deuuidify(obd2cli_tgt(imp->imp_obd), NULL,
172 &target_start, &target_len);
174 if (imp->imp_replayable) {
175 LCONSOLE_WARN("%s: Connection to %.*s (at %s) was "
176 "lost; in progress operations using this "
177 "service will wait for recovery to complete\n",
178 imp->imp_obd->obd_name, target_len, target_start,
179 libcfs_nid2str(imp->imp_connection->c_peer.nid));
181 LCONSOLE_ERROR_MSG(0x166, "%s: Connection to "
182 "%.*s (at %s) was lost; in progress "
183 "operations using this service will fail\n",
184 imp->imp_obd->obd_name,
185 target_len, target_start,
186 libcfs_nid2str(imp->imp_connection->c_peer.nid));
188 IMPORT_SET_STATE_NOLOCK(imp, LUSTRE_IMP_DISCON);
189 spin_unlock(&imp->imp_lock);
191 if (obd_dump_on_timeout)
192 libcfs_debug_dumplog();
194 obd_import_event(imp->imp_obd, imp, IMP_EVENT_DISCON);
197 spin_unlock(&imp->imp_lock);
198 CDEBUG(D_HA, "%s: import %p already %s (conn %u, was %u): %s\n",
199 imp->imp_client->cli_name, imp,
200 (imp->imp_state == LUSTRE_IMP_FULL &&
201 imp->imp_conn_cnt > conn_cnt) ?
202 "reconnected" : "not connected", imp->imp_conn_cnt,
203 conn_cnt, ptlrpc_import_state_name(imp->imp_state));
209 /* Must be called with imp_lock held! */
210 static void ptlrpc_deactivate_and_unlock_import(struct obd_import *imp)
213 assert_spin_locked(&imp->imp_lock);
215 CDEBUG(D_HA, "setting import %s INVALID\n", obd2cli_tgt(imp->imp_obd));
216 imp->imp_invalid = 1;
217 imp->imp_generation++;
218 spin_unlock(&imp->imp_lock);
220 ptlrpc_abort_inflight(imp);
221 obd_import_event(imp->imp_obd, imp, IMP_EVENT_INACTIVE);
227 * This acts as a barrier; all existing requests are rejected, and
228 * no new requests will be accepted until the import is valid again.
230 void ptlrpc_deactivate_import(struct obd_import *imp)
232 spin_lock(&imp->imp_lock);
233 ptlrpc_deactivate_and_unlock_import(imp);
235 EXPORT_SYMBOL(ptlrpc_deactivate_import);
238 ptlrpc_inflight_deadline(struct ptlrpc_request *req, time64_t now)
242 if (!(((req->rq_phase == RQ_PHASE_RPC) && !req->rq_waiting) ||
243 (req->rq_phase == RQ_PHASE_BULK) ||
244 (req->rq_phase == RQ_PHASE_NEW)))
247 if (req->rq_timedout)
250 if (req->rq_phase == RQ_PHASE_NEW)
253 dl = req->rq_deadline;
261 static unsigned int ptlrpc_inflight_timeout(struct obd_import *imp)
263 time64_t now = ktime_get_real_seconds();
264 struct list_head *tmp, *n;
265 struct ptlrpc_request *req;
266 unsigned int timeout = 0;
268 spin_lock(&imp->imp_lock);
269 list_for_each_safe(tmp, n, &imp->imp_sending_list) {
270 req = list_entry(tmp, struct ptlrpc_request, rq_list);
271 timeout = max(ptlrpc_inflight_deadline(req, now), timeout);
273 spin_unlock(&imp->imp_lock);
278 * This function will invalidate the import, if necessary, then block
279 * for all the RPC completions, and finally notify the obd to
280 * invalidate its state (ie cancel locks, clear pending requests,
283 void ptlrpc_invalidate_import(struct obd_import *imp)
285 struct list_head *tmp, *n;
286 struct ptlrpc_request *req;
287 struct l_wait_info lwi;
288 unsigned int timeout;
291 atomic_inc(&imp->imp_inval_count);
293 if (!imp->imp_invalid || imp->imp_obd->obd_no_recov)
294 ptlrpc_deactivate_import(imp);
296 if (OBD_FAIL_PRECHECK(OBD_FAIL_PTLRPC_CONNECT_RACE)) {
297 OBD_RACE(OBD_FAIL_PTLRPC_CONNECT_RACE);
298 msleep(10 * MSEC_PER_SEC);
300 CFS_FAIL_TIMEOUT(OBD_FAIL_MGS_CONNECT_NET, 3 * cfs_fail_val / 2);
301 LASSERT(imp->imp_invalid);
303 /* Wait forever until inflight == 0. We really can't do it another
304 * way because in some cases we need to wait for very long reply
305 * unlink. We can't do anything before that because there is really
306 * no guarantee that some rdma transfer is not in progress right now. */
308 /* Calculate max timeout for waiting on rpcs to error
309 * out. Use obd_timeout if calculated value is smaller
311 if (!OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_REPL_UNLINK)) {
312 timeout = ptlrpc_inflight_timeout(imp);
313 timeout += timeout / 3;
316 timeout = obd_timeout;
318 /* decrease the interval to increase race condition */
322 CDEBUG(D_RPCTRACE,"Sleeping %d sec for inflight to error out\n",
325 /* Wait for all requests to error out and call completion
326 * callbacks. Cap it at obd_timeout -- these should all
327 * have been locally cancelled by ptlrpc_abort_inflight. */
328 lwi = LWI_TIMEOUT_INTERVAL(
329 cfs_timeout_cap(cfs_time_seconds(timeout)),
330 (timeout > 1)?cfs_time_seconds(1):cfs_time_seconds(1)/2,
332 rc = l_wait_event(imp->imp_recovery_waitq,
333 (atomic_read(&imp->imp_inflight) == 0),
336 const char *cli_tgt = obd2cli_tgt(imp->imp_obd);
338 CERROR("%s: rc = %d waiting for callback (%d != 0)\n",
339 cli_tgt, rc, atomic_read(&imp->imp_inflight));
341 spin_lock(&imp->imp_lock);
342 if (atomic_read(&imp->imp_inflight) == 0) {
343 int count = atomic_read(&imp->imp_unregistering);
345 /* We know that "unregistering" rpcs only can
346 * survive in sending or delaying lists (they
347 * maybe waiting for long reply unlink in
348 * sluggish nets). Let's check this. If there
349 * is no inflight and unregistering != 0, this
351 LASSERTF(count == 0, "Some RPCs are still "
352 "unregistering: %d\n", count);
354 /* Let's save one loop as soon as inflight have
355 * dropped to zero. No new inflights possible at
359 list_for_each_safe(tmp, n,
360 &imp->imp_sending_list) {
361 req = list_entry(tmp,
362 struct ptlrpc_request,
364 DEBUG_REQ(D_ERROR, req,
365 "still on sending list");
367 list_for_each_safe(tmp, n,
368 &imp->imp_delayed_list) {
369 req = list_entry(tmp,
370 struct ptlrpc_request,
372 DEBUG_REQ(D_ERROR, req,
373 "still on delayed list");
376 CERROR("%s: Unregistering RPCs found (%d). "
377 "Network is sluggish? Waiting them "
378 "to error out.\n", cli_tgt,
379 atomic_read(&imp->imp_unregistering));
381 spin_unlock(&imp->imp_lock);
386 * Let's additionally check that no new rpcs added to import in
387 * "invalidate" state.
389 LASSERT(atomic_read(&imp->imp_inflight) == 0);
390 obd_import_event(imp->imp_obd, imp, IMP_EVENT_INVALIDATE);
391 sptlrpc_import_flush_all_ctx(imp);
393 atomic_dec(&imp->imp_inval_count);
394 wake_up_all(&imp->imp_recovery_waitq);
396 EXPORT_SYMBOL(ptlrpc_invalidate_import);
398 /* unset imp_invalid */
399 void ptlrpc_activate_import(struct obd_import *imp)
401 struct obd_device *obd = imp->imp_obd;
403 spin_lock(&imp->imp_lock);
404 if (imp->imp_deactive != 0) {
405 spin_unlock(&imp->imp_lock);
409 imp->imp_invalid = 0;
410 spin_unlock(&imp->imp_lock);
411 obd_import_event(obd, imp, IMP_EVENT_ACTIVE);
413 EXPORT_SYMBOL(ptlrpc_activate_import);
415 void ptlrpc_pinger_force(struct obd_import *imp)
417 CDEBUG(D_HA, "%s: waking up pinger s:%s\n", obd2cli_tgt(imp->imp_obd),
418 ptlrpc_import_state_name(imp->imp_state));
420 spin_lock(&imp->imp_lock);
421 imp->imp_force_verify = 1;
422 spin_unlock(&imp->imp_lock);
424 if (imp->imp_state != LUSTRE_IMP_CONNECTING)
425 ptlrpc_pinger_wake_up();
427 EXPORT_SYMBOL(ptlrpc_pinger_force);
429 void ptlrpc_fail_import(struct obd_import *imp, __u32 conn_cnt)
433 LASSERT(!imp->imp_dlm_fake);
435 if (ptlrpc_set_import_discon(imp, conn_cnt)) {
436 if (!imp->imp_replayable) {
437 CDEBUG(D_HA, "import %s@%s for %s not replayable, "
438 "auto-deactivating\n",
439 obd2cli_tgt(imp->imp_obd),
440 imp->imp_connection->c_remote_uuid.uuid,
441 imp->imp_obd->obd_name);
442 ptlrpc_deactivate_import(imp);
445 ptlrpc_pinger_force(imp);
450 int ptlrpc_reconnect_import(struct obd_import *imp)
453 struct l_wait_info lwi;
454 int secs = cfs_time_seconds(obd_timeout);
457 ptlrpc_pinger_force(imp);
459 CDEBUG(D_HA, "%s: recovery started, waiting %u seconds\n",
460 obd2cli_tgt(imp->imp_obd), secs);
462 lwi = LWI_TIMEOUT(secs, NULL, NULL);
463 rc = l_wait_event(imp->imp_recovery_waitq,
464 !ptlrpc_import_in_recovery(imp), &lwi);
465 CDEBUG(D_HA, "%s: recovery finished s:%s\n", obd2cli_tgt(imp->imp_obd),
466 ptlrpc_import_state_name(imp->imp_state));
469 ptlrpc_set_import_discon(imp, 0);
470 /* Force a new connect attempt */
471 ptlrpc_invalidate_import(imp);
472 /* Do a fresh connect next time by zeroing the handle */
473 ptlrpc_disconnect_import(imp, 1);
474 /* Wait for all invalidate calls to finish */
475 if (atomic_read(&imp->imp_inval_count) > 0) {
476 struct l_wait_info lwi = LWI_INTR(LWI_ON_SIGNAL_NOOP, NULL);
479 rc = l_wait_event(imp->imp_recovery_waitq,
480 (atomic_read(&imp->imp_inval_count) == 0),
483 CERROR("Interrupted, inval=%d\n",
484 atomic_read(&imp->imp_inval_count));
487 /* Allow reconnect attempts */
488 imp->imp_obd->obd_no_recov = 0;
489 /* Remove 'invalid' flag */
490 ptlrpc_activate_import(imp);
491 /* Attempt a new connect */
492 ptlrpc_recover_import(imp, NULL, 0);
496 EXPORT_SYMBOL(ptlrpc_reconnect_import);
499 * Connection on import \a imp is changed to another one (if more than one is
500 * present). We typically chose connection that we have not tried to connect to
503 static int import_select_connection(struct obd_import *imp)
505 struct obd_import_conn *imp_conn = NULL, *conn;
506 struct obd_export *dlmexp;
508 int target_len, tried_all = 1;
511 spin_lock(&imp->imp_lock);
513 if (list_empty(&imp->imp_conn_list)) {
514 CERROR("%s: no connections available\n",
515 imp->imp_obd->obd_name);
516 spin_unlock(&imp->imp_lock);
520 list_for_each_entry(conn, &imp->imp_conn_list, oic_item) {
521 CDEBUG(D_HA, "%s: connect to NID %s last attempt %llu\n",
522 imp->imp_obd->obd_name,
523 libcfs_nid2str(conn->oic_conn->c_peer.nid),
524 conn->oic_last_attempt);
526 /* If we have not tried this connection since
527 the last successful attempt, go with this one */
528 if ((conn->oic_last_attempt == 0) ||
529 cfs_time_beforeq_64(conn->oic_last_attempt,
530 imp->imp_last_success_conn)) {
536 /* If all of the connections have already been tried
537 since the last successful connection; just choose the
538 least recently used */
541 else if (cfs_time_before_64(conn->oic_last_attempt,
542 imp_conn->oic_last_attempt))
546 /* if not found, simply choose the current one */
547 if (!imp_conn || imp->imp_force_reconnect) {
548 LASSERT(imp->imp_conn_current);
549 imp_conn = imp->imp_conn_current;
552 LASSERT(imp_conn->oic_conn);
554 /* If we've tried everything, and we're back to the beginning of the
555 list, increase our timeout and try again. It will be reset when
556 we do finally connect. (FIXME: really we should wait for all network
557 state associated with the last connection attempt to drain before
558 trying to reconnect on it.) */
559 if (tried_all && (imp->imp_conn_list.next == &imp_conn->oic_item)) {
560 struct adaptive_timeout *at = &imp->imp_at.iat_net_latency;
561 if (at_get(at) < CONNECTION_SWITCH_MAX) {
562 at_measured(at, at_get(at) + CONNECTION_SWITCH_INC);
563 if (at_get(at) > CONNECTION_SWITCH_MAX)
564 at_reset(at, CONNECTION_SWITCH_MAX);
566 LASSERT(imp_conn->oic_last_attempt);
567 CDEBUG(D_HA, "%s: tried all connections, increasing latency "
568 "to %ds\n", imp->imp_obd->obd_name, at_get(at));
571 imp_conn->oic_last_attempt = cfs_time_current_64();
573 /* switch connection, don't mind if it's same as the current one */
574 if (imp->imp_connection)
575 ptlrpc_connection_put(imp->imp_connection);
576 imp->imp_connection = ptlrpc_connection_addref(imp_conn->oic_conn);
578 dlmexp = class_conn2export(&imp->imp_dlm_handle);
579 LASSERT(dlmexp != NULL);
580 if (dlmexp->exp_connection)
581 ptlrpc_connection_put(dlmexp->exp_connection);
582 dlmexp->exp_connection = ptlrpc_connection_addref(imp_conn->oic_conn);
583 class_export_put(dlmexp);
585 if (imp->imp_conn_current != imp_conn) {
586 if (imp->imp_conn_current) {
587 deuuidify(obd2cli_tgt(imp->imp_obd), NULL,
588 &target_start, &target_len);
590 CDEBUG(D_HA, "%s: Connection changing to"
592 imp->imp_obd->obd_name,
593 target_len, target_start,
594 libcfs_nid2str(imp_conn->oic_conn->c_peer.nid));
597 imp->imp_conn_current = imp_conn;
600 CDEBUG(D_HA, "%s: import %p using connection %s/%s\n",
601 imp->imp_obd->obd_name, imp, imp_conn->oic_uuid.uuid,
602 libcfs_nid2str(imp_conn->oic_conn->c_peer.nid));
604 spin_unlock(&imp->imp_lock);
610 * must be called under imp_lock
612 static int ptlrpc_first_transno(struct obd_import *imp, __u64 *transno)
614 struct ptlrpc_request *req;
615 struct list_head *tmp;
617 /* The requests in committed_list always have smaller transnos than
618 * the requests in replay_list */
619 if (!list_empty(&imp->imp_committed_list)) {
620 tmp = imp->imp_committed_list.next;
621 req = list_entry(tmp, struct ptlrpc_request, rq_replay_list);
622 *transno = req->rq_transno;
623 if (req->rq_transno == 0) {
624 DEBUG_REQ(D_ERROR, req, "zero transno in committed_list");
629 if (!list_empty(&imp->imp_replay_list)) {
630 tmp = imp->imp_replay_list.next;
631 req = list_entry(tmp, struct ptlrpc_request, rq_replay_list);
632 *transno = req->rq_transno;
633 if (req->rq_transno == 0) {
634 DEBUG_REQ(D_ERROR, req, "zero transno in replay_list");
643 * Attempt to (re)connect import \a imp. This includes all preparations,
644 * initializing CONNECT RPC request and passing it to ptlrpcd for
646 * Returns 0 on success or error code.
648 int ptlrpc_connect_import(struct obd_import *imp)
650 struct obd_device *obd = imp->imp_obd;
651 int initial_connect = 0;
653 __u64 committed_before_reconnect = 0;
654 struct ptlrpc_request *request;
655 char *bufs[] = { NULL,
656 obd2cli_tgt(imp->imp_obd),
658 (char *)&imp->imp_dlm_handle,
659 (char *)&imp->imp_connect_data };
660 struct ptlrpc_connect_async_args *aa;
664 spin_lock(&imp->imp_lock);
665 if (imp->imp_state == LUSTRE_IMP_CLOSED) {
666 spin_unlock(&imp->imp_lock);
667 CERROR("can't connect to a closed import\n");
669 } else if (imp->imp_state == LUSTRE_IMP_FULL) {
670 spin_unlock(&imp->imp_lock);
671 CERROR("already connected\n");
673 } else if (imp->imp_state == LUSTRE_IMP_CONNECTING ||
674 imp->imp_state == LUSTRE_IMP_EVICTED ||
675 imp->imp_connected) {
676 spin_unlock(&imp->imp_lock);
677 CERROR("already connecting\n");
681 IMPORT_SET_STATE_NOLOCK(imp, LUSTRE_IMP_CONNECTING);
684 imp->imp_resend_replay = 0;
686 if (!lustre_handle_is_used(&imp->imp_remote_handle))
689 committed_before_reconnect = imp->imp_peer_committed_transno;
691 set_transno = ptlrpc_first_transno(imp,
692 &imp->imp_connect_data.ocd_transno);
693 spin_unlock(&imp->imp_lock);
695 rc = import_select_connection(imp);
699 rc = sptlrpc_import_sec_adapt(imp, NULL, NULL);
703 /* Reset connect flags to the originally requested flags, in case
704 * the server is updated on-the-fly we will get the new features. */
705 imp->imp_connect_data.ocd_connect_flags = imp->imp_connect_flags_orig;
706 imp->imp_connect_data.ocd_connect_flags2 = imp->imp_connect_flags2_orig;
707 /* Reset ocd_version each time so the server knows the exact versions */
708 imp->imp_connect_data.ocd_version = LUSTRE_VERSION_CODE;
709 imp->imp_msghdr_flags &= ~MSGHDR_AT_SUPPORT;
710 imp->imp_msghdr_flags &= ~MSGHDR_CKSUM_INCOMPAT18;
712 rc = obd_reconnect(NULL, imp->imp_obd->obd_self_export, obd,
713 &obd->obd_uuid, &imp->imp_connect_data, NULL);
717 request = ptlrpc_request_alloc(imp, &RQF_MDS_CONNECT);
719 GOTO(out, rc = -ENOMEM);
721 rc = ptlrpc_request_bufs_pack(request, LUSTRE_OBD_VERSION,
722 imp->imp_connect_op, bufs, NULL);
724 ptlrpc_request_free(request);
728 /* Report the rpc service time to the server so that it knows how long
729 * to wait for clients to join recovery */
730 lustre_msg_set_service_time(request->rq_reqmsg,
731 at_timeout2est(request->rq_timeout));
733 /* The amount of time we give the server to process the connect req.
734 * import_select_connection will increase the net latency on
735 * repeated reconnect attempts to cover slow networks.
736 * We override/ignore the server rpc completion estimate here,
737 * which may be large if this is a reconnect attempt */
738 request->rq_timeout = INITIAL_CONNECT_TIMEOUT;
739 lustre_msg_set_timeout(request->rq_reqmsg, request->rq_timeout);
741 request->rq_no_resend = request->rq_no_delay = 1;
742 request->rq_send_state = LUSTRE_IMP_CONNECTING;
743 /* Allow a slightly larger reply for future growth compatibility */
744 req_capsule_set_size(&request->rq_pill, &RMF_CONNECT_DATA, RCL_SERVER,
745 sizeof(struct obd_connect_data)+16*sizeof(__u64));
746 ptlrpc_request_set_replen(request);
747 request->rq_interpret_reply = ptlrpc_connect_interpret;
749 CLASSERT(sizeof(*aa) <= sizeof(request->rq_async_args));
750 aa = ptlrpc_req_async_args(request);
751 memset(aa, 0, sizeof *aa);
753 aa->pcaa_peer_committed = committed_before_reconnect;
754 aa->pcaa_initial_connect = initial_connect;
756 if (aa->pcaa_initial_connect) {
757 spin_lock(&imp->imp_lock);
758 imp->imp_replayable = 1;
759 spin_unlock(&imp->imp_lock);
760 lustre_msg_add_op_flags(request->rq_reqmsg,
761 MSG_CONNECT_INITIAL);
765 lustre_msg_add_op_flags(request->rq_reqmsg,
766 MSG_CONNECT_TRANSNO);
768 DEBUG_REQ(D_RPCTRACE, request, "(re)connect request (timeout %d)",
769 request->rq_timeout);
770 ptlrpcd_add_req(request);
774 IMPORT_SET_STATE(imp, LUSTRE_IMP_DISCON);
778 EXPORT_SYMBOL(ptlrpc_connect_import);
780 static void ptlrpc_maybe_ping_import_soon(struct obd_import *imp)
784 spin_lock(&imp->imp_lock);
785 force_verify = imp->imp_force_verify != 0;
786 spin_unlock(&imp->imp_lock);
789 ptlrpc_pinger_wake_up();
792 static int ptlrpc_busy_reconnect(int rc)
794 return (rc == -EBUSY) || (rc == -EAGAIN);
797 static int ptlrpc_connect_set_flags(struct obd_import *imp,
798 struct obd_connect_data *ocd,
799 __u64 old_connect_flags,
800 struct obd_export *exp, int init_connect)
803 struct client_obd *cli = &imp->imp_obd->u.cli;
805 spin_lock(&imp->imp_lock);
806 list_del(&imp->imp_conn_current->oic_item);
807 list_add(&imp->imp_conn_current->oic_item,
808 &imp->imp_conn_list);
809 imp->imp_last_success_conn =
810 imp->imp_conn_current->oic_last_attempt;
812 spin_unlock(&imp->imp_lock);
815 if (!warned && (ocd->ocd_connect_flags & OBD_CONNECT_VERSION) &&
816 (ocd->ocd_version > LUSTRE_VERSION_CODE +
817 LUSTRE_VERSION_OFFSET_WARN ||
818 ocd->ocd_version < LUSTRE_VERSION_CODE -
819 LUSTRE_VERSION_OFFSET_WARN)) {
820 /* Sigh, some compilers do not like #ifdef in the middle
821 of macro arguments */
822 const char *older = "older than client. "
823 "Consider upgrading server";
824 const char *newer = "newer than client. "
825 "Consider recompiling application";
827 LCONSOLE_WARN("Server %s version (%d.%d.%d.%d) "
829 obd2cli_tgt(imp->imp_obd),
830 OBD_OCD_VERSION_MAJOR(ocd->ocd_version),
831 OBD_OCD_VERSION_MINOR(ocd->ocd_version),
832 OBD_OCD_VERSION_PATCH(ocd->ocd_version),
833 OBD_OCD_VERSION_FIX(ocd->ocd_version),
834 ocd->ocd_version > LUSTRE_VERSION_CODE ?
835 newer : older, LUSTRE_VERSION_STRING);
839 #if LUSTRE_VERSION_CODE < OBD_OCD_VERSION(3, 0, 53, 0)
840 /* Check if server has LU-1252 fix applied to not always swab
841 * the IR MNE entries. Do this only once per connection. This
842 * fixup is version-limited, because we don't want to carry the
843 * OBD_CONNECT_MNE_SWAB flag around forever, just so long as we
844 * need interop with unpatched 2.2 servers. For newer servers,
845 * the client will do MNE swabbing only as needed. LU-1644 */
846 if (unlikely((ocd->ocd_connect_flags & OBD_CONNECT_VERSION) &&
847 !(ocd->ocd_connect_flags & OBD_CONNECT_MNE_SWAB) &&
848 OBD_OCD_VERSION_MAJOR(ocd->ocd_version) == 2 &&
849 OBD_OCD_VERSION_MINOR(ocd->ocd_version) == 2 &&
850 OBD_OCD_VERSION_PATCH(ocd->ocd_version) < 55 &&
851 strcmp(imp->imp_obd->obd_type->typ_name,
852 LUSTRE_MGC_NAME) == 0))
853 imp->imp_need_mne_swab = 1;
854 else /* clear if server was upgraded since last connect */
855 imp->imp_need_mne_swab = 0;
858 if (ocd->ocd_connect_flags & OBD_CONNECT_CKSUM) {
859 /* We sent to the server ocd_cksum_types with bits set
860 * for algorithms we understand. The server masked off
861 * the checksum types it doesn't support */
862 if ((ocd->ocd_cksum_types &
863 cksum_types_supported_client()) == 0) {
864 LCONSOLE_ERROR("The negotiation of the checksum "
865 "alogrithm to use with server %s "
867 obd2cli_tgt(imp->imp_obd),
868 ocd->ocd_cksum_types,
869 cksum_types_supported_client());
872 cli->cl_supp_cksum_types = ocd->ocd_cksum_types;
875 /* The server does not support OBD_CONNECT_CKSUM.
876 * Enforce ADLER for backward compatibility*/
877 cli->cl_supp_cksum_types = OBD_CKSUM_ADLER;
879 cli->cl_cksum_type = cksum_type_select(cli->cl_supp_cksum_types);
881 if (ocd->ocd_connect_flags & OBD_CONNECT_BRW_SIZE)
882 cli->cl_max_pages_per_rpc =
883 min(ocd->ocd_brw_size >> PAGE_SHIFT,
884 cli->cl_max_pages_per_rpc);
885 else if (imp->imp_connect_op == MDS_CONNECT ||
886 imp->imp_connect_op == MGS_CONNECT)
887 cli->cl_max_pages_per_rpc = 1;
889 LASSERT((cli->cl_max_pages_per_rpc <= PTLRPC_MAX_BRW_PAGES) &&
890 (cli->cl_max_pages_per_rpc > 0));
892 client_adjust_max_dirty(cli);
894 /* Update client max modify RPCs in flight with value returned
896 if (ocd->ocd_connect_flags & OBD_CONNECT_MULTIMODRPCS)
897 cli->cl_max_mod_rpcs_in_flight = min(
898 cli->cl_max_mod_rpcs_in_flight,
899 ocd->ocd_maxmodrpcs);
901 cli->cl_max_mod_rpcs_in_flight = 1;
903 /* Reset ns_connect_flags only for initial connect. It might be
904 * changed in while using FS and if we reset it in reconnect
905 * this leads to losing user settings done before such as
906 * disable lru_resize, etc. */
907 if (old_connect_flags != exp_connect_flags(exp) || init_connect) {
908 CDEBUG(D_HA, "%s: Resetting ns_connect_flags to server "
909 "flags: %#llx\n", imp->imp_obd->obd_name,
910 ocd->ocd_connect_flags);
911 imp->imp_obd->obd_namespace->ns_connect_flags =
912 ocd->ocd_connect_flags;
913 imp->imp_obd->obd_namespace->ns_orig_connect_flags =
914 ocd->ocd_connect_flags;
917 if (ocd->ocd_connect_flags & OBD_CONNECT_AT)
918 /* We need a per-message support flag, because
919 * a. we don't know if the incoming connect reply
920 * supports AT or not (in reply_in_callback)
921 * until we unpack it.
922 * b. failovered server means export and flags are gone
923 * (in ptlrpc_send_reply).
924 * Can only be set when we know AT is supported at
926 imp->imp_msghdr_flags |= MSGHDR_AT_SUPPORT;
928 imp->imp_msghdr_flags &= ~MSGHDR_AT_SUPPORT;
930 imp->imp_msghdr_flags |= MSGHDR_CKSUM_INCOMPAT18;
936 * Add all replay requests back to unreplied list before start replay,
937 * so that we can make sure the known replied XID is always increased
938 * only even if when replaying requests.
940 static void ptlrpc_prepare_replay(struct obd_import *imp)
942 struct ptlrpc_request *req;
944 if (imp->imp_state != LUSTRE_IMP_REPLAY ||
945 imp->imp_resend_replay)
948 /* If the server was restart during repaly, the requests may
949 * have been added to the unreplied list in former replay. */
950 spin_lock(&imp->imp_lock);
952 list_for_each_entry(req, &imp->imp_committed_list, rq_replay_list) {
953 if (list_empty(&req->rq_unreplied_list))
954 ptlrpc_add_unreplied(req);
957 list_for_each_entry(req, &imp->imp_replay_list, rq_replay_list) {
958 if (list_empty(&req->rq_unreplied_list))
959 ptlrpc_add_unreplied(req);
962 imp->imp_known_replied_xid = ptlrpc_known_replied_xid(imp);
963 spin_unlock(&imp->imp_lock);
967 * interpret_reply callback for connect RPCs.
968 * Looks into returned status of connect operation and decides
969 * what to do with the import - i.e enter recovery, promote it to
970 * full state for normal operations of disconnect it due to an error.
972 static int ptlrpc_connect_interpret(const struct lu_env *env,
973 struct ptlrpc_request *request,
976 struct ptlrpc_connect_async_args *aa = data;
977 struct obd_import *imp = request->rq_import;
978 struct lustre_handle old_hdl;
979 __u64 old_connect_flags;
981 struct obd_connect_data *ocd;
982 struct obd_export *exp = NULL;
986 spin_lock(&imp->imp_lock);
987 if (imp->imp_state == LUSTRE_IMP_CLOSED) {
988 imp->imp_connect_tried = 1;
989 spin_unlock(&imp->imp_lock);
994 /* if this reconnect to busy export - not need select new target
996 imp->imp_force_reconnect = ptlrpc_busy_reconnect(rc);
997 spin_unlock(&imp->imp_lock);
998 ptlrpc_maybe_ping_import_soon(imp);
1002 /* LU-7558: indicate that we are interpretting connect reply,
1003 * pltrpc_connect_import() will not try to reconnect until
1004 * interpret will finish. */
1005 imp->imp_connected = 1;
1006 spin_unlock(&imp->imp_lock);
1008 LASSERT(imp->imp_conn_current);
1010 msg_flags = lustre_msg_get_op_flags(request->rq_repmsg);
1012 ret = req_capsule_get_size(&request->rq_pill, &RMF_CONNECT_DATA,
1014 /* server replied obd_connect_data is always bigger */
1015 ocd = req_capsule_server_sized_get(&request->rq_pill,
1016 &RMF_CONNECT_DATA, ret);
1019 CERROR("%s: no connect data from server\n",
1020 imp->imp_obd->obd_name);
1025 spin_lock(&imp->imp_lock);
1027 /* All imports are pingable */
1028 imp->imp_pingable = 1;
1029 imp->imp_force_reconnect = 0;
1030 imp->imp_force_verify = 0;
1032 imp->imp_connect_data = *ocd;
1034 CDEBUG(D_HA, "%s: connect to target with instance %u\n",
1035 imp->imp_obd->obd_name, ocd->ocd_instance);
1036 exp = class_conn2export(&imp->imp_dlm_handle);
1038 spin_unlock(&imp->imp_lock);
1041 /* This could happen if export is cleaned during the
1043 CERROR("%s: missing export after connect\n",
1044 imp->imp_obd->obd_name);
1045 GOTO(out, rc = -ENODEV);
1048 /* check that server granted subset of flags we asked for. */
1049 if ((ocd->ocd_connect_flags & imp->imp_connect_flags_orig) !=
1050 ocd->ocd_connect_flags) {
1051 CERROR("%s: Server didn't grant requested subset of flags: "
1052 "asked=%#llx granted=%#llx\n",
1053 imp->imp_obd->obd_name, imp->imp_connect_flags_orig,
1054 ocd->ocd_connect_flags);
1055 GOTO(out, rc = -EPROTO);
1058 if ((ocd->ocd_connect_flags2 & imp->imp_connect_flags2_orig) !=
1059 ocd->ocd_connect_flags2) {
1060 CERROR("%s: Server didn't grant requested subset of flags2: "
1061 "asked=%#llx granted=%#llx\n",
1062 imp->imp_obd->obd_name, imp->imp_connect_flags2_orig,
1063 ocd->ocd_connect_flags2);
1064 GOTO(out, rc = -EPROTO);
1067 if (!(imp->imp_connect_flags_orig & OBD_CONNECT_LIGHTWEIGHT) &&
1068 (imp->imp_connect_flags_orig & OBD_CONNECT_MDS_MDS) &&
1069 (imp->imp_connect_flags_orig & OBD_CONNECT_FID) &&
1070 (ocd->ocd_connect_flags & OBD_CONNECT_VERSION)) {
1071 __u32 major = OBD_OCD_VERSION_MAJOR(ocd->ocd_version);
1072 __u32 minor = OBD_OCD_VERSION_MINOR(ocd->ocd_version);
1073 __u32 patch = OBD_OCD_VERSION_PATCH(ocd->ocd_version);
1075 /* We do not support the MDT-MDT interoperations with
1076 * different version MDT because of protocol changes. */
1077 if (unlikely(major != LUSTRE_MAJOR ||
1078 minor != LUSTRE_MINOR ||
1079 abs(patch - LUSTRE_PATCH) > 3)) {
1080 LCONSOLE_WARN("%s: import %p (%u.%u.%u.%u) tried the "
1081 "connection to different version MDT "
1082 "(%d.%d.%d.%d) %s\n",
1083 imp->imp_obd->obd_name, imp, LUSTRE_MAJOR,
1084 LUSTRE_MINOR, LUSTRE_PATCH, LUSTRE_FIX,
1085 major, minor, patch,
1086 OBD_OCD_VERSION_FIX(ocd->ocd_version),
1087 imp->imp_connection->c_remote_uuid.uuid);
1089 GOTO(out, rc = -EPROTO);
1093 old_connect_flags = exp_connect_flags(exp);
1094 exp->exp_connect_data = *ocd;
1095 imp->imp_obd->obd_self_export->exp_connect_data = *ocd;
1097 /* The net statistics after (re-)connect is not valid anymore,
1098 * because may reflect other routing, etc. */
1099 at_reinit(&imp->imp_at.iat_net_latency, 0, 0);
1100 ptlrpc_at_adj_net_latency(request,
1101 lustre_msg_get_service_time(request->rq_repmsg));
1103 /* Import flags should be updated before waking import at FULL state */
1104 rc = ptlrpc_connect_set_flags(imp, ocd, old_connect_flags, exp,
1105 aa->pcaa_initial_connect);
1106 class_export_put(exp);
1112 obd_import_event(imp->imp_obd, imp, IMP_EVENT_OCD);
1114 if (aa->pcaa_initial_connect) {
1115 spin_lock(&imp->imp_lock);
1116 if (msg_flags & MSG_CONNECT_REPLAYABLE) {
1117 imp->imp_replayable = 1;
1118 spin_unlock(&imp->imp_lock);
1119 CDEBUG(D_HA, "connected to replayable target: %s\n",
1120 obd2cli_tgt(imp->imp_obd));
1122 imp->imp_replayable = 0;
1123 spin_unlock(&imp->imp_lock);
1126 /* if applies, adjust the imp->imp_msg_magic here
1127 * according to reply flags */
1129 imp->imp_remote_handle =
1130 *lustre_msg_get_handle(request->rq_repmsg);
1132 /* Initial connects are allowed for clients with non-random
1133 * uuids when servers are in recovery. Simply signal the
1134 * servers replay is complete and wait in REPLAY_WAIT. */
1135 if (msg_flags & MSG_CONNECT_RECOVERING) {
1136 CDEBUG(D_HA, "connect to %s during recovery\n",
1137 obd2cli_tgt(imp->imp_obd));
1138 IMPORT_SET_STATE(imp, LUSTRE_IMP_REPLAY_LOCKS);
1140 IMPORT_SET_STATE(imp, LUSTRE_IMP_FULL);
1141 ptlrpc_activate_import(imp);
1144 GOTO(finish, rc = 0);
1147 /* Determine what recovery state to move the import to. */
1148 if (MSG_CONNECT_RECONNECT & msg_flags) {
1149 memset(&old_hdl, 0, sizeof(old_hdl));
1150 if (!memcmp(&old_hdl, lustre_msg_get_handle(request->rq_repmsg),
1151 sizeof (old_hdl))) {
1152 LCONSOLE_WARN("Reconnect to %s (at @%s) failed due "
1153 "bad handle %#llx\n",
1154 obd2cli_tgt(imp->imp_obd),
1155 imp->imp_connection->c_remote_uuid.uuid,
1156 imp->imp_dlm_handle.cookie);
1157 GOTO(out, rc = -ENOTCONN);
1160 if (memcmp(&imp->imp_remote_handle,
1161 lustre_msg_get_handle(request->rq_repmsg),
1162 sizeof(imp->imp_remote_handle))) {
1163 int level = msg_flags & MSG_CONNECT_RECOVERING ?
1166 /* Bug 16611/14775: if server handle have changed,
1167 * that means some sort of disconnection happened.
1168 * If the server is not in recovery, that also means it
1169 * already erased all of our state because of previous
1170 * eviction. If it is in recovery - we are safe to
1171 * participate since we can reestablish all of our state
1172 * with server again */
1173 if ((MSG_CONNECT_RECOVERING & msg_flags)) {
1174 CDEBUG(level,"%s@%s changed server handle from "
1176 " but is still in recovery\n",
1177 obd2cli_tgt(imp->imp_obd),
1178 imp->imp_connection->c_remote_uuid.uuid,
1179 imp->imp_remote_handle.cookie,
1180 lustre_msg_get_handle(
1181 request->rq_repmsg)->cookie);
1183 LCONSOLE_WARN("Evicted from %s (at %s) "
1184 "after server handle changed from "
1186 obd2cli_tgt(imp->imp_obd),
1187 imp->imp_connection-> \
1189 imp->imp_remote_handle.cookie,
1190 lustre_msg_get_handle(
1191 request->rq_repmsg)->cookie);
1195 imp->imp_remote_handle =
1196 *lustre_msg_get_handle(request->rq_repmsg);
1198 if (!(MSG_CONNECT_RECOVERING & msg_flags)) {
1199 IMPORT_SET_STATE(imp, LUSTRE_IMP_EVICTED);
1200 GOTO(finish, rc = 0);
1204 CDEBUG(D_HA, "reconnected to %s@%s after partition\n",
1205 obd2cli_tgt(imp->imp_obd),
1206 imp->imp_connection->c_remote_uuid.uuid);
1209 if (imp->imp_invalid) {
1210 CDEBUG(D_HA, "%s: reconnected but import is invalid; "
1211 "marking evicted\n", imp->imp_obd->obd_name);
1212 IMPORT_SET_STATE(imp, LUSTRE_IMP_EVICTED);
1213 } else if (MSG_CONNECT_RECOVERING & msg_flags) {
1214 CDEBUG(D_HA, "%s: reconnected to %s during replay\n",
1215 imp->imp_obd->obd_name,
1216 obd2cli_tgt(imp->imp_obd));
1218 spin_lock(&imp->imp_lock);
1219 imp->imp_resend_replay = 1;
1220 spin_unlock(&imp->imp_lock);
1222 IMPORT_SET_STATE(imp, imp->imp_replay_state);
1224 IMPORT_SET_STATE(imp, LUSTRE_IMP_RECOVER);
1226 } else if ((MSG_CONNECT_RECOVERING & msg_flags) && !imp->imp_invalid) {
1227 LASSERT(imp->imp_replayable);
1228 imp->imp_remote_handle =
1229 *lustre_msg_get_handle(request->rq_repmsg);
1230 imp->imp_last_replay_transno = 0;
1231 imp->imp_replay_cursor = &imp->imp_committed_list;
1232 IMPORT_SET_STATE(imp, LUSTRE_IMP_REPLAY);
1234 DEBUG_REQ(D_HA, request, "%s: evicting (reconnect/recover flags"
1235 " not set: %x)", imp->imp_obd->obd_name, msg_flags);
1236 imp->imp_remote_handle =
1237 *lustre_msg_get_handle(request->rq_repmsg);
1238 IMPORT_SET_STATE(imp, LUSTRE_IMP_EVICTED);
1241 /* Sanity checks for a reconnected import. */
1242 if (!(imp->imp_replayable) != !(msg_flags & MSG_CONNECT_REPLAYABLE)) {
1243 CERROR("imp_replayable flag does not match server "
1244 "after reconnect. We should LBUG right here.\n");
1247 if (lustre_msg_get_last_committed(request->rq_repmsg) > 0 &&
1248 lustre_msg_get_last_committed(request->rq_repmsg) <
1249 aa->pcaa_peer_committed) {
1250 CERROR("%s went back in time (transno %lld"
1251 " was previously committed, server now claims %lld"
1252 ")! See https://bugzilla.lustre.org/show_bug.cgi?"
1254 obd2cli_tgt(imp->imp_obd), aa->pcaa_peer_committed,
1255 lustre_msg_get_last_committed(request->rq_repmsg));
1259 ptlrpc_prepare_replay(imp);
1260 rc = ptlrpc_import_recovery_state_machine(imp);
1261 if (rc == -ENOTCONN) {
1262 CDEBUG(D_HA, "evicted/aborted by %s@%s during recovery;"
1263 "invalidating and reconnecting\n",
1264 obd2cli_tgt(imp->imp_obd),
1265 imp->imp_connection->c_remote_uuid.uuid);
1266 ptlrpc_connect_import(imp);
1267 spin_lock(&imp->imp_lock);
1268 imp->imp_connected = 0;
1269 imp->imp_connect_tried = 1;
1270 spin_unlock(&imp->imp_lock);
1275 spin_lock(&imp->imp_lock);
1276 imp->imp_connected = 0;
1277 imp->imp_connect_tried = 1;
1278 spin_unlock(&imp->imp_lock);
1281 class_export_put(exp);
1284 IMPORT_SET_STATE(imp, LUSTRE_IMP_DISCON);
1285 if (rc == -EACCES) {
1287 * Give up trying to reconnect
1288 * EACCES means client has no permission for connection
1290 imp->imp_obd->obd_no_recov = 1;
1291 ptlrpc_deactivate_import(imp);
1294 if (rc == -EPROTO) {
1295 struct obd_connect_data *ocd;
1297 /* reply message might not be ready */
1298 if (request->rq_repmsg == NULL)
1301 ocd = req_capsule_server_get(&request->rq_pill,
1304 (ocd->ocd_connect_flags & OBD_CONNECT_VERSION) &&
1305 (ocd->ocd_version != LUSTRE_VERSION_CODE)) {
1306 /* Actually servers are only supposed to refuse
1307 connection from liblustre clients, so we should
1308 never see this from VFS context */
1309 LCONSOLE_ERROR_MSG(0x16a, "Server %s version "
1311 " refused connection from this client "
1312 "with an incompatible version (%s). "
1313 "Client must be recompiled\n",
1314 obd2cli_tgt(imp->imp_obd),
1315 OBD_OCD_VERSION_MAJOR(ocd->ocd_version),
1316 OBD_OCD_VERSION_MINOR(ocd->ocd_version),
1317 OBD_OCD_VERSION_PATCH(ocd->ocd_version),
1318 OBD_OCD_VERSION_FIX(ocd->ocd_version),
1319 LUSTRE_VERSION_STRING);
1320 ptlrpc_deactivate_import(imp);
1321 IMPORT_SET_STATE(imp, LUSTRE_IMP_CLOSED);
1326 ptlrpc_maybe_ping_import_soon(imp);
1328 CDEBUG(D_HA, "recovery of %s on %s failed (%d)\n",
1329 obd2cli_tgt(imp->imp_obd),
1330 (char *)imp->imp_connection->c_remote_uuid.uuid, rc);
1333 wake_up_all(&imp->imp_recovery_waitq);
1338 * interpret callback for "completed replay" RPCs.
1339 * \see signal_completed_replay
1341 static int completed_replay_interpret(const struct lu_env *env,
1342 struct ptlrpc_request *req,
1343 void * data, int rc)
1346 atomic_dec(&req->rq_import->imp_replay_inflight);
1347 if (req->rq_status == 0 &&
1348 !req->rq_import->imp_vbr_failed) {
1349 ptlrpc_import_recovery_state_machine(req->rq_import);
1351 if (req->rq_import->imp_vbr_failed) {
1353 "%s: version recovery fails, reconnecting\n",
1354 req->rq_import->imp_obd->obd_name);
1356 CDEBUG(D_HA, "%s: LAST_REPLAY message error: %d, "
1358 req->rq_import->imp_obd->obd_name,
1361 ptlrpc_connect_import(req->rq_import);
1368 * Let server know that we have no requests to replay anymore.
1369 * Achieved by just sending a PING request
1371 static int signal_completed_replay(struct obd_import *imp)
1373 struct ptlrpc_request *req;
1376 if (unlikely(OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_FINISH_REPLAY)))
1379 LASSERT(atomic_read(&imp->imp_replay_inflight) == 0);
1380 atomic_inc(&imp->imp_replay_inflight);
1382 req = ptlrpc_request_alloc_pack(imp, &RQF_OBD_PING, LUSTRE_OBD_VERSION,
1385 atomic_dec(&imp->imp_replay_inflight);
1389 ptlrpc_request_set_replen(req);
1390 req->rq_send_state = LUSTRE_IMP_REPLAY_WAIT;
1391 lustre_msg_add_flags(req->rq_reqmsg,
1392 MSG_LOCK_REPLAY_DONE | MSG_REQ_REPLAY_DONE);
1394 req->rq_timeout *= 3;
1395 req->rq_interpret_reply = completed_replay_interpret;
1397 ptlrpcd_add_req(req);
1402 * In kernel code all import invalidation happens in its own
1403 * separate thread, so that whatever application happened to encounter
1404 * a problem could still be killed or otherwise continue
1406 static int ptlrpc_invalidate_import_thread(void *data)
1408 struct obd_import *imp = data;
1412 unshare_fs_struct();
1414 CDEBUG(D_HA, "thread invalidate import %s to %s@%s\n",
1415 imp->imp_obd->obd_name, obd2cli_tgt(imp->imp_obd),
1416 imp->imp_connection->c_remote_uuid.uuid);
1418 ptlrpc_invalidate_import(imp);
1420 if (obd_dump_on_eviction) {
1421 CERROR("dump the log upon eviction\n");
1422 libcfs_debug_dumplog();
1425 IMPORT_SET_STATE(imp, LUSTRE_IMP_RECOVER);
1426 ptlrpc_import_recovery_state_machine(imp);
1428 class_import_put(imp);
1433 * This is the state machine for client-side recovery on import.
1435 * Typicaly we have two possibly paths. If we came to server and it is not
1436 * in recovery, we just enter IMP_EVICTED state, invalidate our import
1437 * state and reconnect from scratch.
1438 * If we came to server that is in recovery, we enter IMP_REPLAY import state.
1439 * We go through our list of requests to replay and send them to server one by
1441 * After sending all request from the list we change import state to
1442 * IMP_REPLAY_LOCKS and re-request all the locks we believe we have from server
1443 * and also all the locks we don't yet have and wait for server to grant us.
1444 * After that we send a special "replay completed" request and change import
1445 * state to IMP_REPLAY_WAIT.
1446 * Upon receiving reply to that "replay completed" RPC we enter IMP_RECOVER
1447 * state and resend all requests from sending list.
1448 * After that we promote import to FULL state and send all delayed requests
1449 * and import is fully operational after that.
1452 int ptlrpc_import_recovery_state_machine(struct obd_import *imp)
1460 if (imp->imp_state == LUSTRE_IMP_EVICTED) {
1461 deuuidify(obd2cli_tgt(imp->imp_obd), NULL,
1462 &target_start, &target_len);
1463 /* Don't care about MGC eviction */
1464 if (strcmp(imp->imp_obd->obd_type->typ_name,
1465 LUSTRE_MGC_NAME) != 0) {
1466 LCONSOLE_ERROR_MSG(0x167, "%s: This client was evicted "
1467 "by %.*s; in progress operations "
1468 "using this service will fail.\n",
1469 imp->imp_obd->obd_name, target_len,
1472 CDEBUG(D_HA, "evicted from %s@%s; invalidating\n",
1473 obd2cli_tgt(imp->imp_obd),
1474 imp->imp_connection->c_remote_uuid.uuid);
1475 /* reset vbr_failed flag upon eviction */
1476 spin_lock(&imp->imp_lock);
1477 imp->imp_vbr_failed = 0;
1478 spin_unlock(&imp->imp_lock);
1481 struct task_struct *task;
1482 /* bug 17802: XXX client_disconnect_export vs connect request
1483 * race. if client is evicted at this time then we start
1484 * invalidate thread without reference to import and import can
1485 * be freed at same time. */
1486 class_import_get(imp);
1487 task = kthread_run(ptlrpc_invalidate_import_thread, imp,
1490 class_import_put(imp);
1491 CERROR("error starting invalidate thread: %d\n", rc);
1500 if (imp->imp_state == LUSTRE_IMP_REPLAY) {
1501 CDEBUG(D_HA, "replay requested by %s\n",
1502 obd2cli_tgt(imp->imp_obd));
1503 rc = ptlrpc_replay_next(imp, &inflight);
1504 if (inflight == 0 &&
1505 atomic_read(&imp->imp_replay_inflight) == 0) {
1506 IMPORT_SET_STATE(imp, LUSTRE_IMP_REPLAY_LOCKS);
1507 rc = ldlm_replay_locks(imp);
1514 if (imp->imp_state == LUSTRE_IMP_REPLAY_LOCKS) {
1515 if (atomic_read(&imp->imp_replay_inflight) == 0) {
1516 IMPORT_SET_STATE(imp, LUSTRE_IMP_REPLAY_WAIT);
1517 rc = signal_completed_replay(imp);
1523 if (imp->imp_state == LUSTRE_IMP_REPLAY_WAIT) {
1524 if (atomic_read(&imp->imp_replay_inflight) == 0) {
1525 IMPORT_SET_STATE(imp, LUSTRE_IMP_RECOVER);
1529 if (imp->imp_state == LUSTRE_IMP_RECOVER) {
1530 struct ptlrpc_connection *conn = imp->imp_connection;
1532 rc = ptlrpc_resend(imp);
1535 IMPORT_SET_STATE(imp, LUSTRE_IMP_FULL);
1536 ptlrpc_activate_import(imp);
1538 LCONSOLE_INFO("%s: Connection restored to %s (at %s)\n",
1539 imp->imp_obd->obd_name,
1540 obd_uuid2str(&conn->c_remote_uuid),
1541 libcfs_nid2str(imp->imp_connection->c_peer.nid));
1544 if (imp->imp_state == LUSTRE_IMP_FULL) {
1545 wake_up_all(&imp->imp_recovery_waitq);
1546 ptlrpc_wake_delayed(imp);
1553 int ptlrpc_disconnect_import(struct obd_import *imp, int noclose)
1555 struct ptlrpc_request *req;
1559 if (imp->imp_obd->obd_force)
1560 GOTO(set_state, rc);
1562 switch (imp->imp_connect_op) {
1564 rq_opc = OST_DISCONNECT;
1567 rq_opc = MDS_DISCONNECT;
1570 rq_opc = MGS_DISCONNECT;
1574 CERROR("%s: don't know how to disconnect from %s "
1575 "(connect_op %d): rc = %d\n",
1576 imp->imp_obd->obd_name, obd2cli_tgt(imp->imp_obd),
1577 imp->imp_connect_op, rc);
1581 if (ptlrpc_import_in_recovery(imp)) {
1582 struct l_wait_info lwi;
1583 cfs_duration_t timeout;
1586 if (imp->imp_server_timeout)
1587 timeout = cfs_time_seconds(obd_timeout / 2);
1589 timeout = cfs_time_seconds(obd_timeout);
1591 int idx = import_at_get_index(imp,
1592 imp->imp_client->cli_request_portal);
1593 timeout = cfs_time_seconds(
1594 at_get(&imp->imp_at.iat_service_estimate[idx]));
1597 lwi = LWI_TIMEOUT_INTR(cfs_timeout_cap(timeout),
1598 back_to_sleep, LWI_ON_SIGNAL_NOOP, NULL);
1599 rc = l_wait_event(imp->imp_recovery_waitq,
1600 !ptlrpc_import_in_recovery(imp), &lwi);
1604 spin_lock(&imp->imp_lock);
1605 if (imp->imp_state != LUSTRE_IMP_FULL)
1607 spin_unlock(&imp->imp_lock);
1609 req = ptlrpc_request_alloc_pack(imp, &RQF_MDS_DISCONNECT,
1610 LUSTRE_OBD_VERSION, rq_opc);
1612 /* We are disconnecting, do not retry a failed DISCONNECT rpc if
1613 * it fails. We can get through the above with a down server
1614 * if the client doesn't know the server is gone yet. */
1615 req->rq_no_resend = 1;
1617 /* We want client umounts to happen quickly, no matter the
1619 req->rq_timeout = min_t(int, req->rq_timeout,
1620 INITIAL_CONNECT_TIMEOUT);
1622 IMPORT_SET_STATE(imp, LUSTRE_IMP_CONNECTING);
1623 req->rq_send_state = LUSTRE_IMP_CONNECTING;
1624 ptlrpc_request_set_replen(req);
1625 rc = ptlrpc_queue_wait(req);
1626 ptlrpc_req_finished(req);
1630 spin_lock(&imp->imp_lock);
1633 IMPORT_SET_STATE_NOLOCK(imp, LUSTRE_IMP_DISCON);
1635 IMPORT_SET_STATE_NOLOCK(imp, LUSTRE_IMP_CLOSED);
1636 memset(&imp->imp_remote_handle, 0, sizeof(imp->imp_remote_handle));
1637 spin_unlock(&imp->imp_lock);
1639 if (rc == -ETIMEDOUT || rc == -ENOTCONN || rc == -ESHUTDOWN)
1643 EXPORT_SYMBOL(ptlrpc_disconnect_import);
1645 void ptlrpc_cleanup_imp(struct obd_import *imp)
1649 spin_lock(&imp->imp_lock);
1650 IMPORT_SET_STATE_NOLOCK(imp, LUSTRE_IMP_CLOSED);
1651 imp->imp_generation++;
1652 spin_unlock(&imp->imp_lock);
1653 ptlrpc_abort_inflight(imp);
1658 /* Adaptive Timeout utils */
1659 extern unsigned int at_min, at_max, at_history;
1661 /* Update at_current with the specified value (bounded by at_min and at_max),
1662 * as well as the AT history "bins".
1663 * - Bin into timeslices using AT_BINS bins.
1664 * - This gives us a max of the last at_history seconds without the storage,
1665 * but still smoothing out a return to normalcy from a slow response.
1666 * - (E.g. remember the maximum latency in each minute of the last 4 minutes.)
1668 int at_measured(struct adaptive_timeout *at, unsigned int val)
1670 unsigned int old = at->at_current;
1671 time64_t now = ktime_get_real_seconds();
1672 long binlimit = max_t(long, at_history / AT_BINS, 1);
1675 CDEBUG(D_OTHER, "add %u to %p time=%lu v=%u (%u %u %u %u)\n",
1676 val, at, (long)(now - at->at_binstart), at->at_current,
1677 at->at_hist[0], at->at_hist[1], at->at_hist[2], at->at_hist[3]);
1680 /* 0's don't count, because we never want our timeout to
1681 drop to 0, and because 0 could mean an error */
1684 spin_lock(&at->at_lock);
1686 if (unlikely(at->at_binstart == 0)) {
1687 /* Special case to remove default from history */
1688 at->at_current = val;
1689 at->at_worst_ever = val;
1690 at->at_worst_time = now;
1691 at->at_hist[0] = val;
1692 at->at_binstart = now;
1693 } else if (now - at->at_binstart < binlimit ) {
1695 at->at_hist[0] = max(val, at->at_hist[0]);
1696 at->at_current = max(val, at->at_current);
1699 unsigned int maxv = val;
1701 /* move bins over */
1702 shift = (u32)(now - at->at_binstart) / binlimit;
1704 for(i = AT_BINS - 1; i >= 0; i--) {
1706 at->at_hist[i] = at->at_hist[i - shift];
1707 maxv = max(maxv, at->at_hist[i]);
1712 at->at_hist[0] = val;
1713 at->at_current = maxv;
1714 at->at_binstart += shift * binlimit;
1717 if (at->at_current > at->at_worst_ever) {
1718 at->at_worst_ever = at->at_current;
1719 at->at_worst_time = now;
1722 if (at->at_flags & AT_FLG_NOHIST)
1723 /* Only keep last reported val; keeping the rest of the history
1725 at->at_current = val;
1728 at->at_current = min(at->at_current, at_max);
1729 at->at_current = max(at->at_current, at_min);
1731 if (at->at_current != old)
1732 CDEBUG(D_OTHER, "AT %p change: old=%u new=%u delta=%d "
1733 "(val=%u) hist %u %u %u %u\n", at,
1734 old, at->at_current, at->at_current - old, val,
1735 at->at_hist[0], at->at_hist[1], at->at_hist[2],
1738 /* if we changed, report the old value */
1739 old = (at->at_current != old) ? old : 0;
1741 spin_unlock(&at->at_lock);
1745 /* Find the imp_at index for a given portal; assign if space available */
1746 int import_at_get_index(struct obd_import *imp, int portal)
1748 struct imp_at *at = &imp->imp_at;
1751 for (i = 0; i < IMP_AT_MAX_PORTALS; i++) {
1752 if (at->iat_portal[i] == portal)
1754 if (at->iat_portal[i] == 0)
1759 /* Not found in list, add it under a lock */
1760 spin_lock(&imp->imp_lock);
1762 /* Check unused under lock */
1763 for (; i < IMP_AT_MAX_PORTALS; i++) {
1764 if (at->iat_portal[i] == portal)
1766 if (at->iat_portal[i] == 0)
1771 /* Not enough portals? */
1772 LASSERT(i < IMP_AT_MAX_PORTALS);
1774 at->iat_portal[i] = portal;
1776 spin_unlock(&imp->imp_lock);