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, 2017, 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
61 static void import_set_state_nolock(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;
78 /* A CLOSED import should remain so. */
79 if (state == LUSTRE_IMP_CLOSED)
82 if (imp->imp_state != LUSTRE_IMP_NEW) {
83 CDEBUG(D_HA, "%p %s: changing import state from %s to %s\n",
84 imp, obd2cli_tgt(imp->imp_obd),
85 ptlrpc_import_state_name(imp->imp_state),
86 ptlrpc_import_state_name(state));
89 imp->imp_state = state;
90 imp->imp_state_hist[imp->imp_state_hist_idx].ish_state = state;
91 imp->imp_state_hist[imp->imp_state_hist_idx].ish_time =
92 ktime_get_real_seconds();
93 imp->imp_state_hist_idx = (imp->imp_state_hist_idx + 1) %
97 static void import_set_state(struct obd_import *imp,
98 enum lustre_imp_state new_state)
100 spin_lock(&imp->imp_lock);
101 import_set_state_nolock(imp, new_state);
102 spin_unlock(&imp->imp_lock);
105 void ptlrpc_import_enter_resend(struct obd_import *imp)
107 import_set_state(imp, LUSTRE_IMP_RECOVER);
109 EXPORT_SYMBOL(ptlrpc_import_enter_resend);
112 static int ptlrpc_connect_interpret(const struct lu_env *env,
113 struct ptlrpc_request *request,
115 int ptlrpc_import_recovery_state_machine(struct obd_import *imp);
117 /* Only this function is allowed to change the import state when it is
118 * CLOSED. I would rather refcount the import and free it after
119 * disconnection like we do with exports. To do that, the client_obd
120 * will need to save the peer info somewhere other than in the import,
122 int ptlrpc_init_import(struct obd_import *imp)
124 spin_lock(&imp->imp_lock);
126 imp->imp_generation++;
127 imp->imp_state = LUSTRE_IMP_NEW;
129 spin_unlock(&imp->imp_lock);
133 EXPORT_SYMBOL(ptlrpc_init_import);
135 #define UUID_STR "_UUID"
136 void deuuidify(char *uuid, const char *prefix, char **uuid_start, int *uuid_len)
138 *uuid_start = !prefix || strncmp(uuid, prefix, strlen(prefix))
139 ? uuid : uuid + strlen(prefix);
141 *uuid_len = strlen(*uuid_start);
143 if (*uuid_len < strlen(UUID_STR))
146 if (!strncmp(*uuid_start + *uuid_len - strlen(UUID_STR),
147 UUID_STR, strlen(UUID_STR)))
148 *uuid_len -= strlen(UUID_STR);
151 /* Must be called with imp_lock held! */
152 static void ptlrpc_deactivate_import_nolock(struct obd_import *imp)
156 assert_spin_locked(&imp->imp_lock);
157 CDEBUG(D_HA, "setting import %s INVALID\n", obd2cli_tgt(imp->imp_obd));
158 imp->imp_invalid = 1;
159 imp->imp_generation++;
161 ptlrpc_abort_inflight(imp);
167 * Returns true if import was FULL, false if import was already not
169 * @imp - import to be disconnected
170 * @conn_cnt - connection count (epoch) of the request that timed out
171 * and caused the disconnection. In some cases, multiple
172 * inflight requests can fail to a single target (e.g. OST
173 * bulk requests) and if one has already caused a reconnection
174 * (increasing the import->conn_cnt) the older failure should
175 * not also cause a reconnection. If zero it forces a reconnect.
176 * @invalid - set import invalid flag
178 int ptlrpc_set_import_discon(struct obd_import *imp,
179 __u32 conn_cnt, bool invalid)
183 spin_lock(&imp->imp_lock);
185 if (imp->imp_state == LUSTRE_IMP_FULL &&
186 (conn_cnt == 0 || conn_cnt == imp->imp_conn_cnt)) {
191 deuuidify(obd2cli_tgt(imp->imp_obd), NULL,
192 &target_start, &target_len);
194 import_set_state_nolock(imp, LUSTRE_IMP_DISCON);
195 if (imp->imp_replayable) {
196 LCONSOLE_WARN("%s: Connection to %.*s (at %s) was "
197 "lost; in progress operations using this "
198 "service will wait for recovery to complete\n",
199 imp->imp_obd->obd_name, target_len, target_start,
200 obd_import_nid2str(imp));
202 LCONSOLE_ERROR_MSG(0x166, "%s: Connection to "
203 "%.*s (at %s) was lost; in progress "
204 "operations using this service will fail\n",
205 imp->imp_obd->obd_name, target_len, target_start,
206 obd_import_nid2str(imp));
208 CDEBUG(D_HA, "import %s@%s for %s not "
209 "replayable, auto-deactivating\n",
210 obd2cli_tgt(imp->imp_obd),
211 imp->imp_connection->c_remote_uuid.uuid,
212 imp->imp_obd->obd_name);
213 ptlrpc_deactivate_import_nolock(imp);
217 spin_unlock(&imp->imp_lock);
219 if (obd_dump_on_timeout)
220 libcfs_debug_dumplog();
222 obd_import_event(imp->imp_obd, imp, IMP_EVENT_DISCON);
225 obd_import_event(imp->imp_obd, imp, IMP_EVENT_INACTIVE);
228 spin_unlock(&imp->imp_lock);
229 CDEBUG(D_HA, "%s: import %p already %s (conn %u, was %u): %s\n",
230 imp->imp_client->cli_name, imp,
231 (imp->imp_state == LUSTRE_IMP_FULL &&
232 imp->imp_conn_cnt > conn_cnt) ?
233 "reconnected" : "not connected", imp->imp_conn_cnt,
234 conn_cnt, ptlrpc_import_state_name(imp->imp_state));
241 * This acts as a barrier; all existing requests are rejected, and
242 * no new requests will be accepted until the import is valid again.
244 void ptlrpc_deactivate_import(struct obd_import *imp)
246 spin_lock(&imp->imp_lock);
247 ptlrpc_deactivate_import_nolock(imp);
248 spin_unlock(&imp->imp_lock);
250 obd_import_event(imp->imp_obd, imp, IMP_EVENT_INACTIVE);
252 EXPORT_SYMBOL(ptlrpc_deactivate_import);
254 static time64_t ptlrpc_inflight_deadline(struct ptlrpc_request *req,
259 if (!(((req->rq_phase == RQ_PHASE_RPC) && !req->rq_waiting) ||
260 (req->rq_phase == RQ_PHASE_BULK) ||
261 (req->rq_phase == RQ_PHASE_NEW)))
264 if (req->rq_timedout)
267 if (req->rq_phase == RQ_PHASE_NEW)
270 dl = req->rq_deadline;
278 static time64_t ptlrpc_inflight_timeout(struct obd_import *imp)
280 time64_t now = ktime_get_real_seconds();
281 struct list_head *tmp, *n;
282 struct ptlrpc_request *req;
283 time64_t timeout = 0;
285 spin_lock(&imp->imp_lock);
286 list_for_each_safe(tmp, n, &imp->imp_sending_list) {
287 req = list_entry(tmp, struct ptlrpc_request, rq_list);
288 timeout = max(ptlrpc_inflight_deadline(req, now), timeout);
290 spin_unlock(&imp->imp_lock);
295 * This function will invalidate the import, if necessary, then block
296 * for all the RPC completions, and finally notify the obd to
297 * invalidate its state (ie cancel locks, clear pending requests,
300 void ptlrpc_invalidate_import(struct obd_import *imp)
302 struct list_head *tmp, *n;
303 struct ptlrpc_request *req;
304 struct l_wait_info lwi;
308 atomic_inc(&imp->imp_inval_count);
310 if (!imp->imp_invalid || imp->imp_obd->obd_no_recov)
311 ptlrpc_deactivate_import(imp);
313 if (OBD_FAIL_PRECHECK(OBD_FAIL_PTLRPC_CONNECT_RACE)) {
314 OBD_RACE(OBD_FAIL_PTLRPC_CONNECT_RACE);
315 msleep(10 * MSEC_PER_SEC);
317 CFS_FAIL_TIMEOUT(OBD_FAIL_MGS_CONNECT_NET, 3 * cfs_fail_val / 2);
318 LASSERT(imp->imp_invalid);
320 /* Wait forever until inflight == 0. We really can't do it another
321 * way because in some cases we need to wait for very long reply
322 * unlink. We can't do anything before that because there is really
323 * no guarantee that some rdma transfer is not in progress right now.
326 long timeout_jiffies;
328 /* Calculate max timeout for waiting on rpcs to error
329 * out. Use obd_timeout if calculated value is smaller
332 if (!OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_REPL_UNLINK)) {
333 timeout = ptlrpc_inflight_timeout(imp);
334 timeout += div_u64(timeout, 3);
337 timeout = obd_timeout;
339 /* decrease the interval to increase race condition */
343 CDEBUG(D_RPCTRACE, "Sleeping %llds for inflight to error out\n",
346 /* Wait for all requests to error out and call completion
347 * callbacks. Cap it at obd_timeout -- these should all
348 * have been locally cancelled by ptlrpc_abort_inflight.
350 timeout_jiffies = max_t(long, cfs_time_seconds(timeout), 1);
351 lwi = LWI_TIMEOUT_INTERVAL(timeout_jiffies,
352 (timeout > 1) ? cfs_time_seconds(1) :
353 cfs_time_seconds(1) / 2,
355 rc = l_wait_event(imp->imp_recovery_waitq,
356 (atomic_read(&imp->imp_inflight) == 0),
359 const char *cli_tgt = obd2cli_tgt(imp->imp_obd);
361 CERROR("%s: rc = %d waiting for callback (%d != 0)\n",
362 cli_tgt, rc, atomic_read(&imp->imp_inflight));
364 spin_lock(&imp->imp_lock);
365 if (atomic_read(&imp->imp_inflight) == 0) {
366 int count = atomic_read(&imp->imp_unregistering);
368 /* We know that "unregistering" rpcs only can
369 * survive in sending or delaying lists (they
370 * maybe waiting for long reply unlink in
371 * sluggish nets). Let's check this. If there
372 * is no inflight and unregistering != 0, this
374 LASSERTF(count == 0, "Some RPCs are still "
375 "unregistering: %d\n", count);
377 /* Let's save one loop as soon as inflight have
378 * dropped to zero. No new inflights possible at
382 list_for_each_safe(tmp, n,
383 &imp->imp_sending_list) {
384 req = list_entry(tmp,
385 struct ptlrpc_request,
387 DEBUG_REQ(D_ERROR, req,
388 "still on sending list");
390 list_for_each_safe(tmp, n,
391 &imp->imp_delayed_list) {
392 req = list_entry(tmp,
393 struct ptlrpc_request,
395 DEBUG_REQ(D_ERROR, req,
396 "still on delayed list");
399 CERROR("%s: Unregistering RPCs found (%d). "
400 "Network is sluggish? Waiting for them "
401 "to error out.\n", cli_tgt,
402 atomic_read(&imp->imp_unregistering));
404 spin_unlock(&imp->imp_lock);
409 * Let's additionally check that no new rpcs added to import in
410 * "invalidate" state.
412 LASSERT(atomic_read(&imp->imp_inflight) == 0);
413 obd_import_event(imp->imp_obd, imp, IMP_EVENT_INVALIDATE);
414 sptlrpc_import_flush_all_ctx(imp);
416 atomic_dec(&imp->imp_inval_count);
417 wake_up_all(&imp->imp_recovery_waitq);
419 EXPORT_SYMBOL(ptlrpc_invalidate_import);
421 /* unset imp_invalid */
422 void ptlrpc_activate_import(struct obd_import *imp, bool set_state_full)
424 struct obd_device *obd = imp->imp_obd;
426 spin_lock(&imp->imp_lock);
427 if (imp->imp_deactive != 0) {
428 LASSERT(imp->imp_state != LUSTRE_IMP_FULL);
429 if (imp->imp_state != LUSTRE_IMP_DISCON)
430 import_set_state_nolock(imp, LUSTRE_IMP_DISCON);
431 spin_unlock(&imp->imp_lock);
435 import_set_state_nolock(imp, LUSTRE_IMP_FULL);
437 imp->imp_invalid = 0;
439 spin_unlock(&imp->imp_lock);
440 obd_import_event(obd, imp, IMP_EVENT_ACTIVE);
442 EXPORT_SYMBOL(ptlrpc_activate_import);
444 void ptlrpc_pinger_force(struct obd_import *imp)
446 CDEBUG(D_HA, "%s: waking up pinger s:%s\n", obd2cli_tgt(imp->imp_obd),
447 ptlrpc_import_state_name(imp->imp_state));
449 spin_lock(&imp->imp_lock);
450 imp->imp_force_verify = 1;
451 spin_unlock(&imp->imp_lock);
453 if (imp->imp_state != LUSTRE_IMP_CONNECTING)
454 ptlrpc_pinger_wake_up();
456 EXPORT_SYMBOL(ptlrpc_pinger_force);
458 void ptlrpc_fail_import(struct obd_import *imp, __u32 conn_cnt)
462 LASSERT(!imp->imp_dlm_fake);
464 if (ptlrpc_set_import_discon(imp, conn_cnt, true))
465 ptlrpc_pinger_force(imp);
470 int ptlrpc_reconnect_import(struct obd_import *imp)
473 long timeout_jiffies = cfs_time_seconds(obd_timeout);
474 struct l_wait_info lwi;
477 ptlrpc_pinger_force(imp);
479 CDEBUG(D_HA, "%s: recovery started, waiting %u seconds\n",
480 obd2cli_tgt(imp->imp_obd), obd_timeout);
482 lwi = LWI_TIMEOUT(timeout_jiffies, NULL, NULL);
483 rc = l_wait_event(imp->imp_recovery_waitq,
484 !ptlrpc_import_in_recovery(imp), &lwi);
485 CDEBUG(D_HA, "%s: recovery finished s:%s\n", obd2cli_tgt(imp->imp_obd),
486 ptlrpc_import_state_name(imp->imp_state));
489 ptlrpc_set_import_discon(imp, 0, false);
490 /* Force a new connect attempt */
491 ptlrpc_invalidate_import(imp);
492 /* Do a fresh connect next time by zeroing the handle */
493 ptlrpc_disconnect_import(imp, 1);
494 /* Wait for all invalidate calls to finish */
495 if (atomic_read(&imp->imp_inval_count) > 0) {
496 struct l_wait_info lwi = LWI_INTR(LWI_ON_SIGNAL_NOOP, NULL);
499 rc = l_wait_event(imp->imp_recovery_waitq,
500 (atomic_read(&imp->imp_inval_count) == 0),
503 CERROR("Interrupted, inval=%d\n",
504 atomic_read(&imp->imp_inval_count));
507 /* Allow reconnect attempts */
508 imp->imp_obd->obd_no_recov = 0;
509 /* Remove 'invalid' flag */
510 ptlrpc_activate_import(imp, false);
511 /* Attempt a new connect */
512 ptlrpc_recover_import(imp, NULL, 0);
516 EXPORT_SYMBOL(ptlrpc_reconnect_import);
519 * Connection on import \a imp is changed to another one (if more than one is
520 * present). We typically chose connection that we have not tried to connect to
523 static int import_select_connection(struct obd_import *imp)
525 struct obd_import_conn *imp_conn = NULL, *conn;
526 struct obd_export *dlmexp;
528 int target_len, tried_all = 1;
531 spin_lock(&imp->imp_lock);
533 if (list_empty(&imp->imp_conn_list)) {
534 CERROR("%s: no connections available\n",
535 imp->imp_obd->obd_name);
536 spin_unlock(&imp->imp_lock);
540 list_for_each_entry(conn, &imp->imp_conn_list, oic_item) {
541 CDEBUG(D_HA, "%s: connect to NID %s last attempt %lld\n",
542 imp->imp_obd->obd_name,
543 libcfs_nid2str(conn->oic_conn->c_peer.nid),
544 conn->oic_last_attempt);
546 /* If we have not tried this connection since
547 the last successful attempt, go with this one */
548 if ((conn->oic_last_attempt == 0) ||
549 conn->oic_last_attempt <= imp->imp_last_success_conn) {
555 /* If all of the connections have already been tried
556 since the last successful connection; just choose the
557 least recently used */
560 else if (imp_conn->oic_last_attempt > conn->oic_last_attempt)
564 /* if not found, simply choose the current one */
565 if (!imp_conn || imp->imp_force_reconnect) {
566 LASSERT(imp->imp_conn_current);
567 imp_conn = imp->imp_conn_current;
570 LASSERT(imp_conn->oic_conn);
572 /* If we've tried everything, and we're back to the beginning of the
573 list, increase our timeout and try again. It will be reset when
574 we do finally connect. (FIXME: really we should wait for all network
575 state associated with the last connection attempt to drain before
576 trying to reconnect on it.) */
577 if (tried_all && (imp->imp_conn_list.next == &imp_conn->oic_item)) {
578 struct adaptive_timeout *at = &imp->imp_at.iat_net_latency;
579 if (at_get(at) < CONNECTION_SWITCH_MAX) {
580 at_measured(at, at_get(at) + CONNECTION_SWITCH_INC);
581 if (at_get(at) > CONNECTION_SWITCH_MAX)
582 at_reset(at, CONNECTION_SWITCH_MAX);
584 LASSERT(imp_conn->oic_last_attempt);
585 CDEBUG(D_HA, "%s: tried all connections, increasing latency "
586 "to %ds\n", imp->imp_obd->obd_name, at_get(at));
589 imp_conn->oic_last_attempt = ktime_get_seconds();
591 /* switch connection, don't mind if it's same as the current one */
592 if (imp->imp_connection)
593 ptlrpc_connection_put(imp->imp_connection);
594 imp->imp_connection = ptlrpc_connection_addref(imp_conn->oic_conn);
596 dlmexp = class_conn2export(&imp->imp_dlm_handle);
597 LASSERT(dlmexp != NULL);
598 if (dlmexp->exp_connection)
599 ptlrpc_connection_put(dlmexp->exp_connection);
600 dlmexp->exp_connection = ptlrpc_connection_addref(imp_conn->oic_conn);
601 class_export_put(dlmexp);
603 if (imp->imp_conn_current != imp_conn) {
604 if (imp->imp_conn_current) {
605 deuuidify(obd2cli_tgt(imp->imp_obd), NULL,
606 &target_start, &target_len);
608 CDEBUG(D_HA, "%s: Connection changing to"
610 imp->imp_obd->obd_name,
611 target_len, target_start,
612 libcfs_nid2str(imp_conn->oic_conn->c_peer.nid));
615 imp->imp_conn_current = imp_conn;
618 /* The below message is checked in conf-sanity.sh test_35[ab] */
619 CDEBUG(D_HA, "%s: import %p using connection %s/%s\n",
620 imp->imp_obd->obd_name, imp, imp_conn->oic_uuid.uuid,
621 libcfs_nid2str(imp_conn->oic_conn->c_peer.nid));
623 spin_unlock(&imp->imp_lock);
629 * must be called under imp_lock
631 static int ptlrpc_first_transno(struct obd_import *imp, __u64 *transno)
633 struct ptlrpc_request *req;
634 struct list_head *tmp;
636 /* The requests in committed_list always have smaller transnos than
637 * the requests in replay_list */
638 if (!list_empty(&imp->imp_committed_list)) {
639 tmp = imp->imp_committed_list.next;
640 req = list_entry(tmp, struct ptlrpc_request, rq_replay_list);
641 *transno = req->rq_transno;
642 if (req->rq_transno == 0) {
643 DEBUG_REQ(D_ERROR, req,
644 "zero transno in committed_list");
649 if (!list_empty(&imp->imp_replay_list)) {
650 tmp = imp->imp_replay_list.next;
651 req = list_entry(tmp, struct ptlrpc_request, rq_replay_list);
652 *transno = req->rq_transno;
653 if (req->rq_transno == 0) {
654 DEBUG_REQ(D_ERROR, req, "zero transno in replay_list");
662 int ptlrpc_connect_import(struct obd_import *imp)
664 spin_lock(&imp->imp_lock);
665 return ptlrpc_connect_import_locked(imp);
669 * Attempt to (re)connect import \a imp. This includes all preparations,
670 * initializing CONNECT RPC request and passing it to ptlrpcd for
673 * Assumes imp->imp_lock is held, and releases it.
675 * Returns 0 on success or error code.
677 int ptlrpc_connect_import_locked(struct obd_import *imp)
679 struct obd_device *obd = imp->imp_obd;
680 int initial_connect = 0;
682 __u64 committed_before_reconnect = 0;
683 struct ptlrpc_request *request;
684 struct obd_connect_data ocd;
685 char *bufs[] = { NULL,
686 obd2cli_tgt(imp->imp_obd),
688 (char *)&imp->imp_dlm_handle,
691 struct ptlrpc_connect_async_args *aa;
695 assert_spin_locked(&imp->imp_lock);
697 if (imp->imp_state == LUSTRE_IMP_CLOSED) {
698 spin_unlock(&imp->imp_lock);
699 CERROR("can't connect to a closed import\n");
701 } else if (imp->imp_state == LUSTRE_IMP_FULL) {
702 spin_unlock(&imp->imp_lock);
703 CERROR("already connected\n");
705 } else if (imp->imp_state == LUSTRE_IMP_CONNECTING ||
706 imp->imp_state == LUSTRE_IMP_EVICTED ||
707 imp->imp_connected) {
708 spin_unlock(&imp->imp_lock);
709 CERROR("already connecting\n");
713 import_set_state_nolock(imp, LUSTRE_IMP_CONNECTING);
716 imp->imp_resend_replay = 0;
718 if (!lustre_handle_is_used(&imp->imp_remote_handle))
721 committed_before_reconnect = imp->imp_peer_committed_transno;
723 set_transno = ptlrpc_first_transno(imp,
724 &imp->imp_connect_data.ocd_transno);
725 spin_unlock(&imp->imp_lock);
727 rc = import_select_connection(imp);
731 rc = sptlrpc_import_sec_adapt(imp, NULL, NULL);
735 /* Reset connect flags to the originally requested flags, in case
736 * the server is updated on-the-fly we will get the new features. */
737 ocd = imp->imp_connect_data;
738 ocd.ocd_connect_flags = imp->imp_connect_flags_orig;
739 ocd.ocd_connect_flags2 = imp->imp_connect_flags2_orig;
740 /* Reset ocd_version each time so the server knows the exact versions */
741 ocd.ocd_version = LUSTRE_VERSION_CODE;
742 imp->imp_msghdr_flags &= ~MSGHDR_AT_SUPPORT;
743 imp->imp_msghdr_flags &= ~MSGHDR_CKSUM_INCOMPAT18;
745 rc = obd_reconnect(NULL, imp->imp_obd->obd_self_export, obd,
746 &obd->obd_uuid, &ocd, NULL);
750 request = ptlrpc_request_alloc(imp, &RQF_MDS_CONNECT);
752 GOTO(out, rc = -ENOMEM);
754 /* get SELinux policy info if any */
755 rc = sptlrpc_get_sepol(request);
757 ptlrpc_request_free(request);
761 bufs[5] = request->rq_sepol;
763 req_capsule_set_size(&request->rq_pill, &RMF_SELINUX_POL, RCL_CLIENT,
764 strlen(request->rq_sepol) ?
765 strlen(request->rq_sepol) + 1 : 0);
767 rc = ptlrpc_request_bufs_pack(request, LUSTRE_OBD_VERSION,
768 imp->imp_connect_op, bufs, NULL);
770 ptlrpc_request_free(request);
774 /* Report the rpc service time to the server so that it knows how long
775 * to wait for clients to join recovery */
776 lustre_msg_set_service_time(request->rq_reqmsg,
777 at_timeout2est(request->rq_timeout));
779 /* The amount of time we give the server to process the connect req.
780 * import_select_connection will increase the net latency on
781 * repeated reconnect attempts to cover slow networks.
782 * We override/ignore the server rpc completion estimate here,
783 * which may be large if this is a reconnect attempt */
784 request->rq_timeout = INITIAL_CONNECT_TIMEOUT;
785 lustre_msg_set_timeout(request->rq_reqmsg, request->rq_timeout);
787 request->rq_no_resend = request->rq_no_delay = 1;
788 request->rq_send_state = LUSTRE_IMP_CONNECTING;
789 /* Allow a slightly larger reply for future growth compatibility */
790 req_capsule_set_size(&request->rq_pill, &RMF_CONNECT_DATA, RCL_SERVER,
791 sizeof(struct obd_connect_data)+16*sizeof(__u64));
792 ptlrpc_request_set_replen(request);
793 request->rq_interpret_reply = ptlrpc_connect_interpret;
795 aa = ptlrpc_req_async_args(aa, request);
796 memset(aa, 0, sizeof *aa);
798 aa->pcaa_peer_committed = committed_before_reconnect;
799 aa->pcaa_initial_connect = initial_connect;
801 if (aa->pcaa_initial_connect) {
802 spin_lock(&imp->imp_lock);
803 imp->imp_replayable = 1;
804 spin_unlock(&imp->imp_lock);
805 lustre_msg_add_op_flags(request->rq_reqmsg,
806 MSG_CONNECT_INITIAL);
810 lustre_msg_add_op_flags(request->rq_reqmsg,
811 MSG_CONNECT_TRANSNO);
813 DEBUG_REQ(D_RPCTRACE, request, "(re)connect request (timeout %ld)",
814 request->rq_timeout);
815 ptlrpcd_add_req(request);
819 import_set_state(imp, LUSTRE_IMP_DISCON);
823 EXPORT_SYMBOL(ptlrpc_connect_import);
825 static void ptlrpc_maybe_ping_import_soon(struct obd_import *imp)
829 spin_lock(&imp->imp_lock);
830 force_verify = imp->imp_force_verify != 0;
831 spin_unlock(&imp->imp_lock);
834 ptlrpc_pinger_wake_up();
837 static int ptlrpc_busy_reconnect(int rc)
839 return (rc == -EBUSY) || (rc == -EAGAIN);
842 static int ptlrpc_connect_set_flags(struct obd_import *imp,
843 struct obd_connect_data *ocd,
844 __u64 old_connect_flags,
845 struct obd_export *exp, int init_connect)
848 struct client_obd *cli = &imp->imp_obd->u.cli;
850 spin_lock(&imp->imp_lock);
851 list_move(&imp->imp_conn_current->oic_item,
852 &imp->imp_conn_list);
853 imp->imp_last_success_conn =
854 imp->imp_conn_current->oic_last_attempt;
856 spin_unlock(&imp->imp_lock);
858 if (!warned && (ocd->ocd_connect_flags & OBD_CONNECT_VERSION) &&
859 (ocd->ocd_version > LUSTRE_VERSION_CODE +
860 LUSTRE_VERSION_OFFSET_WARN ||
861 ocd->ocd_version < LUSTRE_VERSION_CODE -
862 LUSTRE_VERSION_OFFSET_WARN)) {
863 /* Sigh, some compilers do not like #ifdef in the middle
864 of macro arguments */
865 const char *older = "older than client. "
866 "Consider upgrading server";
867 const char *newer = "newer than client. "
868 "Consider upgrading client";
870 LCONSOLE_WARN("Server %s version (%d.%d.%d.%d) "
872 obd2cli_tgt(imp->imp_obd),
873 OBD_OCD_VERSION_MAJOR(ocd->ocd_version),
874 OBD_OCD_VERSION_MINOR(ocd->ocd_version),
875 OBD_OCD_VERSION_PATCH(ocd->ocd_version),
876 OBD_OCD_VERSION_FIX(ocd->ocd_version),
877 ocd->ocd_version > LUSTRE_VERSION_CODE ?
878 newer : older, LUSTRE_VERSION_STRING);
882 if (ocd->ocd_connect_flags & OBD_CONNECT_CKSUM) {
883 /* We sent to the server ocd_cksum_types with bits set
884 * for algorithms we understand. The server masked off
885 * the checksum types it doesn't support */
886 if ((ocd->ocd_cksum_types &
887 obd_cksum_types_supported_client()) == 0) {
888 LCONSOLE_ERROR("The negotiation of the checksum "
889 "alogrithm to use with server %s "
891 obd2cli_tgt(imp->imp_obd),
892 ocd->ocd_cksum_types,
893 obd_cksum_types_supported_client());
896 cli->cl_supp_cksum_types = ocd->ocd_cksum_types;
899 /* The server does not support OBD_CONNECT_CKSUM.
900 * Enforce ADLER for backward compatibility*/
901 cli->cl_supp_cksum_types = OBD_CKSUM_ADLER;
903 cli->cl_cksum_type = obd_cksum_type_select(imp->imp_obd->obd_name,
904 cli->cl_supp_cksum_types,
905 cli->cl_preferred_cksum_type);
907 if (ocd->ocd_connect_flags & OBD_CONNECT_BRW_SIZE)
908 cli->cl_max_pages_per_rpc =
909 min(ocd->ocd_brw_size >> PAGE_SHIFT,
910 cli->cl_max_pages_per_rpc);
911 else if (imp->imp_connect_op == MDS_CONNECT ||
912 imp->imp_connect_op == MGS_CONNECT)
913 cli->cl_max_pages_per_rpc = 1;
915 LASSERT((cli->cl_max_pages_per_rpc <= PTLRPC_MAX_BRW_PAGES) &&
916 (cli->cl_max_pages_per_rpc > 0));
918 client_adjust_max_dirty(cli);
920 /* Update client max modify RPCs in flight with value returned
922 if (ocd->ocd_connect_flags & OBD_CONNECT_MULTIMODRPCS)
923 cli->cl_max_mod_rpcs_in_flight = min(
924 cli->cl_max_mod_rpcs_in_flight,
925 ocd->ocd_maxmodrpcs);
927 cli->cl_max_mod_rpcs_in_flight = 1;
929 /* Reset ns_connect_flags only for initial connect. It might be
930 * changed in while using FS and if we reset it in reconnect
931 * this leads to losing user settings done before such as
932 * disable lru_resize, etc. */
933 if (old_connect_flags != exp_connect_flags(exp) || init_connect) {
934 struct ldlm_namespace *ns = imp->imp_obd->obd_namespace;
938 ns->ns_connect_flags ^ ns->ns_orig_connect_flags;
939 CDEBUG(D_HA, "%s: Resetting ns_connect_flags to server "
940 "flags: %#llx\n", imp->imp_obd->obd_name,
941 ocd->ocd_connect_flags);
942 ns->ns_connect_flags = (ns->ns_connect_flags & changed_flags) |
943 (ocd->ocd_connect_flags & ~changed_flags);
944 ns->ns_orig_connect_flags = ocd->ocd_connect_flags;
947 if (ocd->ocd_connect_flags & OBD_CONNECT_AT)
948 /* We need a per-message support flag, because
949 * a. we don't know if the incoming connect reply
950 * supports AT or not (in reply_in_callback)
951 * until we unpack it.
952 * b. failovered server means export and flags are gone
953 * (in ptlrpc_send_reply).
954 * Can only be set when we know AT is supported at
956 imp->imp_msghdr_flags |= MSGHDR_AT_SUPPORT;
958 imp->imp_msghdr_flags &= ~MSGHDR_AT_SUPPORT;
960 imp->imp_msghdr_flags |= MSGHDR_CKSUM_INCOMPAT18;
966 * Add all replay requests back to unreplied list before start replay,
967 * so that we can make sure the known replied XID is always increased
968 * only even if when replaying requests.
970 static void ptlrpc_prepare_replay(struct obd_import *imp)
972 struct ptlrpc_request *req;
974 if (imp->imp_state != LUSTRE_IMP_REPLAY ||
975 imp->imp_resend_replay)
978 /* If the server was restart during repaly, the requests may
979 * have been added to the unreplied list in former replay. */
980 spin_lock(&imp->imp_lock);
982 list_for_each_entry(req, &imp->imp_committed_list, rq_replay_list) {
983 if (list_empty(&req->rq_unreplied_list))
984 ptlrpc_add_unreplied(req);
987 list_for_each_entry(req, &imp->imp_replay_list, rq_replay_list) {
988 if (list_empty(&req->rq_unreplied_list))
989 ptlrpc_add_unreplied(req);
992 imp->imp_known_replied_xid = ptlrpc_known_replied_xid(imp);
993 spin_unlock(&imp->imp_lock);
997 * interpret_reply callback for connect RPCs.
998 * Looks into returned status of connect operation and decides
999 * what to do with the import - i.e enter recovery, promote it to
1000 * full state for normal operations of disconnect it due to an error.
1002 static int ptlrpc_connect_interpret(const struct lu_env *env,
1003 struct ptlrpc_request *request,
1006 struct ptlrpc_connect_async_args *aa = data;
1007 struct obd_import *imp = request->rq_import;
1008 struct lustre_handle old_hdl;
1009 __u64 old_connect_flags;
1011 struct obd_connect_data *ocd;
1012 struct obd_export *exp = NULL;
1016 spin_lock(&imp->imp_lock);
1017 if (imp->imp_state == LUSTRE_IMP_CLOSED) {
1018 imp->imp_connect_tried = 1;
1019 spin_unlock(&imp->imp_lock);
1023 imp->imp_connect_error = rc;
1025 struct ptlrpc_request *free_req;
1026 struct ptlrpc_request *tmp;
1028 /* abort all delayed requests initiated connection */
1029 list_for_each_entry_safe(free_req, tmp, &imp->imp_delayed_list,
1031 spin_lock(&free_req->rq_lock);
1032 if (free_req->rq_no_resend) {
1033 free_req->rq_err = 1;
1034 free_req->rq_status = -EIO;
1035 ptlrpc_client_wake_req(free_req);
1037 spin_unlock(&free_req->rq_lock);
1040 /* if this reconnect to busy export - not need select new target
1042 imp->imp_force_reconnect = ptlrpc_busy_reconnect(rc);
1043 spin_unlock(&imp->imp_lock);
1044 ptlrpc_maybe_ping_import_soon(imp);
1048 /* LU-7558: indicate that we are interpretting connect reply,
1049 * pltrpc_connect_import() will not try to reconnect until
1050 * interpret will finish. */
1051 imp->imp_connected = 1;
1052 spin_unlock(&imp->imp_lock);
1054 LASSERT(imp->imp_conn_current);
1056 msg_flags = lustre_msg_get_op_flags(request->rq_repmsg);
1058 ret = req_capsule_get_size(&request->rq_pill, &RMF_CONNECT_DATA,
1060 /* server replied obd_connect_data is always bigger */
1061 ocd = req_capsule_server_sized_get(&request->rq_pill,
1062 &RMF_CONNECT_DATA, ret);
1065 CERROR("%s: no connect data from server\n",
1066 imp->imp_obd->obd_name);
1071 spin_lock(&imp->imp_lock);
1073 /* All imports are pingable */
1074 imp->imp_pingable = 1;
1075 imp->imp_force_reconnect = 0;
1076 imp->imp_force_verify = 0;
1078 imp->imp_connect_data = *ocd;
1080 CDEBUG(D_HA, "%s: connect to target with instance %u\n",
1081 imp->imp_obd->obd_name, ocd->ocd_instance);
1082 exp = class_conn2export(&imp->imp_dlm_handle);
1084 spin_unlock(&imp->imp_lock);
1087 /* This could happen if export is cleaned during the
1089 CERROR("%s: missing export after connect\n",
1090 imp->imp_obd->obd_name);
1091 GOTO(out, rc = -ENODEV);
1094 /* check that server granted subset of flags we asked for. */
1095 if ((ocd->ocd_connect_flags & imp->imp_connect_flags_orig) !=
1096 ocd->ocd_connect_flags) {
1097 CERROR("%s: Server didn't grant requested subset of flags: "
1098 "asked=%#llx granted=%#llx\n",
1099 imp->imp_obd->obd_name, imp->imp_connect_flags_orig,
1100 ocd->ocd_connect_flags);
1101 GOTO(out, rc = -EPROTO);
1104 if ((ocd->ocd_connect_flags2 & imp->imp_connect_flags2_orig) !=
1105 ocd->ocd_connect_flags2) {
1106 CERROR("%s: Server didn't grant requested subset of flags2: "
1107 "asked=%#llx granted=%#llx\n",
1108 imp->imp_obd->obd_name, imp->imp_connect_flags2_orig,
1109 ocd->ocd_connect_flags2);
1110 GOTO(out, rc = -EPROTO);
1113 if (!(imp->imp_connect_flags_orig & OBD_CONNECT_LIGHTWEIGHT) &&
1114 (imp->imp_connect_flags_orig & OBD_CONNECT_MDS_MDS) &&
1115 (imp->imp_connect_flags_orig & OBD_CONNECT_FID) &&
1116 (ocd->ocd_connect_flags & OBD_CONNECT_VERSION)) {
1117 __u32 major = OBD_OCD_VERSION_MAJOR(ocd->ocd_version);
1118 __u32 minor = OBD_OCD_VERSION_MINOR(ocd->ocd_version);
1119 __u32 patch = OBD_OCD_VERSION_PATCH(ocd->ocd_version);
1121 /* We do not support the MDT-MDT interoperations with
1122 * different version MDT because of protocol changes. */
1123 if (unlikely(major != LUSTRE_MAJOR ||
1124 minor != LUSTRE_MINOR ||
1125 abs(patch - LUSTRE_PATCH) > 3)) {
1126 LCONSOLE_WARN("%s: import %p (%u.%u.%u.%u) tried the "
1127 "connection to different version MDT "
1128 "(%d.%d.%d.%d) %s\n",
1129 imp->imp_obd->obd_name, imp, LUSTRE_MAJOR,
1130 LUSTRE_MINOR, LUSTRE_PATCH, LUSTRE_FIX,
1131 major, minor, patch,
1132 OBD_OCD_VERSION_FIX(ocd->ocd_version),
1133 imp->imp_connection->c_remote_uuid.uuid);
1135 GOTO(out, rc = -EPROTO);
1139 old_connect_flags = exp_connect_flags(exp);
1140 exp->exp_connect_data = *ocd;
1141 imp->imp_obd->obd_self_export->exp_connect_data = *ocd;
1143 /* The net statistics after (re-)connect is not valid anymore,
1144 * because may reflect other routing, etc. */
1145 at_reinit(&imp->imp_at.iat_net_latency, 0, 0);
1146 ptlrpc_at_adj_net_latency(request,
1147 lustre_msg_get_service_time(
1148 request->rq_repmsg));
1150 /* Import flags should be updated before waking import at FULL state */
1151 rc = ptlrpc_connect_set_flags(imp, ocd, old_connect_flags, exp,
1152 aa->pcaa_initial_connect);
1153 class_export_put(exp);
1159 obd_import_event(imp->imp_obd, imp, IMP_EVENT_OCD);
1161 if (aa->pcaa_initial_connect) {
1162 spin_lock(&imp->imp_lock);
1163 if (msg_flags & MSG_CONNECT_REPLAYABLE) {
1164 imp->imp_replayable = 1;
1165 CDEBUG(D_HA, "connected to replayable target: %s\n",
1166 obd2cli_tgt(imp->imp_obd));
1168 imp->imp_replayable = 0;
1171 /* if applies, adjust the imp->imp_msg_magic here
1172 * according to reply flags
1175 imp->imp_remote_handle =
1176 *lustre_msg_get_handle(request->rq_repmsg);
1178 /* Initial connects are allowed for clients with non-random
1179 * uuids when servers are in recovery. Simply signal the
1180 * servers replay is complete and wait in REPLAY_WAIT.
1182 if (msg_flags & MSG_CONNECT_RECOVERING) {
1183 CDEBUG(D_HA, "connect to %s during recovery\n",
1184 obd2cli_tgt(imp->imp_obd));
1185 import_set_state_nolock(imp, LUSTRE_IMP_REPLAY_LOCKS);
1186 spin_unlock(&imp->imp_lock);
1188 spin_unlock(&imp->imp_lock);
1189 ptlrpc_activate_import(imp, true);
1192 GOTO(finish, rc = 0);
1195 /* Determine what recovery state to move the import to. */
1196 if (MSG_CONNECT_RECONNECT & msg_flags) {
1197 memset(&old_hdl, 0, sizeof(old_hdl));
1198 if (!memcmp(&old_hdl, lustre_msg_get_handle(request->rq_repmsg),
1200 LCONSOLE_WARN("Reconnect to %s (at @%s) failed due "
1201 "bad handle %#llx\n",
1202 obd2cli_tgt(imp->imp_obd),
1203 imp->imp_connection->c_remote_uuid.uuid,
1204 imp->imp_dlm_handle.cookie);
1205 GOTO(out, rc = -ENOTCONN);
1208 if (memcmp(&imp->imp_remote_handle,
1209 lustre_msg_get_handle(request->rq_repmsg),
1210 sizeof(imp->imp_remote_handle))) {
1211 int level = msg_flags & MSG_CONNECT_RECOVERING ?
1214 /* Bug 16611/14775: if server handle have changed,
1215 * that means some sort of disconnection happened.
1216 * If the server is not in recovery, that also means it
1217 * already erased all of our state because of previous
1218 * eviction. If it is in recovery - we are safe to
1219 * participate since we can reestablish all of our state
1222 if ((MSG_CONNECT_RECOVERING & msg_flags)) {
1224 "%s@%s changed server handle from "
1226 " but is still in recovery\n",
1227 obd2cli_tgt(imp->imp_obd),
1228 imp->imp_connection->c_remote_uuid.uuid,
1229 imp->imp_remote_handle.cookie,
1230 lustre_msg_get_handle(
1231 request->rq_repmsg)->cookie);
1233 LCONSOLE_WARN("Evicted from %s (at %s) "
1234 "after server handle changed from "
1236 obd2cli_tgt(imp->imp_obd),
1237 imp->imp_connection->
1239 imp->imp_remote_handle.cookie,
1240 lustre_msg_get_handle(
1241 request->rq_repmsg)->cookie);
1244 imp->imp_remote_handle =
1245 *lustre_msg_get_handle(request->rq_repmsg);
1247 if (!(MSG_CONNECT_RECOVERING & msg_flags)) {
1248 import_set_state(imp, LUSTRE_IMP_EVICTED);
1249 GOTO(finish, rc = 0);
1252 CDEBUG(D_HA, "reconnected to %s@%s after partition\n",
1253 obd2cli_tgt(imp->imp_obd),
1254 imp->imp_connection->c_remote_uuid.uuid);
1257 if (imp->imp_invalid) {
1258 CDEBUG(D_HA, "%s: reconnected but import is invalid; "
1259 "marking evicted\n", imp->imp_obd->obd_name);
1260 import_set_state(imp, LUSTRE_IMP_EVICTED);
1261 } else if (MSG_CONNECT_RECOVERING & msg_flags) {
1262 CDEBUG(D_HA, "%s: reconnected to %s during replay\n",
1263 imp->imp_obd->obd_name,
1264 obd2cli_tgt(imp->imp_obd));
1266 spin_lock(&imp->imp_lock);
1267 imp->imp_resend_replay = 1;
1268 spin_unlock(&imp->imp_lock);
1270 import_set_state(imp, imp->imp_replay_state);
1272 import_set_state(imp, LUSTRE_IMP_RECOVER);
1274 } else if ((MSG_CONNECT_RECOVERING & msg_flags) && !imp->imp_invalid) {
1275 LASSERT(imp->imp_replayable);
1276 imp->imp_remote_handle =
1277 *lustre_msg_get_handle(request->rq_repmsg);
1278 imp->imp_last_replay_transno = 0;
1279 imp->imp_replay_cursor = &imp->imp_committed_list;
1280 import_set_state(imp, LUSTRE_IMP_REPLAY);
1281 } else if ((ocd->ocd_connect_flags & OBD_CONNECT_LIGHTWEIGHT) != 0 &&
1282 !imp->imp_invalid) {
1284 obd_import_event(imp->imp_obd, imp, IMP_EVENT_INVALIDATE);
1285 /* The below message is checked in recovery-small.sh test_106 */
1286 DEBUG_REQ(D_HA, request, "%s: lwp recover",
1287 imp->imp_obd->obd_name);
1288 imp->imp_remote_handle =
1289 *lustre_msg_get_handle(request->rq_repmsg);
1290 import_set_state(imp, LUSTRE_IMP_RECOVER);
1292 DEBUG_REQ(D_HA, request,
1293 "%s: evicting (reconnect/recover flags not set: %x)",
1294 imp->imp_obd->obd_name, msg_flags);
1295 imp->imp_remote_handle =
1296 *lustre_msg_get_handle(request->rq_repmsg);
1297 import_set_state(imp, LUSTRE_IMP_EVICTED);
1300 /* Sanity checks for a reconnected import. */
1301 if (!(imp->imp_replayable) != !(msg_flags & MSG_CONNECT_REPLAYABLE))
1302 CERROR("imp_replayable flag does not match server after reconnect. We should LBUG right here.\n");
1304 if (lustre_msg_get_last_committed(request->rq_repmsg) > 0 &&
1305 lustre_msg_get_last_committed(request->rq_repmsg) <
1306 aa->pcaa_peer_committed) {
1307 static bool printed;
1309 /* The below message is checked in recovery-small.sh test_54 */
1310 CERROR("%s: went back in time (transno %lld was previously committed, server now claims %lld)!\n",
1311 obd2cli_tgt(imp->imp_obd), aa->pcaa_peer_committed,
1312 lustre_msg_get_last_committed(request->rq_repmsg));
1314 CERROR("For further information, see http://doc.lustre.org/lustre_manual.xhtml#went_back_in_time\n");
1320 ptlrpc_prepare_replay(imp);
1321 rc = ptlrpc_import_recovery_state_machine(imp);
1322 if (rc == -ENOTCONN) {
1323 CDEBUG(D_HA, "evicted/aborted by %s@%s during recovery;"
1324 "invalidating and reconnecting\n",
1325 obd2cli_tgt(imp->imp_obd),
1326 imp->imp_connection->c_remote_uuid.uuid);
1327 ptlrpc_connect_import(imp);
1328 spin_lock(&imp->imp_lock);
1329 imp->imp_connected = 0;
1330 imp->imp_connect_tried = 1;
1331 spin_unlock(&imp->imp_lock);
1337 class_export_put(exp);
1339 spin_lock(&imp->imp_lock);
1340 imp->imp_connected = 0;
1341 imp->imp_connect_tried = 1;
1346 import_set_state_nolock(imp, LUSTRE_IMP_DISCON);
1347 if (rc == -EACCES) {
1349 * Give up trying to reconnect
1350 * EACCES means client has no permission for connection
1352 imp->imp_obd->obd_no_recov = 1;
1353 ptlrpc_deactivate_import_nolock(imp);
1355 } else if (rc == -EPROTO) {
1356 struct obd_connect_data *ocd;
1358 /* reply message might not be ready */
1359 if (request->rq_repmsg == NULL) {
1360 spin_unlock(&imp->imp_lock);
1364 ocd = req_capsule_server_get(&request->rq_pill,
1366 /* Servers are not supposed to refuse connections from
1367 * clients based on version, only connection feature
1368 * flags. We should never see this from llite, but it
1369 * may be useful for debugging in the future. */
1371 (ocd->ocd_connect_flags & OBD_CONNECT_VERSION) &&
1372 (ocd->ocd_version != LUSTRE_VERSION_CODE)) {
1373 LCONSOLE_ERROR_MSG(0x16a, "Server %s version "
1375 " refused connection from this client "
1376 "with an incompatible version (%s). "
1377 "Client must be recompiled\n",
1378 obd2cli_tgt(imp->imp_obd),
1379 OBD_OCD_VERSION_MAJOR(ocd->ocd_version),
1380 OBD_OCD_VERSION_MINOR(ocd->ocd_version),
1381 OBD_OCD_VERSION_PATCH(ocd->ocd_version),
1382 OBD_OCD_VERSION_FIX(ocd->ocd_version),
1383 LUSTRE_VERSION_STRING);
1384 ptlrpc_deactivate_import_nolock(imp);
1385 import_set_state_nolock(imp, LUSTRE_IMP_CLOSED);
1389 spin_unlock(&imp->imp_lock);
1392 obd_import_event(imp->imp_obd, imp, IMP_EVENT_INACTIVE);
1397 ptlrpc_maybe_ping_import_soon(imp);
1399 CDEBUG(D_HA, "recovery of %s on %s failed (%d)\n",
1400 obd2cli_tgt(imp->imp_obd),
1401 (char *)imp->imp_connection->c_remote_uuid.uuid, rc);
1403 spin_unlock(&imp->imp_lock);
1406 wake_up_all(&imp->imp_recovery_waitq);
1411 * interpret callback for "completed replay" RPCs.
1412 * \see signal_completed_replay
1414 static int completed_replay_interpret(const struct lu_env *env,
1415 struct ptlrpc_request *req,
1419 atomic_dec(&req->rq_import->imp_replay_inflight);
1420 if (req->rq_status == 0 && !req->rq_import->imp_vbr_failed) {
1421 ptlrpc_import_recovery_state_machine(req->rq_import);
1423 if (req->rq_import->imp_vbr_failed) {
1425 "%s: version recovery fails, reconnecting\n",
1426 req->rq_import->imp_obd->obd_name);
1428 CDEBUG(D_HA, "%s: LAST_REPLAY message error: %d, "
1430 req->rq_import->imp_obd->obd_name,
1433 ptlrpc_connect_import(req->rq_import);
1440 * Let server know that we have no requests to replay anymore.
1441 * Achieved by just sending a PING request
1443 static int signal_completed_replay(struct obd_import *imp)
1445 struct ptlrpc_request *req;
1448 if (unlikely(OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_FINISH_REPLAY)))
1451 LASSERT(atomic_read(&imp->imp_replay_inflight) == 0);
1452 atomic_inc(&imp->imp_replay_inflight);
1454 req = ptlrpc_request_alloc_pack(imp, &RQF_OBD_PING, LUSTRE_OBD_VERSION,
1457 atomic_dec(&imp->imp_replay_inflight);
1461 ptlrpc_request_set_replen(req);
1462 req->rq_send_state = LUSTRE_IMP_REPLAY_WAIT;
1463 lustre_msg_add_flags(req->rq_reqmsg,
1464 MSG_LOCK_REPLAY_DONE | MSG_REQ_REPLAY_DONE);
1466 req->rq_timeout *= 3;
1467 req->rq_interpret_reply = completed_replay_interpret;
1469 ptlrpcd_add_req(req);
1474 * In kernel code all import invalidation happens in its own
1475 * separate thread, so that whatever application happened to encounter
1476 * a problem could still be killed or otherwise continue
1478 static int ptlrpc_invalidate_import_thread(void *data)
1480 struct obd_import *imp = data;
1484 unshare_fs_struct();
1486 CDEBUG(D_HA, "thread invalidate import %s to %s@%s\n",
1487 imp->imp_obd->obd_name, obd2cli_tgt(imp->imp_obd),
1488 imp->imp_connection->c_remote_uuid.uuid);
1490 ptlrpc_invalidate_import(imp);
1492 if (obd_dump_on_eviction) {
1493 CERROR("dump the log upon eviction\n");
1494 libcfs_debug_dumplog();
1497 import_set_state(imp, LUSTRE_IMP_RECOVER);
1498 ptlrpc_import_recovery_state_machine(imp);
1500 class_import_put(imp);
1505 * This is the state machine for client-side recovery on import.
1507 * Typicaly we have two possibly paths. If we came to server and it is not
1508 * in recovery, we just enter IMP_EVICTED state, invalidate our import
1509 * state and reconnect from scratch.
1510 * If we came to server that is in recovery, we enter IMP_REPLAY import state.
1511 * We go through our list of requests to replay and send them to server one by
1513 * After sending all request from the list we change import state to
1514 * IMP_REPLAY_LOCKS and re-request all the locks we believe we have from server
1515 * and also all the locks we don't yet have and wait for server to grant us.
1516 * After that we send a special "replay completed" request and change import
1517 * state to IMP_REPLAY_WAIT.
1518 * Upon receiving reply to that "replay completed" RPC we enter IMP_RECOVER
1519 * state and resend all requests from sending list.
1520 * After that we promote import to FULL state and send all delayed requests
1521 * and import is fully operational after that.
1524 int ptlrpc_import_recovery_state_machine(struct obd_import *imp)
1532 if (imp->imp_state == LUSTRE_IMP_EVICTED) {
1533 deuuidify(obd2cli_tgt(imp->imp_obd), NULL,
1534 &target_start, &target_len);
1535 /* Don't care about MGC eviction */
1536 if (strcmp(imp->imp_obd->obd_type->typ_name,
1537 LUSTRE_MGC_NAME) != 0) {
1538 LCONSOLE_ERROR_MSG(0x167, "%s: This client was evicted "
1539 "by %.*s; in progress operations "
1540 "using this service will fail.\n",
1541 imp->imp_obd->obd_name, target_len,
1543 LASSERTF(!obd_lbug_on_eviction, "LBUG upon eviction");
1545 CDEBUG(D_HA, "evicted from %s@%s; invalidating\n",
1546 obd2cli_tgt(imp->imp_obd),
1547 imp->imp_connection->c_remote_uuid.uuid);
1548 /* reset vbr_failed flag upon eviction */
1549 spin_lock(&imp->imp_lock);
1550 imp->imp_vbr_failed = 0;
1551 spin_unlock(&imp->imp_lock);
1554 struct task_struct *task;
1555 /* bug 17802: XXX client_disconnect_export vs connect request
1556 * race. if client is evicted at this time then we start
1557 * invalidate thread without reference to import and import can
1558 * be freed at same time. */
1559 class_import_get(imp);
1560 task = kthread_run(ptlrpc_invalidate_import_thread, imp,
1563 class_import_put(imp);
1564 CERROR("error starting invalidate thread: %d\n", rc);
1573 if (imp->imp_state == LUSTRE_IMP_REPLAY) {
1574 CDEBUG(D_HA, "replay requested by %s\n",
1575 obd2cli_tgt(imp->imp_obd));
1576 rc = ptlrpc_replay_next(imp, &inflight);
1577 if (inflight == 0 &&
1578 atomic_read(&imp->imp_replay_inflight) == 0) {
1579 import_set_state(imp, LUSTRE_IMP_REPLAY_LOCKS);
1580 rc = ldlm_replay_locks(imp);
1587 if (imp->imp_state == LUSTRE_IMP_REPLAY_LOCKS) {
1588 if (atomic_read(&imp->imp_replay_inflight) == 0) {
1589 import_set_state(imp, LUSTRE_IMP_REPLAY_WAIT);
1590 rc = signal_completed_replay(imp);
1596 if (imp->imp_state == LUSTRE_IMP_REPLAY_WAIT) {
1597 if (atomic_read(&imp->imp_replay_inflight) == 0) {
1598 import_set_state(imp, LUSTRE_IMP_RECOVER);
1602 if (imp->imp_state == LUSTRE_IMP_RECOVER) {
1603 struct ptlrpc_connection *conn = imp->imp_connection;
1605 rc = ptlrpc_resend(imp);
1608 ptlrpc_activate_import(imp, true);
1610 LCONSOLE_INFO("%s: Connection restored to %s (at %s)\n",
1611 imp->imp_obd->obd_name,
1612 obd_uuid2str(&conn->c_remote_uuid),
1613 obd_import_nid2str(imp));
1616 if (imp->imp_state == LUSTRE_IMP_FULL) {
1617 wake_up_all(&imp->imp_recovery_waitq);
1618 ptlrpc_wake_delayed(imp);
1625 static struct ptlrpc_request *ptlrpc_disconnect_prep_req(struct obd_import *imp)
1627 struct ptlrpc_request *req;
1631 switch (imp->imp_connect_op) {
1633 rq_opc = OST_DISCONNECT;
1636 rq_opc = MDS_DISCONNECT;
1639 rq_opc = MGS_DISCONNECT;
1643 CERROR("%s: don't know how to disconnect from %s "
1644 "(connect_op %d): rc = %d\n",
1645 imp->imp_obd->obd_name, obd2cli_tgt(imp->imp_obd),
1646 imp->imp_connect_op, rc);
1647 RETURN(ERR_PTR(rc));
1650 req = ptlrpc_request_alloc_pack(imp, &RQF_MDS_DISCONNECT,
1651 LUSTRE_OBD_VERSION, rq_opc);
1653 RETURN(ERR_PTR(-ENOMEM));
1655 /* We are disconnecting, do not retry a failed DISCONNECT rpc if
1656 * it fails. We can get through the above with a down server
1657 * if the client doesn't know the server is gone yet. */
1658 req->rq_no_resend = 1;
1660 /* We want client umounts to happen quickly, no matter the
1662 req->rq_timeout = min_t(int, req->rq_timeout,
1663 INITIAL_CONNECT_TIMEOUT);
1665 import_set_state(imp, LUSTRE_IMP_CONNECTING);
1666 req->rq_send_state = LUSTRE_IMP_CONNECTING;
1667 ptlrpc_request_set_replen(req);
1672 int ptlrpc_disconnect_import(struct obd_import *imp, int noclose)
1674 struct ptlrpc_request *req;
1678 if (imp->imp_obd->obd_force)
1679 GOTO(set_state, rc);
1681 /* probably the import has been disconnected already being idle */
1682 spin_lock(&imp->imp_lock);
1683 if (imp->imp_state == LUSTRE_IMP_IDLE)
1685 spin_unlock(&imp->imp_lock);
1687 if (ptlrpc_import_in_recovery(imp)) {
1688 struct l_wait_info lwi;
1689 long timeout_jiffies;
1693 if (imp->imp_server_timeout)
1694 timeout = obd_timeout >> 1;
1696 timeout = obd_timeout;
1701 req_portal = imp->imp_client->cli_request_portal;
1702 idx = import_at_get_index(imp, req_portal);
1703 timeout = at_get(&imp->imp_at.iat_service_estimate[idx]);
1706 timeout_jiffies = cfs_time_seconds(timeout);
1707 lwi = LWI_TIMEOUT_INTR(max_t(long, timeout_jiffies, 1),
1708 back_to_sleep, LWI_ON_SIGNAL_NOOP, NULL);
1709 rc = l_wait_event(imp->imp_recovery_waitq,
1710 !ptlrpc_import_in_recovery(imp), &lwi);
1714 spin_lock(&imp->imp_lock);
1715 if (imp->imp_state != LUSTRE_IMP_FULL)
1717 spin_unlock(&imp->imp_lock);
1719 req = ptlrpc_disconnect_prep_req(imp);
1721 GOTO(set_state, rc = PTR_ERR(req));
1722 rc = ptlrpc_queue_wait(req);
1723 ptlrpc_req_finished(req);
1726 spin_lock(&imp->imp_lock);
1729 import_set_state_nolock(imp, LUSTRE_IMP_DISCON);
1731 import_set_state_nolock(imp, LUSTRE_IMP_CLOSED);
1732 memset(&imp->imp_remote_handle, 0, sizeof(imp->imp_remote_handle));
1733 spin_unlock(&imp->imp_lock);
1735 if (rc == -ETIMEDOUT || rc == -ENOTCONN || rc == -ESHUTDOWN)
1739 EXPORT_SYMBOL(ptlrpc_disconnect_import);
1741 static void ptlrpc_reset_reqs_generation(struct obd_import *imp)
1743 struct ptlrpc_request *old, *tmp;
1745 /* tag all resendable requests generated before disconnection
1746 * notice this code is part of disconnect-at-idle path only */
1747 list_for_each_entry_safe(old, tmp, &imp->imp_delayed_list,
1749 spin_lock(&old->rq_lock);
1750 if (old->rq_import_generation == imp->imp_generation - 1 &&
1751 ((imp->imp_initiated_at == imp->imp_generation) ||
1752 !old->rq_no_resend))
1753 old->rq_import_generation = imp->imp_generation;
1754 spin_unlock(&old->rq_lock);
1758 static int ptlrpc_disconnect_idle_interpret(const struct lu_env *env,
1759 struct ptlrpc_request *req,
1762 struct obd_import *imp = req->rq_import;
1765 DEBUG_REQ(D_HA, req, "inflight=%d, refcount=%d: rc = %d",
1766 atomic_read(&imp->imp_inflight),
1767 atomic_read(&imp->imp_refcount), rc);
1769 spin_lock(&imp->imp_lock);
1770 /* DISCONNECT reply can be late and another connection can just
1771 * be initiated. so we have to abort disconnection. */
1772 if (req->rq_import_generation == imp->imp_generation &&
1773 imp->imp_state != LUSTRE_IMP_CLOSED) {
1774 LASSERTF(imp->imp_state == LUSTRE_IMP_CONNECTING,
1775 "%s\n", ptlrpc_import_state_name(imp->imp_state));
1776 imp->imp_state = LUSTRE_IMP_IDLE;
1777 memset(&imp->imp_remote_handle, 0,
1778 sizeof(imp->imp_remote_handle));
1779 /* take our DISCONNECT into account */
1780 if (atomic_read(&imp->imp_inflight) > 1) {
1781 imp->imp_generation++;
1782 imp->imp_initiated_at = imp->imp_generation;
1783 import_set_state_nolock(imp, LUSTRE_IMP_NEW);
1784 ptlrpc_reset_reqs_generation(imp);
1790 rc = ptlrpc_connect_import_locked(imp);
1792 ptlrpc_pinger_add_import(imp);
1794 spin_unlock(&imp->imp_lock);
1800 int ptlrpc_disconnect_and_idle_import(struct obd_import *imp)
1802 struct ptlrpc_request *req;
1805 if (imp->imp_obd->obd_force)
1808 if (ptlrpc_import_in_recovery(imp))
1811 spin_lock(&imp->imp_lock);
1812 if (imp->imp_state != LUSTRE_IMP_FULL) {
1813 spin_unlock(&imp->imp_lock);
1816 spin_unlock(&imp->imp_lock);
1818 req = ptlrpc_disconnect_prep_req(imp);
1820 RETURN(PTR_ERR(req));
1822 CDEBUG_LIMIT(imp->imp_idle_debug, "%s: disconnect after %llus idle\n",
1823 imp->imp_obd->obd_name,
1824 ktime_get_real_seconds() - imp->imp_last_reply_time);
1825 req->rq_interpret_reply = ptlrpc_disconnect_idle_interpret;
1826 ptlrpcd_add_req(req);
1830 EXPORT_SYMBOL(ptlrpc_disconnect_and_idle_import);
1832 void ptlrpc_cleanup_imp(struct obd_import *imp)
1836 spin_lock(&imp->imp_lock);
1838 import_set_state_nolock(imp, LUSTRE_IMP_CLOSED);
1839 imp->imp_generation++;
1840 ptlrpc_abort_inflight(imp);
1842 spin_unlock(&imp->imp_lock);
1847 /* Adaptive Timeout utils */
1849 /* Update at_current with the specified value (bounded by at_min and at_max),
1850 * as well as the AT history "bins".
1851 * - Bin into timeslices using AT_BINS bins.
1852 * - This gives us a max of the last at_history seconds without the storage,
1853 * but still smoothing out a return to normalcy from a slow response.
1854 * - (E.g. remember the maximum latency in each minute of the last 4 minutes.)
1856 int at_measured(struct adaptive_timeout *at, unsigned int val)
1858 unsigned int old = at->at_current;
1859 time64_t now = ktime_get_real_seconds();
1860 long binlimit = max_t(long, at_history / AT_BINS, 1);
1863 CDEBUG(D_OTHER, "add %u to %p time=%lu v=%u (%u %u %u %u)\n",
1864 val, at, (long)(now - at->at_binstart), at->at_current,
1865 at->at_hist[0], at->at_hist[1], at->at_hist[2], at->at_hist[3]);
1868 /* 0's don't count, because we never want our timeout to
1869 drop to 0, and because 0 could mean an error */
1872 spin_lock(&at->at_lock);
1874 if (unlikely(at->at_binstart == 0)) {
1875 /* Special case to remove default from history */
1876 at->at_current = val;
1877 at->at_worst_ever = val;
1878 at->at_worst_time = now;
1879 at->at_hist[0] = val;
1880 at->at_binstart = now;
1881 } else if (now - at->at_binstart < binlimit ) {
1883 at->at_hist[0] = max(val, at->at_hist[0]);
1884 at->at_current = max(val, at->at_current);
1887 unsigned int maxv = val;
1889 /* move bins over */
1890 shift = (u32)(now - at->at_binstart) / binlimit;
1892 for(i = AT_BINS - 1; i >= 0; i--) {
1894 at->at_hist[i] = at->at_hist[i - shift];
1895 maxv = max(maxv, at->at_hist[i]);
1900 at->at_hist[0] = val;
1901 at->at_current = maxv;
1902 at->at_binstart += shift * binlimit;
1905 if (at->at_current > at->at_worst_ever) {
1906 at->at_worst_ever = at->at_current;
1907 at->at_worst_time = now;
1910 if (at->at_flags & AT_FLG_NOHIST)
1911 /* Only keep last reported val; keeping the rest of the history
1913 at->at_current = val;
1916 at->at_current = min(at->at_current, at_max);
1917 at->at_current = max(at->at_current, at_min);
1919 if (at->at_current != old)
1920 CDEBUG(D_OTHER, "AT %p change: old=%u new=%u delta=%d "
1921 "(val=%u) hist %u %u %u %u\n", at,
1922 old, at->at_current, at->at_current - old, val,
1923 at->at_hist[0], at->at_hist[1], at->at_hist[2],
1926 /* if we changed, report the old value */
1927 old = (at->at_current != old) ? old : 0;
1929 spin_unlock(&at->at_lock);
1933 /* Find the imp_at index for a given portal; assign if space available */
1934 int import_at_get_index(struct obd_import *imp, int portal)
1936 struct imp_at *at = &imp->imp_at;
1939 for (i = 0; i < IMP_AT_MAX_PORTALS; i++) {
1940 if (at->iat_portal[i] == portal)
1942 if (at->iat_portal[i] == 0)
1947 /* Not found in list, add it under a lock */
1948 spin_lock(&imp->imp_lock);
1950 /* Check unused under lock */
1951 for (; i < IMP_AT_MAX_PORTALS; i++) {
1952 if (at->iat_portal[i] == portal)
1954 if (at->iat_portal[i] == 0)
1959 /* Not enough portals? */
1960 LASSERT(i < IMP_AT_MAX_PORTALS);
1962 at->iat_portal[i] = portal;
1964 spin_unlock(&imp->imp_lock);