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);
476 ptlrpc_pinger_force(imp);
478 CDEBUG(D_HA, "%s: recovery started, waiting %u seconds\n",
479 obd2cli_tgt(imp->imp_obd), obd_timeout);
481 rc = wait_event_idle_timeout(imp->imp_recovery_waitq,
482 !ptlrpc_import_in_recovery(imp),
488 CDEBUG(D_HA, "%s: recovery finished s:%s\n", obd2cli_tgt(imp->imp_obd),
489 ptlrpc_import_state_name(imp->imp_state));
492 ptlrpc_set_import_discon(imp, 0, false);
493 /* Force a new connect attempt */
494 ptlrpc_invalidate_import(imp);
495 /* Do a fresh connect next time by zeroing the handle */
496 ptlrpc_disconnect_import(imp, 1);
497 /* Wait for all invalidate calls to finish */
498 if (atomic_read(&imp->imp_inval_count) > 0) {
501 rc = l_wait_event_abortable(
502 imp->imp_recovery_waitq,
503 (atomic_read(&imp->imp_inval_count) == 0));
505 CERROR("Interrupted, inval=%d\n",
506 atomic_read(&imp->imp_inval_count));
509 /* Allow reconnect attempts */
510 imp->imp_obd->obd_no_recov = 0;
511 /* Remove 'invalid' flag */
512 ptlrpc_activate_import(imp, false);
513 /* Attempt a new connect */
514 ptlrpc_recover_import(imp, NULL, 0);
518 EXPORT_SYMBOL(ptlrpc_reconnect_import);
521 * Connection on import \a imp is changed to another one (if more than one is
522 * present). We typically chose connection that we have not tried to connect to
525 static int import_select_connection(struct obd_import *imp)
527 struct obd_import_conn *imp_conn = NULL, *conn;
528 struct obd_export *dlmexp;
530 int target_len, tried_all = 1;
533 spin_lock(&imp->imp_lock);
535 if (list_empty(&imp->imp_conn_list)) {
536 CERROR("%s: no connections available\n",
537 imp->imp_obd->obd_name);
538 spin_unlock(&imp->imp_lock);
542 list_for_each_entry(conn, &imp->imp_conn_list, oic_item) {
543 CDEBUG(D_HA, "%s: connect to NID %s last attempt %lld\n",
544 imp->imp_obd->obd_name,
545 libcfs_nid2str(conn->oic_conn->c_peer.nid),
546 conn->oic_last_attempt);
548 /* If we have not tried this connection since
549 the last successful attempt, go with this one */
550 if ((conn->oic_last_attempt == 0) ||
551 conn->oic_last_attempt <= imp->imp_last_success_conn) {
557 /* If all of the connections have already been tried
558 since the last successful connection; just choose the
559 least recently used */
562 else if (imp_conn->oic_last_attempt > conn->oic_last_attempt)
566 /* if not found, simply choose the current one */
567 if (!imp_conn || imp->imp_force_reconnect) {
568 LASSERT(imp->imp_conn_current);
569 imp_conn = imp->imp_conn_current;
572 LASSERT(imp_conn->oic_conn);
574 /* If we've tried everything, and we're back to the beginning of the
575 list, increase our timeout and try again. It will be reset when
576 we do finally connect. (FIXME: really we should wait for all network
577 state associated with the last connection attempt to drain before
578 trying to reconnect on it.) */
579 if (tried_all && (imp->imp_conn_list.next == &imp_conn->oic_item)) {
580 struct adaptive_timeout *at = &imp->imp_at.iat_net_latency;
581 if (at_get(at) < CONNECTION_SWITCH_MAX) {
582 at_measured(at, at_get(at) + CONNECTION_SWITCH_INC);
583 if (at_get(at) > CONNECTION_SWITCH_MAX)
584 at_reset(at, CONNECTION_SWITCH_MAX);
586 LASSERT(imp_conn->oic_last_attempt);
587 CDEBUG(D_HA, "%s: tried all connections, increasing latency "
588 "to %ds\n", imp->imp_obd->obd_name, at_get(at));
591 imp_conn->oic_last_attempt = ktime_get_seconds();
593 /* switch connection, don't mind if it's same as the current one */
594 if (imp->imp_connection)
595 ptlrpc_connection_put(imp->imp_connection);
596 imp->imp_connection = ptlrpc_connection_addref(imp_conn->oic_conn);
598 dlmexp = class_conn2export(&imp->imp_dlm_handle);
599 LASSERT(dlmexp != NULL);
600 if (dlmexp->exp_connection)
601 ptlrpc_connection_put(dlmexp->exp_connection);
602 dlmexp->exp_connection = ptlrpc_connection_addref(imp_conn->oic_conn);
603 class_export_put(dlmexp);
605 if (imp->imp_conn_current != imp_conn) {
606 if (imp->imp_conn_current) {
607 deuuidify(obd2cli_tgt(imp->imp_obd), NULL,
608 &target_start, &target_len);
610 CDEBUG(D_HA, "%s: Connection changing to"
612 imp->imp_obd->obd_name,
613 target_len, target_start,
614 libcfs_nid2str(imp_conn->oic_conn->c_peer.nid));
617 imp->imp_conn_current = imp_conn;
620 /* The below message is checked in conf-sanity.sh test_35[ab] */
621 CDEBUG(D_HA, "%s: import %p using connection %s/%s\n",
622 imp->imp_obd->obd_name, imp, imp_conn->oic_uuid.uuid,
623 libcfs_nid2str(imp_conn->oic_conn->c_peer.nid));
625 spin_unlock(&imp->imp_lock);
631 * must be called under imp_lock
633 static int ptlrpc_first_transno(struct obd_import *imp, __u64 *transno)
635 struct ptlrpc_request *req;
636 struct list_head *tmp;
638 /* The requests in committed_list always have smaller transnos than
639 * the requests in replay_list */
640 if (!list_empty(&imp->imp_committed_list)) {
641 tmp = imp->imp_committed_list.next;
642 req = list_entry(tmp, struct ptlrpc_request, rq_replay_list);
643 *transno = req->rq_transno;
644 if (req->rq_transno == 0) {
645 DEBUG_REQ(D_ERROR, req,
646 "zero transno in committed_list");
651 if (!list_empty(&imp->imp_replay_list)) {
652 tmp = imp->imp_replay_list.next;
653 req = list_entry(tmp, struct ptlrpc_request, rq_replay_list);
654 *transno = req->rq_transno;
655 if (req->rq_transno == 0) {
656 DEBUG_REQ(D_ERROR, req, "zero transno in replay_list");
664 int ptlrpc_connect_import(struct obd_import *imp)
666 spin_lock(&imp->imp_lock);
667 return ptlrpc_connect_import_locked(imp);
671 * Attempt to (re)connect import \a imp. This includes all preparations,
672 * initializing CONNECT RPC request and passing it to ptlrpcd for
675 * Assumes imp->imp_lock is held, and releases it.
677 * Returns 0 on success or error code.
679 int ptlrpc_connect_import_locked(struct obd_import *imp)
681 struct obd_device *obd = imp->imp_obd;
682 int initial_connect = 0;
684 __u64 committed_before_reconnect = 0;
685 struct ptlrpc_request *request;
686 struct obd_connect_data ocd;
687 char *bufs[] = { NULL,
688 obd2cli_tgt(imp->imp_obd),
690 (char *)&imp->imp_dlm_handle,
693 struct ptlrpc_connect_async_args *aa;
697 assert_spin_locked(&imp->imp_lock);
699 if (imp->imp_state == LUSTRE_IMP_CLOSED) {
700 spin_unlock(&imp->imp_lock);
701 CERROR("can't connect to a closed import\n");
703 } else if (imp->imp_state == LUSTRE_IMP_FULL) {
704 spin_unlock(&imp->imp_lock);
705 CERROR("already connected\n");
707 } else if (imp->imp_state == LUSTRE_IMP_CONNECTING ||
708 imp->imp_state == LUSTRE_IMP_EVICTED ||
709 imp->imp_connected) {
710 spin_unlock(&imp->imp_lock);
711 CERROR("already connecting\n");
715 import_set_state_nolock(imp, LUSTRE_IMP_CONNECTING);
718 imp->imp_resend_replay = 0;
720 if (!lustre_handle_is_used(&imp->imp_remote_handle))
723 committed_before_reconnect = imp->imp_peer_committed_transno;
725 set_transno = ptlrpc_first_transno(imp,
726 &imp->imp_connect_data.ocd_transno);
727 spin_unlock(&imp->imp_lock);
729 rc = import_select_connection(imp);
733 rc = sptlrpc_import_sec_adapt(imp, NULL, NULL);
737 /* Reset connect flags to the originally requested flags, in case
738 * the server is updated on-the-fly we will get the new features. */
739 ocd = imp->imp_connect_data;
740 ocd.ocd_connect_flags = imp->imp_connect_flags_orig;
741 ocd.ocd_connect_flags2 = imp->imp_connect_flags2_orig;
742 /* Reset ocd_version each time so the server knows the exact versions */
743 ocd.ocd_version = LUSTRE_VERSION_CODE;
744 imp->imp_msghdr_flags &= ~MSGHDR_AT_SUPPORT;
745 imp->imp_msghdr_flags &= ~MSGHDR_CKSUM_INCOMPAT18;
747 rc = obd_reconnect(NULL, imp->imp_obd->obd_self_export, obd,
748 &obd->obd_uuid, &ocd, NULL);
752 request = ptlrpc_request_alloc(imp, &RQF_MDS_CONNECT);
754 GOTO(out, rc = -ENOMEM);
756 /* get SELinux policy info if any */
757 rc = sptlrpc_get_sepol(request);
759 ptlrpc_request_free(request);
763 bufs[5] = request->rq_sepol;
765 req_capsule_set_size(&request->rq_pill, &RMF_SELINUX_POL, RCL_CLIENT,
766 strlen(request->rq_sepol) ?
767 strlen(request->rq_sepol) + 1 : 0);
769 rc = ptlrpc_request_bufs_pack(request, LUSTRE_OBD_VERSION,
770 imp->imp_connect_op, bufs, NULL);
772 ptlrpc_request_free(request);
776 /* Report the rpc service time to the server so that it knows how long
777 * to wait for clients to join recovery */
778 lustre_msg_set_service_time(request->rq_reqmsg,
779 at_timeout2est(request->rq_timeout));
781 /* The amount of time we give the server to process the connect req.
782 * import_select_connection will increase the net latency on
783 * repeated reconnect attempts to cover slow networks.
784 * We override/ignore the server rpc completion estimate here,
785 * which may be large if this is a reconnect attempt */
786 request->rq_timeout = INITIAL_CONNECT_TIMEOUT;
787 lustre_msg_set_timeout(request->rq_reqmsg, request->rq_timeout);
789 request->rq_no_resend = request->rq_no_delay = 1;
790 request->rq_send_state = LUSTRE_IMP_CONNECTING;
791 /* Allow a slightly larger reply for future growth compatibility */
792 req_capsule_set_size(&request->rq_pill, &RMF_CONNECT_DATA, RCL_SERVER,
793 sizeof(struct obd_connect_data)+16*sizeof(__u64));
794 ptlrpc_request_set_replen(request);
795 request->rq_interpret_reply = ptlrpc_connect_interpret;
797 aa = ptlrpc_req_async_args(aa, request);
798 memset(aa, 0, sizeof *aa);
800 aa->pcaa_peer_committed = committed_before_reconnect;
801 aa->pcaa_initial_connect = initial_connect;
803 if (aa->pcaa_initial_connect) {
804 spin_lock(&imp->imp_lock);
805 imp->imp_replayable = 1;
806 spin_unlock(&imp->imp_lock);
807 lustre_msg_add_op_flags(request->rq_reqmsg,
808 MSG_CONNECT_INITIAL);
812 lustre_msg_add_op_flags(request->rq_reqmsg,
813 MSG_CONNECT_TRANSNO);
815 DEBUG_REQ(D_RPCTRACE, request, "(re)connect request (timeout %ld)",
816 request->rq_timeout);
817 ptlrpcd_add_req(request);
821 import_set_state(imp, LUSTRE_IMP_DISCON);
825 EXPORT_SYMBOL(ptlrpc_connect_import);
827 static void ptlrpc_maybe_ping_import_soon(struct obd_import *imp)
831 spin_lock(&imp->imp_lock);
832 force_verify = imp->imp_force_verify != 0;
833 spin_unlock(&imp->imp_lock);
836 ptlrpc_pinger_wake_up();
839 static int ptlrpc_busy_reconnect(int rc)
841 return (rc == -EBUSY) || (rc == -EAGAIN);
844 static int ptlrpc_connect_set_flags(struct obd_import *imp,
845 struct obd_connect_data *ocd,
846 __u64 old_connect_flags,
847 struct obd_export *exp, int init_connect)
850 struct client_obd *cli = &imp->imp_obd->u.cli;
852 spin_lock(&imp->imp_lock);
853 list_move(&imp->imp_conn_current->oic_item,
854 &imp->imp_conn_list);
855 imp->imp_last_success_conn =
856 imp->imp_conn_current->oic_last_attempt;
858 spin_unlock(&imp->imp_lock);
860 if (!warned && (ocd->ocd_connect_flags & OBD_CONNECT_VERSION) &&
861 (ocd->ocd_version > LUSTRE_VERSION_CODE +
862 LUSTRE_VERSION_OFFSET_WARN ||
863 ocd->ocd_version < LUSTRE_VERSION_CODE -
864 LUSTRE_VERSION_OFFSET_WARN)) {
865 /* Sigh, some compilers do not like #ifdef in the middle
866 of macro arguments */
867 const char *older = "older than client. "
868 "Consider upgrading server";
869 const char *newer = "newer than client. "
870 "Consider upgrading client";
872 LCONSOLE_WARN("Server %s version (%d.%d.%d.%d) "
874 obd2cli_tgt(imp->imp_obd),
875 OBD_OCD_VERSION_MAJOR(ocd->ocd_version),
876 OBD_OCD_VERSION_MINOR(ocd->ocd_version),
877 OBD_OCD_VERSION_PATCH(ocd->ocd_version),
878 OBD_OCD_VERSION_FIX(ocd->ocd_version),
879 ocd->ocd_version > LUSTRE_VERSION_CODE ?
880 newer : older, LUSTRE_VERSION_STRING);
884 if (ocd->ocd_connect_flags & OBD_CONNECT_CKSUM) {
885 /* We sent to the server ocd_cksum_types with bits set
886 * for algorithms we understand. The server masked off
887 * the checksum types it doesn't support */
888 if ((ocd->ocd_cksum_types &
889 obd_cksum_types_supported_client()) == 0) {
890 LCONSOLE_ERROR("The negotiation of the checksum "
891 "alogrithm to use with server %s "
893 obd2cli_tgt(imp->imp_obd),
894 ocd->ocd_cksum_types,
895 obd_cksum_types_supported_client());
898 cli->cl_supp_cksum_types = ocd->ocd_cksum_types;
901 /* The server does not support OBD_CONNECT_CKSUM.
902 * Enforce ADLER for backward compatibility*/
903 cli->cl_supp_cksum_types = OBD_CKSUM_ADLER;
905 cli->cl_cksum_type = obd_cksum_type_select(imp->imp_obd->obd_name,
906 cli->cl_supp_cksum_types,
907 cli->cl_preferred_cksum_type);
909 if (ocd->ocd_connect_flags & OBD_CONNECT_BRW_SIZE)
910 cli->cl_max_pages_per_rpc =
911 min(ocd->ocd_brw_size >> PAGE_SHIFT,
912 cli->cl_max_pages_per_rpc);
913 else if (imp->imp_connect_op == MDS_CONNECT ||
914 imp->imp_connect_op == MGS_CONNECT)
915 cli->cl_max_pages_per_rpc = 1;
917 LASSERT((cli->cl_max_pages_per_rpc <= PTLRPC_MAX_BRW_PAGES) &&
918 (cli->cl_max_pages_per_rpc > 0));
920 client_adjust_max_dirty(cli);
922 /* Update client max modify RPCs in flight with value returned
924 if (ocd->ocd_connect_flags & OBD_CONNECT_MULTIMODRPCS)
925 cli->cl_max_mod_rpcs_in_flight = min(
926 cli->cl_max_mod_rpcs_in_flight,
927 ocd->ocd_maxmodrpcs);
929 cli->cl_max_mod_rpcs_in_flight = 1;
931 /* Reset ns_connect_flags only for initial connect. It might be
932 * changed in while using FS and if we reset it in reconnect
933 * this leads to losing user settings done before such as
934 * disable lru_resize, etc. */
935 if (old_connect_flags != exp_connect_flags(exp) || init_connect) {
936 struct ldlm_namespace *ns = imp->imp_obd->obd_namespace;
940 ns->ns_connect_flags ^ ns->ns_orig_connect_flags;
941 CDEBUG(D_HA, "%s: Resetting ns_connect_flags to server "
942 "flags: %#llx\n", imp->imp_obd->obd_name,
943 ocd->ocd_connect_flags);
944 ns->ns_connect_flags = (ns->ns_connect_flags & changed_flags) |
945 (ocd->ocd_connect_flags & ~changed_flags);
946 ns->ns_orig_connect_flags = ocd->ocd_connect_flags;
949 if (ocd->ocd_connect_flags & OBD_CONNECT_AT)
950 /* We need a per-message support flag, because
951 * a. we don't know if the incoming connect reply
952 * supports AT or not (in reply_in_callback)
953 * until we unpack it.
954 * b. failovered server means export and flags are gone
955 * (in ptlrpc_send_reply).
956 * Can only be set when we know AT is supported at
958 imp->imp_msghdr_flags |= MSGHDR_AT_SUPPORT;
960 imp->imp_msghdr_flags &= ~MSGHDR_AT_SUPPORT;
962 imp->imp_msghdr_flags |= MSGHDR_CKSUM_INCOMPAT18;
968 * Add all replay requests back to unreplied list before start replay,
969 * so that we can make sure the known replied XID is always increased
970 * only even if when replaying requests.
972 static void ptlrpc_prepare_replay(struct obd_import *imp)
974 struct ptlrpc_request *req;
976 if (imp->imp_state != LUSTRE_IMP_REPLAY ||
977 imp->imp_resend_replay)
980 /* If the server was restart during repaly, the requests may
981 * have been added to the unreplied list in former replay. */
982 spin_lock(&imp->imp_lock);
984 list_for_each_entry(req, &imp->imp_committed_list, rq_replay_list) {
985 if (list_empty(&req->rq_unreplied_list))
986 ptlrpc_add_unreplied(req);
989 list_for_each_entry(req, &imp->imp_replay_list, rq_replay_list) {
990 if (list_empty(&req->rq_unreplied_list))
991 ptlrpc_add_unreplied(req);
994 imp->imp_known_replied_xid = ptlrpc_known_replied_xid(imp);
995 spin_unlock(&imp->imp_lock);
999 * interpret_reply callback for connect RPCs.
1000 * Looks into returned status of connect operation and decides
1001 * what to do with the import - i.e enter recovery, promote it to
1002 * full state for normal operations of disconnect it due to an error.
1004 static int ptlrpc_connect_interpret(const struct lu_env *env,
1005 struct ptlrpc_request *request,
1008 struct ptlrpc_connect_async_args *aa = data;
1009 struct obd_import *imp = request->rq_import;
1010 struct lustre_handle old_hdl;
1011 __u64 old_connect_flags;
1013 struct obd_connect_data *ocd;
1014 struct obd_export *exp = NULL;
1018 spin_lock(&imp->imp_lock);
1019 if (imp->imp_state == LUSTRE_IMP_CLOSED) {
1020 imp->imp_connect_tried = 1;
1021 spin_unlock(&imp->imp_lock);
1025 imp->imp_connect_error = rc;
1027 struct ptlrpc_request *free_req;
1028 struct ptlrpc_request *tmp;
1030 /* abort all delayed requests initiated connection */
1031 list_for_each_entry_safe(free_req, tmp, &imp->imp_delayed_list,
1033 spin_lock(&free_req->rq_lock);
1034 if (free_req->rq_no_resend) {
1035 free_req->rq_err = 1;
1036 free_req->rq_status = -EIO;
1037 ptlrpc_client_wake_req(free_req);
1039 spin_unlock(&free_req->rq_lock);
1042 /* if this reconnect to busy export - not need select new target
1044 imp->imp_force_reconnect = ptlrpc_busy_reconnect(rc);
1045 spin_unlock(&imp->imp_lock);
1046 ptlrpc_maybe_ping_import_soon(imp);
1050 /* LU-7558: indicate that we are interpretting connect reply,
1051 * pltrpc_connect_import() will not try to reconnect until
1052 * interpret will finish. */
1053 imp->imp_connected = 1;
1054 spin_unlock(&imp->imp_lock);
1056 LASSERT(imp->imp_conn_current);
1058 msg_flags = lustre_msg_get_op_flags(request->rq_repmsg);
1060 ret = req_capsule_get_size(&request->rq_pill, &RMF_CONNECT_DATA,
1062 /* server replied obd_connect_data is always bigger */
1063 ocd = req_capsule_server_sized_get(&request->rq_pill,
1064 &RMF_CONNECT_DATA, ret);
1067 CERROR("%s: no connect data from server\n",
1068 imp->imp_obd->obd_name);
1073 spin_lock(&imp->imp_lock);
1075 /* All imports are pingable */
1076 imp->imp_pingable = 1;
1077 imp->imp_force_reconnect = 0;
1078 imp->imp_force_verify = 0;
1080 imp->imp_connect_data = *ocd;
1082 CDEBUG(D_HA, "%s: connect to target with instance %u\n",
1083 imp->imp_obd->obd_name, ocd->ocd_instance);
1084 exp = class_conn2export(&imp->imp_dlm_handle);
1086 spin_unlock(&imp->imp_lock);
1089 /* This could happen if export is cleaned during the
1091 CERROR("%s: missing export after connect\n",
1092 imp->imp_obd->obd_name);
1093 GOTO(out, rc = -ENODEV);
1096 /* check that server granted subset of flags we asked for. */
1097 if ((ocd->ocd_connect_flags & imp->imp_connect_flags_orig) !=
1098 ocd->ocd_connect_flags) {
1099 CERROR("%s: Server didn't grant requested subset of flags: "
1100 "asked=%#llx granted=%#llx\n",
1101 imp->imp_obd->obd_name, imp->imp_connect_flags_orig,
1102 ocd->ocd_connect_flags);
1103 GOTO(out, rc = -EPROTO);
1106 if ((ocd->ocd_connect_flags2 & imp->imp_connect_flags2_orig) !=
1107 ocd->ocd_connect_flags2) {
1108 CERROR("%s: Server didn't grant requested subset of flags2: "
1109 "asked=%#llx granted=%#llx\n",
1110 imp->imp_obd->obd_name, imp->imp_connect_flags2_orig,
1111 ocd->ocd_connect_flags2);
1112 GOTO(out, rc = -EPROTO);
1115 if (!(imp->imp_connect_flags_orig & OBD_CONNECT_LIGHTWEIGHT) &&
1116 (imp->imp_connect_flags_orig & OBD_CONNECT_MDS_MDS) &&
1117 (imp->imp_connect_flags_orig & OBD_CONNECT_FID) &&
1118 (ocd->ocd_connect_flags & OBD_CONNECT_VERSION)) {
1119 __u32 major = OBD_OCD_VERSION_MAJOR(ocd->ocd_version);
1120 __u32 minor = OBD_OCD_VERSION_MINOR(ocd->ocd_version);
1121 __u32 patch = OBD_OCD_VERSION_PATCH(ocd->ocd_version);
1123 /* We do not support the MDT-MDT interoperations with
1124 * different version MDT because of protocol changes. */
1125 if (unlikely(major != LUSTRE_MAJOR ||
1126 minor != LUSTRE_MINOR ||
1127 abs(patch - LUSTRE_PATCH) > 3)) {
1128 LCONSOLE_WARN("%s: import %p (%u.%u.%u.%u) tried the "
1129 "connection to different version MDT "
1130 "(%d.%d.%d.%d) %s\n",
1131 imp->imp_obd->obd_name, imp, LUSTRE_MAJOR,
1132 LUSTRE_MINOR, LUSTRE_PATCH, LUSTRE_FIX,
1133 major, minor, patch,
1134 OBD_OCD_VERSION_FIX(ocd->ocd_version),
1135 imp->imp_connection->c_remote_uuid.uuid);
1137 GOTO(out, rc = -EPROTO);
1141 old_connect_flags = exp_connect_flags(exp);
1142 exp->exp_connect_data = *ocd;
1143 imp->imp_obd->obd_self_export->exp_connect_data = *ocd;
1145 /* The net statistics after (re-)connect is not valid anymore,
1146 * because may reflect other routing, etc. */
1147 at_reinit(&imp->imp_at.iat_net_latency, 0, 0);
1148 ptlrpc_at_adj_net_latency(request,
1149 lustre_msg_get_service_time(
1150 request->rq_repmsg));
1152 /* Import flags should be updated before waking import at FULL state */
1153 rc = ptlrpc_connect_set_flags(imp, ocd, old_connect_flags, exp,
1154 aa->pcaa_initial_connect);
1155 class_export_put(exp);
1161 obd_import_event(imp->imp_obd, imp, IMP_EVENT_OCD);
1163 if (aa->pcaa_initial_connect) {
1164 spin_lock(&imp->imp_lock);
1165 if (msg_flags & MSG_CONNECT_REPLAYABLE) {
1166 imp->imp_replayable = 1;
1167 CDEBUG(D_HA, "connected to replayable target: %s\n",
1168 obd2cli_tgt(imp->imp_obd));
1170 imp->imp_replayable = 0;
1173 /* if applies, adjust the imp->imp_msg_magic here
1174 * according to reply flags
1177 imp->imp_remote_handle =
1178 *lustre_msg_get_handle(request->rq_repmsg);
1180 /* Initial connects are allowed for clients with non-random
1181 * uuids when servers are in recovery. Simply signal the
1182 * servers replay is complete and wait in REPLAY_WAIT.
1184 if (msg_flags & MSG_CONNECT_RECOVERING) {
1185 CDEBUG(D_HA, "connect to %s during recovery\n",
1186 obd2cli_tgt(imp->imp_obd));
1187 import_set_state_nolock(imp, LUSTRE_IMP_REPLAY_LOCKS);
1188 spin_unlock(&imp->imp_lock);
1190 spin_unlock(&imp->imp_lock);
1191 ptlrpc_activate_import(imp, true);
1194 GOTO(finish, rc = 0);
1197 /* Determine what recovery state to move the import to. */
1198 if (MSG_CONNECT_RECONNECT & msg_flags) {
1199 memset(&old_hdl, 0, sizeof(old_hdl));
1200 if (!memcmp(&old_hdl, lustre_msg_get_handle(request->rq_repmsg),
1202 LCONSOLE_WARN("Reconnect to %s (at @%s) failed due "
1203 "bad handle %#llx\n",
1204 obd2cli_tgt(imp->imp_obd),
1205 imp->imp_connection->c_remote_uuid.uuid,
1206 imp->imp_dlm_handle.cookie);
1207 GOTO(out, rc = -ENOTCONN);
1210 if (memcmp(&imp->imp_remote_handle,
1211 lustre_msg_get_handle(request->rq_repmsg),
1212 sizeof(imp->imp_remote_handle))) {
1213 int level = msg_flags & MSG_CONNECT_RECOVERING ?
1216 /* Bug 16611/14775: if server handle have changed,
1217 * that means some sort of disconnection happened.
1218 * If the server is not in recovery, that also means it
1219 * already erased all of our state because of previous
1220 * eviction. If it is in recovery - we are safe to
1221 * participate since we can reestablish all of our state
1224 if ((MSG_CONNECT_RECOVERING & msg_flags)) {
1226 "%s@%s changed server handle from "
1228 " but is still in recovery\n",
1229 obd2cli_tgt(imp->imp_obd),
1230 imp->imp_connection->c_remote_uuid.uuid,
1231 imp->imp_remote_handle.cookie,
1232 lustre_msg_get_handle(
1233 request->rq_repmsg)->cookie);
1235 LCONSOLE_WARN("Evicted from %s (at %s) "
1236 "after server handle changed from "
1238 obd2cli_tgt(imp->imp_obd),
1239 imp->imp_connection->
1241 imp->imp_remote_handle.cookie,
1242 lustre_msg_get_handle(
1243 request->rq_repmsg)->cookie);
1246 imp->imp_remote_handle =
1247 *lustre_msg_get_handle(request->rq_repmsg);
1249 if (!(MSG_CONNECT_RECOVERING & msg_flags)) {
1250 import_set_state(imp, LUSTRE_IMP_EVICTED);
1251 GOTO(finish, rc = 0);
1254 CDEBUG(D_HA, "reconnected to %s@%s after partition\n",
1255 obd2cli_tgt(imp->imp_obd),
1256 imp->imp_connection->c_remote_uuid.uuid);
1259 if (imp->imp_invalid) {
1260 CDEBUG(D_HA, "%s: reconnected but import is invalid; "
1261 "marking evicted\n", imp->imp_obd->obd_name);
1262 import_set_state(imp, LUSTRE_IMP_EVICTED);
1263 } else if (MSG_CONNECT_RECOVERING & msg_flags) {
1264 CDEBUG(D_HA, "%s: reconnected to %s during replay\n",
1265 imp->imp_obd->obd_name,
1266 obd2cli_tgt(imp->imp_obd));
1268 spin_lock(&imp->imp_lock);
1269 imp->imp_resend_replay = 1;
1270 spin_unlock(&imp->imp_lock);
1272 import_set_state(imp, imp->imp_replay_state);
1274 import_set_state(imp, LUSTRE_IMP_RECOVER);
1276 } else if ((MSG_CONNECT_RECOVERING & msg_flags) && !imp->imp_invalid) {
1277 LASSERT(imp->imp_replayable);
1278 imp->imp_remote_handle =
1279 *lustre_msg_get_handle(request->rq_repmsg);
1280 imp->imp_last_replay_transno = 0;
1281 imp->imp_replay_cursor = &imp->imp_committed_list;
1282 import_set_state(imp, LUSTRE_IMP_REPLAY);
1283 } else if ((ocd->ocd_connect_flags & OBD_CONNECT_LIGHTWEIGHT) != 0 &&
1284 !imp->imp_invalid) {
1286 obd_import_event(imp->imp_obd, imp, IMP_EVENT_INVALIDATE);
1287 /* The below message is checked in recovery-small.sh test_106 */
1288 DEBUG_REQ(D_HA, request, "%s: lwp recover",
1289 imp->imp_obd->obd_name);
1290 imp->imp_remote_handle =
1291 *lustre_msg_get_handle(request->rq_repmsg);
1292 import_set_state(imp, LUSTRE_IMP_RECOVER);
1294 DEBUG_REQ(D_HA, request,
1295 "%s: evicting (reconnect/recover flags not set: %x)",
1296 imp->imp_obd->obd_name, msg_flags);
1297 imp->imp_remote_handle =
1298 *lustre_msg_get_handle(request->rq_repmsg);
1299 import_set_state(imp, LUSTRE_IMP_EVICTED);
1302 /* Sanity checks for a reconnected import. */
1303 if (!(imp->imp_replayable) != !(msg_flags & MSG_CONNECT_REPLAYABLE))
1304 CERROR("imp_replayable flag does not match server after reconnect. We should LBUG right here.\n");
1306 if (lustre_msg_get_last_committed(request->rq_repmsg) > 0 &&
1307 lustre_msg_get_last_committed(request->rq_repmsg) <
1308 aa->pcaa_peer_committed) {
1309 static bool printed;
1311 /* The below message is checked in recovery-small.sh test_54 */
1312 CERROR("%s: went back in time (transno %lld was previously committed, server now claims %lld)!\n",
1313 obd2cli_tgt(imp->imp_obd), aa->pcaa_peer_committed,
1314 lustre_msg_get_last_committed(request->rq_repmsg));
1316 CERROR("For further information, see http://doc.lustre.org/lustre_manual.xhtml#went_back_in_time\n");
1322 ptlrpc_prepare_replay(imp);
1323 rc = ptlrpc_import_recovery_state_machine(imp);
1324 if (rc == -ENOTCONN) {
1325 CDEBUG(D_HA, "evicted/aborted by %s@%s during recovery;"
1326 "invalidating and reconnecting\n",
1327 obd2cli_tgt(imp->imp_obd),
1328 imp->imp_connection->c_remote_uuid.uuid);
1329 ptlrpc_connect_import(imp);
1330 spin_lock(&imp->imp_lock);
1331 imp->imp_connected = 0;
1332 imp->imp_connect_tried = 1;
1333 spin_unlock(&imp->imp_lock);
1339 class_export_put(exp);
1341 spin_lock(&imp->imp_lock);
1342 imp->imp_connected = 0;
1343 imp->imp_connect_tried = 1;
1348 import_set_state_nolock(imp, LUSTRE_IMP_DISCON);
1349 if (rc == -EACCES) {
1351 * Give up trying to reconnect
1352 * EACCES means client has no permission for connection
1354 imp->imp_obd->obd_no_recov = 1;
1355 ptlrpc_deactivate_import_nolock(imp);
1357 } else if (rc == -EPROTO) {
1358 struct obd_connect_data *ocd;
1360 /* reply message might not be ready */
1361 if (request->rq_repmsg == NULL) {
1362 spin_unlock(&imp->imp_lock);
1366 ocd = req_capsule_server_get(&request->rq_pill,
1368 /* Servers are not supposed to refuse connections from
1369 * clients based on version, only connection feature
1370 * flags. We should never see this from llite, but it
1371 * may be useful for debugging in the future. */
1373 (ocd->ocd_connect_flags & OBD_CONNECT_VERSION) &&
1374 (ocd->ocd_version != LUSTRE_VERSION_CODE)) {
1375 LCONSOLE_ERROR_MSG(0x16a, "Server %s version "
1377 " refused connection from this client "
1378 "with an incompatible version (%s). "
1379 "Client must be recompiled\n",
1380 obd2cli_tgt(imp->imp_obd),
1381 OBD_OCD_VERSION_MAJOR(ocd->ocd_version),
1382 OBD_OCD_VERSION_MINOR(ocd->ocd_version),
1383 OBD_OCD_VERSION_PATCH(ocd->ocd_version),
1384 OBD_OCD_VERSION_FIX(ocd->ocd_version),
1385 LUSTRE_VERSION_STRING);
1386 ptlrpc_deactivate_import_nolock(imp);
1387 import_set_state_nolock(imp, LUSTRE_IMP_CLOSED);
1391 spin_unlock(&imp->imp_lock);
1394 obd_import_event(imp->imp_obd, imp, IMP_EVENT_INACTIVE);
1399 ptlrpc_maybe_ping_import_soon(imp);
1401 CDEBUG(D_HA, "recovery of %s on %s failed (%d)\n",
1402 obd2cli_tgt(imp->imp_obd),
1403 (char *)imp->imp_connection->c_remote_uuid.uuid, rc);
1405 spin_unlock(&imp->imp_lock);
1408 wake_up_all(&imp->imp_recovery_waitq);
1413 * interpret callback for "completed replay" RPCs.
1414 * \see signal_completed_replay
1416 static int completed_replay_interpret(const struct lu_env *env,
1417 struct ptlrpc_request *req,
1421 atomic_dec(&req->rq_import->imp_replay_inflight);
1422 if (req->rq_status == 0 && !req->rq_import->imp_vbr_failed) {
1423 ptlrpc_import_recovery_state_machine(req->rq_import);
1425 if (req->rq_import->imp_vbr_failed) {
1427 "%s: version recovery fails, reconnecting\n",
1428 req->rq_import->imp_obd->obd_name);
1430 CDEBUG(D_HA, "%s: LAST_REPLAY message error: %d, "
1432 req->rq_import->imp_obd->obd_name,
1435 ptlrpc_connect_import(req->rq_import);
1442 * Let server know that we have no requests to replay anymore.
1443 * Achieved by just sending a PING request
1445 static int signal_completed_replay(struct obd_import *imp)
1447 struct ptlrpc_request *req;
1450 if (unlikely(OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_FINISH_REPLAY)))
1453 LASSERT(atomic_read(&imp->imp_replay_inflight) == 0);
1454 atomic_inc(&imp->imp_replay_inflight);
1456 req = ptlrpc_request_alloc_pack(imp, &RQF_OBD_PING, LUSTRE_OBD_VERSION,
1459 atomic_dec(&imp->imp_replay_inflight);
1463 ptlrpc_request_set_replen(req);
1464 req->rq_send_state = LUSTRE_IMP_REPLAY_WAIT;
1465 lustre_msg_add_flags(req->rq_reqmsg,
1466 MSG_LOCK_REPLAY_DONE | MSG_REQ_REPLAY_DONE);
1468 req->rq_timeout *= 3;
1469 req->rq_interpret_reply = completed_replay_interpret;
1471 ptlrpcd_add_req(req);
1476 * In kernel code all import invalidation happens in its own
1477 * separate thread, so that whatever application happened to encounter
1478 * a problem could still be killed or otherwise continue
1480 static int ptlrpc_invalidate_import_thread(void *data)
1482 struct obd_import *imp = data;
1486 unshare_fs_struct();
1488 CDEBUG(D_HA, "thread invalidate import %s to %s@%s\n",
1489 imp->imp_obd->obd_name, obd2cli_tgt(imp->imp_obd),
1490 imp->imp_connection->c_remote_uuid.uuid);
1492 ptlrpc_invalidate_import(imp);
1494 if (obd_dump_on_eviction) {
1495 CERROR("dump the log upon eviction\n");
1496 libcfs_debug_dumplog();
1499 import_set_state(imp, LUSTRE_IMP_RECOVER);
1500 ptlrpc_import_recovery_state_machine(imp);
1502 class_import_put(imp);
1507 * This is the state machine for client-side recovery on import.
1509 * Typicaly we have two possibly paths. If we came to server and it is not
1510 * in recovery, we just enter IMP_EVICTED state, invalidate our import
1511 * state and reconnect from scratch.
1512 * If we came to server that is in recovery, we enter IMP_REPLAY import state.
1513 * We go through our list of requests to replay and send them to server one by
1515 * After sending all request from the list we change import state to
1516 * IMP_REPLAY_LOCKS and re-request all the locks we believe we have from server
1517 * and also all the locks we don't yet have and wait for server to grant us.
1518 * After that we send a special "replay completed" request and change import
1519 * state to IMP_REPLAY_WAIT.
1520 * Upon receiving reply to that "replay completed" RPC we enter IMP_RECOVER
1521 * state and resend all requests from sending list.
1522 * After that we promote import to FULL state and send all delayed requests
1523 * and import is fully operational after that.
1526 int ptlrpc_import_recovery_state_machine(struct obd_import *imp)
1534 if (imp->imp_state == LUSTRE_IMP_EVICTED) {
1535 deuuidify(obd2cli_tgt(imp->imp_obd), NULL,
1536 &target_start, &target_len);
1537 /* Don't care about MGC eviction */
1538 if (strcmp(imp->imp_obd->obd_type->typ_name,
1539 LUSTRE_MGC_NAME) != 0) {
1540 LCONSOLE_ERROR_MSG(0x167, "%s: This client was evicted "
1541 "by %.*s; in progress operations "
1542 "using this service will fail.\n",
1543 imp->imp_obd->obd_name, target_len,
1545 LASSERTF(!obd_lbug_on_eviction, "LBUG upon eviction");
1547 CDEBUG(D_HA, "evicted from %s@%s; invalidating\n",
1548 obd2cli_tgt(imp->imp_obd),
1549 imp->imp_connection->c_remote_uuid.uuid);
1550 /* reset vbr_failed flag upon eviction */
1551 spin_lock(&imp->imp_lock);
1552 imp->imp_vbr_failed = 0;
1553 spin_unlock(&imp->imp_lock);
1556 struct task_struct *task;
1557 /* bug 17802: XXX client_disconnect_export vs connect request
1558 * race. if client is evicted at this time then we start
1559 * invalidate thread without reference to import and import can
1560 * be freed at same time. */
1561 class_import_get(imp);
1562 task = kthread_run(ptlrpc_invalidate_import_thread, imp,
1565 class_import_put(imp);
1566 CERROR("error starting invalidate thread: %d\n", rc);
1575 if (imp->imp_state == LUSTRE_IMP_REPLAY) {
1576 CDEBUG(D_HA, "replay requested by %s\n",
1577 obd2cli_tgt(imp->imp_obd));
1578 rc = ptlrpc_replay_next(imp, &inflight);
1579 if (inflight == 0 &&
1580 atomic_read(&imp->imp_replay_inflight) == 0) {
1581 import_set_state(imp, LUSTRE_IMP_REPLAY_LOCKS);
1582 rc = ldlm_replay_locks(imp);
1589 if (imp->imp_state == LUSTRE_IMP_REPLAY_LOCKS) {
1590 if (atomic_read(&imp->imp_replay_inflight) == 0) {
1591 import_set_state(imp, LUSTRE_IMP_REPLAY_WAIT);
1592 rc = signal_completed_replay(imp);
1598 if (imp->imp_state == LUSTRE_IMP_REPLAY_WAIT) {
1599 if (atomic_read(&imp->imp_replay_inflight) == 0) {
1600 import_set_state(imp, LUSTRE_IMP_RECOVER);
1604 if (imp->imp_state == LUSTRE_IMP_RECOVER) {
1605 struct ptlrpc_connection *conn = imp->imp_connection;
1607 rc = ptlrpc_resend(imp);
1610 ptlrpc_activate_import(imp, true);
1612 CDEBUG_LIMIT(imp->imp_was_idle ?
1613 imp->imp_idle_debug : D_CONSOLE,
1614 "%s: Connection restored to %s (at %s)\n",
1615 imp->imp_obd->obd_name,
1616 obd_uuid2str(&conn->c_remote_uuid),
1617 obd_import_nid2str(imp));
1618 spin_lock(&imp->imp_lock);
1619 imp->imp_was_idle = 0;
1620 spin_unlock(&imp->imp_lock);
1623 if (imp->imp_state == LUSTRE_IMP_FULL) {
1624 wake_up_all(&imp->imp_recovery_waitq);
1625 ptlrpc_wake_delayed(imp);
1632 static struct ptlrpc_request *ptlrpc_disconnect_prep_req(struct obd_import *imp)
1634 struct ptlrpc_request *req;
1638 switch (imp->imp_connect_op) {
1640 rq_opc = OST_DISCONNECT;
1643 rq_opc = MDS_DISCONNECT;
1646 rq_opc = MGS_DISCONNECT;
1650 CERROR("%s: don't know how to disconnect from %s "
1651 "(connect_op %d): rc = %d\n",
1652 imp->imp_obd->obd_name, obd2cli_tgt(imp->imp_obd),
1653 imp->imp_connect_op, rc);
1654 RETURN(ERR_PTR(rc));
1657 req = ptlrpc_request_alloc_pack(imp, &RQF_MDS_DISCONNECT,
1658 LUSTRE_OBD_VERSION, rq_opc);
1660 RETURN(ERR_PTR(-ENOMEM));
1662 /* We are disconnecting, do not retry a failed DISCONNECT rpc if
1663 * it fails. We can get through the above with a down server
1664 * if the client doesn't know the server is gone yet. */
1665 req->rq_no_resend = 1;
1667 /* We want client umounts to happen quickly, no matter the
1669 req->rq_timeout = min_t(int, req->rq_timeout,
1670 INITIAL_CONNECT_TIMEOUT);
1672 import_set_state(imp, LUSTRE_IMP_CONNECTING);
1673 req->rq_send_state = LUSTRE_IMP_CONNECTING;
1674 ptlrpc_request_set_replen(req);
1679 int ptlrpc_disconnect_import(struct obd_import *imp, int noclose)
1681 struct ptlrpc_request *req;
1685 if (imp->imp_obd->obd_force)
1686 GOTO(set_state, rc);
1688 /* probably the import has been disconnected already being idle */
1689 spin_lock(&imp->imp_lock);
1690 if (imp->imp_state == LUSTRE_IMP_IDLE)
1692 spin_unlock(&imp->imp_lock);
1694 if (ptlrpc_import_in_recovery(imp)) {
1695 long timeout_jiffies;
1699 if (imp->imp_server_timeout)
1700 timeout = obd_timeout >> 1;
1702 timeout = obd_timeout;
1707 req_portal = imp->imp_client->cli_request_portal;
1708 idx = import_at_get_index(imp, req_portal);
1709 timeout = at_get(&imp->imp_at.iat_service_estimate[idx]);
1712 timeout_jiffies = cfs_time_seconds(timeout);
1713 if (wait_event_idle_timeout(imp->imp_recovery_waitq,
1714 !ptlrpc_import_in_recovery(imp),
1715 timeout_jiffies) == 0 &&
1716 l_wait_event_abortable(imp->imp_recovery_waitq,
1717 !ptlrpc_import_in_recovery(imp)) < 0)
1721 spin_lock(&imp->imp_lock);
1722 if (imp->imp_state != LUSTRE_IMP_FULL)
1724 spin_unlock(&imp->imp_lock);
1726 req = ptlrpc_disconnect_prep_req(imp);
1728 GOTO(set_state, rc = PTR_ERR(req));
1729 rc = ptlrpc_queue_wait(req);
1730 ptlrpc_req_finished(req);
1733 spin_lock(&imp->imp_lock);
1736 import_set_state_nolock(imp, LUSTRE_IMP_DISCON);
1738 import_set_state_nolock(imp, LUSTRE_IMP_CLOSED);
1739 memset(&imp->imp_remote_handle, 0, sizeof(imp->imp_remote_handle));
1740 spin_unlock(&imp->imp_lock);
1742 if (rc == -ETIMEDOUT || rc == -ENOTCONN || rc == -ESHUTDOWN)
1746 EXPORT_SYMBOL(ptlrpc_disconnect_import);
1748 static void ptlrpc_reset_reqs_generation(struct obd_import *imp)
1750 struct ptlrpc_request *old, *tmp;
1752 /* tag all resendable requests generated before disconnection
1753 * notice this code is part of disconnect-at-idle path only */
1754 list_for_each_entry_safe(old, tmp, &imp->imp_delayed_list,
1756 spin_lock(&old->rq_lock);
1757 if (old->rq_import_generation == imp->imp_generation - 1 &&
1758 ((imp->imp_initiated_at == imp->imp_generation) ||
1759 !old->rq_no_resend))
1760 old->rq_import_generation = imp->imp_generation;
1761 spin_unlock(&old->rq_lock);
1765 static int ptlrpc_disconnect_idle_interpret(const struct lu_env *env,
1766 struct ptlrpc_request *req,
1769 struct obd_import *imp = req->rq_import;
1772 DEBUG_REQ(D_HA, req, "inflight=%d, refcount=%d: rc = %d",
1773 atomic_read(&imp->imp_inflight),
1774 atomic_read(&imp->imp_refcount), rc);
1776 spin_lock(&imp->imp_lock);
1777 /* DISCONNECT reply can be late and another connection can just
1778 * be initiated. so we have to abort disconnection. */
1779 if (req->rq_import_generation == imp->imp_generation &&
1780 imp->imp_state != LUSTRE_IMP_CLOSED) {
1781 LASSERTF(imp->imp_state == LUSTRE_IMP_CONNECTING,
1782 "%s\n", ptlrpc_import_state_name(imp->imp_state));
1783 imp->imp_state = LUSTRE_IMP_IDLE;
1784 memset(&imp->imp_remote_handle, 0,
1785 sizeof(imp->imp_remote_handle));
1786 /* take our DISCONNECT into account */
1787 if (atomic_read(&imp->imp_inflight) > 1) {
1788 imp->imp_generation++;
1789 imp->imp_initiated_at = imp->imp_generation;
1790 import_set_state_nolock(imp, LUSTRE_IMP_NEW);
1791 ptlrpc_reset_reqs_generation(imp);
1797 rc = ptlrpc_connect_import_locked(imp);
1799 ptlrpc_pinger_add_import(imp);
1801 spin_unlock(&imp->imp_lock);
1807 int ptlrpc_disconnect_and_idle_import(struct obd_import *imp)
1809 struct ptlrpc_request *req;
1812 if (imp->imp_obd->obd_force)
1815 if (ptlrpc_import_in_recovery(imp))
1818 spin_lock(&imp->imp_lock);
1819 if (imp->imp_state != LUSTRE_IMP_FULL) {
1820 spin_unlock(&imp->imp_lock);
1823 spin_unlock(&imp->imp_lock);
1825 req = ptlrpc_disconnect_prep_req(imp);
1827 RETURN(PTR_ERR(req));
1829 CDEBUG_LIMIT(imp->imp_idle_debug, "%s: disconnect after %llus idle\n",
1830 imp->imp_obd->obd_name,
1831 ktime_get_real_seconds() - imp->imp_last_reply_time);
1833 /* don't make noise at reconnection */
1834 spin_lock(&imp->imp_lock);
1835 imp->imp_was_idle = 1;
1836 spin_unlock(&imp->imp_lock);
1838 req->rq_interpret_reply = ptlrpc_disconnect_idle_interpret;
1839 ptlrpcd_add_req(req);
1843 EXPORT_SYMBOL(ptlrpc_disconnect_and_idle_import);
1845 void ptlrpc_cleanup_imp(struct obd_import *imp)
1849 spin_lock(&imp->imp_lock);
1851 import_set_state_nolock(imp, LUSTRE_IMP_CLOSED);
1852 imp->imp_generation++;
1853 ptlrpc_abort_inflight(imp);
1855 spin_unlock(&imp->imp_lock);
1860 /* Adaptive Timeout utils */
1862 /* Update at_current with the specified value (bounded by at_min and at_max),
1863 * as well as the AT history "bins".
1864 * - Bin into timeslices using AT_BINS bins.
1865 * - This gives us a max of the last at_history seconds without the storage,
1866 * but still smoothing out a return to normalcy from a slow response.
1867 * - (E.g. remember the maximum latency in each minute of the last 4 minutes.)
1869 int at_measured(struct adaptive_timeout *at, unsigned int val)
1871 unsigned int old = at->at_current;
1872 time64_t now = ktime_get_real_seconds();
1873 long binlimit = max_t(long, at_history / AT_BINS, 1);
1876 CDEBUG(D_OTHER, "add %u to %p time=%lu v=%u (%u %u %u %u)\n",
1877 val, at, (long)(now - at->at_binstart), at->at_current,
1878 at->at_hist[0], at->at_hist[1], at->at_hist[2], at->at_hist[3]);
1881 /* 0's don't count, because we never want our timeout to
1882 drop to 0, and because 0 could mean an error */
1885 spin_lock(&at->at_lock);
1887 if (unlikely(at->at_binstart == 0)) {
1888 /* Special case to remove default from history */
1889 at->at_current = val;
1890 at->at_worst_ever = val;
1891 at->at_worst_time = now;
1892 at->at_hist[0] = val;
1893 at->at_binstart = now;
1894 } else if (now - at->at_binstart < binlimit ) {
1896 at->at_hist[0] = max(val, at->at_hist[0]);
1897 at->at_current = max(val, at->at_current);
1900 unsigned int maxv = val;
1902 /* move bins over */
1903 shift = (u32)(now - at->at_binstart) / binlimit;
1905 for(i = AT_BINS - 1; i >= 0; i--) {
1907 at->at_hist[i] = at->at_hist[i - shift];
1908 maxv = max(maxv, at->at_hist[i]);
1913 at->at_hist[0] = val;
1914 at->at_current = maxv;
1915 at->at_binstart += shift * binlimit;
1918 if (at->at_current > at->at_worst_ever) {
1919 at->at_worst_ever = at->at_current;
1920 at->at_worst_time = now;
1923 if (at->at_flags & AT_FLG_NOHIST)
1924 /* Only keep last reported val; keeping the rest of the history
1926 at->at_current = val;
1929 at->at_current = min(at->at_current, at_max);
1930 at->at_current = max(at->at_current, at_min);
1932 if (at->at_current != old)
1933 CDEBUG(D_OTHER, "AT %p change: old=%u new=%u delta=%d "
1934 "(val=%u) hist %u %u %u %u\n", at,
1935 old, at->at_current, at->at_current - old, val,
1936 at->at_hist[0], at->at_hist[1], at->at_hist[2],
1939 /* if we changed, report the old value */
1940 old = (at->at_current != old) ? old : 0;
1942 spin_unlock(&at->at_lock);
1946 /* Find the imp_at index for a given portal; assign if space available */
1947 int import_at_get_index(struct obd_import *imp, int portal)
1949 struct imp_at *at = &imp->imp_at;
1952 for (i = 0; i < IMP_AT_MAX_PORTALS; i++) {
1953 if (at->iat_portal[i] == portal)
1955 if (at->iat_portal[i] == 0)
1960 /* Not found in list, add it under a lock */
1961 spin_lock(&imp->imp_lock);
1963 /* Check unused under lock */
1964 for (; i < IMP_AT_MAX_PORTALS; i++) {
1965 if (at->iat_portal[i] == portal)
1967 if (at->iat_portal[i] == 0)
1972 /* Not enough portals? */
1973 LASSERT(i < IMP_AT_MAX_PORTALS);
1975 at->iat_portal[i] = portal;
1977 spin_unlock(&imp->imp_lock);