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. */
325 long timeout_jiffies;
327 /* Calculate max timeout for waiting on rpcs to error
328 * out. Use obd_timeout if calculated value is smaller
331 if (!OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_REPL_UNLINK)) {
332 timeout = ptlrpc_inflight_timeout(imp);
333 timeout += div_u64(timeout, 3);
336 timeout = obd_timeout;
338 /* decrease the interval to increase race condition */
342 CDEBUG(D_RPCTRACE, "Sleeping %llds for inflight to error out\n",
345 /* Wait for all requests to error out and call completion
346 * callbacks. Cap it at obd_timeout -- these should all
347 * have been locally cancelled by ptlrpc_abort_inflight.
349 timeout_jiffies = max_t(long, cfs_time_seconds(timeout), 1);
350 lwi = LWI_TIMEOUT_INTERVAL(timeout_jiffies,
351 (timeout > 1) ? cfs_time_seconds(1) :
352 cfs_time_seconds(1) / 2,
354 rc = l_wait_event(imp->imp_recovery_waitq,
355 (atomic_read(&imp->imp_inflight) == 0),
358 const char *cli_tgt = obd2cli_tgt(imp->imp_obd);
360 CERROR("%s: rc = %d waiting for callback (%d != 0)\n",
361 cli_tgt, rc, atomic_read(&imp->imp_inflight));
363 spin_lock(&imp->imp_lock);
364 if (atomic_read(&imp->imp_inflight) == 0) {
365 int count = atomic_read(&imp->imp_unregistering);
367 /* We know that "unregistering" rpcs only can
368 * survive in sending or delaying lists (they
369 * maybe waiting for long reply unlink in
370 * sluggish nets). Let's check this. If there
371 * is no inflight and unregistering != 0, this
373 LASSERTF(count == 0, "Some RPCs are still "
374 "unregistering: %d\n", count);
376 /* Let's save one loop as soon as inflight have
377 * dropped to zero. No new inflights possible at
381 list_for_each_safe(tmp, n,
382 &imp->imp_sending_list) {
383 req = list_entry(tmp,
384 struct ptlrpc_request,
386 DEBUG_REQ(D_ERROR, req,
387 "still on sending list");
389 list_for_each_safe(tmp, n,
390 &imp->imp_delayed_list) {
391 req = list_entry(tmp,
392 struct ptlrpc_request,
394 DEBUG_REQ(D_ERROR, req,
395 "still on delayed list");
398 CERROR("%s: Unregistering RPCs found (%d). "
399 "Network is sluggish? Waiting them "
400 "to error out.\n", cli_tgt,
401 atomic_read(&imp->imp_unregistering));
403 spin_unlock(&imp->imp_lock);
408 * Let's additionally check that no new rpcs added to import in
409 * "invalidate" state.
411 LASSERT(atomic_read(&imp->imp_inflight) == 0);
412 obd_import_event(imp->imp_obd, imp, IMP_EVENT_INVALIDATE);
413 sptlrpc_import_flush_all_ctx(imp);
415 atomic_dec(&imp->imp_inval_count);
416 wake_up_all(&imp->imp_recovery_waitq);
418 EXPORT_SYMBOL(ptlrpc_invalidate_import);
420 /* unset imp_invalid */
421 void ptlrpc_activate_import(struct obd_import *imp, bool set_state_full)
423 struct obd_device *obd = imp->imp_obd;
425 spin_lock(&imp->imp_lock);
426 if (imp->imp_deactive != 0) {
427 LASSERT(imp->imp_state != LUSTRE_IMP_FULL);
428 if (imp->imp_state != LUSTRE_IMP_DISCON)
429 import_set_state_nolock(imp, LUSTRE_IMP_DISCON);
430 spin_unlock(&imp->imp_lock);
434 import_set_state_nolock(imp, LUSTRE_IMP_FULL);
436 imp->imp_invalid = 0;
438 spin_unlock(&imp->imp_lock);
439 obd_import_event(obd, imp, IMP_EVENT_ACTIVE);
441 EXPORT_SYMBOL(ptlrpc_activate_import);
443 void ptlrpc_pinger_force(struct obd_import *imp)
445 CDEBUG(D_HA, "%s: waking up pinger s:%s\n", obd2cli_tgt(imp->imp_obd),
446 ptlrpc_import_state_name(imp->imp_state));
448 spin_lock(&imp->imp_lock);
449 imp->imp_force_verify = 1;
450 spin_unlock(&imp->imp_lock);
452 if (imp->imp_state != LUSTRE_IMP_CONNECTING)
453 ptlrpc_pinger_wake_up();
455 EXPORT_SYMBOL(ptlrpc_pinger_force);
457 void ptlrpc_fail_import(struct obd_import *imp, __u32 conn_cnt)
461 LASSERT(!imp->imp_dlm_fake);
463 if (ptlrpc_set_import_discon(imp, conn_cnt, true))
464 ptlrpc_pinger_force(imp);
469 int ptlrpc_reconnect_import(struct obd_import *imp)
472 long timeout_jiffies = cfs_time_seconds(obd_timeout);
473 struct l_wait_info lwi;
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 lwi = LWI_TIMEOUT(timeout_jiffies, NULL, NULL);
482 rc = l_wait_event(imp->imp_recovery_waitq,
483 !ptlrpc_import_in_recovery(imp), &lwi);
484 CDEBUG(D_HA, "%s: recovery finished s:%s\n", obd2cli_tgt(imp->imp_obd),
485 ptlrpc_import_state_name(imp->imp_state));
488 ptlrpc_set_import_discon(imp, 0, false);
489 /* Force a new connect attempt */
490 ptlrpc_invalidate_import(imp);
491 /* Do a fresh connect next time by zeroing the handle */
492 ptlrpc_disconnect_import(imp, 1);
493 /* Wait for all invalidate calls to finish */
494 if (atomic_read(&imp->imp_inval_count) > 0) {
495 struct l_wait_info lwi = LWI_INTR(LWI_ON_SIGNAL_NOOP, NULL);
498 rc = l_wait_event(imp->imp_recovery_waitq,
499 (atomic_read(&imp->imp_inval_count) == 0),
502 CERROR("Interrupted, inval=%d\n",
503 atomic_read(&imp->imp_inval_count));
506 /* Allow reconnect attempts */
507 imp->imp_obd->obd_no_recov = 0;
508 /* Remove 'invalid' flag */
509 ptlrpc_activate_import(imp, false);
510 /* Attempt a new connect */
511 ptlrpc_recover_import(imp, NULL, 0);
515 EXPORT_SYMBOL(ptlrpc_reconnect_import);
518 * Connection on import \a imp is changed to another one (if more than one is
519 * present). We typically chose connection that we have not tried to connect to
522 static int import_select_connection(struct obd_import *imp)
524 struct obd_import_conn *imp_conn = NULL, *conn;
525 struct obd_export *dlmexp;
527 int target_len, tried_all = 1;
530 spin_lock(&imp->imp_lock);
532 if (list_empty(&imp->imp_conn_list)) {
533 CERROR("%s: no connections available\n",
534 imp->imp_obd->obd_name);
535 spin_unlock(&imp->imp_lock);
539 list_for_each_entry(conn, &imp->imp_conn_list, oic_item) {
540 CDEBUG(D_HA, "%s: connect to NID %s last attempt %lld\n",
541 imp->imp_obd->obd_name,
542 libcfs_nid2str(conn->oic_conn->c_peer.nid),
543 conn->oic_last_attempt);
545 /* If we have not tried this connection since
546 the last successful attempt, go with this one */
547 if ((conn->oic_last_attempt == 0) ||
548 conn->oic_last_attempt <= imp->imp_last_success_conn) {
554 /* If all of the connections have already been tried
555 since the last successful connection; just choose the
556 least recently used */
559 else if (imp_conn->oic_last_attempt > conn->oic_last_attempt)
563 /* if not found, simply choose the current one */
564 if (!imp_conn || imp->imp_force_reconnect) {
565 LASSERT(imp->imp_conn_current);
566 imp_conn = imp->imp_conn_current;
569 LASSERT(imp_conn->oic_conn);
571 /* If we've tried everything, and we're back to the beginning of the
572 list, increase our timeout and try again. It will be reset when
573 we do finally connect. (FIXME: really we should wait for all network
574 state associated with the last connection attempt to drain before
575 trying to reconnect on it.) */
576 if (tried_all && (imp->imp_conn_list.next == &imp_conn->oic_item)) {
577 struct adaptive_timeout *at = &imp->imp_at.iat_net_latency;
578 if (at_get(at) < CONNECTION_SWITCH_MAX) {
579 at_measured(at, at_get(at) + CONNECTION_SWITCH_INC);
580 if (at_get(at) > CONNECTION_SWITCH_MAX)
581 at_reset(at, CONNECTION_SWITCH_MAX);
583 LASSERT(imp_conn->oic_last_attempt);
584 CDEBUG(D_HA, "%s: tried all connections, increasing latency "
585 "to %ds\n", imp->imp_obd->obd_name, at_get(at));
588 imp_conn->oic_last_attempt = ktime_get_seconds();
590 /* switch connection, don't mind if it's same as the current one */
591 if (imp->imp_connection)
592 ptlrpc_connection_put(imp->imp_connection);
593 imp->imp_connection = ptlrpc_connection_addref(imp_conn->oic_conn);
595 dlmexp = class_conn2export(&imp->imp_dlm_handle);
596 LASSERT(dlmexp != NULL);
597 if (dlmexp->exp_connection)
598 ptlrpc_connection_put(dlmexp->exp_connection);
599 dlmexp->exp_connection = ptlrpc_connection_addref(imp_conn->oic_conn);
600 class_export_put(dlmexp);
602 if (imp->imp_conn_current != imp_conn) {
603 if (imp->imp_conn_current) {
604 deuuidify(obd2cli_tgt(imp->imp_obd), NULL,
605 &target_start, &target_len);
607 CDEBUG(D_HA, "%s: Connection changing to"
609 imp->imp_obd->obd_name,
610 target_len, target_start,
611 libcfs_nid2str(imp_conn->oic_conn->c_peer.nid));
614 imp->imp_conn_current = imp_conn;
617 /* The below message is checked in conf-sanity.sh test_35[ab] */
618 CDEBUG(D_HA, "%s: import %p using connection %s/%s\n",
619 imp->imp_obd->obd_name, imp, imp_conn->oic_uuid.uuid,
620 libcfs_nid2str(imp_conn->oic_conn->c_peer.nid));
622 spin_unlock(&imp->imp_lock);
628 * must be called under imp_lock
630 static int ptlrpc_first_transno(struct obd_import *imp, __u64 *transno)
632 struct ptlrpc_request *req;
633 struct list_head *tmp;
635 /* The requests in committed_list always have smaller transnos than
636 * the requests in replay_list */
637 if (!list_empty(&imp->imp_committed_list)) {
638 tmp = imp->imp_committed_list.next;
639 req = list_entry(tmp, struct ptlrpc_request, rq_replay_list);
640 *transno = req->rq_transno;
641 if (req->rq_transno == 0) {
642 DEBUG_REQ(D_ERROR, req,
643 "zero transno in committed_list");
648 if (!list_empty(&imp->imp_replay_list)) {
649 tmp = imp->imp_replay_list.next;
650 req = list_entry(tmp, struct ptlrpc_request, rq_replay_list);
651 *transno = req->rq_transno;
652 if (req->rq_transno == 0) {
653 DEBUG_REQ(D_ERROR, req, "zero transno in replay_list");
661 int ptlrpc_connect_import(struct obd_import *imp)
663 spin_lock(&imp->imp_lock);
664 return ptlrpc_connect_import_locked(imp);
668 * Attempt to (re)connect import \a imp. This includes all preparations,
669 * initializing CONNECT RPC request and passing it to ptlrpcd for
672 * Assumes imp->imp_lock is held, and releases it.
674 * Returns 0 on success or error code.
676 int ptlrpc_connect_import_locked(struct obd_import *imp)
678 struct obd_device *obd = imp->imp_obd;
679 int initial_connect = 0;
681 __u64 committed_before_reconnect = 0;
682 struct ptlrpc_request *request;
683 struct obd_connect_data ocd;
684 char *bufs[] = { NULL,
685 obd2cli_tgt(imp->imp_obd),
687 (char *)&imp->imp_dlm_handle,
690 struct ptlrpc_connect_async_args *aa;
694 assert_spin_locked(&imp->imp_lock);
696 if (imp->imp_state == LUSTRE_IMP_CLOSED) {
697 spin_unlock(&imp->imp_lock);
698 CERROR("can't connect to a closed import\n");
700 } else if (imp->imp_state == LUSTRE_IMP_FULL) {
701 spin_unlock(&imp->imp_lock);
702 CERROR("already connected\n");
704 } else if (imp->imp_state == LUSTRE_IMP_CONNECTING ||
705 imp->imp_state == LUSTRE_IMP_EVICTED ||
706 imp->imp_connected) {
707 spin_unlock(&imp->imp_lock);
708 CERROR("already connecting\n");
712 import_set_state_nolock(imp, LUSTRE_IMP_CONNECTING);
715 imp->imp_resend_replay = 0;
717 if (!lustre_handle_is_used(&imp->imp_remote_handle))
720 committed_before_reconnect = imp->imp_peer_committed_transno;
722 set_transno = ptlrpc_first_transno(imp,
723 &imp->imp_connect_data.ocd_transno);
724 spin_unlock(&imp->imp_lock);
726 rc = import_select_connection(imp);
730 rc = sptlrpc_import_sec_adapt(imp, NULL, NULL);
734 /* Reset connect flags to the originally requested flags, in case
735 * the server is updated on-the-fly we will get the new features. */
736 ocd = imp->imp_connect_data;
737 ocd.ocd_connect_flags = imp->imp_connect_flags_orig;
738 ocd.ocd_connect_flags2 = imp->imp_connect_flags2_orig;
739 /* Reset ocd_version each time so the server knows the exact versions */
740 ocd.ocd_version = LUSTRE_VERSION_CODE;
741 imp->imp_msghdr_flags &= ~MSGHDR_AT_SUPPORT;
742 imp->imp_msghdr_flags &= ~MSGHDR_CKSUM_INCOMPAT18;
744 rc = obd_reconnect(NULL, imp->imp_obd->obd_self_export, obd,
745 &obd->obd_uuid, &ocd, NULL);
749 request = ptlrpc_request_alloc(imp, &RQF_MDS_CONNECT);
751 GOTO(out, rc = -ENOMEM);
753 /* get SELinux policy info if any */
754 rc = sptlrpc_get_sepol(request);
756 ptlrpc_request_free(request);
760 bufs[5] = request->rq_sepol;
762 req_capsule_set_size(&request->rq_pill, &RMF_SELINUX_POL, RCL_CLIENT,
763 strlen(request->rq_sepol) ?
764 strlen(request->rq_sepol) + 1 : 0);
766 rc = ptlrpc_request_bufs_pack(request, LUSTRE_OBD_VERSION,
767 imp->imp_connect_op, bufs, NULL);
769 ptlrpc_request_free(request);
773 /* Report the rpc service time to the server so that it knows how long
774 * to wait for clients to join recovery */
775 lustre_msg_set_service_time(request->rq_reqmsg,
776 at_timeout2est(request->rq_timeout));
778 /* The amount of time we give the server to process the connect req.
779 * import_select_connection will increase the net latency on
780 * repeated reconnect attempts to cover slow networks.
781 * We override/ignore the server rpc completion estimate here,
782 * which may be large if this is a reconnect attempt */
783 request->rq_timeout = INITIAL_CONNECT_TIMEOUT;
784 lustre_msg_set_timeout(request->rq_reqmsg, request->rq_timeout);
786 request->rq_no_resend = request->rq_no_delay = 1;
787 request->rq_send_state = LUSTRE_IMP_CONNECTING;
788 /* Allow a slightly larger reply for future growth compatibility */
789 req_capsule_set_size(&request->rq_pill, &RMF_CONNECT_DATA, RCL_SERVER,
790 sizeof(struct obd_connect_data)+16*sizeof(__u64));
791 ptlrpc_request_set_replen(request);
792 request->rq_interpret_reply = ptlrpc_connect_interpret;
794 aa = ptlrpc_req_async_args(aa, request);
795 memset(aa, 0, sizeof *aa);
797 aa->pcaa_peer_committed = committed_before_reconnect;
798 aa->pcaa_initial_connect = initial_connect;
800 if (aa->pcaa_initial_connect) {
801 spin_lock(&imp->imp_lock);
802 imp->imp_replayable = 1;
803 spin_unlock(&imp->imp_lock);
804 lustre_msg_add_op_flags(request->rq_reqmsg,
805 MSG_CONNECT_INITIAL);
809 lustre_msg_add_op_flags(request->rq_reqmsg,
810 MSG_CONNECT_TRANSNO);
812 DEBUG_REQ(D_RPCTRACE, request, "(re)connect request (timeout %ld)",
813 request->rq_timeout);
814 ptlrpcd_add_req(request);
818 import_set_state(imp, LUSTRE_IMP_DISCON);
822 EXPORT_SYMBOL(ptlrpc_connect_import);
824 static void ptlrpc_maybe_ping_import_soon(struct obd_import *imp)
828 spin_lock(&imp->imp_lock);
829 force_verify = imp->imp_force_verify != 0;
830 spin_unlock(&imp->imp_lock);
833 ptlrpc_pinger_wake_up();
836 static int ptlrpc_busy_reconnect(int rc)
838 return (rc == -EBUSY) || (rc == -EAGAIN);
841 static int ptlrpc_connect_set_flags(struct obd_import *imp,
842 struct obd_connect_data *ocd,
843 __u64 old_connect_flags,
844 struct obd_export *exp, int init_connect)
847 struct client_obd *cli = &imp->imp_obd->u.cli;
849 spin_lock(&imp->imp_lock);
850 list_move(&imp->imp_conn_current->oic_item,
851 &imp->imp_conn_list);
852 imp->imp_last_success_conn =
853 imp->imp_conn_current->oic_last_attempt;
855 spin_unlock(&imp->imp_lock);
857 if (!warned && (ocd->ocd_connect_flags & OBD_CONNECT_VERSION) &&
858 (ocd->ocd_version > LUSTRE_VERSION_CODE +
859 LUSTRE_VERSION_OFFSET_WARN ||
860 ocd->ocd_version < LUSTRE_VERSION_CODE -
861 LUSTRE_VERSION_OFFSET_WARN)) {
862 /* Sigh, some compilers do not like #ifdef in the middle
863 of macro arguments */
864 const char *older = "older than client. "
865 "Consider upgrading server";
866 const char *newer = "newer than client. "
867 "Consider upgrading client";
869 LCONSOLE_WARN("Server %s version (%d.%d.%d.%d) "
871 obd2cli_tgt(imp->imp_obd),
872 OBD_OCD_VERSION_MAJOR(ocd->ocd_version),
873 OBD_OCD_VERSION_MINOR(ocd->ocd_version),
874 OBD_OCD_VERSION_PATCH(ocd->ocd_version),
875 OBD_OCD_VERSION_FIX(ocd->ocd_version),
876 ocd->ocd_version > LUSTRE_VERSION_CODE ?
877 newer : older, LUSTRE_VERSION_STRING);
881 if (ocd->ocd_connect_flags & OBD_CONNECT_CKSUM) {
882 /* We sent to the server ocd_cksum_types with bits set
883 * for algorithms we understand. The server masked off
884 * the checksum types it doesn't support */
885 if ((ocd->ocd_cksum_types &
886 obd_cksum_types_supported_client()) == 0) {
887 LCONSOLE_ERROR("The negotiation of the checksum "
888 "alogrithm to use with server %s "
890 obd2cli_tgt(imp->imp_obd),
891 ocd->ocd_cksum_types,
892 obd_cksum_types_supported_client());
895 cli->cl_supp_cksum_types = ocd->ocd_cksum_types;
898 /* The server does not support OBD_CONNECT_CKSUM.
899 * Enforce ADLER for backward compatibility*/
900 cli->cl_supp_cksum_types = OBD_CKSUM_ADLER;
902 cli->cl_cksum_type = obd_cksum_type_select(imp->imp_obd->obd_name,
903 cli->cl_supp_cksum_types,
904 cli->cl_preferred_cksum_type);
906 if (ocd->ocd_connect_flags & OBD_CONNECT_BRW_SIZE)
907 cli->cl_max_pages_per_rpc =
908 min(ocd->ocd_brw_size >> PAGE_SHIFT,
909 cli->cl_max_pages_per_rpc);
910 else if (imp->imp_connect_op == MDS_CONNECT ||
911 imp->imp_connect_op == MGS_CONNECT)
912 cli->cl_max_pages_per_rpc = 1;
914 LASSERT((cli->cl_max_pages_per_rpc <= PTLRPC_MAX_BRW_PAGES) &&
915 (cli->cl_max_pages_per_rpc > 0));
917 client_adjust_max_dirty(cli);
919 /* Update client max modify RPCs in flight with value returned
921 if (ocd->ocd_connect_flags & OBD_CONNECT_MULTIMODRPCS)
922 cli->cl_max_mod_rpcs_in_flight = min(
923 cli->cl_max_mod_rpcs_in_flight,
924 ocd->ocd_maxmodrpcs);
926 cli->cl_max_mod_rpcs_in_flight = 1;
928 /* Reset ns_connect_flags only for initial connect. It might be
929 * changed in while using FS and if we reset it in reconnect
930 * this leads to losing user settings done before such as
931 * disable lru_resize, etc. */
932 if (old_connect_flags != exp_connect_flags(exp) || init_connect) {
933 struct ldlm_namespace *ns = imp->imp_obd->obd_namespace;
937 ns->ns_connect_flags ^ ns->ns_orig_connect_flags;
938 CDEBUG(D_HA, "%s: Resetting ns_connect_flags to server "
939 "flags: %#llx\n", imp->imp_obd->obd_name,
940 ocd->ocd_connect_flags);
941 ns->ns_connect_flags = (ns->ns_connect_flags & changed_flags) |
942 (ocd->ocd_connect_flags & ~changed_flags);
943 ns->ns_orig_connect_flags = ocd->ocd_connect_flags;
946 if (ocd->ocd_connect_flags & OBD_CONNECT_AT)
947 /* We need a per-message support flag, because
948 * a. we don't know if the incoming connect reply
949 * supports AT or not (in reply_in_callback)
950 * until we unpack it.
951 * b. failovered server means export and flags are gone
952 * (in ptlrpc_send_reply).
953 * Can only be set when we know AT is supported at
955 imp->imp_msghdr_flags |= MSGHDR_AT_SUPPORT;
957 imp->imp_msghdr_flags &= ~MSGHDR_AT_SUPPORT;
959 imp->imp_msghdr_flags |= MSGHDR_CKSUM_INCOMPAT18;
965 * Add all replay requests back to unreplied list before start replay,
966 * so that we can make sure the known replied XID is always increased
967 * only even if when replaying requests.
969 static void ptlrpc_prepare_replay(struct obd_import *imp)
971 struct ptlrpc_request *req;
973 if (imp->imp_state != LUSTRE_IMP_REPLAY ||
974 imp->imp_resend_replay)
977 /* If the server was restart during repaly, the requests may
978 * have been added to the unreplied list in former replay. */
979 spin_lock(&imp->imp_lock);
981 list_for_each_entry(req, &imp->imp_committed_list, rq_replay_list) {
982 if (list_empty(&req->rq_unreplied_list))
983 ptlrpc_add_unreplied(req);
986 list_for_each_entry(req, &imp->imp_replay_list, rq_replay_list) {
987 if (list_empty(&req->rq_unreplied_list))
988 ptlrpc_add_unreplied(req);
991 imp->imp_known_replied_xid = ptlrpc_known_replied_xid(imp);
992 spin_unlock(&imp->imp_lock);
996 * interpret_reply callback for connect RPCs.
997 * Looks into returned status of connect operation and decides
998 * what to do with the import - i.e enter recovery, promote it to
999 * full state for normal operations of disconnect it due to an error.
1001 static int ptlrpc_connect_interpret(const struct lu_env *env,
1002 struct ptlrpc_request *request,
1005 struct ptlrpc_connect_async_args *aa = data;
1006 struct obd_import *imp = request->rq_import;
1007 struct lustre_handle old_hdl;
1008 __u64 old_connect_flags;
1010 struct obd_connect_data *ocd;
1011 struct obd_export *exp = NULL;
1015 spin_lock(&imp->imp_lock);
1016 if (imp->imp_state == LUSTRE_IMP_CLOSED) {
1017 imp->imp_connect_tried = 1;
1018 spin_unlock(&imp->imp_lock);
1022 imp->imp_connect_error = rc;
1024 struct ptlrpc_request *free_req;
1025 struct ptlrpc_request *tmp;
1027 /* abort all delayed requests initiated connection */
1028 list_for_each_entry_safe(free_req, tmp, &imp->imp_delayed_list,
1030 spin_lock(&free_req->rq_lock);
1031 if (free_req->rq_no_resend) {
1032 free_req->rq_err = 1;
1033 free_req->rq_status = -EIO;
1034 ptlrpc_client_wake_req(free_req);
1036 spin_unlock(&free_req->rq_lock);
1039 /* if this reconnect to busy export - not need select new target
1041 imp->imp_force_reconnect = ptlrpc_busy_reconnect(rc);
1042 spin_unlock(&imp->imp_lock);
1043 ptlrpc_maybe_ping_import_soon(imp);
1047 /* LU-7558: indicate that we are interpretting connect reply,
1048 * pltrpc_connect_import() will not try to reconnect until
1049 * interpret will finish. */
1050 imp->imp_connected = 1;
1051 spin_unlock(&imp->imp_lock);
1053 LASSERT(imp->imp_conn_current);
1055 msg_flags = lustre_msg_get_op_flags(request->rq_repmsg);
1057 ret = req_capsule_get_size(&request->rq_pill, &RMF_CONNECT_DATA,
1059 /* server replied obd_connect_data is always bigger */
1060 ocd = req_capsule_server_sized_get(&request->rq_pill,
1061 &RMF_CONNECT_DATA, ret);
1064 CERROR("%s: no connect data from server\n",
1065 imp->imp_obd->obd_name);
1070 spin_lock(&imp->imp_lock);
1072 /* All imports are pingable */
1073 imp->imp_pingable = 1;
1074 imp->imp_force_reconnect = 0;
1075 imp->imp_force_verify = 0;
1077 imp->imp_connect_data = *ocd;
1079 CDEBUG(D_HA, "%s: connect to target with instance %u\n",
1080 imp->imp_obd->obd_name, ocd->ocd_instance);
1081 exp = class_conn2export(&imp->imp_dlm_handle);
1083 spin_unlock(&imp->imp_lock);
1086 /* This could happen if export is cleaned during the
1088 CERROR("%s: missing export after connect\n",
1089 imp->imp_obd->obd_name);
1090 GOTO(out, rc = -ENODEV);
1093 /* check that server granted subset of flags we asked for. */
1094 if ((ocd->ocd_connect_flags & imp->imp_connect_flags_orig) !=
1095 ocd->ocd_connect_flags) {
1096 CERROR("%s: Server didn't grant requested subset of flags: "
1097 "asked=%#llx granted=%#llx\n",
1098 imp->imp_obd->obd_name, imp->imp_connect_flags_orig,
1099 ocd->ocd_connect_flags);
1100 GOTO(out, rc = -EPROTO);
1103 if ((ocd->ocd_connect_flags2 & imp->imp_connect_flags2_orig) !=
1104 ocd->ocd_connect_flags2) {
1105 CERROR("%s: Server didn't grant requested subset of flags2: "
1106 "asked=%#llx granted=%#llx\n",
1107 imp->imp_obd->obd_name, imp->imp_connect_flags2_orig,
1108 ocd->ocd_connect_flags2);
1109 GOTO(out, rc = -EPROTO);
1112 if (!(imp->imp_connect_flags_orig & OBD_CONNECT_LIGHTWEIGHT) &&
1113 (imp->imp_connect_flags_orig & OBD_CONNECT_MDS_MDS) &&
1114 (imp->imp_connect_flags_orig & OBD_CONNECT_FID) &&
1115 (ocd->ocd_connect_flags & OBD_CONNECT_VERSION)) {
1116 __u32 major = OBD_OCD_VERSION_MAJOR(ocd->ocd_version);
1117 __u32 minor = OBD_OCD_VERSION_MINOR(ocd->ocd_version);
1118 __u32 patch = OBD_OCD_VERSION_PATCH(ocd->ocd_version);
1120 /* We do not support the MDT-MDT interoperations with
1121 * different version MDT because of protocol changes. */
1122 if (unlikely(major != LUSTRE_MAJOR ||
1123 minor != LUSTRE_MINOR ||
1124 abs(patch - LUSTRE_PATCH) > 3)) {
1125 LCONSOLE_WARN("%s: import %p (%u.%u.%u.%u) tried the "
1126 "connection to different version MDT "
1127 "(%d.%d.%d.%d) %s\n",
1128 imp->imp_obd->obd_name, imp, LUSTRE_MAJOR,
1129 LUSTRE_MINOR, LUSTRE_PATCH, LUSTRE_FIX,
1130 major, minor, patch,
1131 OBD_OCD_VERSION_FIX(ocd->ocd_version),
1132 imp->imp_connection->c_remote_uuid.uuid);
1134 GOTO(out, rc = -EPROTO);
1138 old_connect_flags = exp_connect_flags(exp);
1139 exp->exp_connect_data = *ocd;
1140 imp->imp_obd->obd_self_export->exp_connect_data = *ocd;
1142 /* The net statistics after (re-)connect is not valid anymore,
1143 * because may reflect other routing, etc. */
1144 at_reinit(&imp->imp_at.iat_net_latency, 0, 0);
1145 ptlrpc_at_adj_net_latency(request,
1146 lustre_msg_get_service_time(request->rq_repmsg));
1148 /* Import flags should be updated before waking import at FULL state */
1149 rc = ptlrpc_connect_set_flags(imp, ocd, old_connect_flags, exp,
1150 aa->pcaa_initial_connect);
1151 class_export_put(exp);
1157 obd_import_event(imp->imp_obd, imp, IMP_EVENT_OCD);
1159 if (aa->pcaa_initial_connect) {
1160 spin_lock(&imp->imp_lock);
1161 if (msg_flags & MSG_CONNECT_REPLAYABLE) {
1162 imp->imp_replayable = 1;
1163 CDEBUG(D_HA, "connected to replayable target: %s\n",
1164 obd2cli_tgt(imp->imp_obd));
1166 imp->imp_replayable = 0;
1169 /* if applies, adjust the imp->imp_msg_magic here
1170 * according to reply flags */
1172 imp->imp_remote_handle =
1173 *lustre_msg_get_handle(request->rq_repmsg);
1175 /* Initial connects are allowed for clients with non-random
1176 * uuids when servers are in recovery. Simply signal the
1177 * servers replay is complete and wait in REPLAY_WAIT. */
1178 if (msg_flags & MSG_CONNECT_RECOVERING) {
1179 CDEBUG(D_HA, "connect to %s during recovery\n",
1180 obd2cli_tgt(imp->imp_obd));
1181 import_set_state_nolock(imp, LUSTRE_IMP_REPLAY_LOCKS);
1182 spin_unlock(&imp->imp_lock);
1184 spin_unlock(&imp->imp_lock);
1185 ptlrpc_activate_import(imp, true);
1188 GOTO(finish, rc = 0);
1191 /* Determine what recovery state to move the import to. */
1192 if (MSG_CONNECT_RECONNECT & msg_flags) {
1193 memset(&old_hdl, 0, sizeof(old_hdl));
1194 if (!memcmp(&old_hdl, lustre_msg_get_handle(request->rq_repmsg),
1195 sizeof (old_hdl))) {
1196 LCONSOLE_WARN("Reconnect to %s (at @%s) failed due "
1197 "bad handle %#llx\n",
1198 obd2cli_tgt(imp->imp_obd),
1199 imp->imp_connection->c_remote_uuid.uuid,
1200 imp->imp_dlm_handle.cookie);
1201 GOTO(out, rc = -ENOTCONN);
1204 if (memcmp(&imp->imp_remote_handle,
1205 lustre_msg_get_handle(request->rq_repmsg),
1206 sizeof(imp->imp_remote_handle))) {
1207 int level = msg_flags & MSG_CONNECT_RECOVERING ?
1210 /* Bug 16611/14775: if server handle have changed,
1211 * that means some sort of disconnection happened.
1212 * If the server is not in recovery, that also means it
1213 * already erased all of our state because of previous
1214 * eviction. If it is in recovery - we are safe to
1215 * participate since we can reestablish all of our state
1216 * with server again */
1217 if ((MSG_CONNECT_RECOVERING & msg_flags)) {
1218 CDEBUG(level,"%s@%s changed server handle from "
1220 " but is still in recovery\n",
1221 obd2cli_tgt(imp->imp_obd),
1222 imp->imp_connection->c_remote_uuid.uuid,
1223 imp->imp_remote_handle.cookie,
1224 lustre_msg_get_handle(
1225 request->rq_repmsg)->cookie);
1227 LCONSOLE_WARN("Evicted from %s (at %s) "
1228 "after server handle changed from "
1230 obd2cli_tgt(imp->imp_obd),
1231 imp->imp_connection-> \
1233 imp->imp_remote_handle.cookie,
1234 lustre_msg_get_handle(
1235 request->rq_repmsg)->cookie);
1239 imp->imp_remote_handle =
1240 *lustre_msg_get_handle(request->rq_repmsg);
1242 if (!(MSG_CONNECT_RECOVERING & msg_flags)) {
1243 import_set_state(imp, LUSTRE_IMP_EVICTED);
1244 GOTO(finish, rc = 0);
1248 CDEBUG(D_HA, "reconnected to %s@%s after partition\n",
1249 obd2cli_tgt(imp->imp_obd),
1250 imp->imp_connection->c_remote_uuid.uuid);
1253 if (imp->imp_invalid) {
1254 CDEBUG(D_HA, "%s: reconnected but import is invalid; "
1255 "marking evicted\n", imp->imp_obd->obd_name);
1256 import_set_state(imp, LUSTRE_IMP_EVICTED);
1257 } else if (MSG_CONNECT_RECOVERING & msg_flags) {
1258 CDEBUG(D_HA, "%s: reconnected to %s during replay\n",
1259 imp->imp_obd->obd_name,
1260 obd2cli_tgt(imp->imp_obd));
1262 spin_lock(&imp->imp_lock);
1263 imp->imp_resend_replay = 1;
1264 spin_unlock(&imp->imp_lock);
1266 import_set_state(imp, imp->imp_replay_state);
1268 import_set_state(imp, LUSTRE_IMP_RECOVER);
1270 } else if ((MSG_CONNECT_RECOVERING & msg_flags) && !imp->imp_invalid) {
1271 LASSERT(imp->imp_replayable);
1272 imp->imp_remote_handle =
1273 *lustre_msg_get_handle(request->rq_repmsg);
1274 imp->imp_last_replay_transno = 0;
1275 imp->imp_replay_cursor = &imp->imp_committed_list;
1276 import_set_state(imp, LUSTRE_IMP_REPLAY);
1277 } else if ((ocd->ocd_connect_flags & OBD_CONNECT_LIGHTWEIGHT) != 0 &&
1278 !imp->imp_invalid) {
1280 obd_import_event(imp->imp_obd, imp, IMP_EVENT_INVALIDATE);
1281 /* The below message is checked in recovery-small.sh test_106 */
1282 DEBUG_REQ(D_HA, request, "%s: lwp recover",
1283 imp->imp_obd->obd_name);
1284 imp->imp_remote_handle =
1285 *lustre_msg_get_handle(request->rq_repmsg);
1286 import_set_state(imp, LUSTRE_IMP_RECOVER);
1288 DEBUG_REQ(D_HA, request,
1289 "%s: evicting (reconnect/recover flags not set: %x)",
1290 imp->imp_obd->obd_name, msg_flags);
1291 imp->imp_remote_handle =
1292 *lustre_msg_get_handle(request->rq_repmsg);
1293 import_set_state(imp, LUSTRE_IMP_EVICTED);
1296 /* Sanity checks for a reconnected import. */
1297 if (!(imp->imp_replayable) != !(msg_flags & MSG_CONNECT_REPLAYABLE))
1298 CERROR("imp_replayable flag does not match server after reconnect. We should LBUG right here.\n");
1300 if (lustre_msg_get_last_committed(request->rq_repmsg) > 0 &&
1301 lustre_msg_get_last_committed(request->rq_repmsg) <
1302 aa->pcaa_peer_committed) {
1303 static bool printed;
1305 /* The below message is checked in recovery-small.sh test_54 */
1306 CERROR("%s: went back in time (transno %lld was previously committed, server now claims %lld)!\n",
1307 obd2cli_tgt(imp->imp_obd), aa->pcaa_peer_committed,
1308 lustre_msg_get_last_committed(request->rq_repmsg));
1310 CERROR("For further information, see http://doc.lustre.org/lustre_manual.xhtml#went_back_in_time\n");
1316 ptlrpc_prepare_replay(imp);
1317 rc = ptlrpc_import_recovery_state_machine(imp);
1318 if (rc == -ENOTCONN) {
1319 CDEBUG(D_HA, "evicted/aborted by %s@%s during recovery;"
1320 "invalidating and reconnecting\n",
1321 obd2cli_tgt(imp->imp_obd),
1322 imp->imp_connection->c_remote_uuid.uuid);
1323 ptlrpc_connect_import(imp);
1324 spin_lock(&imp->imp_lock);
1325 imp->imp_connected = 0;
1326 imp->imp_connect_tried = 1;
1327 spin_unlock(&imp->imp_lock);
1333 class_export_put(exp);
1335 spin_lock(&imp->imp_lock);
1336 imp->imp_connected = 0;
1337 imp->imp_connect_tried = 1;
1342 import_set_state_nolock(imp, LUSTRE_IMP_DISCON);
1343 if (rc == -EACCES) {
1345 * Give up trying to reconnect
1346 * EACCES means client has no permission for connection
1348 imp->imp_obd->obd_no_recov = 1;
1349 ptlrpc_deactivate_import_nolock(imp);
1351 } else if (rc == -EPROTO) {
1352 struct obd_connect_data *ocd;
1354 /* reply message might not be ready */
1355 if (request->rq_repmsg == NULL) {
1356 spin_unlock(&imp->imp_lock);
1360 ocd = req_capsule_server_get(&request->rq_pill,
1362 /* Servers are not supposed to refuse connections from
1363 * clients based on version, only connection feature
1364 * flags. We should never see this from llite, but it
1365 * may be useful for debugging in the future. */
1367 (ocd->ocd_connect_flags & OBD_CONNECT_VERSION) &&
1368 (ocd->ocd_version != LUSTRE_VERSION_CODE)) {
1369 LCONSOLE_ERROR_MSG(0x16a, "Server %s version "
1371 " refused connection from this client "
1372 "with an incompatible version (%s). "
1373 "Client must be recompiled\n",
1374 obd2cli_tgt(imp->imp_obd),
1375 OBD_OCD_VERSION_MAJOR(ocd->ocd_version),
1376 OBD_OCD_VERSION_MINOR(ocd->ocd_version),
1377 OBD_OCD_VERSION_PATCH(ocd->ocd_version),
1378 OBD_OCD_VERSION_FIX(ocd->ocd_version),
1379 LUSTRE_VERSION_STRING);
1380 ptlrpc_deactivate_import_nolock(imp);
1381 import_set_state_nolock(imp, LUSTRE_IMP_CLOSED);
1385 spin_unlock(&imp->imp_lock);
1388 obd_import_event(imp->imp_obd, imp, IMP_EVENT_INACTIVE);
1393 ptlrpc_maybe_ping_import_soon(imp);
1395 CDEBUG(D_HA, "recovery of %s on %s failed (%d)\n",
1396 obd2cli_tgt(imp->imp_obd),
1397 (char *)imp->imp_connection->c_remote_uuid.uuid, rc);
1399 spin_unlock(&imp->imp_lock);
1402 wake_up_all(&imp->imp_recovery_waitq);
1407 * interpret callback for "completed replay" RPCs.
1408 * \see signal_completed_replay
1410 static int completed_replay_interpret(const struct lu_env *env,
1411 struct ptlrpc_request *req,
1415 atomic_dec(&req->rq_import->imp_replay_inflight);
1416 if (req->rq_status == 0 && !req->rq_import->imp_vbr_failed) {
1417 ptlrpc_import_recovery_state_machine(req->rq_import);
1419 if (req->rq_import->imp_vbr_failed) {
1421 "%s: version recovery fails, reconnecting\n",
1422 req->rq_import->imp_obd->obd_name);
1424 CDEBUG(D_HA, "%s: LAST_REPLAY message error: %d, "
1426 req->rq_import->imp_obd->obd_name,
1429 ptlrpc_connect_import(req->rq_import);
1436 * Let server know that we have no requests to replay anymore.
1437 * Achieved by just sending a PING request
1439 static int signal_completed_replay(struct obd_import *imp)
1441 struct ptlrpc_request *req;
1444 if (unlikely(OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_FINISH_REPLAY)))
1447 LASSERT(atomic_read(&imp->imp_replay_inflight) == 0);
1448 atomic_inc(&imp->imp_replay_inflight);
1450 req = ptlrpc_request_alloc_pack(imp, &RQF_OBD_PING, LUSTRE_OBD_VERSION,
1453 atomic_dec(&imp->imp_replay_inflight);
1457 ptlrpc_request_set_replen(req);
1458 req->rq_send_state = LUSTRE_IMP_REPLAY_WAIT;
1459 lustre_msg_add_flags(req->rq_reqmsg,
1460 MSG_LOCK_REPLAY_DONE | MSG_REQ_REPLAY_DONE);
1462 req->rq_timeout *= 3;
1463 req->rq_interpret_reply = completed_replay_interpret;
1465 ptlrpcd_add_req(req);
1470 * In kernel code all import invalidation happens in its own
1471 * separate thread, so that whatever application happened to encounter
1472 * a problem could still be killed or otherwise continue
1474 static int ptlrpc_invalidate_import_thread(void *data)
1476 struct obd_import *imp = data;
1480 unshare_fs_struct();
1482 CDEBUG(D_HA, "thread invalidate import %s to %s@%s\n",
1483 imp->imp_obd->obd_name, obd2cli_tgt(imp->imp_obd),
1484 imp->imp_connection->c_remote_uuid.uuid);
1486 ptlrpc_invalidate_import(imp);
1488 if (obd_dump_on_eviction) {
1489 CERROR("dump the log upon eviction\n");
1490 libcfs_debug_dumplog();
1493 import_set_state(imp, LUSTRE_IMP_RECOVER);
1494 ptlrpc_import_recovery_state_machine(imp);
1496 class_import_put(imp);
1501 * This is the state machine for client-side recovery on import.
1503 * Typicaly we have two possibly paths. If we came to server and it is not
1504 * in recovery, we just enter IMP_EVICTED state, invalidate our import
1505 * state and reconnect from scratch.
1506 * If we came to server that is in recovery, we enter IMP_REPLAY import state.
1507 * We go through our list of requests to replay and send them to server one by
1509 * After sending all request from the list we change import state to
1510 * IMP_REPLAY_LOCKS and re-request all the locks we believe we have from server
1511 * and also all the locks we don't yet have and wait for server to grant us.
1512 * After that we send a special "replay completed" request and change import
1513 * state to IMP_REPLAY_WAIT.
1514 * Upon receiving reply to that "replay completed" RPC we enter IMP_RECOVER
1515 * state and resend all requests from sending list.
1516 * After that we promote import to FULL state and send all delayed requests
1517 * and import is fully operational after that.
1520 int ptlrpc_import_recovery_state_machine(struct obd_import *imp)
1528 if (imp->imp_state == LUSTRE_IMP_EVICTED) {
1529 deuuidify(obd2cli_tgt(imp->imp_obd), NULL,
1530 &target_start, &target_len);
1531 /* Don't care about MGC eviction */
1532 if (strcmp(imp->imp_obd->obd_type->typ_name,
1533 LUSTRE_MGC_NAME) != 0) {
1534 LCONSOLE_ERROR_MSG(0x167, "%s: This client was evicted "
1535 "by %.*s; in progress operations "
1536 "using this service will fail.\n",
1537 imp->imp_obd->obd_name, target_len,
1539 LASSERTF(!obd_lbug_on_eviction, "LBUG upon eviction");
1541 CDEBUG(D_HA, "evicted from %s@%s; invalidating\n",
1542 obd2cli_tgt(imp->imp_obd),
1543 imp->imp_connection->c_remote_uuid.uuid);
1544 /* reset vbr_failed flag upon eviction */
1545 spin_lock(&imp->imp_lock);
1546 imp->imp_vbr_failed = 0;
1547 spin_unlock(&imp->imp_lock);
1550 struct task_struct *task;
1551 /* bug 17802: XXX client_disconnect_export vs connect request
1552 * race. if client is evicted at this time then we start
1553 * invalidate thread without reference to import and import can
1554 * be freed at same time. */
1555 class_import_get(imp);
1556 task = kthread_run(ptlrpc_invalidate_import_thread, imp,
1559 class_import_put(imp);
1560 CERROR("error starting invalidate thread: %d\n", rc);
1569 if (imp->imp_state == LUSTRE_IMP_REPLAY) {
1570 CDEBUG(D_HA, "replay requested by %s\n",
1571 obd2cli_tgt(imp->imp_obd));
1572 rc = ptlrpc_replay_next(imp, &inflight);
1573 if (inflight == 0 &&
1574 atomic_read(&imp->imp_replay_inflight) == 0) {
1575 import_set_state(imp, LUSTRE_IMP_REPLAY_LOCKS);
1576 rc = ldlm_replay_locks(imp);
1583 if (imp->imp_state == LUSTRE_IMP_REPLAY_LOCKS) {
1584 if (atomic_read(&imp->imp_replay_inflight) == 0) {
1585 import_set_state(imp, LUSTRE_IMP_REPLAY_WAIT);
1586 rc = signal_completed_replay(imp);
1592 if (imp->imp_state == LUSTRE_IMP_REPLAY_WAIT) {
1593 if (atomic_read(&imp->imp_replay_inflight) == 0) {
1594 import_set_state(imp, LUSTRE_IMP_RECOVER);
1598 if (imp->imp_state == LUSTRE_IMP_RECOVER) {
1599 struct ptlrpc_connection *conn = imp->imp_connection;
1601 rc = ptlrpc_resend(imp);
1604 ptlrpc_activate_import(imp, true);
1606 LCONSOLE_INFO("%s: Connection restored to %s (at %s)\n",
1607 imp->imp_obd->obd_name,
1608 obd_uuid2str(&conn->c_remote_uuid),
1609 obd_import_nid2str(imp));
1612 if (imp->imp_state == LUSTRE_IMP_FULL) {
1613 wake_up_all(&imp->imp_recovery_waitq);
1614 ptlrpc_wake_delayed(imp);
1621 static struct ptlrpc_request *ptlrpc_disconnect_prep_req(struct obd_import *imp)
1623 struct ptlrpc_request *req;
1627 switch (imp->imp_connect_op) {
1629 rq_opc = OST_DISCONNECT;
1632 rq_opc = MDS_DISCONNECT;
1635 rq_opc = MGS_DISCONNECT;
1639 CERROR("%s: don't know how to disconnect from %s "
1640 "(connect_op %d): rc = %d\n",
1641 imp->imp_obd->obd_name, obd2cli_tgt(imp->imp_obd),
1642 imp->imp_connect_op, rc);
1643 RETURN(ERR_PTR(rc));
1646 req = ptlrpc_request_alloc_pack(imp, &RQF_MDS_DISCONNECT,
1647 LUSTRE_OBD_VERSION, rq_opc);
1649 RETURN(ERR_PTR(-ENOMEM));
1651 /* We are disconnecting, do not retry a failed DISCONNECT rpc if
1652 * it fails. We can get through the above with a down server
1653 * if the client doesn't know the server is gone yet. */
1654 req->rq_no_resend = 1;
1656 /* We want client umounts to happen quickly, no matter the
1658 req->rq_timeout = min_t(int, req->rq_timeout,
1659 INITIAL_CONNECT_TIMEOUT);
1661 import_set_state(imp, LUSTRE_IMP_CONNECTING);
1662 req->rq_send_state = LUSTRE_IMP_CONNECTING;
1663 ptlrpc_request_set_replen(req);
1668 int ptlrpc_disconnect_import(struct obd_import *imp, int noclose)
1670 struct ptlrpc_request *req;
1674 if (imp->imp_obd->obd_force)
1675 GOTO(set_state, rc);
1677 /* probably the import has been disconnected already being idle */
1678 spin_lock(&imp->imp_lock);
1679 if (imp->imp_state == LUSTRE_IMP_IDLE)
1681 spin_unlock(&imp->imp_lock);
1683 if (ptlrpc_import_in_recovery(imp)) {
1684 struct l_wait_info lwi;
1685 long timeout_jiffies;
1689 if (imp->imp_server_timeout)
1690 timeout = obd_timeout >> 1;
1692 timeout = obd_timeout;
1697 req_portal = imp->imp_client->cli_request_portal;
1698 idx = import_at_get_index(imp, req_portal);
1699 timeout = at_get(&imp->imp_at.iat_service_estimate[idx]);
1702 timeout_jiffies = cfs_time_seconds(timeout);
1703 lwi = LWI_TIMEOUT_INTR(max_t(long, timeout_jiffies, 1),
1704 back_to_sleep, LWI_ON_SIGNAL_NOOP, NULL);
1705 rc = l_wait_event(imp->imp_recovery_waitq,
1706 !ptlrpc_import_in_recovery(imp), &lwi);
1710 spin_lock(&imp->imp_lock);
1711 if (imp->imp_state != LUSTRE_IMP_FULL)
1713 spin_unlock(&imp->imp_lock);
1715 req = ptlrpc_disconnect_prep_req(imp);
1717 GOTO(set_state, rc = PTR_ERR(req));
1718 rc = ptlrpc_queue_wait(req);
1719 ptlrpc_req_finished(req);
1722 spin_lock(&imp->imp_lock);
1725 import_set_state_nolock(imp, LUSTRE_IMP_DISCON);
1727 import_set_state_nolock(imp, LUSTRE_IMP_CLOSED);
1728 memset(&imp->imp_remote_handle, 0, sizeof(imp->imp_remote_handle));
1729 spin_unlock(&imp->imp_lock);
1731 if (rc == -ETIMEDOUT || rc == -ENOTCONN || rc == -ESHUTDOWN)
1735 EXPORT_SYMBOL(ptlrpc_disconnect_import);
1737 static void ptlrpc_reset_reqs_generation(struct obd_import *imp)
1739 struct ptlrpc_request *old, *tmp;
1741 /* tag all resendable requests generated before disconnection
1742 * notice this code is part of disconnect-at-idle path only */
1743 list_for_each_entry_safe(old, tmp, &imp->imp_delayed_list,
1745 spin_lock(&old->rq_lock);
1746 if (old->rq_import_generation == imp->imp_generation - 1 &&
1748 old->rq_import_generation = imp->imp_generation;
1749 spin_unlock(&old->rq_lock);
1753 static int ptlrpc_disconnect_idle_interpret(const struct lu_env *env,
1754 struct ptlrpc_request *req,
1757 struct obd_import *imp = req->rq_import;
1760 DEBUG_REQ(D_HA, req, "inflight=%d, refcount=%d: rc = %d",
1761 atomic_read(&imp->imp_inflight),
1762 atomic_read(&imp->imp_refcount), rc);
1764 spin_lock(&imp->imp_lock);
1765 /* DISCONNECT reply can be late and another connection can just
1766 * be initiated. so we have to abort disconnection. */
1767 if (req->rq_import_generation == imp->imp_generation &&
1768 imp->imp_state != LUSTRE_IMP_CLOSED) {
1769 LASSERTF(imp->imp_state == LUSTRE_IMP_CONNECTING,
1770 "%s\n", ptlrpc_import_state_name(imp->imp_state));
1771 imp->imp_state = LUSTRE_IMP_IDLE;
1772 memset(&imp->imp_remote_handle, 0,
1773 sizeof(imp->imp_remote_handle));
1774 /* take our DISCONNECT into account */
1775 if (atomic_read(&imp->imp_inflight) > 1) {
1776 imp->imp_generation++;
1777 imp->imp_initiated_at = imp->imp_generation;
1778 import_set_state_nolock(imp, LUSTRE_IMP_NEW);
1779 ptlrpc_reset_reqs_generation(imp);
1785 rc = ptlrpc_connect_import_locked(imp);
1787 ptlrpc_pinger_add_import(imp);
1789 spin_unlock(&imp->imp_lock);
1795 int ptlrpc_disconnect_and_idle_import(struct obd_import *imp)
1797 struct ptlrpc_request *req;
1800 if (imp->imp_obd->obd_force)
1803 if (ptlrpc_import_in_recovery(imp))
1806 spin_lock(&imp->imp_lock);
1807 if (imp->imp_state != LUSTRE_IMP_FULL) {
1808 spin_unlock(&imp->imp_lock);
1811 spin_unlock(&imp->imp_lock);
1813 req = ptlrpc_disconnect_prep_req(imp);
1815 RETURN(PTR_ERR(req));
1817 CDEBUG_LIMIT(imp->imp_idle_debug, "%s: disconnect after %llus idle\n",
1818 imp->imp_obd->obd_name,
1819 ktime_get_real_seconds() - imp->imp_last_reply_time);
1820 req->rq_interpret_reply = ptlrpc_disconnect_idle_interpret;
1821 ptlrpcd_add_req(req);
1825 EXPORT_SYMBOL(ptlrpc_disconnect_and_idle_import);
1827 void ptlrpc_cleanup_imp(struct obd_import *imp)
1831 spin_lock(&imp->imp_lock);
1833 import_set_state_nolock(imp, LUSTRE_IMP_CLOSED);
1834 imp->imp_generation++;
1835 ptlrpc_abort_inflight(imp);
1837 spin_unlock(&imp->imp_lock);
1842 /* Adaptive Timeout utils */
1844 /* Update at_current with the specified value (bounded by at_min and at_max),
1845 * as well as the AT history "bins".
1846 * - Bin into timeslices using AT_BINS bins.
1847 * - This gives us a max of the last at_history seconds without the storage,
1848 * but still smoothing out a return to normalcy from a slow response.
1849 * - (E.g. remember the maximum latency in each minute of the last 4 minutes.)
1851 int at_measured(struct adaptive_timeout *at, unsigned int val)
1853 unsigned int old = at->at_current;
1854 time64_t now = ktime_get_real_seconds();
1855 long binlimit = max_t(long, at_history / AT_BINS, 1);
1858 CDEBUG(D_OTHER, "add %u to %p time=%lu v=%u (%u %u %u %u)\n",
1859 val, at, (long)(now - at->at_binstart), at->at_current,
1860 at->at_hist[0], at->at_hist[1], at->at_hist[2], at->at_hist[3]);
1863 /* 0's don't count, because we never want our timeout to
1864 drop to 0, and because 0 could mean an error */
1867 spin_lock(&at->at_lock);
1869 if (unlikely(at->at_binstart == 0)) {
1870 /* Special case to remove default from history */
1871 at->at_current = val;
1872 at->at_worst_ever = val;
1873 at->at_worst_time = now;
1874 at->at_hist[0] = val;
1875 at->at_binstart = now;
1876 } else if (now - at->at_binstart < binlimit ) {
1878 at->at_hist[0] = max(val, at->at_hist[0]);
1879 at->at_current = max(val, at->at_current);
1882 unsigned int maxv = val;
1884 /* move bins over */
1885 shift = (u32)(now - at->at_binstart) / binlimit;
1887 for(i = AT_BINS - 1; i >= 0; i--) {
1889 at->at_hist[i] = at->at_hist[i - shift];
1890 maxv = max(maxv, at->at_hist[i]);
1895 at->at_hist[0] = val;
1896 at->at_current = maxv;
1897 at->at_binstart += shift * binlimit;
1900 if (at->at_current > at->at_worst_ever) {
1901 at->at_worst_ever = at->at_current;
1902 at->at_worst_time = now;
1905 if (at->at_flags & AT_FLG_NOHIST)
1906 /* Only keep last reported val; keeping the rest of the history
1908 at->at_current = val;
1911 at->at_current = min(at->at_current, at_max);
1912 at->at_current = max(at->at_current, at_min);
1914 if (at->at_current != old)
1915 CDEBUG(D_OTHER, "AT %p change: old=%u new=%u delta=%d "
1916 "(val=%u) hist %u %u %u %u\n", at,
1917 old, at->at_current, at->at_current - old, val,
1918 at->at_hist[0], at->at_hist[1], at->at_hist[2],
1921 /* if we changed, report the old value */
1922 old = (at->at_current != old) ? old : 0;
1924 spin_unlock(&at->at_lock);
1928 /* Find the imp_at index for a given portal; assign if space available */
1929 int import_at_get_index(struct obd_import *imp, int portal)
1931 struct imp_at *at = &imp->imp_at;
1934 for (i = 0; i < IMP_AT_MAX_PORTALS; i++) {
1935 if (at->iat_portal[i] == portal)
1937 if (at->iat_portal[i] == 0)
1942 /* Not found in list, add it under a lock */
1943 spin_lock(&imp->imp_lock);
1945 /* Check unused under lock */
1946 for (; i < IMP_AT_MAX_PORTALS; i++) {
1947 if (at->iat_portal[i] == portal)
1949 if (at->iat_portal[i] == 0)
1954 /* Not enough portals? */
1955 LASSERT(i < IMP_AT_MAX_PORTALS);
1957 at->iat_portal[i] = portal;
1959 spin_unlock(&imp->imp_lock);