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 (imp->imp_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;
307 atomic_inc(&imp->imp_inval_count);
309 if (!imp->imp_invalid || imp->imp_obd->obd_no_recov)
310 ptlrpc_deactivate_import(imp);
312 if (OBD_FAIL_PRECHECK(OBD_FAIL_PTLRPC_CONNECT_RACE)) {
313 OBD_RACE(OBD_FAIL_PTLRPC_CONNECT_RACE);
314 msleep(10 * MSEC_PER_SEC);
316 CFS_FAIL_TIMEOUT(OBD_FAIL_MGS_CONNECT_NET, 3 * cfs_fail_val / 2);
317 LASSERT(imp->imp_invalid);
319 /* Wait forever until inflight == 0. We really can't do it another
320 * way because in some cases we need to wait for very long reply
321 * unlink. We can't do anything before that because there is really
322 * 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 rc = wait_event_idle_timeout(
351 imp->imp_recovery_waitq,
352 (atomic_read(&imp->imp_inflight) == 0),
356 const char *cli_tgt = obd2cli_tgt(imp->imp_obd);
358 CERROR("%s: timeout waiting for callback (%d != 0)\n",
359 cli_tgt, atomic_read(&imp->imp_inflight));
361 spin_lock(&imp->imp_lock);
362 if (atomic_read(&imp->imp_inflight) == 0) {
363 int count = atomic_read(&imp->imp_unregistering);
365 /* We know that "unregistering" rpcs only can
366 * survive in sending or delaying lists (they
367 * maybe waiting for long reply unlink in
368 * sluggish nets). Let's check this. If there
369 * is no inflight and unregistering != 0, this
371 LASSERTF(count == 0, "Some RPCs are still "
372 "unregistering: %d\n", count);
374 /* Let's save one loop as soon as inflight have
375 * dropped to zero. No new inflights possible at
379 list_for_each_safe(tmp, n,
380 &imp->imp_sending_list) {
381 req = list_entry(tmp,
382 struct ptlrpc_request,
384 DEBUG_REQ(D_ERROR, req,
385 "still on sending list");
387 list_for_each_safe(tmp, n,
388 &imp->imp_delayed_list) {
389 req = list_entry(tmp,
390 struct ptlrpc_request,
392 DEBUG_REQ(D_ERROR, req,
393 "still on delayed list");
396 CERROR("%s: Unregistering RPCs found (%d). "
397 "Network is sluggish? Waiting for them "
398 "to error out.\n", cli_tgt,
399 atomic_read(&imp->imp_unregistering));
401 spin_unlock(&imp->imp_lock);
406 * Let's additionally check that no new rpcs added to import in
407 * "invalidate" state.
409 LASSERT(atomic_read(&imp->imp_inflight) == 0);
410 obd_import_event(imp->imp_obd, imp, IMP_EVENT_INVALIDATE);
411 sptlrpc_import_flush_all_ctx(imp);
413 atomic_dec(&imp->imp_inval_count);
414 wake_up(&imp->imp_recovery_waitq);
416 EXPORT_SYMBOL(ptlrpc_invalidate_import);
418 /* unset imp_invalid */
419 void ptlrpc_activate_import(struct obd_import *imp, bool set_state_full)
421 struct obd_device *obd = imp->imp_obd;
423 spin_lock(&imp->imp_lock);
424 if (imp->imp_deactive != 0) {
425 LASSERT(imp->imp_state != LUSTRE_IMP_FULL);
426 if (imp->imp_state != LUSTRE_IMP_DISCON)
427 import_set_state_nolock(imp, LUSTRE_IMP_DISCON);
428 spin_unlock(&imp->imp_lock);
432 import_set_state_nolock(imp, LUSTRE_IMP_FULL);
434 imp->imp_invalid = 0;
436 spin_unlock(&imp->imp_lock);
437 obd_import_event(obd, imp, IMP_EVENT_ACTIVE);
439 EXPORT_SYMBOL(ptlrpc_activate_import);
441 void ptlrpc_pinger_force(struct obd_import *imp)
443 CDEBUG(D_HA, "%s: waking up pinger s:%s\n", obd2cli_tgt(imp->imp_obd),
444 ptlrpc_import_state_name(imp->imp_state));
446 spin_lock(&imp->imp_lock);
447 imp->imp_force_verify = 1;
448 spin_unlock(&imp->imp_lock);
450 if (imp->imp_state != LUSTRE_IMP_CONNECTING)
451 ptlrpc_pinger_wake_up();
453 EXPORT_SYMBOL(ptlrpc_pinger_force);
455 void ptlrpc_fail_import(struct obd_import *imp, __u32 conn_cnt)
459 LASSERT(!imp->imp_dlm_fake);
461 if (ptlrpc_set_import_discon(imp, conn_cnt, true))
462 ptlrpc_pinger_force(imp);
467 int ptlrpc_reconnect_import(struct obd_import *imp)
469 #ifdef CONFIG_LUSTRE_FS_PINGER
470 long timeout_jiffies = cfs_time_seconds(obd_timeout);
473 ptlrpc_pinger_force(imp);
475 CDEBUG(D_HA, "%s: recovery started, waiting %u seconds\n",
476 obd2cli_tgt(imp->imp_obd), obd_timeout);
478 rc = wait_event_idle_timeout(imp->imp_recovery_waitq,
479 !ptlrpc_import_in_recovery(imp),
485 CDEBUG(D_HA, "%s: recovery finished s:%s\n", obd2cli_tgt(imp->imp_obd),
486 ptlrpc_import_state_name(imp->imp_state));
489 ptlrpc_set_import_discon(imp, 0, false);
490 /* Force a new connect attempt */
491 ptlrpc_invalidate_import(imp);
492 /* Do a fresh connect next time by zeroing the handle */
493 ptlrpc_disconnect_import(imp, 1);
494 /* Wait for all invalidate calls to finish */
495 if (atomic_read(&imp->imp_inval_count) > 0) {
498 rc = l_wait_event_abortable(
499 imp->imp_recovery_waitq,
500 (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;
531 spin_lock(&imp->imp_lock);
533 if (list_empty(&imp->imp_conn_list)) {
535 CERROR("%s: no connections available: rc = %d\n",
536 imp->imp_obd->obd_name, rc);
537 GOTO(out_unlock, rc);
540 list_for_each_entry(conn, &imp->imp_conn_list, oic_item) {
541 CDEBUG(D_HA, "%s: connect to NID %s last attempt %lld\n",
542 imp->imp_obd->obd_name,
543 libcfs_nid2str(conn->oic_conn->c_peer.nid),
544 conn->oic_last_attempt);
546 /* If we have not tried this connection since
547 * the last successful attempt, go with this one
549 if ((conn->oic_last_attempt == 0) ||
550 conn->oic_last_attempt <= imp->imp_last_success_conn) {
556 /* If all of the connections have already been tried
557 * since the last successful connection; just choose the
558 * 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.)
580 if (tried_all && (imp->imp_conn_list.next == &imp_conn->oic_item)) {
581 struct adaptive_timeout *at = &imp->imp_at.iat_net_latency;
583 if (at_get(at) < CONNECTION_SWITCH_MAX) {
584 at_measured(at, at_get(at) + CONNECTION_SWITCH_INC);
585 if (at_get(at) > CONNECTION_SWITCH_MAX)
586 at_reset(at, CONNECTION_SWITCH_MAX);
588 LASSERT(imp_conn->oic_last_attempt);
590 "%s: tried all connections, increasing latency to %ds\n",
591 imp->imp_obd->obd_name, at_get(at));
594 imp_conn->oic_last_attempt = ktime_get_seconds();
596 /* switch connection, don't mind if it's same as the current one */
597 ptlrpc_connection_put(imp->imp_connection);
598 imp->imp_connection = ptlrpc_connection_addref(imp_conn->oic_conn);
600 dlmexp = class_conn2export(&imp->imp_dlm_handle);
602 GOTO(out_unlock, rc = -EINVAL);
603 ptlrpc_connection_put(dlmexp->exp_connection);
604 dlmexp->exp_connection = ptlrpc_connection_addref(imp_conn->oic_conn);
605 class_export_put(dlmexp);
607 if (imp->imp_conn_current != imp_conn) {
608 if (imp->imp_conn_current) {
609 deuuidify(obd2cli_tgt(imp->imp_obd), NULL,
610 &target_start, &target_len);
612 CDEBUG(D_HA, "%s: Connection changing to"
614 imp->imp_obd->obd_name,
615 target_len, target_start,
616 libcfs_nid2str(imp_conn->oic_conn->c_peer.nid));
619 imp->imp_conn_current = imp_conn;
622 /* The below message is checked in conf-sanity.sh test_35[ab] */
623 CDEBUG(D_HA, "%s: import %p using connection %s/%s\n",
624 imp->imp_obd->obd_name, imp, imp_conn->oic_uuid.uuid,
625 libcfs_nid2str(imp_conn->oic_conn->c_peer.nid));
628 spin_unlock(&imp->imp_lock);
633 * must be called under imp_lock
635 static int ptlrpc_first_transno(struct obd_import *imp, __u64 *transno)
637 struct ptlrpc_request *req;
638 struct list_head *tmp;
640 /* The requests in committed_list always have smaller transnos than
641 * the requests in replay_list */
642 if (!list_empty(&imp->imp_committed_list)) {
643 tmp = imp->imp_committed_list.next;
644 req = list_entry(tmp, struct ptlrpc_request, rq_replay_list);
645 *transno = req->rq_transno;
646 if (req->rq_transno == 0) {
647 DEBUG_REQ(D_ERROR, req,
648 "zero transno in committed_list");
653 if (!list_empty(&imp->imp_replay_list)) {
654 tmp = imp->imp_replay_list.next;
655 req = list_entry(tmp, struct ptlrpc_request, rq_replay_list);
656 *transno = req->rq_transno;
657 if (req->rq_transno == 0) {
658 DEBUG_REQ(D_ERROR, req, "zero transno in replay_list");
666 int ptlrpc_connect_import(struct obd_import *imp)
668 spin_lock(&imp->imp_lock);
669 return ptlrpc_connect_import_locked(imp);
673 * Attempt to (re)connect import \a imp. This includes all preparations,
674 * initializing CONNECT RPC request and passing it to ptlrpcd for
677 * Assumes imp->imp_lock is held, and releases it.
679 * Returns 0 on success or error code.
681 int ptlrpc_connect_import_locked(struct obd_import *imp)
683 struct obd_device *obd = imp->imp_obd;
684 int initial_connect = 0;
686 __u64 committed_before_reconnect = 0;
687 struct ptlrpc_request *request;
688 struct obd_connect_data ocd;
689 char *bufs[] = { NULL,
690 obd2cli_tgt(imp->imp_obd),
692 (char *)&imp->imp_dlm_handle,
695 struct ptlrpc_connect_async_args *aa;
699 assert_spin_locked(&imp->imp_lock);
701 if (imp->imp_state == LUSTRE_IMP_CLOSED) {
702 spin_unlock(&imp->imp_lock);
703 CERROR("can't connect to a closed import\n");
705 } else if (imp->imp_state == LUSTRE_IMP_FULL) {
706 spin_unlock(&imp->imp_lock);
707 CERROR("already connected\n");
709 } else if (imp->imp_state == LUSTRE_IMP_CONNECTING ||
710 imp->imp_state == LUSTRE_IMP_EVICTED ||
711 imp->imp_connected) {
712 spin_unlock(&imp->imp_lock);
713 CERROR("already connecting\n");
717 import_set_state_nolock(imp, LUSTRE_IMP_CONNECTING);
720 imp->imp_resend_replay = 0;
722 if (!lustre_handle_is_used(&imp->imp_remote_handle))
725 committed_before_reconnect = imp->imp_peer_committed_transno;
727 set_transno = ptlrpc_first_transno(imp,
728 &imp->imp_connect_data.ocd_transno);
729 spin_unlock(&imp->imp_lock);
731 rc = import_select_connection(imp);
735 rc = sptlrpc_import_sec_adapt(imp, NULL, NULL);
739 /* Reset connect flags to the originally requested flags, in case
740 * the server is updated on-the-fly we will get the new features. */
741 ocd = imp->imp_connect_data;
742 ocd.ocd_connect_flags = imp->imp_connect_flags_orig;
743 ocd.ocd_connect_flags2 = imp->imp_connect_flags2_orig;
744 /* Reset ocd_version each time so the server knows the exact versions */
745 ocd.ocd_version = LUSTRE_VERSION_CODE;
746 imp->imp_msghdr_flags &= ~MSGHDR_AT_SUPPORT;
747 imp->imp_msghdr_flags &= ~MSGHDR_CKSUM_INCOMPAT18;
749 rc = obd_reconnect(NULL, imp->imp_obd->obd_self_export, obd,
750 &obd->obd_uuid, &ocd, NULL);
754 request = ptlrpc_request_alloc(imp, &RQF_MDS_CONNECT);
756 GOTO(out, rc = -ENOMEM);
758 /* get SELinux policy info if any */
759 rc = sptlrpc_get_sepol(request);
761 ptlrpc_request_free(request);
765 bufs[5] = request->rq_sepol;
767 req_capsule_set_size(&request->rq_pill, &RMF_SELINUX_POL, RCL_CLIENT,
768 strlen(request->rq_sepol) ?
769 strlen(request->rq_sepol) + 1 : 0);
771 rc = ptlrpc_request_bufs_pack(request, LUSTRE_OBD_VERSION,
772 imp->imp_connect_op, bufs, NULL);
774 ptlrpc_request_free(request);
778 /* Report the rpc service time to the server so that it knows how long
779 * to wait for clients to join recovery */
780 lustre_msg_set_service_timeout(request->rq_reqmsg,
781 at_timeout2est(request->rq_timeout));
783 /* The amount of time we give the server to process the connect req.
784 * import_select_connection will increase the net latency on
785 * repeated reconnect attempts to cover slow networks.
786 * We override/ignore the server rpc completion estimate here,
787 * which may be large if this is a reconnect attempt */
788 request->rq_timeout = INITIAL_CONNECT_TIMEOUT;
789 lustre_msg_set_timeout(request->rq_reqmsg, request->rq_timeout);
791 request->rq_no_resend = request->rq_no_delay = 1;
792 request->rq_send_state = LUSTRE_IMP_CONNECTING;
793 /* Allow a slightly larger reply for future growth compatibility */
794 req_capsule_set_size(&request->rq_pill, &RMF_CONNECT_DATA, RCL_SERVER,
795 sizeof(struct obd_connect_data)+16*sizeof(__u64));
796 ptlrpc_request_set_replen(request);
797 request->rq_interpret_reply = ptlrpc_connect_interpret;
799 aa = ptlrpc_req_async_args(aa, request);
800 memset(aa, 0, sizeof *aa);
802 aa->pcaa_peer_committed = committed_before_reconnect;
803 aa->pcaa_initial_connect = initial_connect;
805 if (aa->pcaa_initial_connect) {
806 spin_lock(&imp->imp_lock);
807 imp->imp_replayable = 1;
808 spin_unlock(&imp->imp_lock);
809 lustre_msg_add_op_flags(request->rq_reqmsg,
810 MSG_CONNECT_INITIAL);
814 lustre_msg_add_op_flags(request->rq_reqmsg,
815 MSG_CONNECT_TRANSNO);
817 DEBUG_REQ(D_RPCTRACE, request, "(re)connect request (timeout %d)",
818 request->rq_timeout);
819 ptlrpcd_add_req(request);
823 import_set_state(imp, LUSTRE_IMP_DISCON);
827 EXPORT_SYMBOL(ptlrpc_connect_import);
829 static void ptlrpc_maybe_ping_import_soon(struct obd_import *imp)
833 spin_lock(&imp->imp_lock);
834 force_verify = imp->imp_force_verify != 0;
835 spin_unlock(&imp->imp_lock);
838 ptlrpc_pinger_wake_up();
841 static int ptlrpc_busy_reconnect(int rc)
843 return (rc == -EBUSY) || (rc == -EAGAIN);
846 static int ptlrpc_connect_set_flags(struct obd_import *imp,
847 struct obd_connect_data *ocd,
848 __u64 old_connect_flags,
849 struct obd_export *exp, int init_connect)
852 struct client_obd *cli = &imp->imp_obd->u.cli;
854 spin_lock(&imp->imp_lock);
855 list_move(&imp->imp_conn_current->oic_item,
856 &imp->imp_conn_list);
857 imp->imp_last_success_conn =
858 imp->imp_conn_current->oic_last_attempt;
860 spin_unlock(&imp->imp_lock);
862 if (!warned && (ocd->ocd_connect_flags & OBD_CONNECT_VERSION) &&
863 (ocd->ocd_version > LUSTRE_VERSION_CODE +
864 LUSTRE_VERSION_OFFSET_WARN ||
865 ocd->ocd_version < LUSTRE_VERSION_CODE -
866 LUSTRE_VERSION_OFFSET_WARN)) {
867 /* Sigh, some compilers do not like #ifdef in the middle
868 of macro arguments */
869 const char *older = "older than client. "
870 "Consider upgrading server";
871 const char *newer = "newer than client. "
872 "Consider upgrading client";
874 LCONSOLE_WARN("Server %s version (%d.%d.%d.%d) "
876 obd2cli_tgt(imp->imp_obd),
877 OBD_OCD_VERSION_MAJOR(ocd->ocd_version),
878 OBD_OCD_VERSION_MINOR(ocd->ocd_version),
879 OBD_OCD_VERSION_PATCH(ocd->ocd_version),
880 OBD_OCD_VERSION_FIX(ocd->ocd_version),
881 ocd->ocd_version > LUSTRE_VERSION_CODE ?
882 newer : older, LUSTRE_VERSION_STRING);
886 if (ocd->ocd_connect_flags & OBD_CONNECT_CKSUM) {
887 /* We sent to the server ocd_cksum_types with bits set
888 * for algorithms we understand. The server masked off
889 * the checksum types it doesn't support */
890 if ((ocd->ocd_cksum_types &
891 obd_cksum_types_supported_client()) == 0) {
892 LCONSOLE_ERROR("The negotiation of the checksum "
893 "alogrithm to use with server %s "
895 obd2cli_tgt(imp->imp_obd),
896 ocd->ocd_cksum_types,
897 obd_cksum_types_supported_client());
900 cli->cl_supp_cksum_types = ocd->ocd_cksum_types;
903 /* The server does not support OBD_CONNECT_CKSUM.
904 * Enforce ADLER for backward compatibility*/
905 cli->cl_supp_cksum_types = OBD_CKSUM_ADLER;
907 cli->cl_cksum_type = obd_cksum_type_select(imp->imp_obd->obd_name,
908 cli->cl_supp_cksum_types,
909 cli->cl_preferred_cksum_type);
911 if (ocd->ocd_connect_flags & OBD_CONNECT_BRW_SIZE)
912 cli->cl_max_pages_per_rpc =
913 min(ocd->ocd_brw_size >> PAGE_SHIFT,
914 cli->cl_max_pages_per_rpc);
915 else if (imp->imp_connect_op == MDS_CONNECT ||
916 imp->imp_connect_op == MGS_CONNECT)
917 cli->cl_max_pages_per_rpc = 1;
919 LASSERT((cli->cl_max_pages_per_rpc <= PTLRPC_MAX_BRW_PAGES) &&
920 (cli->cl_max_pages_per_rpc > 0));
922 client_adjust_max_dirty(cli);
924 /* Update client max modify RPCs in flight with value returned
926 if (ocd->ocd_connect_flags & OBD_CONNECT_MULTIMODRPCS)
927 cli->cl_max_mod_rpcs_in_flight = min(
928 cli->cl_max_mod_rpcs_in_flight,
929 ocd->ocd_maxmodrpcs);
931 cli->cl_max_mod_rpcs_in_flight = 1;
933 /* Reset ns_connect_flags only for initial connect. It might be
934 * changed in while using FS and if we reset it in reconnect
935 * this leads to losing user settings done before such as
936 * disable lru_resize, etc. */
937 if (old_connect_flags != exp_connect_flags(exp) || init_connect) {
938 struct ldlm_namespace *ns = imp->imp_obd->obd_namespace;
942 ns->ns_connect_flags ^ ns->ns_orig_connect_flags;
943 CDEBUG(D_HA, "%s: Resetting ns_connect_flags to server "
944 "flags: %#llx\n", imp->imp_obd->obd_name,
945 ocd->ocd_connect_flags);
946 ns->ns_connect_flags = (ns->ns_connect_flags & changed_flags) |
947 (ocd->ocd_connect_flags & ~changed_flags);
948 ns->ns_orig_connect_flags = ocd->ocd_connect_flags;
951 if (ocd->ocd_connect_flags & OBD_CONNECT_AT)
952 /* We need a per-message support flag, because
953 * a. we don't know if the incoming connect reply
954 * supports AT or not (in reply_in_callback)
955 * until we unpack it.
956 * b. failovered server means export and flags are gone
957 * (in ptlrpc_send_reply).
958 * Can only be set when we know AT is supported at
960 imp->imp_msghdr_flags |= MSGHDR_AT_SUPPORT;
962 imp->imp_msghdr_flags &= ~MSGHDR_AT_SUPPORT;
964 imp->imp_msghdr_flags |= MSGHDR_CKSUM_INCOMPAT18;
970 * Add all replay requests back to unreplied list before start replay,
971 * so that we can make sure the known replied XID is always increased
972 * only even if when replaying requests.
974 static void ptlrpc_prepare_replay(struct obd_import *imp)
976 struct ptlrpc_request *req;
978 if (imp->imp_state != LUSTRE_IMP_REPLAY ||
979 imp->imp_resend_replay)
982 /* If the server was restart during repaly, the requests may
983 * have been added to the unreplied list in former replay. */
984 spin_lock(&imp->imp_lock);
986 list_for_each_entry(req, &imp->imp_committed_list, rq_replay_list) {
987 if (list_empty(&req->rq_unreplied_list))
988 ptlrpc_add_unreplied(req);
991 list_for_each_entry(req, &imp->imp_replay_list, rq_replay_list) {
992 if (list_empty(&req->rq_unreplied_list))
993 ptlrpc_add_unreplied(req);
996 imp->imp_known_replied_xid = ptlrpc_known_replied_xid(imp);
997 spin_unlock(&imp->imp_lock);
1001 * interpret_reply callback for connect RPCs.
1002 * Looks into returned status of connect operation and decides
1003 * what to do with the import - i.e enter recovery, promote it to
1004 * full state for normal operations of disconnect it due to an error.
1006 static int ptlrpc_connect_interpret(const struct lu_env *env,
1007 struct ptlrpc_request *request,
1010 struct ptlrpc_connect_async_args *aa = data;
1011 struct obd_import *imp = request->rq_import;
1012 struct lustre_handle old_hdl;
1013 __u64 old_connect_flags;
1014 timeout_t service_timeout;
1016 struct obd_connect_data *ocd;
1017 struct obd_export *exp = NULL;
1021 spin_lock(&imp->imp_lock);
1022 if (imp->imp_state == LUSTRE_IMP_CLOSED) {
1023 imp->imp_connect_tried = 1;
1024 spin_unlock(&imp->imp_lock);
1028 imp->imp_connect_error = rc;
1030 struct ptlrpc_request *free_req;
1031 struct ptlrpc_request *tmp;
1033 /* abort all delayed requests initiated connection */
1034 list_for_each_entry_safe(free_req, tmp, &imp->imp_delayed_list,
1036 spin_lock(&free_req->rq_lock);
1037 if (free_req->rq_no_resend) {
1038 free_req->rq_err = 1;
1039 free_req->rq_status = -EIO;
1040 ptlrpc_client_wake_req(free_req);
1042 spin_unlock(&free_req->rq_lock);
1045 /* if this reconnect to busy export - not need select new target
1047 imp->imp_force_reconnect = ptlrpc_busy_reconnect(rc);
1048 spin_unlock(&imp->imp_lock);
1049 ptlrpc_maybe_ping_import_soon(imp);
1053 /* LU-7558: indicate that we are interpretting connect reply,
1054 * pltrpc_connect_import() will not try to reconnect until
1055 * interpret will finish. */
1056 imp->imp_connected = 1;
1057 spin_unlock(&imp->imp_lock);
1059 LASSERT(imp->imp_conn_current);
1061 msg_flags = lustre_msg_get_op_flags(request->rq_repmsg);
1063 ret = req_capsule_get_size(&request->rq_pill, &RMF_CONNECT_DATA,
1065 /* server replied obd_connect_data is always bigger */
1066 ocd = req_capsule_server_sized_get(&request->rq_pill,
1067 &RMF_CONNECT_DATA, ret);
1070 CERROR("%s: no connect data from server\n",
1071 imp->imp_obd->obd_name);
1076 spin_lock(&imp->imp_lock);
1078 /* All imports are pingable */
1079 imp->imp_pingable = 1;
1080 imp->imp_force_reconnect = 0;
1081 imp->imp_force_verify = 0;
1083 imp->imp_connect_data = *ocd;
1085 CDEBUG(D_HA, "%s: connect to target with instance %u\n",
1086 imp->imp_obd->obd_name, ocd->ocd_instance);
1087 exp = class_conn2export(&imp->imp_dlm_handle);
1089 spin_unlock(&imp->imp_lock);
1092 /* This could happen if export is cleaned during the
1094 CERROR("%s: missing export after connect\n",
1095 imp->imp_obd->obd_name);
1096 GOTO(out, rc = -ENODEV);
1099 /* check that server granted subset of flags we asked for. */
1100 if ((ocd->ocd_connect_flags & imp->imp_connect_flags_orig) !=
1101 ocd->ocd_connect_flags) {
1102 CERROR("%s: Server didn't grant requested subset of flags: "
1103 "asked=%#llx granted=%#llx\n",
1104 imp->imp_obd->obd_name, imp->imp_connect_flags_orig,
1105 ocd->ocd_connect_flags);
1106 GOTO(out, rc = -EPROTO);
1109 if ((ocd->ocd_connect_flags2 & imp->imp_connect_flags2_orig) !=
1110 ocd->ocd_connect_flags2) {
1111 CERROR("%s: Server didn't grant requested subset of flags2: "
1112 "asked=%#llx granted=%#llx\n",
1113 imp->imp_obd->obd_name, imp->imp_connect_flags2_orig,
1114 ocd->ocd_connect_flags2);
1115 GOTO(out, rc = -EPROTO);
1118 if (!(imp->imp_connect_flags_orig & OBD_CONNECT_LIGHTWEIGHT) &&
1119 (imp->imp_connect_flags_orig & OBD_CONNECT_MDS_MDS) &&
1120 (imp->imp_connect_flags_orig & OBD_CONNECT_FID) &&
1121 (ocd->ocd_connect_flags & OBD_CONNECT_VERSION)) {
1122 __u32 major = OBD_OCD_VERSION_MAJOR(ocd->ocd_version);
1123 __u32 minor = OBD_OCD_VERSION_MINOR(ocd->ocd_version);
1124 __u32 patch = OBD_OCD_VERSION_PATCH(ocd->ocd_version);
1126 /* We do not support the MDT-MDT interoperations with
1127 * different version MDT because of protocol changes. */
1128 if (unlikely(major != LUSTRE_MAJOR ||
1129 minor != LUSTRE_MINOR ||
1130 abs(patch - LUSTRE_PATCH) > 3)) {
1131 LCONSOLE_WARN("%s: import %p (%u.%u.%u.%u) tried the "
1132 "connection to different version MDT "
1133 "(%d.%d.%d.%d) %s\n",
1134 imp->imp_obd->obd_name, imp, LUSTRE_MAJOR,
1135 LUSTRE_MINOR, LUSTRE_PATCH, LUSTRE_FIX,
1136 major, minor, patch,
1137 OBD_OCD_VERSION_FIX(ocd->ocd_version),
1138 imp->imp_connection->c_remote_uuid.uuid);
1140 GOTO(out, rc = -EPROTO);
1144 old_connect_flags = exp_connect_flags(exp);
1145 exp->exp_connect_data = *ocd;
1146 imp->imp_obd->obd_self_export->exp_connect_data = *ocd;
1148 /* The net statistics after (re-)connect is not valid anymore,
1149 * because may reflect other routing, etc.
1151 service_timeout = lustre_msg_get_service_timeout(request->rq_repmsg);
1152 at_reinit(&imp->imp_at.iat_net_latency, 0, 0);
1153 ptlrpc_at_adj_net_latency(request, service_timeout);
1155 /* Import flags should be updated before waking import at FULL state */
1156 rc = ptlrpc_connect_set_flags(imp, ocd, old_connect_flags, exp,
1157 aa->pcaa_initial_connect);
1158 class_export_put(exp);
1164 obd_import_event(imp->imp_obd, imp, IMP_EVENT_OCD);
1166 if (aa->pcaa_initial_connect) {
1167 spin_lock(&imp->imp_lock);
1168 if (msg_flags & MSG_CONNECT_REPLAYABLE) {
1169 imp->imp_replayable = 1;
1170 CDEBUG(D_HA, "connected to replayable target: %s\n",
1171 obd2cli_tgt(imp->imp_obd));
1173 imp->imp_replayable = 0;
1176 /* if applies, adjust the imp->imp_msg_magic here
1177 * according to reply flags
1180 imp->imp_remote_handle =
1181 *lustre_msg_get_handle(request->rq_repmsg);
1183 /* Initial connects are allowed for clients with non-random
1184 * uuids when servers are in recovery. Simply signal the
1185 * servers replay is complete and wait in REPLAY_WAIT.
1187 if (msg_flags & MSG_CONNECT_RECOVERING) {
1188 CDEBUG(D_HA, "connect to %s during recovery\n",
1189 obd2cli_tgt(imp->imp_obd));
1190 import_set_state_nolock(imp, LUSTRE_IMP_REPLAY_LOCKS);
1191 spin_unlock(&imp->imp_lock);
1193 spin_unlock(&imp->imp_lock);
1194 ptlrpc_activate_import(imp, true);
1197 GOTO(finish, rc = 0);
1200 /* Determine what recovery state to move the import to. */
1201 if (MSG_CONNECT_RECONNECT & msg_flags) {
1202 memset(&old_hdl, 0, sizeof(old_hdl));
1203 if (!memcmp(&old_hdl, lustre_msg_get_handle(request->rq_repmsg),
1205 LCONSOLE_WARN("Reconnect to %s (at @%s) failed due "
1206 "bad handle %#llx\n",
1207 obd2cli_tgt(imp->imp_obd),
1208 imp->imp_connection->c_remote_uuid.uuid,
1209 imp->imp_dlm_handle.cookie);
1210 GOTO(out, rc = -ENOTCONN);
1213 if (memcmp(&imp->imp_remote_handle,
1214 lustre_msg_get_handle(request->rq_repmsg),
1215 sizeof(imp->imp_remote_handle))) {
1216 int level = msg_flags & MSG_CONNECT_RECOVERING ?
1219 /* Bug 16611/14775: if server handle have changed,
1220 * that means some sort of disconnection happened.
1221 * If the server is not in recovery, that also means it
1222 * already erased all of our state because of previous
1223 * eviction. If it is in recovery - we are safe to
1224 * participate since we can reestablish all of our state
1227 if ((MSG_CONNECT_RECOVERING & msg_flags)) {
1229 "%s@%s changed server handle from "
1231 " but is still in recovery\n",
1232 obd2cli_tgt(imp->imp_obd),
1233 imp->imp_connection->c_remote_uuid.uuid,
1234 imp->imp_remote_handle.cookie,
1235 lustre_msg_get_handle(
1236 request->rq_repmsg)->cookie);
1238 LCONSOLE_WARN("Evicted from %s (at %s) "
1239 "after server handle changed from "
1241 obd2cli_tgt(imp->imp_obd),
1242 imp->imp_connection->
1244 imp->imp_remote_handle.cookie,
1245 lustre_msg_get_handle(
1246 request->rq_repmsg)->cookie);
1249 imp->imp_remote_handle =
1250 *lustre_msg_get_handle(request->rq_repmsg);
1252 if (!(MSG_CONNECT_RECOVERING & msg_flags)) {
1253 import_set_state(imp, LUSTRE_IMP_EVICTED);
1254 GOTO(finish, rc = 0);
1257 CDEBUG(D_HA, "reconnected to %s@%s after partition\n",
1258 obd2cli_tgt(imp->imp_obd),
1259 imp->imp_connection->c_remote_uuid.uuid);
1262 if (imp->imp_invalid) {
1263 CDEBUG(D_HA, "%s: reconnected but import is invalid; "
1264 "marking evicted\n", imp->imp_obd->obd_name);
1265 import_set_state(imp, LUSTRE_IMP_EVICTED);
1266 } else if (MSG_CONNECT_RECOVERING & msg_flags) {
1267 CDEBUG(D_HA, "%s: reconnected to %s during replay\n",
1268 imp->imp_obd->obd_name,
1269 obd2cli_tgt(imp->imp_obd));
1271 spin_lock(&imp->imp_lock);
1272 imp->imp_resend_replay = 1;
1273 spin_unlock(&imp->imp_lock);
1275 import_set_state(imp, imp->imp_replay_state);
1277 import_set_state(imp, LUSTRE_IMP_RECOVER);
1279 } else if ((MSG_CONNECT_RECOVERING & msg_flags) && !imp->imp_invalid) {
1280 LASSERT(imp->imp_replayable);
1281 imp->imp_remote_handle =
1282 *lustre_msg_get_handle(request->rq_repmsg);
1283 imp->imp_last_replay_transno = 0;
1284 imp->imp_replay_cursor = &imp->imp_committed_list;
1285 import_set_state(imp, LUSTRE_IMP_REPLAY);
1286 } else if ((ocd->ocd_connect_flags & OBD_CONNECT_LIGHTWEIGHT) != 0 &&
1287 !imp->imp_invalid) {
1289 obd_import_event(imp->imp_obd, imp, IMP_EVENT_INVALIDATE);
1290 /* The below message is checked in recovery-small.sh test_106 */
1291 DEBUG_REQ(D_HA, request, "%s: lwp recover",
1292 imp->imp_obd->obd_name);
1293 imp->imp_remote_handle =
1294 *lustre_msg_get_handle(request->rq_repmsg);
1295 import_set_state(imp, LUSTRE_IMP_RECOVER);
1297 DEBUG_REQ(D_HA, request,
1298 "%s: evicting (reconnect/recover flags not set: %x)",
1299 imp->imp_obd->obd_name, msg_flags);
1300 imp->imp_remote_handle =
1301 *lustre_msg_get_handle(request->rq_repmsg);
1302 import_set_state(imp, LUSTRE_IMP_EVICTED);
1305 /* Sanity checks for a reconnected import. */
1306 if (!(imp->imp_replayable) != !(msg_flags & MSG_CONNECT_REPLAYABLE))
1307 CERROR("imp_replayable flag does not match server after reconnect. We should LBUG right here.\n");
1309 if (lustre_msg_get_last_committed(request->rq_repmsg) > 0 &&
1310 lustre_msg_get_last_committed(request->rq_repmsg) <
1311 aa->pcaa_peer_committed) {
1312 static bool printed;
1314 /* The below message is checked in recovery-small.sh test_54 */
1315 CERROR("%s: went back in time (transno %lld was previously committed, server now claims %lld)!\n",
1316 obd2cli_tgt(imp->imp_obd), aa->pcaa_peer_committed,
1317 lustre_msg_get_last_committed(request->rq_repmsg));
1319 CERROR("For further information, see http://doc.lustre.org/lustre_manual.xhtml#went_back_in_time\n");
1325 ptlrpc_prepare_replay(imp);
1326 rc = ptlrpc_import_recovery_state_machine(imp);
1327 if (rc == -ENOTCONN) {
1329 "evicted/aborted by %s@%s during recovery; invalidating and reconnecting\n",
1330 obd2cli_tgt(imp->imp_obd),
1331 imp->imp_connection->c_remote_uuid.uuid);
1332 ptlrpc_connect_import(imp);
1333 spin_lock(&imp->imp_lock);
1334 imp->imp_connected = 0;
1335 imp->imp_connect_tried = 1;
1336 spin_unlock(&imp->imp_lock);
1342 class_export_put(exp);
1344 spin_lock(&imp->imp_lock);
1345 imp->imp_connected = 0;
1346 imp->imp_connect_tried = 1;
1351 import_set_state_nolock(imp, LUSTRE_IMP_DISCON);
1352 if (rc == -EACCES) {
1354 * Give up trying to reconnect
1355 * EACCES means client has no permission for connection
1357 imp->imp_obd->obd_no_recov = 1;
1358 ptlrpc_deactivate_import_nolock(imp);
1360 } else if (rc == -EPROTO) {
1361 struct obd_connect_data *ocd;
1363 /* reply message might not be ready */
1364 if (request->rq_repmsg == NULL) {
1365 spin_unlock(&imp->imp_lock);
1369 ocd = req_capsule_server_get(&request->rq_pill,
1371 /* Servers are not supposed to refuse connections from
1372 * clients based on version, only connection feature
1373 * flags. We should never see this from llite, but it
1374 * may be useful for debugging in the future. */
1376 (ocd->ocd_connect_flags & OBD_CONNECT_VERSION) &&
1377 (ocd->ocd_version != LUSTRE_VERSION_CODE)) {
1378 LCONSOLE_ERROR_MSG(0x16a, "Server %s version "
1380 " refused connection from this client "
1381 "with an incompatible version (%s). "
1382 "Client must be recompiled\n",
1383 obd2cli_tgt(imp->imp_obd),
1384 OBD_OCD_VERSION_MAJOR(ocd->ocd_version),
1385 OBD_OCD_VERSION_MINOR(ocd->ocd_version),
1386 OBD_OCD_VERSION_PATCH(ocd->ocd_version),
1387 OBD_OCD_VERSION_FIX(ocd->ocd_version),
1388 LUSTRE_VERSION_STRING);
1389 ptlrpc_deactivate_import_nolock(imp);
1390 import_set_state_nolock(imp, LUSTRE_IMP_CLOSED);
1394 spin_unlock(&imp->imp_lock);
1397 obd_import_event(imp->imp_obd, imp, IMP_EVENT_INACTIVE);
1402 ptlrpc_maybe_ping_import_soon(imp);
1404 CDEBUG(D_HA, "recovery of %s on %s failed (%d)\n",
1405 obd2cli_tgt(imp->imp_obd),
1406 (char *)imp->imp_connection->c_remote_uuid.uuid, rc);
1408 spin_unlock(&imp->imp_lock);
1411 wake_up(&imp->imp_recovery_waitq);
1416 * interpret callback for "completed replay" RPCs.
1417 * \see signal_completed_replay
1419 static int completed_replay_interpret(const struct lu_env *env,
1420 struct ptlrpc_request *req,
1424 atomic_dec(&req->rq_import->imp_replay_inflight);
1425 if (req->rq_status == 0 && !req->rq_import->imp_vbr_failed) {
1426 ptlrpc_import_recovery_state_machine(req->rq_import);
1428 if (req->rq_import->imp_vbr_failed) {
1430 "%s: version recovery fails, reconnecting\n",
1431 req->rq_import->imp_obd->obd_name);
1433 CDEBUG(D_HA, "%s: LAST_REPLAY message error: %d, "
1435 req->rq_import->imp_obd->obd_name,
1438 ptlrpc_connect_import(req->rq_import);
1445 * Let server know that we have no requests to replay anymore.
1446 * Achieved by just sending a PING request
1448 static int signal_completed_replay(struct obd_import *imp)
1450 struct ptlrpc_request *req;
1453 if (unlikely(OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_FINISH_REPLAY)))
1456 if (!atomic_add_unless(&imp->imp_replay_inflight, 1, 1))
1459 req = ptlrpc_request_alloc_pack(imp, &RQF_OBD_PING, LUSTRE_OBD_VERSION,
1462 atomic_dec(&imp->imp_replay_inflight);
1466 ptlrpc_request_set_replen(req);
1467 req->rq_send_state = LUSTRE_IMP_REPLAY_WAIT;
1468 lustre_msg_add_flags(req->rq_reqmsg,
1469 MSG_LOCK_REPLAY_DONE | MSG_REQ_REPLAY_DONE);
1471 req->rq_timeout *= 3;
1472 req->rq_interpret_reply = completed_replay_interpret;
1474 ptlrpcd_add_req(req);
1479 * In kernel code all import invalidation happens in its own
1480 * separate thread, so that whatever application happened to encounter
1481 * a problem could still be killed or otherwise continue
1483 static int ptlrpc_invalidate_import_thread(void *data)
1485 struct obd_import *imp = data;
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 struct task_struct *task;
1537 deuuidify(obd2cli_tgt(imp->imp_obd), NULL,
1538 &target_start, &target_len);
1539 /* Don't care about MGC eviction */
1540 if (strcmp(imp->imp_obd->obd_type->typ_name,
1541 LUSTRE_MGC_NAME) != 0) {
1542 LCONSOLE_ERROR_MSG(0x167, "%s: This client was evicted "
1543 "by %.*s; in progress operations "
1544 "using this service will fail.\n",
1545 imp->imp_obd->obd_name, target_len,
1547 LASSERTF(!obd_lbug_on_eviction, "LBUG upon eviction\n");
1549 CDEBUG(D_HA, "evicted from %s@%s; invalidating\n",
1550 obd2cli_tgt(imp->imp_obd),
1551 imp->imp_connection->c_remote_uuid.uuid);
1552 /* reset vbr_failed flag upon eviction */
1553 spin_lock(&imp->imp_lock);
1554 imp->imp_vbr_failed = 0;
1555 spin_unlock(&imp->imp_lock);
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);
1567 CERROR("%s: can't start invalidate thread: rc = %d\n",
1568 imp->imp_obd->obd_name, 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(&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(timeout_t, 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 refcount_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 memset(&imp->imp_remote_handle, 0,
1784 sizeof(imp->imp_remote_handle));
1785 /* take our DISCONNECT into account */
1786 if (atomic_read(&imp->imp_reqs) > 1) {
1787 imp->imp_generation++;
1788 imp->imp_initiated_at = imp->imp_generation;
1789 import_set_state_nolock(imp, LUSTRE_IMP_NEW);
1790 ptlrpc_reset_reqs_generation(imp);
1793 /* do not expose transient IDLE state */
1794 import_set_state_nolock(imp, LUSTRE_IMP_IDLE);
1799 rc = ptlrpc_connect_import_locked(imp);
1801 ptlrpc_pinger_add_import(imp);
1803 spin_unlock(&imp->imp_lock);
1809 int ptlrpc_disconnect_and_idle_import(struct obd_import *imp)
1811 struct ptlrpc_request *req;
1814 if (imp->imp_obd->obd_force)
1817 if (ptlrpc_import_in_recovery(imp))
1820 spin_lock(&imp->imp_lock);
1821 if (imp->imp_state != LUSTRE_IMP_FULL) {
1822 spin_unlock(&imp->imp_lock);
1825 spin_unlock(&imp->imp_lock);
1827 req = ptlrpc_disconnect_prep_req(imp);
1829 RETURN(PTR_ERR(req));
1831 CDEBUG_LIMIT(imp->imp_idle_debug, "%s: disconnect after %llus idle\n",
1832 imp->imp_obd->obd_name,
1833 ktime_get_real_seconds() - imp->imp_last_reply_time);
1835 /* don't make noise at reconnection */
1836 spin_lock(&imp->imp_lock);
1837 imp->imp_was_idle = 1;
1838 spin_unlock(&imp->imp_lock);
1840 req->rq_interpret_reply = ptlrpc_disconnect_idle_interpret;
1841 ptlrpcd_add_req(req);
1845 EXPORT_SYMBOL(ptlrpc_disconnect_and_idle_import);
1847 void ptlrpc_cleanup_imp(struct obd_import *imp)
1851 spin_lock(&imp->imp_lock);
1853 import_set_state_nolock(imp, LUSTRE_IMP_CLOSED);
1854 imp->imp_generation++;
1855 ptlrpc_abort_inflight(imp);
1857 spin_unlock(&imp->imp_lock);
1862 /* Adaptive Timeout utils */
1864 /* Update at_current_timeout with the specified value (bounded by at_min and
1865 * at_max), as well as the AT history "bins".
1866 * - Bin into timeslices using AT_BINS bins.
1867 * - This gives us a max of the last at_history seconds without the storage,
1868 * but still smoothing out a return to normalcy from a slow response.
1869 * - (E.g. remember the maximum latency in each minute of the last 4 minutes.)
1871 timeout_t at_measured(struct adaptive_timeout *at, timeout_t timeout)
1873 timeout_t old_timeout = at->at_current_timeout;
1874 time64_t now = ktime_get_real_seconds();
1875 long binlimit = max_t(long, at_history / AT_BINS, 1);
1878 CDEBUG(D_OTHER, "add %u to %p time=%lld v=%u (%u %u %u %u)\n",
1879 timeout, at, now - at->at_binstart, at->at_current_timeout,
1880 at->at_hist[0], at->at_hist[1], at->at_hist[2], at->at_hist[3]);
1883 /* Negative timeouts and 0's don't count, because we never
1884 * want our timeout to drop to 0 or below, and because 0 could
1889 spin_lock(&at->at_lock);
1891 if (unlikely(at->at_binstart == 0)) {
1892 /* Special case to remove default from history */
1893 at->at_current_timeout = timeout;
1894 at->at_worst_timeout_ever = timeout;
1895 at->at_worst_timestamp = now;
1896 at->at_hist[0] = timeout;
1897 at->at_binstart = now;
1898 } else if (now - at->at_binstart < binlimit ) {
1900 at->at_hist[0] = max_t(timeout_t, timeout, at->at_hist[0]);
1901 at->at_current_timeout = max_t(timeout_t, timeout,
1902 at->at_current_timeout);
1905 timeout_t maxv = timeout;
1907 /* move bins over */
1908 shift = (u32)(now - at->at_binstart) / binlimit;
1910 for(i = AT_BINS - 1; i >= 0; i--) {
1912 at->at_hist[i] = at->at_hist[i - shift];
1913 maxv = max_t(timeout_t, maxv, at->at_hist[i]);
1918 at->at_hist[0] = timeout;
1919 at->at_current_timeout = maxv;
1920 at->at_binstart += shift * binlimit;
1923 if (at->at_current_timeout > at->at_worst_timeout_ever) {
1924 at->at_worst_timeout_ever = at->at_current_timeout;
1925 at->at_worst_timestamp = now;
1928 if (at->at_flags & AT_FLG_NOHIST)
1929 /* Only keep last reported val; keeping the rest of the history
1932 at->at_current_timeout = timeout;
1935 at->at_current_timeout = min_t(timeout_t,
1936 at->at_current_timeout, at_max);
1937 at->at_current_timeout = max_t(timeout_t, at->at_current_timeout,
1939 if (at->at_current_timeout != old_timeout)
1941 "AT %p change: old=%u new=%u delta=%d (val=%d) hist %u %u %u %u\n",
1942 at, old_timeout, at->at_current_timeout,
1943 at->at_current_timeout - old_timeout, timeout,
1944 at->at_hist[0], at->at_hist[1], at->at_hist[2],
1947 /* if we changed, report the old timeout value */
1948 old_timeout = (at->at_current_timeout != old_timeout) ? old_timeout : 0;
1950 spin_unlock(&at->at_lock);
1954 /* Find the imp_at index for a given portal; assign if space available */
1955 int import_at_get_index(struct obd_import *imp, int portal)
1957 struct imp_at *at = &imp->imp_at;
1960 for (i = 0; i < IMP_AT_MAX_PORTALS; i++) {
1961 if (at->iat_portal[i] == portal)
1963 if (at->iat_portal[i] == 0)
1968 /* Not found in list, add it under a lock */
1969 spin_lock(&imp->imp_lock);
1971 /* Check unused under lock */
1972 for (; i < IMP_AT_MAX_PORTALS; i++) {
1973 if (at->iat_portal[i] == portal)
1975 if (at->iat_portal[i] == 0)
1980 /* Not enough portals? */
1981 LASSERT(i < IMP_AT_MAX_PORTALS);
1983 at->iat_portal[i] = portal;
1985 spin_unlock(&imp->imp_lock);