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
59 * Helper function. Must be called under imp_lock.
61 static void __import_set_state(struct obd_import *imp,
62 enum lustre_imp_state state)
65 case LUSTRE_IMP_CLOSED:
67 case LUSTRE_IMP_DISCON:
68 case LUSTRE_IMP_CONNECTING:
70 case LUSTRE_IMP_REPLAY_WAIT:
71 imp->imp_replay_state = LUSTRE_IMP_REPLAY_LOCKS;
74 imp->imp_replay_state = LUSTRE_IMP_REPLAY;
76 imp->imp_state = state;
77 imp->imp_state_hist[imp->imp_state_hist_idx].ish_state = state;
78 imp->imp_state_hist[imp->imp_state_hist_idx].ish_time =
79 ktime_get_real_seconds();
80 imp->imp_state_hist_idx = (imp->imp_state_hist_idx + 1) %
84 /* A CLOSED import should remain so. */
85 #define IMPORT_SET_STATE_NOLOCK(imp, state) \
87 if (imp->imp_state != LUSTRE_IMP_CLOSED) { \
88 CDEBUG(D_HA, "%p %s: changing import state from %s to %s\n", \
89 imp, obd2cli_tgt(imp->imp_obd), \
90 ptlrpc_import_state_name(imp->imp_state), \
91 ptlrpc_import_state_name(state)); \
92 __import_set_state(imp, state); \
96 #define IMPORT_SET_STATE(imp, state) \
98 spin_lock(&imp->imp_lock); \
99 IMPORT_SET_STATE_NOLOCK(imp, state); \
100 spin_unlock(&imp->imp_lock); \
103 void ptlrpc_import_enter_resend(struct obd_import *imp)
105 IMPORT_SET_STATE(imp, LUSTRE_IMP_RECOVER);
107 EXPORT_SYMBOL(ptlrpc_import_enter_resend);
110 static int ptlrpc_connect_interpret(const struct lu_env *env,
111 struct ptlrpc_request *request,
113 int ptlrpc_import_recovery_state_machine(struct obd_import *imp);
115 /* Only this function is allowed to change the import state when it is
116 * CLOSED. I would rather refcount the import and free it after
117 * disconnection like we do with exports. To do that, the client_obd
118 * will need to save the peer info somewhere other than in the import,
120 int ptlrpc_init_import(struct obd_import *imp)
122 spin_lock(&imp->imp_lock);
124 imp->imp_generation++;
125 imp->imp_state = LUSTRE_IMP_NEW;
127 spin_unlock(&imp->imp_lock);
131 EXPORT_SYMBOL(ptlrpc_init_import);
133 #define UUID_STR "_UUID"
134 void deuuidify(char *uuid, const char *prefix, char **uuid_start, int *uuid_len)
136 *uuid_start = !prefix || strncmp(uuid, prefix, strlen(prefix))
137 ? uuid : uuid + strlen(prefix);
139 *uuid_len = strlen(*uuid_start);
141 if (*uuid_len < strlen(UUID_STR))
144 if (!strncmp(*uuid_start + *uuid_len - strlen(UUID_STR),
145 UUID_STR, strlen(UUID_STR)))
146 *uuid_len -= strlen(UUID_STR);
150 * Returns true if import was FULL, false if import was already not
152 * @imp - import to be disconnected
153 * @conn_cnt - connection count (epoch) of the request that timed out
154 * and caused the disconnection. In some cases, multiple
155 * inflight requests can fail to a single target (e.g. OST
156 * bulk requests) and if one has already caused a reconnection
157 * (increasing the import->conn_cnt) the older failure should
158 * not also cause a reconnection. If zero it forces a reconnect.
160 int ptlrpc_set_import_discon(struct obd_import *imp, __u32 conn_cnt)
164 spin_lock(&imp->imp_lock);
166 if (imp->imp_state == LUSTRE_IMP_FULL &&
167 (conn_cnt == 0 || conn_cnt == imp->imp_conn_cnt)) {
171 deuuidify(obd2cli_tgt(imp->imp_obd), NULL,
172 &target_start, &target_len);
174 if (imp->imp_replayable) {
175 LCONSOLE_WARN("%s: Connection to %.*s (at %s) was "
176 "lost; in progress operations using this "
177 "service will wait for recovery to complete\n",
178 imp->imp_obd->obd_name, target_len, target_start,
179 obd_import_nid2str(imp));
181 LCONSOLE_ERROR_MSG(0x166, "%s: Connection to "
182 "%.*s (at %s) was lost; in progress "
183 "operations using this service will fail\n",
184 imp->imp_obd->obd_name, target_len, target_start,
185 obd_import_nid2str(imp));
187 IMPORT_SET_STATE_NOLOCK(imp, LUSTRE_IMP_DISCON);
188 spin_unlock(&imp->imp_lock);
190 if (obd_dump_on_timeout)
191 libcfs_debug_dumplog();
193 obd_import_event(imp->imp_obd, imp, IMP_EVENT_DISCON);
196 spin_unlock(&imp->imp_lock);
197 CDEBUG(D_HA, "%s: import %p already %s (conn %u, was %u): %s\n",
198 imp->imp_client->cli_name, imp,
199 (imp->imp_state == LUSTRE_IMP_FULL &&
200 imp->imp_conn_cnt > conn_cnt) ?
201 "reconnected" : "not connected", imp->imp_conn_cnt,
202 conn_cnt, ptlrpc_import_state_name(imp->imp_state));
208 /* Must be called with imp_lock held! */
209 static void ptlrpc_deactivate_and_unlock_import(struct obd_import *imp)
212 assert_spin_locked(&imp->imp_lock);
214 CDEBUG(D_HA, "setting import %s INVALID\n", obd2cli_tgt(imp->imp_obd));
215 imp->imp_invalid = 1;
216 imp->imp_generation++;
217 spin_unlock(&imp->imp_lock);
219 ptlrpc_abort_inflight(imp);
220 obd_import_event(imp->imp_obd, imp, IMP_EVENT_INACTIVE);
226 * This acts as a barrier; all existing requests are rejected, and
227 * no new requests will be accepted until the import is valid again.
229 void ptlrpc_deactivate_import(struct obd_import *imp)
231 spin_lock(&imp->imp_lock);
232 ptlrpc_deactivate_and_unlock_import(imp);
234 EXPORT_SYMBOL(ptlrpc_deactivate_import);
236 static time64_t ptlrpc_inflight_deadline(struct ptlrpc_request *req,
241 if (!(((req->rq_phase == RQ_PHASE_RPC) && !req->rq_waiting) ||
242 (req->rq_phase == RQ_PHASE_BULK) ||
243 (req->rq_phase == RQ_PHASE_NEW)))
246 if (req->rq_timedout)
249 if (req->rq_phase == RQ_PHASE_NEW)
252 dl = req->rq_deadline;
260 static time64_t ptlrpc_inflight_timeout(struct obd_import *imp)
262 time64_t now = ktime_get_real_seconds();
263 struct list_head *tmp, *n;
264 struct ptlrpc_request *req;
265 time64_t timeout = 0;
267 spin_lock(&imp->imp_lock);
268 list_for_each_safe(tmp, n, &imp->imp_sending_list) {
269 req = list_entry(tmp, struct ptlrpc_request, rq_list);
270 timeout = max(ptlrpc_inflight_deadline(req, now), timeout);
272 spin_unlock(&imp->imp_lock);
277 * This function will invalidate the import, if necessary, then block
278 * for all the RPC completions, and finally notify the obd to
279 * invalidate its state (ie cancel locks, clear pending requests,
282 void ptlrpc_invalidate_import(struct obd_import *imp)
284 struct list_head *tmp, *n;
285 struct ptlrpc_request *req;
286 struct l_wait_info lwi;
290 atomic_inc(&imp->imp_inval_count);
292 if (!imp->imp_invalid || imp->imp_obd->obd_no_recov)
293 ptlrpc_deactivate_import(imp);
295 if (OBD_FAIL_PRECHECK(OBD_FAIL_PTLRPC_CONNECT_RACE)) {
296 OBD_RACE(OBD_FAIL_PTLRPC_CONNECT_RACE);
297 msleep(10 * MSEC_PER_SEC);
299 CFS_FAIL_TIMEOUT(OBD_FAIL_MGS_CONNECT_NET, 3 * cfs_fail_val / 2);
300 LASSERT(imp->imp_invalid);
302 /* Wait forever until inflight == 0. We really can't do it another
303 * way because in some cases we need to wait for very long reply
304 * unlink. We can't do anything before that because there is really
305 * no guarantee that some rdma transfer is not in progress right now. */
307 long timeout_jiffies;
309 /* Calculate max timeout for waiting on rpcs to error
310 * out. Use obd_timeout if calculated value is smaller
313 if (!OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_REPL_UNLINK)) {
314 timeout = ptlrpc_inflight_timeout(imp);
315 timeout += div_u64(timeout, 3);
318 timeout = obd_timeout;
320 /* decrease the interval to increase race condition */
324 CDEBUG(D_RPCTRACE, "Sleeping %llds for inflight to error out\n",
327 /* Wait for all requests to error out and call completion
328 * callbacks. Cap it at obd_timeout -- these should all
329 * have been locally cancelled by ptlrpc_abort_inflight.
331 timeout_jiffies = max_t(long, cfs_time_seconds(timeout), 1);
332 lwi = LWI_TIMEOUT_INTERVAL(timeout_jiffies,
333 (timeout > 1) ? cfs_time_seconds(1) :
334 cfs_time_seconds(1) / 2,
336 rc = l_wait_event(imp->imp_recovery_waitq,
337 (atomic_read(&imp->imp_inflight) == 0),
340 const char *cli_tgt = obd2cli_tgt(imp->imp_obd);
342 CERROR("%s: rc = %d waiting for callback (%d != 0)\n",
343 cli_tgt, rc, atomic_read(&imp->imp_inflight));
345 spin_lock(&imp->imp_lock);
346 if (atomic_read(&imp->imp_inflight) == 0) {
347 int count = atomic_read(&imp->imp_unregistering);
349 /* We know that "unregistering" rpcs only can
350 * survive in sending or delaying lists (they
351 * maybe waiting for long reply unlink in
352 * sluggish nets). Let's check this. If there
353 * is no inflight and unregistering != 0, this
355 LASSERTF(count == 0, "Some RPCs are still "
356 "unregistering: %d\n", count);
358 /* Let's save one loop as soon as inflight have
359 * dropped to zero. No new inflights possible at
363 list_for_each_safe(tmp, n,
364 &imp->imp_sending_list) {
365 req = list_entry(tmp,
366 struct ptlrpc_request,
368 DEBUG_REQ(D_ERROR, req,
369 "still on sending list");
371 list_for_each_safe(tmp, n,
372 &imp->imp_delayed_list) {
373 req = list_entry(tmp,
374 struct ptlrpc_request,
376 DEBUG_REQ(D_ERROR, req,
377 "still on delayed list");
380 CERROR("%s: Unregistering RPCs found (%d). "
381 "Network is sluggish? Waiting them "
382 "to error out.\n", cli_tgt,
383 atomic_read(&imp->imp_unregistering));
385 spin_unlock(&imp->imp_lock);
390 * Let's additionally check that no new rpcs added to import in
391 * "invalidate" state.
393 LASSERT(atomic_read(&imp->imp_inflight) == 0);
394 obd_import_event(imp->imp_obd, imp, IMP_EVENT_INVALIDATE);
395 sptlrpc_import_flush_all_ctx(imp);
397 atomic_dec(&imp->imp_inval_count);
398 wake_up_all(&imp->imp_recovery_waitq);
400 EXPORT_SYMBOL(ptlrpc_invalidate_import);
402 /* unset imp_invalid */
403 void ptlrpc_activate_import(struct obd_import *imp)
405 struct obd_device *obd = imp->imp_obd;
407 spin_lock(&imp->imp_lock);
408 if (imp->imp_deactive != 0) {
409 spin_unlock(&imp->imp_lock);
413 imp->imp_invalid = 0;
414 spin_unlock(&imp->imp_lock);
415 obd_import_event(obd, imp, IMP_EVENT_ACTIVE);
417 EXPORT_SYMBOL(ptlrpc_activate_import);
419 void ptlrpc_pinger_force(struct obd_import *imp)
421 CDEBUG(D_HA, "%s: waking up pinger s:%s\n", obd2cli_tgt(imp->imp_obd),
422 ptlrpc_import_state_name(imp->imp_state));
424 spin_lock(&imp->imp_lock);
425 imp->imp_force_verify = 1;
426 spin_unlock(&imp->imp_lock);
428 if (imp->imp_state != LUSTRE_IMP_CONNECTING)
429 ptlrpc_pinger_wake_up();
431 EXPORT_SYMBOL(ptlrpc_pinger_force);
433 void ptlrpc_fail_import(struct obd_import *imp, __u32 conn_cnt)
437 LASSERT(!imp->imp_dlm_fake);
439 if (ptlrpc_set_import_discon(imp, conn_cnt)) {
440 if (!imp->imp_replayable) {
441 CDEBUG(D_HA, "import %s@%s for %s not replayable, "
442 "auto-deactivating\n",
443 obd2cli_tgt(imp->imp_obd),
444 imp->imp_connection->c_remote_uuid.uuid,
445 imp->imp_obd->obd_name);
446 ptlrpc_deactivate_import(imp);
449 ptlrpc_pinger_force(imp);
454 int ptlrpc_reconnect_import(struct obd_import *imp)
457 long timeout_jiffies = cfs_time_seconds(obd_timeout);
458 struct l_wait_info lwi;
461 ptlrpc_pinger_force(imp);
463 CDEBUG(D_HA, "%s: recovery started, waiting %u seconds\n",
464 obd2cli_tgt(imp->imp_obd), obd_timeout);
466 lwi = LWI_TIMEOUT(timeout_jiffies, NULL, NULL);
467 rc = l_wait_event(imp->imp_recovery_waitq,
468 !ptlrpc_import_in_recovery(imp), &lwi);
469 CDEBUG(D_HA, "%s: recovery finished s:%s\n", obd2cli_tgt(imp->imp_obd),
470 ptlrpc_import_state_name(imp->imp_state));
473 ptlrpc_set_import_discon(imp, 0);
474 /* Force a new connect attempt */
475 ptlrpc_invalidate_import(imp);
476 /* Do a fresh connect next time by zeroing the handle */
477 ptlrpc_disconnect_import(imp, 1);
478 /* Wait for all invalidate calls to finish */
479 if (atomic_read(&imp->imp_inval_count) > 0) {
480 struct l_wait_info lwi = LWI_INTR(LWI_ON_SIGNAL_NOOP, NULL);
483 rc = l_wait_event(imp->imp_recovery_waitq,
484 (atomic_read(&imp->imp_inval_count) == 0),
487 CERROR("Interrupted, inval=%d\n",
488 atomic_read(&imp->imp_inval_count));
491 /* Allow reconnect attempts */
492 imp->imp_obd->obd_no_recov = 0;
493 /* Remove 'invalid' flag */
494 ptlrpc_activate_import(imp);
495 /* Attempt a new connect */
496 ptlrpc_recover_import(imp, NULL, 0);
500 EXPORT_SYMBOL(ptlrpc_reconnect_import);
503 * Connection on import \a imp is changed to another one (if more than one is
504 * present). We typically chose connection that we have not tried to connect to
507 static int import_select_connection(struct obd_import *imp)
509 struct obd_import_conn *imp_conn = NULL, *conn;
510 struct obd_export *dlmexp;
512 int target_len, tried_all = 1;
515 spin_lock(&imp->imp_lock);
517 if (list_empty(&imp->imp_conn_list)) {
518 CERROR("%s: no connections available\n",
519 imp->imp_obd->obd_name);
520 spin_unlock(&imp->imp_lock);
524 list_for_each_entry(conn, &imp->imp_conn_list, oic_item) {
525 CDEBUG(D_HA, "%s: connect to NID %s last attempt %lld\n",
526 imp->imp_obd->obd_name,
527 libcfs_nid2str(conn->oic_conn->c_peer.nid),
528 conn->oic_last_attempt);
530 /* If we have not tried this connection since
531 the last successful attempt, go with this one */
532 if ((conn->oic_last_attempt == 0) ||
533 conn->oic_last_attempt <= imp->imp_last_success_conn) {
539 /* If all of the connections have already been tried
540 since the last successful connection; just choose the
541 least recently used */
544 else if (imp_conn->oic_last_attempt > conn->oic_last_attempt)
548 /* if not found, simply choose the current one */
549 if (!imp_conn || imp->imp_force_reconnect) {
550 LASSERT(imp->imp_conn_current);
551 imp_conn = imp->imp_conn_current;
554 LASSERT(imp_conn->oic_conn);
556 /* If we've tried everything, and we're back to the beginning of the
557 list, increase our timeout and try again. It will be reset when
558 we do finally connect. (FIXME: really we should wait for all network
559 state associated with the last connection attempt to drain before
560 trying to reconnect on it.) */
561 if (tried_all && (imp->imp_conn_list.next == &imp_conn->oic_item)) {
562 struct adaptive_timeout *at = &imp->imp_at.iat_net_latency;
563 if (at_get(at) < CONNECTION_SWITCH_MAX) {
564 at_measured(at, at_get(at) + CONNECTION_SWITCH_INC);
565 if (at_get(at) > CONNECTION_SWITCH_MAX)
566 at_reset(at, CONNECTION_SWITCH_MAX);
568 LASSERT(imp_conn->oic_last_attempt);
569 CDEBUG(D_HA, "%s: tried all connections, increasing latency "
570 "to %ds\n", imp->imp_obd->obd_name, at_get(at));
573 imp_conn->oic_last_attempt = ktime_get_seconds();
575 /* switch connection, don't mind if it's same as the current one */
576 if (imp->imp_connection)
577 ptlrpc_connection_put(imp->imp_connection);
578 imp->imp_connection = ptlrpc_connection_addref(imp_conn->oic_conn);
580 dlmexp = class_conn2export(&imp->imp_dlm_handle);
581 LASSERT(dlmexp != NULL);
582 if (dlmexp->exp_connection)
583 ptlrpc_connection_put(dlmexp->exp_connection);
584 dlmexp->exp_connection = ptlrpc_connection_addref(imp_conn->oic_conn);
585 class_export_put(dlmexp);
587 if (imp->imp_conn_current != imp_conn) {
588 if (imp->imp_conn_current) {
589 deuuidify(obd2cli_tgt(imp->imp_obd), NULL,
590 &target_start, &target_len);
592 CDEBUG(D_HA, "%s: Connection changing to"
594 imp->imp_obd->obd_name,
595 target_len, target_start,
596 libcfs_nid2str(imp_conn->oic_conn->c_peer.nid));
599 imp->imp_conn_current = imp_conn;
602 CDEBUG(D_HA, "%s: import %p using connection %s/%s\n",
603 imp->imp_obd->obd_name, imp, imp_conn->oic_uuid.uuid,
604 libcfs_nid2str(imp_conn->oic_conn->c_peer.nid));
606 spin_unlock(&imp->imp_lock);
612 * must be called under imp_lock
614 static int ptlrpc_first_transno(struct obd_import *imp, __u64 *transno)
616 struct ptlrpc_request *req;
617 struct list_head *tmp;
619 /* The requests in committed_list always have smaller transnos than
620 * the requests in replay_list */
621 if (!list_empty(&imp->imp_committed_list)) {
622 tmp = imp->imp_committed_list.next;
623 req = list_entry(tmp, struct ptlrpc_request, rq_replay_list);
624 *transno = req->rq_transno;
625 if (req->rq_transno == 0) {
626 DEBUG_REQ(D_ERROR, req, "zero transno in committed_list");
631 if (!list_empty(&imp->imp_replay_list)) {
632 tmp = imp->imp_replay_list.next;
633 req = list_entry(tmp, struct ptlrpc_request, rq_replay_list);
634 *transno = req->rq_transno;
635 if (req->rq_transno == 0) {
636 DEBUG_REQ(D_ERROR, req, "zero transno in replay_list");
645 * Attempt to (re)connect import \a imp. This includes all preparations,
646 * initializing CONNECT RPC request and passing it to ptlrpcd for
648 * Returns 0 on success or error code.
650 int ptlrpc_connect_import(struct obd_import *imp)
652 struct obd_device *obd = imp->imp_obd;
653 int initial_connect = 0;
655 __u64 committed_before_reconnect = 0;
656 struct ptlrpc_request *request;
657 char *bufs[] = { NULL,
658 obd2cli_tgt(imp->imp_obd),
660 (char *)&imp->imp_dlm_handle,
661 (char *)&imp->imp_connect_data };
662 struct ptlrpc_connect_async_args *aa;
666 spin_lock(&imp->imp_lock);
667 if (imp->imp_state == LUSTRE_IMP_CLOSED) {
668 spin_unlock(&imp->imp_lock);
669 CERROR("can't connect to a closed import\n");
671 } else if (imp->imp_state == LUSTRE_IMP_FULL) {
672 spin_unlock(&imp->imp_lock);
673 CERROR("already connected\n");
675 } else if (imp->imp_state == LUSTRE_IMP_CONNECTING ||
676 imp->imp_state == LUSTRE_IMP_EVICTED ||
677 imp->imp_connected) {
678 spin_unlock(&imp->imp_lock);
679 CERROR("already connecting\n");
683 IMPORT_SET_STATE_NOLOCK(imp, LUSTRE_IMP_CONNECTING);
686 imp->imp_resend_replay = 0;
688 if (!lustre_handle_is_used(&imp->imp_remote_handle))
691 committed_before_reconnect = imp->imp_peer_committed_transno;
693 set_transno = ptlrpc_first_transno(imp,
694 &imp->imp_connect_data.ocd_transno);
695 spin_unlock(&imp->imp_lock);
697 rc = import_select_connection(imp);
701 rc = sptlrpc_import_sec_adapt(imp, NULL, NULL);
705 /* Reset connect flags to the originally requested flags, in case
706 * the server is updated on-the-fly we will get the new features. */
707 imp->imp_connect_data.ocd_connect_flags = imp->imp_connect_flags_orig;
708 imp->imp_connect_data.ocd_connect_flags2 = imp->imp_connect_flags2_orig;
709 /* Reset ocd_version each time so the server knows the exact versions */
710 imp->imp_connect_data.ocd_version = LUSTRE_VERSION_CODE;
711 imp->imp_msghdr_flags &= ~MSGHDR_AT_SUPPORT;
712 imp->imp_msghdr_flags &= ~MSGHDR_CKSUM_INCOMPAT18;
714 rc = obd_reconnect(NULL, imp->imp_obd->obd_self_export, obd,
715 &obd->obd_uuid, &imp->imp_connect_data, NULL);
719 request = ptlrpc_request_alloc(imp, &RQF_MDS_CONNECT);
721 GOTO(out, rc = -ENOMEM);
723 rc = ptlrpc_request_bufs_pack(request, LUSTRE_OBD_VERSION,
724 imp->imp_connect_op, bufs, NULL);
726 ptlrpc_request_free(request);
730 /* Report the rpc service time to the server so that it knows how long
731 * to wait for clients to join recovery */
732 lustre_msg_set_service_time(request->rq_reqmsg,
733 at_timeout2est(request->rq_timeout));
735 /* The amount of time we give the server to process the connect req.
736 * import_select_connection will increase the net latency on
737 * repeated reconnect attempts to cover slow networks.
738 * We override/ignore the server rpc completion estimate here,
739 * which may be large if this is a reconnect attempt */
740 request->rq_timeout = INITIAL_CONNECT_TIMEOUT;
741 lustre_msg_set_timeout(request->rq_reqmsg, request->rq_timeout);
743 request->rq_no_resend = request->rq_no_delay = 1;
744 request->rq_send_state = LUSTRE_IMP_CONNECTING;
745 /* Allow a slightly larger reply for future growth compatibility */
746 req_capsule_set_size(&request->rq_pill, &RMF_CONNECT_DATA, RCL_SERVER,
747 sizeof(struct obd_connect_data)+16*sizeof(__u64));
748 ptlrpc_request_set_replen(request);
749 request->rq_interpret_reply = ptlrpc_connect_interpret;
751 CLASSERT(sizeof(*aa) <= sizeof(request->rq_async_args));
752 aa = ptlrpc_req_async_args(request);
753 memset(aa, 0, sizeof *aa);
755 aa->pcaa_peer_committed = committed_before_reconnect;
756 aa->pcaa_initial_connect = initial_connect;
758 if (aa->pcaa_initial_connect) {
759 spin_lock(&imp->imp_lock);
760 imp->imp_replayable = 1;
761 spin_unlock(&imp->imp_lock);
762 lustre_msg_add_op_flags(request->rq_reqmsg,
763 MSG_CONNECT_INITIAL);
767 lustre_msg_add_op_flags(request->rq_reqmsg,
768 MSG_CONNECT_TRANSNO);
770 DEBUG_REQ(D_RPCTRACE, request, "(re)connect request (timeout %ld)",
771 request->rq_timeout);
772 ptlrpcd_add_req(request);
776 IMPORT_SET_STATE(imp, LUSTRE_IMP_DISCON);
780 EXPORT_SYMBOL(ptlrpc_connect_import);
782 static void ptlrpc_maybe_ping_import_soon(struct obd_import *imp)
786 spin_lock(&imp->imp_lock);
787 force_verify = imp->imp_force_verify != 0;
788 spin_unlock(&imp->imp_lock);
791 ptlrpc_pinger_wake_up();
794 static int ptlrpc_busy_reconnect(int rc)
796 return (rc == -EBUSY) || (rc == -EAGAIN);
799 static int ptlrpc_connect_set_flags(struct obd_import *imp,
800 struct obd_connect_data *ocd,
801 __u64 old_connect_flags,
802 struct obd_export *exp, int init_connect)
805 struct client_obd *cli = &imp->imp_obd->u.cli;
807 spin_lock(&imp->imp_lock);
808 list_del(&imp->imp_conn_current->oic_item);
809 list_add(&imp->imp_conn_current->oic_item,
810 &imp->imp_conn_list);
811 imp->imp_last_success_conn =
812 imp->imp_conn_current->oic_last_attempt;
814 spin_unlock(&imp->imp_lock);
816 if (!warned && (ocd->ocd_connect_flags & OBD_CONNECT_VERSION) &&
817 (ocd->ocd_version > LUSTRE_VERSION_CODE +
818 LUSTRE_VERSION_OFFSET_WARN ||
819 ocd->ocd_version < LUSTRE_VERSION_CODE -
820 LUSTRE_VERSION_OFFSET_WARN)) {
821 /* Sigh, some compilers do not like #ifdef in the middle
822 of macro arguments */
823 const char *older = "older than client. "
824 "Consider upgrading server";
825 const char *newer = "newer than client. "
826 "Consider upgrading client";
828 LCONSOLE_WARN("Server %s version (%d.%d.%d.%d) "
830 obd2cli_tgt(imp->imp_obd),
831 OBD_OCD_VERSION_MAJOR(ocd->ocd_version),
832 OBD_OCD_VERSION_MINOR(ocd->ocd_version),
833 OBD_OCD_VERSION_PATCH(ocd->ocd_version),
834 OBD_OCD_VERSION_FIX(ocd->ocd_version),
835 ocd->ocd_version > LUSTRE_VERSION_CODE ?
836 newer : older, LUSTRE_VERSION_STRING);
840 if (ocd->ocd_connect_flags & OBD_CONNECT_CKSUM) {
841 /* We sent to the server ocd_cksum_types with bits set
842 * for algorithms we understand. The server masked off
843 * the checksum types it doesn't support */
844 if ((ocd->ocd_cksum_types &
845 obd_cksum_types_supported_client()) == 0) {
846 LCONSOLE_ERROR("The negotiation of the checksum "
847 "alogrithm to use with server %s "
849 obd2cli_tgt(imp->imp_obd),
850 ocd->ocd_cksum_types,
851 obd_cksum_types_supported_client());
854 cli->cl_supp_cksum_types = ocd->ocd_cksum_types;
857 /* The server does not support OBD_CONNECT_CKSUM.
858 * Enforce ADLER for backward compatibility*/
859 cli->cl_supp_cksum_types = OBD_CKSUM_ADLER;
861 cli->cl_cksum_type = obd_cksum_type_select(imp->imp_obd->obd_name,
862 cli->cl_supp_cksum_types);
864 if (ocd->ocd_connect_flags & OBD_CONNECT_BRW_SIZE)
865 cli->cl_max_pages_per_rpc =
866 min(ocd->ocd_brw_size >> PAGE_SHIFT,
867 cli->cl_max_pages_per_rpc);
868 else if (imp->imp_connect_op == MDS_CONNECT ||
869 imp->imp_connect_op == MGS_CONNECT)
870 cli->cl_max_pages_per_rpc = 1;
872 LASSERT((cli->cl_max_pages_per_rpc <= PTLRPC_MAX_BRW_PAGES) &&
873 (cli->cl_max_pages_per_rpc > 0));
875 client_adjust_max_dirty(cli);
877 /* Update client max modify RPCs in flight with value returned
879 if (ocd->ocd_connect_flags & OBD_CONNECT_MULTIMODRPCS)
880 cli->cl_max_mod_rpcs_in_flight = min(
881 cli->cl_max_mod_rpcs_in_flight,
882 ocd->ocd_maxmodrpcs);
884 cli->cl_max_mod_rpcs_in_flight = 1;
886 /* Reset ns_connect_flags only for initial connect. It might be
887 * changed in while using FS and if we reset it in reconnect
888 * this leads to losing user settings done before such as
889 * disable lru_resize, etc. */
890 if (old_connect_flags != exp_connect_flags(exp) || init_connect) {
891 CDEBUG(D_HA, "%s: Resetting ns_connect_flags to server "
892 "flags: %#llx\n", imp->imp_obd->obd_name,
893 ocd->ocd_connect_flags);
894 imp->imp_obd->obd_namespace->ns_connect_flags =
895 ocd->ocd_connect_flags;
896 imp->imp_obd->obd_namespace->ns_orig_connect_flags =
897 ocd->ocd_connect_flags;
900 if (ocd->ocd_connect_flags & OBD_CONNECT_AT)
901 /* We need a per-message support flag, because
902 * a. we don't know if the incoming connect reply
903 * supports AT or not (in reply_in_callback)
904 * until we unpack it.
905 * b. failovered server means export and flags are gone
906 * (in ptlrpc_send_reply).
907 * Can only be set when we know AT is supported at
909 imp->imp_msghdr_flags |= MSGHDR_AT_SUPPORT;
911 imp->imp_msghdr_flags &= ~MSGHDR_AT_SUPPORT;
913 imp->imp_msghdr_flags |= MSGHDR_CKSUM_INCOMPAT18;
919 * Add all replay requests back to unreplied list before start replay,
920 * so that we can make sure the known replied XID is always increased
921 * only even if when replaying requests.
923 static void ptlrpc_prepare_replay(struct obd_import *imp)
925 struct ptlrpc_request *req;
927 if (imp->imp_state != LUSTRE_IMP_REPLAY ||
928 imp->imp_resend_replay)
931 /* If the server was restart during repaly, the requests may
932 * have been added to the unreplied list in former replay. */
933 spin_lock(&imp->imp_lock);
935 list_for_each_entry(req, &imp->imp_committed_list, rq_replay_list) {
936 if (list_empty(&req->rq_unreplied_list))
937 ptlrpc_add_unreplied(req);
940 list_for_each_entry(req, &imp->imp_replay_list, rq_replay_list) {
941 if (list_empty(&req->rq_unreplied_list))
942 ptlrpc_add_unreplied(req);
945 imp->imp_known_replied_xid = ptlrpc_known_replied_xid(imp);
946 spin_unlock(&imp->imp_lock);
950 * interpret_reply callback for connect RPCs.
951 * Looks into returned status of connect operation and decides
952 * what to do with the import - i.e enter recovery, promote it to
953 * full state for normal operations of disconnect it due to an error.
955 static int ptlrpc_connect_interpret(const struct lu_env *env,
956 struct ptlrpc_request *request,
959 struct ptlrpc_connect_async_args *aa = data;
960 struct obd_import *imp = request->rq_import;
961 struct lustre_handle old_hdl;
962 __u64 old_connect_flags;
964 struct obd_connect_data *ocd;
965 struct obd_export *exp = NULL;
969 spin_lock(&imp->imp_lock);
970 if (imp->imp_state == LUSTRE_IMP_CLOSED) {
971 imp->imp_connect_tried = 1;
972 spin_unlock(&imp->imp_lock);
977 struct ptlrpc_request *free_req;
978 struct ptlrpc_request *tmp;
980 /* abort all delayed requests initiated connection */
981 list_for_each_entry_safe(free_req, tmp, &imp->imp_delayed_list,
983 spin_lock(&free_req->rq_lock);
984 if (free_req->rq_no_resend) {
985 free_req->rq_err = 1;
986 free_req->rq_status = -EIO;
987 ptlrpc_client_wake_req(free_req);
989 spin_unlock(&free_req->rq_lock);
992 /* if this reconnect to busy export - not need select new target
994 imp->imp_force_reconnect = ptlrpc_busy_reconnect(rc);
995 spin_unlock(&imp->imp_lock);
996 ptlrpc_maybe_ping_import_soon(imp);
1000 /* LU-7558: indicate that we are interpretting connect reply,
1001 * pltrpc_connect_import() will not try to reconnect until
1002 * interpret will finish. */
1003 imp->imp_connected = 1;
1004 spin_unlock(&imp->imp_lock);
1006 LASSERT(imp->imp_conn_current);
1008 msg_flags = lustre_msg_get_op_flags(request->rq_repmsg);
1010 ret = req_capsule_get_size(&request->rq_pill, &RMF_CONNECT_DATA,
1012 /* server replied obd_connect_data is always bigger */
1013 ocd = req_capsule_server_sized_get(&request->rq_pill,
1014 &RMF_CONNECT_DATA, ret);
1017 CERROR("%s: no connect data from server\n",
1018 imp->imp_obd->obd_name);
1023 spin_lock(&imp->imp_lock);
1025 /* All imports are pingable */
1026 imp->imp_pingable = 1;
1027 imp->imp_force_reconnect = 0;
1028 imp->imp_force_verify = 0;
1030 imp->imp_connect_data = *ocd;
1032 CDEBUG(D_HA, "%s: connect to target with instance %u\n",
1033 imp->imp_obd->obd_name, ocd->ocd_instance);
1034 exp = class_conn2export(&imp->imp_dlm_handle);
1036 spin_unlock(&imp->imp_lock);
1039 /* This could happen if export is cleaned during the
1041 CERROR("%s: missing export after connect\n",
1042 imp->imp_obd->obd_name);
1043 GOTO(out, rc = -ENODEV);
1046 /* check that server granted subset of flags we asked for. */
1047 if ((ocd->ocd_connect_flags & imp->imp_connect_flags_orig) !=
1048 ocd->ocd_connect_flags) {
1049 CERROR("%s: Server didn't grant requested subset of flags: "
1050 "asked=%#llx granted=%#llx\n",
1051 imp->imp_obd->obd_name, imp->imp_connect_flags_orig,
1052 ocd->ocd_connect_flags);
1053 GOTO(out, rc = -EPROTO);
1056 if ((ocd->ocd_connect_flags2 & imp->imp_connect_flags2_orig) !=
1057 ocd->ocd_connect_flags2) {
1058 CERROR("%s: Server didn't grant requested subset of flags2: "
1059 "asked=%#llx granted=%#llx\n",
1060 imp->imp_obd->obd_name, imp->imp_connect_flags2_orig,
1061 ocd->ocd_connect_flags2);
1062 GOTO(out, rc = -EPROTO);
1065 if (!(imp->imp_connect_flags_orig & OBD_CONNECT_LIGHTWEIGHT) &&
1066 (imp->imp_connect_flags_orig & OBD_CONNECT_MDS_MDS) &&
1067 (imp->imp_connect_flags_orig & OBD_CONNECT_FID) &&
1068 (ocd->ocd_connect_flags & OBD_CONNECT_VERSION)) {
1069 __u32 major = OBD_OCD_VERSION_MAJOR(ocd->ocd_version);
1070 __u32 minor = OBD_OCD_VERSION_MINOR(ocd->ocd_version);
1071 __u32 patch = OBD_OCD_VERSION_PATCH(ocd->ocd_version);
1073 /* We do not support the MDT-MDT interoperations with
1074 * different version MDT because of protocol changes. */
1075 if (unlikely(major != LUSTRE_MAJOR ||
1076 minor != LUSTRE_MINOR ||
1077 abs(patch - LUSTRE_PATCH) > 3)) {
1078 LCONSOLE_WARN("%s: import %p (%u.%u.%u.%u) tried the "
1079 "connection to different version MDT "
1080 "(%d.%d.%d.%d) %s\n",
1081 imp->imp_obd->obd_name, imp, LUSTRE_MAJOR,
1082 LUSTRE_MINOR, LUSTRE_PATCH, LUSTRE_FIX,
1083 major, minor, patch,
1084 OBD_OCD_VERSION_FIX(ocd->ocd_version),
1085 imp->imp_connection->c_remote_uuid.uuid);
1087 GOTO(out, rc = -EPROTO);
1091 old_connect_flags = exp_connect_flags(exp);
1092 exp->exp_connect_data = *ocd;
1093 imp->imp_obd->obd_self_export->exp_connect_data = *ocd;
1095 /* The net statistics after (re-)connect is not valid anymore,
1096 * because may reflect other routing, etc. */
1097 at_reinit(&imp->imp_at.iat_net_latency, 0, 0);
1098 ptlrpc_at_adj_net_latency(request,
1099 lustre_msg_get_service_time(request->rq_repmsg));
1101 /* Import flags should be updated before waking import at FULL state */
1102 rc = ptlrpc_connect_set_flags(imp, ocd, old_connect_flags, exp,
1103 aa->pcaa_initial_connect);
1104 class_export_put(exp);
1110 obd_import_event(imp->imp_obd, imp, IMP_EVENT_OCD);
1112 if (aa->pcaa_initial_connect) {
1113 spin_lock(&imp->imp_lock);
1114 if (msg_flags & MSG_CONNECT_REPLAYABLE) {
1115 imp->imp_replayable = 1;
1116 spin_unlock(&imp->imp_lock);
1117 CDEBUG(D_HA, "connected to replayable target: %s\n",
1118 obd2cli_tgt(imp->imp_obd));
1120 imp->imp_replayable = 0;
1121 spin_unlock(&imp->imp_lock);
1124 /* if applies, adjust the imp->imp_msg_magic here
1125 * according to reply flags */
1127 imp->imp_remote_handle =
1128 *lustre_msg_get_handle(request->rq_repmsg);
1130 /* Initial connects are allowed for clients with non-random
1131 * uuids when servers are in recovery. Simply signal the
1132 * servers replay is complete and wait in REPLAY_WAIT. */
1133 if (msg_flags & MSG_CONNECT_RECOVERING) {
1134 CDEBUG(D_HA, "connect to %s during recovery\n",
1135 obd2cli_tgt(imp->imp_obd));
1136 IMPORT_SET_STATE(imp, LUSTRE_IMP_REPLAY_LOCKS);
1138 IMPORT_SET_STATE(imp, LUSTRE_IMP_FULL);
1139 ptlrpc_activate_import(imp);
1142 GOTO(finish, rc = 0);
1145 /* Determine what recovery state to move the import to. */
1146 if (MSG_CONNECT_RECONNECT & msg_flags) {
1147 memset(&old_hdl, 0, sizeof(old_hdl));
1148 if (!memcmp(&old_hdl, lustre_msg_get_handle(request->rq_repmsg),
1149 sizeof (old_hdl))) {
1150 LCONSOLE_WARN("Reconnect to %s (at @%s) failed due "
1151 "bad handle %#llx\n",
1152 obd2cli_tgt(imp->imp_obd),
1153 imp->imp_connection->c_remote_uuid.uuid,
1154 imp->imp_dlm_handle.cookie);
1155 GOTO(out, rc = -ENOTCONN);
1158 if (memcmp(&imp->imp_remote_handle,
1159 lustre_msg_get_handle(request->rq_repmsg),
1160 sizeof(imp->imp_remote_handle))) {
1161 int level = msg_flags & MSG_CONNECT_RECOVERING ?
1164 /* Bug 16611/14775: if server handle have changed,
1165 * that means some sort of disconnection happened.
1166 * If the server is not in recovery, that also means it
1167 * already erased all of our state because of previous
1168 * eviction. If it is in recovery - we are safe to
1169 * participate since we can reestablish all of our state
1170 * with server again */
1171 if ((MSG_CONNECT_RECOVERING & msg_flags)) {
1172 CDEBUG(level,"%s@%s changed server handle from "
1174 " but is still in recovery\n",
1175 obd2cli_tgt(imp->imp_obd),
1176 imp->imp_connection->c_remote_uuid.uuid,
1177 imp->imp_remote_handle.cookie,
1178 lustre_msg_get_handle(
1179 request->rq_repmsg)->cookie);
1181 LCONSOLE_WARN("Evicted from %s (at %s) "
1182 "after server handle changed from "
1184 obd2cli_tgt(imp->imp_obd),
1185 imp->imp_connection-> \
1187 imp->imp_remote_handle.cookie,
1188 lustre_msg_get_handle(
1189 request->rq_repmsg)->cookie);
1193 imp->imp_remote_handle =
1194 *lustre_msg_get_handle(request->rq_repmsg);
1196 if (!(MSG_CONNECT_RECOVERING & msg_flags)) {
1197 IMPORT_SET_STATE(imp, LUSTRE_IMP_EVICTED);
1198 GOTO(finish, rc = 0);
1202 CDEBUG(D_HA, "reconnected to %s@%s after partition\n",
1203 obd2cli_tgt(imp->imp_obd),
1204 imp->imp_connection->c_remote_uuid.uuid);
1207 if (imp->imp_invalid) {
1208 CDEBUG(D_HA, "%s: reconnected but import is invalid; "
1209 "marking evicted\n", imp->imp_obd->obd_name);
1210 IMPORT_SET_STATE(imp, LUSTRE_IMP_EVICTED);
1211 } else if (MSG_CONNECT_RECOVERING & msg_flags) {
1212 CDEBUG(D_HA, "%s: reconnected to %s during replay\n",
1213 imp->imp_obd->obd_name,
1214 obd2cli_tgt(imp->imp_obd));
1216 spin_lock(&imp->imp_lock);
1217 imp->imp_resend_replay = 1;
1218 spin_unlock(&imp->imp_lock);
1220 IMPORT_SET_STATE(imp, imp->imp_replay_state);
1222 IMPORT_SET_STATE(imp, LUSTRE_IMP_RECOVER);
1224 } else if ((MSG_CONNECT_RECOVERING & msg_flags) && !imp->imp_invalid) {
1225 LASSERT(imp->imp_replayable);
1226 imp->imp_remote_handle =
1227 *lustre_msg_get_handle(request->rq_repmsg);
1228 imp->imp_last_replay_transno = 0;
1229 imp->imp_replay_cursor = &imp->imp_committed_list;
1230 IMPORT_SET_STATE(imp, LUSTRE_IMP_REPLAY);
1231 } else if ((ocd->ocd_connect_flags & OBD_CONNECT_LIGHTWEIGHT) != 0 &&
1232 !imp->imp_invalid) {
1234 obd_import_event(imp->imp_obd, imp, IMP_EVENT_INVALIDATE);
1235 DEBUG_REQ(D_HA, request, "%s: lwp recover",
1236 imp->imp_obd->obd_name);
1237 imp->imp_remote_handle =
1238 *lustre_msg_get_handle(request->rq_repmsg);
1239 IMPORT_SET_STATE(imp, LUSTRE_IMP_RECOVER);
1241 DEBUG_REQ(D_HA, request, "%s: evicting (reconnect/recover flags"
1242 " not set: %x)", imp->imp_obd->obd_name, msg_flags);
1243 imp->imp_remote_handle =
1244 *lustre_msg_get_handle(request->rq_repmsg);
1245 IMPORT_SET_STATE(imp, LUSTRE_IMP_EVICTED);
1248 /* Sanity checks for a reconnected import. */
1249 if (!(imp->imp_replayable) != !(msg_flags & MSG_CONNECT_REPLAYABLE)) {
1250 CERROR("imp_replayable flag does not match server "
1251 "after reconnect. We should LBUG right here.\n");
1254 if (lustre_msg_get_last_committed(request->rq_repmsg) > 0 &&
1255 lustre_msg_get_last_committed(request->rq_repmsg) <
1256 aa->pcaa_peer_committed) {
1257 CERROR("%s went back in time (transno %lld"
1258 " was previously committed, server now claims %lld"
1259 ")! See https://bugzilla.lustre.org/show_bug.cgi?"
1261 obd2cli_tgt(imp->imp_obd), aa->pcaa_peer_committed,
1262 lustre_msg_get_last_committed(request->rq_repmsg));
1266 ptlrpc_prepare_replay(imp);
1267 rc = ptlrpc_import_recovery_state_machine(imp);
1268 if (rc == -ENOTCONN) {
1269 CDEBUG(D_HA, "evicted/aborted by %s@%s during recovery;"
1270 "invalidating and reconnecting\n",
1271 obd2cli_tgt(imp->imp_obd),
1272 imp->imp_connection->c_remote_uuid.uuid);
1273 ptlrpc_connect_import(imp);
1274 spin_lock(&imp->imp_lock);
1275 imp->imp_connected = 0;
1276 imp->imp_connect_tried = 1;
1277 spin_unlock(&imp->imp_lock);
1282 spin_lock(&imp->imp_lock);
1283 imp->imp_connected = 0;
1284 imp->imp_connect_tried = 1;
1285 spin_unlock(&imp->imp_lock);
1288 class_export_put(exp);
1291 IMPORT_SET_STATE(imp, LUSTRE_IMP_DISCON);
1292 if (rc == -EACCES) {
1294 * Give up trying to reconnect
1295 * EACCES means client has no permission for connection
1297 imp->imp_obd->obd_no_recov = 1;
1298 ptlrpc_deactivate_import(imp);
1301 if (rc == -EPROTO) {
1302 struct obd_connect_data *ocd;
1304 /* reply message might not be ready */
1305 if (request->rq_repmsg == NULL)
1308 ocd = req_capsule_server_get(&request->rq_pill,
1310 /* Servers are not supposed to refuse connections from
1311 * clients based on version, only connection feature
1312 * flags. We should never see this from llite, but it
1313 * may be useful for debugging in the future. */
1315 (ocd->ocd_connect_flags & OBD_CONNECT_VERSION) &&
1316 (ocd->ocd_version != LUSTRE_VERSION_CODE)) {
1317 LCONSOLE_ERROR_MSG(0x16a, "Server %s version "
1319 " refused connection from this client "
1320 "with an incompatible version (%s). "
1321 "Client must be recompiled\n",
1322 obd2cli_tgt(imp->imp_obd),
1323 OBD_OCD_VERSION_MAJOR(ocd->ocd_version),
1324 OBD_OCD_VERSION_MINOR(ocd->ocd_version),
1325 OBD_OCD_VERSION_PATCH(ocd->ocd_version),
1326 OBD_OCD_VERSION_FIX(ocd->ocd_version),
1327 LUSTRE_VERSION_STRING);
1328 ptlrpc_deactivate_import(imp);
1329 IMPORT_SET_STATE(imp, LUSTRE_IMP_CLOSED);
1334 ptlrpc_maybe_ping_import_soon(imp);
1336 CDEBUG(D_HA, "recovery of %s on %s failed (%d)\n",
1337 obd2cli_tgt(imp->imp_obd),
1338 (char *)imp->imp_connection->c_remote_uuid.uuid, rc);
1341 wake_up_all(&imp->imp_recovery_waitq);
1346 * interpret callback for "completed replay" RPCs.
1347 * \see signal_completed_replay
1349 static int completed_replay_interpret(const struct lu_env *env,
1350 struct ptlrpc_request *req,
1354 atomic_dec(&req->rq_import->imp_replay_inflight);
1355 if (req->rq_status == 0 && !req->rq_import->imp_vbr_failed) {
1356 ptlrpc_import_recovery_state_machine(req->rq_import);
1358 if (req->rq_import->imp_vbr_failed) {
1360 "%s: version recovery fails, reconnecting\n",
1361 req->rq_import->imp_obd->obd_name);
1363 CDEBUG(D_HA, "%s: LAST_REPLAY message error: %d, "
1365 req->rq_import->imp_obd->obd_name,
1368 ptlrpc_connect_import(req->rq_import);
1375 * Let server know that we have no requests to replay anymore.
1376 * Achieved by just sending a PING request
1378 static int signal_completed_replay(struct obd_import *imp)
1380 struct ptlrpc_request *req;
1383 if (unlikely(OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_FINISH_REPLAY)))
1386 LASSERT(atomic_read(&imp->imp_replay_inflight) == 0);
1387 atomic_inc(&imp->imp_replay_inflight);
1389 req = ptlrpc_request_alloc_pack(imp, &RQF_OBD_PING, LUSTRE_OBD_VERSION,
1392 atomic_dec(&imp->imp_replay_inflight);
1396 ptlrpc_request_set_replen(req);
1397 req->rq_send_state = LUSTRE_IMP_REPLAY_WAIT;
1398 lustre_msg_add_flags(req->rq_reqmsg,
1399 MSG_LOCK_REPLAY_DONE | MSG_REQ_REPLAY_DONE);
1401 req->rq_timeout *= 3;
1402 req->rq_interpret_reply = completed_replay_interpret;
1404 ptlrpcd_add_req(req);
1409 * In kernel code all import invalidation happens in its own
1410 * separate thread, so that whatever application happened to encounter
1411 * a problem could still be killed or otherwise continue
1413 static int ptlrpc_invalidate_import_thread(void *data)
1415 struct obd_import *imp = data;
1419 unshare_fs_struct();
1421 CDEBUG(D_HA, "thread invalidate import %s to %s@%s\n",
1422 imp->imp_obd->obd_name, obd2cli_tgt(imp->imp_obd),
1423 imp->imp_connection->c_remote_uuid.uuid);
1425 ptlrpc_invalidate_import(imp);
1427 if (obd_dump_on_eviction) {
1428 CERROR("dump the log upon eviction\n");
1429 libcfs_debug_dumplog();
1432 IMPORT_SET_STATE(imp, LUSTRE_IMP_RECOVER);
1433 ptlrpc_import_recovery_state_machine(imp);
1435 class_import_put(imp);
1440 * This is the state machine for client-side recovery on import.
1442 * Typicaly we have two possibly paths. If we came to server and it is not
1443 * in recovery, we just enter IMP_EVICTED state, invalidate our import
1444 * state and reconnect from scratch.
1445 * If we came to server that is in recovery, we enter IMP_REPLAY import state.
1446 * We go through our list of requests to replay and send them to server one by
1448 * After sending all request from the list we change import state to
1449 * IMP_REPLAY_LOCKS and re-request all the locks we believe we have from server
1450 * and also all the locks we don't yet have and wait for server to grant us.
1451 * After that we send a special "replay completed" request and change import
1452 * state to IMP_REPLAY_WAIT.
1453 * Upon receiving reply to that "replay completed" RPC we enter IMP_RECOVER
1454 * state and resend all requests from sending list.
1455 * After that we promote import to FULL state and send all delayed requests
1456 * and import is fully operational after that.
1459 int ptlrpc_import_recovery_state_machine(struct obd_import *imp)
1467 if (imp->imp_state == LUSTRE_IMP_EVICTED) {
1468 deuuidify(obd2cli_tgt(imp->imp_obd), NULL,
1469 &target_start, &target_len);
1470 /* Don't care about MGC eviction */
1471 if (strcmp(imp->imp_obd->obd_type->typ_name,
1472 LUSTRE_MGC_NAME) != 0) {
1473 LCONSOLE_ERROR_MSG(0x167, "%s: This client was evicted "
1474 "by %.*s; in progress operations "
1475 "using this service will fail.\n",
1476 imp->imp_obd->obd_name, target_len,
1478 LASSERTF(!obd_lbug_on_eviction, "LBUG upon eviction");
1480 CDEBUG(D_HA, "evicted from %s@%s; invalidating\n",
1481 obd2cli_tgt(imp->imp_obd),
1482 imp->imp_connection->c_remote_uuid.uuid);
1483 /* reset vbr_failed flag upon eviction */
1484 spin_lock(&imp->imp_lock);
1485 imp->imp_vbr_failed = 0;
1486 spin_unlock(&imp->imp_lock);
1489 struct task_struct *task;
1490 /* bug 17802: XXX client_disconnect_export vs connect request
1491 * race. if client is evicted at this time then we start
1492 * invalidate thread without reference to import and import can
1493 * be freed at same time. */
1494 class_import_get(imp);
1495 task = kthread_run(ptlrpc_invalidate_import_thread, imp,
1498 class_import_put(imp);
1499 CERROR("error starting invalidate thread: %d\n", rc);
1508 if (imp->imp_state == LUSTRE_IMP_REPLAY) {
1509 CDEBUG(D_HA, "replay requested by %s\n",
1510 obd2cli_tgt(imp->imp_obd));
1511 rc = ptlrpc_replay_next(imp, &inflight);
1512 if (inflight == 0 &&
1513 atomic_read(&imp->imp_replay_inflight) == 0) {
1514 IMPORT_SET_STATE(imp, LUSTRE_IMP_REPLAY_LOCKS);
1515 rc = ldlm_replay_locks(imp);
1522 if (imp->imp_state == LUSTRE_IMP_REPLAY_LOCKS) {
1523 if (atomic_read(&imp->imp_replay_inflight) == 0) {
1524 IMPORT_SET_STATE(imp, LUSTRE_IMP_REPLAY_WAIT);
1525 rc = signal_completed_replay(imp);
1531 if (imp->imp_state == LUSTRE_IMP_REPLAY_WAIT) {
1532 if (atomic_read(&imp->imp_replay_inflight) == 0) {
1533 IMPORT_SET_STATE(imp, LUSTRE_IMP_RECOVER);
1537 if (imp->imp_state == LUSTRE_IMP_RECOVER) {
1538 struct ptlrpc_connection *conn = imp->imp_connection;
1540 rc = ptlrpc_resend(imp);
1543 IMPORT_SET_STATE(imp, LUSTRE_IMP_FULL);
1544 ptlrpc_activate_import(imp);
1546 LCONSOLE_INFO("%s: Connection restored to %s (at %s)\n",
1547 imp->imp_obd->obd_name,
1548 obd_uuid2str(&conn->c_remote_uuid),
1549 obd_import_nid2str(imp));
1552 if (imp->imp_state == LUSTRE_IMP_FULL) {
1553 wake_up_all(&imp->imp_recovery_waitq);
1554 ptlrpc_wake_delayed(imp);
1561 static struct ptlrpc_request *ptlrpc_disconnect_prep_req(struct obd_import *imp)
1563 struct ptlrpc_request *req;
1567 switch (imp->imp_connect_op) {
1569 rq_opc = OST_DISCONNECT;
1572 rq_opc = MDS_DISCONNECT;
1575 rq_opc = MGS_DISCONNECT;
1579 CERROR("%s: don't know how to disconnect from %s "
1580 "(connect_op %d): rc = %d\n",
1581 imp->imp_obd->obd_name, obd2cli_tgt(imp->imp_obd),
1582 imp->imp_connect_op, rc);
1583 RETURN(ERR_PTR(rc));
1586 req = ptlrpc_request_alloc_pack(imp, &RQF_MDS_DISCONNECT,
1587 LUSTRE_OBD_VERSION, rq_opc);
1591 /* We are disconnecting, do not retry a failed DISCONNECT rpc if
1592 * it fails. We can get through the above with a down server
1593 * if the client doesn't know the server is gone yet. */
1594 req->rq_no_resend = 1;
1596 /* We want client umounts to happen quickly, no matter the
1598 req->rq_timeout = min_t(int, req->rq_timeout,
1599 INITIAL_CONNECT_TIMEOUT);
1601 IMPORT_SET_STATE(imp, LUSTRE_IMP_CONNECTING);
1602 req->rq_send_state = LUSTRE_IMP_CONNECTING;
1603 ptlrpc_request_set_replen(req);
1608 int ptlrpc_disconnect_import(struct obd_import *imp, int noclose)
1610 struct ptlrpc_request *req;
1614 if (imp->imp_obd->obd_force)
1615 GOTO(set_state, rc);
1617 /* probably the import has been disconnected already being idle */
1618 spin_lock(&imp->imp_lock);
1619 if (imp->imp_state == LUSTRE_IMP_IDLE)
1621 spin_unlock(&imp->imp_lock);
1623 if (ptlrpc_import_in_recovery(imp)) {
1624 struct l_wait_info lwi;
1625 long timeout_jiffies;
1629 if (imp->imp_server_timeout)
1630 timeout = obd_timeout >> 1;
1632 timeout = obd_timeout;
1637 req_portal = imp->imp_client->cli_request_portal;
1638 idx = import_at_get_index(imp, req_portal);
1639 timeout = at_get(&imp->imp_at.iat_service_estimate[idx]);
1642 timeout_jiffies = cfs_time_seconds(timeout);
1643 lwi = LWI_TIMEOUT_INTR(max_t(long, timeout_jiffies, 1),
1644 back_to_sleep, LWI_ON_SIGNAL_NOOP, NULL);
1645 rc = l_wait_event(imp->imp_recovery_waitq,
1646 !ptlrpc_import_in_recovery(imp), &lwi);
1650 spin_lock(&imp->imp_lock);
1651 if (imp->imp_state != LUSTRE_IMP_FULL)
1653 spin_unlock(&imp->imp_lock);
1655 req = ptlrpc_disconnect_prep_req(imp);
1657 GOTO(set_state, rc = PTR_ERR(req));
1658 rc = ptlrpc_queue_wait(req);
1659 ptlrpc_req_finished(req);
1662 spin_lock(&imp->imp_lock);
1665 IMPORT_SET_STATE_NOLOCK(imp, LUSTRE_IMP_DISCON);
1667 IMPORT_SET_STATE_NOLOCK(imp, LUSTRE_IMP_CLOSED);
1668 memset(&imp->imp_remote_handle, 0, sizeof(imp->imp_remote_handle));
1669 spin_unlock(&imp->imp_lock);
1671 if (rc == -ETIMEDOUT || rc == -ENOTCONN || rc == -ESHUTDOWN)
1675 EXPORT_SYMBOL(ptlrpc_disconnect_import);
1677 static int ptlrpc_disconnect_idle_interpret(const struct lu_env *env,
1678 struct ptlrpc_request *req,
1681 struct obd_import *imp = req->rq_import;
1684 DEBUG_REQ(D_HA, req, "inflight=%d, refcount=%d: rc = %d\n",
1685 atomic_read(&imp->imp_inflight),
1686 atomic_read(&imp->imp_refcount), rc);
1688 spin_lock(&imp->imp_lock);
1689 /* DISCONNECT reply can be late and another connection can just
1690 * be initiated. so we have to abort disconnection. */
1691 if (req->rq_import_generation == imp->imp_generation &&
1692 imp->imp_state != LUSTRE_IMP_CLOSED) {
1693 LASSERTF(imp->imp_state == LUSTRE_IMP_CONNECTING,
1694 "%s\n", ptlrpc_import_state_name(imp->imp_state));
1695 imp->imp_state = LUSTRE_IMP_IDLE;
1696 memset(&imp->imp_remote_handle, 0,
1697 sizeof(imp->imp_remote_handle));
1698 /* take our DISCONNECT into account */
1699 if (atomic_read(&imp->imp_inflight) > 1) {
1700 imp->imp_generation++;
1701 imp->imp_initiated_at = imp->imp_generation;
1702 IMPORT_SET_STATE_NOLOCK(imp, LUSTRE_IMP_NEW);
1706 spin_unlock(&imp->imp_lock);
1709 rc = ptlrpc_connect_import(imp);
1711 ptlrpc_pinger_add_import(imp);
1717 int ptlrpc_disconnect_and_idle_import(struct obd_import *imp)
1719 struct ptlrpc_request *req;
1722 if (imp->imp_obd->obd_force)
1725 if (ptlrpc_import_in_recovery(imp))
1728 spin_lock(&imp->imp_lock);
1729 if (imp->imp_state != LUSTRE_IMP_FULL) {
1730 spin_unlock(&imp->imp_lock);
1733 spin_unlock(&imp->imp_lock);
1735 req = ptlrpc_disconnect_prep_req(imp);
1737 RETURN(PTR_ERR(req));
1739 CDEBUG_LIMIT(imp->imp_idle_debug, "%s: disconnect after %llus idle\n",
1740 imp->imp_obd->obd_name,
1741 ktime_get_real_seconds() - imp->imp_last_reply_time);
1742 req->rq_interpret_reply = ptlrpc_disconnect_idle_interpret;
1743 ptlrpcd_add_req(req);
1747 EXPORT_SYMBOL(ptlrpc_disconnect_and_idle_import);
1749 void ptlrpc_cleanup_imp(struct obd_import *imp)
1753 spin_lock(&imp->imp_lock);
1754 IMPORT_SET_STATE_NOLOCK(imp, LUSTRE_IMP_CLOSED);
1755 imp->imp_generation++;
1756 spin_unlock(&imp->imp_lock);
1757 ptlrpc_abort_inflight(imp);
1762 /* Adaptive Timeout utils */
1763 extern unsigned int at_min, at_max, at_history;
1765 /* Update at_current with the specified value (bounded by at_min and at_max),
1766 * as well as the AT history "bins".
1767 * - Bin into timeslices using AT_BINS bins.
1768 * - This gives us a max of the last at_history seconds without the storage,
1769 * but still smoothing out a return to normalcy from a slow response.
1770 * - (E.g. remember the maximum latency in each minute of the last 4 minutes.)
1772 int at_measured(struct adaptive_timeout *at, unsigned int val)
1774 unsigned int old = at->at_current;
1775 time64_t now = ktime_get_real_seconds();
1776 long binlimit = max_t(long, at_history / AT_BINS, 1);
1779 CDEBUG(D_OTHER, "add %u to %p time=%lu v=%u (%u %u %u %u)\n",
1780 val, at, (long)(now - at->at_binstart), at->at_current,
1781 at->at_hist[0], at->at_hist[1], at->at_hist[2], at->at_hist[3]);
1784 /* 0's don't count, because we never want our timeout to
1785 drop to 0, and because 0 could mean an error */
1788 spin_lock(&at->at_lock);
1790 if (unlikely(at->at_binstart == 0)) {
1791 /* Special case to remove default from history */
1792 at->at_current = val;
1793 at->at_worst_ever = val;
1794 at->at_worst_time = now;
1795 at->at_hist[0] = val;
1796 at->at_binstart = now;
1797 } else if (now - at->at_binstart < binlimit ) {
1799 at->at_hist[0] = max(val, at->at_hist[0]);
1800 at->at_current = max(val, at->at_current);
1803 unsigned int maxv = val;
1805 /* move bins over */
1806 shift = (u32)(now - at->at_binstart) / binlimit;
1808 for(i = AT_BINS - 1; i >= 0; i--) {
1810 at->at_hist[i] = at->at_hist[i - shift];
1811 maxv = max(maxv, at->at_hist[i]);
1816 at->at_hist[0] = val;
1817 at->at_current = maxv;
1818 at->at_binstart += shift * binlimit;
1821 if (at->at_current > at->at_worst_ever) {
1822 at->at_worst_ever = at->at_current;
1823 at->at_worst_time = now;
1826 if (at->at_flags & AT_FLG_NOHIST)
1827 /* Only keep last reported val; keeping the rest of the history
1829 at->at_current = val;
1832 at->at_current = min(at->at_current, at_max);
1833 at->at_current = max(at->at_current, at_min);
1835 if (at->at_current != old)
1836 CDEBUG(D_OTHER, "AT %p change: old=%u new=%u delta=%d "
1837 "(val=%u) hist %u %u %u %u\n", at,
1838 old, at->at_current, at->at_current - old, val,
1839 at->at_hist[0], at->at_hist[1], at->at_hist[2],
1842 /* if we changed, report the old value */
1843 old = (at->at_current != old) ? old : 0;
1845 spin_unlock(&at->at_lock);
1849 /* Find the imp_at index for a given portal; assign if space available */
1850 int import_at_get_index(struct obd_import *imp, int portal)
1852 struct imp_at *at = &imp->imp_at;
1855 for (i = 0; i < IMP_AT_MAX_PORTALS; i++) {
1856 if (at->iat_portal[i] == portal)
1858 if (at->iat_portal[i] == 0)
1863 /* Not found in list, add it under a lock */
1864 spin_lock(&imp->imp_lock);
1866 /* Check unused under lock */
1867 for (; i < IMP_AT_MAX_PORTALS; i++) {
1868 if (at->iat_portal[i] == portal)
1870 if (at->iat_portal[i] == 0)
1875 /* Not enough portals? */
1876 LASSERT(i < IMP_AT_MAX_PORTALS);
1878 at->iat_portal[i] = portal;
1880 spin_unlock(&imp->imp_lock);