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.sun.com/software/products/lustre/docs/GPLv2.pdf
20 * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
21 * CA 95054 USA or visit www.sun.com if you need additional information or
27 * Copyright (c) 2003, 2010, Oracle and/or its affiliates. All rights reserved.
28 * Use is subject to license terms.
30 * Copyright (c) 2011, 2015, Intel Corporation.
33 * This file is part of Lustre, http://www.lustre.org/
34 * Lustre is a trademark of Sun Microsystems, Inc.
36 * lustre/ptlrpc/import.c
38 * Author: Mike Shaver <shaver@clusterfs.com>
41 #define DEBUG_SUBSYSTEM S_RPC
43 #include <linux/kthread.h>
44 #include <obd_support.h>
45 #include <lustre_ha.h>
46 #include <lustre_net.h>
47 #include <lustre_import.h>
48 #include <lustre_export.h>
50 #include <obd_cksum.h>
51 #include <obd_class.h>
53 #include "ptlrpc_internal.h"
55 struct ptlrpc_connect_async_args {
56 __u64 pcaa_peer_committed;
57 int pcaa_initial_connect;
61 * Updates import \a imp current state to provided \a state value
62 * Helper function. Must be called under imp_lock.
64 static void __import_set_state(struct obd_import *imp,
65 enum lustre_imp_state state)
68 case LUSTRE_IMP_CLOSED:
70 case LUSTRE_IMP_DISCON:
71 case LUSTRE_IMP_CONNECTING:
73 case LUSTRE_IMP_REPLAY_WAIT:
74 imp->imp_replay_state = LUSTRE_IMP_REPLAY_LOCKS;
77 imp->imp_replay_state = LUSTRE_IMP_REPLAY;
79 imp->imp_state = state;
80 imp->imp_state_hist[imp->imp_state_hist_idx].ish_state = state;
81 imp->imp_state_hist[imp->imp_state_hist_idx].ish_time =
82 cfs_time_current_sec();
83 imp->imp_state_hist_idx = (imp->imp_state_hist_idx + 1) %
87 /* A CLOSED import should remain so. */
88 #define IMPORT_SET_STATE_NOLOCK(imp, state) \
90 if (imp->imp_state != LUSTRE_IMP_CLOSED) { \
91 CDEBUG(D_HA, "%p %s: changing import state from %s to %s\n", \
92 imp, obd2cli_tgt(imp->imp_obd), \
93 ptlrpc_import_state_name(imp->imp_state), \
94 ptlrpc_import_state_name(state)); \
95 __import_set_state(imp, state); \
99 #define IMPORT_SET_STATE(imp, state) \
101 spin_lock(&imp->imp_lock); \
102 IMPORT_SET_STATE_NOLOCK(imp, state); \
103 spin_unlock(&imp->imp_lock); \
106 void ptlrpc_import_enter_resend(struct obd_import *imp)
108 IMPORT_SET_STATE(imp, LUSTRE_IMP_RECOVER);
110 EXPORT_SYMBOL(ptlrpc_import_enter_resend);
113 static int ptlrpc_connect_interpret(const struct lu_env *env,
114 struct ptlrpc_request *request,
115 void * data, int rc);
116 int ptlrpc_import_recovery_state_machine(struct obd_import *imp);
118 /* Only this function is allowed to change the import state when it is
119 * CLOSED. I would rather refcount the import and free it after
120 * disconnection like we do with exports. To do that, the client_obd
121 * will need to save the peer info somewhere other than in the import,
123 int ptlrpc_init_import(struct obd_import *imp)
125 spin_lock(&imp->imp_lock);
127 imp->imp_generation++;
128 imp->imp_state = LUSTRE_IMP_NEW;
130 spin_unlock(&imp->imp_lock);
134 EXPORT_SYMBOL(ptlrpc_init_import);
136 #define UUID_STR "_UUID"
137 void deuuidify(char *uuid, const char *prefix, char **uuid_start, int *uuid_len)
139 *uuid_start = !prefix || strncmp(uuid, prefix, strlen(prefix))
140 ? uuid : uuid + strlen(prefix);
142 *uuid_len = strlen(*uuid_start);
144 if (*uuid_len < strlen(UUID_STR))
147 if (!strncmp(*uuid_start + *uuid_len - strlen(UUID_STR),
148 UUID_STR, strlen(UUID_STR)))
149 *uuid_len -= strlen(UUID_STR);
153 * Returns true if import was FULL, false if import was already not
155 * @imp - import to be disconnected
156 * @conn_cnt - connection count (epoch) of the request that timed out
157 * and caused the disconnection. In some cases, multiple
158 * inflight requests can fail to a single target (e.g. OST
159 * bulk requests) and if one has already caused a reconnection
160 * (increasing the import->conn_cnt) the older failure should
161 * not also cause a reconnection. If zero it forces a reconnect.
163 int ptlrpc_set_import_discon(struct obd_import *imp, __u32 conn_cnt)
167 spin_lock(&imp->imp_lock);
169 if (imp->imp_state == LUSTRE_IMP_FULL &&
170 (conn_cnt == 0 || conn_cnt == imp->imp_conn_cnt)) {
174 deuuidify(obd2cli_tgt(imp->imp_obd), NULL,
175 &target_start, &target_len);
177 if (imp->imp_replayable) {
178 LCONSOLE_WARN("%s: Connection to %.*s (at %s) was "
179 "lost; in progress operations using this "
180 "service will wait for recovery to complete\n",
181 imp->imp_obd->obd_name, target_len, target_start,
182 libcfs_nid2str(imp->imp_connection->c_peer.nid));
184 LCONSOLE_ERROR_MSG(0x166, "%s: Connection to "
185 "%.*s (at %s) was lost; in progress "
186 "operations using this service will fail\n",
187 imp->imp_obd->obd_name,
188 target_len, target_start,
189 libcfs_nid2str(imp->imp_connection->c_peer.nid));
191 IMPORT_SET_STATE_NOLOCK(imp, LUSTRE_IMP_DISCON);
192 spin_unlock(&imp->imp_lock);
194 if (obd_dump_on_timeout)
195 libcfs_debug_dumplog();
197 obd_import_event(imp->imp_obd, imp, IMP_EVENT_DISCON);
200 spin_unlock(&imp->imp_lock);
201 CDEBUG(D_HA, "%s: import %p already %s (conn %u, was %u): %s\n",
202 imp->imp_client->cli_name, imp,
203 (imp->imp_state == LUSTRE_IMP_FULL &&
204 imp->imp_conn_cnt > conn_cnt) ?
205 "reconnected" : "not connected", imp->imp_conn_cnt,
206 conn_cnt, ptlrpc_import_state_name(imp->imp_state));
212 /* Must be called with imp_lock held! */
213 static void ptlrpc_deactivate_and_unlock_import(struct obd_import *imp)
216 assert_spin_locked(&imp->imp_lock);
218 CDEBUG(D_HA, "setting import %s INVALID\n", obd2cli_tgt(imp->imp_obd));
219 imp->imp_invalid = 1;
220 imp->imp_generation++;
221 spin_unlock(&imp->imp_lock);
223 ptlrpc_abort_inflight(imp);
224 obd_import_event(imp->imp_obd, imp, IMP_EVENT_INACTIVE);
230 * This acts as a barrier; all existing requests are rejected, and
231 * no new requests will be accepted until the import is valid again.
233 void ptlrpc_deactivate_import(struct obd_import *imp)
235 spin_lock(&imp->imp_lock);
236 ptlrpc_deactivate_and_unlock_import(imp);
238 EXPORT_SYMBOL(ptlrpc_deactivate_import);
241 ptlrpc_inflight_deadline(struct ptlrpc_request *req, time_t now)
245 if (!(((req->rq_phase == RQ_PHASE_RPC) && !req->rq_waiting) ||
246 (req->rq_phase == RQ_PHASE_BULK) ||
247 (req->rq_phase == RQ_PHASE_NEW)))
250 if (req->rq_timedout)
253 if (req->rq_phase == RQ_PHASE_NEW)
256 dl = req->rq_deadline;
264 static unsigned int ptlrpc_inflight_timeout(struct obd_import *imp)
266 time_t now = cfs_time_current_sec();
267 struct list_head *tmp, *n;
268 struct ptlrpc_request *req;
269 unsigned int timeout = 0;
271 spin_lock(&imp->imp_lock);
272 list_for_each_safe(tmp, n, &imp->imp_sending_list) {
273 req = list_entry(tmp, struct ptlrpc_request, rq_list);
274 timeout = max(ptlrpc_inflight_deadline(req, now), timeout);
276 spin_unlock(&imp->imp_lock);
281 * This function will invalidate the import, if necessary, then block
282 * for all the RPC completions, and finally notify the obd to
283 * invalidate its state (ie cancel locks, clear pending requests,
286 void ptlrpc_invalidate_import(struct obd_import *imp)
288 struct list_head *tmp, *n;
289 struct ptlrpc_request *req;
290 struct l_wait_info lwi;
291 unsigned int timeout;
294 atomic_inc(&imp->imp_inval_count);
296 if (!imp->imp_invalid || imp->imp_obd->obd_no_recov)
297 ptlrpc_deactivate_import(imp);
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 /* Calculate max timeout for waiting on rpcs to error
308 * out. Use obd_timeout if calculated value is smaller
310 if (!OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_REPL_UNLINK)) {
311 timeout = ptlrpc_inflight_timeout(imp);
312 timeout += timeout / 3;
315 timeout = obd_timeout;
317 /* decrease the interval to increase race condition */
321 CDEBUG(D_RPCTRACE,"Sleeping %d sec for inflight to error out\n",
324 /* Wait for all requests to error out and call completion
325 * callbacks. Cap it at obd_timeout -- these should all
326 * have been locally cancelled by ptlrpc_abort_inflight. */
327 lwi = LWI_TIMEOUT_INTERVAL(
328 cfs_timeout_cap(cfs_time_seconds(timeout)),
329 (timeout > 1)?cfs_time_seconds(1):cfs_time_seconds(1)/2,
331 rc = l_wait_event(imp->imp_recovery_waitq,
332 (atomic_read(&imp->imp_inflight) == 0),
335 const char *cli_tgt = obd2cli_tgt(imp->imp_obd);
337 CERROR("%s: rc = %d waiting for callback (%d != 0)\n",
338 cli_tgt, rc, atomic_read(&imp->imp_inflight));
340 spin_lock(&imp->imp_lock);
341 if (atomic_read(&imp->imp_inflight) == 0) {
342 int count = atomic_read(&imp->imp_unregistering);
344 /* We know that "unregistering" rpcs only can
345 * survive in sending or delaying lists (they
346 * maybe waiting for long reply unlink in
347 * sluggish nets). Let's check this. If there
348 * is no inflight and unregistering != 0, this
350 LASSERTF(count == 0, "Some RPCs are still "
351 "unregistering: %d\n", count);
353 /* Let's save one loop as soon as inflight have
354 * dropped to zero. No new inflights possible at
358 list_for_each_safe(tmp, n,
359 &imp->imp_sending_list) {
360 req = list_entry(tmp,
361 struct ptlrpc_request,
363 DEBUG_REQ(D_ERROR, req,
364 "still on sending list");
366 list_for_each_safe(tmp, n,
367 &imp->imp_delayed_list) {
368 req = list_entry(tmp,
369 struct ptlrpc_request,
371 DEBUG_REQ(D_ERROR, req,
372 "still on delayed list");
375 CERROR("%s: Unregistering RPCs found (%d). "
376 "Network is sluggish? Waiting them "
377 "to error out.\n", cli_tgt,
378 atomic_read(&imp->imp_unregistering));
380 spin_unlock(&imp->imp_lock);
385 * Let's additionally check that no new rpcs added to import in
386 * "invalidate" state.
388 LASSERT(atomic_read(&imp->imp_inflight) == 0);
389 obd_import_event(imp->imp_obd, imp, IMP_EVENT_INVALIDATE);
390 sptlrpc_import_flush_all_ctx(imp);
392 atomic_dec(&imp->imp_inval_count);
393 wake_up_all(&imp->imp_recovery_waitq);
395 EXPORT_SYMBOL(ptlrpc_invalidate_import);
397 /* unset imp_invalid */
398 void ptlrpc_activate_import(struct obd_import *imp)
400 struct obd_device *obd = imp->imp_obd;
402 spin_lock(&imp->imp_lock);
403 if (imp->imp_deactive != 0) {
404 spin_unlock(&imp->imp_lock);
408 imp->imp_invalid = 0;
409 spin_unlock(&imp->imp_lock);
410 obd_import_event(obd, imp, IMP_EVENT_ACTIVE);
412 EXPORT_SYMBOL(ptlrpc_activate_import);
414 void ptlrpc_pinger_force(struct obd_import *imp)
416 CDEBUG(D_HA, "%s: waking up pinger s:%s\n", obd2cli_tgt(imp->imp_obd),
417 ptlrpc_import_state_name(imp->imp_state));
419 spin_lock(&imp->imp_lock);
420 imp->imp_force_verify = 1;
421 spin_unlock(&imp->imp_lock);
423 if (imp->imp_state != LUSTRE_IMP_CONNECTING)
424 ptlrpc_pinger_wake_up();
426 EXPORT_SYMBOL(ptlrpc_pinger_force);
428 void ptlrpc_fail_import(struct obd_import *imp, __u32 conn_cnt)
432 LASSERT(!imp->imp_dlm_fake);
434 if (ptlrpc_set_import_discon(imp, conn_cnt)) {
435 if (!imp->imp_replayable) {
436 CDEBUG(D_HA, "import %s@%s for %s not replayable, "
437 "auto-deactivating\n",
438 obd2cli_tgt(imp->imp_obd),
439 imp->imp_connection->c_remote_uuid.uuid,
440 imp->imp_obd->obd_name);
441 ptlrpc_deactivate_import(imp);
444 ptlrpc_pinger_force(imp);
449 int ptlrpc_reconnect_import(struct obd_import *imp)
452 struct l_wait_info lwi;
453 int secs = cfs_time_seconds(obd_timeout);
456 ptlrpc_pinger_force(imp);
458 CDEBUG(D_HA, "%s: recovery started, waiting %u seconds\n",
459 obd2cli_tgt(imp->imp_obd), secs);
461 lwi = LWI_TIMEOUT(secs, NULL, NULL);
462 rc = l_wait_event(imp->imp_recovery_waitq,
463 !ptlrpc_import_in_recovery(imp), &lwi);
464 CDEBUG(D_HA, "%s: recovery finished s:%s\n", obd2cli_tgt(imp->imp_obd),
465 ptlrpc_import_state_name(imp->imp_state));
468 ptlrpc_set_import_discon(imp, 0);
469 /* Force a new connect attempt */
470 ptlrpc_invalidate_import(imp);
471 /* Do a fresh connect next time by zeroing the handle */
472 ptlrpc_disconnect_import(imp, 1);
473 /* Wait for all invalidate calls to finish */
474 if (atomic_read(&imp->imp_inval_count) > 0) {
475 struct l_wait_info lwi = LWI_INTR(LWI_ON_SIGNAL_NOOP, NULL);
478 rc = l_wait_event(imp->imp_recovery_waitq,
479 (atomic_read(&imp->imp_inval_count) == 0),
482 CERROR("Interrupted, inval=%d\n",
483 atomic_read(&imp->imp_inval_count));
486 /* Allow reconnect attempts */
487 imp->imp_obd->obd_no_recov = 0;
488 /* Remove 'invalid' flag */
489 ptlrpc_activate_import(imp);
490 /* Attempt a new connect */
491 ptlrpc_recover_import(imp, NULL, 0);
495 EXPORT_SYMBOL(ptlrpc_reconnect_import);
498 * Connection on import \a imp is changed to another one (if more than one is
499 * present). We typically chose connection that we have not tried to connect to
502 static int import_select_connection(struct obd_import *imp)
504 struct obd_import_conn *imp_conn = NULL, *conn;
505 struct obd_export *dlmexp;
507 int target_len, tried_all = 1;
510 spin_lock(&imp->imp_lock);
512 if (list_empty(&imp->imp_conn_list)) {
513 CERROR("%s: no connections available\n",
514 imp->imp_obd->obd_name);
515 spin_unlock(&imp->imp_lock);
519 list_for_each_entry(conn, &imp->imp_conn_list, oic_item) {
520 CDEBUG(D_HA, "%s: connect to NID %s last attempt %llu\n",
521 imp->imp_obd->obd_name,
522 libcfs_nid2str(conn->oic_conn->c_peer.nid),
523 conn->oic_last_attempt);
525 /* If we have not tried this connection since
526 the last successful attempt, go with this one */
527 if ((conn->oic_last_attempt == 0) ||
528 cfs_time_beforeq_64(conn->oic_last_attempt,
529 imp->imp_last_success_conn)) {
535 /* If all of the connections have already been tried
536 since the last successful connection; just choose the
537 least recently used */
540 else if (cfs_time_before_64(conn->oic_last_attempt,
541 imp_conn->oic_last_attempt))
545 /* if not found, simply choose the current one */
546 if (!imp_conn || imp->imp_force_reconnect) {
547 LASSERT(imp->imp_conn_current);
548 imp_conn = imp->imp_conn_current;
551 LASSERT(imp_conn->oic_conn);
553 /* If we've tried everything, and we're back to the beginning of the
554 list, increase our timeout and try again. It will be reset when
555 we do finally connect. (FIXME: really we should wait for all network
556 state associated with the last connection attempt to drain before
557 trying to reconnect on it.) */
558 if (tried_all && (imp->imp_conn_list.next == &imp_conn->oic_item)) {
559 struct adaptive_timeout *at = &imp->imp_at.iat_net_latency;
560 if (at_get(at) < CONNECTION_SWITCH_MAX) {
561 at_measured(at, at_get(at) + CONNECTION_SWITCH_INC);
562 if (at_get(at) > CONNECTION_SWITCH_MAX)
563 at_reset(at, CONNECTION_SWITCH_MAX);
565 LASSERT(imp_conn->oic_last_attempt);
566 CDEBUG(D_HA, "%s: tried all connections, increasing latency "
567 "to %ds\n", imp->imp_obd->obd_name, at_get(at));
570 imp_conn->oic_last_attempt = cfs_time_current_64();
572 /* switch connection, don't mind if it's same as the current one */
573 if (imp->imp_connection)
574 ptlrpc_connection_put(imp->imp_connection);
575 imp->imp_connection = ptlrpc_connection_addref(imp_conn->oic_conn);
577 dlmexp = class_conn2export(&imp->imp_dlm_handle);
578 LASSERT(dlmexp != NULL);
579 if (dlmexp->exp_connection)
580 ptlrpc_connection_put(dlmexp->exp_connection);
581 dlmexp->exp_connection = ptlrpc_connection_addref(imp_conn->oic_conn);
582 class_export_put(dlmexp);
584 if (imp->imp_conn_current != imp_conn) {
585 if (imp->imp_conn_current) {
586 deuuidify(obd2cli_tgt(imp->imp_obd), NULL,
587 &target_start, &target_len);
589 CDEBUG(D_HA, "%s: Connection changing to"
591 imp->imp_obd->obd_name,
592 target_len, target_start,
593 libcfs_nid2str(imp_conn->oic_conn->c_peer.nid));
596 imp->imp_conn_current = imp_conn;
599 CDEBUG(D_HA, "%s: import %p using connection %s/%s\n",
600 imp->imp_obd->obd_name, imp, imp_conn->oic_uuid.uuid,
601 libcfs_nid2str(imp_conn->oic_conn->c_peer.nid));
603 spin_unlock(&imp->imp_lock);
609 * must be called under imp_lock
611 static int ptlrpc_first_transno(struct obd_import *imp, __u64 *transno)
613 struct ptlrpc_request *req;
614 struct list_head *tmp;
616 /* The requests in committed_list always have smaller transnos than
617 * the requests in replay_list */
618 if (!list_empty(&imp->imp_committed_list)) {
619 tmp = imp->imp_committed_list.next;
620 req = list_entry(tmp, struct ptlrpc_request, rq_replay_list);
621 *transno = req->rq_transno;
622 if (req->rq_transno == 0) {
623 DEBUG_REQ(D_ERROR, req, "zero transno in committed_list");
628 if (!list_empty(&imp->imp_replay_list)) {
629 tmp = imp->imp_replay_list.next;
630 req = list_entry(tmp, struct ptlrpc_request, rq_replay_list);
631 *transno = req->rq_transno;
632 if (req->rq_transno == 0) {
633 DEBUG_REQ(D_ERROR, req, "zero transno in replay_list");
642 * Attempt to (re)connect import \a imp. This includes all preparations,
643 * initializing CONNECT RPC request and passing it to ptlrpcd for
645 * Returns 0 on success or error code.
647 int ptlrpc_connect_import(struct obd_import *imp)
649 struct obd_device *obd = imp->imp_obd;
650 int initial_connect = 0;
652 __u64 committed_before_reconnect = 0;
653 struct ptlrpc_request *request;
654 char *bufs[] = { NULL,
655 obd2cli_tgt(imp->imp_obd),
657 (char *)&imp->imp_dlm_handle,
658 (char *)&imp->imp_connect_data };
659 struct ptlrpc_connect_async_args *aa;
663 spin_lock(&imp->imp_lock);
664 if (imp->imp_state == LUSTRE_IMP_CLOSED) {
665 spin_unlock(&imp->imp_lock);
666 CERROR("can't connect to a closed import\n");
668 } else if (imp->imp_state == LUSTRE_IMP_FULL) {
669 spin_unlock(&imp->imp_lock);
670 CERROR("already connected\n");
672 } else if (imp->imp_state == LUSTRE_IMP_CONNECTING ||
673 imp->imp_connected) {
674 spin_unlock(&imp->imp_lock);
675 CERROR("already connecting\n");
679 IMPORT_SET_STATE_NOLOCK(imp, LUSTRE_IMP_CONNECTING);
682 imp->imp_resend_replay = 0;
684 if (!lustre_handle_is_used(&imp->imp_remote_handle))
687 committed_before_reconnect = imp->imp_peer_committed_transno;
689 set_transno = ptlrpc_first_transno(imp,
690 &imp->imp_connect_data.ocd_transno);
691 spin_unlock(&imp->imp_lock);
693 rc = import_select_connection(imp);
697 rc = sptlrpc_import_sec_adapt(imp, NULL, NULL);
701 /* Reset connect flags to the originally requested flags, in case
702 * the server is updated on-the-fly we will get the new features. */
703 imp->imp_connect_data.ocd_connect_flags = imp->imp_connect_flags_orig;
704 imp->imp_connect_data.ocd_connect_flags2 = imp->imp_connect_flags2_orig;
705 /* Reset ocd_version each time so the server knows the exact versions */
706 imp->imp_connect_data.ocd_version = LUSTRE_VERSION_CODE;
707 imp->imp_msghdr_flags &= ~MSGHDR_AT_SUPPORT;
708 imp->imp_msghdr_flags &= ~MSGHDR_CKSUM_INCOMPAT18;
710 rc = obd_reconnect(NULL, imp->imp_obd->obd_self_export, obd,
711 &obd->obd_uuid, &imp->imp_connect_data, NULL);
715 request = ptlrpc_request_alloc(imp, &RQF_MDS_CONNECT);
717 GOTO(out, rc = -ENOMEM);
719 rc = ptlrpc_request_bufs_pack(request, LUSTRE_OBD_VERSION,
720 imp->imp_connect_op, bufs, NULL);
722 ptlrpc_request_free(request);
726 /* Report the rpc service time to the server so that it knows how long
727 * to wait for clients to join recovery */
728 lustre_msg_set_service_time(request->rq_reqmsg,
729 at_timeout2est(request->rq_timeout));
731 /* The amount of time we give the server to process the connect req.
732 * import_select_connection will increase the net latency on
733 * repeated reconnect attempts to cover slow networks.
734 * We override/ignore the server rpc completion estimate here,
735 * which may be large if this is a reconnect attempt */
736 request->rq_timeout = INITIAL_CONNECT_TIMEOUT;
737 lustre_msg_set_timeout(request->rq_reqmsg, request->rq_timeout);
739 request->rq_no_resend = request->rq_no_delay = 1;
740 request->rq_send_state = LUSTRE_IMP_CONNECTING;
741 /* Allow a slightly larger reply for future growth compatibility */
742 req_capsule_set_size(&request->rq_pill, &RMF_CONNECT_DATA, RCL_SERVER,
743 sizeof(struct obd_connect_data)+16*sizeof(__u64));
744 ptlrpc_request_set_replen(request);
745 request->rq_interpret_reply = ptlrpc_connect_interpret;
747 CLASSERT(sizeof(*aa) <= sizeof(request->rq_async_args));
748 aa = ptlrpc_req_async_args(request);
749 memset(aa, 0, sizeof *aa);
751 aa->pcaa_peer_committed = committed_before_reconnect;
752 aa->pcaa_initial_connect = initial_connect;
754 if (aa->pcaa_initial_connect) {
755 spin_lock(&imp->imp_lock);
756 imp->imp_replayable = 1;
757 spin_unlock(&imp->imp_lock);
758 lustre_msg_add_op_flags(request->rq_reqmsg,
759 MSG_CONNECT_INITIAL);
763 lustre_msg_add_op_flags(request->rq_reqmsg,
764 MSG_CONNECT_TRANSNO);
766 DEBUG_REQ(D_RPCTRACE, request, "(re)connect request (timeout %d)",
767 request->rq_timeout);
768 ptlrpcd_add_req(request);
772 IMPORT_SET_STATE(imp, LUSTRE_IMP_DISCON);
776 EXPORT_SYMBOL(ptlrpc_connect_import);
778 static void ptlrpc_maybe_ping_import_soon(struct obd_import *imp)
782 spin_lock(&imp->imp_lock);
783 force_verify = imp->imp_force_verify != 0;
784 spin_unlock(&imp->imp_lock);
787 ptlrpc_pinger_wake_up();
790 static int ptlrpc_busy_reconnect(int rc)
792 return (rc == -EBUSY) || (rc == -EAGAIN);
795 static int ptlrpc_connect_set_flags(struct obd_import *imp,
796 struct obd_connect_data *ocd,
797 __u64 old_connect_flags,
798 struct obd_export *exp, int init_connect)
801 struct client_obd *cli = &imp->imp_obd->u.cli;
803 if ((imp->imp_connect_flags_orig & OBD_CONNECT_IBITS) &&
804 !(ocd->ocd_connect_flags & OBD_CONNECT_IBITS)) {
805 LCONSOLE_WARN("%s: MDS %s does not support ibits "
806 "lock, either very old or invalid: "
807 "requested %#llx, replied %#llx\n",
808 imp->imp_obd->obd_name,
809 imp->imp_connection->c_remote_uuid.uuid,
810 imp->imp_connect_flags_orig,
811 ocd->ocd_connect_flags);
815 spin_lock(&imp->imp_lock);
816 list_del(&imp->imp_conn_current->oic_item);
817 list_add(&imp->imp_conn_current->oic_item,
818 &imp->imp_conn_list);
819 imp->imp_last_success_conn =
820 imp->imp_conn_current->oic_last_attempt;
822 spin_unlock(&imp->imp_lock);
825 if (!warned && (ocd->ocd_connect_flags & OBD_CONNECT_VERSION) &&
826 (ocd->ocd_version > LUSTRE_VERSION_CODE +
827 LUSTRE_VERSION_OFFSET_WARN ||
828 ocd->ocd_version < LUSTRE_VERSION_CODE -
829 LUSTRE_VERSION_OFFSET_WARN)) {
830 /* Sigh, some compilers do not like #ifdef in the middle
831 of macro arguments */
832 const char *older = "older than client. "
833 "Consider upgrading server";
834 const char *newer = "newer than client. "
835 "Consider recompiling application";
837 LCONSOLE_WARN("Server %s version (%d.%d.%d.%d) "
839 obd2cli_tgt(imp->imp_obd),
840 OBD_OCD_VERSION_MAJOR(ocd->ocd_version),
841 OBD_OCD_VERSION_MINOR(ocd->ocd_version),
842 OBD_OCD_VERSION_PATCH(ocd->ocd_version),
843 OBD_OCD_VERSION_FIX(ocd->ocd_version),
844 ocd->ocd_version > LUSTRE_VERSION_CODE ?
845 newer : older, LUSTRE_VERSION_STRING);
849 #if LUSTRE_VERSION_CODE < OBD_OCD_VERSION(3, 0, 53, 0)
850 /* Check if server has LU-1252 fix applied to not always swab
851 * the IR MNE entries. Do this only once per connection. This
852 * fixup is version-limited, because we don't want to carry the
853 * OBD_CONNECT_MNE_SWAB flag around forever, just so long as we
854 * need interop with unpatched 2.2 servers. For newer servers,
855 * the client will do MNE swabbing only as needed. LU-1644 */
856 if (unlikely((ocd->ocd_connect_flags & OBD_CONNECT_VERSION) &&
857 !(ocd->ocd_connect_flags & OBD_CONNECT_MNE_SWAB) &&
858 OBD_OCD_VERSION_MAJOR(ocd->ocd_version) == 2 &&
859 OBD_OCD_VERSION_MINOR(ocd->ocd_version) == 2 &&
860 OBD_OCD_VERSION_PATCH(ocd->ocd_version) < 55 &&
861 strcmp(imp->imp_obd->obd_type->typ_name,
862 LUSTRE_MGC_NAME) == 0))
863 imp->imp_need_mne_swab = 1;
864 else /* clear if server was upgraded since last connect */
865 imp->imp_need_mne_swab = 0;
868 if (ocd->ocd_connect_flags & OBD_CONNECT_CKSUM) {
869 /* We sent to the server ocd_cksum_types with bits set
870 * for algorithms we understand. The server masked off
871 * the checksum types it doesn't support */
872 if ((ocd->ocd_cksum_types &
873 cksum_types_supported_client()) == 0) {
874 LCONSOLE_WARN("The negotiation of the checksum "
875 "alogrithm to use with server %s "
876 "failed (%x/%x), disabling "
878 obd2cli_tgt(imp->imp_obd),
879 ocd->ocd_cksum_types,
880 cksum_types_supported_client());
881 cli->cl_checksum = 0;
882 cli->cl_supp_cksum_types = OBD_CKSUM_ADLER;
884 cli->cl_supp_cksum_types = ocd->ocd_cksum_types;
887 /* The server does not support OBD_CONNECT_CKSUM.
888 * Enforce ADLER for backward compatibility*/
889 cli->cl_supp_cksum_types = OBD_CKSUM_ADLER;
891 cli->cl_cksum_type = cksum_type_select(cli->cl_supp_cksum_types);
893 if (ocd->ocd_connect_flags & OBD_CONNECT_BRW_SIZE)
894 cli->cl_max_pages_per_rpc =
895 min(ocd->ocd_brw_size >> PAGE_CACHE_SHIFT,
896 cli->cl_max_pages_per_rpc);
897 else if (imp->imp_connect_op == MDS_CONNECT ||
898 imp->imp_connect_op == MGS_CONNECT)
899 cli->cl_max_pages_per_rpc = 1;
901 LASSERT((cli->cl_max_pages_per_rpc <= PTLRPC_MAX_BRW_PAGES) &&
902 (cli->cl_max_pages_per_rpc > 0));
904 client_adjust_max_dirty(cli);
906 /* Update client max modify RPCs in flight with value returned
908 if (ocd->ocd_connect_flags & OBD_CONNECT_MULTIMODRPCS)
909 cli->cl_max_mod_rpcs_in_flight = min(
910 cli->cl_max_mod_rpcs_in_flight,
911 ocd->ocd_maxmodrpcs);
913 cli->cl_max_mod_rpcs_in_flight = 1;
915 /* Reset ns_connect_flags only for initial connect. It might be
916 * changed in while using FS and if we reset it in reconnect
917 * this leads to losing user settings done before such as
918 * disable lru_resize, etc. */
919 if (old_connect_flags != exp_connect_flags(exp) || init_connect) {
920 CDEBUG(D_HA, "%s: Resetting ns_connect_flags to server "
921 "flags: %#llx\n", imp->imp_obd->obd_name,
922 ocd->ocd_connect_flags);
923 imp->imp_obd->obd_namespace->ns_connect_flags =
924 ocd->ocd_connect_flags;
925 imp->imp_obd->obd_namespace->ns_orig_connect_flags =
926 ocd->ocd_connect_flags;
929 if (ocd->ocd_connect_flags & OBD_CONNECT_AT)
930 /* We need a per-message support flag, because
931 * a. we don't know if the incoming connect reply
932 * supports AT or not (in reply_in_callback)
933 * until we unpack it.
934 * b. failovered server means export and flags are gone
935 * (in ptlrpc_send_reply).
936 * Can only be set when we know AT is supported at
938 imp->imp_msghdr_flags |= MSGHDR_AT_SUPPORT;
940 imp->imp_msghdr_flags &= ~MSGHDR_AT_SUPPORT;
942 imp->imp_msghdr_flags |= MSGHDR_CKSUM_INCOMPAT18;
948 * Add all replay requests back to unreplied list before start replay,
949 * so that we can make sure the known replied XID is always increased
950 * only even if when replaying requests.
952 static void ptlrpc_prepare_replay(struct obd_import *imp)
954 struct ptlrpc_request *req;
956 if (imp->imp_state != LUSTRE_IMP_REPLAY ||
957 imp->imp_resend_replay)
960 /* If the server was restart during repaly, the requests may
961 * have been added to the unreplied list in former replay. */
962 spin_lock(&imp->imp_lock);
964 list_for_each_entry(req, &imp->imp_committed_list, rq_replay_list) {
965 if (list_empty(&req->rq_unreplied_list))
966 ptlrpc_add_unreplied(req);
969 list_for_each_entry(req, &imp->imp_replay_list, rq_replay_list) {
970 if (list_empty(&req->rq_unreplied_list))
971 ptlrpc_add_unreplied(req);
974 imp->imp_known_replied_xid = ptlrpc_known_replied_xid(imp);
975 spin_unlock(&imp->imp_lock);
979 * interpret_reply callback for connect RPCs.
980 * Looks into returned status of connect operation and decides
981 * what to do with the import - i.e enter recovery, promote it to
982 * full state for normal operations of disconnect it due to an error.
984 static int ptlrpc_connect_interpret(const struct lu_env *env,
985 struct ptlrpc_request *request,
988 struct ptlrpc_connect_async_args *aa = data;
989 struct obd_import *imp = request->rq_import;
990 struct lustre_handle old_hdl;
991 __u64 old_connect_flags;
993 struct obd_connect_data *ocd;
994 struct obd_export *exp;
998 spin_lock(&imp->imp_lock);
999 if (imp->imp_state == LUSTRE_IMP_CLOSED) {
1000 imp->imp_connect_tried = 1;
1001 spin_unlock(&imp->imp_lock);
1006 /* if this reconnect to busy export - not need select new target
1008 imp->imp_force_reconnect = ptlrpc_busy_reconnect(rc);
1009 spin_unlock(&imp->imp_lock);
1010 ptlrpc_maybe_ping_import_soon(imp);
1014 /* LU-7558: indicate that we are interpretting connect reply,
1015 * pltrpc_connect_import() will not try to reconnect until
1016 * interpret will finish. */
1017 imp->imp_connected = 1;
1018 spin_unlock(&imp->imp_lock);
1020 LASSERT(imp->imp_conn_current);
1022 msg_flags = lustre_msg_get_op_flags(request->rq_repmsg);
1024 ret = req_capsule_get_size(&request->rq_pill, &RMF_CONNECT_DATA,
1026 /* server replied obd_connect_data is always bigger */
1027 ocd = req_capsule_server_sized_get(&request->rq_pill,
1028 &RMF_CONNECT_DATA, ret);
1031 CERROR("%s: no connect data from server\n",
1032 imp->imp_obd->obd_name);
1037 spin_lock(&imp->imp_lock);
1039 /* All imports are pingable */
1040 imp->imp_pingable = 1;
1041 imp->imp_force_reconnect = 0;
1042 imp->imp_force_verify = 0;
1044 imp->imp_connect_data = *ocd;
1046 CDEBUG(D_HA, "%s: connect to target with instance %u\n",
1047 imp->imp_obd->obd_name, ocd->ocd_instance);
1048 exp = class_conn2export(&imp->imp_dlm_handle);
1050 spin_unlock(&imp->imp_lock);
1052 /* check that server granted subset of flags we asked for. */
1053 if ((ocd->ocd_connect_flags & imp->imp_connect_flags_orig) !=
1054 ocd->ocd_connect_flags) {
1055 CERROR("%s: Server didn't grant requested subset of flags: "
1056 "asked=%#llx granted=%#llx\n",
1057 imp->imp_obd->obd_name, imp->imp_connect_flags_orig,
1058 ocd->ocd_connect_flags);
1059 GOTO(out, rc = -EPROTO);
1062 if ((ocd->ocd_connect_flags2 & imp->imp_connect_flags2_orig) !=
1063 ocd->ocd_connect_flags2) {
1064 CERROR("%s: Server didn't grant requested subset of flags2: "
1065 "asked=%#llx granted=%#llx\n",
1066 imp->imp_obd->obd_name, imp->imp_connect_flags2_orig,
1067 ocd->ocd_connect_flags2);
1068 GOTO(out, rc = -EPROTO);
1071 if (!(imp->imp_connect_flags_orig & OBD_CONNECT_LIGHTWEIGHT) &&
1072 (imp->imp_connect_flags_orig & OBD_CONNECT_MDS_MDS) &&
1073 (imp->imp_connect_flags_orig & OBD_CONNECT_FID) &&
1074 (ocd->ocd_connect_flags & OBD_CONNECT_VERSION)) {
1075 __u32 major = OBD_OCD_VERSION_MAJOR(ocd->ocd_version);
1076 __u32 minor = OBD_OCD_VERSION_MINOR(ocd->ocd_version);
1077 __u32 patch = OBD_OCD_VERSION_PATCH(ocd->ocd_version);
1079 /* We do not support the MDT-MDT interoperations with
1080 * different version MDT because of protocol changes. */
1081 if (unlikely(major != LUSTRE_MAJOR ||
1082 minor != LUSTRE_MINOR ||
1083 abs(patch - LUSTRE_PATCH) > 3)) {
1084 LCONSOLE_WARN("%s: import %p (%u.%u.%u.%u) tried the "
1085 "connection to different version MDT "
1086 "(%d.%d.%d.%d) %s\n",
1087 imp->imp_obd->obd_name, imp, LUSTRE_MAJOR,
1088 LUSTRE_MINOR, LUSTRE_PATCH, LUSTRE_FIX,
1089 major, minor, patch,
1090 OBD_OCD_VERSION_FIX(ocd->ocd_version),
1091 imp->imp_connection->c_remote_uuid.uuid);
1093 GOTO(out, rc = -EPROTO);
1098 /* This could happen if export is cleaned during the
1100 CERROR("%s: missing export after connect\n",
1101 imp->imp_obd->obd_name);
1102 GOTO(out, rc = -ENODEV);
1104 old_connect_flags = exp_connect_flags(exp);
1105 exp->exp_connect_data = *ocd;
1106 imp->imp_obd->obd_self_export->exp_connect_data = *ocd;
1108 /* The net statistics after (re-)connect is not valid anymore,
1109 * because may reflect other routing, etc. */
1110 at_reinit(&imp->imp_at.iat_net_latency, 0, 0);
1111 ptlrpc_at_adj_net_latency(request,
1112 lustre_msg_get_service_time(request->rq_repmsg));
1114 /* Import flags should be updated before waking import at FULL state */
1115 rc = ptlrpc_connect_set_flags(imp, ocd, old_connect_flags, exp,
1116 aa->pcaa_initial_connect);
1117 class_export_put(exp);
1121 obd_import_event(imp->imp_obd, imp, IMP_EVENT_OCD);
1123 if (aa->pcaa_initial_connect) {
1124 spin_lock(&imp->imp_lock);
1125 if (msg_flags & MSG_CONNECT_REPLAYABLE) {
1126 imp->imp_replayable = 1;
1127 spin_unlock(&imp->imp_lock);
1128 CDEBUG(D_HA, "connected to replayable target: %s\n",
1129 obd2cli_tgt(imp->imp_obd));
1131 imp->imp_replayable = 0;
1132 spin_unlock(&imp->imp_lock);
1135 /* if applies, adjust the imp->imp_msg_magic here
1136 * according to reply flags */
1138 imp->imp_remote_handle =
1139 *lustre_msg_get_handle(request->rq_repmsg);
1141 /* Initial connects are allowed for clients with non-random
1142 * uuids when servers are in recovery. Simply signal the
1143 * servers replay is complete and wait in REPLAY_WAIT. */
1144 if (msg_flags & MSG_CONNECT_RECOVERING) {
1145 CDEBUG(D_HA, "connect to %s during recovery\n",
1146 obd2cli_tgt(imp->imp_obd));
1147 IMPORT_SET_STATE(imp, LUSTRE_IMP_REPLAY_LOCKS);
1149 IMPORT_SET_STATE(imp, LUSTRE_IMP_FULL);
1150 ptlrpc_activate_import(imp);
1153 GOTO(finish, rc = 0);
1156 /* Determine what recovery state to move the import to. */
1157 if (MSG_CONNECT_RECONNECT & msg_flags) {
1158 memset(&old_hdl, 0, sizeof(old_hdl));
1159 if (!memcmp(&old_hdl, lustre_msg_get_handle(request->rq_repmsg),
1160 sizeof (old_hdl))) {
1161 LCONSOLE_WARN("Reconnect to %s (at @%s) failed due "
1162 "bad handle %#llx\n",
1163 obd2cli_tgt(imp->imp_obd),
1164 imp->imp_connection->c_remote_uuid.uuid,
1165 imp->imp_dlm_handle.cookie);
1166 GOTO(out, rc = -ENOTCONN);
1169 if (memcmp(&imp->imp_remote_handle,
1170 lustre_msg_get_handle(request->rq_repmsg),
1171 sizeof(imp->imp_remote_handle))) {
1172 int level = msg_flags & MSG_CONNECT_RECOVERING ?
1175 /* Bug 16611/14775: if server handle have changed,
1176 * that means some sort of disconnection happened.
1177 * If the server is not in recovery, that also means it
1178 * already erased all of our state because of previous
1179 * eviction. If it is in recovery - we are safe to
1180 * participate since we can reestablish all of our state
1181 * with server again */
1182 if ((MSG_CONNECT_RECOVERING & msg_flags)) {
1183 CDEBUG(level,"%s@%s changed server handle from "
1185 " but is still in recovery\n",
1186 obd2cli_tgt(imp->imp_obd),
1187 imp->imp_connection->c_remote_uuid.uuid,
1188 imp->imp_remote_handle.cookie,
1189 lustre_msg_get_handle(
1190 request->rq_repmsg)->cookie);
1192 LCONSOLE_WARN("Evicted from %s (at %s) "
1193 "after server handle changed from "
1195 obd2cli_tgt(imp->imp_obd),
1196 imp->imp_connection-> \
1198 imp->imp_remote_handle.cookie,
1199 lustre_msg_get_handle(
1200 request->rq_repmsg)->cookie);
1204 imp->imp_remote_handle =
1205 *lustre_msg_get_handle(request->rq_repmsg);
1207 if (!(MSG_CONNECT_RECOVERING & msg_flags)) {
1208 IMPORT_SET_STATE(imp, LUSTRE_IMP_EVICTED);
1209 GOTO(finish, rc = 0);
1213 CDEBUG(D_HA, "reconnected to %s@%s after partition\n",
1214 obd2cli_tgt(imp->imp_obd),
1215 imp->imp_connection->c_remote_uuid.uuid);
1218 if (imp->imp_invalid) {
1219 CDEBUG(D_HA, "%s: reconnected but import is invalid; "
1220 "marking evicted\n", imp->imp_obd->obd_name);
1221 IMPORT_SET_STATE(imp, LUSTRE_IMP_EVICTED);
1222 } else if (MSG_CONNECT_RECOVERING & msg_flags) {
1223 CDEBUG(D_HA, "%s: reconnected to %s during replay\n",
1224 imp->imp_obd->obd_name,
1225 obd2cli_tgt(imp->imp_obd));
1227 spin_lock(&imp->imp_lock);
1228 imp->imp_resend_replay = 1;
1229 spin_unlock(&imp->imp_lock);
1231 IMPORT_SET_STATE(imp, imp->imp_replay_state);
1233 IMPORT_SET_STATE(imp, LUSTRE_IMP_RECOVER);
1235 } else if ((MSG_CONNECT_RECOVERING & msg_flags) && !imp->imp_invalid) {
1236 LASSERT(imp->imp_replayable);
1237 imp->imp_remote_handle =
1238 *lustre_msg_get_handle(request->rq_repmsg);
1239 imp->imp_last_replay_transno = 0;
1240 imp->imp_replay_cursor = &imp->imp_committed_list;
1241 IMPORT_SET_STATE(imp, LUSTRE_IMP_REPLAY);
1243 DEBUG_REQ(D_HA, request, "%s: evicting (reconnect/recover flags"
1244 " not set: %x)", imp->imp_obd->obd_name, msg_flags);
1245 imp->imp_remote_handle =
1246 *lustre_msg_get_handle(request->rq_repmsg);
1247 IMPORT_SET_STATE(imp, LUSTRE_IMP_EVICTED);
1250 /* Sanity checks for a reconnected import. */
1251 if (!(imp->imp_replayable) != !(msg_flags & MSG_CONNECT_REPLAYABLE)) {
1252 CERROR("imp_replayable flag does not match server "
1253 "after reconnect. We should LBUG right here.\n");
1256 if (lustre_msg_get_last_committed(request->rq_repmsg) > 0 &&
1257 lustre_msg_get_last_committed(request->rq_repmsg) <
1258 aa->pcaa_peer_committed) {
1259 CERROR("%s went back in time (transno %lld"
1260 " was previously committed, server now claims %lld"
1261 ")! See https://bugzilla.lustre.org/show_bug.cgi?"
1263 obd2cli_tgt(imp->imp_obd), aa->pcaa_peer_committed,
1264 lustre_msg_get_last_committed(request->rq_repmsg));
1268 ptlrpc_prepare_replay(imp);
1269 rc = ptlrpc_import_recovery_state_machine(imp);
1270 if (rc == -ENOTCONN) {
1271 CDEBUG(D_HA, "evicted/aborted by %s@%s during recovery;"
1272 "invalidating and reconnecting\n",
1273 obd2cli_tgt(imp->imp_obd),
1274 imp->imp_connection->c_remote_uuid.uuid);
1275 ptlrpc_connect_import(imp);
1276 spin_lock(&imp->imp_lock);
1277 imp->imp_connected = 0;
1278 imp->imp_connect_tried = 1;
1279 spin_unlock(&imp->imp_lock);
1284 spin_lock(&imp->imp_lock);
1285 imp->imp_connected = 0;
1286 imp->imp_connect_tried = 1;
1287 spin_unlock(&imp->imp_lock);
1290 IMPORT_SET_STATE(imp, LUSTRE_IMP_DISCON);
1291 if (rc == -EACCES) {
1293 * Give up trying to reconnect
1294 * EACCES means client has no permission for connection
1296 imp->imp_obd->obd_no_recov = 1;
1297 ptlrpc_deactivate_import(imp);
1300 if (rc == -EPROTO) {
1301 struct obd_connect_data *ocd;
1303 /* reply message might not be ready */
1304 if (request->rq_repmsg == NULL)
1307 ocd = req_capsule_server_get(&request->rq_pill,
1310 (ocd->ocd_connect_flags & OBD_CONNECT_VERSION) &&
1311 (ocd->ocd_version != LUSTRE_VERSION_CODE)) {
1312 /* Actually servers are only supposed to refuse
1313 connection from liblustre clients, so we should
1314 never see this from VFS context */
1315 LCONSOLE_ERROR_MSG(0x16a, "Server %s version "
1317 " refused connection from this client "
1318 "with an incompatible version (%s). "
1319 "Client must be recompiled\n",
1320 obd2cli_tgt(imp->imp_obd),
1321 OBD_OCD_VERSION_MAJOR(ocd->ocd_version),
1322 OBD_OCD_VERSION_MINOR(ocd->ocd_version),
1323 OBD_OCD_VERSION_PATCH(ocd->ocd_version),
1324 OBD_OCD_VERSION_FIX(ocd->ocd_version),
1325 LUSTRE_VERSION_STRING);
1326 ptlrpc_deactivate_import(imp);
1327 IMPORT_SET_STATE(imp, LUSTRE_IMP_CLOSED);
1332 ptlrpc_maybe_ping_import_soon(imp);
1334 CDEBUG(D_HA, "recovery of %s on %s failed (%d)\n",
1335 obd2cli_tgt(imp->imp_obd),
1336 (char *)imp->imp_connection->c_remote_uuid.uuid, rc);
1339 wake_up_all(&imp->imp_recovery_waitq);
1344 * interpret callback for "completed replay" RPCs.
1345 * \see signal_completed_replay
1347 static int completed_replay_interpret(const struct lu_env *env,
1348 struct ptlrpc_request *req,
1349 void * data, int rc)
1352 atomic_dec(&req->rq_import->imp_replay_inflight);
1353 if (req->rq_status == 0 &&
1354 !req->rq_import->imp_vbr_failed) {
1355 ptlrpc_import_recovery_state_machine(req->rq_import);
1357 if (req->rq_import->imp_vbr_failed) {
1359 "%s: version recovery fails, reconnecting\n",
1360 req->rq_import->imp_obd->obd_name);
1362 CDEBUG(D_HA, "%s: LAST_REPLAY message error: %d, "
1364 req->rq_import->imp_obd->obd_name,
1367 ptlrpc_connect_import(req->rq_import);
1374 * Let server know that we have no requests to replay anymore.
1375 * Achieved by just sending a PING request
1377 static int signal_completed_replay(struct obd_import *imp)
1379 struct ptlrpc_request *req;
1382 if (unlikely(OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_FINISH_REPLAY)))
1385 LASSERT(atomic_read(&imp->imp_replay_inflight) == 0);
1386 atomic_inc(&imp->imp_replay_inflight);
1388 req = ptlrpc_request_alloc_pack(imp, &RQF_OBD_PING, LUSTRE_OBD_VERSION,
1391 atomic_dec(&imp->imp_replay_inflight);
1395 ptlrpc_request_set_replen(req);
1396 req->rq_send_state = LUSTRE_IMP_REPLAY_WAIT;
1397 lustre_msg_add_flags(req->rq_reqmsg,
1398 MSG_LOCK_REPLAY_DONE | MSG_REQ_REPLAY_DONE);
1400 req->rq_timeout *= 3;
1401 req->rq_interpret_reply = completed_replay_interpret;
1403 ptlrpcd_add_req(req);
1408 * In kernel code all import invalidation happens in its own
1409 * separate thread, so that whatever application happened to encounter
1410 * a problem could still be killed or otherwise continue
1412 static int ptlrpc_invalidate_import_thread(void *data)
1414 struct obd_import *imp = data;
1418 unshare_fs_struct();
1420 CDEBUG(D_HA, "thread invalidate import %s to %s@%s\n",
1421 imp->imp_obd->obd_name, obd2cli_tgt(imp->imp_obd),
1422 imp->imp_connection->c_remote_uuid.uuid);
1424 ptlrpc_invalidate_import(imp);
1426 if (obd_dump_on_eviction) {
1427 CERROR("dump the log upon eviction\n");
1428 libcfs_debug_dumplog();
1431 IMPORT_SET_STATE(imp, LUSTRE_IMP_RECOVER);
1432 ptlrpc_import_recovery_state_machine(imp);
1434 class_import_put(imp);
1439 * This is the state machine for client-side recovery on import.
1441 * Typicaly we have two possibly paths. If we came to server and it is not
1442 * in recovery, we just enter IMP_EVICTED state, invalidate our import
1443 * state and reconnect from scratch.
1444 * If we came to server that is in recovery, we enter IMP_REPLAY import state.
1445 * We go through our list of requests to replay and send them to server one by
1447 * After sending all request from the list we change import state to
1448 * IMP_REPLAY_LOCKS and re-request all the locks we believe we have from server
1449 * and also all the locks we don't yet have and wait for server to grant us.
1450 * After that we send a special "replay completed" request and change import
1451 * state to IMP_REPLAY_WAIT.
1452 * Upon receiving reply to that "replay completed" RPC we enter IMP_RECOVER
1453 * state and resend all requests from sending list.
1454 * After that we promote import to FULL state and send all delayed requests
1455 * and import is fully operational after that.
1458 int ptlrpc_import_recovery_state_machine(struct obd_import *imp)
1466 if (imp->imp_state == LUSTRE_IMP_EVICTED) {
1467 deuuidify(obd2cli_tgt(imp->imp_obd), NULL,
1468 &target_start, &target_len);
1469 /* Don't care about MGC eviction */
1470 if (strcmp(imp->imp_obd->obd_type->typ_name,
1471 LUSTRE_MGC_NAME) != 0) {
1472 LCONSOLE_ERROR_MSG(0x167, "%s: This client was evicted "
1473 "by %.*s; in progress operations "
1474 "using this service will fail.\n",
1475 imp->imp_obd->obd_name, target_len,
1478 CDEBUG(D_HA, "evicted from %s@%s; invalidating\n",
1479 obd2cli_tgt(imp->imp_obd),
1480 imp->imp_connection->c_remote_uuid.uuid);
1481 /* reset vbr_failed flag upon eviction */
1482 spin_lock(&imp->imp_lock);
1483 imp->imp_vbr_failed = 0;
1484 spin_unlock(&imp->imp_lock);
1487 struct task_struct *task;
1488 /* bug 17802: XXX client_disconnect_export vs connect request
1489 * race. if client is evicted at this time then we start
1490 * invalidate thread without reference to import and import can
1491 * be freed at same time. */
1492 class_import_get(imp);
1493 task = kthread_run(ptlrpc_invalidate_import_thread, imp,
1496 class_import_put(imp);
1497 CERROR("error starting invalidate thread: %d\n", rc);
1506 if (imp->imp_state == LUSTRE_IMP_REPLAY) {
1507 CDEBUG(D_HA, "replay requested by %s\n",
1508 obd2cli_tgt(imp->imp_obd));
1509 rc = ptlrpc_replay_next(imp, &inflight);
1510 if (inflight == 0 &&
1511 atomic_read(&imp->imp_replay_inflight) == 0) {
1512 IMPORT_SET_STATE(imp, LUSTRE_IMP_REPLAY_LOCKS);
1513 rc = ldlm_replay_locks(imp);
1520 if (imp->imp_state == LUSTRE_IMP_REPLAY_LOCKS) {
1521 if (atomic_read(&imp->imp_replay_inflight) == 0) {
1522 IMPORT_SET_STATE(imp, LUSTRE_IMP_REPLAY_WAIT);
1523 rc = signal_completed_replay(imp);
1529 if (imp->imp_state == LUSTRE_IMP_REPLAY_WAIT) {
1530 if (atomic_read(&imp->imp_replay_inflight) == 0) {
1531 IMPORT_SET_STATE(imp, LUSTRE_IMP_RECOVER);
1535 if (imp->imp_state == LUSTRE_IMP_RECOVER) {
1536 struct ptlrpc_connection *conn = imp->imp_connection;
1538 rc = ptlrpc_resend(imp);
1541 IMPORT_SET_STATE(imp, LUSTRE_IMP_FULL);
1542 ptlrpc_activate_import(imp);
1544 LCONSOLE_INFO("%s: Connection restored to %s (at %s)\n",
1545 imp->imp_obd->obd_name,
1546 obd_uuid2str(&conn->c_remote_uuid),
1547 libcfs_nid2str(imp->imp_connection->c_peer.nid));
1550 if (imp->imp_state == LUSTRE_IMP_FULL) {
1551 wake_up_all(&imp->imp_recovery_waitq);
1552 ptlrpc_wake_delayed(imp);
1559 int ptlrpc_disconnect_import(struct obd_import *imp, int noclose)
1561 struct ptlrpc_request *req;
1565 if (imp->imp_obd->obd_force)
1566 GOTO(set_state, rc);
1568 switch (imp->imp_connect_op) {
1570 rq_opc = OST_DISCONNECT;
1573 rq_opc = MDS_DISCONNECT;
1576 rq_opc = MGS_DISCONNECT;
1580 CERROR("%s: don't know how to disconnect from %s "
1581 "(connect_op %d): rc = %d\n",
1582 imp->imp_obd->obd_name, obd2cli_tgt(imp->imp_obd),
1583 imp->imp_connect_op, rc);
1587 if (ptlrpc_import_in_recovery(imp)) {
1588 struct l_wait_info lwi;
1589 cfs_duration_t timeout;
1592 if (imp->imp_server_timeout)
1593 timeout = cfs_time_seconds(obd_timeout / 2);
1595 timeout = cfs_time_seconds(obd_timeout);
1597 int idx = import_at_get_index(imp,
1598 imp->imp_client->cli_request_portal);
1599 timeout = cfs_time_seconds(
1600 at_get(&imp->imp_at.iat_service_estimate[idx]));
1603 lwi = LWI_TIMEOUT_INTR(cfs_timeout_cap(timeout),
1604 back_to_sleep, LWI_ON_SIGNAL_NOOP, NULL);
1605 rc = l_wait_event(imp->imp_recovery_waitq,
1606 !ptlrpc_import_in_recovery(imp), &lwi);
1610 spin_lock(&imp->imp_lock);
1611 if (imp->imp_state != LUSTRE_IMP_FULL)
1613 spin_unlock(&imp->imp_lock);
1615 req = ptlrpc_request_alloc_pack(imp, &RQF_MDS_DISCONNECT,
1616 LUSTRE_OBD_VERSION, rq_opc);
1618 /* We are disconnecting, do not retry a failed DISCONNECT rpc if
1619 * it fails. We can get through the above with a down server
1620 * if the client doesn't know the server is gone yet. */
1621 req->rq_no_resend = 1;
1623 /* We want client umounts to happen quickly, no matter the
1625 req->rq_timeout = min_t(int, req->rq_timeout,
1626 INITIAL_CONNECT_TIMEOUT);
1628 IMPORT_SET_STATE(imp, LUSTRE_IMP_CONNECTING);
1629 req->rq_send_state = LUSTRE_IMP_CONNECTING;
1630 ptlrpc_request_set_replen(req);
1631 rc = ptlrpc_queue_wait(req);
1632 ptlrpc_req_finished(req);
1636 spin_lock(&imp->imp_lock);
1639 IMPORT_SET_STATE_NOLOCK(imp, LUSTRE_IMP_DISCON);
1641 IMPORT_SET_STATE_NOLOCK(imp, LUSTRE_IMP_CLOSED);
1642 memset(&imp->imp_remote_handle, 0, sizeof(imp->imp_remote_handle));
1643 spin_unlock(&imp->imp_lock);
1645 if (rc == -ETIMEDOUT || rc == -ENOTCONN || rc == -ESHUTDOWN)
1649 EXPORT_SYMBOL(ptlrpc_disconnect_import);
1651 void ptlrpc_cleanup_imp(struct obd_import *imp)
1655 spin_lock(&imp->imp_lock);
1656 IMPORT_SET_STATE_NOLOCK(imp, LUSTRE_IMP_CLOSED);
1657 imp->imp_generation++;
1658 spin_unlock(&imp->imp_lock);
1659 ptlrpc_abort_inflight(imp);
1664 /* Adaptive Timeout utils */
1665 extern unsigned int at_min, at_max, at_history;
1667 /* Update at_current with the specified value (bounded by at_min and at_max),
1668 * as well as the AT history "bins".
1669 * - Bin into timeslices using AT_BINS bins.
1670 * - This gives us a max of the last at_history seconds without the storage,
1671 * but still smoothing out a return to normalcy from a slow response.
1672 * - (E.g. remember the maximum latency in each minute of the last 4 minutes.)
1674 int at_measured(struct adaptive_timeout *at, unsigned int val)
1676 unsigned int old = at->at_current;
1677 time_t now = cfs_time_current_sec();
1678 time_t binlimit = max_t(time_t, at_history / AT_BINS, 1);
1681 CDEBUG(D_OTHER, "add %u to %p time=%lu v=%u (%u %u %u %u)\n",
1682 val, at, now - at->at_binstart, at->at_current,
1683 at->at_hist[0], at->at_hist[1], at->at_hist[2], at->at_hist[3]);
1686 /* 0's don't count, because we never want our timeout to
1687 drop to 0, and because 0 could mean an error */
1690 spin_lock(&at->at_lock);
1692 if (unlikely(at->at_binstart == 0)) {
1693 /* Special case to remove default from history */
1694 at->at_current = val;
1695 at->at_worst_ever = val;
1696 at->at_worst_time = now;
1697 at->at_hist[0] = val;
1698 at->at_binstart = now;
1699 } else if (now - at->at_binstart < binlimit ) {
1701 at->at_hist[0] = max(val, at->at_hist[0]);
1702 at->at_current = max(val, at->at_current);
1705 unsigned int maxv = val;
1706 /* move bins over */
1707 shift = (now - at->at_binstart) / binlimit;
1709 for(i = AT_BINS - 1; i >= 0; i--) {
1711 at->at_hist[i] = at->at_hist[i - shift];
1712 maxv = max(maxv, at->at_hist[i]);
1717 at->at_hist[0] = val;
1718 at->at_current = maxv;
1719 at->at_binstart += shift * binlimit;
1722 if (at->at_current > at->at_worst_ever) {
1723 at->at_worst_ever = at->at_current;
1724 at->at_worst_time = now;
1727 if (at->at_flags & AT_FLG_NOHIST)
1728 /* Only keep last reported val; keeping the rest of the history
1730 at->at_current = val;
1733 at->at_current = min(at->at_current, at_max);
1734 at->at_current = max(at->at_current, at_min);
1736 if (at->at_current != old)
1737 CDEBUG(D_OTHER, "AT %p change: old=%u new=%u delta=%d "
1738 "(val=%u) hist %u %u %u %u\n", at,
1739 old, at->at_current, at->at_current - old, val,
1740 at->at_hist[0], at->at_hist[1], at->at_hist[2],
1743 /* if we changed, report the old value */
1744 old = (at->at_current != old) ? old : 0;
1746 spin_unlock(&at->at_lock);
1750 /* Find the imp_at index for a given portal; assign if space available */
1751 int import_at_get_index(struct obd_import *imp, int portal)
1753 struct imp_at *at = &imp->imp_at;
1756 for (i = 0; i < IMP_AT_MAX_PORTALS; i++) {
1757 if (at->iat_portal[i] == portal)
1759 if (at->iat_portal[i] == 0)
1764 /* Not found in list, add it under a lock */
1765 spin_lock(&imp->imp_lock);
1767 /* Check unused under lock */
1768 for (; i < IMP_AT_MAX_PORTALS; i++) {
1769 if (at->iat_portal[i] == portal)
1771 if (at->iat_portal[i] == 0)
1776 /* Not enough portals? */
1777 LASSERT(i < IMP_AT_MAX_PORTALS);
1779 at->iat_portal[i] = portal;
1781 spin_unlock(&imp->imp_lock);