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, 2012, 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 "../include/obd_support.h"
44 #include "../include/lustre_ha.h"
45 #include "../include/lustre_net.h"
46 #include "../include/lustre_import.h"
47 #include "../include/lustre_export.h"
48 #include "../include/obd.h"
49 #include "../include/obd_cksum.h"
50 #include "../include/obd_class.h"
52 #include "ptlrpc_internal.h"
54 struct ptlrpc_connect_async_args {
55 __u64 pcaa_peer_committed;
56 int pcaa_initial_connect;
60 * Updates import \a imp current state to provided \a state value
61 * Helper function. Must be called under imp_lock.
63 static void __import_set_state(struct obd_import *imp,
64 enum lustre_imp_state state)
66 imp->imp_state = state;
67 imp->imp_state_hist[imp->imp_state_hist_idx].ish_state = state;
68 imp->imp_state_hist[imp->imp_state_hist_idx].ish_time =
70 imp->imp_state_hist_idx = (imp->imp_state_hist_idx + 1) %
74 /* A CLOSED import should remain so. */
75 #define IMPORT_SET_STATE_NOLOCK(imp, state) \
77 if (imp->imp_state != LUSTRE_IMP_CLOSED) { \
78 CDEBUG(D_HA, "%p %s: changing import state from %s to %s\n", \
79 imp, obd2cli_tgt(imp->imp_obd), \
80 ptlrpc_import_state_name(imp->imp_state), \
81 ptlrpc_import_state_name(state)); \
82 __import_set_state(imp, state); \
86 #define IMPORT_SET_STATE(imp, state) \
88 spin_lock(&imp->imp_lock); \
89 IMPORT_SET_STATE_NOLOCK(imp, state); \
90 spin_unlock(&imp->imp_lock); \
94 static int ptlrpc_connect_interpret(const struct lu_env *env,
95 struct ptlrpc_request *request,
97 int ptlrpc_import_recovery_state_machine(struct obd_import *imp);
99 /* Only this function is allowed to change the import state when it is
100 * CLOSED. I would rather refcount the import and free it after
101 * disconnection like we do with exports. To do that, the client_obd
102 * will need to save the peer info somewhere other than in the import,
104 int ptlrpc_init_import(struct obd_import *imp)
106 spin_lock(&imp->imp_lock);
108 imp->imp_generation++;
109 imp->imp_state = LUSTRE_IMP_NEW;
111 spin_unlock(&imp->imp_lock);
115 EXPORT_SYMBOL(ptlrpc_init_import);
117 #define UUID_STR "_UUID"
118 void deuuidify(char *uuid, const char *prefix, char **uuid_start, int *uuid_len)
120 *uuid_start = !prefix || strncmp(uuid, prefix, strlen(prefix))
121 ? uuid : uuid + strlen(prefix);
123 *uuid_len = strlen(*uuid_start);
125 if (*uuid_len < strlen(UUID_STR))
128 if (!strncmp(*uuid_start + *uuid_len - strlen(UUID_STR),
129 UUID_STR, strlen(UUID_STR)))
130 *uuid_len -= strlen(UUID_STR);
132 EXPORT_SYMBOL(deuuidify);
135 * Returns true if import was FULL, false if import was already not
137 * @imp - import to be disconnected
138 * @conn_cnt - connection count (epoch) of the request that timed out
139 * and caused the disconnection. In some cases, multiple
140 * inflight requests can fail to a single target (e.g. OST
141 * bulk requests) and if one has already caused a reconnection
142 * (increasing the import->conn_cnt) the older failure should
143 * not also cause a reconnection. If zero it forces a reconnect.
145 int ptlrpc_set_import_discon(struct obd_import *imp, __u32 conn_cnt)
149 spin_lock(&imp->imp_lock);
151 if (imp->imp_state == LUSTRE_IMP_FULL &&
152 (conn_cnt == 0 || conn_cnt == imp->imp_conn_cnt)) {
156 deuuidify(obd2cli_tgt(imp->imp_obd), NULL,
157 &target_start, &target_len);
159 if (imp->imp_replayable) {
160 LCONSOLE_WARN("%s: Connection to %.*s (at %s) was lost; in progress operations using this service will wait for recovery to complete\n",
161 imp->imp_obd->obd_name, target_len, target_start,
162 libcfs_nid2str(imp->imp_connection->c_peer.nid));
164 LCONSOLE_ERROR_MSG(0x166, "%s: Connection to %.*s (at %s) was lost; in progress operations using this service will fail\n",
165 imp->imp_obd->obd_name,
166 target_len, target_start,
167 libcfs_nid2str(imp->imp_connection->c_peer.nid));
169 IMPORT_SET_STATE_NOLOCK(imp, LUSTRE_IMP_DISCON);
170 spin_unlock(&imp->imp_lock);
172 if (obd_dump_on_timeout)
173 libcfs_debug_dumplog();
175 obd_import_event(imp->imp_obd, imp, IMP_EVENT_DISCON);
178 spin_unlock(&imp->imp_lock);
179 CDEBUG(D_HA, "%s: import %p already %s (conn %u, was %u): %s\n",
180 imp->imp_client->cli_name, imp,
181 (imp->imp_state == LUSTRE_IMP_FULL &&
182 imp->imp_conn_cnt > conn_cnt) ?
183 "reconnected" : "not connected", imp->imp_conn_cnt,
184 conn_cnt, ptlrpc_import_state_name(imp->imp_state));
190 /* Must be called with imp_lock held! */
191 static void ptlrpc_deactivate_and_unlock_import(struct obd_import *imp)
193 assert_spin_locked(&imp->imp_lock);
195 CDEBUG(D_HA, "setting import %s INVALID\n", obd2cli_tgt(imp->imp_obd));
196 imp->imp_invalid = 1;
197 imp->imp_generation++;
198 spin_unlock(&imp->imp_lock);
200 ptlrpc_abort_inflight(imp);
201 obd_import_event(imp->imp_obd, imp, IMP_EVENT_INACTIVE);
205 * This acts as a barrier; all existing requests are rejected, and
206 * no new requests will be accepted until the import is valid again.
208 void ptlrpc_deactivate_import(struct obd_import *imp)
210 spin_lock(&imp->imp_lock);
211 ptlrpc_deactivate_and_unlock_import(imp);
213 EXPORT_SYMBOL(ptlrpc_deactivate_import);
216 ptlrpc_inflight_deadline(struct ptlrpc_request *req, time_t now)
220 if (!(((req->rq_phase == RQ_PHASE_RPC) && !req->rq_waiting) ||
221 (req->rq_phase == RQ_PHASE_BULK) ||
222 (req->rq_phase == RQ_PHASE_NEW)))
225 if (req->rq_timedout)
228 if (req->rq_phase == RQ_PHASE_NEW)
231 dl = req->rq_deadline;
239 static unsigned int ptlrpc_inflight_timeout(struct obd_import *imp)
241 time_t now = get_seconds();
242 struct list_head *tmp, *n;
243 struct ptlrpc_request *req;
244 unsigned int timeout = 0;
246 spin_lock(&imp->imp_lock);
247 list_for_each_safe(tmp, n, &imp->imp_sending_list) {
248 req = list_entry(tmp, struct ptlrpc_request, rq_list);
249 timeout = max(ptlrpc_inflight_deadline(req, now), timeout);
251 spin_unlock(&imp->imp_lock);
256 * This function will invalidate the import, if necessary, then block
257 * for all the RPC completions, and finally notify the obd to
258 * invalidate its state (ie cancel locks, clear pending requests,
261 void ptlrpc_invalidate_import(struct obd_import *imp)
263 struct list_head *tmp, *n;
264 struct ptlrpc_request *req;
265 struct l_wait_info lwi;
266 unsigned int timeout;
269 atomic_inc(&imp->imp_inval_count);
271 if (!imp->imp_invalid || imp->imp_obd->obd_no_recov)
272 ptlrpc_deactivate_import(imp);
274 CFS_FAIL_TIMEOUT(OBD_FAIL_MGS_CONNECT_NET, 3 * cfs_fail_val / 2);
275 LASSERT(imp->imp_invalid);
277 /* Wait forever until inflight == 0. We really can't do it another
278 * way because in some cases we need to wait for very long reply
279 * unlink. We can't do anything before that because there is really
280 * no guarantee that some rdma transfer is not in progress right now. */
282 /* Calculate max timeout for waiting on rpcs to error
283 * out. Use obd_timeout if calculated value is smaller
285 if (!OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_REPL_UNLINK)) {
286 timeout = ptlrpc_inflight_timeout(imp);
287 timeout += timeout / 3;
290 timeout = obd_timeout;
292 /* decrease the interval to increase race condition */
297 "Sleeping %d sec for inflight to error out\n",
300 /* Wait for all requests to error out and call completion
301 * callbacks. Cap it at obd_timeout -- these should all
302 * have been locally cancelled by ptlrpc_abort_inflight. */
303 lwi = LWI_TIMEOUT_INTERVAL(
304 cfs_timeout_cap(cfs_time_seconds(timeout)),
305 (timeout > 1)?cfs_time_seconds(1):cfs_time_seconds(1)/2,
307 rc = l_wait_event(imp->imp_recovery_waitq,
308 (atomic_read(&imp->imp_inflight) == 0),
311 const char *cli_tgt = obd2cli_tgt(imp->imp_obd);
313 CERROR("%s: rc = %d waiting for callback (%d != 0)\n",
315 atomic_read(&imp->imp_inflight));
317 spin_lock(&imp->imp_lock);
318 if (atomic_read(&imp->imp_inflight) == 0) {
319 int count = atomic_read(&imp->imp_unregistering);
321 /* We know that "unregistering" rpcs only can
322 * survive in sending or delaying lists (they
323 * maybe waiting for long reply unlink in
324 * sluggish nets). Let's check this. If there
325 * is no inflight and unregistering != 0, this
327 LASSERTF(count == 0, "Some RPCs are still unregistering: %d\n",
330 /* Let's save one loop as soon as inflight have
331 * dropped to zero. No new inflights possible at
335 list_for_each_safe(tmp, n,
336 &imp->imp_sending_list) {
337 req = list_entry(tmp,
338 struct ptlrpc_request,
340 DEBUG_REQ(D_ERROR, req,
341 "still on sending list");
343 list_for_each_safe(tmp, n,
344 &imp->imp_delayed_list) {
345 req = list_entry(tmp,
346 struct ptlrpc_request,
348 DEBUG_REQ(D_ERROR, req,
349 "still on delayed list");
352 CERROR("%s: RPCs in \"%s\" phase found (%d). Network is sluggish? Waiting them to error out.\n",
354 ptlrpc_phase2str(RQ_PHASE_UNREGISTERING),
358 spin_unlock(&imp->imp_lock);
363 * Let's additionally check that no new rpcs added to import in
364 * "invalidate" state.
366 LASSERT(atomic_read(&imp->imp_inflight) == 0);
367 obd_import_event(imp->imp_obd, imp, IMP_EVENT_INVALIDATE);
368 sptlrpc_import_flush_all_ctx(imp);
370 atomic_dec(&imp->imp_inval_count);
371 wake_up_all(&imp->imp_recovery_waitq);
373 EXPORT_SYMBOL(ptlrpc_invalidate_import);
375 /* unset imp_invalid */
376 void ptlrpc_activate_import(struct obd_import *imp)
378 struct obd_device *obd = imp->imp_obd;
380 spin_lock(&imp->imp_lock);
381 if (imp->imp_deactive != 0) {
382 spin_unlock(&imp->imp_lock);
386 imp->imp_invalid = 0;
387 spin_unlock(&imp->imp_lock);
388 obd_import_event(obd, imp, IMP_EVENT_ACTIVE);
390 EXPORT_SYMBOL(ptlrpc_activate_import);
392 static void ptlrpc_pinger_force(struct obd_import *imp)
394 CDEBUG(D_HA, "%s: waking up pinger s:%s\n", obd2cli_tgt(imp->imp_obd),
395 ptlrpc_import_state_name(imp->imp_state));
397 spin_lock(&imp->imp_lock);
398 imp->imp_force_verify = 1;
399 spin_unlock(&imp->imp_lock);
401 if (imp->imp_state != LUSTRE_IMP_CONNECTING)
402 ptlrpc_pinger_wake_up();
405 void ptlrpc_fail_import(struct obd_import *imp, __u32 conn_cnt)
407 LASSERT(!imp->imp_dlm_fake);
409 if (ptlrpc_set_import_discon(imp, conn_cnt)) {
410 if (!imp->imp_replayable) {
411 CDEBUG(D_HA, "import %s@%s for %s not replayable, auto-deactivating\n",
412 obd2cli_tgt(imp->imp_obd),
413 imp->imp_connection->c_remote_uuid.uuid,
414 imp->imp_obd->obd_name);
415 ptlrpc_deactivate_import(imp);
418 ptlrpc_pinger_force(imp);
421 EXPORT_SYMBOL(ptlrpc_fail_import);
423 int ptlrpc_reconnect_import(struct obd_import *imp)
426 struct l_wait_info lwi;
427 int secs = cfs_time_seconds(obd_timeout);
430 ptlrpc_pinger_force(imp);
432 CDEBUG(D_HA, "%s: recovery started, waiting %u seconds\n",
433 obd2cli_tgt(imp->imp_obd), secs);
435 lwi = LWI_TIMEOUT(secs, NULL, NULL);
436 rc = l_wait_event(imp->imp_recovery_waitq,
437 !ptlrpc_import_in_recovery(imp), &lwi);
438 CDEBUG(D_HA, "%s: recovery finished s:%s\n", obd2cli_tgt(imp->imp_obd),
439 ptlrpc_import_state_name(imp->imp_state));
442 ptlrpc_set_import_discon(imp, 0);
443 /* Force a new connect attempt */
444 ptlrpc_invalidate_import(imp);
445 /* Do a fresh connect next time by zeroing the handle */
446 ptlrpc_disconnect_import(imp, 1);
447 /* Wait for all invalidate calls to finish */
448 if (atomic_read(&imp->imp_inval_count) > 0) {
450 struct l_wait_info lwi = LWI_INTR(LWI_ON_SIGNAL_NOOP, NULL);
451 rc = l_wait_event(imp->imp_recovery_waitq,
452 (atomic_read(&imp->imp_inval_count) == 0),
455 CERROR("Interrupted, inval=%d\n",
456 atomic_read(&imp->imp_inval_count));
459 /* Allow reconnect attempts */
460 imp->imp_obd->obd_no_recov = 0;
461 /* Remove 'invalid' flag */
462 ptlrpc_activate_import(imp);
463 /* Attempt a new connect */
464 ptlrpc_recover_import(imp, NULL, 0);
468 EXPORT_SYMBOL(ptlrpc_reconnect_import);
471 * Connection on import \a imp is changed to another one (if more than one is
472 * present). We typically chose connection that we have not tried to connect to
475 static int import_select_connection(struct obd_import *imp)
477 struct obd_import_conn *imp_conn = NULL, *conn;
478 struct obd_export *dlmexp;
480 int target_len, tried_all = 1;
482 spin_lock(&imp->imp_lock);
484 if (list_empty(&imp->imp_conn_list)) {
485 CERROR("%s: no connections available\n",
486 imp->imp_obd->obd_name);
487 spin_unlock(&imp->imp_lock);
491 list_for_each_entry(conn, &imp->imp_conn_list, oic_item) {
492 CDEBUG(D_HA, "%s: connect to NID %s last attempt %llu\n",
493 imp->imp_obd->obd_name,
494 libcfs_nid2str(conn->oic_conn->c_peer.nid),
495 conn->oic_last_attempt);
497 /* If we have not tried this connection since
498 the last successful attempt, go with this one */
499 if ((conn->oic_last_attempt == 0) ||
500 cfs_time_beforeq_64(conn->oic_last_attempt,
501 imp->imp_last_success_conn)) {
507 /* If all of the connections have already been tried
508 since the last successful connection; just choose the
509 least recently used */
512 else if (cfs_time_before_64(conn->oic_last_attempt,
513 imp_conn->oic_last_attempt))
517 /* if not found, simply choose the current one */
518 if (!imp_conn || imp->imp_force_reconnect) {
519 LASSERT(imp->imp_conn_current);
520 imp_conn = imp->imp_conn_current;
523 LASSERT(imp_conn->oic_conn);
525 /* If we've tried everything, and we're back to the beginning of the
526 list, increase our timeout and try again. It will be reset when
527 we do finally connect. (FIXME: really we should wait for all network
528 state associated with the last connection attempt to drain before
529 trying to reconnect on it.) */
530 if (tried_all && (imp->imp_conn_list.next == &imp_conn->oic_item)) {
531 struct adaptive_timeout *at = &imp->imp_at.iat_net_latency;
532 if (at_get(at) < CONNECTION_SWITCH_MAX) {
533 at_measured(at, at_get(at) + CONNECTION_SWITCH_INC);
534 if (at_get(at) > CONNECTION_SWITCH_MAX)
535 at_reset(at, CONNECTION_SWITCH_MAX);
537 LASSERT(imp_conn->oic_last_attempt);
538 CDEBUG(D_HA, "%s: tried all connections, increasing latency to %ds\n",
539 imp->imp_obd->obd_name, at_get(at));
542 imp_conn->oic_last_attempt = cfs_time_current_64();
544 /* switch connection, don't mind if it's same as the current one */
545 if (imp->imp_connection)
546 ptlrpc_connection_put(imp->imp_connection);
547 imp->imp_connection = ptlrpc_connection_addref(imp_conn->oic_conn);
549 dlmexp = class_conn2export(&imp->imp_dlm_handle);
550 LASSERT(dlmexp != NULL);
551 if (dlmexp->exp_connection)
552 ptlrpc_connection_put(dlmexp->exp_connection);
553 dlmexp->exp_connection = ptlrpc_connection_addref(imp_conn->oic_conn);
554 class_export_put(dlmexp);
556 if (imp->imp_conn_current != imp_conn) {
557 if (imp->imp_conn_current) {
558 deuuidify(obd2cli_tgt(imp->imp_obd), NULL,
559 &target_start, &target_len);
561 CDEBUG(D_HA, "%s: Connection changing to %.*s (at %s)\n",
562 imp->imp_obd->obd_name,
563 target_len, target_start,
564 libcfs_nid2str(imp_conn->oic_conn->c_peer.nid));
567 imp->imp_conn_current = imp_conn;
570 CDEBUG(D_HA, "%s: import %p using connection %s/%s\n",
571 imp->imp_obd->obd_name, imp, imp_conn->oic_uuid.uuid,
572 libcfs_nid2str(imp_conn->oic_conn->c_peer.nid));
574 spin_unlock(&imp->imp_lock);
580 * must be called under imp_lock
582 static int ptlrpc_first_transno(struct obd_import *imp, __u64 *transno)
584 struct ptlrpc_request *req;
585 struct list_head *tmp;
587 /* The requests in committed_list always have smaller transnos than
588 * the requests in replay_list */
589 if (!list_empty(&imp->imp_committed_list)) {
590 tmp = imp->imp_committed_list.next;
591 req = list_entry(tmp, struct ptlrpc_request, rq_replay_list);
592 *transno = req->rq_transno;
593 if (req->rq_transno == 0) {
594 DEBUG_REQ(D_ERROR, req,
595 "zero transno in committed_list");
600 if (!list_empty(&imp->imp_replay_list)) {
601 tmp = imp->imp_replay_list.next;
602 req = list_entry(tmp, struct ptlrpc_request, rq_replay_list);
603 *transno = req->rq_transno;
604 if (req->rq_transno == 0) {
605 DEBUG_REQ(D_ERROR, req, "zero transno in replay_list");
614 * Attempt to (re)connect import \a imp. This includes all preparations,
615 * initializing CONNECT RPC request and passing it to ptlrpcd for
617 * Returns 0 on success or error code.
619 int ptlrpc_connect_import(struct obd_import *imp)
621 struct obd_device *obd = imp->imp_obd;
622 int initial_connect = 0;
624 __u64 committed_before_reconnect = 0;
625 struct ptlrpc_request *request;
626 char *bufs[] = { NULL,
627 obd2cli_tgt(imp->imp_obd),
629 (char *)&imp->imp_dlm_handle,
630 (char *)&imp->imp_connect_data };
631 struct ptlrpc_connect_async_args *aa;
634 spin_lock(&imp->imp_lock);
635 if (imp->imp_state == LUSTRE_IMP_CLOSED) {
636 spin_unlock(&imp->imp_lock);
637 CERROR("can't connect to a closed import\n");
639 } else if (imp->imp_state == LUSTRE_IMP_FULL) {
640 spin_unlock(&imp->imp_lock);
641 CERROR("already connected\n");
643 } else if (imp->imp_state == LUSTRE_IMP_CONNECTING) {
644 spin_unlock(&imp->imp_lock);
645 CERROR("already connecting\n");
649 IMPORT_SET_STATE_NOLOCK(imp, LUSTRE_IMP_CONNECTING);
652 imp->imp_resend_replay = 0;
654 if (!lustre_handle_is_used(&imp->imp_remote_handle))
657 committed_before_reconnect = imp->imp_peer_committed_transno;
659 set_transno = ptlrpc_first_transno(imp,
660 &imp->imp_connect_data.ocd_transno);
661 spin_unlock(&imp->imp_lock);
663 rc = import_select_connection(imp);
667 rc = sptlrpc_import_sec_adapt(imp, NULL, NULL);
671 /* Reset connect flags to the originally requested flags, in case
672 * the server is updated on-the-fly we will get the new features. */
673 imp->imp_connect_data.ocd_connect_flags = imp->imp_connect_flags_orig;
674 /* Reset ocd_version each time so the server knows the exact versions */
675 imp->imp_connect_data.ocd_version = LUSTRE_VERSION_CODE;
676 imp->imp_msghdr_flags &= ~MSGHDR_AT_SUPPORT;
677 imp->imp_msghdr_flags &= ~MSGHDR_CKSUM_INCOMPAT18;
679 rc = obd_reconnect(NULL, imp->imp_obd->obd_self_export, obd,
680 &obd->obd_uuid, &imp->imp_connect_data, NULL);
684 request = ptlrpc_request_alloc(imp, &RQF_MDS_CONNECT);
685 if (request == NULL) {
690 rc = ptlrpc_request_bufs_pack(request, LUSTRE_OBD_VERSION,
691 imp->imp_connect_op, bufs, NULL);
693 ptlrpc_request_free(request);
697 /* Report the rpc service time to the server so that it knows how long
698 * to wait for clients to join recovery */
699 lustre_msg_set_service_time(request->rq_reqmsg,
700 at_timeout2est(request->rq_timeout));
702 /* The amount of time we give the server to process the connect req.
703 * import_select_connection will increase the net latency on
704 * repeated reconnect attempts to cover slow networks.
705 * We override/ignore the server rpc completion estimate here,
706 * which may be large if this is a reconnect attempt */
707 request->rq_timeout = INITIAL_CONNECT_TIMEOUT;
708 lustre_msg_set_timeout(request->rq_reqmsg, request->rq_timeout);
710 lustre_msg_add_op_flags(request->rq_reqmsg, MSG_CONNECT_NEXT_VER);
712 request->rq_no_resend = request->rq_no_delay = 1;
713 request->rq_send_state = LUSTRE_IMP_CONNECTING;
714 /* Allow a slightly larger reply for future growth compatibility */
715 req_capsule_set_size(&request->rq_pill, &RMF_CONNECT_DATA, RCL_SERVER,
716 sizeof(struct obd_connect_data)+16*sizeof(__u64));
717 ptlrpc_request_set_replen(request);
718 request->rq_interpret_reply = ptlrpc_connect_interpret;
720 CLASSERT(sizeof(*aa) <= sizeof(request->rq_async_args));
721 aa = ptlrpc_req_async_args(request);
722 memset(aa, 0, sizeof(*aa));
724 aa->pcaa_peer_committed = committed_before_reconnect;
725 aa->pcaa_initial_connect = initial_connect;
727 if (aa->pcaa_initial_connect) {
728 spin_lock(&imp->imp_lock);
729 imp->imp_replayable = 1;
730 spin_unlock(&imp->imp_lock);
731 lustre_msg_add_op_flags(request->rq_reqmsg,
732 MSG_CONNECT_INITIAL);
736 lustre_msg_add_op_flags(request->rq_reqmsg,
737 MSG_CONNECT_TRANSNO);
739 DEBUG_REQ(D_RPCTRACE, request, "(re)connect request (timeout %d)",
740 request->rq_timeout);
741 ptlrpcd_add_req(request, PDL_POLICY_ROUND, -1);
745 IMPORT_SET_STATE(imp, LUSTRE_IMP_DISCON);
750 EXPORT_SYMBOL(ptlrpc_connect_import);
752 static void ptlrpc_maybe_ping_import_soon(struct obd_import *imp)
756 spin_lock(&imp->imp_lock);
757 force_verify = imp->imp_force_verify != 0;
758 spin_unlock(&imp->imp_lock);
761 ptlrpc_pinger_wake_up();
764 static int ptlrpc_busy_reconnect(int rc)
766 return (rc == -EBUSY) || (rc == -EAGAIN);
770 * interpret_reply callback for connect RPCs.
771 * Looks into returned status of connect operation and decides
772 * what to do with the import - i.e enter recovery, promote it to
773 * full state for normal operations of disconnect it due to an error.
775 static int ptlrpc_connect_interpret(const struct lu_env *env,
776 struct ptlrpc_request *request,
779 struct ptlrpc_connect_async_args *aa = data;
780 struct obd_import *imp = request->rq_import;
781 struct client_obd *cli = &imp->imp_obd->u.cli;
782 struct lustre_handle old_hdl;
783 __u64 old_connect_flags;
785 struct obd_connect_data *ocd;
786 struct obd_export *exp;
789 spin_lock(&imp->imp_lock);
790 if (imp->imp_state == LUSTRE_IMP_CLOSED) {
791 imp->imp_connect_tried = 1;
792 spin_unlock(&imp->imp_lock);
797 /* if this reconnect to busy export - not need select new target
799 imp->imp_force_reconnect = ptlrpc_busy_reconnect(rc);
800 spin_unlock(&imp->imp_lock);
801 ptlrpc_maybe_ping_import_soon(imp);
804 spin_unlock(&imp->imp_lock);
806 LASSERT(imp->imp_conn_current);
808 msg_flags = lustre_msg_get_op_flags(request->rq_repmsg);
810 ret = req_capsule_get_size(&request->rq_pill, &RMF_CONNECT_DATA,
812 /* server replied obd_connect_data is always bigger */
813 ocd = req_capsule_server_sized_get(&request->rq_pill,
814 &RMF_CONNECT_DATA, ret);
817 CERROR("%s: no connect data from server\n",
818 imp->imp_obd->obd_name);
823 spin_lock(&imp->imp_lock);
825 /* All imports are pingable */
826 imp->imp_pingable = 1;
827 imp->imp_force_reconnect = 0;
828 imp->imp_force_verify = 0;
830 imp->imp_connect_data = *ocd;
832 CDEBUG(D_HA, "%s: connect to target with instance %u\n",
833 imp->imp_obd->obd_name, ocd->ocd_instance);
834 exp = class_conn2export(&imp->imp_dlm_handle);
836 spin_unlock(&imp->imp_lock);
838 /* check that server granted subset of flags we asked for. */
839 if ((ocd->ocd_connect_flags & imp->imp_connect_flags_orig) !=
840 ocd->ocd_connect_flags) {
841 CERROR("%s: Server didn't granted asked subset of flags: asked=%#llx grranted=%#llx\n",
842 imp->imp_obd->obd_name, imp->imp_connect_flags_orig,
843 ocd->ocd_connect_flags);
849 /* This could happen if export is cleaned during the
851 CERROR("%s: missing export after connect\n",
852 imp->imp_obd->obd_name);
856 old_connect_flags = exp_connect_flags(exp);
857 exp->exp_connect_data = *ocd;
858 imp->imp_obd->obd_self_export->exp_connect_data = *ocd;
859 class_export_put(exp);
861 obd_import_event(imp->imp_obd, imp, IMP_EVENT_OCD);
863 if (aa->pcaa_initial_connect) {
864 spin_lock(&imp->imp_lock);
865 if (msg_flags & MSG_CONNECT_REPLAYABLE) {
866 imp->imp_replayable = 1;
867 spin_unlock(&imp->imp_lock);
868 CDEBUG(D_HA, "connected to replayable target: %s\n",
869 obd2cli_tgt(imp->imp_obd));
871 imp->imp_replayable = 0;
872 spin_unlock(&imp->imp_lock);
875 /* if applies, adjust the imp->imp_msg_magic here
876 * according to reply flags */
878 imp->imp_remote_handle =
879 *lustre_msg_get_handle(request->rq_repmsg);
881 /* Initial connects are allowed for clients with non-random
882 * uuids when servers are in recovery. Simply signal the
883 * servers replay is complete and wait in REPLAY_WAIT. */
884 if (msg_flags & MSG_CONNECT_RECOVERING) {
885 CDEBUG(D_HA, "connect to %s during recovery\n",
886 obd2cli_tgt(imp->imp_obd));
887 IMPORT_SET_STATE(imp, LUSTRE_IMP_REPLAY_LOCKS);
889 IMPORT_SET_STATE(imp, LUSTRE_IMP_FULL);
890 ptlrpc_activate_import(imp);
897 /* Determine what recovery state to move the import to. */
898 if (MSG_CONNECT_RECONNECT & msg_flags) {
899 memset(&old_hdl, 0, sizeof(old_hdl));
900 if (!memcmp(&old_hdl, lustre_msg_get_handle(request->rq_repmsg),
902 LCONSOLE_WARN("Reconnect to %s (at @%s) failed due bad handle %#llx\n",
903 obd2cli_tgt(imp->imp_obd),
904 imp->imp_connection->c_remote_uuid.uuid,
905 imp->imp_dlm_handle.cookie);
910 if (memcmp(&imp->imp_remote_handle,
911 lustre_msg_get_handle(request->rq_repmsg),
912 sizeof(imp->imp_remote_handle))) {
913 int level = msg_flags & MSG_CONNECT_RECOVERING ?
916 /* Bug 16611/14775: if server handle have changed,
917 * that means some sort of disconnection happened.
918 * If the server is not in recovery, that also means it
919 * already erased all of our state because of previous
920 * eviction. If it is in recovery - we are safe to
921 * participate since we can reestablish all of our state
922 * with server again */
923 if ((MSG_CONNECT_RECOVERING & msg_flags)) {
924 CDEBUG(level, "%s@%s changed server handle from %#llx to %#llx but is still in recovery\n",
925 obd2cli_tgt(imp->imp_obd),
926 imp->imp_connection->c_remote_uuid.uuid,
927 imp->imp_remote_handle.cookie,
928 lustre_msg_get_handle(
929 request->rq_repmsg)->cookie);
931 LCONSOLE_WARN("Evicted from %s (at %s) after server handle changed from %#llx to %#llx\n",
932 obd2cli_tgt(imp->imp_obd),
933 imp->imp_connection-> \
935 imp->imp_remote_handle.cookie,
936 lustre_msg_get_handle(
937 request->rq_repmsg)->cookie);
941 imp->imp_remote_handle =
942 *lustre_msg_get_handle(request->rq_repmsg);
944 if (!(MSG_CONNECT_RECOVERING & msg_flags)) {
945 IMPORT_SET_STATE(imp, LUSTRE_IMP_EVICTED);
951 CDEBUG(D_HA, "reconnected to %s@%s after partition\n",
952 obd2cli_tgt(imp->imp_obd),
953 imp->imp_connection->c_remote_uuid.uuid);
956 if (imp->imp_invalid) {
957 CDEBUG(D_HA, "%s: reconnected but import is invalid; marking evicted\n",
958 imp->imp_obd->obd_name);
959 IMPORT_SET_STATE(imp, LUSTRE_IMP_EVICTED);
960 } else if (MSG_CONNECT_RECOVERING & msg_flags) {
961 CDEBUG(D_HA, "%s: reconnected to %s during replay\n",
962 imp->imp_obd->obd_name,
963 obd2cli_tgt(imp->imp_obd));
965 spin_lock(&imp->imp_lock);
966 imp->imp_resend_replay = 1;
967 spin_unlock(&imp->imp_lock);
969 IMPORT_SET_STATE(imp, LUSTRE_IMP_REPLAY);
971 IMPORT_SET_STATE(imp, LUSTRE_IMP_RECOVER);
973 } else if ((MSG_CONNECT_RECOVERING & msg_flags) && !imp->imp_invalid) {
974 LASSERT(imp->imp_replayable);
975 imp->imp_remote_handle =
976 *lustre_msg_get_handle(request->rq_repmsg);
977 imp->imp_last_replay_transno = 0;
978 IMPORT_SET_STATE(imp, LUSTRE_IMP_REPLAY);
980 DEBUG_REQ(D_HA, request, "%s: evicting (reconnect/recover flags not set: %x)",
981 imp->imp_obd->obd_name, msg_flags);
982 imp->imp_remote_handle =
983 *lustre_msg_get_handle(request->rq_repmsg);
984 IMPORT_SET_STATE(imp, LUSTRE_IMP_EVICTED);
987 /* Sanity checks for a reconnected import. */
988 if (!(imp->imp_replayable) != !(msg_flags & MSG_CONNECT_REPLAYABLE)) {
989 CERROR("imp_replayable flag does not match server after reconnect. We should LBUG right here.\n");
992 if (lustre_msg_get_last_committed(request->rq_repmsg) > 0 &&
993 lustre_msg_get_last_committed(request->rq_repmsg) <
994 aa->pcaa_peer_committed) {
995 CERROR("%s went back in time (transno %lld was previously committed, server now claims %lld)! See https://bugzilla.lustre.org/show_bug.cgi?id=9646\n",
996 obd2cli_tgt(imp->imp_obd), aa->pcaa_peer_committed,
997 lustre_msg_get_last_committed(request->rq_repmsg));
1001 rc = ptlrpc_import_recovery_state_machine(imp);
1003 if (rc == -ENOTCONN) {
1004 CDEBUG(D_HA, "evicted/aborted by %s@%s during recovery; invalidating and reconnecting\n",
1005 obd2cli_tgt(imp->imp_obd),
1006 imp->imp_connection->c_remote_uuid.uuid);
1007 ptlrpc_connect_import(imp);
1008 imp->imp_connect_tried = 1;
1013 spin_lock(&imp->imp_lock);
1014 list_del(&imp->imp_conn_current->oic_item);
1015 list_add(&imp->imp_conn_current->oic_item,
1016 &imp->imp_conn_list);
1017 imp->imp_last_success_conn =
1018 imp->imp_conn_current->oic_last_attempt;
1020 spin_unlock(&imp->imp_lock);
1022 if ((imp->imp_connect_flags_orig & OBD_CONNECT_IBITS) &&
1023 !(ocd->ocd_connect_flags & OBD_CONNECT_IBITS)) {
1024 LCONSOLE_WARN("%s: MDS %s does not support ibits lock, either very old or invalid: requested %llx, replied %llx\n",
1025 imp->imp_obd->obd_name,
1026 imp->imp_connection->c_remote_uuid.uuid,
1027 imp->imp_connect_flags_orig,
1028 ocd->ocd_connect_flags);
1033 if ((ocd->ocd_connect_flags & OBD_CONNECT_VERSION) &&
1034 (ocd->ocd_version > LUSTRE_VERSION_CODE +
1035 LUSTRE_VERSION_OFFSET_WARN ||
1036 ocd->ocd_version < LUSTRE_VERSION_CODE -
1037 LUSTRE_VERSION_OFFSET_WARN)) {
1038 /* Sigh, some compilers do not like #ifdef in the middle
1039 of macro arguments */
1040 const char *older = "older. Consider upgrading server or downgrading client"
1042 const char *newer = "newer than client version. Consider upgrading client"
1045 LCONSOLE_WARN("Server %s version (%d.%d.%d.%d) is much %s (%s)\n",
1046 obd2cli_tgt(imp->imp_obd),
1047 OBD_OCD_VERSION_MAJOR(ocd->ocd_version),
1048 OBD_OCD_VERSION_MINOR(ocd->ocd_version),
1049 OBD_OCD_VERSION_PATCH(ocd->ocd_version),
1050 OBD_OCD_VERSION_FIX(ocd->ocd_version),
1051 ocd->ocd_version > LUSTRE_VERSION_CODE ?
1052 newer : older, LUSTRE_VERSION_STRING);
1055 #if LUSTRE_VERSION_CODE < OBD_OCD_VERSION(3, 2, 50, 0)
1056 /* Check if server has LU-1252 fix applied to not always swab
1057 * the IR MNE entries. Do this only once per connection. This
1058 * fixup is version-limited, because we don't want to carry the
1059 * OBD_CONNECT_MNE_SWAB flag around forever, just so long as we
1060 * need interop with unpatched 2.2 servers. For newer servers,
1061 * the client will do MNE swabbing only as needed. LU-1644 */
1062 if (unlikely((ocd->ocd_connect_flags & OBD_CONNECT_VERSION) &&
1063 !(ocd->ocd_connect_flags & OBD_CONNECT_MNE_SWAB) &&
1064 OBD_OCD_VERSION_MAJOR(ocd->ocd_version) == 2 &&
1065 OBD_OCD_VERSION_MINOR(ocd->ocd_version) == 2 &&
1066 OBD_OCD_VERSION_PATCH(ocd->ocd_version) < 55 &&
1067 strcmp(imp->imp_obd->obd_type->typ_name,
1068 LUSTRE_MGC_NAME) == 0))
1069 imp->imp_need_mne_swab = 1;
1070 else /* clear if server was upgraded since last connect */
1071 imp->imp_need_mne_swab = 0;
1073 #warning "LU-1644: Remove old OBD_CONNECT_MNE_SWAB fixup and imp_need_mne_swab"
1076 if (ocd->ocd_connect_flags & OBD_CONNECT_CKSUM) {
1077 /* We sent to the server ocd_cksum_types with bits set
1078 * for algorithms we understand. The server masked off
1079 * the checksum types it doesn't support */
1080 if ((ocd->ocd_cksum_types &
1081 cksum_types_supported_client()) == 0) {
1082 LCONSOLE_WARN("The negotiation of the checksum algorithm to use with server %s failed (%x/%x), disabling checksums\n",
1083 obd2cli_tgt(imp->imp_obd),
1084 ocd->ocd_cksum_types,
1085 cksum_types_supported_client());
1086 cli->cl_checksum = 0;
1087 cli->cl_supp_cksum_types = OBD_CKSUM_ADLER;
1089 cli->cl_supp_cksum_types = ocd->ocd_cksum_types;
1092 /* The server does not support OBD_CONNECT_CKSUM.
1093 * Enforce ADLER for backward compatibility*/
1094 cli->cl_supp_cksum_types = OBD_CKSUM_ADLER;
1096 cli->cl_cksum_type = cksum_type_select(cli->cl_supp_cksum_types);
1098 if (ocd->ocd_connect_flags & OBD_CONNECT_BRW_SIZE)
1099 cli->cl_max_pages_per_rpc =
1100 min(ocd->ocd_brw_size >> PAGE_CACHE_SHIFT,
1101 cli->cl_max_pages_per_rpc);
1102 else if (imp->imp_connect_op == MDS_CONNECT ||
1103 imp->imp_connect_op == MGS_CONNECT)
1104 cli->cl_max_pages_per_rpc = 1;
1106 /* Reset ns_connect_flags only for initial connect. It might be
1107 * changed in while using FS and if we reset it in reconnect
1108 * this leads to losing user settings done before such as
1109 * disable lru_resize, etc. */
1110 if (old_connect_flags != exp_connect_flags(exp) ||
1111 aa->pcaa_initial_connect) {
1112 CDEBUG(D_HA, "%s: Resetting ns_connect_flags to server flags: %#llx\n",
1113 imp->imp_obd->obd_name, ocd->ocd_connect_flags);
1114 imp->imp_obd->obd_namespace->ns_connect_flags =
1115 ocd->ocd_connect_flags;
1116 imp->imp_obd->obd_namespace->ns_orig_connect_flags =
1117 ocd->ocd_connect_flags;
1120 if ((ocd->ocd_connect_flags & OBD_CONNECT_AT) &&
1121 (imp->imp_msg_magic == LUSTRE_MSG_MAGIC_V2))
1122 /* We need a per-message support flag, because
1123 a. we don't know if the incoming connect reply
1124 supports AT or not (in reply_in_callback)
1126 b. failovered server means export and flags are gone
1127 (in ptlrpc_send_reply).
1128 Can only be set when we know AT is supported at
1130 imp->imp_msghdr_flags |= MSGHDR_AT_SUPPORT;
1132 imp->imp_msghdr_flags &= ~MSGHDR_AT_SUPPORT;
1134 if ((ocd->ocd_connect_flags & OBD_CONNECT_FULL20) &&
1135 (imp->imp_msg_magic == LUSTRE_MSG_MAGIC_V2))
1136 imp->imp_msghdr_flags |= MSGHDR_CKSUM_INCOMPAT18;
1138 imp->imp_msghdr_flags &= ~MSGHDR_CKSUM_INCOMPAT18;
1140 LASSERT((cli->cl_max_pages_per_rpc <= PTLRPC_MAX_BRW_PAGES) &&
1141 (cli->cl_max_pages_per_rpc > 0));
1145 imp->imp_connect_tried = 1;
1148 IMPORT_SET_STATE(imp, LUSTRE_IMP_DISCON);
1149 if (rc == -EACCES) {
1151 * Give up trying to reconnect
1152 * EACCES means client has no permission for connection
1154 imp->imp_obd->obd_no_recov = 1;
1155 ptlrpc_deactivate_import(imp);
1158 if (rc == -EPROTO) {
1159 struct obd_connect_data *ocd;
1161 /* reply message might not be ready */
1162 if (request->rq_repmsg == NULL)
1165 ocd = req_capsule_server_get(&request->rq_pill,
1168 (ocd->ocd_connect_flags & OBD_CONNECT_VERSION) &&
1169 (ocd->ocd_version != LUSTRE_VERSION_CODE)) {
1171 * Actually servers are only supposed to refuse
1172 * connection from liblustre clients, so we
1173 * should never see this from VFS context
1175 LCONSOLE_ERROR_MSG(0x16a, "Server %s version (%d.%d.%d.%d) refused connection from this client with an incompatible version (%s). Client must be recompiled\n",
1176 obd2cli_tgt(imp->imp_obd),
1177 OBD_OCD_VERSION_MAJOR(ocd->ocd_version),
1178 OBD_OCD_VERSION_MINOR(ocd->ocd_version),
1179 OBD_OCD_VERSION_PATCH(ocd->ocd_version),
1180 OBD_OCD_VERSION_FIX(ocd->ocd_version),
1181 LUSTRE_VERSION_STRING);
1182 ptlrpc_deactivate_import(imp);
1183 IMPORT_SET_STATE(imp, LUSTRE_IMP_CLOSED);
1188 ptlrpc_maybe_ping_import_soon(imp);
1190 CDEBUG(D_HA, "recovery of %s on %s failed (%d)\n",
1191 obd2cli_tgt(imp->imp_obd),
1192 (char *)imp->imp_connection->c_remote_uuid.uuid, rc);
1195 wake_up_all(&imp->imp_recovery_waitq);
1200 * interpret callback for "completed replay" RPCs.
1201 * \see signal_completed_replay
1203 static int completed_replay_interpret(const struct lu_env *env,
1204 struct ptlrpc_request *req,
1207 atomic_dec(&req->rq_import->imp_replay_inflight);
1208 if (req->rq_status == 0 &&
1209 !req->rq_import->imp_vbr_failed) {
1210 ptlrpc_import_recovery_state_machine(req->rq_import);
1212 if (req->rq_import->imp_vbr_failed) {
1214 "%s: version recovery fails, reconnecting\n",
1215 req->rq_import->imp_obd->obd_name);
1217 CDEBUG(D_HA, "%s: LAST_REPLAY message error: %d, reconnecting\n",
1218 req->rq_import->imp_obd->obd_name,
1221 ptlrpc_connect_import(req->rq_import);
1228 * Let server know that we have no requests to replay anymore.
1229 * Achieved by just sending a PING request
1231 static int signal_completed_replay(struct obd_import *imp)
1233 struct ptlrpc_request *req;
1235 if (unlikely(OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_FINISH_REPLAY)))
1238 LASSERT(atomic_read(&imp->imp_replay_inflight) == 0);
1239 atomic_inc(&imp->imp_replay_inflight);
1241 req = ptlrpc_request_alloc_pack(imp, &RQF_OBD_PING, LUSTRE_OBD_VERSION,
1244 atomic_dec(&imp->imp_replay_inflight);
1248 ptlrpc_request_set_replen(req);
1249 req->rq_send_state = LUSTRE_IMP_REPLAY_WAIT;
1250 lustre_msg_add_flags(req->rq_reqmsg,
1251 MSG_LOCK_REPLAY_DONE | MSG_REQ_REPLAY_DONE);
1253 req->rq_timeout *= 3;
1254 req->rq_interpret_reply = completed_replay_interpret;
1256 ptlrpcd_add_req(req, PDL_POLICY_ROUND, -1);
1261 * In kernel code all import invalidation happens in its own
1262 * separate thread, so that whatever application happened to encounter
1263 * a problem could still be killed or otherwise continue
1265 static int ptlrpc_invalidate_import_thread(void *data)
1267 struct obd_import *imp = data;
1269 unshare_fs_struct();
1271 CDEBUG(D_HA, "thread invalidate import %s to %s@%s\n",
1272 imp->imp_obd->obd_name, obd2cli_tgt(imp->imp_obd),
1273 imp->imp_connection->c_remote_uuid.uuid);
1275 ptlrpc_invalidate_import(imp);
1277 if (obd_dump_on_eviction) {
1278 CERROR("dump the log upon eviction\n");
1279 libcfs_debug_dumplog();
1282 IMPORT_SET_STATE(imp, LUSTRE_IMP_RECOVER);
1283 ptlrpc_import_recovery_state_machine(imp);
1285 class_import_put(imp);
1290 * This is the state machine for client-side recovery on import.
1292 * Typically we have two possibly paths. If we came to server and it is not
1293 * in recovery, we just enter IMP_EVICTED state, invalidate our import
1294 * state and reconnect from scratch.
1295 * If we came to server that is in recovery, we enter IMP_REPLAY import state.
1296 * We go through our list of requests to replay and send them to server one by
1298 * After sending all request from the list we change import state to
1299 * IMP_REPLAY_LOCKS and re-request all the locks we believe we have from server
1300 * and also all the locks we don't yet have and wait for server to grant us.
1301 * After that we send a special "replay completed" request and change import
1302 * state to IMP_REPLAY_WAIT.
1303 * Upon receiving reply to that "replay completed" RPC we enter IMP_RECOVER
1304 * state and resend all requests from sending list.
1305 * After that we promote import to FULL state and send all delayed requests
1306 * and import is fully operational after that.
1309 int ptlrpc_import_recovery_state_machine(struct obd_import *imp)
1316 if (imp->imp_state == LUSTRE_IMP_EVICTED) {
1317 deuuidify(obd2cli_tgt(imp->imp_obd), NULL,
1318 &target_start, &target_len);
1319 /* Don't care about MGC eviction */
1320 if (strcmp(imp->imp_obd->obd_type->typ_name,
1321 LUSTRE_MGC_NAME) != 0) {
1322 LCONSOLE_ERROR_MSG(0x167, "%s: This client was evicted by %.*s; in progress operations using this service will fail.\n",
1323 imp->imp_obd->obd_name, target_len,
1326 CDEBUG(D_HA, "evicted from %s@%s; invalidating\n",
1327 obd2cli_tgt(imp->imp_obd),
1328 imp->imp_connection->c_remote_uuid.uuid);
1329 /* reset vbr_failed flag upon eviction */
1330 spin_lock(&imp->imp_lock);
1331 imp->imp_vbr_failed = 0;
1332 spin_unlock(&imp->imp_lock);
1335 struct task_struct *task;
1336 /* bug 17802: XXX client_disconnect_export vs connect request
1337 * race. if client will evicted at this time, we start
1338 * invalidate thread without reference to import and import can
1339 * be freed at same time. */
1340 class_import_get(imp);
1341 task = kthread_run(ptlrpc_invalidate_import_thread, imp,
1344 class_import_put(imp);
1345 CERROR("error starting invalidate thread: %d\n", rc);
1354 if (imp->imp_state == LUSTRE_IMP_REPLAY) {
1355 CDEBUG(D_HA, "replay requested by %s\n",
1356 obd2cli_tgt(imp->imp_obd));
1357 rc = ptlrpc_replay_next(imp, &inflight);
1358 if (inflight == 0 &&
1359 atomic_read(&imp->imp_replay_inflight) == 0) {
1360 IMPORT_SET_STATE(imp, LUSTRE_IMP_REPLAY_LOCKS);
1361 rc = ldlm_replay_locks(imp);
1368 if (imp->imp_state == LUSTRE_IMP_REPLAY_LOCKS) {
1369 if (atomic_read(&imp->imp_replay_inflight) == 0) {
1370 IMPORT_SET_STATE(imp, LUSTRE_IMP_REPLAY_WAIT);
1371 rc = signal_completed_replay(imp);
1378 if (imp->imp_state == LUSTRE_IMP_REPLAY_WAIT) {
1379 if (atomic_read(&imp->imp_replay_inflight) == 0) {
1380 IMPORT_SET_STATE(imp, LUSTRE_IMP_RECOVER);
1384 if (imp->imp_state == LUSTRE_IMP_RECOVER) {
1385 CDEBUG(D_HA, "reconnected to %s@%s\n",
1386 obd2cli_tgt(imp->imp_obd),
1387 imp->imp_connection->c_remote_uuid.uuid);
1389 rc = ptlrpc_resend(imp);
1392 IMPORT_SET_STATE(imp, LUSTRE_IMP_FULL);
1393 ptlrpc_activate_import(imp);
1395 deuuidify(obd2cli_tgt(imp->imp_obd), NULL,
1396 &target_start, &target_len);
1397 LCONSOLE_INFO("%s: Connection restored to %.*s (at %s)\n",
1398 imp->imp_obd->obd_name,
1399 target_len, target_start,
1400 libcfs_nid2str(imp->imp_connection->c_peer.nid));
1403 if (imp->imp_state == LUSTRE_IMP_FULL) {
1404 wake_up_all(&imp->imp_recovery_waitq);
1405 ptlrpc_wake_delayed(imp);
1412 int ptlrpc_disconnect_import(struct obd_import *imp, int noclose)
1414 struct ptlrpc_request *req;
1417 if (imp->imp_obd->obd_force)
1420 switch (imp->imp_connect_op) {
1422 rq_opc = OST_DISCONNECT;
1425 rq_opc = MDS_DISCONNECT;
1428 rq_opc = MGS_DISCONNECT;
1432 CERROR("%s: don't know how to disconnect from %s (connect_op %d): rc = %d\n",
1433 imp->imp_obd->obd_name, obd2cli_tgt(imp->imp_obd),
1434 imp->imp_connect_op, rc);
1438 if (ptlrpc_import_in_recovery(imp)) {
1439 struct l_wait_info lwi;
1443 if (imp->imp_server_timeout)
1444 timeout = cfs_time_seconds(obd_timeout / 2);
1446 timeout = cfs_time_seconds(obd_timeout);
1448 int idx = import_at_get_index(imp,
1449 imp->imp_client->cli_request_portal);
1450 timeout = cfs_time_seconds(
1451 at_get(&imp->imp_at.iat_service_estimate[idx]));
1454 lwi = LWI_TIMEOUT_INTR(cfs_timeout_cap(timeout),
1455 back_to_sleep, LWI_ON_SIGNAL_NOOP, NULL);
1456 rc = l_wait_event(imp->imp_recovery_waitq,
1457 !ptlrpc_import_in_recovery(imp), &lwi);
1461 spin_lock(&imp->imp_lock);
1462 if (imp->imp_state != LUSTRE_IMP_FULL)
1464 spin_unlock(&imp->imp_lock);
1466 req = ptlrpc_request_alloc_pack(imp, &RQF_MDS_DISCONNECT,
1467 LUSTRE_OBD_VERSION, rq_opc);
1469 /* We are disconnecting, do not retry a failed DISCONNECT rpc if
1470 * it fails. We can get through the above with a down server
1471 * if the client doesn't know the server is gone yet. */
1472 req->rq_no_resend = 1;
1474 /* We want client umounts to happen quickly, no matter the
1476 req->rq_timeout = min_t(int, req->rq_timeout,
1477 INITIAL_CONNECT_TIMEOUT);
1479 IMPORT_SET_STATE(imp, LUSTRE_IMP_CONNECTING);
1480 req->rq_send_state = LUSTRE_IMP_CONNECTING;
1481 ptlrpc_request_set_replen(req);
1482 rc = ptlrpc_queue_wait(req);
1483 ptlrpc_req_finished(req);
1487 spin_lock(&imp->imp_lock);
1490 IMPORT_SET_STATE_NOLOCK(imp, LUSTRE_IMP_DISCON);
1492 IMPORT_SET_STATE_NOLOCK(imp, LUSTRE_IMP_CLOSED);
1493 memset(&imp->imp_remote_handle, 0, sizeof(imp->imp_remote_handle));
1494 spin_unlock(&imp->imp_lock);
1496 if (rc == -ETIMEDOUT || rc == -ENOTCONN || rc == -ESHUTDOWN)
1501 EXPORT_SYMBOL(ptlrpc_disconnect_import);
1503 void ptlrpc_cleanup_imp(struct obd_import *imp)
1505 spin_lock(&imp->imp_lock);
1506 IMPORT_SET_STATE_NOLOCK(imp, LUSTRE_IMP_CLOSED);
1507 imp->imp_generation++;
1508 spin_unlock(&imp->imp_lock);
1509 ptlrpc_abort_inflight(imp);
1511 EXPORT_SYMBOL(ptlrpc_cleanup_imp);
1513 /* Adaptive Timeout utils */
1514 extern unsigned int at_min, at_max, at_history;
1516 /* Bin into timeslices using AT_BINS bins.
1517 This gives us a max of the last binlimit*AT_BINS secs without the storage,
1518 but still smoothing out a return to normalcy from a slow response.
1519 (E.g. remember the maximum latency in each minute of the last 4 minutes.) */
1520 int at_measured(struct adaptive_timeout *at, unsigned int val)
1522 unsigned int old = at->at_current;
1523 time_t now = get_seconds();
1524 time_t binlimit = max_t(time_t, at_history / AT_BINS, 1);
1527 CDEBUG(D_OTHER, "add %u to %p time=%lu v=%u (%u %u %u %u)\n",
1528 val, at, now - at->at_binstart, at->at_current,
1529 at->at_hist[0], at->at_hist[1], at->at_hist[2], at->at_hist[3]);
1532 /* 0's don't count, because we never want our timeout to
1533 drop to 0, and because 0 could mean an error */
1536 spin_lock(&at->at_lock);
1538 if (unlikely(at->at_binstart == 0)) {
1539 /* Special case to remove default from history */
1540 at->at_current = val;
1541 at->at_worst_ever = val;
1542 at->at_worst_time = now;
1543 at->at_hist[0] = val;
1544 at->at_binstart = now;
1545 } else if (now - at->at_binstart < binlimit) {
1547 at->at_hist[0] = max(val, at->at_hist[0]);
1548 at->at_current = max(val, at->at_current);
1551 unsigned int maxv = val;
1552 /* move bins over */
1553 shift = (now - at->at_binstart) / binlimit;
1555 for (i = AT_BINS - 1; i >= 0; i--) {
1557 at->at_hist[i] = at->at_hist[i - shift];
1558 maxv = max(maxv, at->at_hist[i]);
1563 at->at_hist[0] = val;
1564 at->at_current = maxv;
1565 at->at_binstart += shift * binlimit;
1568 if (at->at_current > at->at_worst_ever) {
1569 at->at_worst_ever = at->at_current;
1570 at->at_worst_time = now;
1573 if (at->at_flags & AT_FLG_NOHIST)
1574 /* Only keep last reported val; keeping the rest of the history
1576 at->at_current = val;
1579 at->at_current = min(at->at_current, at_max);
1580 at->at_current = max(at->at_current, at_min);
1582 if (at->at_current != old)
1583 CDEBUG(D_OTHER, "AT %p change: old=%u new=%u delta=%d (val=%u) hist %u %u %u %u\n",
1585 old, at->at_current, at->at_current - old, val,
1586 at->at_hist[0], at->at_hist[1], at->at_hist[2],
1589 /* if we changed, report the old value */
1590 old = (at->at_current != old) ? old : 0;
1592 spin_unlock(&at->at_lock);
1596 /* Find the imp_at index for a given portal; assign if space available */
1597 int import_at_get_index(struct obd_import *imp, int portal)
1599 struct imp_at *at = &imp->imp_at;
1602 for (i = 0; i < IMP_AT_MAX_PORTALS; i++) {
1603 if (at->iat_portal[i] == portal)
1605 if (at->iat_portal[i] == 0)
1610 /* Not found in list, add it under a lock */
1611 spin_lock(&imp->imp_lock);
1613 /* Check unused under lock */
1614 for (; i < IMP_AT_MAX_PORTALS; i++) {
1615 if (at->iat_portal[i] == portal)
1617 if (at->iat_portal[i] == 0)
1622 /* Not enough portals? */
1623 LASSERT(i < IMP_AT_MAX_PORTALS);
1625 at->iat_portal[i] = portal;
1627 spin_unlock(&imp->imp_lock);