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, 2014, 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); \
107 static int ptlrpc_connect_interpret(const struct lu_env *env,
108 struct ptlrpc_request *request,
109 void * data, int rc);
110 int ptlrpc_import_recovery_state_machine(struct obd_import *imp);
112 /* Only this function is allowed to change the import state when it is
113 * CLOSED. I would rather refcount the import and free it after
114 * disconnection like we do with exports. To do that, the client_obd
115 * will need to save the peer info somewhere other than in the import,
117 int ptlrpc_init_import(struct obd_import *imp)
119 spin_lock(&imp->imp_lock);
121 imp->imp_generation++;
122 imp->imp_state = LUSTRE_IMP_NEW;
124 spin_unlock(&imp->imp_lock);
128 EXPORT_SYMBOL(ptlrpc_init_import);
130 #define UUID_STR "_UUID"
131 void deuuidify(char *uuid, const char *prefix, char **uuid_start, int *uuid_len)
133 *uuid_start = !prefix || strncmp(uuid, prefix, strlen(prefix))
134 ? uuid : uuid + strlen(prefix);
136 *uuid_len = strlen(*uuid_start);
138 if (*uuid_len < strlen(UUID_STR))
141 if (!strncmp(*uuid_start + *uuid_len - strlen(UUID_STR),
142 UUID_STR, strlen(UUID_STR)))
143 *uuid_len -= strlen(UUID_STR);
147 * Returns true if import was FULL, false if import was already not
149 * @imp - import to be disconnected
150 * @conn_cnt - connection count (epoch) of the request that timed out
151 * and caused the disconnection. In some cases, multiple
152 * inflight requests can fail to a single target (e.g. OST
153 * bulk requests) and if one has already caused a reconnection
154 * (increasing the import->conn_cnt) the older failure should
155 * not also cause a reconnection. If zero it forces a reconnect.
157 int ptlrpc_set_import_discon(struct obd_import *imp, __u32 conn_cnt)
161 spin_lock(&imp->imp_lock);
163 if (imp->imp_state == LUSTRE_IMP_FULL &&
164 (conn_cnt == 0 || conn_cnt == imp->imp_conn_cnt)) {
168 deuuidify(obd2cli_tgt(imp->imp_obd), NULL,
169 &target_start, &target_len);
171 if (imp->imp_replayable) {
172 LCONSOLE_WARN("%s: Connection to %.*s (at %s) was "
173 "lost; in progress operations using this "
174 "service will wait for recovery to complete\n",
175 imp->imp_obd->obd_name, target_len, target_start,
176 libcfs_nid2str(imp->imp_connection->c_peer.nid));
178 LCONSOLE_ERROR_MSG(0x166, "%s: Connection to "
179 "%.*s (at %s) was lost; in progress "
180 "operations using this service will fail\n",
181 imp->imp_obd->obd_name,
182 target_len, target_start,
183 libcfs_nid2str(imp->imp_connection->c_peer.nid));
185 IMPORT_SET_STATE_NOLOCK(imp, LUSTRE_IMP_DISCON);
186 spin_unlock(&imp->imp_lock);
188 if (obd_dump_on_timeout)
189 libcfs_debug_dumplog();
191 obd_import_event(imp->imp_obd, imp, IMP_EVENT_DISCON);
194 spin_unlock(&imp->imp_lock);
195 CDEBUG(D_HA, "%s: import %p already %s (conn %u, was %u): %s\n",
196 imp->imp_client->cli_name, imp,
197 (imp->imp_state == LUSTRE_IMP_FULL &&
198 imp->imp_conn_cnt > conn_cnt) ?
199 "reconnected" : "not connected", imp->imp_conn_cnt,
200 conn_cnt, ptlrpc_import_state_name(imp->imp_state));
206 /* Must be called with imp_lock held! */
207 static void ptlrpc_deactivate_and_unlock_import(struct obd_import *imp)
210 assert_spin_locked(&imp->imp_lock);
212 CDEBUG(D_HA, "setting import %s INVALID\n", obd2cli_tgt(imp->imp_obd));
213 imp->imp_invalid = 1;
214 imp->imp_generation++;
215 spin_unlock(&imp->imp_lock);
217 ptlrpc_abort_inflight(imp);
218 obd_import_event(imp->imp_obd, imp, IMP_EVENT_INACTIVE);
224 * This acts as a barrier; all existing requests are rejected, and
225 * no new requests will be accepted until the import is valid again.
227 void ptlrpc_deactivate_import(struct obd_import *imp)
229 spin_lock(&imp->imp_lock);
230 ptlrpc_deactivate_and_unlock_import(imp);
232 EXPORT_SYMBOL(ptlrpc_deactivate_import);
235 ptlrpc_inflight_deadline(struct ptlrpc_request *req, time_t now)
239 if (!(((req->rq_phase == RQ_PHASE_RPC) && !req->rq_waiting) ||
240 (req->rq_phase == RQ_PHASE_BULK) ||
241 (req->rq_phase == RQ_PHASE_NEW)))
244 if (req->rq_timedout)
247 if (req->rq_phase == RQ_PHASE_NEW)
250 dl = req->rq_deadline;
258 static unsigned int ptlrpc_inflight_timeout(struct obd_import *imp)
260 time_t now = cfs_time_current_sec();
261 struct list_head *tmp, *n;
262 struct ptlrpc_request *req;
263 unsigned int timeout = 0;
265 spin_lock(&imp->imp_lock);
266 list_for_each_safe(tmp, n, &imp->imp_sending_list) {
267 req = list_entry(tmp, struct ptlrpc_request, rq_list);
268 timeout = max(ptlrpc_inflight_deadline(req, now), timeout);
270 spin_unlock(&imp->imp_lock);
275 * This function will invalidate the import, if necessary, then block
276 * for all the RPC completions, and finally notify the obd to
277 * invalidate its state (ie cancel locks, clear pending requests,
280 void ptlrpc_invalidate_import(struct obd_import *imp)
282 struct list_head *tmp, *n;
283 struct ptlrpc_request *req;
284 struct l_wait_info lwi;
285 unsigned int timeout;
288 atomic_inc(&imp->imp_inval_count);
290 if (!imp->imp_invalid || imp->imp_obd->obd_no_recov)
291 ptlrpc_deactivate_import(imp);
293 CFS_FAIL_TIMEOUT(OBD_FAIL_MGS_CONNECT_NET, 3 * cfs_fail_val / 2);
294 LASSERT(imp->imp_invalid);
296 /* Wait forever until inflight == 0. We really can't do it another
297 * way because in some cases we need to wait for very long reply
298 * unlink. We can't do anything before that because there is really
299 * no guarantee that some rdma transfer is not in progress right now. */
301 /* Calculate max timeout for waiting on rpcs to error
302 * out. Use obd_timeout if calculated value is smaller
304 if (!OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_REPL_UNLINK)) {
305 timeout = ptlrpc_inflight_timeout(imp);
306 timeout += timeout / 3;
309 timeout = obd_timeout;
311 /* decrease the interval to increase race condition */
315 CDEBUG(D_RPCTRACE,"Sleeping %d sec for inflight to error out\n",
318 /* Wait for all requests to error out and call completion
319 * callbacks. Cap it at obd_timeout -- these should all
320 * have been locally cancelled by ptlrpc_abort_inflight. */
321 lwi = LWI_TIMEOUT_INTERVAL(
322 cfs_timeout_cap(cfs_time_seconds(timeout)),
323 (timeout > 1)?cfs_time_seconds(1):cfs_time_seconds(1)/2,
325 rc = l_wait_event(imp->imp_recovery_waitq,
326 (atomic_read(&imp->imp_inflight) == 0),
329 const char *cli_tgt = obd2cli_tgt(imp->imp_obd);
331 CERROR("%s: rc = %d waiting for callback (%d != 0)\n",
332 cli_tgt, rc, atomic_read(&imp->imp_inflight));
334 spin_lock(&imp->imp_lock);
335 if (atomic_read(&imp->imp_inflight) == 0) {
336 int count = atomic_read(&imp->imp_unregistering);
338 /* We know that "unregistering" rpcs only can
339 * survive in sending or delaying lists (they
340 * maybe waiting for long reply unlink in
341 * sluggish nets). Let's check this. If there
342 * is no inflight and unregistering != 0, this
344 LASSERTF(count == 0, "Some RPCs are still "
345 "unregistering: %d\n", count);
347 /* Let's save one loop as soon as inflight have
348 * dropped to zero. No new inflights possible at
352 list_for_each_safe(tmp, n,
353 &imp->imp_sending_list) {
354 req = list_entry(tmp,
355 struct ptlrpc_request,
357 DEBUG_REQ(D_ERROR, req,
358 "still on sending list");
360 list_for_each_safe(tmp, n,
361 &imp->imp_delayed_list) {
362 req = list_entry(tmp,
363 struct ptlrpc_request,
365 DEBUG_REQ(D_ERROR, req,
366 "still on delayed list");
369 CERROR("%s: RPCs in \"%s\" phase found (%d). "
370 "Network is sluggish? Waiting them "
371 "to error out.\n", cli_tgt,
372 ptlrpc_phase2str(RQ_PHASE_UNREGISTERING),
373 atomic_read(&imp->imp_unregistering));
375 spin_unlock(&imp->imp_lock);
380 * Let's additionally check that no new rpcs added to import in
381 * "invalidate" state.
383 LASSERT(atomic_read(&imp->imp_inflight) == 0);
384 obd_import_event(imp->imp_obd, imp, IMP_EVENT_INVALIDATE);
385 sptlrpc_import_flush_all_ctx(imp);
387 atomic_dec(&imp->imp_inval_count);
388 wake_up_all(&imp->imp_recovery_waitq);
390 EXPORT_SYMBOL(ptlrpc_invalidate_import);
392 /* unset imp_invalid */
393 void ptlrpc_activate_import(struct obd_import *imp)
395 struct obd_device *obd = imp->imp_obd;
397 spin_lock(&imp->imp_lock);
398 if (imp->imp_deactive != 0) {
399 spin_unlock(&imp->imp_lock);
403 imp->imp_invalid = 0;
404 spin_unlock(&imp->imp_lock);
405 obd_import_event(obd, imp, IMP_EVENT_ACTIVE);
407 EXPORT_SYMBOL(ptlrpc_activate_import);
409 void ptlrpc_pinger_force(struct obd_import *imp)
411 CDEBUG(D_HA, "%s: waking up pinger s:%s\n", obd2cli_tgt(imp->imp_obd),
412 ptlrpc_import_state_name(imp->imp_state));
414 spin_lock(&imp->imp_lock);
415 imp->imp_force_verify = 1;
416 spin_unlock(&imp->imp_lock);
418 if (imp->imp_state != LUSTRE_IMP_CONNECTING)
419 ptlrpc_pinger_wake_up();
421 EXPORT_SYMBOL(ptlrpc_pinger_force);
423 void ptlrpc_fail_import(struct obd_import *imp, __u32 conn_cnt)
427 LASSERT(!imp->imp_dlm_fake);
429 if (ptlrpc_set_import_discon(imp, conn_cnt)) {
430 if (!imp->imp_replayable) {
431 CDEBUG(D_HA, "import %s@%s for %s not replayable, "
432 "auto-deactivating\n",
433 obd2cli_tgt(imp->imp_obd),
434 imp->imp_connection->c_remote_uuid.uuid,
435 imp->imp_obd->obd_name);
436 ptlrpc_deactivate_import(imp);
439 ptlrpc_pinger_force(imp);
444 int ptlrpc_reconnect_import(struct obd_import *imp)
447 struct l_wait_info lwi;
448 int secs = cfs_time_seconds(obd_timeout);
451 ptlrpc_pinger_force(imp);
453 CDEBUG(D_HA, "%s: recovery started, waiting %u seconds\n",
454 obd2cli_tgt(imp->imp_obd), secs);
456 lwi = LWI_TIMEOUT(secs, NULL, NULL);
457 rc = l_wait_event(imp->imp_recovery_waitq,
458 !ptlrpc_import_in_recovery(imp), &lwi);
459 CDEBUG(D_HA, "%s: recovery finished s:%s\n", obd2cli_tgt(imp->imp_obd),
460 ptlrpc_import_state_name(imp->imp_state));
463 ptlrpc_set_import_discon(imp, 0);
464 /* Force a new connect attempt */
465 ptlrpc_invalidate_import(imp);
466 /* Do a fresh connect next time by zeroing the handle */
467 ptlrpc_disconnect_import(imp, 1);
468 /* Wait for all invalidate calls to finish */
469 if (atomic_read(&imp->imp_inval_count) > 0) {
470 struct l_wait_info lwi = LWI_INTR(LWI_ON_SIGNAL_NOOP, NULL);
473 rc = l_wait_event(imp->imp_recovery_waitq,
474 (atomic_read(&imp->imp_inval_count) == 0),
477 CERROR("Interrupted, inval=%d\n",
478 atomic_read(&imp->imp_inval_count));
481 /* Allow reconnect attempts */
482 imp->imp_obd->obd_no_recov = 0;
483 /* Remove 'invalid' flag */
484 ptlrpc_activate_import(imp);
485 /* Attempt a new connect */
486 ptlrpc_recover_import(imp, NULL, 0);
490 EXPORT_SYMBOL(ptlrpc_reconnect_import);
493 * Connection on import \a imp is changed to another one (if more than one is
494 * present). We typically chose connection that we have not tried to connect to
497 static int import_select_connection(struct obd_import *imp)
499 struct obd_import_conn *imp_conn = NULL, *conn;
500 struct obd_export *dlmexp;
502 int target_len, tried_all = 1;
505 spin_lock(&imp->imp_lock);
507 if (list_empty(&imp->imp_conn_list)) {
508 CERROR("%s: no connections available\n",
509 imp->imp_obd->obd_name);
510 spin_unlock(&imp->imp_lock);
514 list_for_each_entry(conn, &imp->imp_conn_list, oic_item) {
515 CDEBUG(D_HA, "%s: connect to NID %s last attempt "LPU64"\n",
516 imp->imp_obd->obd_name,
517 libcfs_nid2str(conn->oic_conn->c_peer.nid),
518 conn->oic_last_attempt);
520 /* If we have not tried this connection since
521 the last successful attempt, go with this one */
522 if ((conn->oic_last_attempt == 0) ||
523 cfs_time_beforeq_64(conn->oic_last_attempt,
524 imp->imp_last_success_conn)) {
530 /* If all of the connections have already been tried
531 since the last successful connection; just choose the
532 least recently used */
535 else if (cfs_time_before_64(conn->oic_last_attempt,
536 imp_conn->oic_last_attempt))
540 /* if not found, simply choose the current one */
541 if (!imp_conn || imp->imp_force_reconnect) {
542 LASSERT(imp->imp_conn_current);
543 imp_conn = imp->imp_conn_current;
546 LASSERT(imp_conn->oic_conn);
548 /* If we've tried everything, and we're back to the beginning of the
549 list, increase our timeout and try again. It will be reset when
550 we do finally connect. (FIXME: really we should wait for all network
551 state associated with the last connection attempt to drain before
552 trying to reconnect on it.) */
553 if (tried_all && (imp->imp_conn_list.next == &imp_conn->oic_item)) {
554 struct adaptive_timeout *at = &imp->imp_at.iat_net_latency;
555 if (at_get(at) < CONNECTION_SWITCH_MAX) {
556 at_measured(at, at_get(at) + CONNECTION_SWITCH_INC);
557 if (at_get(at) > CONNECTION_SWITCH_MAX)
558 at_reset(at, CONNECTION_SWITCH_MAX);
560 LASSERT(imp_conn->oic_last_attempt);
561 CDEBUG(D_HA, "%s: tried all connections, increasing latency "
562 "to %ds\n", imp->imp_obd->obd_name, at_get(at));
565 imp_conn->oic_last_attempt = cfs_time_current_64();
567 /* switch connection, don't mind if it's same as the current one */
568 if (imp->imp_connection)
569 ptlrpc_connection_put(imp->imp_connection);
570 imp->imp_connection = ptlrpc_connection_addref(imp_conn->oic_conn);
572 dlmexp = class_conn2export(&imp->imp_dlm_handle);
573 LASSERT(dlmexp != NULL);
574 if (dlmexp->exp_connection)
575 ptlrpc_connection_put(dlmexp->exp_connection);
576 dlmexp->exp_connection = ptlrpc_connection_addref(imp_conn->oic_conn);
577 class_export_put(dlmexp);
579 if (imp->imp_conn_current != imp_conn) {
580 if (imp->imp_conn_current) {
581 deuuidify(obd2cli_tgt(imp->imp_obd), NULL,
582 &target_start, &target_len);
584 CDEBUG(D_HA, "%s: Connection changing to"
586 imp->imp_obd->obd_name,
587 target_len, target_start,
588 libcfs_nid2str(imp_conn->oic_conn->c_peer.nid));
591 imp->imp_conn_current = imp_conn;
594 CDEBUG(D_HA, "%s: import %p using connection %s/%s\n",
595 imp->imp_obd->obd_name, imp, imp_conn->oic_uuid.uuid,
596 libcfs_nid2str(imp_conn->oic_conn->c_peer.nid));
598 spin_unlock(&imp->imp_lock);
604 * must be called under imp_lock
606 static int ptlrpc_first_transno(struct obd_import *imp, __u64 *transno)
608 struct ptlrpc_request *req;
609 struct list_head *tmp;
611 /* The requests in committed_list always have smaller transnos than
612 * the requests in replay_list */
613 if (!list_empty(&imp->imp_committed_list)) {
614 tmp = imp->imp_committed_list.next;
615 req = list_entry(tmp, struct ptlrpc_request, rq_replay_list);
616 *transno = req->rq_transno;
617 if (req->rq_transno == 0) {
618 DEBUG_REQ(D_ERROR, req, "zero transno in committed_list");
623 if (!list_empty(&imp->imp_replay_list)) {
624 tmp = imp->imp_replay_list.next;
625 req = list_entry(tmp, struct ptlrpc_request, rq_replay_list);
626 *transno = req->rq_transno;
627 if (req->rq_transno == 0) {
628 DEBUG_REQ(D_ERROR, req, "zero transno in replay_list");
637 * Attempt to (re)connect import \a imp. This includes all preparations,
638 * initializing CONNECT RPC request and passing it to ptlrpcd for
640 * Returns 0 on success or error code.
642 int ptlrpc_connect_import(struct obd_import *imp)
644 struct obd_device *obd = imp->imp_obd;
645 int initial_connect = 0;
647 __u64 committed_before_reconnect = 0;
648 struct ptlrpc_request *request;
649 char *bufs[] = { NULL,
650 obd2cli_tgt(imp->imp_obd),
652 (char *)&imp->imp_dlm_handle,
653 (char *)&imp->imp_connect_data };
654 struct ptlrpc_connect_async_args *aa;
658 spin_lock(&imp->imp_lock);
659 if (imp->imp_state == LUSTRE_IMP_CLOSED) {
660 spin_unlock(&imp->imp_lock);
661 CERROR("can't connect to a closed import\n");
663 } else if (imp->imp_state == LUSTRE_IMP_FULL) {
664 spin_unlock(&imp->imp_lock);
665 CERROR("already connected\n");
667 } else if (imp->imp_state == LUSTRE_IMP_CONNECTING) {
668 spin_unlock(&imp->imp_lock);
669 CERROR("already connecting\n");
673 IMPORT_SET_STATE_NOLOCK(imp, LUSTRE_IMP_CONNECTING);
676 imp->imp_resend_replay = 0;
678 if (!lustre_handle_is_used(&imp->imp_remote_handle))
681 committed_before_reconnect = imp->imp_peer_committed_transno;
683 set_transno = ptlrpc_first_transno(imp,
684 &imp->imp_connect_data.ocd_transno);
685 spin_unlock(&imp->imp_lock);
687 rc = import_select_connection(imp);
691 rc = sptlrpc_import_sec_adapt(imp, NULL, NULL);
695 /* Reset connect flags to the originally requested flags, in case
696 * the server is updated on-the-fly we will get the new features. */
697 imp->imp_connect_data.ocd_connect_flags = imp->imp_connect_flags_orig;
698 /* Reset ocd_version each time so the server knows the exact versions */
699 imp->imp_connect_data.ocd_version = LUSTRE_VERSION_CODE;
700 imp->imp_msghdr_flags &= ~MSGHDR_AT_SUPPORT;
701 imp->imp_msghdr_flags &= ~MSGHDR_CKSUM_INCOMPAT18;
703 rc = obd_reconnect(NULL, imp->imp_obd->obd_self_export, obd,
704 &obd->obd_uuid, &imp->imp_connect_data, NULL);
708 request = ptlrpc_request_alloc(imp, &RQF_MDS_CONNECT);
710 GOTO(out, rc = -ENOMEM);
712 rc = ptlrpc_request_bufs_pack(request, LUSTRE_OBD_VERSION,
713 imp->imp_connect_op, bufs, NULL);
715 ptlrpc_request_free(request);
719 /* Report the rpc service time to the server so that it knows how long
720 * to wait for clients to join recovery */
721 lustre_msg_set_service_time(request->rq_reqmsg,
722 at_timeout2est(request->rq_timeout));
724 /* The amount of time we give the server to process the connect req.
725 * import_select_connection will increase the net latency on
726 * repeated reconnect attempts to cover slow networks.
727 * We override/ignore the server rpc completion estimate here,
728 * which may be large if this is a reconnect attempt */
729 request->rq_timeout = INITIAL_CONNECT_TIMEOUT;
730 lustre_msg_set_timeout(request->rq_reqmsg, request->rq_timeout);
732 request->rq_no_resend = request->rq_no_delay = 1;
733 request->rq_send_state = LUSTRE_IMP_CONNECTING;
734 /* Allow a slightly larger reply for future growth compatibility */
735 req_capsule_set_size(&request->rq_pill, &RMF_CONNECT_DATA, RCL_SERVER,
736 sizeof(struct obd_connect_data)+16*sizeof(__u64));
737 ptlrpc_request_set_replen(request);
738 request->rq_interpret_reply = ptlrpc_connect_interpret;
740 CLASSERT(sizeof (*aa) <= sizeof (request->rq_async_args));
741 aa = ptlrpc_req_async_args(request);
742 memset(aa, 0, sizeof *aa);
744 aa->pcaa_peer_committed = committed_before_reconnect;
745 aa->pcaa_initial_connect = initial_connect;
747 if (aa->pcaa_initial_connect) {
748 spin_lock(&imp->imp_lock);
749 imp->imp_replayable = 1;
750 spin_unlock(&imp->imp_lock);
751 lustre_msg_add_op_flags(request->rq_reqmsg,
752 MSG_CONNECT_INITIAL);
756 lustre_msg_add_op_flags(request->rq_reqmsg,
757 MSG_CONNECT_TRANSNO);
759 DEBUG_REQ(D_RPCTRACE, request, "(re)connect request (timeout %d)",
760 request->rq_timeout);
761 ptlrpcd_add_req(request);
765 IMPORT_SET_STATE(imp, LUSTRE_IMP_DISCON);
769 EXPORT_SYMBOL(ptlrpc_connect_import);
771 static void ptlrpc_maybe_ping_import_soon(struct obd_import *imp)
775 spin_lock(&imp->imp_lock);
776 force_verify = imp->imp_force_verify != 0;
777 spin_unlock(&imp->imp_lock);
780 ptlrpc_pinger_wake_up();
783 static int ptlrpc_busy_reconnect(int rc)
785 return (rc == -EBUSY) || (rc == -EAGAIN);
788 static int ptlrpc_connect_set_flags(struct obd_import *imp,
789 struct obd_connect_data *ocd,
790 __u64 old_connect_flags,
791 struct obd_export *exp, int init_connect)
794 struct client_obd *cli = &imp->imp_obd->u.cli;
796 if ((imp->imp_connect_flags_orig & OBD_CONNECT_IBITS) &&
797 !(ocd->ocd_connect_flags & OBD_CONNECT_IBITS)) {
798 LCONSOLE_WARN("%s: MDS %s does not support ibits "
799 "lock, either very old or invalid: "
800 "requested "LPX64", replied "LPX64"\n",
801 imp->imp_obd->obd_name,
802 imp->imp_connection->c_remote_uuid.uuid,
803 imp->imp_connect_flags_orig,
804 ocd->ocd_connect_flags);
808 spin_lock(&imp->imp_lock);
809 list_del(&imp->imp_conn_current->oic_item);
810 list_add(&imp->imp_conn_current->oic_item,
811 &imp->imp_conn_list);
812 imp->imp_last_success_conn =
813 imp->imp_conn_current->oic_last_attempt;
815 spin_unlock(&imp->imp_lock);
818 if (!warned && (ocd->ocd_connect_flags & OBD_CONNECT_VERSION) &&
819 (ocd->ocd_version > LUSTRE_VERSION_CODE +
820 LUSTRE_VERSION_OFFSET_WARN ||
821 ocd->ocd_version < LUSTRE_VERSION_CODE -
822 LUSTRE_VERSION_OFFSET_WARN)) {
823 /* Sigh, some compilers do not like #ifdef in the middle
824 of macro arguments */
825 const char *older = "older than client. "
826 "Consider upgrading server";
827 const char *newer = "newer than client. "
828 "Consider recompiling application";
830 LCONSOLE_WARN("Server %s version (%d.%d.%d.%d) "
832 obd2cli_tgt(imp->imp_obd),
833 OBD_OCD_VERSION_MAJOR(ocd->ocd_version),
834 OBD_OCD_VERSION_MINOR(ocd->ocd_version),
835 OBD_OCD_VERSION_PATCH(ocd->ocd_version),
836 OBD_OCD_VERSION_FIX(ocd->ocd_version),
837 ocd->ocd_version > LUSTRE_VERSION_CODE ?
838 newer : older, LUSTRE_VERSION_STRING);
842 #if LUSTRE_VERSION_CODE < OBD_OCD_VERSION(3, 0, 53, 0)
843 /* Check if server has LU-1252 fix applied to not always swab
844 * the IR MNE entries. Do this only once per connection. This
845 * fixup is version-limited, because we don't want to carry the
846 * OBD_CONNECT_MNE_SWAB flag around forever, just so long as we
847 * need interop with unpatched 2.2 servers. For newer servers,
848 * the client will do MNE swabbing only as needed. LU-1644 */
849 if (unlikely((ocd->ocd_connect_flags & OBD_CONNECT_VERSION) &&
850 !(ocd->ocd_connect_flags & OBD_CONNECT_MNE_SWAB) &&
851 OBD_OCD_VERSION_MAJOR(ocd->ocd_version) == 2 &&
852 OBD_OCD_VERSION_MINOR(ocd->ocd_version) == 2 &&
853 OBD_OCD_VERSION_PATCH(ocd->ocd_version) < 55 &&
854 strcmp(imp->imp_obd->obd_type->typ_name,
855 LUSTRE_MGC_NAME) == 0))
856 imp->imp_need_mne_swab = 1;
857 else /* clear if server was upgraded since last connect */
858 imp->imp_need_mne_swab = 0;
861 if (ocd->ocd_connect_flags & OBD_CONNECT_CKSUM) {
862 /* We sent to the server ocd_cksum_types with bits set
863 * for algorithms we understand. The server masked off
864 * the checksum types it doesn't support */
865 if ((ocd->ocd_cksum_types &
866 cksum_types_supported_client()) == 0) {
867 LCONSOLE_WARN("The negotiation of the checksum "
868 "alogrithm to use with server %s "
869 "failed (%x/%x), disabling "
871 obd2cli_tgt(imp->imp_obd),
872 ocd->ocd_cksum_types,
873 cksum_types_supported_client());
874 cli->cl_checksum = 0;
875 cli->cl_supp_cksum_types = OBD_CKSUM_ADLER;
877 cli->cl_supp_cksum_types = ocd->ocd_cksum_types;
880 /* The server does not support OBD_CONNECT_CKSUM.
881 * Enforce ADLER for backward compatibility*/
882 cli->cl_supp_cksum_types = OBD_CKSUM_ADLER;
884 cli->cl_cksum_type = cksum_type_select(cli->cl_supp_cksum_types);
886 if (ocd->ocd_connect_flags & OBD_CONNECT_BRW_SIZE)
887 cli->cl_max_pages_per_rpc =
888 min(ocd->ocd_brw_size >> PAGE_CACHE_SHIFT,
889 cli->cl_max_pages_per_rpc);
890 else if (imp->imp_connect_op == MDS_CONNECT ||
891 imp->imp_connect_op == MGS_CONNECT)
892 cli->cl_max_pages_per_rpc = 1;
894 LASSERT((cli->cl_max_pages_per_rpc <= PTLRPC_MAX_BRW_PAGES) &&
895 (cli->cl_max_pages_per_rpc > 0));
897 client_adjust_max_dirty(cli);
899 /* Update client max modify RPCs in flight with value returned
901 if (ocd->ocd_connect_flags & OBD_CONNECT_MULTIMODRPCS)
902 cli->cl_max_mod_rpcs_in_flight = min(
903 cli->cl_max_mod_rpcs_in_flight,
904 ocd->ocd_maxmodrpcs);
906 cli->cl_max_mod_rpcs_in_flight = 1;
908 /* Reset ns_connect_flags only for initial connect. It might be
909 * changed in while using FS and if we reset it in reconnect
910 * this leads to losing user settings done before such as
911 * disable lru_resize, etc. */
912 if (old_connect_flags != exp_connect_flags(exp) || init_connect) {
913 CDEBUG(D_HA, "%s: Resetting ns_connect_flags to server "
914 "flags: "LPX64"\n", imp->imp_obd->obd_name,
915 ocd->ocd_connect_flags);
916 imp->imp_obd->obd_namespace->ns_connect_flags =
917 ocd->ocd_connect_flags;
918 imp->imp_obd->obd_namespace->ns_orig_connect_flags =
919 ocd->ocd_connect_flags;
922 if (ocd->ocd_connect_flags & OBD_CONNECT_AT)
923 /* We need a per-message support flag, because
924 * a. we don't know if the incoming connect reply
925 * supports AT or not (in reply_in_callback)
926 * until we unpack it.
927 * b. failovered server means export and flags are gone
928 * (in ptlrpc_send_reply).
929 * Can only be set when we know AT is supported at
931 imp->imp_msghdr_flags |= MSGHDR_AT_SUPPORT;
933 imp->imp_msghdr_flags &= ~MSGHDR_AT_SUPPORT;
935 imp->imp_msghdr_flags |= MSGHDR_CKSUM_INCOMPAT18;
941 * interpret_reply callback for connect RPCs.
942 * Looks into returned status of connect operation and decides
943 * what to do with the import - i.e enter recovery, promote it to
944 * full state for normal operations of disconnect it due to an error.
946 static int ptlrpc_connect_interpret(const struct lu_env *env,
947 struct ptlrpc_request *request,
950 struct ptlrpc_connect_async_args *aa = data;
951 struct obd_import *imp = request->rq_import;
952 struct lustre_handle old_hdl;
953 __u64 old_connect_flags;
955 struct obd_connect_data *ocd;
956 struct obd_export *exp;
960 spin_lock(&imp->imp_lock);
961 if (imp->imp_state == LUSTRE_IMP_CLOSED) {
962 imp->imp_connect_tried = 1;
963 spin_unlock(&imp->imp_lock);
968 /* if this reconnect to busy export - not need select new target
970 imp->imp_force_reconnect = ptlrpc_busy_reconnect(rc);
971 spin_unlock(&imp->imp_lock);
972 ptlrpc_maybe_ping_import_soon(imp);
975 spin_unlock(&imp->imp_lock);
977 LASSERT(imp->imp_conn_current);
979 msg_flags = lustre_msg_get_op_flags(request->rq_repmsg);
981 ret = req_capsule_get_size(&request->rq_pill, &RMF_CONNECT_DATA,
983 /* server replied obd_connect_data is always bigger */
984 ocd = req_capsule_server_sized_get(&request->rq_pill,
985 &RMF_CONNECT_DATA, ret);
988 CERROR("%s: no connect data from server\n",
989 imp->imp_obd->obd_name);
994 spin_lock(&imp->imp_lock);
996 /* All imports are pingable */
997 imp->imp_pingable = 1;
998 imp->imp_force_reconnect = 0;
999 imp->imp_force_verify = 0;
1001 imp->imp_connect_data = *ocd;
1003 CDEBUG(D_HA, "%s: connect to target with instance %u\n",
1004 imp->imp_obd->obd_name, ocd->ocd_instance);
1005 exp = class_conn2export(&imp->imp_dlm_handle);
1007 spin_unlock(&imp->imp_lock);
1009 /* check that server granted subset of flags we asked for. */
1010 if ((ocd->ocd_connect_flags & imp->imp_connect_flags_orig) !=
1011 ocd->ocd_connect_flags) {
1012 CERROR("%s: Server didn't granted asked subset of flags: "
1013 "asked="LPX64" grranted="LPX64"\n",
1014 imp->imp_obd->obd_name,imp->imp_connect_flags_orig,
1015 ocd->ocd_connect_flags);
1016 GOTO(out, rc = -EPROTO);
1019 if (!(imp->imp_connect_flags_orig & OBD_CONNECT_LIGHTWEIGHT) &&
1020 (imp->imp_connect_flags_orig & OBD_CONNECT_MDS_MDS) &&
1021 (imp->imp_connect_flags_orig & OBD_CONNECT_FID) &&
1022 (ocd->ocd_connect_flags & OBD_CONNECT_VERSION)) {
1023 __u32 major = OBD_OCD_VERSION_MAJOR(ocd->ocd_version);
1024 __u32 minor = OBD_OCD_VERSION_MINOR(ocd->ocd_version);
1025 __u32 patch = OBD_OCD_VERSION_PATCH(ocd->ocd_version);
1027 /* We do not support the MDT-MDT interoperations with
1028 * different version MDT because of protocol changes. */
1029 if (unlikely(major != LUSTRE_MAJOR ||
1030 minor != LUSTRE_MINOR ||
1031 abs(patch - LUSTRE_PATCH) > 3)) {
1032 LCONSOLE_WARN("%s: import %p (%u.%u.%u.%u) tried the "
1033 "connection to different version MDT "
1034 "(%d.%d.%d.%d) %s\n",
1035 imp->imp_obd->obd_name, imp, LUSTRE_MAJOR,
1036 LUSTRE_MINOR, LUSTRE_PATCH, LUSTRE_FIX,
1037 major, minor, patch,
1038 OBD_OCD_VERSION_FIX(ocd->ocd_version),
1039 imp->imp_connection->c_remote_uuid.uuid);
1041 GOTO(out, rc = -EPROTO);
1046 /* This could happen if export is cleaned during the
1048 CERROR("%s: missing export after connect\n",
1049 imp->imp_obd->obd_name);
1050 GOTO(out, rc = -ENODEV);
1052 old_connect_flags = exp_connect_flags(exp);
1053 exp->exp_connect_data = *ocd;
1054 imp->imp_obd->obd_self_export->exp_connect_data = *ocd;
1056 /* The net statistics after (re-)connect is not valid anymore,
1057 * because may reflect other routing, etc. */
1058 at_reinit(&imp->imp_at.iat_net_latency, 0, 0);
1059 ptlrpc_at_adj_net_latency(request,
1060 lustre_msg_get_service_time(request->rq_repmsg));
1062 /* Import flags should be updated before waking import at FULL state */
1063 rc = ptlrpc_connect_set_flags(imp, ocd, old_connect_flags, exp,
1064 aa->pcaa_initial_connect);
1065 class_export_put(exp);
1069 obd_import_event(imp->imp_obd, imp, IMP_EVENT_OCD);
1071 if (aa->pcaa_initial_connect) {
1072 spin_lock(&imp->imp_lock);
1073 if (msg_flags & MSG_CONNECT_REPLAYABLE) {
1074 imp->imp_replayable = 1;
1075 spin_unlock(&imp->imp_lock);
1076 CDEBUG(D_HA, "connected to replayable target: %s\n",
1077 obd2cli_tgt(imp->imp_obd));
1079 imp->imp_replayable = 0;
1080 spin_unlock(&imp->imp_lock);
1083 /* if applies, adjust the imp->imp_msg_magic here
1084 * according to reply flags */
1086 imp->imp_remote_handle =
1087 *lustre_msg_get_handle(request->rq_repmsg);
1089 /* Initial connects are allowed for clients with non-random
1090 * uuids when servers are in recovery. Simply signal the
1091 * servers replay is complete and wait in REPLAY_WAIT. */
1092 if (msg_flags & MSG_CONNECT_RECOVERING) {
1093 CDEBUG(D_HA, "connect to %s during recovery\n",
1094 obd2cli_tgt(imp->imp_obd));
1095 IMPORT_SET_STATE(imp, LUSTRE_IMP_REPLAY_LOCKS);
1097 IMPORT_SET_STATE(imp, LUSTRE_IMP_FULL);
1098 ptlrpc_activate_import(imp);
1101 GOTO(finish, rc = 0);
1104 /* Determine what recovery state to move the import to. */
1105 if (MSG_CONNECT_RECONNECT & msg_flags) {
1106 memset(&old_hdl, 0, sizeof(old_hdl));
1107 if (!memcmp(&old_hdl, lustre_msg_get_handle(request->rq_repmsg),
1108 sizeof (old_hdl))) {
1109 LCONSOLE_WARN("Reconnect to %s (at @%s) failed due "
1110 "bad handle "LPX64"\n",
1111 obd2cli_tgt(imp->imp_obd),
1112 imp->imp_connection->c_remote_uuid.uuid,
1113 imp->imp_dlm_handle.cookie);
1114 GOTO(out, rc = -ENOTCONN);
1117 if (memcmp(&imp->imp_remote_handle,
1118 lustre_msg_get_handle(request->rq_repmsg),
1119 sizeof(imp->imp_remote_handle))) {
1120 int level = msg_flags & MSG_CONNECT_RECOVERING ?
1123 /* Bug 16611/14775: if server handle have changed,
1124 * that means some sort of disconnection happened.
1125 * If the server is not in recovery, that also means it
1126 * already erased all of our state because of previous
1127 * eviction. If it is in recovery - we are safe to
1128 * participate since we can reestablish all of our state
1129 * with server again */
1130 if ((MSG_CONNECT_RECOVERING & msg_flags)) {
1131 CDEBUG(level,"%s@%s changed server handle from "
1133 " but is still in recovery\n",
1134 obd2cli_tgt(imp->imp_obd),
1135 imp->imp_connection->c_remote_uuid.uuid,
1136 imp->imp_remote_handle.cookie,
1137 lustre_msg_get_handle(
1138 request->rq_repmsg)->cookie);
1140 LCONSOLE_WARN("Evicted from %s (at %s) "
1141 "after server handle changed from "
1142 LPX64" to "LPX64"\n",
1143 obd2cli_tgt(imp->imp_obd),
1144 imp->imp_connection-> \
1146 imp->imp_remote_handle.cookie,
1147 lustre_msg_get_handle(
1148 request->rq_repmsg)->cookie);
1152 imp->imp_remote_handle =
1153 *lustre_msg_get_handle(request->rq_repmsg);
1155 if (!(MSG_CONNECT_RECOVERING & msg_flags)) {
1156 IMPORT_SET_STATE(imp, LUSTRE_IMP_EVICTED);
1157 GOTO(finish, rc = 0);
1161 CDEBUG(D_HA, "reconnected to %s@%s after partition\n",
1162 obd2cli_tgt(imp->imp_obd),
1163 imp->imp_connection->c_remote_uuid.uuid);
1166 if (imp->imp_invalid) {
1167 CDEBUG(D_HA, "%s: reconnected but import is invalid; "
1168 "marking evicted\n", imp->imp_obd->obd_name);
1169 IMPORT_SET_STATE(imp, LUSTRE_IMP_EVICTED);
1170 } else if (MSG_CONNECT_RECOVERING & msg_flags) {
1171 CDEBUG(D_HA, "%s: reconnected to %s during replay\n",
1172 imp->imp_obd->obd_name,
1173 obd2cli_tgt(imp->imp_obd));
1175 spin_lock(&imp->imp_lock);
1176 imp->imp_resend_replay = 1;
1177 spin_unlock(&imp->imp_lock);
1179 IMPORT_SET_STATE(imp, imp->imp_replay_state);
1181 IMPORT_SET_STATE(imp, LUSTRE_IMP_RECOVER);
1183 } else if ((MSG_CONNECT_RECOVERING & msg_flags) && !imp->imp_invalid) {
1184 LASSERT(imp->imp_replayable);
1185 imp->imp_remote_handle =
1186 *lustre_msg_get_handle(request->rq_repmsg);
1187 imp->imp_last_replay_transno = 0;
1188 IMPORT_SET_STATE(imp, LUSTRE_IMP_REPLAY);
1190 DEBUG_REQ(D_HA, request, "%s: evicting (reconnect/recover flags"
1191 " not set: %x)", imp->imp_obd->obd_name, msg_flags);
1192 imp->imp_remote_handle =
1193 *lustre_msg_get_handle(request->rq_repmsg);
1194 IMPORT_SET_STATE(imp, LUSTRE_IMP_EVICTED);
1197 /* Sanity checks for a reconnected import. */
1198 if (!(imp->imp_replayable) != !(msg_flags & MSG_CONNECT_REPLAYABLE)) {
1199 CERROR("imp_replayable flag does not match server "
1200 "after reconnect. We should LBUG right here.\n");
1203 if (lustre_msg_get_last_committed(request->rq_repmsg) > 0 &&
1204 lustre_msg_get_last_committed(request->rq_repmsg) <
1205 aa->pcaa_peer_committed) {
1206 CERROR("%s went back in time (transno "LPD64
1207 " was previously committed, server now claims "LPD64
1208 ")! See https://bugzilla.lustre.org/show_bug.cgi?"
1210 obd2cli_tgt(imp->imp_obd), aa->pcaa_peer_committed,
1211 lustre_msg_get_last_committed(request->rq_repmsg));
1215 rc = ptlrpc_import_recovery_state_machine(imp);
1216 if (rc == -ENOTCONN) {
1217 CDEBUG(D_HA, "evicted/aborted by %s@%s during recovery;"
1218 "invalidating and reconnecting\n",
1219 obd2cli_tgt(imp->imp_obd),
1220 imp->imp_connection->c_remote_uuid.uuid);
1221 ptlrpc_connect_import(imp);
1222 imp->imp_connect_tried = 1;
1227 imp->imp_connect_tried = 1;
1230 IMPORT_SET_STATE(imp, LUSTRE_IMP_DISCON);
1231 if (rc == -EACCES) {
1233 * Give up trying to reconnect
1234 * EACCES means client has no permission for connection
1236 imp->imp_obd->obd_no_recov = 1;
1237 ptlrpc_deactivate_import(imp);
1240 if (rc == -EPROTO) {
1241 struct obd_connect_data *ocd;
1243 /* reply message might not be ready */
1244 if (request->rq_repmsg == NULL)
1247 ocd = req_capsule_server_get(&request->rq_pill,
1250 (ocd->ocd_connect_flags & OBD_CONNECT_VERSION) &&
1251 (ocd->ocd_version != LUSTRE_VERSION_CODE)) {
1252 /* Actually servers are only supposed to refuse
1253 connection from liblustre clients, so we should
1254 never see this from VFS context */
1255 LCONSOLE_ERROR_MSG(0x16a, "Server %s version "
1257 " refused connection from this client "
1258 "with an incompatible version (%s). "
1259 "Client must be recompiled\n",
1260 obd2cli_tgt(imp->imp_obd),
1261 OBD_OCD_VERSION_MAJOR(ocd->ocd_version),
1262 OBD_OCD_VERSION_MINOR(ocd->ocd_version),
1263 OBD_OCD_VERSION_PATCH(ocd->ocd_version),
1264 OBD_OCD_VERSION_FIX(ocd->ocd_version),
1265 LUSTRE_VERSION_STRING);
1266 ptlrpc_deactivate_import(imp);
1267 IMPORT_SET_STATE(imp, LUSTRE_IMP_CLOSED);
1272 ptlrpc_maybe_ping_import_soon(imp);
1274 CDEBUG(D_HA, "recovery of %s on %s failed (%d)\n",
1275 obd2cli_tgt(imp->imp_obd),
1276 (char *)imp->imp_connection->c_remote_uuid.uuid, rc);
1279 wake_up_all(&imp->imp_recovery_waitq);
1284 * interpret callback for "completed replay" RPCs.
1285 * \see signal_completed_replay
1287 static int completed_replay_interpret(const struct lu_env *env,
1288 struct ptlrpc_request *req,
1289 void * data, int rc)
1292 atomic_dec(&req->rq_import->imp_replay_inflight);
1293 if (req->rq_status == 0 &&
1294 !req->rq_import->imp_vbr_failed) {
1295 ptlrpc_import_recovery_state_machine(req->rq_import);
1297 if (req->rq_import->imp_vbr_failed) {
1299 "%s: version recovery fails, reconnecting\n",
1300 req->rq_import->imp_obd->obd_name);
1302 CDEBUG(D_HA, "%s: LAST_REPLAY message error: %d, "
1304 req->rq_import->imp_obd->obd_name,
1307 ptlrpc_connect_import(req->rq_import);
1314 * Let server know that we have no requests to replay anymore.
1315 * Achieved by just sending a PING request
1317 static int signal_completed_replay(struct obd_import *imp)
1319 struct ptlrpc_request *req;
1322 if (unlikely(OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_FINISH_REPLAY)))
1325 LASSERT(atomic_read(&imp->imp_replay_inflight) == 0);
1326 atomic_inc(&imp->imp_replay_inflight);
1328 req = ptlrpc_request_alloc_pack(imp, &RQF_OBD_PING, LUSTRE_OBD_VERSION,
1331 atomic_dec(&imp->imp_replay_inflight);
1335 ptlrpc_request_set_replen(req);
1336 req->rq_send_state = LUSTRE_IMP_REPLAY_WAIT;
1337 lustre_msg_add_flags(req->rq_reqmsg,
1338 MSG_LOCK_REPLAY_DONE | MSG_REQ_REPLAY_DONE);
1340 req->rq_timeout *= 3;
1341 req->rq_interpret_reply = completed_replay_interpret;
1343 ptlrpcd_add_req(req);
1348 * In kernel code all import invalidation happens in its own
1349 * separate thread, so that whatever application happened to encounter
1350 * a problem could still be killed or otherwise continue
1352 static int ptlrpc_invalidate_import_thread(void *data)
1354 struct obd_import *imp = data;
1358 unshare_fs_struct();
1360 CDEBUG(D_HA, "thread invalidate import %s to %s@%s\n",
1361 imp->imp_obd->obd_name, obd2cli_tgt(imp->imp_obd),
1362 imp->imp_connection->c_remote_uuid.uuid);
1364 ptlrpc_invalidate_import(imp);
1366 if (obd_dump_on_eviction) {
1367 CERROR("dump the log upon eviction\n");
1368 libcfs_debug_dumplog();
1371 IMPORT_SET_STATE(imp, LUSTRE_IMP_RECOVER);
1372 ptlrpc_import_recovery_state_machine(imp);
1374 class_import_put(imp);
1379 * This is the state machine for client-side recovery on import.
1381 * Typicaly we have two possibly paths. If we came to server and it is not
1382 * in recovery, we just enter IMP_EVICTED state, invalidate our import
1383 * state and reconnect from scratch.
1384 * If we came to server that is in recovery, we enter IMP_REPLAY import state.
1385 * We go through our list of requests to replay and send them to server one by
1387 * After sending all request from the list we change import state to
1388 * IMP_REPLAY_LOCKS and re-request all the locks we believe we have from server
1389 * and also all the locks we don't yet have and wait for server to grant us.
1390 * After that we send a special "replay completed" request and change import
1391 * state to IMP_REPLAY_WAIT.
1392 * Upon receiving reply to that "replay completed" RPC we enter IMP_RECOVER
1393 * state and resend all requests from sending list.
1394 * After that we promote import to FULL state and send all delayed requests
1395 * and import is fully operational after that.
1398 int ptlrpc_import_recovery_state_machine(struct obd_import *imp)
1406 if (imp->imp_state == LUSTRE_IMP_EVICTED) {
1407 deuuidify(obd2cli_tgt(imp->imp_obd), NULL,
1408 &target_start, &target_len);
1409 /* Don't care about MGC eviction */
1410 if (strcmp(imp->imp_obd->obd_type->typ_name,
1411 LUSTRE_MGC_NAME) != 0) {
1412 LCONSOLE_ERROR_MSG(0x167, "%s: This client was evicted "
1413 "by %.*s; in progress operations "
1414 "using this service will fail.\n",
1415 imp->imp_obd->obd_name, target_len,
1418 CDEBUG(D_HA, "evicted from %s@%s; invalidating\n",
1419 obd2cli_tgt(imp->imp_obd),
1420 imp->imp_connection->c_remote_uuid.uuid);
1421 /* reset vbr_failed flag upon eviction */
1422 spin_lock(&imp->imp_lock);
1423 imp->imp_vbr_failed = 0;
1424 spin_unlock(&imp->imp_lock);
1427 struct task_struct *task;
1428 /* bug 17802: XXX client_disconnect_export vs connect request
1429 * race. if client is evicted at this time then we start
1430 * invalidate thread without reference to import and import can
1431 * be freed at same time. */
1432 class_import_get(imp);
1433 task = kthread_run(ptlrpc_invalidate_import_thread, imp,
1436 class_import_put(imp);
1437 CERROR("error starting invalidate thread: %d\n", rc);
1446 if (imp->imp_state == LUSTRE_IMP_REPLAY) {
1447 CDEBUG(D_HA, "replay requested by %s\n",
1448 obd2cli_tgt(imp->imp_obd));
1449 rc = ptlrpc_replay_next(imp, &inflight);
1450 if (inflight == 0 &&
1451 atomic_read(&imp->imp_replay_inflight) == 0) {
1452 IMPORT_SET_STATE(imp, LUSTRE_IMP_REPLAY_LOCKS);
1453 rc = ldlm_replay_locks(imp);
1460 if (imp->imp_state == LUSTRE_IMP_REPLAY_LOCKS) {
1461 if (atomic_read(&imp->imp_replay_inflight) == 0) {
1462 IMPORT_SET_STATE(imp, LUSTRE_IMP_REPLAY_WAIT);
1463 rc = signal_completed_replay(imp);
1469 if (imp->imp_state == LUSTRE_IMP_REPLAY_WAIT) {
1470 if (atomic_read(&imp->imp_replay_inflight) == 0) {
1471 IMPORT_SET_STATE(imp, LUSTRE_IMP_RECOVER);
1475 if (imp->imp_state == LUSTRE_IMP_RECOVER) {
1476 CDEBUG(D_HA, "reconnected to %s@%s\n",
1477 obd2cli_tgt(imp->imp_obd),
1478 imp->imp_connection->c_remote_uuid.uuid);
1480 rc = ptlrpc_resend(imp);
1483 IMPORT_SET_STATE(imp, LUSTRE_IMP_FULL);
1484 ptlrpc_activate_import(imp);
1486 deuuidify(obd2cli_tgt(imp->imp_obd), NULL,
1487 &target_start, &target_len);
1488 LCONSOLE_INFO("%s: Connection restored to %.*s (at %s)\n",
1489 imp->imp_obd->obd_name,
1490 target_len, target_start,
1491 libcfs_nid2str(imp->imp_connection->c_peer.nid));
1494 if (imp->imp_state == LUSTRE_IMP_FULL) {
1495 wake_up_all(&imp->imp_recovery_waitq);
1496 ptlrpc_wake_delayed(imp);
1503 int ptlrpc_disconnect_import(struct obd_import *imp, int noclose)
1505 struct ptlrpc_request *req;
1509 if (imp->imp_obd->obd_force)
1510 GOTO(set_state, rc);
1512 switch (imp->imp_connect_op) {
1514 rq_opc = OST_DISCONNECT;
1517 rq_opc = MDS_DISCONNECT;
1520 rq_opc = MGS_DISCONNECT;
1524 CERROR("%s: don't know how to disconnect from %s "
1525 "(connect_op %d): rc = %d\n",
1526 imp->imp_obd->obd_name, obd2cli_tgt(imp->imp_obd),
1527 imp->imp_connect_op, rc);
1531 if (ptlrpc_import_in_recovery(imp)) {
1532 struct l_wait_info lwi;
1533 cfs_duration_t timeout;
1536 if (imp->imp_server_timeout)
1537 timeout = cfs_time_seconds(obd_timeout / 2);
1539 timeout = cfs_time_seconds(obd_timeout);
1541 int idx = import_at_get_index(imp,
1542 imp->imp_client->cli_request_portal);
1543 timeout = cfs_time_seconds(
1544 at_get(&imp->imp_at.iat_service_estimate[idx]));
1547 lwi = LWI_TIMEOUT_INTR(cfs_timeout_cap(timeout),
1548 back_to_sleep, LWI_ON_SIGNAL_NOOP, NULL);
1549 rc = l_wait_event(imp->imp_recovery_waitq,
1550 !ptlrpc_import_in_recovery(imp), &lwi);
1554 spin_lock(&imp->imp_lock);
1555 if (imp->imp_state != LUSTRE_IMP_FULL)
1557 spin_unlock(&imp->imp_lock);
1559 req = ptlrpc_request_alloc_pack(imp, &RQF_MDS_DISCONNECT,
1560 LUSTRE_OBD_VERSION, rq_opc);
1562 /* We are disconnecting, do not retry a failed DISCONNECT rpc if
1563 * it fails. We can get through the above with a down server
1564 * if the client doesn't know the server is gone yet. */
1565 req->rq_no_resend = 1;
1567 /* We want client umounts to happen quickly, no matter the
1569 req->rq_timeout = min_t(int, req->rq_timeout,
1570 INITIAL_CONNECT_TIMEOUT);
1572 IMPORT_SET_STATE(imp, LUSTRE_IMP_CONNECTING);
1573 req->rq_send_state = LUSTRE_IMP_CONNECTING;
1574 ptlrpc_request_set_replen(req);
1575 rc = ptlrpc_queue_wait(req);
1576 ptlrpc_req_finished(req);
1580 spin_lock(&imp->imp_lock);
1583 IMPORT_SET_STATE_NOLOCK(imp, LUSTRE_IMP_DISCON);
1585 IMPORT_SET_STATE_NOLOCK(imp, LUSTRE_IMP_CLOSED);
1586 memset(&imp->imp_remote_handle, 0, sizeof(imp->imp_remote_handle));
1587 spin_unlock(&imp->imp_lock);
1589 if (rc == -ETIMEDOUT || rc == -ENOTCONN || rc == -ESHUTDOWN)
1593 EXPORT_SYMBOL(ptlrpc_disconnect_import);
1595 void ptlrpc_cleanup_imp(struct obd_import *imp)
1599 spin_lock(&imp->imp_lock);
1600 IMPORT_SET_STATE_NOLOCK(imp, LUSTRE_IMP_CLOSED);
1601 imp->imp_generation++;
1602 spin_unlock(&imp->imp_lock);
1603 ptlrpc_abort_inflight(imp);
1608 /* Adaptive Timeout utils */
1609 extern unsigned int at_min, at_max, at_history;
1611 /* Update at_current with the specified value (bounded by at_min and at_max),
1612 * as well as the AT history "bins".
1613 * - Bin into timeslices using AT_BINS bins.
1614 * - This gives us a max of the last at_history seconds without the storage,
1615 * but still smoothing out a return to normalcy from a slow response.
1616 * - (E.g. remember the maximum latency in each minute of the last 4 minutes.)
1618 int at_measured(struct adaptive_timeout *at, unsigned int val)
1620 unsigned int old = at->at_current;
1621 time_t now = cfs_time_current_sec();
1622 time_t binlimit = max_t(time_t, at_history / AT_BINS, 1);
1625 CDEBUG(D_OTHER, "add %u to %p time=%lu v=%u (%u %u %u %u)\n",
1626 val, at, now - at->at_binstart, at->at_current,
1627 at->at_hist[0], at->at_hist[1], at->at_hist[2], at->at_hist[3]);
1630 /* 0's don't count, because we never want our timeout to
1631 drop to 0, and because 0 could mean an error */
1634 spin_lock(&at->at_lock);
1636 if (unlikely(at->at_binstart == 0)) {
1637 /* Special case to remove default from history */
1638 at->at_current = val;
1639 at->at_worst_ever = val;
1640 at->at_worst_time = now;
1641 at->at_hist[0] = val;
1642 at->at_binstart = now;
1643 } else if (now - at->at_binstart < binlimit ) {
1645 at->at_hist[0] = max(val, at->at_hist[0]);
1646 at->at_current = max(val, at->at_current);
1649 unsigned int maxv = val;
1650 /* move bins over */
1651 shift = (now - at->at_binstart) / binlimit;
1653 for(i = AT_BINS - 1; i >= 0; i--) {
1655 at->at_hist[i] = at->at_hist[i - shift];
1656 maxv = max(maxv, at->at_hist[i]);
1661 at->at_hist[0] = val;
1662 at->at_current = maxv;
1663 at->at_binstart += shift * binlimit;
1666 if (at->at_current > at->at_worst_ever) {
1667 at->at_worst_ever = at->at_current;
1668 at->at_worst_time = now;
1671 if (at->at_flags & AT_FLG_NOHIST)
1672 /* Only keep last reported val; keeping the rest of the history
1674 at->at_current = val;
1677 at->at_current = min(at->at_current, at_max);
1678 at->at_current = max(at->at_current, at_min);
1680 if (at->at_current != old)
1681 CDEBUG(D_OTHER, "AT %p change: old=%u new=%u delta=%d "
1682 "(val=%u) hist %u %u %u %u\n", at,
1683 old, at->at_current, at->at_current - old, val,
1684 at->at_hist[0], at->at_hist[1], at->at_hist[2],
1687 /* if we changed, report the old value */
1688 old = (at->at_current != old) ? old : 0;
1690 spin_unlock(&at->at_lock);
1694 /* Find the imp_at index for a given portal; assign if space available */
1695 int import_at_get_index(struct obd_import *imp, int portal)
1697 struct imp_at *at = &imp->imp_at;
1700 for (i = 0; i < IMP_AT_MAX_PORTALS; i++) {
1701 if (at->iat_portal[i] == portal)
1703 if (at->iat_portal[i] == 0)
1708 /* Not found in list, add it under a lock */
1709 spin_lock(&imp->imp_lock);
1711 /* Check unused under lock */
1712 for (; i < IMP_AT_MAX_PORTALS; i++) {
1713 if (at->iat_portal[i] == portal)
1715 if (at->iat_portal[i] == 0)
1720 /* Not enough portals? */
1721 LASSERT(i < IMP_AT_MAX_PORTALS);
1723 at->iat_portal[i] = portal;
1725 spin_unlock(&imp->imp_lock);