4 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
6 * This program is free software; you can redistribute it and/or modify
7 * it under the terms of the GNU General Public License version 2 only,
8 * as published by the Free Software Foundation.
10 * This program is distributed in the hope that it will be useful, but
11 * WITHOUT ANY WARRANTY; without even the implied warranty of
12 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
13 * General Public License version 2 for more details (a copy is included
14 * in the LICENSE file that accompanied this code).
16 * You should have received a copy of the GNU General Public License
17 * version 2 along with this program; If not, see
18 * http://www.sun.com/software/products/lustre/docs/GPLv2.pdf
20 * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
21 * CA 95054 USA or visit www.sun.com if you need additional information or
27 * Copyright (c) 2003, 2010, Oracle and/or its affiliates. All rights reserved.
28 * Use is subject to license terms.
30 * Copyright (c) 2011, 2015, Intel Corporation.
33 * This file is part of Lustre, http://www.lustre.org/
34 * Lustre is a trademark of Sun Microsystems, Inc.
36 * lustre/ptlrpc/import.c
38 * Author: Mike Shaver <shaver@clusterfs.com>
41 #define DEBUG_SUBSYSTEM S_RPC
43 #include <linux/kthread.h>
44 #include <obd_support.h>
45 #include <lustre_ha.h>
46 #include <lustre_net.h>
47 #include <lustre_import.h>
48 #include <lustre_export.h>
50 #include <obd_cksum.h>
51 #include <obd_class.h>
53 #include "ptlrpc_internal.h"
55 struct ptlrpc_connect_async_args {
56 __u64 pcaa_peer_committed;
57 int pcaa_initial_connect;
61 * Updates import \a imp current state to provided \a state value
62 * Helper function. Must be called under imp_lock.
64 static void __import_set_state(struct obd_import *imp,
65 enum lustre_imp_state state)
68 case LUSTRE_IMP_CLOSED:
70 case LUSTRE_IMP_DISCON:
71 case LUSTRE_IMP_CONNECTING:
73 case LUSTRE_IMP_REPLAY_WAIT:
74 imp->imp_replay_state = LUSTRE_IMP_REPLAY_LOCKS;
77 imp->imp_replay_state = LUSTRE_IMP_REPLAY;
79 imp->imp_state = state;
80 imp->imp_state_hist[imp->imp_state_hist_idx].ish_state = state;
81 imp->imp_state_hist[imp->imp_state_hist_idx].ish_time =
82 cfs_time_current_sec();
83 imp->imp_state_hist_idx = (imp->imp_state_hist_idx + 1) %
87 /* A CLOSED import should remain so. */
88 #define IMPORT_SET_STATE_NOLOCK(imp, state) \
90 if (imp->imp_state != LUSTRE_IMP_CLOSED) { \
91 CDEBUG(D_HA, "%p %s: changing import state from %s to %s\n", \
92 imp, obd2cli_tgt(imp->imp_obd), \
93 ptlrpc_import_state_name(imp->imp_state), \
94 ptlrpc_import_state_name(state)); \
95 __import_set_state(imp, state); \
99 #define IMPORT_SET_STATE(imp, state) \
101 spin_lock(&imp->imp_lock); \
102 IMPORT_SET_STATE_NOLOCK(imp, state); \
103 spin_unlock(&imp->imp_lock); \
106 void ptlrpc_import_enter_resend(struct obd_import *imp)
108 IMPORT_SET_STATE(imp, LUSTRE_IMP_RECOVER);
110 EXPORT_SYMBOL(ptlrpc_import_enter_resend);
113 static int ptlrpc_connect_interpret(const struct lu_env *env,
114 struct ptlrpc_request *request,
115 void * data, int rc);
116 int ptlrpc_import_recovery_state_machine(struct obd_import *imp);
118 /* Only this function is allowed to change the import state when it is
119 * CLOSED. I would rather refcount the import and free it after
120 * disconnection like we do with exports. To do that, the client_obd
121 * will need to save the peer info somewhere other than in the import,
123 int ptlrpc_init_import(struct obd_import *imp)
125 spin_lock(&imp->imp_lock);
127 imp->imp_generation++;
128 imp->imp_state = LUSTRE_IMP_NEW;
130 spin_unlock(&imp->imp_lock);
134 EXPORT_SYMBOL(ptlrpc_init_import);
136 #define UUID_STR "_UUID"
137 void deuuidify(char *uuid, const char *prefix, char **uuid_start, int *uuid_len)
139 *uuid_start = !prefix || strncmp(uuid, prefix, strlen(prefix))
140 ? uuid : uuid + strlen(prefix);
142 *uuid_len = strlen(*uuid_start);
144 if (*uuid_len < strlen(UUID_STR))
147 if (!strncmp(*uuid_start + *uuid_len - strlen(UUID_STR),
148 UUID_STR, strlen(UUID_STR)))
149 *uuid_len -= strlen(UUID_STR);
153 * Returns true if import was FULL, false if import was already not
155 * @imp - import to be disconnected
156 * @conn_cnt - connection count (epoch) of the request that timed out
157 * and caused the disconnection. In some cases, multiple
158 * inflight requests can fail to a single target (e.g. OST
159 * bulk requests) and if one has already caused a reconnection
160 * (increasing the import->conn_cnt) the older failure should
161 * not also cause a reconnection. If zero it forces a reconnect.
163 int ptlrpc_set_import_discon(struct obd_import *imp, __u32 conn_cnt)
167 spin_lock(&imp->imp_lock);
169 if (imp->imp_state == LUSTRE_IMP_FULL &&
170 (conn_cnt == 0 || conn_cnt == imp->imp_conn_cnt)) {
174 deuuidify(obd2cli_tgt(imp->imp_obd), NULL,
175 &target_start, &target_len);
177 if (imp->imp_replayable) {
178 LCONSOLE_WARN("%s: Connection to %.*s (at %s) was "
179 "lost; in progress operations using this "
180 "service will wait for recovery to complete\n",
181 imp->imp_obd->obd_name, target_len, target_start,
182 libcfs_nid2str(imp->imp_connection->c_peer.nid));
184 LCONSOLE_ERROR_MSG(0x166, "%s: Connection to "
185 "%.*s (at %s) was lost; in progress "
186 "operations using this service will fail\n",
187 imp->imp_obd->obd_name,
188 target_len, target_start,
189 libcfs_nid2str(imp->imp_connection->c_peer.nid));
191 IMPORT_SET_STATE_NOLOCK(imp, LUSTRE_IMP_DISCON);
192 spin_unlock(&imp->imp_lock);
194 if (obd_dump_on_timeout)
195 libcfs_debug_dumplog();
197 obd_import_event(imp->imp_obd, imp, IMP_EVENT_DISCON);
200 spin_unlock(&imp->imp_lock);
201 CDEBUG(D_HA, "%s: import %p already %s (conn %u, was %u): %s\n",
202 imp->imp_client->cli_name, imp,
203 (imp->imp_state == LUSTRE_IMP_FULL &&
204 imp->imp_conn_cnt > conn_cnt) ?
205 "reconnected" : "not connected", imp->imp_conn_cnt,
206 conn_cnt, ptlrpc_import_state_name(imp->imp_state));
212 /* Must be called with imp_lock held! */
213 static void ptlrpc_deactivate_and_unlock_import(struct obd_import *imp)
216 assert_spin_locked(&imp->imp_lock);
218 CDEBUG(D_HA, "setting import %s INVALID\n", obd2cli_tgt(imp->imp_obd));
219 imp->imp_invalid = 1;
220 imp->imp_generation++;
221 spin_unlock(&imp->imp_lock);
223 ptlrpc_abort_inflight(imp);
224 obd_import_event(imp->imp_obd, imp, IMP_EVENT_INACTIVE);
230 * This acts as a barrier; all existing requests are rejected, and
231 * no new requests will be accepted until the import is valid again.
233 void ptlrpc_deactivate_import(struct obd_import *imp)
235 spin_lock(&imp->imp_lock);
236 ptlrpc_deactivate_and_unlock_import(imp);
238 EXPORT_SYMBOL(ptlrpc_deactivate_import);
241 ptlrpc_inflight_deadline(struct ptlrpc_request *req, time_t now)
245 if (!(((req->rq_phase == RQ_PHASE_RPC) && !req->rq_waiting) ||
246 (req->rq_phase == RQ_PHASE_BULK) ||
247 (req->rq_phase == RQ_PHASE_NEW)))
250 if (req->rq_timedout)
253 if (req->rq_phase == RQ_PHASE_NEW)
256 dl = req->rq_deadline;
264 static unsigned int ptlrpc_inflight_timeout(struct obd_import *imp)
266 time_t now = cfs_time_current_sec();
267 struct list_head *tmp, *n;
268 struct ptlrpc_request *req;
269 unsigned int timeout = 0;
271 spin_lock(&imp->imp_lock);
272 list_for_each_safe(tmp, n, &imp->imp_sending_list) {
273 req = list_entry(tmp, struct ptlrpc_request, rq_list);
274 timeout = max(ptlrpc_inflight_deadline(req, now), timeout);
276 spin_unlock(&imp->imp_lock);
281 * This function will invalidate the import, if necessary, then block
282 * for all the RPC completions, and finally notify the obd to
283 * invalidate its state (ie cancel locks, clear pending requests,
286 void ptlrpc_invalidate_import(struct obd_import *imp)
288 struct list_head *tmp, *n;
289 struct ptlrpc_request *req;
290 struct l_wait_info lwi;
291 unsigned int timeout;
294 atomic_inc(&imp->imp_inval_count);
296 if (!imp->imp_invalid || imp->imp_obd->obd_no_recov)
297 ptlrpc_deactivate_import(imp);
299 CFS_FAIL_TIMEOUT(OBD_FAIL_MGS_CONNECT_NET, 3 * cfs_fail_val / 2);
300 LASSERT(imp->imp_invalid);
302 /* Wait forever until inflight == 0. We really can't do it another
303 * way because in some cases we need to wait for very long reply
304 * unlink. We can't do anything before that because there is really
305 * no guarantee that some rdma transfer is not in progress right now. */
307 /* Calculate max timeout for waiting on rpcs to error
308 * out. Use obd_timeout if calculated value is smaller
310 if (!OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_REPL_UNLINK)) {
311 timeout = ptlrpc_inflight_timeout(imp);
312 timeout += timeout / 3;
315 timeout = obd_timeout;
317 /* decrease the interval to increase race condition */
321 CDEBUG(D_RPCTRACE,"Sleeping %d sec for inflight to error out\n",
324 /* Wait for all requests to error out and call completion
325 * callbacks. Cap it at obd_timeout -- these should all
326 * have been locally cancelled by ptlrpc_abort_inflight. */
327 lwi = LWI_TIMEOUT_INTERVAL(
328 cfs_timeout_cap(cfs_time_seconds(timeout)),
329 (timeout > 1)?cfs_time_seconds(1):cfs_time_seconds(1)/2,
331 rc = l_wait_event(imp->imp_recovery_waitq,
332 (atomic_read(&imp->imp_inflight) == 0),
335 const char *cli_tgt = obd2cli_tgt(imp->imp_obd);
337 CERROR("%s: rc = %d waiting for callback (%d != 0)\n",
338 cli_tgt, rc, atomic_read(&imp->imp_inflight));
340 spin_lock(&imp->imp_lock);
341 if (atomic_read(&imp->imp_inflight) == 0) {
342 int count = atomic_read(&imp->imp_unregistering);
344 /* We know that "unregistering" rpcs only can
345 * survive in sending or delaying lists (they
346 * maybe waiting for long reply unlink in
347 * sluggish nets). Let's check this. If there
348 * is no inflight and unregistering != 0, this
350 LASSERTF(count == 0, "Some RPCs are still "
351 "unregistering: %d\n", count);
353 /* Let's save one loop as soon as inflight have
354 * dropped to zero. No new inflights possible at
358 list_for_each_safe(tmp, n,
359 &imp->imp_sending_list) {
360 req = list_entry(tmp,
361 struct ptlrpc_request,
363 DEBUG_REQ(D_ERROR, req,
364 "still on sending list");
366 list_for_each_safe(tmp, n,
367 &imp->imp_delayed_list) {
368 req = list_entry(tmp,
369 struct ptlrpc_request,
371 DEBUG_REQ(D_ERROR, req,
372 "still on delayed list");
375 CERROR("%s: RPCs in \"%s\" phase found (%d). "
376 "Network is sluggish? Waiting them "
377 "to error out.\n", cli_tgt,
378 ptlrpc_phase2str(RQ_PHASE_UNREGISTERING),
379 atomic_read(&imp->imp_unregistering));
381 spin_unlock(&imp->imp_lock);
386 * Let's additionally check that no new rpcs added to import in
387 * "invalidate" state.
389 LASSERT(atomic_read(&imp->imp_inflight) == 0);
390 obd_import_event(imp->imp_obd, imp, IMP_EVENT_INVALIDATE);
391 sptlrpc_import_flush_all_ctx(imp);
393 atomic_dec(&imp->imp_inval_count);
394 wake_up_all(&imp->imp_recovery_waitq);
396 EXPORT_SYMBOL(ptlrpc_invalidate_import);
398 /* unset imp_invalid */
399 void ptlrpc_activate_import(struct obd_import *imp)
401 struct obd_device *obd = imp->imp_obd;
403 spin_lock(&imp->imp_lock);
404 if (imp->imp_deactive != 0) {
405 spin_unlock(&imp->imp_lock);
409 imp->imp_invalid = 0;
410 spin_unlock(&imp->imp_lock);
411 obd_import_event(obd, imp, IMP_EVENT_ACTIVE);
413 EXPORT_SYMBOL(ptlrpc_activate_import);
415 void ptlrpc_pinger_force(struct obd_import *imp)
417 CDEBUG(D_HA, "%s: waking up pinger s:%s\n", obd2cli_tgt(imp->imp_obd),
418 ptlrpc_import_state_name(imp->imp_state));
420 spin_lock(&imp->imp_lock);
421 imp->imp_force_verify = 1;
422 spin_unlock(&imp->imp_lock);
424 if (imp->imp_state != LUSTRE_IMP_CONNECTING)
425 ptlrpc_pinger_wake_up();
427 EXPORT_SYMBOL(ptlrpc_pinger_force);
429 void ptlrpc_fail_import(struct obd_import *imp, __u32 conn_cnt)
433 LASSERT(!imp->imp_dlm_fake);
435 if (ptlrpc_set_import_discon(imp, conn_cnt)) {
436 if (!imp->imp_replayable) {
437 CDEBUG(D_HA, "import %s@%s for %s not replayable, "
438 "auto-deactivating\n",
439 obd2cli_tgt(imp->imp_obd),
440 imp->imp_connection->c_remote_uuid.uuid,
441 imp->imp_obd->obd_name);
442 ptlrpc_deactivate_import(imp);
445 ptlrpc_pinger_force(imp);
450 int ptlrpc_reconnect_import(struct obd_import *imp)
453 struct l_wait_info lwi;
454 int secs = cfs_time_seconds(obd_timeout);
457 ptlrpc_pinger_force(imp);
459 CDEBUG(D_HA, "%s: recovery started, waiting %u seconds\n",
460 obd2cli_tgt(imp->imp_obd), secs);
462 lwi = LWI_TIMEOUT(secs, NULL, NULL);
463 rc = l_wait_event(imp->imp_recovery_waitq,
464 !ptlrpc_import_in_recovery(imp), &lwi);
465 CDEBUG(D_HA, "%s: recovery finished s:%s\n", obd2cli_tgt(imp->imp_obd),
466 ptlrpc_import_state_name(imp->imp_state));
469 ptlrpc_set_import_discon(imp, 0);
470 /* Force a new connect attempt */
471 ptlrpc_invalidate_import(imp);
472 /* Do a fresh connect next time by zeroing the handle */
473 ptlrpc_disconnect_import(imp, 1);
474 /* Wait for all invalidate calls to finish */
475 if (atomic_read(&imp->imp_inval_count) > 0) {
476 struct l_wait_info lwi = LWI_INTR(LWI_ON_SIGNAL_NOOP, NULL);
479 rc = l_wait_event(imp->imp_recovery_waitq,
480 (atomic_read(&imp->imp_inval_count) == 0),
483 CERROR("Interrupted, inval=%d\n",
484 atomic_read(&imp->imp_inval_count));
487 /* Allow reconnect attempts */
488 imp->imp_obd->obd_no_recov = 0;
489 /* Remove 'invalid' flag */
490 ptlrpc_activate_import(imp);
491 /* Attempt a new connect */
492 ptlrpc_recover_import(imp, NULL, 0);
496 EXPORT_SYMBOL(ptlrpc_reconnect_import);
499 * Connection on import \a imp is changed to another one (if more than one is
500 * present). We typically chose connection that we have not tried to connect to
503 static int import_select_connection(struct obd_import *imp)
505 struct obd_import_conn *imp_conn = NULL, *conn;
506 struct obd_export *dlmexp;
508 int target_len, tried_all = 1;
511 spin_lock(&imp->imp_lock);
513 if (list_empty(&imp->imp_conn_list)) {
514 CERROR("%s: no connections available\n",
515 imp->imp_obd->obd_name);
516 spin_unlock(&imp->imp_lock);
520 list_for_each_entry(conn, &imp->imp_conn_list, oic_item) {
521 CDEBUG(D_HA, "%s: connect to NID %s last attempt "LPU64"\n",
522 imp->imp_obd->obd_name,
523 libcfs_nid2str(conn->oic_conn->c_peer.nid),
524 conn->oic_last_attempt);
526 /* If we have not tried this connection since
527 the last successful attempt, go with this one */
528 if ((conn->oic_last_attempt == 0) ||
529 cfs_time_beforeq_64(conn->oic_last_attempt,
530 imp->imp_last_success_conn)) {
536 /* If all of the connections have already been tried
537 since the last successful connection; just choose the
538 least recently used */
541 else if (cfs_time_before_64(conn->oic_last_attempt,
542 imp_conn->oic_last_attempt))
546 /* if not found, simply choose the current one */
547 if (!imp_conn || imp->imp_force_reconnect) {
548 LASSERT(imp->imp_conn_current);
549 imp_conn = imp->imp_conn_current;
552 LASSERT(imp_conn->oic_conn);
554 /* If we've tried everything, and we're back to the beginning of the
555 list, increase our timeout and try again. It will be reset when
556 we do finally connect. (FIXME: really we should wait for all network
557 state associated with the last connection attempt to drain before
558 trying to reconnect on it.) */
559 if (tried_all && (imp->imp_conn_list.next == &imp_conn->oic_item)) {
560 struct adaptive_timeout *at = &imp->imp_at.iat_net_latency;
561 if (at_get(at) < CONNECTION_SWITCH_MAX) {
562 at_measured(at, at_get(at) + CONNECTION_SWITCH_INC);
563 if (at_get(at) > CONNECTION_SWITCH_MAX)
564 at_reset(at, CONNECTION_SWITCH_MAX);
566 LASSERT(imp_conn->oic_last_attempt);
567 CDEBUG(D_HA, "%s: tried all connections, increasing latency "
568 "to %ds\n", imp->imp_obd->obd_name, at_get(at));
571 imp_conn->oic_last_attempt = cfs_time_current_64();
573 /* switch connection, don't mind if it's same as the current one */
574 if (imp->imp_connection)
575 ptlrpc_connection_put(imp->imp_connection);
576 imp->imp_connection = ptlrpc_connection_addref(imp_conn->oic_conn);
578 dlmexp = class_conn2export(&imp->imp_dlm_handle);
579 LASSERT(dlmexp != NULL);
580 if (dlmexp->exp_connection)
581 ptlrpc_connection_put(dlmexp->exp_connection);
582 dlmexp->exp_connection = ptlrpc_connection_addref(imp_conn->oic_conn);
583 class_export_put(dlmexp);
585 if (imp->imp_conn_current != imp_conn) {
586 if (imp->imp_conn_current) {
587 deuuidify(obd2cli_tgt(imp->imp_obd), NULL,
588 &target_start, &target_len);
590 CDEBUG(D_HA, "%s: Connection changing to"
592 imp->imp_obd->obd_name,
593 target_len, target_start,
594 libcfs_nid2str(imp_conn->oic_conn->c_peer.nid));
597 imp->imp_conn_current = imp_conn;
600 CDEBUG(D_HA, "%s: import %p using connection %s/%s\n",
601 imp->imp_obd->obd_name, imp, imp_conn->oic_uuid.uuid,
602 libcfs_nid2str(imp_conn->oic_conn->c_peer.nid));
604 spin_unlock(&imp->imp_lock);
610 * must be called under imp_lock
612 static int ptlrpc_first_transno(struct obd_import *imp, __u64 *transno)
614 struct ptlrpc_request *req;
615 struct list_head *tmp;
617 /* The requests in committed_list always have smaller transnos than
618 * the requests in replay_list */
619 if (!list_empty(&imp->imp_committed_list)) {
620 tmp = imp->imp_committed_list.next;
621 req = list_entry(tmp, struct ptlrpc_request, rq_replay_list);
622 *transno = req->rq_transno;
623 if (req->rq_transno == 0) {
624 DEBUG_REQ(D_ERROR, req, "zero transno in committed_list");
629 if (!list_empty(&imp->imp_replay_list)) {
630 tmp = imp->imp_replay_list.next;
631 req = list_entry(tmp, struct ptlrpc_request, rq_replay_list);
632 *transno = req->rq_transno;
633 if (req->rq_transno == 0) {
634 DEBUG_REQ(D_ERROR, req, "zero transno in replay_list");
643 * Attempt to (re)connect import \a imp. This includes all preparations,
644 * initializing CONNECT RPC request and passing it to ptlrpcd for
646 * Returns 0 on success or error code.
648 int ptlrpc_connect_import(struct obd_import *imp)
650 struct obd_device *obd = imp->imp_obd;
651 int initial_connect = 0;
653 __u64 committed_before_reconnect = 0;
654 struct ptlrpc_request *request;
655 char *bufs[] = { NULL,
656 obd2cli_tgt(imp->imp_obd),
658 (char *)&imp->imp_dlm_handle,
659 (char *)&imp->imp_connect_data };
660 struct ptlrpc_connect_async_args *aa;
664 spin_lock(&imp->imp_lock);
665 if (imp->imp_state == LUSTRE_IMP_CLOSED) {
666 spin_unlock(&imp->imp_lock);
667 CERROR("can't connect to a closed import\n");
669 } else if (imp->imp_state == LUSTRE_IMP_FULL) {
670 spin_unlock(&imp->imp_lock);
671 CERROR("already connected\n");
673 } else if (imp->imp_state == LUSTRE_IMP_CONNECTING) {
674 spin_unlock(&imp->imp_lock);
675 CERROR("already connecting\n");
679 IMPORT_SET_STATE_NOLOCK(imp, LUSTRE_IMP_CONNECTING);
682 imp->imp_resend_replay = 0;
684 if (!lustre_handle_is_used(&imp->imp_remote_handle))
687 committed_before_reconnect = imp->imp_peer_committed_transno;
689 set_transno = ptlrpc_first_transno(imp,
690 &imp->imp_connect_data.ocd_transno);
691 spin_unlock(&imp->imp_lock);
693 rc = import_select_connection(imp);
697 rc = sptlrpc_import_sec_adapt(imp, NULL, NULL);
701 /* Reset connect flags to the originally requested flags, in case
702 * the server is updated on-the-fly we will get the new features. */
703 imp->imp_connect_data.ocd_connect_flags = imp->imp_connect_flags_orig;
704 /* Reset ocd_version each time so the server knows the exact versions */
705 imp->imp_connect_data.ocd_version = LUSTRE_VERSION_CODE;
706 imp->imp_msghdr_flags &= ~MSGHDR_AT_SUPPORT;
707 imp->imp_msghdr_flags &= ~MSGHDR_CKSUM_INCOMPAT18;
709 rc = obd_reconnect(NULL, imp->imp_obd->obd_self_export, obd,
710 &obd->obd_uuid, &imp->imp_connect_data, NULL);
714 request = ptlrpc_request_alloc(imp, &RQF_MDS_CONNECT);
716 GOTO(out, rc = -ENOMEM);
718 rc = ptlrpc_request_bufs_pack(request, LUSTRE_OBD_VERSION,
719 imp->imp_connect_op, bufs, NULL);
721 ptlrpc_request_free(request);
725 /* Report the rpc service time to the server so that it knows how long
726 * to wait for clients to join recovery */
727 lustre_msg_set_service_time(request->rq_reqmsg,
728 at_timeout2est(request->rq_timeout));
730 /* The amount of time we give the server to process the connect req.
731 * import_select_connection will increase the net latency on
732 * repeated reconnect attempts to cover slow networks.
733 * We override/ignore the server rpc completion estimate here,
734 * which may be large if this is a reconnect attempt */
735 request->rq_timeout = INITIAL_CONNECT_TIMEOUT;
736 lustre_msg_set_timeout(request->rq_reqmsg, request->rq_timeout);
738 request->rq_no_resend = request->rq_no_delay = 1;
739 request->rq_send_state = LUSTRE_IMP_CONNECTING;
740 /* Allow a slightly larger reply for future growth compatibility */
741 req_capsule_set_size(&request->rq_pill, &RMF_CONNECT_DATA, RCL_SERVER,
742 sizeof(struct obd_connect_data)+16*sizeof(__u64));
743 ptlrpc_request_set_replen(request);
744 request->rq_interpret_reply = ptlrpc_connect_interpret;
746 CLASSERT(sizeof (*aa) <= sizeof (request->rq_async_args));
747 aa = ptlrpc_req_async_args(request);
748 memset(aa, 0, sizeof *aa);
750 aa->pcaa_peer_committed = committed_before_reconnect;
751 aa->pcaa_initial_connect = initial_connect;
753 if (aa->pcaa_initial_connect) {
754 spin_lock(&imp->imp_lock);
755 imp->imp_replayable = 1;
756 spin_unlock(&imp->imp_lock);
757 lustre_msg_add_op_flags(request->rq_reqmsg,
758 MSG_CONNECT_INITIAL);
762 lustre_msg_add_op_flags(request->rq_reqmsg,
763 MSG_CONNECT_TRANSNO);
765 DEBUG_REQ(D_RPCTRACE, request, "(re)connect request (timeout %d)",
766 request->rq_timeout);
767 ptlrpcd_add_req(request);
771 IMPORT_SET_STATE(imp, LUSTRE_IMP_DISCON);
775 EXPORT_SYMBOL(ptlrpc_connect_import);
777 static void ptlrpc_maybe_ping_import_soon(struct obd_import *imp)
781 spin_lock(&imp->imp_lock);
782 force_verify = imp->imp_force_verify != 0;
783 spin_unlock(&imp->imp_lock);
786 ptlrpc_pinger_wake_up();
789 static int ptlrpc_busy_reconnect(int rc)
791 return (rc == -EBUSY) || (rc == -EAGAIN);
794 static int ptlrpc_connect_set_flags(struct obd_import *imp,
795 struct obd_connect_data *ocd,
796 __u64 old_connect_flags,
797 struct obd_export *exp, int init_connect)
800 struct client_obd *cli = &imp->imp_obd->u.cli;
802 if ((imp->imp_connect_flags_orig & OBD_CONNECT_IBITS) &&
803 !(ocd->ocd_connect_flags & OBD_CONNECT_IBITS)) {
804 LCONSOLE_WARN("%s: MDS %s does not support ibits "
805 "lock, either very old or invalid: "
806 "requested "LPX64", replied "LPX64"\n",
807 imp->imp_obd->obd_name,
808 imp->imp_connection->c_remote_uuid.uuid,
809 imp->imp_connect_flags_orig,
810 ocd->ocd_connect_flags);
814 spin_lock(&imp->imp_lock);
815 list_del(&imp->imp_conn_current->oic_item);
816 list_add(&imp->imp_conn_current->oic_item,
817 &imp->imp_conn_list);
818 imp->imp_last_success_conn =
819 imp->imp_conn_current->oic_last_attempt;
821 spin_unlock(&imp->imp_lock);
824 if (!warned && (ocd->ocd_connect_flags & OBD_CONNECT_VERSION) &&
825 (ocd->ocd_version > LUSTRE_VERSION_CODE +
826 LUSTRE_VERSION_OFFSET_WARN ||
827 ocd->ocd_version < LUSTRE_VERSION_CODE -
828 LUSTRE_VERSION_OFFSET_WARN)) {
829 /* Sigh, some compilers do not like #ifdef in the middle
830 of macro arguments */
831 const char *older = "older than client. "
832 "Consider upgrading server";
833 const char *newer = "newer than client. "
834 "Consider recompiling application";
836 LCONSOLE_WARN("Server %s version (%d.%d.%d.%d) "
838 obd2cli_tgt(imp->imp_obd),
839 OBD_OCD_VERSION_MAJOR(ocd->ocd_version),
840 OBD_OCD_VERSION_MINOR(ocd->ocd_version),
841 OBD_OCD_VERSION_PATCH(ocd->ocd_version),
842 OBD_OCD_VERSION_FIX(ocd->ocd_version),
843 ocd->ocd_version > LUSTRE_VERSION_CODE ?
844 newer : older, LUSTRE_VERSION_STRING);
848 #if LUSTRE_VERSION_CODE < OBD_OCD_VERSION(3, 0, 53, 0)
849 /* Check if server has LU-1252 fix applied to not always swab
850 * the IR MNE entries. Do this only once per connection. This
851 * fixup is version-limited, because we don't want to carry the
852 * OBD_CONNECT_MNE_SWAB flag around forever, just so long as we
853 * need interop with unpatched 2.2 servers. For newer servers,
854 * the client will do MNE swabbing only as needed. LU-1644 */
855 if (unlikely((ocd->ocd_connect_flags & OBD_CONNECT_VERSION) &&
856 !(ocd->ocd_connect_flags & OBD_CONNECT_MNE_SWAB) &&
857 OBD_OCD_VERSION_MAJOR(ocd->ocd_version) == 2 &&
858 OBD_OCD_VERSION_MINOR(ocd->ocd_version) == 2 &&
859 OBD_OCD_VERSION_PATCH(ocd->ocd_version) < 55 &&
860 strcmp(imp->imp_obd->obd_type->typ_name,
861 LUSTRE_MGC_NAME) == 0))
862 imp->imp_need_mne_swab = 1;
863 else /* clear if server was upgraded since last connect */
864 imp->imp_need_mne_swab = 0;
867 if (ocd->ocd_connect_flags & OBD_CONNECT_CKSUM) {
868 /* We sent to the server ocd_cksum_types with bits set
869 * for algorithms we understand. The server masked off
870 * the checksum types it doesn't support */
871 if ((ocd->ocd_cksum_types &
872 cksum_types_supported_client()) == 0) {
873 LCONSOLE_WARN("The negotiation of the checksum "
874 "alogrithm to use with server %s "
875 "failed (%x/%x), disabling "
877 obd2cli_tgt(imp->imp_obd),
878 ocd->ocd_cksum_types,
879 cksum_types_supported_client());
880 cli->cl_checksum = 0;
881 cli->cl_supp_cksum_types = OBD_CKSUM_ADLER;
883 cli->cl_supp_cksum_types = ocd->ocd_cksum_types;
886 /* The server does not support OBD_CONNECT_CKSUM.
887 * Enforce ADLER for backward compatibility*/
888 cli->cl_supp_cksum_types = OBD_CKSUM_ADLER;
890 cli->cl_cksum_type = cksum_type_select(cli->cl_supp_cksum_types);
892 if (ocd->ocd_connect_flags & OBD_CONNECT_BRW_SIZE)
893 cli->cl_max_pages_per_rpc =
894 min(ocd->ocd_brw_size >> PAGE_CACHE_SHIFT,
895 cli->cl_max_pages_per_rpc);
896 else if (imp->imp_connect_op == MDS_CONNECT ||
897 imp->imp_connect_op == MGS_CONNECT)
898 cli->cl_max_pages_per_rpc = 1;
900 LASSERT((cli->cl_max_pages_per_rpc <= PTLRPC_MAX_BRW_PAGES) &&
901 (cli->cl_max_pages_per_rpc > 0));
903 client_adjust_max_dirty(cli);
905 /* Update client max modify RPCs in flight with value returned
907 if (ocd->ocd_connect_flags & OBD_CONNECT_MULTIMODRPCS)
908 cli->cl_max_mod_rpcs_in_flight = min(
909 cli->cl_max_mod_rpcs_in_flight,
910 ocd->ocd_maxmodrpcs);
912 cli->cl_max_mod_rpcs_in_flight = 1;
914 /* Reset ns_connect_flags only for initial connect. It might be
915 * changed in while using FS and if we reset it in reconnect
916 * this leads to losing user settings done before such as
917 * disable lru_resize, etc. */
918 if (old_connect_flags != exp_connect_flags(exp) || init_connect) {
919 CDEBUG(D_HA, "%s: Resetting ns_connect_flags to server "
920 "flags: "LPX64"\n", imp->imp_obd->obd_name,
921 ocd->ocd_connect_flags);
922 imp->imp_obd->obd_namespace->ns_connect_flags =
923 ocd->ocd_connect_flags;
924 imp->imp_obd->obd_namespace->ns_orig_connect_flags =
925 ocd->ocd_connect_flags;
928 if (ocd->ocd_connect_flags & OBD_CONNECT_AT)
929 /* We need a per-message support flag, because
930 * a. we don't know if the incoming connect reply
931 * supports AT or not (in reply_in_callback)
932 * until we unpack it.
933 * b. failovered server means export and flags are gone
934 * (in ptlrpc_send_reply).
935 * Can only be set when we know AT is supported at
937 imp->imp_msghdr_flags |= MSGHDR_AT_SUPPORT;
939 imp->imp_msghdr_flags &= ~MSGHDR_AT_SUPPORT;
941 imp->imp_msghdr_flags |= MSGHDR_CKSUM_INCOMPAT18;
947 * Add all replay requests back to unreplied list before start replay,
948 * so that we can make sure the known replied XID is always increased
949 * only even if when replaying requests.
951 static void ptlrpc_prepare_replay(struct obd_import *imp)
953 struct ptlrpc_request *req;
955 if (imp->imp_state != LUSTRE_IMP_REPLAY ||
956 imp->imp_resend_replay)
959 /* If the server was restart during repaly, the requests may
960 * have been added to the unreplied list in former replay. */
961 spin_lock(&imp->imp_lock);
963 list_for_each_entry(req, &imp->imp_committed_list, rq_replay_list) {
964 if (list_empty(&req->rq_unreplied_list))
965 ptlrpc_add_unreplied(req);
968 list_for_each_entry(req, &imp->imp_replay_list, rq_replay_list) {
969 if (list_empty(&req->rq_unreplied_list))
970 ptlrpc_add_unreplied(req);
973 imp->imp_known_replied_xid = ptlrpc_known_replied_xid(imp);
974 spin_unlock(&imp->imp_lock);
978 * interpret_reply callback for connect RPCs.
979 * Looks into returned status of connect operation and decides
980 * what to do with the import - i.e enter recovery, promote it to
981 * full state for normal operations of disconnect it due to an error.
983 static int ptlrpc_connect_interpret(const struct lu_env *env,
984 struct ptlrpc_request *request,
987 struct ptlrpc_connect_async_args *aa = data;
988 struct obd_import *imp = request->rq_import;
989 struct lustre_handle old_hdl;
990 __u64 old_connect_flags;
992 struct obd_connect_data *ocd;
993 struct obd_export *exp;
997 spin_lock(&imp->imp_lock);
998 if (imp->imp_state == LUSTRE_IMP_CLOSED) {
999 imp->imp_connect_tried = 1;
1000 spin_unlock(&imp->imp_lock);
1005 /* if this reconnect to busy export - not need select new target
1007 imp->imp_force_reconnect = ptlrpc_busy_reconnect(rc);
1008 spin_unlock(&imp->imp_lock);
1009 ptlrpc_maybe_ping_import_soon(imp);
1012 spin_unlock(&imp->imp_lock);
1014 LASSERT(imp->imp_conn_current);
1016 msg_flags = lustre_msg_get_op_flags(request->rq_repmsg);
1018 ret = req_capsule_get_size(&request->rq_pill, &RMF_CONNECT_DATA,
1020 /* server replied obd_connect_data is always bigger */
1021 ocd = req_capsule_server_sized_get(&request->rq_pill,
1022 &RMF_CONNECT_DATA, ret);
1025 CERROR("%s: no connect data from server\n",
1026 imp->imp_obd->obd_name);
1031 spin_lock(&imp->imp_lock);
1033 /* All imports are pingable */
1034 imp->imp_pingable = 1;
1035 imp->imp_force_reconnect = 0;
1036 imp->imp_force_verify = 0;
1038 imp->imp_connect_data = *ocd;
1040 CDEBUG(D_HA, "%s: connect to target with instance %u\n",
1041 imp->imp_obd->obd_name, ocd->ocd_instance);
1042 exp = class_conn2export(&imp->imp_dlm_handle);
1044 spin_unlock(&imp->imp_lock);
1046 /* check that server granted subset of flags we asked for. */
1047 if ((ocd->ocd_connect_flags & imp->imp_connect_flags_orig) !=
1048 ocd->ocd_connect_flags) {
1049 CERROR("%s: Server didn't granted asked subset of flags: "
1050 "asked="LPX64" grranted="LPX64"\n",
1051 imp->imp_obd->obd_name,imp->imp_connect_flags_orig,
1052 ocd->ocd_connect_flags);
1053 GOTO(out, rc = -EPROTO);
1056 if (!(imp->imp_connect_flags_orig & OBD_CONNECT_LIGHTWEIGHT) &&
1057 (imp->imp_connect_flags_orig & OBD_CONNECT_MDS_MDS) &&
1058 (imp->imp_connect_flags_orig & OBD_CONNECT_FID) &&
1059 (ocd->ocd_connect_flags & OBD_CONNECT_VERSION)) {
1060 __u32 major = OBD_OCD_VERSION_MAJOR(ocd->ocd_version);
1061 __u32 minor = OBD_OCD_VERSION_MINOR(ocd->ocd_version);
1062 __u32 patch = OBD_OCD_VERSION_PATCH(ocd->ocd_version);
1064 /* We do not support the MDT-MDT interoperations with
1065 * different version MDT because of protocol changes. */
1066 if (unlikely(major != LUSTRE_MAJOR ||
1067 minor != LUSTRE_MINOR ||
1068 abs(patch - LUSTRE_PATCH) > 3)) {
1069 LCONSOLE_WARN("%s: import %p (%u.%u.%u.%u) tried the "
1070 "connection to different version MDT "
1071 "(%d.%d.%d.%d) %s\n",
1072 imp->imp_obd->obd_name, imp, LUSTRE_MAJOR,
1073 LUSTRE_MINOR, LUSTRE_PATCH, LUSTRE_FIX,
1074 major, minor, patch,
1075 OBD_OCD_VERSION_FIX(ocd->ocd_version),
1076 imp->imp_connection->c_remote_uuid.uuid);
1078 GOTO(out, rc = -EPROTO);
1083 /* This could happen if export is cleaned during the
1085 CERROR("%s: missing export after connect\n",
1086 imp->imp_obd->obd_name);
1087 GOTO(out, rc = -ENODEV);
1089 old_connect_flags = exp_connect_flags(exp);
1090 exp->exp_connect_data = *ocd;
1091 imp->imp_obd->obd_self_export->exp_connect_data = *ocd;
1093 /* The net statistics after (re-)connect is not valid anymore,
1094 * because may reflect other routing, etc. */
1095 at_reinit(&imp->imp_at.iat_net_latency, 0, 0);
1096 ptlrpc_at_adj_net_latency(request,
1097 lustre_msg_get_service_time(request->rq_repmsg));
1099 /* Import flags should be updated before waking import at FULL state */
1100 rc = ptlrpc_connect_set_flags(imp, ocd, old_connect_flags, exp,
1101 aa->pcaa_initial_connect);
1102 class_export_put(exp);
1106 obd_import_event(imp->imp_obd, imp, IMP_EVENT_OCD);
1108 if (aa->pcaa_initial_connect) {
1109 spin_lock(&imp->imp_lock);
1110 if (msg_flags & MSG_CONNECT_REPLAYABLE) {
1111 imp->imp_replayable = 1;
1112 spin_unlock(&imp->imp_lock);
1113 CDEBUG(D_HA, "connected to replayable target: %s\n",
1114 obd2cli_tgt(imp->imp_obd));
1116 imp->imp_replayable = 0;
1117 spin_unlock(&imp->imp_lock);
1120 /* if applies, adjust the imp->imp_msg_magic here
1121 * according to reply flags */
1123 imp->imp_remote_handle =
1124 *lustre_msg_get_handle(request->rq_repmsg);
1126 /* Initial connects are allowed for clients with non-random
1127 * uuids when servers are in recovery. Simply signal the
1128 * servers replay is complete and wait in REPLAY_WAIT. */
1129 if (msg_flags & MSG_CONNECT_RECOVERING) {
1130 CDEBUG(D_HA, "connect to %s during recovery\n",
1131 obd2cli_tgt(imp->imp_obd));
1132 IMPORT_SET_STATE(imp, LUSTRE_IMP_REPLAY_LOCKS);
1134 IMPORT_SET_STATE(imp, LUSTRE_IMP_FULL);
1135 ptlrpc_activate_import(imp);
1138 GOTO(finish, rc = 0);
1141 /* Determine what recovery state to move the import to. */
1142 if (MSG_CONNECT_RECONNECT & msg_flags) {
1143 memset(&old_hdl, 0, sizeof(old_hdl));
1144 if (!memcmp(&old_hdl, lustre_msg_get_handle(request->rq_repmsg),
1145 sizeof (old_hdl))) {
1146 LCONSOLE_WARN("Reconnect to %s (at @%s) failed due "
1147 "bad handle "LPX64"\n",
1148 obd2cli_tgt(imp->imp_obd),
1149 imp->imp_connection->c_remote_uuid.uuid,
1150 imp->imp_dlm_handle.cookie);
1151 GOTO(out, rc = -ENOTCONN);
1154 if (memcmp(&imp->imp_remote_handle,
1155 lustre_msg_get_handle(request->rq_repmsg),
1156 sizeof(imp->imp_remote_handle))) {
1157 int level = msg_flags & MSG_CONNECT_RECOVERING ?
1160 /* Bug 16611/14775: if server handle have changed,
1161 * that means some sort of disconnection happened.
1162 * If the server is not in recovery, that also means it
1163 * already erased all of our state because of previous
1164 * eviction. If it is in recovery - we are safe to
1165 * participate since we can reestablish all of our state
1166 * with server again */
1167 if ((MSG_CONNECT_RECOVERING & msg_flags)) {
1168 CDEBUG(level,"%s@%s changed server handle from "
1170 " but is still in recovery\n",
1171 obd2cli_tgt(imp->imp_obd),
1172 imp->imp_connection->c_remote_uuid.uuid,
1173 imp->imp_remote_handle.cookie,
1174 lustre_msg_get_handle(
1175 request->rq_repmsg)->cookie);
1177 LCONSOLE_WARN("Evicted from %s (at %s) "
1178 "after server handle changed from "
1179 LPX64" to "LPX64"\n",
1180 obd2cli_tgt(imp->imp_obd),
1181 imp->imp_connection-> \
1183 imp->imp_remote_handle.cookie,
1184 lustre_msg_get_handle(
1185 request->rq_repmsg)->cookie);
1189 imp->imp_remote_handle =
1190 *lustre_msg_get_handle(request->rq_repmsg);
1192 if (!(MSG_CONNECT_RECOVERING & msg_flags)) {
1193 IMPORT_SET_STATE(imp, LUSTRE_IMP_EVICTED);
1194 GOTO(finish, rc = 0);
1198 CDEBUG(D_HA, "reconnected to %s@%s after partition\n",
1199 obd2cli_tgt(imp->imp_obd),
1200 imp->imp_connection->c_remote_uuid.uuid);
1203 if (imp->imp_invalid) {
1204 CDEBUG(D_HA, "%s: reconnected but import is invalid; "
1205 "marking evicted\n", imp->imp_obd->obd_name);
1206 IMPORT_SET_STATE(imp, LUSTRE_IMP_EVICTED);
1207 } else if (MSG_CONNECT_RECOVERING & msg_flags) {
1208 CDEBUG(D_HA, "%s: reconnected to %s during replay\n",
1209 imp->imp_obd->obd_name,
1210 obd2cli_tgt(imp->imp_obd));
1212 spin_lock(&imp->imp_lock);
1213 imp->imp_resend_replay = 1;
1214 spin_unlock(&imp->imp_lock);
1216 IMPORT_SET_STATE(imp, imp->imp_replay_state);
1218 IMPORT_SET_STATE(imp, LUSTRE_IMP_RECOVER);
1220 } else if ((MSG_CONNECT_RECOVERING & msg_flags) && !imp->imp_invalid) {
1221 LASSERT(imp->imp_replayable);
1222 imp->imp_remote_handle =
1223 *lustre_msg_get_handle(request->rq_repmsg);
1224 imp->imp_last_replay_transno = 0;
1225 imp->imp_replay_cursor = &imp->imp_committed_list;
1226 IMPORT_SET_STATE(imp, LUSTRE_IMP_REPLAY);
1228 DEBUG_REQ(D_HA, request, "%s: evicting (reconnect/recover flags"
1229 " not set: %x)", imp->imp_obd->obd_name, msg_flags);
1230 imp->imp_remote_handle =
1231 *lustre_msg_get_handle(request->rq_repmsg);
1232 IMPORT_SET_STATE(imp, LUSTRE_IMP_EVICTED);
1235 /* Sanity checks for a reconnected import. */
1236 if (!(imp->imp_replayable) != !(msg_flags & MSG_CONNECT_REPLAYABLE)) {
1237 CERROR("imp_replayable flag does not match server "
1238 "after reconnect. We should LBUG right here.\n");
1241 if (lustre_msg_get_last_committed(request->rq_repmsg) > 0 &&
1242 lustre_msg_get_last_committed(request->rq_repmsg) <
1243 aa->pcaa_peer_committed) {
1244 CERROR("%s went back in time (transno "LPD64
1245 " was previously committed, server now claims "LPD64
1246 ")! See https://bugzilla.lustre.org/show_bug.cgi?"
1248 obd2cli_tgt(imp->imp_obd), aa->pcaa_peer_committed,
1249 lustre_msg_get_last_committed(request->rq_repmsg));
1253 ptlrpc_prepare_replay(imp);
1254 rc = ptlrpc_import_recovery_state_machine(imp);
1255 if (rc == -ENOTCONN) {
1256 CDEBUG(D_HA, "evicted/aborted by %s@%s during recovery;"
1257 "invalidating and reconnecting\n",
1258 obd2cli_tgt(imp->imp_obd),
1259 imp->imp_connection->c_remote_uuid.uuid);
1260 ptlrpc_connect_import(imp);
1261 imp->imp_connect_tried = 1;
1266 imp->imp_connect_tried = 1;
1269 IMPORT_SET_STATE(imp, LUSTRE_IMP_DISCON);
1270 if (rc == -EACCES) {
1272 * Give up trying to reconnect
1273 * EACCES means client has no permission for connection
1275 imp->imp_obd->obd_no_recov = 1;
1276 ptlrpc_deactivate_import(imp);
1279 if (rc == -EPROTO) {
1280 struct obd_connect_data *ocd;
1282 /* reply message might not be ready */
1283 if (request->rq_repmsg == NULL)
1286 ocd = req_capsule_server_get(&request->rq_pill,
1289 (ocd->ocd_connect_flags & OBD_CONNECT_VERSION) &&
1290 (ocd->ocd_version != LUSTRE_VERSION_CODE)) {
1291 /* Actually servers are only supposed to refuse
1292 connection from liblustre clients, so we should
1293 never see this from VFS context */
1294 LCONSOLE_ERROR_MSG(0x16a, "Server %s version "
1296 " refused connection from this client "
1297 "with an incompatible version (%s). "
1298 "Client must be recompiled\n",
1299 obd2cli_tgt(imp->imp_obd),
1300 OBD_OCD_VERSION_MAJOR(ocd->ocd_version),
1301 OBD_OCD_VERSION_MINOR(ocd->ocd_version),
1302 OBD_OCD_VERSION_PATCH(ocd->ocd_version),
1303 OBD_OCD_VERSION_FIX(ocd->ocd_version),
1304 LUSTRE_VERSION_STRING);
1305 ptlrpc_deactivate_import(imp);
1306 IMPORT_SET_STATE(imp, LUSTRE_IMP_CLOSED);
1311 ptlrpc_maybe_ping_import_soon(imp);
1313 CDEBUG(D_HA, "recovery of %s on %s failed (%d)\n",
1314 obd2cli_tgt(imp->imp_obd),
1315 (char *)imp->imp_connection->c_remote_uuid.uuid, rc);
1318 wake_up_all(&imp->imp_recovery_waitq);
1323 * interpret callback for "completed replay" RPCs.
1324 * \see signal_completed_replay
1326 static int completed_replay_interpret(const struct lu_env *env,
1327 struct ptlrpc_request *req,
1328 void * data, int rc)
1331 atomic_dec(&req->rq_import->imp_replay_inflight);
1332 if (req->rq_status == 0 &&
1333 !req->rq_import->imp_vbr_failed) {
1334 ptlrpc_import_recovery_state_machine(req->rq_import);
1336 if (req->rq_import->imp_vbr_failed) {
1338 "%s: version recovery fails, reconnecting\n",
1339 req->rq_import->imp_obd->obd_name);
1341 CDEBUG(D_HA, "%s: LAST_REPLAY message error: %d, "
1343 req->rq_import->imp_obd->obd_name,
1346 ptlrpc_connect_import(req->rq_import);
1353 * Let server know that we have no requests to replay anymore.
1354 * Achieved by just sending a PING request
1356 static int signal_completed_replay(struct obd_import *imp)
1358 struct ptlrpc_request *req;
1361 if (unlikely(OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_FINISH_REPLAY)))
1364 LASSERT(atomic_read(&imp->imp_replay_inflight) == 0);
1365 atomic_inc(&imp->imp_replay_inflight);
1367 req = ptlrpc_request_alloc_pack(imp, &RQF_OBD_PING, LUSTRE_OBD_VERSION,
1370 atomic_dec(&imp->imp_replay_inflight);
1374 ptlrpc_request_set_replen(req);
1375 req->rq_send_state = LUSTRE_IMP_REPLAY_WAIT;
1376 lustre_msg_add_flags(req->rq_reqmsg,
1377 MSG_LOCK_REPLAY_DONE | MSG_REQ_REPLAY_DONE);
1379 req->rq_timeout *= 3;
1380 req->rq_interpret_reply = completed_replay_interpret;
1382 ptlrpcd_add_req(req);
1387 * In kernel code all import invalidation happens in its own
1388 * separate thread, so that whatever application happened to encounter
1389 * a problem could still be killed or otherwise continue
1391 static int ptlrpc_invalidate_import_thread(void *data)
1393 struct obd_import *imp = data;
1397 unshare_fs_struct();
1399 CDEBUG(D_HA, "thread invalidate import %s to %s@%s\n",
1400 imp->imp_obd->obd_name, obd2cli_tgt(imp->imp_obd),
1401 imp->imp_connection->c_remote_uuid.uuid);
1403 ptlrpc_invalidate_import(imp);
1405 if (obd_dump_on_eviction) {
1406 CERROR("dump the log upon eviction\n");
1407 libcfs_debug_dumplog();
1410 IMPORT_SET_STATE(imp, LUSTRE_IMP_RECOVER);
1411 ptlrpc_import_recovery_state_machine(imp);
1413 class_import_put(imp);
1418 * This is the state machine for client-side recovery on import.
1420 * Typicaly we have two possibly paths. If we came to server and it is not
1421 * in recovery, we just enter IMP_EVICTED state, invalidate our import
1422 * state and reconnect from scratch.
1423 * If we came to server that is in recovery, we enter IMP_REPLAY import state.
1424 * We go through our list of requests to replay and send them to server one by
1426 * After sending all request from the list we change import state to
1427 * IMP_REPLAY_LOCKS and re-request all the locks we believe we have from server
1428 * and also all the locks we don't yet have and wait for server to grant us.
1429 * After that we send a special "replay completed" request and change import
1430 * state to IMP_REPLAY_WAIT.
1431 * Upon receiving reply to that "replay completed" RPC we enter IMP_RECOVER
1432 * state and resend all requests from sending list.
1433 * After that we promote import to FULL state and send all delayed requests
1434 * and import is fully operational after that.
1437 int ptlrpc_import_recovery_state_machine(struct obd_import *imp)
1445 if (imp->imp_state == LUSTRE_IMP_EVICTED) {
1446 deuuidify(obd2cli_tgt(imp->imp_obd), NULL,
1447 &target_start, &target_len);
1448 /* Don't care about MGC eviction */
1449 if (strcmp(imp->imp_obd->obd_type->typ_name,
1450 LUSTRE_MGC_NAME) != 0) {
1451 LCONSOLE_ERROR_MSG(0x167, "%s: This client was evicted "
1452 "by %.*s; in progress operations "
1453 "using this service will fail.\n",
1454 imp->imp_obd->obd_name, target_len,
1457 CDEBUG(D_HA, "evicted from %s@%s; invalidating\n",
1458 obd2cli_tgt(imp->imp_obd),
1459 imp->imp_connection->c_remote_uuid.uuid);
1460 /* reset vbr_failed flag upon eviction */
1461 spin_lock(&imp->imp_lock);
1462 imp->imp_vbr_failed = 0;
1463 spin_unlock(&imp->imp_lock);
1466 struct task_struct *task;
1467 /* bug 17802: XXX client_disconnect_export vs connect request
1468 * race. if client is evicted at this time then we start
1469 * invalidate thread without reference to import and import can
1470 * be freed at same time. */
1471 class_import_get(imp);
1472 task = kthread_run(ptlrpc_invalidate_import_thread, imp,
1475 class_import_put(imp);
1476 CERROR("error starting invalidate thread: %d\n", rc);
1485 if (imp->imp_state == LUSTRE_IMP_REPLAY) {
1486 CDEBUG(D_HA, "replay requested by %s\n",
1487 obd2cli_tgt(imp->imp_obd));
1488 rc = ptlrpc_replay_next(imp, &inflight);
1489 if (inflight == 0 &&
1490 atomic_read(&imp->imp_replay_inflight) == 0) {
1491 IMPORT_SET_STATE(imp, LUSTRE_IMP_REPLAY_LOCKS);
1492 rc = ldlm_replay_locks(imp);
1499 if (imp->imp_state == LUSTRE_IMP_REPLAY_LOCKS) {
1500 if (atomic_read(&imp->imp_replay_inflight) == 0) {
1501 IMPORT_SET_STATE(imp, LUSTRE_IMP_REPLAY_WAIT);
1502 rc = signal_completed_replay(imp);
1508 if (imp->imp_state == LUSTRE_IMP_REPLAY_WAIT) {
1509 if (atomic_read(&imp->imp_replay_inflight) == 0) {
1510 IMPORT_SET_STATE(imp, LUSTRE_IMP_RECOVER);
1514 if (imp->imp_state == LUSTRE_IMP_RECOVER) {
1515 struct ptlrpc_connection *conn = imp->imp_connection;
1517 rc = ptlrpc_resend(imp);
1520 IMPORT_SET_STATE(imp, LUSTRE_IMP_FULL);
1521 ptlrpc_activate_import(imp);
1523 LCONSOLE_INFO("%s: Connection restored to %s (at %s)\n",
1524 imp->imp_obd->obd_name,
1525 obd_uuid2str(&conn->c_remote_uuid),
1526 libcfs_nid2str(imp->imp_connection->c_peer.nid));
1529 if (imp->imp_state == LUSTRE_IMP_FULL) {
1530 wake_up_all(&imp->imp_recovery_waitq);
1531 ptlrpc_wake_delayed(imp);
1538 int ptlrpc_disconnect_import(struct obd_import *imp, int noclose)
1540 struct ptlrpc_request *req;
1544 if (imp->imp_obd->obd_force)
1545 GOTO(set_state, rc);
1547 switch (imp->imp_connect_op) {
1549 rq_opc = OST_DISCONNECT;
1552 rq_opc = MDS_DISCONNECT;
1555 rq_opc = MGS_DISCONNECT;
1559 CERROR("%s: don't know how to disconnect from %s "
1560 "(connect_op %d): rc = %d\n",
1561 imp->imp_obd->obd_name, obd2cli_tgt(imp->imp_obd),
1562 imp->imp_connect_op, rc);
1566 if (ptlrpc_import_in_recovery(imp)) {
1567 struct l_wait_info lwi;
1568 cfs_duration_t timeout;
1571 if (imp->imp_server_timeout)
1572 timeout = cfs_time_seconds(obd_timeout / 2);
1574 timeout = cfs_time_seconds(obd_timeout);
1576 int idx = import_at_get_index(imp,
1577 imp->imp_client->cli_request_portal);
1578 timeout = cfs_time_seconds(
1579 at_get(&imp->imp_at.iat_service_estimate[idx]));
1582 lwi = LWI_TIMEOUT_INTR(cfs_timeout_cap(timeout),
1583 back_to_sleep, LWI_ON_SIGNAL_NOOP, NULL);
1584 rc = l_wait_event(imp->imp_recovery_waitq,
1585 !ptlrpc_import_in_recovery(imp), &lwi);
1589 spin_lock(&imp->imp_lock);
1590 if (imp->imp_state != LUSTRE_IMP_FULL)
1592 spin_unlock(&imp->imp_lock);
1594 req = ptlrpc_request_alloc_pack(imp, &RQF_MDS_DISCONNECT,
1595 LUSTRE_OBD_VERSION, rq_opc);
1597 /* We are disconnecting, do not retry a failed DISCONNECT rpc if
1598 * it fails. We can get through the above with a down server
1599 * if the client doesn't know the server is gone yet. */
1600 req->rq_no_resend = 1;
1602 /* We want client umounts to happen quickly, no matter the
1604 req->rq_timeout = min_t(int, req->rq_timeout,
1605 INITIAL_CONNECT_TIMEOUT);
1607 IMPORT_SET_STATE(imp, LUSTRE_IMP_CONNECTING);
1608 req->rq_send_state = LUSTRE_IMP_CONNECTING;
1609 ptlrpc_request_set_replen(req);
1610 rc = ptlrpc_queue_wait(req);
1611 ptlrpc_req_finished(req);
1615 spin_lock(&imp->imp_lock);
1618 IMPORT_SET_STATE_NOLOCK(imp, LUSTRE_IMP_DISCON);
1620 IMPORT_SET_STATE_NOLOCK(imp, LUSTRE_IMP_CLOSED);
1621 memset(&imp->imp_remote_handle, 0, sizeof(imp->imp_remote_handle));
1622 spin_unlock(&imp->imp_lock);
1624 if (rc == -ETIMEDOUT || rc == -ENOTCONN || rc == -ESHUTDOWN)
1628 EXPORT_SYMBOL(ptlrpc_disconnect_import);
1630 void ptlrpc_cleanup_imp(struct obd_import *imp)
1634 spin_lock(&imp->imp_lock);
1635 IMPORT_SET_STATE_NOLOCK(imp, LUSTRE_IMP_CLOSED);
1636 imp->imp_generation++;
1637 spin_unlock(&imp->imp_lock);
1638 ptlrpc_abort_inflight(imp);
1643 /* Adaptive Timeout utils */
1644 extern unsigned int at_min, at_max, at_history;
1646 /* Update at_current with the specified value (bounded by at_min and at_max),
1647 * as well as the AT history "bins".
1648 * - Bin into timeslices using AT_BINS bins.
1649 * - This gives us a max of the last at_history seconds without the storage,
1650 * but still smoothing out a return to normalcy from a slow response.
1651 * - (E.g. remember the maximum latency in each minute of the last 4 minutes.)
1653 int at_measured(struct adaptive_timeout *at, unsigned int val)
1655 unsigned int old = at->at_current;
1656 time_t now = cfs_time_current_sec();
1657 time_t binlimit = max_t(time_t, at_history / AT_BINS, 1);
1660 CDEBUG(D_OTHER, "add %u to %p time=%lu v=%u (%u %u %u %u)\n",
1661 val, at, now - at->at_binstart, at->at_current,
1662 at->at_hist[0], at->at_hist[1], at->at_hist[2], at->at_hist[3]);
1665 /* 0's don't count, because we never want our timeout to
1666 drop to 0, and because 0 could mean an error */
1669 spin_lock(&at->at_lock);
1671 if (unlikely(at->at_binstart == 0)) {
1672 /* Special case to remove default from history */
1673 at->at_current = val;
1674 at->at_worst_ever = val;
1675 at->at_worst_time = now;
1676 at->at_hist[0] = val;
1677 at->at_binstart = now;
1678 } else if (now - at->at_binstart < binlimit ) {
1680 at->at_hist[0] = max(val, at->at_hist[0]);
1681 at->at_current = max(val, at->at_current);
1684 unsigned int maxv = val;
1685 /* move bins over */
1686 shift = (now - at->at_binstart) / binlimit;
1688 for(i = AT_BINS - 1; i >= 0; i--) {
1690 at->at_hist[i] = at->at_hist[i - shift];
1691 maxv = max(maxv, at->at_hist[i]);
1696 at->at_hist[0] = val;
1697 at->at_current = maxv;
1698 at->at_binstart += shift * binlimit;
1701 if (at->at_current > at->at_worst_ever) {
1702 at->at_worst_ever = at->at_current;
1703 at->at_worst_time = now;
1706 if (at->at_flags & AT_FLG_NOHIST)
1707 /* Only keep last reported val; keeping the rest of the history
1709 at->at_current = val;
1712 at->at_current = min(at->at_current, at_max);
1713 at->at_current = max(at->at_current, at_min);
1715 if (at->at_current != old)
1716 CDEBUG(D_OTHER, "AT %p change: old=%u new=%u delta=%d "
1717 "(val=%u) hist %u %u %u %u\n", at,
1718 old, at->at_current, at->at_current - old, val,
1719 at->at_hist[0], at->at_hist[1], at->at_hist[2],
1722 /* if we changed, report the old value */
1723 old = (at->at_current != old) ? old : 0;
1725 spin_unlock(&at->at_lock);
1729 /* Find the imp_at index for a given portal; assign if space available */
1730 int import_at_get_index(struct obd_import *imp, int portal)
1732 struct imp_at *at = &imp->imp_at;
1735 for (i = 0; i < IMP_AT_MAX_PORTALS; i++) {
1736 if (at->iat_portal[i] == portal)
1738 if (at->iat_portal[i] == 0)
1743 /* Not found in list, add it under a lock */
1744 spin_lock(&imp->imp_lock);
1746 /* Check unused under lock */
1747 for (; i < IMP_AT_MAX_PORTALS; i++) {
1748 if (at->iat_portal[i] == portal)
1750 if (at->iat_portal[i] == 0)
1755 /* Not enough portals? */
1756 LASSERT(i < IMP_AT_MAX_PORTALS);
1758 at->iat_portal[i] = portal;
1760 spin_unlock(&imp->imp_lock);