1 /* -*- mode: c; c-basic-offset: 8; indent-tabs-mode: nil; -*-
2 * vim:expandtab:shiftwidth=8:tabstop=8:
6 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
8 * This program is free software; you can redistribute it and/or modify
9 * it under the terms of the GNU General Public License version 2 only,
10 * as published by the Free Software Foundation.
12 * This program is distributed in the hope that it will be useful, but
13 * WITHOUT ANY WARRANTY; without even the implied warranty of
14 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
15 * General Public License version 2 for more details (a copy is included
16 * in the LICENSE file that accompanied this code).
18 * You should have received a copy of the GNU General Public License
19 * version 2 along with this program; If not, see
20 * http://www.sun.com/software/products/lustre/docs/GPLv2.pdf
22 * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
23 * CA 95054 USA or visit www.sun.com if you need additional information or
29 * Copyright 2008 Sun Microsystems, Inc. All rights reserved
30 * Use is subject to license terms.
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 <liblustre.h>
46 #include <obd_support.h>
47 #include <lustre_ha.h>
48 #include <lustre_net.h>
49 #include <lustre_import.h>
50 #include <lustre_export.h>
52 #include <obd_class.h>
54 #include "ptlrpc_internal.h"
56 struct ptlrpc_connect_async_args {
57 __u64 pcaa_peer_committed;
58 int pcaa_initial_connect;
61 /* A CLOSED import should remain so. */
62 #define IMPORT_SET_STATE_NOLOCK(imp, state) \
64 if (imp->imp_state != LUSTRE_IMP_CLOSED) { \
65 CDEBUG(D_HA, "%p %s: changing import state from %s to %s\n", \
66 imp, obd2cli_tgt(imp->imp_obd), \
67 ptlrpc_import_state_name(imp->imp_state), \
68 ptlrpc_import_state_name(state)); \
69 imp->imp_state = state; \
73 #define IMPORT_SET_STATE(imp, state) \
75 spin_lock(&imp->imp_lock); \
76 IMPORT_SET_STATE_NOLOCK(imp, state); \
77 spin_unlock(&imp->imp_lock); \
81 static int ptlrpc_connect_interpret(struct ptlrpc_request *request,
83 int ptlrpc_import_recovery_state_machine(struct obd_import *imp);
85 /* Only this function is allowed to change the import state when it is
86 * CLOSED. I would rather refcount the import and free it after
87 * disconnection like we do with exports. To do that, the client_obd
88 * will need to save the peer info somewhere other than in the import,
90 int ptlrpc_init_import(struct obd_import *imp)
92 spin_lock(&imp->imp_lock);
94 imp->imp_generation++;
95 imp->imp_state = LUSTRE_IMP_NEW;
97 spin_unlock(&imp->imp_lock);
101 EXPORT_SYMBOL(ptlrpc_init_import);
103 #define UUID_STR "_UUID"
104 static void deuuidify(char *uuid, const char *prefix, char **uuid_start,
107 *uuid_start = !prefix || strncmp(uuid, prefix, strlen(prefix))
108 ? uuid : uuid + strlen(prefix);
110 *uuid_len = strlen(*uuid_start);
112 if (*uuid_len < strlen(UUID_STR))
115 if (!strncmp(*uuid_start + *uuid_len - strlen(UUID_STR),
116 UUID_STR, strlen(UUID_STR)))
117 *uuid_len -= strlen(UUID_STR);
120 /* Returns true if import was FULL, false if import was already not
122 * @imp - import to be disconnected
123 * @conn_cnt - connection count (epoch) of the request that timed out
124 * and caused the disconnection. In some cases, multiple
125 * inflight requests can fail to a single target (e.g. OST
126 * bulk requests) and if one has already caused a reconnection
127 * (increasing the import->conn_cnt) the older failure should
128 * not also cause a reconnection. If zero it forces a reconnect.
130 int ptlrpc_set_import_discon(struct obd_import *imp, __u32 conn_cnt)
134 spin_lock(&imp->imp_lock);
136 if (imp->imp_state == LUSTRE_IMP_FULL &&
137 (conn_cnt == 0 || conn_cnt == imp->imp_conn_cnt)) {
141 deuuidify(obd2cli_tgt(imp->imp_obd), NULL,
142 &target_start, &target_len);
143 if (imp->imp_replayable) {
144 LCONSOLE_WARN("%s: Connection to service %.*s via nid "
145 "%s was lost; in progress operations using this "
146 "service will wait for recovery to complete.\n",
147 imp->imp_obd->obd_name, target_len, target_start,
148 libcfs_nid2str(imp->imp_connection->c_peer.nid));
150 LCONSOLE_ERROR_MSG(0x166, "%s: Connection to service "
151 "%.*s via nid %s was lost; in progress "
152 "operations using this service will fail.\n",
153 imp->imp_obd->obd_name, target_len, target_start,
154 libcfs_nid2str(imp->imp_connection->c_peer.nid));
156 IMPORT_SET_STATE_NOLOCK(imp, LUSTRE_IMP_DISCON);
157 spin_unlock(&imp->imp_lock);
159 if (obd_dump_on_timeout)
160 libcfs_debug_dumplog();
162 obd_import_event(imp->imp_obd, imp, IMP_EVENT_DISCON);
165 spin_unlock(&imp->imp_lock);
166 CDEBUG(D_HA, "%s: import %p already %s (conn %u, was %u): %s\n",
167 imp->imp_client->cli_name, imp,
168 (imp->imp_state == LUSTRE_IMP_FULL &&
169 imp->imp_conn_cnt > conn_cnt) ?
170 "reconnected" : "not connected", imp->imp_conn_cnt,
171 conn_cnt, ptlrpc_import_state_name(imp->imp_state));
177 /* Must be called with imp_lock held! */
178 static void ptlrpc_deactivate_and_unlock_import(struct obd_import *imp)
181 LASSERT_SPIN_LOCKED(&imp->imp_lock);
183 CDEBUG(D_HA, "setting import %s INVALID\n", obd2cli_tgt(imp->imp_obd));
184 imp->imp_invalid = 1;
185 imp->imp_generation++;
186 spin_unlock(&imp->imp_lock);
188 ptlrpc_abort_inflight(imp);
189 obd_import_event(imp->imp_obd, imp, IMP_EVENT_INACTIVE);
193 * This acts as a barrier; all existing requests are rejected, and
194 * no new requests will be accepted until the import is valid again.
196 void ptlrpc_deactivate_import(struct obd_import *imp)
198 spin_lock(&imp->imp_lock);
199 ptlrpc_deactivate_and_unlock_import(imp);
203 ptlrpc_inflight_deadline(struct ptlrpc_request *req, time_t now)
207 if (!(((req->rq_phase == RQ_PHASE_RPC) && !req->rq_waiting) ||
208 (req->rq_phase == RQ_PHASE_BULK) ||
209 (req->rq_phase == RQ_PHASE_NEW)))
212 if (req->rq_timedout)
215 if (req->rq_phase == RQ_PHASE_NEW)
218 dl = req->rq_deadline;
226 static unsigned int ptlrpc_inflight_timeout(struct obd_import *imp)
228 time_t now = cfs_time_current_sec();
229 struct list_head *tmp, *n;
230 struct ptlrpc_request *req;
231 unsigned int timeout = 0;
233 spin_lock(&imp->imp_lock);
234 list_for_each_safe(tmp, n, &imp->imp_sending_list) {
235 req = list_entry(tmp, struct ptlrpc_request, rq_list);
236 timeout = max(ptlrpc_inflight_deadline(req, now), timeout);
238 spin_unlock(&imp->imp_lock);
243 * This function will invalidate the import, if necessary, then block
244 * for all the RPC completions, and finally notify the obd to
245 * invalidate its state (ie cancel locks, clear pending requests,
248 void ptlrpc_invalidate_import(struct obd_import *imp)
250 struct list_head *tmp, *n;
251 struct ptlrpc_request *req;
252 struct l_wait_info lwi;
253 unsigned int timeout;
256 atomic_inc(&imp->imp_inval_count);
259 * If this is an invalid MGC connection, then don't bother
260 * waiting for imp_inflight to drop to 0.
262 if (imp->imp_invalid && imp->imp_recon_bk && !imp->imp_obd->obd_no_recov)
265 if (!imp->imp_invalid || imp->imp_obd->obd_no_recov)
266 ptlrpc_deactivate_import(imp);
268 LASSERT(imp->imp_invalid);
270 /* Wait forever until inflight == 0. We really can't do it another
271 * way because in some cases we need to wait for very long reply
272 * unlink. We can't do anything before that because there is really
273 * no guarantee that some rdma transfer is not in progress right now. */
275 /* Calculate max timeout for waiting on rpcs to error
276 * out. Use obd_timeout if calculated value is smaller
278 timeout = ptlrpc_inflight_timeout(imp);
279 timeout += timeout / 3;
282 timeout = obd_timeout;
284 CDEBUG(D_RPCTRACE, "Sleeping %d sec for inflight to error out\n",
287 /* Wait for all requests to error out and call completion
288 * callbacks. Cap it at obd_timeout -- these should all
289 * have been locally cancelled by ptlrpc_abort_inflight. */
290 lwi = LWI_TIMEOUT_INTERVAL(
291 cfs_timeout_cap(cfs_time_seconds(timeout)),
292 cfs_time_seconds(1), NULL, NULL);
293 rc = l_wait_event(imp->imp_recovery_waitq,
294 (atomic_read(&imp->imp_inflight) == 0), &lwi);
296 const char *cli_tgt = obd2cli_tgt(imp->imp_obd);
298 CERROR("%s: rc = %d waiting for callback (%d != 0)\n",
299 cli_tgt, rc, atomic_read(&imp->imp_inflight));
301 spin_lock(&imp->imp_lock);
302 list_for_each_safe(tmp, n, &imp->imp_sending_list) {
303 req = list_entry(tmp, struct ptlrpc_request,
305 DEBUG_REQ(D_ERROR, req, "still on sending list");
307 list_for_each_safe(tmp, n, &imp->imp_delayed_list) {
308 req = list_entry(tmp, struct ptlrpc_request,
310 DEBUG_REQ(D_ERROR, req, "still on delayed list");
313 if (atomic_read(&imp->imp_unregistering) == 0) {
314 /* We know that only "unregistering" rpcs may
315 * still survive in sending or delaying lists
316 * (They are waiting for long reply unlink in
317 * sluggish nets). Let's check this. If there
318 * is no unregistering and inflight != 0 this
320 LASSERT(atomic_read(&imp->imp_inflight) == 0);
322 /* Let's save one loop as soon as inflight have
323 * dropped to zero. No new inflights possible at
327 CERROR("%s: RPCs in \"%s\" phase found (%d). "
328 "Network is sluggish? Waiting them "
329 "to error out.\n", cli_tgt,
330 ptlrpc_phase2str(RQ_PHASE_UNREGISTERING),
331 atomic_read(&imp->imp_unregistering));
333 spin_unlock(&imp->imp_lock);
337 /* Let's additionally check that no new rpcs added to import in
338 * "invalidate" state. */
339 LASSERT(atomic_read(&imp->imp_inflight) == 0);
342 obd_import_event(imp->imp_obd, imp, IMP_EVENT_INVALIDATE);
344 atomic_dec(&imp->imp_inval_count);
345 cfs_waitq_signal(&imp->imp_recovery_waitq);
348 /* unset imp_invalid */
349 void ptlrpc_activate_import(struct obd_import *imp)
351 struct obd_device *obd = imp->imp_obd;
353 spin_lock(&imp->imp_lock);
354 imp->imp_invalid = 0;
355 spin_unlock(&imp->imp_lock);
357 obd_import_event(obd, imp, IMP_EVENT_ACTIVE);
360 void ptlrpc_fail_import(struct obd_import *imp, __u32 conn_cnt)
364 LASSERT(!imp->imp_dlm_fake);
366 if (ptlrpc_set_import_discon(imp, conn_cnt)) {
367 if (!imp->imp_replayable) {
368 CDEBUG(D_HA, "import %s@%s for %s not replayable, "
369 "auto-deactivating\n",
370 obd2cli_tgt(imp->imp_obd),
371 imp->imp_connection->c_remote_uuid.uuid,
372 imp->imp_obd->obd_name);
373 ptlrpc_deactivate_import(imp);
376 CDEBUG(D_HA, "%s: waking up pinger\n",
377 obd2cli_tgt(imp->imp_obd));
379 spin_lock(&imp->imp_lock);
380 imp->imp_force_verify = 1;
381 spin_unlock(&imp->imp_lock);
383 ptlrpc_pinger_wake_up();
388 int ptlrpc_reconnect_import(struct obd_import *imp)
391 ptlrpc_set_import_discon(imp, 0);
392 /* Force a new connect attempt */
393 ptlrpc_invalidate_import(imp);
394 /* Do a fresh connect next time by zeroing the handle */
395 ptlrpc_disconnect_import(imp, 1);
396 /* Wait for all invalidate calls to finish */
397 if (atomic_read(&imp->imp_inval_count) > 0) {
399 struct l_wait_info lwi = LWI_INTR(LWI_ON_SIGNAL_NOOP, NULL);
400 rc = l_wait_event(imp->imp_recovery_waitq,
401 (atomic_read(&imp->imp_inval_count) == 0),
404 CERROR("Interrupted, inval=%d\n",
405 atomic_read(&imp->imp_inval_count));
409 * Allow reconnect attempts. Note: Currently, the function is
410 * only called by MGC. So assume this is a recoverable import,
411 * and force import to be recoverable. fix this if you need to
414 imp->imp_obd->obd_no_recov = 0;
415 /* Remove 'invalid' flag */
416 ptlrpc_activate_import(imp);
417 /* Attempt a new connect */
418 ptlrpc_recover_import(imp, NULL);
422 EXPORT_SYMBOL(ptlrpc_reconnect_import);
424 static int import_select_connection(struct obd_import *imp)
426 struct obd_import_conn *imp_conn = NULL, *conn;
427 struct obd_export *dlmexp;
431 spin_lock(&imp->imp_lock);
433 if (list_empty(&imp->imp_conn_list)) {
434 CERROR("%s: no connections available\n",
435 imp->imp_obd->obd_name);
436 spin_unlock(&imp->imp_lock);
440 list_for_each_entry(conn, &imp->imp_conn_list, oic_item) {
441 CDEBUG(D_HA, "%s: connect to NID %s last attempt "LPU64"\n",
442 imp->imp_obd->obd_name,
443 libcfs_nid2str(conn->oic_conn->c_peer.nid),
444 conn->oic_last_attempt);
446 /* Don't thrash connections */
447 if (cfs_time_before_64(cfs_time_current_64(),
448 conn->oic_last_attempt +
449 cfs_time_seconds(CONNECTION_SWITCH_MIN))) {
453 /* If we have not tried this connection since the
454 the last successful attempt, go with this one */
455 if ((conn->oic_last_attempt == 0) ||
456 cfs_time_beforeq_64(conn->oic_last_attempt,
457 imp->imp_last_success_conn)) {
463 /* If all of the connections have already been tried
464 since the last successful connection; just choose the
465 least recently used */
468 else if (cfs_time_before_64(conn->oic_last_attempt,
469 imp_conn->oic_last_attempt))
473 /* if not found, simply choose the current one */
475 LASSERT(imp->imp_conn_current);
476 imp_conn = imp->imp_conn_current;
479 LASSERT(imp_conn->oic_conn);
481 /* If we've tried everything, and we're back to the beginning of the
482 list, increase our timeout and try again. It will be reset when
483 we do finally connect. (FIXME: really we should wait for all network
484 state associated with the last connection attempt to drain before
485 trying to reconnect on it.) */
486 if (tried_all && (imp->imp_conn_list.next == &imp_conn->oic_item) &&
487 !imp->imp_recon_bk /* not retrying */) {
488 if (at_get(&imp->imp_at.iat_net_latency) <
489 CONNECTION_SWITCH_MAX) {
490 at_add(&imp->imp_at.iat_net_latency,
491 at_get(&imp->imp_at.iat_net_latency) +
492 CONNECTION_SWITCH_INC);
494 LASSERT(imp_conn->oic_last_attempt);
495 CWARN("%s: tried all connections, increasing latency to %ds\n",
496 imp->imp_obd->obd_name,
497 at_get(&imp->imp_at.iat_net_latency));
500 imp_conn->oic_last_attempt = cfs_time_current_64();
502 /* switch connection, don't mind if it's same as the current one */
503 if (imp->imp_connection)
504 ptlrpc_connection_put(imp->imp_connection);
505 imp->imp_connection = ptlrpc_connection_addref(imp_conn->oic_conn);
507 dlmexp = class_conn2export(&imp->imp_dlm_handle);
508 LASSERT(dlmexp != NULL);
509 if (dlmexp->exp_connection)
510 ptlrpc_connection_put(dlmexp->exp_connection);
511 dlmexp->exp_connection = ptlrpc_connection_addref(imp_conn->oic_conn);
512 class_export_put(dlmexp);
514 if (imp->imp_conn_current != imp_conn) {
515 if (imp->imp_conn_current)
516 LCONSOLE_INFO("Changing connection for %s to %s/%s\n",
517 imp->imp_obd->obd_name,
518 imp_conn->oic_uuid.uuid,
519 libcfs_nid2str(imp_conn->oic_conn->c_peer.nid));
520 imp->imp_conn_current = imp_conn;
523 CDEBUG(D_HA, "%s: import %p using connection %s/%s\n",
524 imp->imp_obd->obd_name, imp, imp_conn->oic_uuid.uuid,
525 libcfs_nid2str(imp_conn->oic_conn->c_peer.nid));
527 spin_unlock(&imp->imp_lock);
532 int ptlrpc_connect_import(struct obd_import *imp, char *new_uuid)
534 struct obd_device *obd = imp->imp_obd;
535 int initial_connect = 0;
537 __u64 committed_before_reconnect = 0;
538 struct ptlrpc_request *request;
539 __u32 size[] = { sizeof(struct ptlrpc_body),
540 sizeof(imp->imp_obd->u.cli.cl_target_uuid),
541 sizeof(obd->obd_uuid),
542 sizeof(imp->imp_dlm_handle),
543 sizeof(imp->imp_connect_data) };
544 char *tmp[] = { NULL,
545 obd2cli_tgt(imp->imp_obd),
547 (char *)&imp->imp_dlm_handle,
548 (char *)&imp->imp_connect_data };
549 struct ptlrpc_connect_async_args *aa;
552 spin_lock(&imp->imp_lock);
553 if (imp->imp_state == LUSTRE_IMP_CLOSED) {
554 spin_unlock(&imp->imp_lock);
555 CERROR("can't connect to a closed import\n");
557 } else if (imp->imp_state == LUSTRE_IMP_FULL) {
558 spin_unlock(&imp->imp_lock);
559 CERROR("already connected\n");
561 } else if (imp->imp_state == LUSTRE_IMP_CONNECTING) {
562 spin_unlock(&imp->imp_lock);
563 CERROR("already connecting\n");
567 IMPORT_SET_STATE_NOLOCK(imp, LUSTRE_IMP_CONNECTING);
570 imp->imp_resend_replay = 0;
572 if (!lustre_handle_is_used(&imp->imp_remote_handle))
575 committed_before_reconnect = imp->imp_peer_committed_transno;
577 spin_unlock(&imp->imp_lock);
580 struct obd_uuid uuid;
582 obd_str2uuid(&uuid, new_uuid);
583 rc = import_set_conn_priority(imp, &uuid);
588 rc = import_select_connection(imp);
592 /* last in connection list */
593 if (imp->imp_conn_current->oic_item.next == &imp->imp_conn_list) {
594 if (imp->imp_initial_recov_bk && initial_connect) {
595 CDEBUG(D_HA, "Last connection attempt (%d) for %s\n",
596 imp->imp_conn_cnt, obd2cli_tgt(imp->imp_obd));
597 /* Don't retry if connect fails */
599 obd_set_info_async(obd->obd_self_export,
600 sizeof(KEY_INIT_RECOV),
602 sizeof(rc), &rc, NULL);
604 if (imp->imp_recon_bk) {
605 CDEBUG(D_HA, "Last reconnection attempt (%d) for %s\n",
606 imp->imp_conn_cnt, obd2cli_tgt(imp->imp_obd));
607 spin_lock(&imp->imp_lock);
608 imp->imp_last_recon = 1;
609 spin_unlock(&imp->imp_lock);
613 /* Reset connect flags to the originally requested flags, in case
614 * the server is updated on-the-fly we will get the new features. */
615 imp->imp_connect_data.ocd_connect_flags = imp->imp_connect_flags_orig;
616 imp->imp_msghdr_flags &= ~MSGHDR_AT_SUPPORT;
618 rc = obd_reconnect(imp->imp_obd->obd_self_export, obd,
619 &obd->obd_uuid, &imp->imp_connect_data, NULL);
623 request = ptlrpc_prep_req(imp, LUSTRE_OBD_VERSION, imp->imp_connect_op,
626 GOTO(out, rc = -ENOMEM);
628 /* Report the rpc service time to the server so that it knows how long
629 * to wait for clients to join recovery */
630 lustre_msg_set_service_time(request->rq_reqmsg,
631 at_timeout2est(request->rq_timeout));
633 /* The amount of time we give the server to process the connect req.
634 * import_select_connection will increase the net latency on
635 * repeated reconnect attempts to cover slow networks.
636 * We override/ignore the server rpc completion estimate here,
637 * which may be large if this is a reconnect attempt */
638 request->rq_timeout = INITIAL_CONNECT_TIMEOUT;
639 lustre_msg_set_timeout(request->rq_reqmsg, request->rq_timeout);
642 lustre_msg_add_op_flags(request->rq_reqmsg, MSG_CONNECT_LIBCLIENT);
644 if (imp->imp_msg_magic == LUSTRE_MSG_MAGIC_V1)
645 lustre_msg_add_op_flags(request->rq_reqmsg,
646 MSG_CONNECT_NEXT_VER);
648 request->rq_no_resend = request->rq_no_delay = 1;
649 request->rq_send_state = LUSTRE_IMP_CONNECTING;
650 /* Allow a slightly larger reply for future growth compatibility */
651 size[REPLY_REC_OFF] = sizeof(struct obd_connect_data) +
653 ptlrpc_req_set_repsize(request, 2, size);
654 request->rq_interpret_reply = ptlrpc_connect_interpret;
656 CLASSERT(sizeof(*aa) <= sizeof(request->rq_async_args));
657 aa = ptlrpc_req_async_args(request);
658 memset(aa, 0, sizeof *aa);
660 aa->pcaa_peer_committed = committed_before_reconnect;
661 aa->pcaa_initial_connect = initial_connect;
662 if (aa->pcaa_initial_connect) {
663 spin_lock(&imp->imp_lock);
664 imp->imp_replayable = 1;
665 spin_unlock(&imp->imp_lock);
668 DEBUG_REQ(D_RPCTRACE, request, "%sconnect request %d",
669 aa->pcaa_initial_connect ? "initial " : "re",
671 ptlrpcd_add_req(request);
675 IMPORT_SET_STATE(imp, LUSTRE_IMP_DISCON);
680 EXPORT_SYMBOL(ptlrpc_connect_import);
682 static void ptlrpc_maybe_ping_import_soon(struct obd_import *imp)
685 struct obd_import_conn *imp_conn;
691 spin_lock(&imp->imp_lock);
692 if (list_empty(&imp->imp_conn_list))
696 imp_conn = list_entry(imp->imp_conn_list.prev,
697 struct obd_import_conn,
700 /* XXX: When the failover node is the primary node, it is possible
701 * to have two identical connections in imp_conn_list. We must
702 * compare not conn's pointers but NIDs, otherwise we can defeat
703 * connection throttling. (See bug 14774.) */
704 if (imp->imp_conn_current->oic_conn->c_peer.nid !=
705 imp_conn->oic_conn->c_peer.nid) {
706 ptlrpc_ping_import_soon(imp);
711 /* liblustre has no pinger thead, so we wakup pinger anyway */
715 spin_unlock(&imp->imp_lock);
718 ptlrpc_pinger_wake_up();
723 static int ptlrpc_connect_interpret(struct ptlrpc_request *request,
726 struct ptlrpc_connect_async_args *aa = data;
727 struct obd_import *imp = request->rq_import;
728 struct client_obd *cli = &imp->imp_obd->u.cli;
729 struct lustre_handle old_hdl;
730 __u64 old_connect_flags;
734 spin_lock(&imp->imp_lock);
735 if (imp->imp_state == LUSTRE_IMP_CLOSED) {
736 spin_unlock(&imp->imp_lock);
739 spin_unlock(&imp->imp_lock);
744 LASSERT(imp->imp_conn_current);
746 msg_flags = lustre_msg_get_op_flags(request->rq_repmsg);
748 /* All imports are pingable */
749 spin_lock(&imp->imp_lock);
750 imp->imp_pingable = 1;
752 if (aa->pcaa_initial_connect) {
753 if (msg_flags & MSG_CONNECT_REPLAYABLE) {
754 imp->imp_replayable = 1;
755 spin_unlock(&imp->imp_lock);
756 CDEBUG(D_HA, "connected to replayable target: %s\n",
757 obd2cli_tgt(imp->imp_obd));
759 imp->imp_replayable = 0;
760 spin_unlock(&imp->imp_lock);
763 if ((request->rq_reqmsg->lm_magic == LUSTRE_MSG_MAGIC_V1 &&
764 msg_flags & MSG_CONNECT_NEXT_VER) ||
765 request->rq_reqmsg->lm_magic == LUSTRE_MSG_MAGIC_V2) {
766 imp->imp_msg_magic = LUSTRE_MSG_MAGIC_V2;
767 CDEBUG(D_RPCTRACE, "connect to %s with lustre_msg_v2\n",
768 obd2cli_tgt(imp->imp_obd));
770 CDEBUG(D_RPCTRACE, "connect to %s with lustre_msg_v1\n",
771 obd2cli_tgt(imp->imp_obd));
774 imp->imp_remote_handle =
775 *lustre_msg_get_handle(request->rq_repmsg);
777 /* Initial connects are allowed for clients with non-random
778 * uuids when servers are in recovery. Simply signal the
779 * servers replay is complete and wait in REPLAY_WAIT. */
780 if (msg_flags & MSG_CONNECT_RECOVERING) {
781 CDEBUG(D_HA, "connect to %s during recovery\n",
782 obd2cli_tgt(imp->imp_obd));
783 IMPORT_SET_STATE(imp, LUSTRE_IMP_REPLAY_LOCKS);
785 IMPORT_SET_STATE(imp, LUSTRE_IMP_FULL);
788 spin_lock(&imp->imp_lock);
789 if (imp->imp_invalid) {
790 spin_unlock(&imp->imp_lock);
791 ptlrpc_activate_import(imp);
793 spin_unlock(&imp->imp_lock);
795 GOTO(finish, rc = 0);
797 spin_unlock(&imp->imp_lock);
800 /* Determine what recovery state to move the import to. */
801 if (MSG_CONNECT_RECONNECT & msg_flags) {
802 memset(&old_hdl, 0, sizeof(old_hdl));
803 if (!memcmp(&old_hdl, lustre_msg_get_handle(request->rq_repmsg),
805 CERROR("%s@%s didn't like our handle "LPX64
806 ", failed\n", obd2cli_tgt(imp->imp_obd),
807 imp->imp_connection->c_remote_uuid.uuid,
808 imp->imp_dlm_handle.cookie);
809 GOTO(out, rc = -ENOTCONN);
812 if (memcmp(&imp->imp_remote_handle,
813 lustre_msg_get_handle(request->rq_repmsg),
814 sizeof(imp->imp_remote_handle))) {
815 int level = msg_flags & MSG_CONNECT_RECOVERING ? D_HA :
818 /* Bug 16611/14775: if server handle have changed,
819 * that means some sort of disconnection happened.
820 * If the server is not in recovery, that also means it
821 * already erased all of our state because of previous
822 * eviction. If it is in recovery - we are safe to
823 * participate since we can reestablish all of our state
824 * with server again */
825 CDEBUG(level,"%s@%s changed server handle from "
826 LPX64" to "LPX64"%s \n" "but is still in recovery \n",
827 obd2cli_tgt(imp->imp_obd),
828 imp->imp_connection->c_remote_uuid.uuid,
829 imp->imp_remote_handle.cookie,
830 lustre_msg_get_handle(request->rq_repmsg)->
832 (MSG_CONNECT_RECOVERING & msg_flags) ?
833 "but is still in recovery" : "");
835 imp->imp_remote_handle =
836 *lustre_msg_get_handle(request->rq_repmsg);
838 if (!(MSG_CONNECT_RECOVERING & msg_flags)) {
839 IMPORT_SET_STATE(imp, LUSTRE_IMP_EVICTED);
840 GOTO(finish, rc = 0);
844 CDEBUG(D_HA, "reconnected to %s@%s after partition\n",
845 obd2cli_tgt(imp->imp_obd),
846 imp->imp_connection->c_remote_uuid.uuid);
849 if (imp->imp_invalid) {
850 CDEBUG(D_HA, "%s: reconnected but import is invalid; "
851 "marking evicted\n", imp->imp_obd->obd_name);
852 IMPORT_SET_STATE(imp, LUSTRE_IMP_EVICTED);
853 } else if (MSG_CONNECT_RECOVERING & msg_flags) {
854 CDEBUG(D_HA, "%s: reconnected to %s during replay\n",
855 imp->imp_obd->obd_name,
856 obd2cli_tgt(imp->imp_obd));
858 spin_lock(&imp->imp_lock);
859 imp->imp_resend_replay = 1;
860 /* VBR: delayed connection */
861 if (MSG_CONNECT_DELAYED & msg_flags) {
862 imp->imp_delayed_recovery = 1;
863 imp->imp_no_lock_replay = 1;
865 spin_unlock(&imp->imp_lock);
867 IMPORT_SET_STATE(imp, LUSTRE_IMP_REPLAY);
869 IMPORT_SET_STATE(imp, LUSTRE_IMP_RECOVER);
871 } else if ((MSG_CONNECT_RECOVERING & msg_flags) && !imp->imp_invalid) {
872 LASSERT(imp->imp_replayable);
873 imp->imp_remote_handle =
874 *lustre_msg_get_handle(request->rq_repmsg);
875 imp->imp_last_replay_transno = 0;
876 /* VBR: delayed connection */
877 if (MSG_CONNECT_DELAYED & msg_flags) {
878 spin_lock(&imp->imp_lock);
879 imp->imp_delayed_recovery = 1;
880 imp->imp_no_lock_replay = 1;
881 spin_unlock(&imp->imp_lock);
883 IMPORT_SET_STATE(imp, LUSTRE_IMP_REPLAY);
885 DEBUG_REQ(D_HA, request, "evicting (not initial connect and "
886 "flags reconnect/recovering not set: %x)",msg_flags);
887 imp->imp_remote_handle =
888 *lustre_msg_get_handle(request->rq_repmsg);
889 IMPORT_SET_STATE(imp, LUSTRE_IMP_EVICTED);
892 /* Sanity checks for a reconnected import. */
893 if (!(imp->imp_replayable) != !(msg_flags & MSG_CONNECT_REPLAYABLE)) {
894 CERROR("imp_replayable flag does not match server "
895 "after reconnect. We should LBUG right here.\n");
898 if (lustre_msg_get_last_committed(request->rq_repmsg) <
899 aa->pcaa_peer_committed) {
900 CERROR("%s went back in time (transno "LPD64
901 " was previously committed, server now claims "LPD64
902 ")! See https://bugzilla.lustre.org/show_bug.cgi?"
904 obd2cli_tgt(imp->imp_obd), aa->pcaa_peer_committed,
905 lustre_msg_get_last_committed(request->rq_repmsg));
909 rc = ptlrpc_import_recovery_state_machine(imp);
911 if (rc == -ENOTCONN) {
912 CDEBUG(D_HA, "evicted/aborted by %s@%s during recovery;"
913 "invalidating and reconnecting\n",
914 obd2cli_tgt(imp->imp_obd),
915 imp->imp_connection->c_remote_uuid.uuid);
916 ptlrpc_connect_import(imp, NULL);
920 struct obd_connect_data *ocd;
921 struct obd_export *exp;
923 ocd = lustre_swab_repbuf(request, REPLY_REC_OFF, sizeof(*ocd),
924 lustre_swab_connect);
925 spin_lock(&imp->imp_lock);
926 list_del(&imp->imp_conn_current->oic_item);
927 list_add(&imp->imp_conn_current->oic_item, &imp->imp_conn_list);
928 imp->imp_last_success_conn =
929 imp->imp_conn_current->oic_last_attempt;
932 spin_unlock(&imp->imp_lock);
933 CERROR("Wrong connect data from server\n");
938 imp->imp_connect_data = *ocd;
940 exp = class_conn2export(&imp->imp_dlm_handle);
941 spin_unlock(&imp->imp_lock);
943 /* check that server granted subset of flags we asked for. */
944 LASSERTF((ocd->ocd_connect_flags &
945 imp->imp_connect_flags_orig) ==
946 ocd->ocd_connect_flags, LPX64" != "LPX64,
947 imp->imp_connect_flags_orig, ocd->ocd_connect_flags);
950 /* This could happen if export is cleaned during the
952 CERROR("Missing export for %s\n",
953 imp->imp_obd->obd_name);
954 GOTO(out, rc = -ENODEV);
956 old_connect_flags = exp->exp_connect_flags;
957 exp->exp_connect_flags = ocd->ocd_connect_flags;
958 imp->imp_obd->obd_self_export->exp_connect_flags = ocd->ocd_connect_flags;
959 class_export_put(exp);
961 obd_import_event(imp->imp_obd, imp, IMP_EVENT_OCD);
963 if (!ocd->ocd_ibits_known &&
964 ocd->ocd_connect_flags & OBD_CONNECT_IBITS)
965 CERROR("Inodebits aware server returned zero compatible"
968 if ((ocd->ocd_connect_flags & OBD_CONNECT_VERSION) &&
969 (ocd->ocd_version > LUSTRE_VERSION_CODE +
970 LUSTRE_VERSION_OFFSET_WARN ||
971 ocd->ocd_version < LUSTRE_VERSION_CODE -
972 LUSTRE_VERSION_OFFSET_WARN)) {
973 /* Sigh, some compilers do not like #ifdef in the middle
974 of macro arguments */
977 "older. Consider upgrading this client";
980 "older. Consider recompiling this application";
982 const char *newer = "newer than client version";
984 LCONSOLE_WARN("Server %s version (%d.%d.%d.%d) "
986 obd2cli_tgt(imp->imp_obd),
987 OBD_OCD_VERSION_MAJOR(ocd->ocd_version),
988 OBD_OCD_VERSION_MINOR(ocd->ocd_version),
989 OBD_OCD_VERSION_PATCH(ocd->ocd_version),
990 OBD_OCD_VERSION_FIX(ocd->ocd_version),
991 ocd->ocd_version > LUSTRE_VERSION_CODE ?
992 newer : older, LUSTRE_VERSION_STRING);
995 if (ocd->ocd_connect_flags & OBD_CONNECT_CKSUM) {
996 /* We sent to the server ocd_cksum_types with bits set
997 * for algorithms we understand. The server masked off
998 * the checksum types it doesn't support */
999 if ((ocd->ocd_cksum_types & OBD_CKSUM_ALL) == 0) {
1000 LCONSOLE_WARN("The negotiation of the checksum "
1001 "alogrithm to use with server %s "
1002 "failed (%x/%x), disabling "
1004 obd2cli_tgt(imp->imp_obd),
1005 ocd->ocd_cksum_types,
1007 cli->cl_checksum = 0;
1008 cli->cl_supp_cksum_types = OBD_CKSUM_CRC32;
1009 cli->cl_cksum_type = OBD_CKSUM_CRC32;
1011 cli->cl_supp_cksum_types = ocd->ocd_cksum_types;
1013 if (ocd->ocd_cksum_types & OSC_DEFAULT_CKSUM)
1014 cli->cl_cksum_type = OSC_DEFAULT_CKSUM;
1015 else if (ocd->ocd_cksum_types & OBD_CKSUM_ADLER)
1016 cli->cl_cksum_type = OBD_CKSUM_ADLER;
1018 cli->cl_cksum_type = OBD_CKSUM_CRC32;
1021 /* The server does not support OBD_CONNECT_CKSUM.
1022 * Enforce CRC32 for backward compatibility*/
1023 cli->cl_supp_cksum_types = OBD_CKSUM_CRC32;
1024 cli->cl_cksum_type = OBD_CKSUM_CRC32;
1027 if (ocd->ocd_connect_flags & OBD_CONNECT_BRW_SIZE) {
1028 cli->cl_max_pages_per_rpc =
1029 ocd->ocd_brw_size >> CFS_PAGE_SHIFT;
1032 /* Reset ns_connect_flags only for initial connect. It might be
1033 * changed in while using FS and if we reset it in reconnect
1034 * this leads to lossing user settings done before such as
1035 * disable lru_resize, etc. */
1036 if (old_connect_flags != exp->exp_connect_flags ||
1037 aa->pcaa_initial_connect) {
1038 CWARN("Reseting ns_connect_flags to server flags: "LPU64"\n",
1039 ocd->ocd_connect_flags);
1040 imp->imp_obd->obd_namespace->ns_connect_flags =
1041 ocd->ocd_connect_flags;
1042 imp->imp_obd->obd_namespace->ns_orig_connect_flags =
1043 ocd->ocd_connect_flags;
1046 if ((ocd->ocd_connect_flags & OBD_CONNECT_AT) &&
1047 (imp->imp_msg_magic == LUSTRE_MSG_MAGIC_V2))
1048 /* We need a per-message support flag, because
1049 a. we don't know if the incoming connect reply
1050 supports AT or not (in reply_in_callback)
1052 b. failovered server means export and flags are gone
1053 (in ptlrpc_send_reply).
1054 Can only be set when we know AT is supported at
1056 imp->imp_msghdr_flags |= MSGHDR_AT_SUPPORT;
1058 imp->imp_msghdr_flags &= ~MSGHDR_AT_SUPPORT;
1060 LASSERT((cli->cl_max_pages_per_rpc <= PTLRPC_MAX_BRW_PAGES) &&
1061 (cli->cl_max_pages_per_rpc > 0));
1066 IMPORT_SET_STATE(imp, LUSTRE_IMP_DISCON);
1067 spin_lock(&imp->imp_lock);
1068 if (aa->pcaa_initial_connect && !imp->imp_initial_recov &&
1069 (request->rq_import_generation == imp->imp_generation))
1070 ptlrpc_deactivate_and_unlock_import(imp);
1072 spin_unlock(&imp->imp_lock);
1074 if (imp->imp_recon_bk && imp->imp_last_recon) {
1075 /* Give up trying to reconnect */
1076 imp->imp_obd->obd_no_recov = 1;
1077 ptlrpc_deactivate_import(imp);
1080 if (rc == -EPROTO) {
1081 struct obd_connect_data *ocd;
1082 ocd = lustre_swab_repbuf(request, REPLY_REC_OFF,
1084 lustre_swab_connect);
1086 (ocd->ocd_connect_flags & OBD_CONNECT_VERSION) &&
1087 (ocd->ocd_version != LUSTRE_VERSION_CODE)) {
1088 /* Actually servers are only supposed to refuse
1089 connection from liblustre clients, so we should
1090 never see this from VFS context */
1091 LCONSOLE_ERROR_MSG(0x16a, "Server %s version "
1093 " refused connection from this client "
1094 "with an incompatible version (%s). "
1095 "Client must be recompiled\n",
1096 obd2cli_tgt(imp->imp_obd),
1097 OBD_OCD_VERSION_MAJOR(ocd->ocd_version),
1098 OBD_OCD_VERSION_MINOR(ocd->ocd_version),
1099 OBD_OCD_VERSION_PATCH(ocd->ocd_version),
1100 OBD_OCD_VERSION_FIX(ocd->ocd_version),
1101 LUSTRE_VERSION_STRING);
1102 ptlrpc_deactivate_import(imp);
1103 IMPORT_SET_STATE(imp, LUSTRE_IMP_CLOSED);
1108 ptlrpc_maybe_ping_import_soon(imp);
1110 CDEBUG(D_HA, "recovery of %s on %s failed (%d)\n",
1111 obd2cli_tgt(imp->imp_obd),
1112 (char *)imp->imp_connection->c_remote_uuid.uuid, rc);
1115 spin_lock(&imp->imp_lock);
1116 imp->imp_last_recon = 0;
1117 spin_unlock(&imp->imp_lock);
1119 cfs_waitq_signal(&imp->imp_recovery_waitq);
1123 static int completed_replay_interpret(struct ptlrpc_request *req,
1124 void * data, int rc)
1127 atomic_dec(&req->rq_import->imp_replay_inflight);
1128 if (req->rq_status == 0 &&
1129 !req->rq_import->imp_vbr_failed) {
1130 ptlrpc_import_recovery_state_machine(req->rq_import);
1132 if (req->rq_import->imp_vbr_failed) {
1134 "%s: version recovery fails, reconnecting\n",
1135 req->rq_import->imp_obd->obd_name);
1136 spin_lock(&req->rq_import->imp_lock);
1137 req->rq_import->imp_vbr_failed = 0;
1138 spin_unlock(&req->rq_import->imp_lock);
1140 CDEBUG(D_HA, "%s: LAST_REPLAY message error: %d, "
1142 req->rq_import->imp_obd->obd_name,
1145 ptlrpc_connect_import(req->rq_import, NULL);
1150 static int signal_completed_replay(struct obd_import *imp)
1152 struct ptlrpc_request *req;
1155 LASSERT(atomic_read(&imp->imp_replay_inflight) == 0);
1156 atomic_inc(&imp->imp_replay_inflight);
1158 req = ptlrpc_prep_req(imp, LUSTRE_OBD_VERSION, OBD_PING, 1, NULL, NULL);
1160 atomic_dec(&imp->imp_replay_inflight);
1164 ptlrpc_req_set_repsize(req, 1, NULL);
1165 req->rq_send_state = LUSTRE_IMP_REPLAY_WAIT;
1166 lustre_msg_add_flags(req->rq_reqmsg, MSG_LAST_REPLAY);
1167 if (imp->imp_delayed_recovery)
1168 lustre_msg_add_flags(req->rq_reqmsg, MSG_DELAY_REPLAY);
1169 req->rq_timeout *= 3;
1170 req->rq_interpret_reply = completed_replay_interpret;
1172 ptlrpcd_add_req(req);
1177 static int ptlrpc_invalidate_import_thread(void *data)
1179 struct obd_import *imp = data;
1184 ptlrpc_daemonize("ll_imp_inval");
1186 CDEBUG(D_HA, "thread invalidate import %s to %s@%s\n",
1187 imp->imp_obd->obd_name, obd2cli_tgt(imp->imp_obd),
1188 imp->imp_connection->c_remote_uuid.uuid);
1190 ptlrpc_invalidate_import(imp);
1192 /* is client_disconnect_export in flight ? */
1193 spin_lock(&imp->imp_lock);
1194 disconnect = imp->imp_deactive;
1195 spin_unlock(&imp->imp_lock);
1199 if (obd_dump_on_eviction) {
1200 CERROR("dump the log upon eviction\n");
1201 libcfs_debug_dumplog();
1204 IMPORT_SET_STATE(imp, LUSTRE_IMP_RECOVER);
1205 ptlrpc_import_recovery_state_machine(imp);
1208 class_import_put(imp);
1213 int ptlrpc_import_recovery_state_machine(struct obd_import *imp)
1221 if (imp->imp_state == LUSTRE_IMP_EVICTED) {
1222 deuuidify(obd2cli_tgt(imp->imp_obd), NULL,
1223 &target_start, &target_len);
1224 /* Don't care about MGC eviction */
1225 if (strcmp(imp->imp_obd->obd_type->typ_name,
1226 LUSTRE_MGC_NAME) != 0) {
1227 LCONSOLE_ERROR_MSG(0x167, "This client was evicted by "
1228 "%.*s; in progress operations using "
1229 "this service will fail.\n",
1230 target_len, target_start);
1232 CDEBUG(D_HA, "evicted from %s@%s; invalidating\n",
1233 obd2cli_tgt(imp->imp_obd),
1234 imp->imp_connection->c_remote_uuid.uuid);
1237 /* bug 17802: XXX client_disconnect_export vs connect request
1238 * race. if client will evicted at this time, we start invalidate
1239 * thread without referece to import and import can be freed
1241 class_import_get(imp);
1242 rc = cfs_kernel_thread(ptlrpc_invalidate_import_thread, imp,
1243 CLONE_VM | CLONE_FILES);
1245 class_import_put(imp);
1246 CERROR("error starting invalidate thread: %d\n", rc);
1252 ptlrpc_invalidate_import(imp);
1254 IMPORT_SET_STATE(imp, LUSTRE_IMP_RECOVER);
1258 if (imp->imp_state == LUSTRE_IMP_REPLAY) {
1259 CDEBUG(D_HA, "replay requested by %s\n",
1260 obd2cli_tgt(imp->imp_obd));
1261 rc = ptlrpc_replay_next(imp, &inflight);
1262 if (inflight == 0 &&
1263 atomic_read(&imp->imp_replay_inflight) == 0) {
1264 IMPORT_SET_STATE(imp, LUSTRE_IMP_REPLAY_LOCKS);
1265 rc = ldlm_replay_locks(imp);
1272 if (imp->imp_state == LUSTRE_IMP_REPLAY_LOCKS) {
1273 if (atomic_read(&imp->imp_replay_inflight) == 0) {
1274 IMPORT_SET_STATE(imp, LUSTRE_IMP_REPLAY_WAIT);
1275 rc = signal_completed_replay(imp);
1282 if (imp->imp_state == LUSTRE_IMP_REPLAY_WAIT) {
1283 if (atomic_read(&imp->imp_replay_inflight) == 0) {
1284 IMPORT_SET_STATE(imp, LUSTRE_IMP_RECOVER);
1288 if (imp->imp_state == LUSTRE_IMP_RECOVER) {
1289 CDEBUG(D_HA, "reconnected to %s@%s\n",
1290 obd2cli_tgt(imp->imp_obd),
1291 imp->imp_connection->c_remote_uuid.uuid);
1293 rc = ptlrpc_resend(imp);
1296 IMPORT_SET_STATE(imp, LUSTRE_IMP_FULL);
1297 ptlrpc_activate_import(imp);
1299 deuuidify(obd2cli_tgt(imp->imp_obd), NULL,
1300 &target_start, &target_len);
1301 LCONSOLE_INFO("%s: Connection restored to service %.*s "
1302 "using nid %s.\n", imp->imp_obd->obd_name,
1303 target_len, target_start,
1304 libcfs_nid2str(imp->imp_connection->c_peer.nid));
1307 if (imp->imp_state == LUSTRE_IMP_FULL) {
1308 cfs_waitq_signal(&imp->imp_recovery_waitq);
1309 ptlrpc_wake_delayed(imp);
1316 static int back_to_sleep(void *unused)
1321 int ptlrpc_disconnect_import(struct obd_import *imp, int noclose)
1323 struct ptlrpc_request *req;
1325 int nowait = imp->imp_obd->obd_force;
1329 GOTO(set_state, rc);
1331 switch (imp->imp_connect_op) {
1332 case OST_CONNECT: rq_opc = OST_DISCONNECT; break;
1333 case MDS_CONNECT: rq_opc = MDS_DISCONNECT; break;
1334 case MGS_CONNECT: rq_opc = MGS_DISCONNECT; break;
1336 CERROR("don't know how to disconnect from %s (connect_op %d)\n",
1337 obd2cli_tgt(imp->imp_obd), imp->imp_connect_op);
1341 if (ptlrpc_import_in_recovery(imp)) {
1342 struct l_wait_info lwi;
1343 cfs_duration_t timeout;
1346 timeout = cfs_time_seconds(obd_timeout);
1348 int idx = import_at_get_index(imp,
1349 imp->imp_client->cli_request_portal);
1350 timeout = cfs_time_seconds(
1351 at_get(&imp->imp_at.iat_service_estimate[idx]));
1353 lwi = LWI_TIMEOUT_INTR(cfs_timeout_cap(timeout),
1354 back_to_sleep, LWI_ON_SIGNAL_NOOP, NULL);
1355 rc = l_wait_event(imp->imp_recovery_waitq,
1356 !ptlrpc_import_in_recovery(imp), &lwi);
1359 spin_lock(&imp->imp_lock);
1360 if (imp->imp_state != LUSTRE_IMP_FULL)
1363 spin_unlock(&imp->imp_lock);
1365 req = ptlrpc_prep_req(imp, LUSTRE_OBD_VERSION, rq_opc, 1, NULL, NULL);
1367 /* We are disconnecting, do not retry a failed DISCONNECT rpc if
1368 * it fails. We can get through the above with a down server
1369 * if the client doesn't know the server is gone yet. */
1370 req->rq_no_resend = 1;
1373 /* We want client umounts to happen quickly, no matter the
1375 req->rq_timeout = min_t(int, req->rq_timeout,
1376 INITIAL_CONNECT_TIMEOUT);
1378 /* ... but we always want liblustre clients to nicely
1379 disconnect, so only use the adaptive value. */
1381 req->rq_timeout = obd_timeout / 3;
1384 IMPORT_SET_STATE(imp, LUSTRE_IMP_CONNECTING);
1385 req->rq_send_state = LUSTRE_IMP_CONNECTING;
1386 ptlrpc_req_set_repsize(req, 1, NULL);
1387 rc = ptlrpc_queue_wait(req);
1388 ptlrpc_req_finished(req);
1392 spin_lock(&imp->imp_lock);
1395 IMPORT_SET_STATE_NOLOCK(imp, LUSTRE_IMP_DISCON);
1397 IMPORT_SET_STATE_NOLOCK(imp, LUSTRE_IMP_CLOSED);
1398 memset(&imp->imp_remote_handle, 0, sizeof(imp->imp_remote_handle));
1399 /* Try all connections in the future - bz 12758 */
1400 imp->imp_last_recon = 0;
1401 spin_unlock(&imp->imp_lock);
1406 /* Sets maximal number of RPCs possible originating from other side of this
1407 import (server) to us and number of async RPC replies that we are not waiting
1409 void ptlrpc_import_setasync(struct obd_import *imp, int count)
1411 LNetSetAsync(imp->imp_connection->c_peer, count);
1415 /* Adaptive Timeout utils */
1416 extern unsigned int at_min, at_max, at_history;
1418 /* Bin into timeslices using AT_BINS bins.
1419 This gives us a max of the last binlimit*AT_BINS secs without the storage,
1420 but still smoothing out a return to normalcy from a slow response.
1421 (E.g. remember the maximum latency in each minute of the last 4 minutes.) */
1422 int at_add(struct adaptive_timeout *at, unsigned int val)
1424 unsigned int old = at->at_current;
1425 time_t now = cfs_time_current_sec();
1426 time_t binlimit = max_t(time_t, at_history / AT_BINS, 1);
1430 CDEBUG(D_INFO, "add %u to %p time=%lu v=%u (%u %u %u %u)\n",
1431 val, at, now - at->at_binstart, at->at_current,
1432 at->at_hist[0], at->at_hist[1], at->at_hist[2], at->at_hist[3]);
1435 /* 0's don't count, because we never want our timeout to
1436 drop to 0, and because 0 could mean an error */
1439 spin_lock(&at->at_lock);
1441 if (unlikely(at->at_binstart == 0)) {
1442 /* Special case to remove default from history */
1443 at->at_current = val;
1444 at->at_worst_ever = val;
1445 at->at_worst_time = now;
1446 at->at_hist[0] = val;
1447 at->at_binstart = now;
1448 } else if (now - at->at_binstart < binlimit ) {
1450 at->at_hist[0] = max(val, at->at_hist[0]);
1451 at->at_current = max(val, at->at_current);
1454 unsigned int maxv = val;
1455 /* move bins over */
1456 shift = (now - at->at_binstart) / binlimit;
1458 for(i = AT_BINS - 1; i >= 0; i--) {
1460 at->at_hist[i] = at->at_hist[i - shift];
1461 maxv = max(maxv, at->at_hist[i]);
1466 at->at_hist[0] = val;
1467 at->at_current = maxv;
1468 at->at_binstart += shift * binlimit;
1471 if (at->at_current > at->at_worst_ever) {
1472 at->at_worst_ever = at->at_current;
1473 at->at_worst_time = now;
1476 if (at->at_flags & AT_FLG_NOHIST)
1477 /* Only keep last reported val; keeping the rest of the history
1479 at->at_current = val;
1482 at->at_current = min(at->at_current, at_max);
1483 at->at_current = max(at->at_current, at_min);
1486 if (at->at_current != old)
1487 CDEBUG(D_ADAPTTO, "AT %p change: old=%u new=%u delta=%d "
1488 "(val=%u) hist %u %u %u %u\n", at,
1489 old, at->at_current, at->at_current - old, val,
1490 at->at_hist[0], at->at_hist[1], at->at_hist[2],
1494 /* if we changed, report the old value */
1495 old = (at->at_current != old) ? old : 0;
1497 spin_unlock(&at->at_lock);
1501 /* Find the imp_at index for a given portal; assign if space available */
1502 int import_at_get_index(struct obd_import *imp, int portal)
1504 struct imp_at *at = &imp->imp_at;
1507 for (i = 0; i < IMP_AT_MAX_PORTALS; i++) {
1508 if (at->iat_portal[i] == portal)
1510 if (at->iat_portal[i] == 0)
1515 /* Not found in list, add it under a lock */
1516 spin_lock(&imp->imp_lock);
1518 /* Check unused under lock */
1519 for (; i < IMP_AT_MAX_PORTALS; i++) {
1520 if (at->iat_portal[i] == portal)
1522 if (at->iat_portal[i] == 0)
1527 /* Not enough portals? */
1528 LASSERT(i < IMP_AT_MAX_PORTALS);
1530 at->iat_portal[i] = portal;
1532 spin_unlock(&imp->imp_lock);