[
https://issues.apache.org/jira/browse/DISPATCH-2046?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=17322095#comment-17322095
]
Alex Ward edited comment on DISPATCH-2046 at 4/15/21, 11:15 AM:
----------------------------------------------------------------
Any suggestions on resolving this race condition? Is there anything
architectuarally that should be preventing the ctx->connection being modified
by the python code?
There is a lock in qd_connector_t that seems to be to prevent this sort of thing
{{144 /* Connector state and ctx can be modified in proactor or management
threads. */}}
{{145 sys_mutex_t *lock;}}
But that is not being held in this case. Should it be somehow? Should it be
held going into handle()?
The connector lock is taken in the management code on line 1059 but nothing is
protecting the connection. Does the connection need it's own lock so the code
would look like:
{{1055 void qd_connection_manager_delete_connector(qd_dispatch_t *qd, void
*impl)}}
1056 {
1057 qd_connector_t ct = (qd_connector_t) impl;
1058 if (ct) {
1059 sys_mutex_lock(ct->lock);
1060 if (ct->ctx) {
{color:#ff0000}new sys_mutex_lock(ct->ctx->lock); {color}
1061 ct->ctx->connector = 0;
was (Author: vorlauf):
Any suggestions on resolving this race condition? Is there anything
architectuarally that should be preventing the ctx->connection being modified
by the python code?
There is a lock in qd_connector_t that seems to be to prevent this sort of thing
{{144 /* Connector state and ctx can be modified in proactor or management
threads. */}}
{{145 sys_mutex_t *lock;}}
But that is not being held in this case. Should it be somehow? Should it be
held going into handle()?
The connector lock is taken in the management code on line 1059 but nothing is
protecting the connection. Does the connection need it's own lock so the code
would look like:
{{1055 void qd_connection_manager_delete_connector(qd_dispatch_t *qd, void
*impl)}}
{{ 1056 { }}
{{ 1057 qd_connector_t ct = (qd_connector_t) impl; }}
{{ 1058 if (ct) { }}
{{ 1059 sys_mutex_lock(ct->lock); }}
{{ 1060 if (ct->ctx) {}}
{color:#FF0000}{{new sys_mutex_lock(ct->ctx->lock); }}{color}
{{ 1061 ct->ctx->connector = 0; }}
> Panic in handle due to deleted connector.
> -----------------------------------------
>
> Key: DISPATCH-2046
> URL: https://issues.apache.org/jira/browse/DISPATCH-2046
> Project: Qpid Dispatch
> Issue Type: Bug
> Components: Routing Engine
> Affects Versions: 1.15.0
> Reporter: Alex Ward
> Priority: Major
>
> I am seeing a segv at line 1063 due to ctx->connector being 0 in
> server.c:handle().
> ugdb-x86_64-7.11-261) p *ctx
> $7 = {prev = 0x8023df490, next = 0x802e7a890, name = 0x0, server =
> 0x80220d0c0, opened = false, closed = false, closed_locally = false, enqueued
> = 0, timer = 0x0, pn_conn = 0x8023ec050, pn_sessions =
> {0x0 <repeats 15 times>}
> , ssl = 0x0, listener = 0x0, connector = 0x0, connector_lock = 0x8023e6600,
> context = 0x0, user_context = 0x0, link_context = 0x0, connection_id = 12,
> user_id = 0x0, free_user_id = false, policy_settings = 0x0, n_sessions = 0,
> n_senders = 0, n_receivers = 0, open_container = 0x0, deferred_calls =
> {head = 0x0, tail = 0x0, scratch = 0x0, size = 0}, deferred_call_lock =
> 0x8023e65c0, policy_counted = false, role = 0x801a17660 "route-container",
> free_link_session_list = \{head = 0x0, tail = 0x0, scratch = 0x0, size = 0}
> , strip_annotations_in = false, strip_annotations_out = false, wake =
> 0x8005d6e20 <connection_wake>, rhost = '\000' <repeats 1024 times>,
> rhost_port = '\000' <repeats 1056 times>}
> As the comment at the start of handle() states, there is only one event being
> handled at a time, but it doesn’t look like there is anything to stop someone
> clearing ctx->connector between lines 1055 and 1063. The python code seems to
> be comin in and deleting the connector while handle is using it
> [https://github.com/apache/qpid-dispatch/blob/main/src/server.c]
> 989 /* Events involving a connection or listener are serialized by the
> proactor so
> 990 * only one event per connection / listener will be processed at a time.
> 991 */
> 992 static bool handle(qd_server_t *qd_server, pn_event_t *e,
> pn_connection_t *pn_conn, qd_connection_t *ctx)
> 993 {
> 994 if (pn_conn && qdr_is_authentication_service_connection(pn_conn))
> {
> 995 qdr_handle_authentication_service_connection_event(e); 996 return true;
> 997 }
> 998
> 999 switch (pn_event_type(e)) {
> …
> 1051 case PN_TRANSPORT_ERROR:
> 1052 {
> 1053 pn_transport_t *transport = pn_event_transport(e);
> 1054 pn_condition_t* condition = transport ?
> pn_transport_condition(transport) : NULL;
> 1055 if (ctx && ctx->connector) { /* Outgoing connection */
> 1056 qd_increment_conn_index(ctx);
> 1057 const qd_server_config_t *config = &ctx->connector->config;
> 1058 ctx->connector->state = CXTR_STATE_FAILED;
> 1059 char conn_msg[300];
> 1060 if (condition && pn_condition_is_set(condition))
> { 1061 qd_format_string(conn_msg, 300, "[C%"PRIu64"] Connection to %s failed:
> %s %s", ctx->co nnection_id, config->host_port, 1062
> pn_condition_get_name(condition), pn_condition_get_description(condition));
> 1063 strcpy(ctx->connector->conn_msg, conn_msg); 1064 1065
> qd_log(qd_server->log_source, QD_LOG_INFO, conn_msg); 1066 }
> else {
> Another thread is at line 1063 in qd_connection_manager_delete_connector so
> it looks like this thread just cleared ctx->connector that is being used in
> handle. What is meant to be preventing this happening?
> [https://github.com/apache/qpid-dispatch/blob/main/src/connection_manager.c]
> 1055 void qd_connection_manager_delete_connector(qd_dispatch_t *qd, void
> *impl)
> 1056 {
> 1057 qd_connector_t *ct = (qd_connector_t*) impl;
> 1058 if (ct) {
> 1059 sys_mutex_lock(ct->lock);
> 1060 if (ct->ctx)
> { 1061 ct->ctx->connector = 0; 1062 if(ct->ctx->pn_conn) 1063
> qd_connection_invoke_deferred(ct->ctx, deferred_close, ct->ctx->pn_conn);
> 1064 1065 }
> 1066 sys_mutex_unlock(ct->lock);
> 1067 DEQ_REMOVE(qd->connection_manager->connectors, ct);
> 1068 qd_connector_decref(ct);
> 1069 }
> 1070 }
> Does anyone have any suggestions for how to prevent this? Doesn't look like I
> could use ct->lock in handle().
--
This message was sent by Atlassian Jira
(v8.3.4#803005)
---------------------------------------------------------------------
To unsubscribe, e-mail: dev-unsubscr...@qpid.apache.org
For additional commands, e-mail: dev-h...@qpid.apache.org
