Do not attempt to manage OpenSSL locking callbacks in libpq
A deadlock occurs when connecting to Postgresql using SSL with postgresql-libs in
a multi-threaded environment with other threads performing SSL independently.
This issue has been causing Module Build Service (https://pagure.io/fm-orchestrator)
outages, which is deployed on RHEL 7.4 but with the latest postgresql-libs package installed.
You can find a reproducer script and a more in-depth description of the issue here:
https://postgrespro.com/list/thread-id/1861629
The upstream patch that resolves this issue is here:
https://commitfest.postgresql.org/4/140/
RHBZ: #1755154
diff -ur postgresql-9.2.24/src/interfaces/libpq/fe-secure.c postgresql-patch/src/interfaces/libpq/fe-secure.c
--- postgresql-9.2.24/src/interfaces/libpq/fe-secure.c 2017-11-06 23:17:39.000000000 +0100
+++ postgresql-patch/src/interfaces/libpq/fe-secure.c 2019-10-29 15:25:28.448918186 +0100
@@ -941,9 +941,12 @@
if (ssl_open_connections++ == 0)
{
- /* These are only required for threaded libcrypto applications */
- CRYPTO_set_id_callback(pq_threadidcallback);
- CRYPTO_set_locking_callback(pq_lockingcallback);
+ /* These are only required for threaded libcrypto applications, but
+ * make sure we don't stomp on them if they're already set. */
+ if (CRYPTO_get_id_callback() == NULL)
+ CRYPTO_set_id_callback(pq_threadidcallback);
+ if (CRYPTO_get_locking_callback() == NULL)
+ CRYPTO_set_locking_callback(pq_lockingcallback);
}
}
#endif /* HAVE_CRYPTO_LOCK */
@@ -997,10 +1000,13 @@
if (pq_init_crypto_lib && ssl_open_connections == 0)
{
- /* No connections left, unregister libcrypto callbacks */
- CRYPTO_set_locking_callback(NULL);
- CRYPTO_set_id_callback(NULL);
+ /* No connections left, unregister libcrypto callbacks, if no one
+ * registered different ones in the meantime. */
+ if (CRYPTO_get_id_callback() == pq_threadidcallback)
+ CRYPTO_set_id_callback(NULL);
+ if (CRYPTO_get_locking_callback() == pq_lockingcallback)
+ CRYPTO_set_locking_callback(NULL);
/*
* We don't free the lock array. If we get another connection in
* this process, we will just re-use them with the existing mutexes.