lokinet/contrib/patches/unbound-delete-crash-fix.patch
2022-11-01 11:06:03 -04:00

34 lines
1.3 KiB
Diff

commit 56d816014d5e8a7eb055169c7e13a303dad5e50f
Author: Jason Rhinelander <jason@imaginary.ca>
Date: Mon Oct 31 22:07:03 2022 -0300
Set tube->ev_listen to NULL to prevent double unregister
On windows when using threaded mode (i.e. `ub_ctx_async(ctx, 1)`)
tube_remove_bg_listen gets called twice: once when the thread does its
own cleanup, then again in `tube_delete()`. Because `ev_listen` doesn't
get cleared, however, we end we calling ub_winsock_unregister_wsaevent
with a freed pointer.
This doesn't always manifest because, apparently, for various compilers
and settings that memory *might* be overwritten in which case the
additional check for ev->magic will prevent anything actually happening,
but in my case under mingw32 that doesn't happen and we end up
eventually crashing.
This fixes the crash by properly NULLing the pointer so that the second
ub_winsock_unregister_wsaevent(...) becomes a no-op.
diff --git a/util/tube.c b/util/tube.c
index 43455fee..a92dfa77 100644
--- a/util/tube.c
+++ b/util/tube.c
@@ -570,6 +570,7 @@ void tube_remove_bg_listen(struct tube* tube)
{
verbose(VERB_ALGO, "tube remove_bg_listen");
ub_winsock_unregister_wsaevent(tube->ev_listen);
+ tube->ev_listen = NULL;
}
void tube_remove_bg_write(struct tube* tube)