[asterisk-dev] Another *8 deadlock :(

[Steve Davies]

I will immediately confess that this deadlock only exists in pre-1.8
code, but pretty please could I have an opinion on whether it is safe
to unlock and relock the 'original' channel in
ast_do_masquerade(original) in order to avoid a potential deadlock?

The deadlock is quite easily caused on our test rig, but the use of
ao2 channel refs in 1.8+ solves it quite elegantly there.
Unfortunately I have a number of older systems that I cannot yet
migrate :(

Basically 99% of the code will get the channel-list lock before a
channel lock. ast_do_masquerade() tries to do the opposite. If this
patch is considered safe, I'll raise it as an issue so that the
licence can be flagged properly on the patch.

Thanks,
Steve


# diff -up main/channel.c~ main/channel.c
--- main/channel.c     2011-06-02 17:23:30.000000000 +0100
+++ main/channel.c      2011-06-02 17:23:09.000000000 +0100
@@ -5072,7 +5072,10 @@ int ast_do_masquerade(struct ast_channel
                        clonechan->hangupcause,
                        ast_cause2str(clonechan->hangupcause)
                        );
+               /* Cannot hold a channel lock when calling
ast_channel_free() as it uses the channel-list lock */
+               ast_channel_unlock(original);
                ast_channel_free(clonechan);
+               ast_channel_lock(original);
        } else {
                ast_debug(1, "Released clone lock on '%s'\n", clonechan->name);
                ast_set_flag(clonechan, AST_FLAG_ZOMBIE);