Re: tcp: Checking a kmemdup() call in tcp_time_wait()

[Eric Dumazet]

On 10/12/19 7:51 AM, Markus Elfring wrote:
> Hello,
> 
> I tried another script for the semantic patch language out.
> This source code analysis approach points out that the implementation
> of the function âtcp_time_waitâ contains also a call of the function âkmemdupâ.
> https://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git/tree/net/ipv4/tcp_minisocks.c?id=1c0cc5f1ae5ee5a6913704c0d75a6e99604ee30a#n306
> https://elixir.bootlin.com/linux/v5.4-rc2/source/net/ipv4/tcp_minisocks.c#L306
> 
> * Do you find the usage of the macro call âBUG_ONâ still appropriate at this place?
>   https://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git/tree/scripts/checkpatch.pl?id=1c0cc5f1ae5ee5a6913704c0d75a6e99604ee30a#n4080
> 
> * Is there a need to adjust the error handling here?

Presumably the BUG would trigger if a really disturbing bug happened.

There is no chance a timewait socket could be created with a MD5 key, 
if the established socket that is the 'parent' of the timewait
has not a MD5 context itself.

The parent socket only could have MD5 context if tcp_md5sig_pool_populated
could have been set to true.

Once tcp_md5sig_pool_populated is true it can never go back to false.

So the bug here would be that a socket  had a successful MD5 context,
and following tcp_alloc_md5sig_pool() would return false.

We can discuss of all BUG() in general, some people simply disable
all of them (cf CONFIG_BUG), but this particular one does not seem
specially bad to me, compared to others.