Re: net/rds: use-after-free in inet_create

[Dmitry Vyukov]

On Tue, Feb 28, 2017 at 5:15 PM, Sowmini Varadhan
<sowmini.varadhan@xxxxxxxxxx> wrote:
> On (02/28/17 16:49), Dmitry Vyukov wrote:
>>
>> Grepping "socket" there, it was doing lots of things with sockets. Are
>> we looking for some particular socket type? If there are few programs
>> that create sockets of that type, then we can narrow down the set:
>
> Yes, we are looking for PF_RDS/AF_RDS - this should be
> #define AF_RDS          21      /* RDS sockets                  */
>
> I see PF_KCM there (value 41) but no instances of 0x15.. how did
> the rds_connect_worker thread get kicked off at all?
>
> the way this is supposed to work is
> 1. someone modprobes rds-tcp
> 2. app tries to do rds_sendmsg to some ip address in a netns - this triggers the
>    creation of an rds_connection, and subsequent kernel socket TCP connection
>    threads (i.e., rds_connect_worker) for that netns
> 3. if you unload rds-tcp, the module_unload should do all the cleanup
>    needed via rds_tcp_conn_paths_destroy. This is done
> Its not clear to me that the test is doing any of this...
>
> is this reproducible? let me check if there is some race window where
> we can restart a connection attempt when rds_tcp_kill_sock assumes
> that the connect worker has been quiesced..


Not reproducible so far.

rds is compiled into kernel (no modules):
CONFIG_RDS=y
CONFIG_RDS_TCP=y

Also fuzzer actively creates and destroys namespaces.

Yes, I don't see socket(0x15) in the log. Probably it was truncated.