[rds-devel] [ofa-general] Re: [PATCH 06/21] RDS: Connection handling

[Andrew Grover]

On Tue, Jan 27, 2009 at 8:28 AM, Steve Wise <swise at opengridcomputing.com> wrote:
> Oliver Neukum wrote:
>>
>> Am Tuesday 27 January 2009 14:34:19 schrieb Evgeniy Polyakov:
>>
>>>
>>> On Mon, Jan 26, 2009 at 06:17:43PM -0800, Andy Grover
>>> (andy.grover at oracle.com) wrote:
>>>
>>>>
>>>> +static inline int rds_conn_is_sending(struct rds_connection *conn)
>>>> +{
>>>> +       int ret = 0;
>>>> +
>>>> +       if (!mutex_trylock(&conn->c_send_lock))
>>>> +               ret = 1;
>>>> +       else
>>>> +               mutex_unlock(&conn->c_send_lock);
>>>> +
>>>> +       return ret;
>>>> +}
>>>> +
>>>>
>>>
>>> This one is eventually invoked under the spin_lock with turned off irqs,
>>> which may freeze the machine:
>>> rds_for_each_conn_info() -> spin_lock_irqsave(global lock) ->
>>> rds_conn_info_visitor() -> rds_conn_info_set() -> rds_conn_is_sending()
>>> -> boom.
>>>
>>
>> Why? This is _trylock. It won't block.
>>
>>
>
> mutex_trylock() uses spin_lock_mutex() which has this in the debug version:
>
> DEBUG_LOCKS_WARN_ON(in_interrupt());

What's the best way to fix this?

This is all so rds-info can print out a nice list of connections, and
if they're sending or not. I don't see an easy way to fix this. A
_trylock-like function that didn't grab it would be nice? I can always
just not report this particular bit of info, that actually might be
easiest.

Regards -- Andy