Re: Socket buffer sizes with autotuning

["Jerry Chu" <hkchu@google.com>]

On Mon, May 12, 2008 at 3:31 PM, David Miller <davem@davemloft.net> wrote:
> From: David Miller <davem@davemloft.net>
>  Date: Mon, 12 May 2008 15:29:00 -0700 (PDT)
>
>
>  > Simply notice, when we're about to decrement in_flight, that the data
>  > reference is one.  You can take appropriate actions if so.
>
>  Actually, there is an even simpler solution.  Since this is a fast
>  clone, when the parent SKB is freed we can see if the fast-clone child
>  is active and has the in_flight atomic_t pointer.  If so, we decrement
>  in_flight and zap it to NULL.

I thought about that too. But what happens if the clone is being freed by
the driver at the same time when TCP is freeing the parent? Isn't there an
intrinsic race condition here? If true, can we use some atomic operation to
zap in_flight, but also return to the caller if in_flight has been nullified
already so that the accounting can be handled correctly? (Although
tp->host_inflight doesn't need to be super accurate, if there is no simple
solution above I'll have to reset tp->host_inflight periodically and convince
myself the pathological case of host_inflight drifting forever won't ever
happen...)

Jerry