> On Jun 7, 2020, at 04:01, Florian Westphal <f...@strlen.de> wrote:
>
> Eric Dumazet <eric.duma...@gmail.com> wrote:
>>> Sure! TCP_DEFER_ACCEPT delays the creation of the socket until data
>>> has been sent by the client *or* the specified time has expired upon
>>> which a last SYN/ACK is sent and if the client replies with an ACK the
>>> socket will be created and presented to the accept()-call. In the
>>> latter case it means that the app gets a socket that does not have any
>>> data to be read - which goes against the intention of TCP_DEFER_ACCEPT
>>> (man-page says: "Allow a listener to be awakened only when data
>>> arrives on the socket.").
>>>
>>> In the original thread the proposal was to kill the connection with a
>>> TCP-RST when the specified timeout expired (after the final SYN/ACK).
>>>
>>> Thus, my question in my first email whether there is a specific reason
>>> to not do this.
>>>
>>> API-breakage does not seem to me to be a concern here. Apps that are
>>> setting DEFER_ACCEPT probably would not expect to get a socket that
>>> does not have data to read.
>>
>> Thanks for the summary ;)
>>
>> I disagree.
>>
>> A server might have two modes :
>>
>> 1) A fast path, expecting data from user in a small amount of time, from
>> peers not too far away.
>>
>> 2) A slow path, for clients far away. Server can implement strategies to
>> control number of sockets
>> that have been accepted() but not yet active (no data yet received).
>
> So we can't change DEFER_ACCEPT behaviour.
> Any objections to adding TCP_DEFER_ACCEPT2 with the behaviour outlined
> by Christoph?
I think that would be useful, although ideally a better, more descriptive name ?
Cheers,
— Leif
