improved TSO interface needed

[Andre Oppermann]

Kip Macy wrote:
> On 2/25/07, Andre Oppermann <andre at freebsd.org> wrote:
>> Kip Macy wrote:
>> > Evidently FreeBSD violates the LSO spec by sending down up to socket
>> > buffer sized TSO segments to the network card. Is there a way to set
>> > this other than reducing net.inet.tcp.sendbuf_max to a compliant value
>> > (64k)? And is there a way to for a device to communicate to the stack
>> > the maximum length of an mbuf chain?
>>
>>  From netinet/ip_output.c:
>>
>>           * When doing TSO limit a burst to TCP_MAXWIN minus the
>>           * IP, TCP and Options length to keep ip->ip_len from
>>           * overflowing.  Prevent the last segment from being
>>           * fractional thus making them all equal sized and set
>>           * the flag to continue sending.
>>
>> TCP_MAXWIN is 64K.
> 
> It is good to know that this how it is supposed to work even if it
> doesn't. You didn't directly answer my last question, since they're
> all supposed to be equal does that mean the max should be 32?

The current TSO code will send mbuf chains of up to 64K (minus headers)
to be split into MSS sized packets as specified by mbuf packet header
usually resulting in an MTU of 1500 bytes.

There is no way at the moment to tell the stack to send shorter mbuf
chain while still using TSO.

A google search for "LSO" doesn't yield any useful answers.  The stack
doesn't send socket buffer sized mbuf chains if it larger than 64K.

Can you explain the problem you're seeing in more details perhaps?  That'd
certainly help in finding a fix for your problem.

-- 
Andre