tcp: Correctly compute the TCP goodput in bits per second by using SEQ_SUB().

[• hselasky]

Summary

MFC after: 3 days
Sponsored by: NVIDIA Networking

[rscheff]

Should we generally use the SEQ_SUB macro throughout, to prevent casting issues? (mechanically replace all instances, where a type tcpseq is subtracted from another type tcpseq?)

[• hselasky]

Maybe. Most places where TCP sequence numbers are subtracted, they end up in integers, so then there is no problem.

[• hselasky]

I need to look at closer at this.

sys/netinet/tcp_input.c

324

I think we need to add SEQ_GET(th->th_ack, tp->gput_seq) here too, just to ensure that the SEQ_SUB() doesn't return a signed negative value.

324

SEQ_GET -> SEQ_GEQ

[• hselasky]

BTW: The current code seems OK, but may be confusing to review.

[rscheff]

In D35505#805182, @hselasky wrote:

BTW: The current code seems OK, but may be confusing to review.

You shouldn't end up doing this calculation on ACKs from below the left edge of the window (th_ack < snd_una)

[• hselasky]

Right, so then the SEQ_SUB() shouldn't become negative?

[rscheff]

In D35505#805188, @hselasky wrote:

Right, so then the SEQ_SUB() shouldn't become negative?

During normal processing, th_ack should always be in the range snd_una ... snd_max. All cases outside of this window should be addressed before normal congestion-control processing. The cases with explicit unsigned int typcasting should be adressed for formal reasons, but generally, those calculations should return positive values.