lwIP is usually known as a good, small enough TCP/IP stack, indeed, but its footprint would be much above 8K. There's also uIP.

I know the footprint could be over 20k with all features enabled, but lwIP has features that make it easier to use and more suitable than uIP (some features I would have to write to use in the kernel anyway, like dynamic interface support), and it has better network performance than uIP (especially by allowing more than one unacknowledged packet at a time). My kernel is only 98k currently, which is still less than two full pages of Flash, so I'm not too concerned about the footprint yet. Then again, its footprint will be nothing if I never get around to adding networking.