On 16/05/14 17:47, Eric Dumazet wrote:

On Fri, 2014-05-16 at 17:29 +0100, Zoltan Kiss wrote:

quoted

On 16/05/14 16:34, Wei Liu wrote:

quoted

It works, at least in this Redis testcase. Could you explain a bit where
this 56000 magic number comes from? :-)

Presumably I can derive it from some constant in core network code?

I guess it just makes more unlikely to have packets with problematic layout. But the following packet would still fail:
linear buffer : 80 bytes, on 2 pages
17 frags, 80 bytes each, each spanning over page boundary.

How would you build such skbs ? Its _very_ difficult, you have to be
very very smart to hit this.

I wouldn't build such skbs, I would expect the network stack to create 
such weird things sometimes :)
The goal here is to prepare and handle the worst case scenarios as well.

Also reducing gso_max_size made sure order-5 allocations would not be
attempted in this unlikely case.

But reducing the gso_max_size would have a bad impact on the general 
network throughput, right?