[ofa-general] [PATCH] IB/mlx4: Use pgprot_writecombine() for BlueFlame pages

[Or Gerlitz]

Roland Dreier wrote:
> Those results make sense.  The only thing you said that I don't understand:
>  > I noted that I have CONFIG_MTRR=y so maybe this can explain the nice latency
>  > even without setting X86_PAT?
>
> But you are getting the worse latency without X86_PAT -- you need X86_PAT and a patch to use PAT in mlx4_ib to get better latency, which is as I expect.  So I'm not sure why you talk about "nice" latency without X86_PAT.
>   
When I said "nice latency" I referred to the 1.4us number, and the "best 
latency" is the 1.1us number - I'm not sure what you referred to in "the 
worse latency without X86_PAT" - unless you wanted to say that 1.4us is 
the worse... anyway, as I wrote you, on one of the benchmarking sessions 
I did last week I was getting about 3.5us with mainline kernel - but I 
don't manage to reproduce this anymore.

Or.