Re: [PATCH net v4 1/2] sfc: use 64-bit writes for PIO.

[Robert Stonehouse <rstonehouse@solarflare.com>]

On 03/06/14 23:58, David Miller wrote:
> Did you do any functional testing of this on such a machine?
>
> I'm extremely disappointed in this patch submission, because you
> didn't even_compile_  test this in the environment where you claim
> the problem exists.

The next patch posted in the series would have meant that this change 
was limited to x86_64 systems.
I am sorry that by posting the patches in an incorrect order/not merging 
them we gave the impression that we do not test these patches (but hands 
up, patch 1/2 did break 32 bit compiles and certainly did need more 
work. Thanks to Sergei, Joe and yourself for comments)

For x86_64 machines we checked writeq() implementations in io.h and 
disassembled the module to ensure that a single 64bit access was used.

I would like to get an improved bug fix upstream, as I know this rare 
hardware issue has been hit in the wild  and with TX checksum offload 
enabled corrupted data can get to the application.

The reason for limiting to x86_64 is that in this code path there is an 
IO write to a write-combined mapping. It was a very hard case to hit 
(believed to depend on the timing of accesses and resulting 
write-combining behaviour), but we never reproduced it in our lab.
Limiting this IO to WC mappings to x86_64 only was a safety measure 
given the amount of hours of testing we have on x86_64 vs.other 64 bit 
platforms that we test with (PPC64)
This feature gives a small latency win on this hardware family for small 
packets.

Unless there are more comments to consider the next post will be a 
merged patch, addressing all points raised so far.

Regards
Rob