On 01/01/15 20:09, Alexander Duyck wrote:

On 12/31/2014 10:33 AM, Vlad Zolotarov wrote:

quoted

On 12/31/14 20:00, Alexander Duyck wrote:

quoted

I suspect this code is badly broken as it doesn't take several things
into account.

First the PF redirection table can have values outside of the range
supported by the VF.  This is allowed as the VF can set how many bits of
the redirection table it actually wants to use.  This is controlled via
the PSRTYPE register.  So for example the PF can be running with 4
queues, and the VF can run either in single queue or as just a pair of
queues.

Second you could compress this data much more tightly by taking
advantage of the bit widths allowed.  So for everything x540 and older
they only use a 4 bit value per entry.  That means you could
theoretically stuff 8 entries per u32 instead of just 4.

Compression is nice but I think ethtool expects it in a certain
format: one entry per byte. And since this patch is targeting the
ethtool the output format should be as ethtool expects it to be and
this is what this patch does. However I agree that masking the
appropriate bits according to PSRTYPE is required. Good catch!

The idea of compression comes into play when you consider there is
significant latency trying to get messages across the mailbox.  By
reducing the number of messages needed to get the redirection table you
should be able to significantly reduce the amount of time needed to
fetch it.  The job of compressing/expanding the values is actually
pretty straight forward when you consider all that should be needed is a
simple loop to perform some shift, and, and or operations.

We are talking about a super slow path here. So, regardless how slow it 
is it should be done only once between boots. Therefore IMHO any code 
complication is not justified here.

- Alex