Re: [E1000-devel] ixgbe MSI-X regression on powerpc

[Tantilov, E. S <emi...@in...>]

Andre Detsch wrote:
> Jeff, Peter,
> 
> I've detected a regression in ixgbe MSI-X initialization when running
> on a ppc64 machine with 2 cpu cores/threads.
> 
> This was the patch that introduced the regression:
> http://git.kernel.org/?p=linux/kernel/git/davem/net-2.6.git;a=commitdiff;h=342bde1b70c79bfc8509b017b3987f3c7541ff8e
> 
> With the current code, MSI-X initialization fails.

Thanks for reporting and apologies for the late reply. 
 
> The reason is that the firmware in power machines only allocate a
> number of queues that is a power of 2. When a request is made for a
> number of queues that is not a power of 2, the value is rounded down.

I had no problem getting 3 vectors on G5 MAC (ppc64). While this is probably
Not the best representation of PPC HW - looks like the `power of 2` rule does
not apply to all PPC systems?
 
> So, after the patch, a typical v_budget value on a system with
> num_online_cpus==2 is 3. The ppc FW gets this request and allocates
> only 2 queues (nearest power of 2 less than 3), which is less than the
> minimum accepted inside the driver (MIN_MSIX_COUNT == 3), causing the
> MSI-X initialization to fail.
> Is there some acceptable way to avoid this problem?
> A lower limit of 4 would have to be imposed to v_budget somehow.
>
> On a side note, it seems that, after the patch, MSI-X will never
> succeed on any machine with only 1 cpu core/thread, since v_budget
> will be no more than 2 (which is less than MIN_MSIX_COUNT). Is this
> intentional? 

On systems with 1 CPU the driver will fall back to using MSI, which is OK.

Thanks,
Emil