Re: [PATCH 06/13] irq: add a helper spread an affinity mask for MSI/MSI-X vectors

Thanks for the responses Bart and Christoph.


On 06/15/2016 07:10 AM, Christoph Hellwig wrote:

On Tue, Jun 14, 2016 at 06:54:22PM -0300, Guilherme G. Piccoli wrote:

quoted

On 06/14/2016 04:58 PM, Christoph Hellwig wrote:

quoted

This is lifted from the blk-mq code and adopted to use the affinity mask
concept just intruced in the irq handling code.

Very nice patch Christoph, thanks. There's a little typo above, on
"intruced".

fixed.

quoted

Another little typo above in "assining".

fixed a swell.

quoted

I take this opportunity to ask you something, since I'm working in a
related code in a specific driver

Which driver?  One of the points here is to get this sort of code out
of drivers and into common code..

A network driver, i40e. I'd be glad to implement/see some common code to raise the topology information I need, but I was implementing on i40e more as a test case/toy example heheh...

quoted

- sorry in advance if my question is
silly or if I misunderstood your code.

The function irq_create_affinity_mask() below deals with the case in which
we have nr_vecs < num_online_cpus(); in this case, wouldn't be a good idea
to trying distribute the vecs among cores?

Example: if we have 128 online cpus, 8 per core (meaning 16 cores) and 64
vecs, I guess would be ideal to distribute 4 vecs _per core_, leaving 4
CPUs in each core without vecs.

There have been some reports about the blk-mq IRQ distribution being
suboptimal, but no one sent patches so far.  This patch just moves the
existing algorithm into the core code to be better bisectable.

I think an algorithm that takes cores into account instead of just SMT
sibling would be very useful.  So if you have a case where this helps
for you an incremental patch (or even one against the current blk-mq
code for now) would be appreciated.

...but now I'll focus on the common/general case! Thanks for the suggestion Christoph. I guess would be even better to have a generic function that retrieves an optimal mask, something like topology_get_optimal_mask(n, *cpumask), in which we get the best distribution of n CPUs among all cores and return such a mask - interesting case is when n < num_online_cpus. So, this function could be used inside your irq_create_affinity_mask() and maybe in other places it is needed.

I was planning to use topology_core_id() to retrieve the core of a CPU, if anybody has a better idea, I'd be glad to hear it.

Cheers,


Guilherme

_______________________________________________
Linux-nvme mailing list
Linux-nvme@lists.infradead.org
http://lists.infradead.org/mailman/listinfo/linux-nvme