On 09/15/2016 08:42 AM, Christoph Hellwig wrote:
On Thu, Sep 15, 2016 at 08:34:42AM -0600, Jens Axboe wrote:
I was going to ask about splitting it, but that looks fine, I can pull
that in.
The series looks fine to me. My only real concern is giving drivers the
flexibility to define mappings, I don't want that to evolve into drivers
(again) doing stupid things wrt mappings. As long as we keep it strictly
as a tunnel for passing mappings defined by the (previous blk-mq) core
code, then that's fine.
So my earlier versions just passed in the affinity mask and left
all the mapping in the core. This doesn't really work anymore with
the sibling aware code so I had to add a method. That being said there
are some drivers that might want slightly different mappings.
For example skd (if converted to blk-mq) has MSI-X vectors for it's up
to four queues, but it also has MSI-X vectors for misc book keeping
before those, so we'd need a version of our PCI mapping that adds an
offset to add to queue number when assining the MSI-X vectors.
That being said I structured the map_queues interface so it can't do
anything crazy - it can just build up the cpu to queue mapping array
so there isn't exactly a whole lot of crazy things a driver could do.
By crazy, I mean even things like thinking it knows better and defining
mappings differently just because it can. But I'm not too worried about
it, it's just something to watch for that we didn't have to care about
before, as the mappings were completely mandated by the core code.
--
Jens Axboe
--
To unsubscribe from this list: send the line "unsubscribe linux-block" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at http://vger.kernel.org/majordomo-info.html
