On 11/19/18 6:32 AM, Christoph Hellwig wrote:

I just saw the patch that avoids the irq disabling show up in your
tree this morning.  I think we can do even better by using slightly
lazy lists that are not updated from ->ki_complete context.

I totally agree, it's just a first step. One restriction you added
is that only one task can be on io_getevents at the time for an
io_context, I didn't want to go that far. One issue with that is
that we can find events that are for other io contexts, if you
saw my patch for the stack list, then that's in there too. So you have
to deal with the case where multiple tasks can be completing events for
the same io_context, regardless of whether you try to block that
behavior at io_getevents() entry time.

Please take a look at the patch below - this replaces patch 3 from
my previous mail, that is it is on top of what you send to the list
plus my first two patches.

Thanks, I'll take a look. I'll split out fhe fops change, as that is
indeed how I wanted to get this done ultimately as well, since it's the
only way to support more complicated setups.

-- 
Jens Axboe