Re: cmwq and dm-crypt devices?

[Milan Broz <mbroz@redhat.com>]

On 11/03/2010 01:33 PM, Andi Kleen wrote:
>> Async crypto is not solved at all by your patch.
> 
> You mean async crypto with 10 stacked devices? Find me a single
> user who does that.

It is just quick reproducer. It can happen with several
non-stacked devices too and just points to the
problem that there was reason for dedicated threads per device.

I would really prefer If we can fix the scalable problem without
dismantling existing stacking support.

There are users who plans to massively use many dm-crypt devices
in system and this will ensure that it will work properly
even in bizarre configurations.

> Traditionally stacking didn't work very 
> well in the kernel due to the limited kernel stack overflows. 
> I don't think that's significantly different here.

Sorry? device-mapper is designed to be stackable. And it works.

> Anyways stacking could be probably fixed, but it's also that 99.999999%
> of all dm-crypt users don't stack or use async but simply need a 
> scalable dm-crypt.  Extreme stacking is extremly low on the 
> priority list.
> 
> Right now dm-crypt seems to have other problems anyways that need
> to be addressed first.

I fixed many bugs in dm-crypt which were caused by reducing problem
to "common situation" (and which appeared later).

Please can you fix you patch instead of this argumentation?

I know that some fix is needed, I am using the patch myself and I am also
receiving mails requesting it.

The whole thread was reopened because we tried to find solution
to this - IOW fix your patch.

Milan