Re: [Xen-devel] [PATCH] don't use mlock() with Solaris tools

[John Levon <levon@movementarian.org>]

On Sun, Oct 22, 2006 at 05:09:35PM -0400, Jimi Xenidis wrote:

> Hollis and I touched on this and has been the bane of our existence  
> since we were required to have Xen interpret user level pointers,  
> thread started here:
>   http://lists.xensource.com/archives/html/xen-devel/2005-08/ 
> msg00719.html
> 
> I cannot speak for Hollis (I think he may actually disagree with me)  
> but see this as an opportunity to design something better, or at  
> least have the debat (again).

Yes, we should have something better. At the moment I'm trying to clear
out our patch queue (we only have a few hacks left really now...)

The "special pages" idea of yours sounds workable if the
xc_get_pfn_list() thing can be fixed.

> What might be a better alternative an to actually have an allocate  
> call rather than an mlock call where the arches and OSes could to  
> what is best for them.
> So what is done on x86 could be:
>   do { x = alloca(len); mlock (x, len); } while (0)

This would be a little uncomfortable, if only because it's an extra
system call every time. Also letting userspace pin translations in
memory would be troublesome.

> BTW: John, how do you solve the minor faults?!

We currently interpret the ioctl in the privcmd driver and softlock the
mappings in. Which isn't very nice for all the reasons you point out.

regards
john