[ cc'ing LKML - since this really belongs there ]

Original thread on linux-ia64 about supporting alternate root for user level emulators:

http://marc.theaimsgroup.com/?t=109708875700003&r=1&w=2

I've reattached the patch here for new readers.

On 10/10/2004 3:10 AM, Christoph Hellwig wrote:
> On Fri, Oct 08, 2004 at 03:37:10PM -0700, Arun Sharma wrote:
>> On 10/8/2004 1:08 AM, David Mosberger wrote:
>> 
>> >Unfortunately, that thread ran out in a rather unhelpful manner, as
>> >far as I can see.  Rusty seemed to agree that the performance-hit of
>> >doing it all in user-level was unacceptably high, but I didn't see any
>> >actual numbers.  There was a suggestion to decouple the altroot from
>> >the personality which makes some sense, but nobody actually did
>> >anything about it?
>> >
>> 
>> I'd really like this issue to be resolved one way or the other. I'm not 
>> sure I've heard a convincing argument on why my original patch(which adds a 
>> new exec domain unconditionally) should not be applied.
>> 
>> I'm fine with the attached patch to set the altroot via a system call as 
>> well.
> 
> I'd still like to see some number on how much smaller a userland emulation
> is.  Best using qumu because that's opensource and we simply don't care
> about intel's propritary stuff.  Else your patch looks pretty okay - but
> I'd need to go through linux-kernel and hooking up to all architectures.

I've prototyped a generic userland solution that covers just open and stat system calls (for completeness, all path walk related syscalls need to be covered) using the LD_PRELOAD approach.

I saw a 16% degradation in system time on this benchmark:

find /usr/src/linux -name '*.[chS]' | xargs grep fsck

mainly due to the doubling of the number of calls to open. Also, there was a slight increase in user time as well, due to malloc/free overhead.

shar file with the benchmark and system call interposer attached.

	-Arun