[RFC] Adding %cs and %ss for AMD64 to GDB

[RFC] Adding %cs and %ss for AMD64 to GDB

[Mark Kettenis]

Folks,

GDB doesn't handle %cs and %ss for AMD64 yet.  I'm not sure why they
were omitted (might be because they're not particularly useful in long
mode), but there are cases where we might need them.  I might add
these registers after the current last register (%mxcsr), but it just
seems so much more logical to have them close to the other segment
registers.  However, since GDB's register numbering influences the
remote interface, I cannot simply do this.

So here's my question: How bad would it be to change the remote
protocol for AMD64?

I'm proposing to add %cs and %ss just after %eflags.  This would mean
that the segment registers, the floating-point registers and the SSE
registers will shift.  Using a new GDB with an old gdbserver, or an
old remote stub, will mean those registers will contain bogus values.

I did a small survey of the Open Source projects that might make use
of GDB's remote interface to see what the impact would be:

* Linux kgdb doesn't use anything beyond %eflags yet.  Impact would be
  zero.

* FreeBSD kgdb supplies %cs and %ss in the slots that are now reserved
  for %ds and %es.  It provides zero for %fs and %gs (which it thinks
  are %ds and %es).  The impact would be positive!

* NetBSD kgdb doesn't seem to support remote debugging for AMD64 yet.

* Gdbserver supplies %cs and %ss in the slots that are now reserved
  for %ds and %es, %ds and %es in the slots for %fs and %gs, and %fs
  and %gs in the slots for %st0 and %st1.  So the situation for the
  segment registers would actually improve!  However, for the
  floating-point and SSE registers the situation would become worse,
  although right now the most important floating-point registers %st0
  and %st1 might be unreliable.

Note that the problems with gdbserver can be solved by simply
upgrading gdbserver on the remote machine, which in most cases won't
bee too difficult.  I can't imagine that there are many embedded AMD64
systems out there yet.

From the above I conclude that it wouldn't be too bad to add the %cs
and %ss registers the way I propose.  However, don't hesitate to tell
me if you think differently.

Mark

Re: [RFC] Adding %cs and %ss for AMD64 to GDB

[Amit S. Kale]

Mark,

Linux kgdb indeed will not have much of a problem. I anyway have posted a gdb 
with kernel module loading feature at kgdb.sourceforge.net. It's going to 
stay there for till the time this feature makes it to official gdb. I can add 
this change also to that gdb.

I guess it's time to introduce a version stamp on remote protocol. That way 
any further modifications to the protocol will not result in surprises from 
gdb-gdbserver mismatches.

On Thursday 01 Jan 2004 6:21 am, Mark Kettenis wrote:
> Folks,
>
> GDB doesn't handle %cs and %ss for AMD64 yet.  I'm not sure why they
> were omitted (might be because they're not particularly useful in long
> mode), but there are cases where we might need them.  I might add
> these registers after the current last register (%mxcsr), but it just
> seems so much more logical to have them close to the other segment
> registers.  However, since GDB's register numbering influences the
> remote interface, I cannot simply do this.
>
> So here's my question: How bad would it be to change the remote
> protocol for AMD64?
>
> I'm proposing to add %cs and %ss just after %eflags.  This would mean
> that the segment registers, the floating-point registers and the SSE
> registers will shift.  Using a new GDB with an old gdbserver, or an
> old remote stub, will mean those registers will contain bogus values.
>
> I did a small survey of the Open Source projects that might make use
> of GDB's remote interface to see what the impact would be:
>
> * Linux kgdb doesn't use anything beyond %eflags yet.  Impact would be
>   zero.
>
> * FreeBSD kgdb supplies %cs and %ss in the slots that are now reserved
>   for %ds and %es.  It provides zero for %fs and %gs (which it thinks
>   are %ds and %es).  The impact would be positive!
>
> * NetBSD kgdb doesn't seem to support remote debugging for AMD64 yet.
>
> * Gdbserver supplies %cs and %ss in the slots that are now reserved
>   for %ds and %es, %ds and %es in the slots for %fs and %gs, and %fs
>   and %gs in the slots for %st0 and %st1.  So the situation for the
>   segment registers would actually improve!  However, for the
>   floating-point and SSE registers the situation would become worse,
>   although right now the most important floating-point registers %st0
>   and %st1 might be unreliable.
>
> Note that the problems with gdbserver can be solved by simply
> upgrading gdbserver on the remote machine, which in most cases won't
> bee too difficult.  I can't imagine that there are many embedded AMD64
> systems out there yet.
>
> From the above I conclude that it wouldn't be too bad to add the %cs
> and %ss registers the way I propose.  However, don't hesitate to tell
> me if you think differently.
>
> Mark

-- 
Amit Kale
EmSysSoft (http://www.emsyssoft.com)
KGDB: Linux Kernel Source Level Debugger (http://kgdb.sourceforge.net)

Re: [RFC] Adding %cs and %ss for AMD64 to GDB

[Daniel Jacobowitz]

On Thu, Jan 01, 2004 at 11:44:29AM +0530, Amit S. Kale wrote:
> Mark,
> 
> Linux kgdb indeed will not have much of a problem. I anyway have posted a gdb 
> with kernel module loading feature at kgdb.sourceforge.net. It's going to 
> stay there for till the time this feature makes it to official gdb. I can add 
> this change also to that gdb.
> 
> I guess it's time to introduce a version stamp on remote protocol. That way 
> any further modifications to the protocol will not result in surprises from 
> gdb-gdbserver mismatches.

Versioning the remote protocol as a whole isn't feasible, since there
are so many variations.  I've tentatively proposed a solution for the
problem of changing register sets; it's just that neither I nor anyone
else has had time to implement it :)

Mark, I agree with your assessment.  I don't think it will be a
problem.

-- 
Daniel Jacobowitz
MontaVista Software                         Debian GNU/Linux Developer

Re: [RFC] Adding %cs and %ss for AMD64 to GDB

[David O'Brien]

On Thu, Jan 01, 2004 at 01:51:05AM +0100, Mark Kettenis wrote:
> * FreeBSD kgdb supplies %cs and %ss in the slots that are now reserved
>   for %ds and %es.  It provides zero for %fs and %gs (which it thinks
>   are %ds and %es).  The impact would be positive!

From a FreeBSD POV, I would like to see the change made.  We can easily
handle the change.

-- 
-- David  (obrien@FreeBSD.org)