RE: [PATCH v3 05/44] gdb, gdbserver, gdbsupport: add 'device' tag to XML target description

["Aktemur, Tankut Baris" <tankut.baris.aktemur@intel.com>]

Hi Simon,

On Monday, December 15, 2025 10:04 PM, Simon Marchi wrote:
> On 12/9/25 4:27 PM, Simon Marchi wrote:
> > On 8/1/25 5:37 AM, Tankut Baris Aktemur wrote:
> >> From: Nils-Christian Kempke <nils-christian.kempke@intel.com>
> >>
> >> Add <device> tag to the XML target description.  The tag is a list of
> >> device attributes.  The extension enables passing device information
> >> within the target description XML sent from gdbserver to gdb.
> >
> > The dtd change allows a single <device> in the target description.
> When
> > (and if) you do multi-device debugging, you connect one inferior per
> > device?

Yes, we represent each device as a separate inferior.

> > From the ROCm perspective: if we ever want to support remote debugging
> > (no concrete plans, but we want to keep the door open to it in the
> > design) and we end up using this attribute, I suppose we'll want to be
> > able to use multiple <device> tags, since we do everything in a single
> > inferior.  If so, it would be easier to support that possibility from
> > the start.
> 
> I talked about that with my team, and what I proposed doesn't make
> sense.  The entire target description describes just one arch, it
> wouldn't make sense to have multiple devices of possibly different
> arches in it.
> 
> In ROCgdb, we have an "info agents" command, which lists the agents
> (devices) available to the current inferior (process).  You can, for
> instance, stop at main(), and then type "info agents" to get the list.
> Therefore, I presume that we would instead need a new kind of query,
> "get me the list of devices", and that query could re-use the XML device
> format proposed in this patch.

I agree, this sounds reasonable to me. 

> Another question I have is whether the low-level details specific to one
> device should really belong there in the target description.  In other
> words, imagine that you debug with two identical devices simultaneously,
> should the target descriptions for both devices be identical?  I tend to
> think that they should.  Perhaps things like pci_slot shouldn't be in
> the target description, but in a separate "device" concept outside
> target descriptions.  Things like vendor_id and target_id allow
> inferring details about the architecture, so I understand why it would
> be in there.
> 
> Simon

In our downstream debugger, we have an "info devices" command that lists
the properties of the devices as obtained from this XML information.
vendor_id and target_id are necessary, like you say, so that the target
description can validate it supports that device and can do device-specific
work, if necessary.  Other information is there for informing the user in the 
"info devices" command.  The pci_slot is useful, because if the system has
multiple identical devices, pci_slot becomes the only distinguishing property.
Although it is not a built-in property of the device (like number of cores),
it is still physical&static -- you wouldn’t/couldn't change the pci_slot
of a device during the debug session.  In that sense, it's a bit in the grey
area.

Is pci_slot the only field that concerns you?  We can remove it from the patch
and consider later.

Btw, at some point we should also align on the "info agents" vs. "info devices"
to provide the users with a uniform presentation, but it's not on the critical
path.

-Baris


Intel Deutschland GmbH
Registered Address: Dornacher Straße 1, 85622 Feldkirchen, Germany
Tel: +49 89 991 430, www.intel.de
Managing Directors: Harry Demas, Jeffrey Schneiderman, Yin Chong Sorrell
Chairperson of the Supervisory Board: Nicole Lau
Registered Seat: Munich
Commercial Register: Amtsgericht München HRB 186928