RE: [PATCH] gdbserver: try selecting a thread first to access memory

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

On Tuesday, July 18, 2023 3:09 PM, Pedro Alves wrote:
> On 2023-06-20 09:35, Tankut Baris Aktemur via Gdb-patches wrote:
> > Since commit
> >
> >   commit 7f8acedeebe295fc8cc1d11ed971cbfc1942618c
> >   Author: Pedro Alves <pedro@palves.net>
> >   Date:   Tue Apr 5 13:57:11 2022 +0100
> >
> >     gdbserver: track current process as well as current thread
> >
> > gdbserver switches to a process, rather than a thread, before
> > processing a memory access request coming from the GDB side.  This
> > switch sets current_thread to null.  Some memory accesses on certain
> > targets, however, may require having a thread context.  Therefore, try
> > switching to the selected thread first.  If this fails for a reason
> > (e.g.  the thread has exited), switch to the process.
> 
> The reason I liked always selecting a process (with thread == null), is
> that it made the corner case be the case that runs all the time.  I.e., any
> target backend that incorrectly assumes there's always a thread selected,
> is immediately exposed to the problem scenario and is forced to fix it.
> Trying to select the thread first loses that.

I'll refer to this comment below as [1].

> If the access is for a thread-specific address space, and the thread exited,
> what is supposed to happen?  Is the backend supposed to error out in that
> scenario?  Or will it misbehave?

I'd expect the memory access at the target side to error out.  Gdbserver would
then convert this to an 'E' response.

> How do you tell what address space the access is targeting?  Are you encoding
> that in the address, or do you have more changes to gdbserver, and maybe
> RSP extensions?

Currently we're encoding the address space in the address (upper 3 bits).

> I'm wondering whether we should have something like:
> 
>   if ((aspace_thread_specific (...) && set_desired_thread ())
>       || set_desired_process ())
>      res = read_inferior_memory (memaddr, myaddr, len);
>   else
>      res = 1;

I believe `aspace_thread_specific` would have to be a target-specific
logic.  What you referred to in the comment [1] above, i.e.  "immediately
exposed to the problem scenario and is forced to fix it", would be caught
and the target would fix the problem by implementing `aspace_thread_specific`
appropriately.  This makes sense to me.  An alternative would be that
the target calls `set_desired_thread` itself, before accessing the memory,
if it detects that the request is thread-specific.  I don't see any low-target
using `set_desired_thread` ever in the gdbserver codebase, though.  So, it
might be considered a breakage of abstraction.  Hence, your suggestion looks
favorable.

Thanks
-Baris


Intel Deutschland GmbH
Registered Address: Am Campeon 10, 85579 Neubiberg, Germany
Tel: +49 89 99 8853-0, www.intel.de <http://www.intel.de>
Managing Directors: Christin Eisenschmid, Sharon Heck, Tiffany Doon Silva  
Chairperson of the Supervisory Board: Nicole Lau
Registered Office: Munich
Commercial Register: Amtsgericht Muenchen HRB 186928