Re: Z packet support in Gdbserver?

Re: Z packet support in Gdbserver?

[Daniel Jacobowitz]

On Tue, May 25, 2004 at 05:23:56PM +0200, Orjan Friberg wrote:
> I noticed there's no support for z/Z packets for hardware watchpoints in 
> the gdbserver.  (I searched the archives for this, but came up 
> empty-handed.)  I'd guess that (from looking at ptrace.c for i386, where 
> it accesses the DR_ registers) that the Gdbserver gets to configure the 
> hardware breakpoint mechanism through ptrace (PTRACE_POKEUSR, ...).  But 
> how does it get there in the first place, if not as a result of a z/Z 
> packet?  Through direct register writes (G packet)?

It's much simpler than that: gdbserver does not support hardware
watchpoints.  No one's ever implemented it.  As a copy/paste exercise
it would not be too hard; I still have pipe dreams of sharing the code
with GDB someday.

-- 
Daniel Jacobowitz

Re: Z packet support in Gdbserver?

[Andrew Cagney]

I noticed there's no support for z/Z packets for hardware watchpoints in the gdbserver.  (I searched the archives for this, but came up empty-handed.)  I'd guess that (from looking at ptrace.c for i386, where it accesses the DR_ registers) that the Gdbserver gets to configure the hardware breakpoint mechanism through ptrace (PTRACE_POKEUSR, ...).  But how does it get there in the first place, if not as a result of a z/Z packet?  Through direct register writes (G packet)?
Yep.

There are two ways to implement h/w watchpoints: in the target (as was 
done for remote.c); or in the per-architecture code (as was done for 
native i386).  Both are correct.  The missing bit is code to get the 
per-architecture logic working with a remote target - this would involve 
juggling things so that the h/w watchpoint registers are included in the 
regcache and the target/arch to select either a local/remote mechanism.

More notes:
http://sources.redhat.com/cgi-bin/gnatsweb.pl?cmd=view%20audit-trail&database=gdb&pr=735
Andrew

Re: Z packet support in Gdbserver?

[Orjan Friberg]

Daniel Jacobowitz wrote:
It's much simpler than that: gdbserver does not support hardware
watchpoints.  No one's ever implemented it.  As a copy/paste exercise
it would not be too hard; I still have pipe dreams of sharing the code
with GDB someday.
Ah, that explains it.  So what's needed is basically two new function 
pointers in struct target_ops (insert_watchpoint/remove_watchpoint) and 
a snippet of code in server.c to recognize the z/Z packets.  (I assume 
that's what you meant by copy/paste exercise.)

--
Orjan Friberg
Axis Communications

Re: Z packet support in Gdbserver?

[Orjan Friberg]

Andrew Cagney wrote:
There are two ways to implement h/w watchpoints: in the target (as was 
done for remote.c); or in the per-architecture code (as was done for 
native i386).  Both are correct.  The missing bit is code to get the 
per-architecture logic working with a remote target - this would involve 
juggling things so that the h/w watchpoint registers are included in the 
regcache and the target/arch to select either a local/remote mechanism.
Ok; I think I understand.  The CRISv32 target is unlikely to ever have a 
native debugger, but as a remote target it will be debuggable in kernel 
mode (using a traditional remote stub) and in user mode (using 
Gdbserver).  For kernel mode, I have implemented the h/w watchpoint 
register setup as a response to z/Z packets, so either I have to 
duplicate that in the Gdbserver, or move it all into GDB, and have it 
done through G packets.

--
Orjan Friberg
Axis Communications

Re: Z packet support in Gdbserver?

[Andrew Cagney]

Andrew Cagney wrote:

There are two ways to implement h/w watchpoints: in the target (as was done for remote.c); or in the per-architecture code (as was done for native i386).  Both are correct.  The missing bit is code to get the per-architecture logic working with a remote target - this would involve juggling things so that the h/w watchpoint registers are included in the regcache and the target/arch to select either a local/remote mechanism.


Ok; I think I understand.  The CRISv32 target is unlikely to ever have a native debugger, but as a remote target it will be debuggable in kernel mode (using a traditional remote stub) and in user mode (using Gdbserver).  For kernel mode, I have implemented the h/w watchpoint register setup as a response to z/Z packets, so either I have to duplicate that in the Gdbserver, or move it all into GDB, and have it done through G packets.
Yep.

Andrew