preemption counter havoc on kgdb-taken faults (was: kgdb Bad IO access)

[Jan Kiszka]

Jason Wessel wrote:
> The only place I could foresee that kgdb could be eating extra cycles in 
> the runtime case is from the die_notifier processing.  Any kind of 
> exception such as a page fault, trap etc... will have a few extra ops 
> and checks of variables so as to determine if the debugger should take 
> the exception.  It looks to me like it would even benefit to add the 
> check at the top of the notify hook for kgdb to exit immediately if the 
> debugger is not attached.
> 
> I have contemplated making some changes to KGDB so as to make the 
> registration to the die_notifier to be dynamic with attaching and 
> detaching of the debugger.  If this is done, I would also make a change 
> to allow for the case where the kernel would wait for the debugger to 
> attach on any fatal fault.

At this chance... Reminds me that this old issue still seems to be
unsolved in current kgdb:

http://www.mail-archive.com/kgdb-bugreport@xxxxxxxxxxxxxxxxxxxxx/msg00442.html

I'm only looking at that spot in kgdb right now and /may/ oversee new
border conditions elsewhere. But my feeling is there are none.

Jan (looking forward to see kgdb merged)

Attachment: signature.asc
Description: OpenPGP digital signature