Re: [patch 06/11] Use virtual debug registers in process/thread handling code

[Ingo Molnar <mingo@elte.hu>]

* Alan Stern <stern@rowland.harvard.edu> wrote:

> On Tue, 10 Mar 2009, Ingo Molnar wrote:
> 
> > * Alan Stern <stern@rowland.harvard.edu> wrote:
> > 
> > > > Speaking of switch_to_thread_hw_breakpoint(), i dont like 
> > > > that function at all:
> > > > 
> > > > - why does it have to do a list of debug registers?
> > > 
> > > I'm not sure I understand the point of this question.  Are you 
> > > asking why the hw_breakpoint structures are stored on a list?  
> > > Because there can be an arbitrarily large number of them.
> > 
> > But that does not make much sense. There's just 4 hardware 
> > registers. There's no sane way to overcommit them hence we 
> > _should not_.
> 
> The number of hardware registers will vary according to the 
> architecture.  Our intention was to make the hardware 
> breakpoint interface architecture-neutral, as nearly as 
> possible.  Hence we decided to let callers register arbitrary 
> numbers of breakpoints, and inform them when the breakpoints 
> actually got installed in or uninstalled from the debug 
> registers.

This may sound as handwaving, but the thing is, it's best to do 
these kinds of things gradually. Keep it clean, design for sane 
hardware first (and x86, as a rare exception i guess, is rather 
sane when it comes to hw debug features), add quirks on an 
as-needed basis.

That principle is _especially_ true when a feature with 
borderline utility is merged. We had to do that with KGDB: had 
to strip down a decade of cruft and it really helped.

> If you think this design decision is a bad one, we can discuss 
> it.  But Roland should be involved, because it is in large 
> part his design.

Sure.

> > > > - why does it worry about IPIs arriving when context-switches on 
> > > >   x86 are always done with interrupts disabled?
> > > 
> > > The routine gets invoked at times other than during a 
> > > context switch.  However you may be right that these times 
> > > are all mutually exclusive.  If so then a good deal of 
> > > complication can be removed.
> > 
> > Yes.
> 
> After looking through it more carefully, I think you're right 
> -- if a kernel breakpoint change does occur while 
> switch_to_thread_hw_breakpoint() is running then the IPI will 
> arrive immediately afterward, so there's no need to check for 
> it explicitly.  (When this was written I probably wasn't aware 
> that interrupts are disabled during context switches.)  So all 
> the stuff involving "goto restart" can be removed.

Good - that certainly makes the code we execute during 
context-switch a lot more palatable.

	Ingo