Blue Swirl writes:

> On Tue, Aug 30, 2011 at 7:36 PM, Lluís <xscr...@gmx.net> wrote:
>> Blue Swirl writes:
>> 
>>> Even better, taking the instrumentation approach, could the test
>>> device be left out completely, just use guest invisible methods (like
>>> watchpoints) to interact with the guest?
>> 
>> I don't get it. Sorry but I've not been closely following the thread,
>> and a quick look at it gave me no clues about what you mean.

> The use case for the proposed test device is that guest and host can
> interact during testing and debugging, preferably not for production.
> But I think instrumentation needs are not unlike this case. I think I
> proposed earlier to you that instead of intrusive code changes, guest
> invisible methods, for example watchpoints and virtual hardware could
> be used. Maybe we could be able to solve several problems at once? If
> the test device would provide a generic way to connect to an external
> program, wouldn't that be useful also for instrumentation?

Ah, now I get it. Well, guest-host interaction is what I call the
backdoor channel, which from my point of view is orthogonal to
instrumentation. Of course, you could implement the former on top of the
latter.

But I still believe using watchpoints is too heavyweight for the kind of
instrumentation that I want (watchpoints go through the slow memory
access path, while I instrument the guest during TCG code generation),
and does not support all the kinds of information that I want to gather
(i.e., the registers that a specific guest instruction is using).

On the other hand, you really convinced me that using virtual devices is
the right thing to do to implement a backdoor channel, as opposed to my
previous approach of adding per-target special instructions. Thus it
will work in all modes (TCG and KVM) and in all targets.


Lluis

-- 
 "And it's much the same thing with knowledge, for whenever you learn
 something new, the whole world becomes that much richer."
 -- The Princess of Pure Reason, as told by Norton Juster in The Phantom
 Tollbooth

Reply via email to