Re: [PATCHv2 4/5] KVM: emulator: move linearize() out of emulator code.

[Avi Kivity <avi@redhat.com>]

On 06/24/2012 05:27 PM, Gleb Natapov wrote:
> On Sun, Jun 24, 2012 at 04:39:22PM +0300, Avi Kivity wrote:
>> On 06/24/2012 04:27 PM, Gleb Natapov wrote:
>> > On Sun, Jun 24, 2012 at 04:12:05PM +0300, Avi Kivity wrote:
>> >> On 06/12/2012 03:01 PM, Gleb Natapov wrote:
>> >> > The function will be used outside of the emulator.
>> >> > 
>> >> >  /*
>> >> >   * x86_emulate_ops:
>> >> >   *
>> >> > @@ -194,6 +199,10 @@ struct x86_emulate_ops {
>> >> >  
>> >> >  	bool (*get_cpuid)(struct x86_emulate_ctxt *ctxt,
>> >> >  			 u32 *eax, u32 *ebx, u32 *ecx, u32 *edx);
>> >> > +
>> >> > +	int (*linearize)(struct x86_emulate_ctxt *ctxt,
>> >> > +			struct segmented_address addr, unsigned size,
>> >> > +			bool write, bool fetch, ulong *linear);
>> >> >  };
>> >> >  
>> >> 
>> >> linearize is defined in terms of the other ops; this means that if we
>> >> get a second user they will have to replicate it.
>> >> 
>> > What do you mean? This patch series adds another user, so now there are two: one
>> > inside the emulator another is outside.
>> 
>> I meant like task switching or real-mode interrupt emulation.
>> 
> You mean code outside of KVM if we ever will make emulator reusable? It will have to
> have its own, much more simple version of the callback.
> 
>> > 
>> >> Why not make the current linearize available to users?
>> >>
>> > Code outside of the emulator does not call the emulator except when
>> > emulation is actually needed. To call linearize() from the emulator.c
>> > almost fully functional emulation ctxt will have to be set up (including
>> > fake instruction decoding, hacky and slower). 
>> 
>> ctxt->d use should be removed for the exported version and replaced by a
>> parameter.  The internal version can still use it (calling the exported
>> version after extracting the parameter).
>>
> IMO we should stick to the pattern we have now: calling generic code from
> the emulator and not vice versa. Lets not create more spaghetti.
> 
>>  To not duplicate the logic
>> > I moved linearize() to generic code and made it available to emulator
>> > via callback. It actually saves a couple of callback invocations when
>> > emulator calls linearize() IIRC.
>> 
>> It's not available to other emulator users (which don't exist yet
>> anyway).  But having linearize() in the emulator is consistent with
>> placing logic in emulate.c and accessors outside.
>> 
> It is the question of where we draw the line. For instance MMU details
> are now hidden from the emulator behind a callback. One can argue that
> emulator should have access to MMU directly via callbacks and
> emulate memory access by itself.

Right now the all segment related operations are behind the line; the
line is linear | physical.  Having a ->linearize op will change that.

> 
>> Regarding initialization, we should eventually initialize nothing and
>> let the emulator bring in needed data via callbacks (including general
>> registers).
>> 
> Some things will have to be initialized (or rather reset to initial value)
> between emulator invocations. Access to registers can be done on demand,
> but this is unrelated to this series optimization.

Right.  But I think we can have x86_linearize() that doesn't take a
context parameter, only ops.

-- 
error compiling committee.c: too many arguments to function