Re: [SeaBIOS] A20: ahci_process_op is called with A20 off

[Xu, Anthony]

> -----Original Message-----
> From: Paolo Bonzini [mailto:[e-mail]]
> Sent: Friday, May 12, 2017 9:29 AM
> To: Kevin O'Connor ; Xu, Anthony
>
> Cc: [e-mail]; [e-mail]; [e-mail]; Zhong,
> Yang 
> Subject: Re: [SeaBIOS] A20: ahci_process_op is called with A20 off
>
> On 12/05/2017 18:26, Kevin O'Connor wrote:
> > On Thu, May 11, 2017 at 11:22:02PM +0000, Xu, Anthony wrote:
> > > > SeaBIOS has a couple of different methods to accomplish this mode
> > > > switching - it can directly switch modes (C16_BIG switch) or it can
> > > > use a helper in SMM mode to perform the switch (C16_SMM).  The
> > > > preferred method is C16_SMM as C16_BIG isn't available in VM86 mode
> > > > and it also clobbers the 16bit segment registers.
> > >
> > > Hi Kevin,
> > >
> > > Why does SeaBios use VM86 mode?
> >
> > As Paolo mentioned, SeaBIOS doesn't use it, but it must support
> > callers that use it.
> > > vtx supports real mode long time ago.
> > >
> > > C16_SMM is slow because it causes several extra switches between VM
>
> and QEMU.
> > > I would prefer C16_BIG for better performance.
> > > Does SeaBios know if it runs in real mode or VM86 mode?
> > > Can SeaBios choose C16_SMM/C16_BIG based on VCPU mode?
> >
> > Yes, SeaBIOS could use the smsw instruction (see x86.h:cr0_vm86_read)
> > to check if it is in VM86 mode.  Using C16_BIG would still clobber the
> > segment registers, but it might be worth that risk to avoid the extra
> > overhead of C16_SMM.
>
> I don't think the SMM overhead is that bad (is SMM only used on QEMU?).
> At least for KVM it's not bad, maybe 4000 clock cycles roundtrip
> (SeaBIOS does two such trips).

For each trip, there are two switches between SeaBios and QEMU,
One is caused by 
outb %%al, $" __stringify(PORT_SMI_CMD)
to enter SMM mode

The other is caused by
u8 cmd = inb(PORT_SMI_CMD)
in smm handler

Anthony
> For TCG I have a patch to reduce it substantially but I haven't had time
> to send it yet.
>
> Paolo