Re: [PATCH] 8250: Hypervisors always export working 16550A UARTs.

[Don Dutile]

On 04/29/2016 11:37 AM, Richard W.M. Jones wrote:
> On Fri, Apr 29, 2016 at 08:16:35AM -0700, Greg KH wrote:
> > On Fri, Apr 29, 2016 at 09:10:06AM +0100, Richard W.M. Jones wrote:
> > > On Thu, Apr 28, 2016 at 03:56:33PM -0700, Greg KH wrote:
> > > > On Thu, Apr 28, 2016 at 11:18:33PM +0100, Richard W.M. Jones wrote:
> > > > > Currently autoconf spends 25ms (on my laptop) testing if the UART
> > > > > exported to it by KVM is an 8250 without FIFO and/or with strange
> > > > > quirks, which it obviously isn't.  Assume it is exported to us by a
> > > > > hypervisor, it's a normal, working 16550A.
> > > > >
> > > > > Signed-off-by: Richard W.M. Jones <rjones@xxxxxxxxxx>
> > > > > ---
> > > > >   drivers/tty/serial/8250/8250_port.c | 7 +++++++
> > > > >   1 file changed, 7 insertions(+)
> > > > >
> > > > > diff --git a/drivers/tty/serial/8250/8250_port.c b/drivers/tty/serial/8250/8250_port.c
> > > > > index 00ad2637..de19924 100644
> > > > > --- a/drivers/tty/serial/8250/8250_port.c
> > > > > +++ b/drivers/tty/serial/8250/8250_port.c
> > > > > @@ -1171,6 +1171,13 @@ static void autoconfig(struct uart_8250_port *up)
> > > > >   	if (!port->iobase && !port->mapbase && !port->membase)
> > > > >   		return;
> > > > >
> > > > > +	/* Hypervisors always export working 16550A devices. */
> > > > > +	if (cpu_has_hypervisor) {
> > > > > +		up->port.type = PORT_16550A;
> > > > > +		up->capabilities |= UART_CAP_FIFO;
> > > > > +		return;
> > > > > +	}
> > > >
> > > > Have you audited vmware, virtualbox, and everyone else that provides a
> > > > virtual uart device that it will work properly here?
> > > >
> > > > qemu isn't all the world :)
> > >
> > > Attached below is a slightly different approach.  If the user passes a
> > > special flag on the kernel command line then we force 16550A and avoid
> > > the 25ms delay.  Since the user chooses the flag, any concerns about
> > > the behaviour of the hypervisor or use of VFIO should be moot.
> >
> > No, no more module parameters, that's crazy, what happens when you have
> > 64 serial ports in a system, which one is this option for?
>
> In this (very special) case, the domain is running under qemu and
> I know it only has one serial port that doesn't need probing.
>
> What's the right way to avoid this 25ms delay?
>
> Rich.
param && cpu_has_hypervisor?
.... that restricts it to your use case.
... you can 1+ which hypervisor it is if you add export for pv_info(.name).
... all of which only works on x86, as cpu_has_hypervisor is defined here:
               arch/x86/include/asm/cpufeature.h

which points to a better design based on config option:
    if(CONFIG_<NEW_OPTION>) <your optimization of choice>
then it can be optimized in & out as needed.
Single, binary kernel for bare-metal & virt (e.g., rhel) would
bear the additional CONFIG_xxx check, but that's during boot/init,
which isn't perf sensitive on real hw.

You could then use the above CONFIG_NEW_OPTION to optimize other kvm-guest
boot init callbacks/paths.