[Qemu-devel] Re: RFC: emulation of system flash

[Gleb Natapov]

On Thu, Mar 10, 2011 at 01:06:14PM +0100, Jan Kiszka wrote:
> On 2011-03-10 12:48, Gleb Natapov wrote:
> > On Thu, Mar 10, 2011 at 12:27:55PM +0100, Jan Kiszka wrote:
> >> On 2011-03-10 10:47, Gleb Natapov wrote:
> >>> On Wed, Mar 09, 2011 at 08:51:23PM -0800, Jordan Justen wrote:
> >>>> Hi all,
> >>>>
> >>>> I have documented a simple flash-like device which I think could be
> >>>> useful for qemu/kvm in some cases.  (Particularly for allowing
> >>>> persistent UEFI non-volatile variables.)
> >>>>
> >>>> http://wiki.qemu.org/Features/System_Flash
> >>>>
> >>>> Let me know if you have any suggestions or concerns.
> >>>>
> >>>
> >>> Two things. First You suggest to replace -bios with -flash. This will
> >>> make firmware upgrade painful process that will have to be performed
> >>> from inside the guest since the same flash image will contain both
> >>> firmware and whatever data was stored on a flash which presumably you
> >>> want to reuse after upgrading a firmware. My suggestion is to extend
> >>> -bios option like this:
> >>>
> >>> -bios bios.bin,flash=flash.bin,flash_base=addr
> >>>
> >>> flash.bin will be mapped at address flash_base, or, if flash_base is not
> >>> present, just below bios.bin.
> >>
> >> ...or define -flash in a way that allows mapping the bios image as an
> >> overlay to the otherwise guest-managed flash image.
> >>
> > It is not much different from what I proposed. The result will be the
> > same. Even option syntax will probably be the same :)
> 
> -bios is PC-centric, the new command should be generic.
> 
Well, I tried to reuse the option we already have instead of introducing
another one. -bios can be extended beyond PC and represent general
firmware specification. But I like the option you proposed in other
email too, so I am not going to defend this one.


> > 
> >>>
> >>> Second. I asked how flash is programmed because interfaces like CFI
> >>> where you write into flash memory address range to issue commands cannot
> >>> be emulated efficiently in KVM. KVM supports either regular memory slots
> >>> or IO memory, but in your proposal the same memory behaves as IO on
> >>> write and regular memory on read. Better idea would be to present
> >>> non-volatile flash as ISA virtio device. Should be simple to implement.
> >>
> >> Why not enhancing KVM memory slots to support direct read access while
> >> writes are trapped and forwarded to a user space device model?
> > Yes we can make memory slot that will be treated as memory on read and
> > IO on write, but first relying on that will prevent using flash interface
> > on older kernels and second it is not enough to implement the proposal.
> > When magic value is written into an address, the address become IO for
> > reading too, but KVM slot granularity is page, not byte, so KVM will
> > have to remove the slot to make it IO, but KVM can't execute code from
> > IO region (yet), so we will not be able to run firmware from flash and
> > simultaneously write into the flash. 
> 
> Yeah, right. I remember that this was also hairy over TCG if you tried
> to optimize flash emulation so that writing doesn't take orders of
> magnitude longer than on real HW.
> 
> BTW, the programming granularity is not bytes but chips with common CFI.
> But that's still tricky if you want to run code from the same chip while
> updating parts of it. The easiest workaround would be handling the
> overlay regions as ROM all the time. Not accurate but realizable without
> kernel changes.
> 
So flash will be always IO and overlay will be always ROM. This will
work, except BIOS upgrade from inside the guest will not be possible,
but since we do not support this today too it doesn't bother me to much.

> > 
> >>                                                                 Virtio
> >> means that you have to patch the guest (which might be something else
> >> than flexible Linux...).
> >>
> > This intended to be used by firmware only and we control that.
> 
> I'm thinking beyond this use case, beyond firmware flashes, beyond x86.
> 
OK, but since both interfaces (virtio and one proposed in the wiki) are PV
I fail to see the difference between them for any use case. If we
implement CFI then it will be another story.

--
                        Gleb.