Re: [PATCH v1 2/2] soc: aspeed: add host-side PCIe BMC device driver

[Andrew Jeffery <andrew@codeconstruct.com.au>]

On Wed, 2026-06-03 at 16:30 +0200, Andrew Lunn wrote:
> On Wed, Jun 03, 2026 at 03:43:36PM +0200, Grégoire Layet wrote:
> > > How virtual is this? Is this directly accessing the hardware via
> > > shared memory? Or is there software on the BMC which traps these
> > > reads/writes and responds?
> > 
> > The VUART is virtual because there is no physical UART link between
> > the host and the BMC.
> > Instead, the AST2600 exposes a 16550-compatible register set on both
> > sides (BMC APB and PCIe host MMIO).
> > The data flows using an internal 16 byte FIFO shared between the two
> > register views.
> > So it's hardware emulated and there is no software in the data path.
> > 
> > The AST2600 has four VUARTs, two of which are accessible via PCIe MMIO.
> > This is based on the following section of the AST2600 datasheet:
> > III.48 VUART and III.64 PCI2VUART.
> > 
> > Because the silicon presents a standard 16550A interface in hardware,
> > the existing 8250 driver works without modification.
> 
> So tell us about security.
> 
> Is only this UART exposed in the shared memory?
> 

No, however the BMC PCI interface provides a PCI-to-LPC bridge, so PCI
accesses can drive cycles in to e.g LPC IO devices exposed by the BMC.

>  So the memory window
> is 8 bytes wide? Or are there other peripherals also exposed? How do
> we know the aspeed is not using the UART itself?
> 

For the "regular" SuperIO-controlled UARTs this is a concern, but it's
not a concern for the VUARTs. Each VUART has a pair of 8250 register
sets, one accessible from the BMC, the other accessible to the host,
where both interfaces share the FIFOs to propagate data.

>  If two drivers are
> using it, are we going to crash one or the other system?

By the above, not for the VUARTs.

> 
> https://en.wikipedia.org/wiki/Core_War
> 
> The advantage of rpmsg is that the aspeed would advertise what
> services it is willing to expose. The security issues are different,
> implementation bugs vs exposing bits of hardware to an attacker.

From my understanding rpmsg seems like a reasonable fit for the mailbox
functionality.

Andrew