On Mon, Apr 20, 2020 at 04:18:58PM +0800, Icenowy Zheng wrote:
> 在 2020-04-06星期一的 10:27 +0200，Maxime Ripard写道：
> > Hi,
> > 
> > On Fri, Apr 03, 2020 at 12:05:49AM +0800, Icenowy Zheng wrote:
> > > The Allwinner H6 SoC uses DesignWare's PCIe controller to provide a
> > > PCIe
> > > host.
> > > 
> > > However, on Allwinner H6, the PCIe host has bad MMIO, which needs
> > > to be
> > > workarounded. A workaround with the EL2 hypervisor functionality of
> > > ARM
> > > Cortex cores is now available, which wraps MMIO operations.
> > > 
> > > This patch is going to add a driver for the DWC PCIe controller
> > > available in Allwinner SoCs, either the H6 one when wrapped by the
> > > hypervisor (so that the driver can consider it as an ordinary PCIe
> > > controller) or further not buggy ones.
> > > 
> > > Signed-off-by: Icenowy Zheng <icenowy@aosc.io>
> > > ---
> > > There's no device tree binding patch available, because I still
> > > have
> > > questions on the device tree compatible string. I want to use it to
> > > describe that this driver doesn't support the "native Allwinner H6
> > > PCIe
> > > controller", but a wrapped version with my hypervisor.
> > > 
> > > I think supporting a "para-physical" device is some new thing, so
> > > this
> > > patch is RFC.
> > > 
> > > My hypervisor is at [1], and some basic usage documentation is at
> > > [2].
> > > 
> > > [1] https://github.com/Icenowy/aw-el2-barebone
> > > [2] 
> > > https://forum.armbian.com/topic/13529-a-try-on-utilizing-h6-pcie-with-virtualization/
> > 
> > I'm a bit concerned to throw yet another mandatory, difficult to
> > update, component in the already quite long boot chain.
> > 
> > Getting fixes deployed in ATF or U-Boot is already pretty long,
> > having
> > another component in there will just make it worse, and it's another
> > hard to debug component that we throw into the mix.
> > 
> > And this prevents any use of virtualisation on the platform.
> > 
> > I haven't found an explanation on what that hypervisor is doing
> > exactly, but from a look at it it seems that it will trap all the
> > accesses to the PCIe memory region to emulate a regular space on top
> > of the restricted one we have?
> > 
> > If so, can't we do that from the kernel directly by using a memory
> > region that always fault with a fault handler like Framebuffer's
> > deferred_io is doing (drivers/video/fbdev/core/fb_defio.c) ?
> 
> I don't know well about the memory management of the kernel. However,
> for PCIe memory space, the kernel allows simple ioremap() on it. So
> wrapping it shouldn't be so easy.

I'm not sure this would cause any trouble, it's worth exploring I guess. This
would solve all the current shortcomings.

Maxime
>