Re: [Qemu-riscv] [Qemu-devel] [PATCH] hw/riscv/virt: re-add machine-specific compatible string to /soc/ node

[Auer, Lukas]

Hi Bin,

On Mon, 2019-03-11 at 23:28 +0800, Bin Meng wrote:
> Hi Lukas,
> 
> On Mon, Mar 11, 2019 at 2:03 AM Auer, Lukas
> <address@hidden> wrote:
> > Hi Bin,
> > 
> > On Sun, 2019-03-10 at 22:57 +0800, Bin Meng wrote:
> > > Hi Lukas,
> > > 
> > > On Sun, Mar 10, 2019 at 9:44 PM Auer, Lukas
> > > <address@hidden> wrote:
> > > > Hi Bin,
> > > > 
> > > > On Sun, 2019-03-10 at 09:07 +0800, Bin Meng wrote:
> > > > > Hi Lukas,
> > > > > 
> > > > > On Mon, Feb 11, 2019 at 6:13 AM Lukas Auer
> > > > > <address@hidden> wrote:
> > > > > > Re-add the previous compatible string "riscv-virtio-soc" to
> > > > > > the
> > > > > > soc
> > > > > > device tree node to allow U-Boot and Linux to bind machine-
> > > > > > specific
> > > > > > drivers to it. The current compatible string "simple-bus"
> > > > > > is
> > > > > > retained.
> > > > > > 
> > > > > > This is required by U-Boot to bind devices early, as part
> > > > > > of
> > > > > > the
> > > > > > pre-relocation driver model.
> > > > > > 
> > > > > 
> > > > > I see no problem with U-Boot working with current compatible
> > > > > string
> > > > > "simple-bus". In fact I had planned to remove the compatible
> > > > > string
> > > > > "riscv-virtio-soc" in U-Boot but did not get time to work on
> > > > > it.
> > > > > 
> > > > 
> > > > It is only required if U-Boot is running in machine-mode. For
> > > > relocation it needs to use the CLINT driver to send appropriate
> > > > IPIs to
> > > > the other harts. To be able to probe the driver, the device and
> > > > its
> > > > parent device tree node (soc) must therefore be available in
> > > > the
> > > > pre-
> > > > relocation device model.
> > > > This patch was the easiest way I could think of for achieving
> > > > this.
> > > > It
> > > > could be that there is a better way of solving this.
> > > > 
> > > 
> > > I tested your SMP U-Boot series in both M-mode and S-mode, using
> > > a 4
> > > core 'virt' target. Works fine. I am using QEMU 3.1.0 so it is
> > > "simple-bus".
> > > 
> > 
> > That is actually my fault, it should not work.
> > What is happening is that U-Boot fails to relocate the secondary
> > harts,
> > because the CLINT driver cannot get the memory address of the CLINT
> > device. This error is currently silently ignored.
> 
> I still don't understand. Why does the CLINT driver fail to get the
> memory address? U-Boot has been supporting "simpile-bus" for a long
> time. It was because QEMU 3.0.0 generated the /soc node with
> "riscv-virtio-soc" compatible string, U-Boot was taught to treat such
> compatible string as a "simple-bus" too (that was the U-Boot commit
> 27dc2c130e29)

That's correct. The problem with the default simple-bus U-Boot driver
is that it does not have the DM_FLAG_PRE_RELOC flag set. The /soc and
/soc/clint nodes are therefore not available before relocation, meaning
that IPIs cannot be sent to relocate the secondary harts.

Thanks,
Lukas

> 
> > The secondary harts are still waiting to receive IPIs, so booting
> > Linux
> > works fine, because U-Boot can now send IPIs. This will however
> > break
> > if U-Boot overwrites the code the secondary harts are running,
> > which
> > could happen when loading an image.
> > 
> > I will update my SMP U-Boot series to print a warning if sending an
> > IPI
> > fails.
> > 
> 
> Regards,
> Bin