Re: [PATCH v1] board: mpfs_icicle: implement board_fdt_blob_setup()

[Conor Dooley <conor.dooley@microchip.com>]

[-- Attachment #1: Type: text/plain, Size: 5853 bytes --]

On Thu, Jun 27, 2024 at 09:36:49AM +0100, Simon Glass wrote:
> 
> On Tue, 25 Jun 2024 at 15:34, Tom Rini <trini@konsulko.com> wrote:
> >
> > On Tue, Jun 25, 2024 at 10:08:06AM +0100, Conor Dooley wrote:
> >
> > > The firmware on the Icicle is capable of providing a devicetree in a1 to
> > > U-Boot, but until now the devicetree has been packaged in a "payload" [1]
> > > alongside U-Boot (or other bootloaders/RTOSes) and appended to the image.
> > > The address of this appended devicetree is placed in a1 by the firmware.
> > > This meant that the mechanism used by OF_SEPARATE to locate the
> > > devicetree at the end of the image would pick up the one provided by the
> > > firmware when u-boot-nodtb.bin was in the payload and U-Boot's devicetree
> > > when u-boot.bin was.
> > >
> > > The firmware is now going to be capable of providing a minimal devicetree
> > > (quite cut down due to severe space constraints), but this devicetree is
> > > linked into the firmware that runs out of the L2 rather than at the end
> > > of the U-Boot image. Implement board_fdt_blob_setup() so that this
> > > devicetree can be optionally used, and the devicetree provided in the
> > > "payload" can be used without relying on "happening" to implement the
> > > same strategy as OF_SEPARATE expects in combination with
> > > u-boot-nodtb.bin. Unlike other RISC-V boards, the firmware provided
> > > devicetree is only used when OF_BOARD is set, so that the almost
> > > certainly more complete devicetree in U-Boot will be used unless
> > > explicitly requested otherwise.
> > >
> > > Link: https://github.com/polarfire-soc/hart-software-services/blob/master/tools/hss-payload-generator/README.md [1]
> > > Signed-off-by: Conor Dooley <conor.dooley@microchip.com>
> > > ---
> > > CC: Ivan Griffin <ivan.griffin@microchip.com>
> > > CC: Padmarao Begari <padmarao.begari@microchip.com>
> > > CC: Cyril Jean <cyril.jean@microchip.com>
> > > CC: Tom Rini <trini@konsulko.com>
> > > CC: Conor Dooley <conor.dooley@microchip.com>
> > > CC: u-boot@lists.denx.de
> > > ---
> > >  board/microchip/mpfs_icicle/mpfs_icicle.c | 19 +++++++++++++++++++
> > >  1 file changed, 19 insertions(+)
> > >
> > > diff --git a/board/microchip/mpfs_icicle/mpfs_icicle.c b/board/microchip/mpfs_icicle/mpfs_icicle.c
> > > index 4d7d843dfa3..2c1f7175f0e 100644
> > > --- a/board/microchip/mpfs_icicle/mpfs_icicle.c
> > > +++ b/board/microchip/mpfs_icicle/mpfs_icicle.c
> > > @@ -9,6 +9,7 @@
> > >  #include <init.h>
> > >  #include <asm/global_data.h>
> > >  #include <asm/io.h>
> > > +#include <asm/sections.h>
> > >
> > >  DECLARE_GLOBAL_DATA_PTR;
> > >
> > > @@ -50,6 +51,24 @@ static void read_device_serial_number(u8 *response, u8 response_size)
> > >               response_buf[idx] = readb(MPFS_SYS_SERVICE_MAILBOX + idx);
> > >  }
> > >
> > > +void *board_fdt_blob_setup(int *err)
> > > +{
> > > +     *err = 0;
> > > +     /*
> > > +      * The devicetree provided by the previous stage is very minimal due to
> > > +      * severe space constraints. The firmware performs no fixups etc.
> > > +      * U-Boot, if providing a devicetree, almost certainly has a better
> > > +      * more complete one than the firmware so that provided by the firmware
> > > +      * is ignored for OF_SEPARATE.
> > > +      */
> > > +     if (IS_ENABLED(CONFIG_OF_BOARD)) {
> > > +             if (gd->arch.firmware_fdt_addr)
> > > +                     return (ulong *)(uintptr_t)gd->arch.firmware_fdt_addr;
> > > +     }
> > > +
> > > +     return (ulong *)_end;
> > > +}
> > > +
> > >  int board_init(void)
> > >  {
> > >       /* For now nothing to do here. */
> >
> > I'm adding in Simon and Ilias as this touches on one of those frequent
> > topics about how device trees can/should be passed along to us.
> 
> The only thing I can think of is implementing bloblist in  the a1 (?)
> firmware, then passing the DT in that.

a1 is the register that is used on riscv to pass the dtb, I think the
corresponding thing on arm64 is x0.

Re-reading the firware handoff spec, it's difficult to see what benefits
it actually provides us when we only ever have a single dtb which the
firmware does not interact with/use.
We are super space constrained in the firmware even carving out 4.5 KiB
for a devicetree blob is a stretch and requires disabling other features
and ripping out anything in the DT not required for U-Boot to load the OS.
Even the ~1 KiB mentioned in bloblist.h for handling a bloblist would be a
challenge.

I think the only way a bloblist could work is if it was created at build
time and linked into the firmware, since the on-disk format seems pretty
minimal. Is there tooling for generating a bloblist at build time that I
could use to check whether or not a bloblist is viable? 
I'd also have to investigate how that would interact with OpenSBI, since
it's integrated into the firmware and involved with loading U-Boot.

> It seems that you still need to be able to turn that on and off in
> U-Boot though. So far we have not agreed the mechanism to do that, I
> have the same problem, with a pending patch here[1]

It seems your patch is trying to do some runtime determination of
whether to examine the bloblist or not, but the ?existing? build-time
check for BLOBLIST being enabled would work equally well/poorly as the
OF_BOARD check the code I am adding. I'm not even really sure what runtime
option could be used here here to check if the passed dtb/bloblist was to
be used. U-Boot only runs here as supervisor mode U-Boot proper and always
has a more complete devicetree. Whether to use the one passed to U-Boot
just depends on what the person with the board wants to do - which, given
this is an FPGA, could be vary significantly.

Cheers,
Conor.

[-- Attachment #2: signature.asc --]
[-- Type: application/pgp-signature, Size: 228 bytes --]