Re: [Qemu-devel] [Qemu-riscv] [PATCH] riscv: sifive_e: Correct various SoC IP block sizes

[Bin Meng]

Hi,

On Thu, Sep 5, 2019 at 2:34 AM Palmer Dabbelt <palmer@sifive.com> wrote:
>
> On Tue, 03 Sep 2019 20:41:52 PDT (-0700), bmeng.cn@gmail.com wrote:
> > Palmer,
> >
> > On Wed, Aug 14, 2019 at 5:34 PM Bin Meng <bmeng.cn@gmail.com> wrote:
> >>
> >> Hi Palmer,
> >>
> >> On Wed, Aug 7, 2019 at 10:53 AM Bin Meng <bmeng.cn@gmail.com> wrote:
> >> >
> >> > On Wed, Aug 7, 2019 at 5:06 AM Philippe Mathieu-Daudé 
> >> > <philmd@redhat.com> wrote:
> >> > >
> >> > > On 8/5/19 8:43 AM, Bin Meng wrote:
> >> > > > On Mon, Aug 5, 2019 at 2:14 PM Chih-Min Chao 
> >> > > > <chihmin.chao@sifive.com> wrote:
> >> > > >> On Sat, Aug 3, 2019 at 8:27 AM Bin Meng <bmeng.cn@gmail.com> wrote:
> >> > > >>>
> >> > > >>> Some of the SoC IP block sizes are wrong. Correct them according
> >> > > >>> to the FE310 manual.
> >> > > >>>
> >> > > >>> Signed-off-by: Bin Meng <bmeng.cn@gmail.com>
> >> > > >>> ---
> >> > > >>>
> >> > > >>>  hw/riscv/sifive_e.c | 6 +++---
> >> > > >>>  1 file changed, 3 insertions(+), 3 deletions(-)
> >> > > >>>
> >> > > >>> diff --git a/hw/riscv/sifive_e.c b/hw/riscv/sifive_e.c
> >> > > >>> index 2a499d8..9655847 100644
> >> > > >>> --- a/hw/riscv/sifive_e.c
> >> > > >>> +++ b/hw/riscv/sifive_e.c
> >> > > >>> @@ -53,13 +53,13 @@ static const struct MemmapEntry {
> >> > > >>>      hwaddr base;
> >> > > >>>      hwaddr size;
> >> > > >>>  } sifive_e_memmap[] = {
> >> > > >>> -    [SIFIVE_E_DEBUG] =    {        0x0,      0x100 },
> >> > > >>> +    [SIFIVE_E_DEBUG] =    {        0x0,     0x1000 },
> >> > > >>>      [SIFIVE_E_MROM] =     {     0x1000,     0x2000 },
> >> > > >>>      [SIFIVE_E_OTP] =      {    0x20000,     0x2000 },
> >> > > >>>      [SIFIVE_E_CLINT] =    {  0x2000000,    0x10000 },
> >> > > >>>      [SIFIVE_E_PLIC] =     {  0xc000000,  0x4000000 },
> >> > > >>> -    [SIFIVE_E_AON] =      { 0x10000000,     0x8000 },
> >> > > >>> -    [SIFIVE_E_PRCI] =     { 0x10008000,     0x8000 },
> >> > > >>> +    [SIFIVE_E_AON] =      { 0x10000000,     0x1000 },
> >> > > >>> +    [SIFIVE_E_PRCI] =     { 0x10008000,     0x1000 },
> >> > > >>>      [SIFIVE_E_OTP_CTRL] = { 0x10010000,     0x1000 },
> >> > > >>>      [SIFIVE_E_GPIO0] =    { 0x10012000,     0x1000 },
> >> > > >>>      [SIFIVE_E_UART0] =    { 0x10013000,     0x1000 },
> >> > > >>> --
> >> > > >>> 2.7.4
> >> > > >>>
> >> > > >>
> >> > > >> It seems the modification follows  E310-G002(Hifive1 Rev B) spec 
> >> > > >> and the origin is for E310-G000(Hifive1) spec.
> >> > > >> There should be some way to specify different board version with 
> >> > > >> different memory map or we have policy, always support the latest 
> >> > > >> spec.
> >> > >
> >> > > I agree with Chao, it would be cleaner to have two different boards
> >> > > (machines).
> >> > > Since the SoCs are very similar, you could add a 'revision' property 
> >> > > and
> >> > > use it to select the correct map.
> >> > >
> >> >
> >> > I am not sure if adding two different machines will bring us a lot of
> >> > benefits, since the only difference is the SoC revision with different
> >> > block sizes.
> >> >
> >> > > >>
> >> > > >
> >> > > > Yes, I checked both specs. The older spec says these bigger sizes,
> >> > > > however their register sizes fit well in the smaller range as well. 
> >> > > > So
> >> > > > I think the modification works well for both.
> >> > >
> >> > > This is OK for the PRCI, since sifive_prci_create() does not use
> >> > > memmap[SIFIVE_E_PRCI].size.
> >> > >
> >> > > However the AON case is borderline, since you shrink it from 32KiB to 
> >> > > 4KiB.
> >> > >
> >> >
> >> > AON is not implemented anyway currently. And I checked the FE310 old
> >> > spec, its register block size is still within the 4KiB range, so
> >> > shrinking the size should be fine for both old and new SoC.
> >> >
> >> > > BTW (not related to this patch) it is odd a function named
> >> > > sifive_mmio_emulate() creates a RAM region with 
> >> > > memory_region_init_ram()
> >> > > and does not use the UnimplementedDevice (see make_unimp_dev() in
> >> > > hw/arm/musca.c).
> >> > >
> >>
> >> What's your suggestion regarding this patch?
> >
> > Ping?
>
> Sorry, I missed this the first time around.  In retrospect, it looks like we
> ended up with the wrong naming scheme for boards: sifive_e is very ambiguous,
> as there are many boards that look like this.  We'd originally chosen a more
> explicit scheme (something like "sifive-fe310-g000"), but that was NAK'd as
> resulting in too many machine types.
>
> Peter: would you be OK deprecating "sifive_e" and adding "sifive-fe310-g000"
> and "sifive-fe310-g002" targets?  We'll end up with a lot of machines this 
> way,
> but I don't see another way to closely match what's out there.  In embedded
> land there isn't really any runtime portability, so if the memory maps don't
> match exactly then it's not a useful target for users.

Just want to restart the discussion for this patch. Now that we have
"revB" support for sifive_e machine, I guess we can do something?

But renaming the sifive_e machine to something like sifive-fe31-g000
is another topic .. Thoughts?

Regards,
Bin