[U-Boot] [PATCH v2 3/5] dm: ns16550: Don't map_physmem for I/O ports

[Simon Glass]

Hi Paul,

On 17 May 2016 at 09:58, Paul Burton <paul.burton at imgtec.com> wrote:
> On Tue, May 17, 2016 at 09:54:21AM -0600, Simon Glass wrote:
>> > >> > diff --git a/drivers/serial/ns16550.c b/drivers/serial/ns16550.c
>> > >> > index 28da9dd..e58e6aa 100644
>> > >> > --- a/drivers/serial/ns16550.c
>> > >> > +++ b/drivers/serial/ns16550.c
>> > >> > @@ -100,7 +100,11 @@ static void ns16550_writeb(NS16550_t port, int offset, int value)
>> > >> >         unsigned char *addr;
>> > >> >
>> > >> >         offset *= 1 << plat->reg_shift;
>> > >> > +#ifdef CONFIG_SYS_NS16550_PORT_MAPPED
>> > >> > +       addr = (unsigned char *)plat->base + offset;
>> > >> > +#else
>> > >> >         addr = map_physmem(plat->base, 0, MAP_NOCACHE) + offset;
>> > >> > +#endif
>> > >>
>> > >> Please don't add CONFIG #ifdefs in these functions. Perhaps it needs
>> > >> to be another parameter? Possibly a flag. But with driver-model we
>> > >> need to be able to support both options in the core code.
>> > >
>> > > Hi Simon,
>> > >
>> > > Are you sure systems rely on using I/O ports with map_physmem? The only
>> > > other systems that define CONFIG_SYS_NS16550_PORT_MAPPED are x86 ones,
>> > > in include/configs/x86-common.h, and so far as I can tell they don't use
>> > > device model which suggests this code has simply been untested before. I
>> > > don't see why you would use map_physmem on an I/O port address that is
>> > > then going to be passed to inb/outb & I think the code here is simply
>> > > wrong to do so.
>> >
>> > the current code looks wrong. serial_in_shift() is expanded to inb()
>> > in case of CONFIG_SYS_NS16550_PORT_MAPPED and to
>> > in_le32()/in_be32()/readl()/readb() otherwise. Only in the latter case
>> > a map_physmem() is required and should be done in serial_in_shift()
>> > itself or preferrably only once in
>> > ns16550_serial_ofdata_to_platdata().
>> >
>> > I think the correct approach would be the following:
>>
>> This is better I think. But how about adding a device tree binding to
>> select I/O access? In principle each device might have its own
>> settings.
>
> Note that's what I worked towards last time I had a crack at this, but
> it just expanded into an attempt to tackle the mess that is ns16550.c &
> rather lost sight of the original goal of making Malta work.
>
> https://patchwork.ozlabs.org/patch/575643/
> https://patchwork.ozlabs.org/patch/577194/

Yes it is tricky. What do you think about the suggestions above?

Regards,
Simon