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

[Daniel Schwierzeck]

2016-05-17 14:48 GMT+02:00 Paul Burton <paul.burton at imgtec.com>:
> On Tue, May 17, 2016 at 06:11:22AM -0600, Simon Glass wrote:
>> Hi Paul,
>>
>> On 16 May 2016 at 11:44, Paul Burton <paul.burton at imgtec.com> wrote:
>> > If the UART is to be accessed using I/O port accessors (inb & outb) then
>> > using map_physmem doesn't make sense, since it operates in a different
>> > memory space. Remove the call to map_physmem when
>> > CONFIG_SYS_NS16550_PORT_MAPPED is defined, allowing I/O port addresses
>> > to not be mangled by the incorrect mapping.
>> >
>> > Signed-off-by: Paul Burton <paul.burton at imgtec.com>
>> > ---
>> >
>> > Changes in v2:
>> > - New patch, part of a simplified approach tackling only a single Malta UART.
>> >
>> >  drivers/serial/ns16550.c | 8 ++++++++
>> >  1 file changed, 8 insertions(+)
>> >
>> > 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:

--- a/drivers/serial/ns16550.c
+++ b/drivers/serial/ns16550.c
@@ -100,7 +100,8 @@ static void ns16550_writeb(NS16550_t port, int
offset, int value)
        unsigned char *addr;

        offset *= 1 << plat->reg_shift;
-       addr = map_physmem(plat->base, 0, MAP_NOCACHE) + offset;
+       addr = plat->base + offset;
+
        /*
         * As far as we know it doesn't make sense to support selection of
         * these options at run-time, so use the existing CONFIG options.
@@ -114,7 +115,7 @@ static int ns16550_readb(NS16550_t port, int offset)
        unsigned char *addr;

        offset *= 1 << plat->reg_shift;
-       addr = map_physmem(plat->base, 0, MAP_NOCACHE) + offset;
+       addr = plat->base + offset;

        return serial_in_shift(addr + plat->reg_offset, plat->reg_shift);
 }
@@ -400,7 +401,12 @@ int ns16550_serial_ofdata_to_platdata(struct udevice *dev)
        if (addr == FDT_ADDR_T_NONE)
                return -EINVAL;

+#ifdef CONFIG_SYS_NS16550_PORT_MAPPED
        plat->base = addr;
+#else
+       plat->base = map_physmem(addr, 0, MAP_NOCACHE);
+#endif
+
        plat->reg_offset = fdtdec_get_int(gd->fdt_blob, dev->of_offset,
                                     "reg-offset", 0);
        plat->reg_shift = fdtdec_get_int(gd->fdt_blob, dev->of_offset,

-- 
- Daniel