Thomas Koeller wrote:
I would like to return to the port type vs. iotype stuff once again.
From what you wrote I seem to understand that the iotype is not just
a method of accessing device registers, but also the primary means of
discrimination between different h/w implementations, and hence every
code to support a nonstandard device must define an iotype of its own,
even though one of the existing iotypes would work just fine?
iotype is all about the access method used to access the registers of
the device, be it by byte or word, and it also takes account of any
variance in the addressing of the registers.
It does not refer to features or bugs in any particular implementation.
Well, the introduction of the UPIO_TSI case seems to contradict this --
it's exactly about the bugs in the particular UART implementation
(otherwise well described by UPIO_MEM). Its only function was to mask 2
hardware issues. And the UUE bit workaround seems like an abuse to me. The
driver could just skip the UUE test altogether based on iotype == UPIO_TSI
(or at least not to ignore writes with UUE set completely like it does but
just mask off UUE bit). With no provision to pass the implicit UART type
for platform devices (and skip the autoconfiguation), the introduction of
UPIO_TSI seems again the necessary evil. Otherwise, we could have this
handled with a distinct TSI UART type...
after a long delay (for which I apologize) I would like to return to
I should also apologize for my discussion style back then -- I just was
short of time (and now I am as well), and might have gotten too nervous for
this topic. Since so much time has passed since the discussion ceased, I
think I should summarize its state at that point:
1. In the RM9000 serial driver patch I submitted, I had introduced a new
port type PORT_RM9000. I set the iotype to UPIO_MEM32. Wherever I
needed to handle any peculiarities of the RM9000 h/w, I did so based
upon the port type. AFAICT this is in agreement with Russel's view as
It was good in theory but the practice has shown that this approach had a
2. Sergei says this is wrong, I need to introduce a new iotype instead,
and make the code conditional upon that, like the AU1000 does (UPIO_AU).
The reason he gives (see quote above) is that there is no way to pass
the port type along with a platform device. Is this argument still valid?
What about including the port type in the information attached to
As they say, "patches welcome". :-)
Btw., I had a look ath the existing code dealing with platform
devices. Is there a particular reason for not using the standard
platform_device.resource mechanism of passing mem/io/irq resources from
the platform to the driver?
It was probably easier to get all info from one place than to parse the
Another topic discussed was the use (or abuse) of dl_serial_read()/
dl_serial_write() to get/set the divisor registers.
More like necessary evil since Pantelis decided to merge support for the
Alchemy UARTs... :-)
3. Russel says (qoute from an earlier mail),
> It's worse than that - this code is there to read the ID from the divisor
> registers implemented in some UARTs. If it isn't one of those UARTs,
> it's expected to return zero.
Could you remind regarding which exactly code Russell has said this,
the one in autoconfig_read_divisor_id()? Are you sure this code gets actually
called for your UART or Alchemy one? From looking at the code, I could
conclude that it shouldn't be -- IIR (sharing address with EFR) shouldn't be
reading as zero unless the modem status interrupt is pending -- that's what
the EFR detection code counts on...
But anyway, I don't see why this code is unsafe to convert to using the
4. I followed the AU1X00 code in this case, using the functions to read/write
the divisor value. My hardware has its UART_DLL and UART_DLM registers at
nonstandard addresses, and so the only other option would have been to
monitor every write to the LCR register for setting or clearing the DLAB
bit, and to switch between different mapping tables accordingly. In the
Why, if we already have a working approach?
light of this, would it still be unacceptable to use the dl_*() functions
tho access the divisor registers?
Why would it be unacceptable, if the like code is already there, in the
Once these issues have been resolved, I will submit an updated version of
I'm afraid you're on your own with resolving the issues now, as the serial
drivers are no longer maintaned by anybody (so, you probably will have to work
with Andrew Morton)...