On 03. 04. 20 17:48, Maarten Brock wrote: > On 2020-04-03 11:51, Michal Simek wrote: >> >> Thanks. I am definitely interested to hear more how this could be done >> differently because that hardcoded limits are painful. >> On FPGAs you can have a lot of uarts for whatever reason and users are >> using DT aliases to have consistent naming. >> Specifically on Xilinx devices we are using uartps which is ttyPS, >> uartlite which is ttyUL, ns16500 which is ttyS and also pl011 which is >> ttyAMA. >> Only ttyAMA or ttyPS on one chip are possible. >> >> And right now you can't have serial0 alias pointed ttyPS0 and another >> serial0 pointed to ttyUL0 or ttyS0. That's why others are shifted and we >> can reach that hardcoded NR_UART limit easily. >> And this was the reason why I have done these patches in past to remove >> any limit from these drivers and if user asks for serial100 alias you >> simply get ttyPS100 node. > > I would argue that the trouble originates from every uart driver using > its own naming scheme and thereby creating separate namespaces. If all > uarts would register as /dev/ttySnn then the serialN alias method would > work. These non-overlapping namespaces is something the linux kernel > driver community has allowed to happen. > > If the namespaces are not abandoned and disallowed, then the serialN > alias method must no longer be used for any driver that does not create > /dev/ttySnn devices. Every namespace will require its own alias base. > Or forget about deriving the number from an alias and set the number in > a property in the device tree node itself. The latter has my preference. Uartlite and as I see ucc_uart are only two driver which are using port-number property for this purpose. And IIRC this property was the part of any spec long time ago. Thanks, Michal
