From: Aleksandar Markovic [mailto:address@hidden]
Sent: Thursday, December 05, 2019 1:27 PM
To: Pavel Dovgalyuk
Cc: Michael Rolnik; address@hidden; address@hidden; Sarah Harris; address@hidden
Subject: Re: [Qemu-devel] [PATCH v27 5/8] target/avr: Add limited support for USART and 16 bit timer peripherals
On Thursday, July 25, 2019, Pavel Dovgalyuk <address@hidden> wrote:
> From: Qemu-devel [mailto:qemu-devel-bounces+patchwork-qemu-
> devel=address@hidden] On Behalf Of Michael Rolnik
> From: Sarah Harris <address@hidden>
>
> These were designed to facilitate testing but should provide enough function to be useful in
> other contexts.
USART is very useful for testing, but to which model of AVR is belongs?
We also started implementation of USART and other devices in our internship program,
using prior version of your patches.
There were other register addresses for the registers and some of them even intersect
making read/write logic more complex (we looked at Atmega8).
You also mix the board and the SoC into one file, making hardware-on-chip harder to reuse.
I think that the structure can be revised in the following way:
Board -> SoC -> Devices By "structure", did you mean structure of patches? I mean structure of files/qobjects. Let's say, after the all ISA instruction patches are introduced, we first introduce one real board of our choice (only infrastructure, with almost empty content, than devices on that board, than the corresponding SoC/MCU infrastucture, than device in that SoC. Additional boards would follow the same pattern, potentially reusing already implemented devices, or whole SoC/MCU. That’s why we tried to join peripherals and CPU into SoC/MCU – for reusing the hardware configuration. We already saw that devices within SoC/MCUs for AVR platform exibit great variation. First, there are around 17 generation of AVR cores (avr1, avr2, ... xmega7). Than, there is, I think 600+ SoC/MCU models (hard to believe, but true). Each SoC defines its devices, and in potentially different way (not only its starting address, but real differences in terms of functionality). I thought that at least for a particular core, the devices would be defined in a consistent way, but even that is not true - for example ADC for a SoC having core X can be significantly different that ADC for another SoC having the same core X. How to deal with such avalanche of devices? How to organize and maintain 27 significantly different versions of ADC; and 53 significantly different versions of Watchdogs (the numbers are invented by me, but are likely to describe the situation well)? I hope, that the differences will be rather small and allow making simple flag switches like in mips/translate.c, when decoding depends on CPU features. In the opposite case we can create several instances of the device implementations. Board includes SoC, loads the firmware, and adds some external peripheral devices, if needed.
SoC includes embedded peripherals. It dispatches IO memory accesses and passes them
to the devices. In this case you can have different register addresses for different SoCs, but
the embedded device emulation code can be mostly the same for simple devices like USART.
> Only a subset of the functions of each peripheral is implemented, mainly due to the lack of a
> standard way to handle electrical connections (like GPIO pins).
We did not got too much results, you can check for our changes here: https://github.com/Dovgalyuk/qemu/tree/avr8
But we can help you in development of better version of the patches and split the work
for making this platform more usable.
Pavel Dovgalyuk
Pavel Dovgalyuk
|