Hi Antal!

I'm asking because I looked into STM32 programming with Rust recently, and platform libraries are starting to show up, but nothing as high level as XPCC. I looked into the SVD files, but I don't think that alternate function pins are encoded in them for example (the thing that enables the type-safe connect() methods).

Have you seen @japaric's post on that topic? http://blog.japaric.io/brave-new-io http://blog.japaric.io/brave-new-io

Unfortunately, SVD files don't contain information beyond the memory map and even worse, the SVD files are often incorrect in some minor way. Btw, so are the official CMSIS header files that ST ships. https://github.com/modm-io/cmsis-header-stm32/commit/1ec448711db342deb4d4bf2... https://github.com/modm-io/cmsis-header-stm32/commit/1ec448711db342deb4d4bf20b670166701a38f26#r27042841

Could someone give me a small description of how the device file generation for STM32 works (both the one currently used in XPCC, and/or the one in MODM)? What I mean is, what information comes from what sources, and how is it put together.

In short: The data comes out of CubeMX, which contains a bunch of XML files. These are transformed, cross-referenced and filtered to generate these device files. https://github.com/modm-io/modm-devices/tree/develop/devices/stm32 https://github.com/modm-io/modm-devices/tree/develop/devices/stm32

I've held a talk about how we use these device files in modm at EmBO++17: https://www.slideshare.net/emBO_Conference/datadriven-hal-generation https://www.slideshare.net/emBO_Conference/datadriven-hal-generation

These device files are more complete and more correct than the ones in xpcc and have several important bug fixes (like correctly supporting STM32F1 group remap). I rewrote the generator from scratch to fix these issue, but unfortunately I've made the decision not to publish it. Let me justify this.

1. The internal data in ST's CubeMX is good, but since it's internal data, the format changes and data sometimes gets removed, added, changed. There is still a lot of manual cleanup that needs to happen. I've even sent ST patches of their own internal data, some of which have been applied, others not. Unfortunately, the CubeMX data does not contain some of the most important data either. This I have to manually extract from dozens of reference manuals/datasheets, compare and add manually. This includes: some of STM32F1 GPIO remap, ALL groups of memory sizes (yes!), ALL peripheral types and all device groups.

This is an insane amount of work, and I'm not willing to put in the hours for this anymore than is absolutely necessary. I fear that if I were to publish this generator and it were to become popular (which the Rust community surrounding @japaric absolutely has the ability to do), I'd be overwhelmed with support requests for all sorts of additional data, regardless of how labor intensive it is to procure.

I've spent the last 4 years looking at this data, if it's not in the above device tree data already, it is _not_ trivial to add. An example is the clock tree data, which exists in the raw data, but it's lacking critical data, which again I'd have to manually add. Ultimately the CubeMX data is incomplete and sometimes just plain wrong.

2. There is a real risk of creating a pseudo-standard of assumptions from my work, since there is no other STM32 data available elsewhere at the moment. I've been following the Linux Device Tree community for a while now, especially the discussion about the Device Tree Specification: https://github.com/devicetree-org/devicetree-specification https://github.com/devicetree-org/devicetree-specification

To caricature the discussion so far: A few (truly very smart) engineers are extrapolating their personal, anecdotal experiences of how *they* used data for a few devices into a grand theory of how this data should be formatted and used for *every* other device out there. And they are just simply wrong about that. You are wrong about what you think you know. You experiences mean nothing, you cannot possibly know all the hardware details of thousands of devices. ST has ~1200 devices, Atmel ~600 AVRs, and their configuration data is not as regular as you'd like to think. Here is an (out-of-data) and incomplete overview of ~1000 STM32 devices: http://salkinium.com/stm32.html http://salkinium.com/stm32.html

A better way would be to formulate your personal experiences as executable assumption checkers, and run them over the raw data. This is actually what I did in modm in the GPIO driver, because the assumptions I made in xpcc regarding GPIO AF were wrong. In fact they were _so_ wrong that our assumption check failed on the majority of STM32 devices and we had to rewrite the API from scratch. https://image.slidesharecdn.com/niklashauser-170228153235/95/datadriven-hal-... https://image.slidesharecdn.com/niklashauser-170228153235/95/datadriven-hal-generation-47-1024.jpg

Now imagine a fuckup like that at device tree level, nobody is rewriting anything. I don't want to be yelled at by Linus.

3. Not a single vendor actually wants to give you this data, it makes them so very deeply comparable to each other. Vendors are already unhappy with the rather detailed Digikey search options. The CubeMX data was a very lucky find, and I have not found data even remotely as extensive as this for other vendors. So this is an ST only solution (the AVR data is even more incomplete), and so I refer back to point 2.

4. I recently quit my job at ARM working on mbed OS and uVisor because they didn't allow any out-of-the-box thinking on embedded software designs, like we do in xpcc/modm. I mean, they don't even generate their linker-/startup scripts, which is absolutely TRIVIAL to do with even the most basic data. It's insane how much unnecessary manual labor goes into porting/maintaining mbed OS. My takeaway was that with the entire IoT craze, nobody in this industry takes the time to fix the foundational issues, like HAL design, tool support, library support (don't get me started on Newlib). Most engineers I met can't even point out any flaws in these, they are just so used to it. I'm still downbeat and angry about that and I need time to gain some distance from that experience.

I've been contributing to xpcc/modm now since 2011, that's a long time. I'm not a student anymore, I need choose more carefully what I want to spend time on now. So I'm taking a step back to reorient myself a little and build up motivation again. There are also interesting and more lucrative opportunities in other industries.

For now I recommend you use the device files as they are. You're not going to be able to extract any more data from CubeMX without unreasonable effort. I'll keep them updated with every quarterly release of modm. I'll reevaluate my position if more projects start using that data and people are willing to take up some of the maintenance effort like what has happened in xpcc. That really is a joy to maintain, I feel like people care about this stuff :-) Make sure to look at modm if you want to generate code/data/documentation from this data, since it was designed for exactly that purpose.

Or use the DeviceFile class directly: https://github.com/modm-io/modm-devices/blob/develop/tools/device/modm/devic... https://github.com/modm-io/modm-devices/blob/develop/tools/device/modm/device_file.py Usage: https://github.com/modm-io/modm/blob/develop/repo.lb#L29-L39 https://github.com/modm-io/modm/blob/develop/repo.lb#L29-L39

Sorry for being in a bad mood today, (sad) Niklas

Hi Niklas,

Thanks for this extremely long answer, it is very interesting, sorry for taking so long to get back to you.

On 8. Feb 2018, at 01:52, Niklas Hauser niklas.hauser@rwth-aachen.de wrote:

Have you seen @japaric's post on that topic? http://blog.japaric.io/brave-new-io

Yes, I'm trying to follow new developments in embedded Rust, and especially @japaric, who is doing most of the work. As far as I know, only the low-level register access code is generated at the moment for Rust, with svd2rust (and that too seems to be less than optimal, for example a distinct type is generated for all of the GPIOs, even if they are completely the same), but not anything higher-level (the HAL for a few STM32 devices are written mostly by hand for a small range of devices if I understand correctly).

Unfortunately, SVD files don't contain information beyond the memory map and even worse, the SVD files are often incorrect in some minor way. Btw, so are the official CMSIS header files that ST ships. https://github.com/modm-io/cmsis-header-stm32/commit/1ec448711db342deb4d4bf2...

I've noticed this too, and it's pretty annoying.

In short: The data comes out of CubeMX, which contains a bunch of XML files. These are transformed, cross-referenced and filtered to generate these device files. https://github.com/modm-io/modm-devices/tree/develop/devices/stm32

This is awesome, I'll try to use these when I get to implementing something. I assumed that the register data is also extracted from the SVDs (or some other source) and all data merged together, but now I realize that's not needed for C++, since you can just use the header files.

I've held a talk about how we use these device files in modm at EmBO++17: https://www.slideshare.net/emBO_Conference/datadriven-hal-generation

Interesting, thanks! I don't quite understand what the huge matrix of devices with different colors mean though.

This is an insane amount of work, and I'm not willing to put in the hours for this anymore than is absolutely necessary. I fear that if I were to publish this generator and it were to become popular (which the Rust community surrounding @japaric absolutely has the ability to do), I'd be overwhelmed with support requests for all sorts of additional data, regardless of how labor intensive it is to procure.

I hope this will not happen at all / in a way that would make you uncomfortable :)

I've spent the last 4 years looking at this data, if it's not in the above device tree data already, it is _not_ trivial to add. An example is the clock tree data, which exists in the raw data, but it's lacking critical data, which again I'd have to manually add. Ultimately the CubeMX data is incomplete and sometimes just plain wrong.

There's an example in the slides that shows automatic clock configuration (slide 50). Is that manually written for that specific target, or are clock trees available for a subset of the devices?

2. There is a real risk of creating a pseudo-standard of assumptions from my work, since there is no other STM32 data available elsewhere at the moment.

I've been following the Linux Device Tree community for a while now, especially the discussion about the Device Tree Specification: https://github.com/devicetree-org/devicetree-specification [...]

I'm not familiar with this specification, but I can see your point.

A better way would be to formulate your personal experiences as executable assumption checkers, and run them over the raw data. This is actually what I did in modm in the GPIO driver, because the assumptions I made in xpcc regarding GPIO AF were wrong. In fact they were _so_ wrong that our assumption check failed on the majority of STM32 devices and we had to rewrite the API from scratch. https://image.slidesharecdn.com/niklashauser-170228153235/95/datadriven-hal-...

That's a nice approach. Is this fixed in modm?

4. I recently quit my job at ARM working on mbed OS and uVisor because they didn't allow any out-of-the-box thinking on embedded software designs, like we do in xpcc/modm. I mean, they don't even generate their linker-/startup scripts, which is absolutely TRIVIAL to do with even the most basic data. It's insane how much unnecessary manual labor goes into porting/maintaining mbed OS.

My takeaway was that with the entire IoT craze, nobody in this industry takes the time to fix the foundational issues, like HAL design, tool support, library support (don't get me started on Newlib). Most engineers I met can't even point out any flaws in these, they are just so used to it. I'm still downbeat and angry about that and I need time to gain some distance from that experience.

I'm sad to hear that, I hope you'll get over it and find something more enjoyable to work on :)

For now I recommend you use the device files as they are. You're not going to be able to extract any more data from CubeMX without unreasonable effort. I'll keep them updated with every quarterly release of modm. I'll reevaluate my position if more projects start using that data and people are willing to take up some of the maintenance effort like what has happened in xpcc. That really is a joy to maintain, I feel like people care about this stuff :-) Make sure to look at modm if you want to generate code/data/documentation from this data, since it was designed for exactly that purpose.

I'll see if I can come up with anything useful using this for Rust. Unfortunately(?) I just finished working on an embedded project (a robot car for a competition), so I think I'm not gonna work on embedded stuff too much, but let's hope.

About the robot car project, I convinced my teammate to use xpcc (he only ever used ST's HAL before, and only used C), and he really liked how easy it is to do mostly anything in xpcc, and we used some of the drivers too, so it was awesome. (One of the factors for choosing it was that there's a driver for the VL53L0X sensor, which we used in our car). We actually finished 1st place in one of the categories, and xpcc helped a lot in that, so thanks! :)

On Fri, Feb 9, 2018 at 2:03 AM, Niklas Hauser niklas.hauser@rwth-aachen.de wrote:

I've had some discussions with my friends and as a result I've changed my mind: I've published the generator as well as some docs to get you started. https://github.com/modm-io/modm-devices

Awesome, though I think I'll be fine with the generated files as a start.

Thanks, Antal

On Mon, Feb 12, 2018 at 10:54 PM, Niklas Hauser niklas.hauser@rwth-aachen.de wrote:

The advantage of the CMSIS header files over the SVD files is that everyone actually uses them to compile code. The SVD files originally were only for the Keil IDE debugger to assign meaning to IO memory addresses. They therefore aren't required to be completely accurate, which is a problem.

So do you think it would be a good idea to generate Rust register definitions (or whatever these are called) from the headers instead of SVD files?

Yeah, it wasn't my best talk, I'm not practiced enough to handle such a complex topic. I'm currently writing a very long blog post about modm, going into all the details of it.

Looking forward to reading it!

That's based on this proof-of-concept implementation: https://github.com/salkinium/save-the-clocktrees https://github.com/roboterclubaachen/xpcc/pull/39

The CubeMX data does contain real clock graphs, which are used as the data model behind the Clock Configuration tab in CubeMX, however they didn't contain the _full_ clock graph. So I added it manually for the F100 and designed an experimental API. But I didn't find a good API for it in the end.

Ah, I knew about these, I just thought you made a new version that works better.

Antal

I published my blog post on extracting and using this data here, btw: http://blog.salkinium.com/modm-devices/

On 13. Feb 2018, at 02:31, Szabó Antal <szabo.antal.92@gmail.commailto:szabo.antal.92@gmail.com> wrote:

On Mon, Feb 12, 2018 at 10:54 PM, Niklas Hauser <niklas.hauser@rwth-aachen.demailto:niklas.hauser@rwth-aachen.de> wrote: The advantage of the CMSIS header files over the SVD files is that everyone actually uses them to compile code. The SVD files originally were only for the Keil IDE debugger to assign meaning to IO memory addresses. They therefore aren't required to be completely accurate, which is a problem.

So do you think it would be a good idea to generate Rust register definitions (or whatever these are called) from the headers instead of SVD files?

Yeah, it wasn't my best talk, I'm not practiced enough to handle such a complex topic. I'm currently writing a very long blog post about modm, going into all the details of it.

Looking forward to reading it!

That's based on this proof-of-concept implementation: https://github.com/salkinium/save-the-clocktrees https://github.com/roboterclubaachen/xpcc/pull/39

The CubeMX data does contain real clock graphs, which are used as the data model behind the Clock Configuration tab in CubeMX, however they didn't contain the _full_ clock graph. So I added it manually for the F100 and designed an experimental API. But I didn't find a good API for it in the end.

Ah, I knew about these, I just thought you made a new version that works better.

Antal _______________________________________________ xpcc-dev mailing list xpcc-dev@lists.rwth-aachen.demailto:xpcc-dev@lists.rwth-aachen.de https://mailman.rwth-aachen.de/mailman/listinfo/xpcc-dev