User Space GPIO Interrupt programming - GSoC-2018

[Dr. Rolf Jansen]

Am 27.11.2020 um 14:32 schrieb Ian Lepore <ian at freebsd.org>:

> On Thu, 2020-11-26 at 22:18 -0300, Dr. Rolf Jansen wrote:
>>> Am 26.11.2020 um 16:56 schrieb Ian Lepore <ian at freebsd.org>:
>>> 
>>> On Tue, 2020-11-24 at 17:14 -0300, Dr. Rolf Jansen wrote:
>>>> Hello
>>>> 
>>>> Has anything of the GSoC-2018 efforts made it into the current
>>>> code
>>>> base?
>>>> 
>>>> 
>>> 
>>> 
> https://wiki.freebsd.org/SummerOfCode2018Projects/UserSpaceGPIOinterrupts
>>>> 
>>>> I installed the recent 13.0-CURRENT snapshot (2020-11-19) on a
>>>> BeagleBone Black which was one of the implementation targets of
>>>> said
>>>> project, but when running the test tools, I either see cannot
>>>> read/kevent/poll/aio_read - Operation not supported by device or
>>>> Inappropriate ioctl for device.
>>>> 
>>>> Perhaps I need to pull the project´s changes into the kernel by
>>>> myself. However, before this I would like to ask whether it is
>>>> worth
>>>> the effort.
>>>> 
>>>> Please, can anyone shed some light on this.
>>>> 
>>>> Best regards
>>>> 
>>>> Rolf
>>>> 
>>> 
>>> I made some time this morning to review the gsoc2018 code.  It
>>> turns
>>> out this code is very high quality, nearly ready to commit as-
>>> is.  The
>>> main thing it needs is some style cleanup in its comment blocks,
>>> and
>>> documentation.  I'd be inclined to commit the code first and write
>>> the
>>> documentation over the next little while and commit it separately.
>>> 
>>> If you'd like to give it a try, here's a diff that should apply and
>>> build cleanly on freebsd 12 or 13:
>>> 
>>> https://people.freebsd.org/~ian/gpio_gsoc2018.diff
>>> 
>>> While there isn't any documentation yet, there is a test program (I
>>> haven't run it yet) that demonstrates all the features:
>>> 
>>> 
> https://github.com/ckraemer/gsoc2018-utils/blob/master/src/gpioc_intr_test.c
>>> 
>>> Right now the code will let you block waiting for a pin-change
>>> event
>>> using select(), poll() or kevents, or to be notified via SIGIO, but
>>> after being notified that something happened, you still have to
>>> call
>>> read() to see which pin changed.  I think if the pin changes state
>>> multiple times between calls to read(), you'll lose track of some
>>> changes (I'm not positive of that, I don't understand the kevent
>>> stuff
>>> well).
>>> 
>>> I'd like to add some features so that you can configure it to track
>>> pin
>>> changes in counting-mode and timestamping-mode.  In counting mode,
>>> when
>>> you do a read() you would get back a pair of values, the pin number
>>> and
>>> how many times its interrupt fired since the last read.  In
>>> timestamping mode, every read would return a pin number and an
>>> associated timespec giving the exact time the interrupt happened
>>> (there
>>> would need to be a way to configure how many events it could
>>> buffer,
>>> but I think even allowing a thousand buffered events would only use
>>> a
>>> few kbytes of memory).
>> 
>> I got it working as well, please see my other post from yesterday. I
>> used gpioc_intr_test.c.
>> 
>> I see hundreds of warning messages when I press the test button a few
>> times. May these warnings be safely ignored. The kernel module of
>> Oskar Holmund works quite nice as well (for what I need), and with
>> that one, I don’t see warnings.
>> 
>> The counting- and timestamping-mode for sure would be very useful.
>> Perhaps by implementing this, there won’t be no unhandled interrupts
>> anymore, and hence there won’t be any warnings either.
>> 
>> Best regards
>> 
>> Rolf
> 
> I'm sorry, I somehow overlooked your previous message about using
> gpioc_intr_test.c.
> 
> Those warning messages are definitely not a good thing, in some changes
> I've made to the original patches they are changed to debugging-only
> output that won't normally show up.  Printing messages from within an
> interrupt handler is pretty much always a bad idea. :)
> 
> I was thinking about the various interrupt options and realized we
> cannot support level-triggered interrupts in this code without a ton of
> work.  It's just a recipe for an interrupt storm (which will happen at
> the rate of tens of thousands of interrupts per second once the printfs
> aren't there to slow things down to console IO speed).  To use level-
> triggered interrupts from userland, the gpioc would have to mask the
> interrupt from within the interrupt handler (we don't have an internal
> API for doing that with gpio interrupts right now), and provide some
> sort of EOI (end-of-interrupt) acknowledgement interface to userland to
> unmask them once it had done something to make the interrupt stop
> asserting.  Basically this would extend the way device drivers handle
> level-triggered interrupts into userland, and I just don't see much
> value in all the work that would be involved to make that happen.
> 
> I think instead we should just run in counting-mode by default, and add
> new code similar to what Vladimir Goncharov has proposed to handle
> detailed reporting of each event if the app requests it.
> 
> I can't make up my mind on the issue of debouncing.  My gut tells me
> that building in some kind of debounce logic as a per-pin configurable
> option would be nice, but it might also get really complicated.  An app
> could handle debouncing for itself if it requested detailed event
> reporting, because the timestamps on the events could help it decide
> what to do.

After working a bit with the various "User Space GPIO Interrupt" incarnations in the last few days, I found that I don't need level interrupts. The hardware colleague of our present project told me that I do not need to bother with button debouncing in software either. There are IC's which manage this, for example: MAX6818 - https://datasheets.maximintegrated.com/en/ds/1896.pdf. So, the GPIO would see only perfect square wave signals when a button has been pressed.

I still think that counting-mode and timestamp-mode could be very useful.

Regarding the memory allocation, I am with you that it should be done once and not for any interrupt trail. As a matter of fact I did kind of this already for a kernel module for FreeBSD-x86-64 targets which I wrote for National Instruments M- and X-Series PCI and PCIe DAQ boards. Here I allocate huge blocks of DMA memory (256 MB for the AI channels and 48 MB for the AO channels) in the course of xxx_attach() and this can be mapped into user space by mmap(). I added sysctls which allow to resize said memory blocks, and ioctls for clearing and synching the buffers. I do all the math on the measurement data in user space by directly accessing the DMA blocks in the course of the measurements.

Thinking about it, wouldn't it be nice to mmap() the interrupt/counter/timestamp table into user space? Programs could read the data in the table directly. It could be a round robin table, perhaps of 100 kB to 1 MB. Perhaps my code in my kernel module for DMA buffer management, which could be easily switched to any-kind-of-memory management, does not exactly meet the high quality and style of FreeBSD code, and I am even hesitant to offer it. However, if you are curious, please send me a note.

Best regards

Rolf