Monday, June 17, 2019

This Is What You Have To Deal With

I'm passing this anecdote along not because I'm asking for help (although any insight would be welcome), but as an example of what a low-level real-time development looks like.

I spent several days troubleshooting what I still assume to be a bug in my software that deals with output from Global Navigation Satellite System (GNSS) - of which the U.S. Global Positioning System (GPS) is one example - receivers, regarding the synchronization of the NMEA, UBX, and RTCM state machines with the input stream from the U-blox UBX-ZED-F9P receiver (a U-Blox 9 device) on my development system named Nickel (Intel NUC5i7RYH, Ubuntu 18.04.1).

NMEA is the typical output of GNSS receivers, consisting of "sentences" of printable ASCII characters. UBX is a proprietary binary format used for messaging to and from GNSS receivers manufactured by U-blox. RTCM is a standard binary format used for messaging to and from GNSS receivers that support differential GNSS (DGNSS). When a receiver like the F9P generates all three types of output, your software has to be able to grok all three formats and switch between them seamlessly.

I tested the same code on Cadmium (NUC7i7BNH, Ubuntu 16.04.5) and saw the same symptom: every tens of seconds or so my software would loose sync with the input stream than regain it shortly afterwards.

I could not reproduce the problem on Gold, an ARM-based Raspberry Pi (Broadcom BCM2837B0 SoC, Raspbian 9.4) using exactly the same application software and the same receiver hardware, nor on Bodega or Mochila, also ARM-based Pis (also Broadcom BCM2837B0, Raspbian 9.8).

So I ran the standard Linux utility socat on Nickel to save about a minute's worth of NMEA data (UBX and RTCM weren't enabled on the F9P for this test). I ran my software against the file (so: no real-time constraints) and recorded the Sync Lost messages that included the offset into the data where the errors occurred. At this point I was pretty sure I'd found a bug in the state machines in my software.

Then I did a hex dump of the original file generated by socat, checked at those offsets, and lo and behold the NMEA stream was corrupted: periodically a spurious handful of bytes, looking suspiciously like a fragment of an NMEA sentence, was inserted at the end of a valid NMEA sentence.

So this corruption occurs without my software being involved at all.

I ran my software on Nickel using the GlobalSat BU353W10 receiver (a U-Blox 8 device), and I saw similar occasional loss of sync due to corruption of the NMEA stream. As before, the same software and GPS hardware on a Pi worked without problems.

I updated Cadmium to Ubuntu 19.04 and observed the same misbehavior with the BU353W10. I updated the BIOS firmware on Cadmium to the latest version and it still lost sync every tens of seconds.

I built the same software and used the BU353W10 receiver on Mercury (Dell OptiPlex 7040, Ubuntu 14.04.4) and saw sync lost every tens of seconds.

The evidence points an issue either with the USB hardware or firmware on all of my Intel boxes, or with the USB software stacks in many versions of Ubuntu (even though both Ubuntu and Raspbian are based on Debian, making it likely that their USB stacks share a common provenance).

I'm not buying it. Either seem unlikely. What is the likelihood that I'm the only developer in existence to have run into this? Gotta be pretty much zero. And my search-fu has turned up nothing.

I'd like this to be a bug in my software - and I am still assuming it is, perhaps a flaw shared with socat - because then I could fix it. But its root cause remains a mystery.

I use the Intel-based NUCs as my development systems (the one Intel-based Dell box is a legacy development system), and the Pis as test systems, typically deploying to ARM-based systems in the field.

I have no clue at this point what this is. But when I figure it out - and I hope that I do - I will have learned something useful.

Update (2019-06-19)

Just by way of providing actual evidence, here is a snippet extracted from about one minute of data collected via the command
cat /dev/ttyACM0 > FILE
running on Nickel, the Intel NUC5i7RYH running Ubuntu 18.04.1, from a BU353W10 (U-blox 8) USB-dongle GPS receiver:


The 7A\r\n ('7', 'A', carriage return, line feed) inserted between two sentences is obvious. It has the appearance of a checksum and standard end matter of another NMEA sentence (and in fact there are several sentences following this corruption that happen have 0x7A as their checksum).

Keep in mind that these corruptions don't appear when I run the same test on an ARM-based Raspberry Pi. This makes me wonder if there is another process that runs on the NUC but not the Pi that is consuming data from the same port. This isn't as crazy as it sounds; I have inadvertently run redundant instances of my gpstool software or the gpsd GPS daemon and run into similar issues (yes, I already checked for that). However, running the commands
fuser /dev/ttyACM0
lsof | grep ttyACM0
(which displays what process has a particular file open, and all open files and the processes using them, respectively) only show the cat command when I run this test.

An interesting difference between the Ubuntu NUC and the Raspbian Pi is that on the former it isn't necessary to configure /dev/ttyACM0 as a serial port (baud rate, raw mode, etc.) in my application; on the latter, it is (although it doesn't seem to matter to what I set the baud rate).