Saturday, July 27, 2019

Time Flies Again

Friday, the European Union Aviation Safety Agency, the EU's equivalent of the U.S. Federal Aviation Administration (FAA), issued a revised mandatory Airworthiness Directive for the Airbus A350-941 jet airliner. Quoting from AD 2017-0129R1, the EASA says:
Prompted by in-service events where a loss of communication occurred between some avionics systems and avionics network, analysis has shown that this may occur after 149 hours of continuous aeroplane power-up. Depending on the affected aeroplane systems or equipment, different consequences have been observed and reported by operators, from redundancy loss to complete loss on a specific function hosted on common remote data concentrator and core processing input/output modules. 
This condition, if not corrected, could lead to partial or total loss of some avionics systems or functions, possibly resulting in an unsafe condition. 
To address this potential unsafe condition, Airbus issued the AOT to provide instructions to reset the internal timer. Consequently, EASA issued AD 2017-0129 to require repetitive on ground power cycles (resets).
This means exactly what you think it does: you need to power cycle your A350-941 aircraft no less often than every 149 hours unless and until a software fix is applied.

So: what's special about 149 hours? The AD doesn't say. Let's see if we can figure that out using a little firmware forensics.

Multiply 149 hours by 60 to get 8940 minutes.

Multiply that by 60 to get 536400 seconds.

Multiply that by 1000 to get 536400000 milliseconds.

Finally, multiply that by 4 to get 2145600000.

We could equivalently have multiplied 536400 seconds by one million to get microseconds, then divided it by 250 to get the same value.

So what?

2145600000 is perilously close to 2147483647 or 0x7FFFFFFF in hexadecimal, which is is (231 - 1), the largest positive number represented in two's compliment binary form that you can store in a 32-bit variable.

So I feel pretty confident in making this prediction: somewhere in the A350-941 firmware or software there is a 32-bit signed variable that is incremented every 250 microseconds. After doing so for exactly 149 hours, 7 minutes, and 50.91175 seconds, the very next 250 microsecond clock tick will make that variable overflow and its value will transition from positive to negative as it increments from 0x7FFFFFFF to 0x80000000.

Wackiness ensues.

(An alternative hypothesis is a 32-bit unsigned variable which is incremented twice as often, or every 125 microseconds. It eventually wraps from 0xFFFFFFFF to 0x00000000, similarly confusing an algorithm.)

Long time readers of my blog will recognize that this counter rollover bug is similar to the one I previously described in the Boeing 787 Dreamliner; that bug could result in the loss of all electrical power on the aircraft after 248 days.

Commercial aircraft in service can stay powered up for a long time. They transition from aircraft power to ground power when at the gate, maintaining power so that the interior can be cleaned, the galleys restocked, the air conditioning kept running while boarding, and so forth. Then they go back to on-board power until the flight lands and the cycle repeats.

248 days is a long time. But keeping an aircraft powered up for 149 hours seems not only plausible, but likely. So how was this not uncovered during testing?

Counter rollover is a class of bug that I spent a significant portion of my career ferreting out of firmware and software for telecommunications systems - the kinds of systems expected to run 24x7 with extraordinarily high reliability - during my time at Bell Labs. It kept cropping up in the products I was helping develop (sometimes, admittedly, in my own code), and also in other vendors' products with which we or our customers were integrating our equipment.

Part of systems engineering - the art and craft of discovering, defining, and specifying requirements for a product, and insuring that they are met - must be deciding how long a product or component is expected to run before it must be power cycled. No human artifact is perfect, and nothing runs forever. I dimly remember during my mainframe days that IBM recommended periodically rebooting OS/360 on our 360/65 because of control block degradation; a euphemism, I suspect, for memory corruption or counter overflow.

But 149 hours? That does not seem like a very long time to me.