Slashdot Mirror
Writing Code for Spacecraft
CowboyRobot writes "In an article subtitled, "And you think *your* operating system needs to be reliable."
Queue has an interview with the developer of the OS that runs on the Mars Rovers. Mike Deliman, chief engineer of operating systems at Wind River Systems, has quotes like, 'Writing the code for spacecraft is no harder than for any other realtime life- or mission-critical application. The thing that is hard is debugging a problem from another planet.' and, 'The operating system and kernel fit in less than 2 megabytes; the rest of the code, plus data space, eventually exceeded 30 megabytes.'"
while (1 = 1) { Dig(); Picture(); }
The interviewer George Neville-Neil co-authored "The Design and Implementation of the FreeBSD Operating System" with Marshall Kirk McKusick.
cpghost at Cordula's Web.
Should have just used WinCE, with a few of the productivity apps cut out. Adding a copy of pocket Auto-route, with some Martian JPEGS would have helped navigation as well.
Yes, here is an email my OS prof sent our class on the subject:
Subject: What really happened on Mars Rover Pathfinder
The Mars Pathfinder mission was widely proclaimed as "flawless" in the early
days after its July 4th, 1997 landing on the Martian surface. Successes
included its unconventional "landing" -- bouncing onto the Martian surface
surrounded by airbags, deploying the Sojourner rover, and gathering and
transmitting voluminous data back to Earth, including the panoramic pictures
that were such a hit on the Web. But a few days into the mission, not long
after Pathfinder started gathering meteorological data, the spacecraft began
experiencing total system resets, each resulting in losses of data. The
press reported these failures in terms such as "software glitches" and "the
computer was trying to do too many things at once".
This week at the IEEE Real-Time Systems Symposium I heard a fascinating
keynote address by David Wilner, Chief Technical Officer of Wind River
Systems. Wind River makes VxWorks, the real-time embedded systems kernel
that was used in the Mars Pathfinder mission. In his talk, he explained in
detail the actual software problems that caused the total system resets of
the Pathfinder spacecraft, how they were diagnosed, and how they were
solved. I wanted to share his story with each of you.
VxWorks provides preemptive priority scheduling of threads. Tasks on the
Pathfinder spacecraft were executed as threads with priorities that were
assigned in the usual manner reflecting the relative urgency of these tasks.
Pathfinder contained an "information bus", which you can think of as a
shared memory area used for passing information between different components
of the spacecraft. A bus management task ran frequently with high priority
to move certain kinds of data in and out of the information bus. Access to
the bus was synchronized with mutual exclusion locks (mutexes).
The meteorological data gathering task ran as an infrequent, low priority
thread, and used the information bus to publish its data. When publishing
its data, it would acquire a mutex, do writes to the bus, and release the
mutex. If an interrupt caused the information bus thread to be scheduled
while this mutex was held, and if the information bus thread then attempted
to acquire this same mutex in order to retrieve published data, this would
cause it to block on the mutex, waiting until the meteorological thread
released the mutex before it could continue. The spacecraft also contained
a communications task that ran with medium priority.
Most of the time this combination worked fine. However, very infrequently
it was possible for an interrupt to occur that caused the (medium priority)
communications task to be scheduled during the short interval while the
(high priority) information bus thread was blocked waiting for the (low
priority) meteorological data thread. In this case, the long-running
communications task, having higher priority than the meteorological task,
would prevent it from running, consequently preventing the blocked
information bus task from running. After some time had passed, a watchdog
timer would go off, notice that the data bus task had not been executed for
some time, conclude that something had gone drastically wrong, and initiate
a total system reset.
This scenario is a classic case of priority inversion.
HOW WAS THIS DEBUGGED?
VxWorks can be run in a mode where it records a total trace of all
interesting system events, including context switches, uses of
synchronization objects, and interrupts. After the failure, JPL engineers
spent hours and hours running the system on the exact spacecraft replica in
their lab with tracing turned on, attempting to replicate the precise
conditions under which they believed that the reset occurred. Early in the
morning, after all but one engineer had gone
Remember sometime ago Spirit was continously rebooting due to a flash memory problem. The usage of FAT file system in the embedded systems was partly responsible for the mess.
The problem, Denise said, was in the file system the rover used. In DOS, a directory structure is actually stored as a file. As that directory tree grows, the directory file grows, as well. The Achilles' heel, Denise said, was that deleting files from the directory tree does not reduce the size of the directory file. Instead, deleted files are represented within the directory by special characters, which tell the OS that the files can be replaced with new data.
By itself, the cancerous file might not have been an issue. Combined with a "feature" of a third-party piece of software used by the onboard Wind River embedded OS, however, the glitch proved nearly fatal.
According to Denise, the Spirit rover contains 256 Mbytes of flash memory, a nonvolatile memory that can be written and rewritten thousands of times. The rover also contains 128 Mbytes of DRAM, 96 Mbytes of which are used for data, such as buffering image files in preparation for transmitting them to Earth. The other 32 Mbytes are used for code storage. An additional 11 Mbytes of EEPROM memory are used for additional program code storage.
The undisclosed software vendor required that data stored in flash memory be mirrored in RAM. Since the rover's flash memory was twice the size of the system RAM, a crash was almost inevitable, Denise said.
Moving an actuator, for example, generates a large number of tiny data files. After the rover rebooted, the OSes heap memory would be a hair's breadth away from a crash, as the system RAM would be nearly full, Denise said. Adding another data file would generate a memory allocation command to a nonexistent memory address, prompting a fatal error.
Source: DOS Glitch Nearly Killed Mars Rover
BTW, there is another interview of Mike Deliman I read sometime ago in PCWorld.
I used to write embedded applications using OS-9 (NOT MacOS 9) on 68000-based systems as a sub-contractor for Nuclear Electric (nuclear power stations company in the UK before it became BNFL). Our development system - complete with OS/Kernel and compilers - had only about a meg of memory; the final embeded systems often only had 512K if we were lucky
Okay, so this was some 14 years ago - but it was doing a lot of work. 2 megabytes is a lot of memory! There's a phenomenal amount of code and data that can be stored in 2 meg. Maybe it's good by current standards, but - personally - I would suggest that current standards is a bad place to start from.
The ways of gods are mysteriously indistinguishable from chance.
Not gonna happen, for one big reason. I could just see the Slashdot headline:
Mars Rover HaX0r3d and OS replaced with Linux.
Shortly thereafter, Micro$oft claims that they can enforce patent infringement on Mars...
For those who are wondering, JPL is very aware of the shortcomings of VxWorks and has seriously considered other alternatives for every mission. Keep in mind that the choice of OS has to be made years before launch, so at the time the OS for the 2004 Mars Rovers was decided on, many options that are possibilities today were not contenders. Also keep in mind that in spite of many shortcomings, VxWorks is a known quantity. JPL has been working with it for years and had a lot of in-house expertise with it.
There are a few groups at JPL that have been actively experimenting with other options, including RTLinux and a few different variants of hard-real-time Java (basically Java with explicit memory management and no garbage collection).
you are in a red rocky landscape..
GO NORTH..
you are in a red rocky landscape..
DIG.
ok. you see some red sand.
it is getting dark.
GO NORTH..
you were eaten by a grue.
"You lied to me! There is a Swansea!"
Dynamically allocating memory is usually a big no-no in real time systems.
I worked on a satellite mission where we had some trouble. Due to an error the satellite wound up pointing 16 degrees away from the sun in a higher-than-expected orbit of 443 miles (714 kilometers) above Earth.
The misalignment meant the spacecraft was unable to look directly at the sun's center to record the amount of radiation streaming toward Earth. To accurately measure sunlight, the darn thing needed to be pointed to within a quarter of a degree of dead center.
It took about four and a half months to fix that problem, due to uplink difficulties. Ground controllers from first had to slow the spacecraft's spin in order to transmit a series of software "patches" and then gradually speed it up to see how well the commands worked.
Then things were fixed.
Moral of the story: it is a tough job indeed!
> "The operating system and kernel fit in less than 2
> megabytes; the rest of the code, plus data space,
> eventually exceeded 30 megabytes." This should be used as
> the example for efficient coding
You've GOT to be kidding, right? 2 meg of OS code? That's ULTRABLOAT compared to most spacecraft. In fact, for the vast majority of the space age, that would have exceeded the resources of the computer by several orders of magnitude.
I've done this kind of programming for a living (for 10 years, moved up to controls design) but the last system I programmed for has 372k of memory, total. That includes data, code, OS, everything. Runs at 432 KIPS. And it performs what it probably one of the most complex in-flight autonomous control operations ever.
Most are even more restrictive. For example, 8K of PROM and 1k of volatile memory (and 28 WORDS) of non-volatile memory. This more than adequate for most applications, if you do it right.
Many spacecraft OS's are more akin to this:
hardware interrupt
external electronics power up processor.
external electronics set PC = 80hex
run
{execute all the code}
halt
power down
Once every 1/4 a second for 15 years.
The project I am currently working on uses VxWorks (and so we were quite interested in the Mars Rover problem) and it's so bloated with unnecessary features it's absurd. This is not a Windows box, it's a spacecraft processor.
I can't argue with the 30 meg of data space. Using the memory as a data recorder would be quite useful and a good picture takes a lot of space. But it's alarming to me that you could figure out how to waste maybe 4-5 meg on code. If you started with a bare home-brew OS, I would guess (and I get paid for this sort of guess) that you could do the entire flight code in 512K, with maybe 8k of data space, excluding the science data.
Only recently have space-qualified rad-hard processors with this kind of capability become available. Until then, if you said you needed 2 meg for the OS alone, you would have gotten fired on the sopt and referred to mental health professionals. The availability of these processors enabled people to use high-level languages with tremendous overhead (like C++) to be used. And this was only done for employee retention purposes during the bubble. For years it was done at the assembler or even machine level. It's still not at all uncommon to do, and we've done MANY flight code patches, with only a processor handbook, an engineering paper pad, and by setting individual bits one-by-one.
Brett
If that was open source, there are so many space nerds who are programmers that flaws of that magnitude would never get by the army of testers.
Many would help out simply because hey it's the *space program* and that's good enough for them. Other would want their name listed next to some obscure bug fix on a NASA site; it's good for the ego or your CV.
Simply put, even a binary distribution of that code would allow unlimited free testing for crashes. Why wouldn't NASA do it?
Because there are still people in washington that think code mysteriously get damaged by being public - even if such code isn't modifiable by the public who reads it.
This is evidence of advanced cluelessness in Washington and maybe independant anti-free-source advocates (spelled M-i-c-r-o-s-o-f-t) are at cause.
But I've learned not to bash. Never explain by Microsoft malice what could be explained by stupidity. Such as using DOS on a space thing...
Microsoft is pure dog-ma. FreeBSD is pure cat-ma.
Perhaps not surprisingly for anyone who has heard about the management at NASA, C++ was selected for the successors to the Remote Agent on the grounds that it is supposed to be more reliable (this despite the fact that the Remote Agent was originally to be developed in C++, an effort that was abandoned after a year of failure). This caused more than a few people to be upset (including a very personal account by one of the aforementioned designers). Clearly the debugging facilities of Common Lisp are far superior to static systems like C++, something which is very useful in diagnosing unexpected error conditions in spacecraft software (read the first question on p. 3 of the interview to see what pains the JPL staff went through to adapt similar, ad-hoc methods to VxWorks). It's also clear from this interview (question: "How is application programming done for a spacecraft?" Answer:"Much the same as for anything elsesoftware requirements are written, with specifications and test plans, then the software is written and tested, problems are fixed, and eventually its sent off to do its job.") that NASA has in no way tried to adapt formal verification methods for it's software, prefering instead to rely on the "tried and true" (at failing, maybe) poke-and-test development "methods."
Clearly, formal verification methods to eliminate bugs before critical software is deployed, and deployment in a system with advanced debugging facilities is a clear win for spacecraft software, and should be adapted as the standard model of development. Unfortunately, like in many other software development enterprises, inertia keeps outdated, inadequate systems going despite a strong failure correlation rate.
In the great CONS chain of life, you can either be the CAR or be in the CDR.
As someone who's worked with VxWorks for the last several years I'm surprised they didn't turn on priority inheritance to begin with for the semaphore. As a rule, we usually turn on priority inheritance for our mutex semaphores.
Other problems in the Mars Pathfinder were related to using the VxWorks filesystem. VxWorks basically only supports FAT on top of flash. For flash, FAT is a poor choice since some areas of the disk like the root directory and FAT tables will quickly wear out. Also, I don't think VxWorks has much support for working around bad sections of flash.
As far as VxWorks memory allocation support, in an ideal world one would statically allocate all memory, but oftentimes things are not ideal. In the product I work on, we have to have dynamic memory allocation, since depending on how the product is being used at the time, different data structures are required with no way of knowing beforehand how many of a particular type are needed, and this changes dynamically. For a simple device, it's easy to statically allocate everything, or if you have enough memory where you can statically allocate everything.
In our case, while we statically allocate memory where we can, however, in many cases we cannot. For example, I have to maintain a data structure keeping track of all of the network gateways connected to an output interface. We can have many thousands of gateways and thousands of output interfaces. There could be anything between one and thousands of gateways on an interface. In this case, I use static arrays for information on each gateway and each output interface, but must use dynamic data structures to list all the gateways connected to an output interface. It would be prohibitive to allocate storage for 30,000 gateways with 30,000 interfaces! I also can't use a linked list of gateways per interface since it doesn't scale, a linked list having access time O(n).
Also, we use third party libraries that perform dynamic memory allocation and it would be prohibitive to change that.
By replacing Wind River's malloc code with Doug Lea's code we eliminated fragmentation problems and saw our startup time jump from 50 minutes to 3 minutes. Doug Lea's malloc code is the basis of malloc in glibc and is very effecient. We also added support for tracing heap memory allocations to keep track of which task allocated a block and where it was allocated. This alone helped tremendously in tracking down a number of memory leaks since we can just walk the heap and see exactly where all the memory is being allocated. This is a sorely missing feature in VxWorks.
The lack of memory protection is another major problem for complex tasks. We have a bug we've spent weeks trying to track down the cause without any luck where random memory locations get corrupted.
Needless to say, all new projects where I work will not run on VxWorks. All of the chip vendors we're looking at are either dropping support for it or have already dropped it and are focusing on Linux.
BTW, this is one feature I would *REALLY* love to see added to Linux. The company I'm working for is looking at writing our next generation platform on top of an embedded Linux. We have not yet decided which one to use, but want something 2.6 based.
With priority inheritance, if a mutex is held by a low priority task and a high priority task tries to grab it, the low priority task is automatically boosted to the highest priority task that has attempted to acquire the semaphore. When the semaphore is released, the low priority task's priority is restored.
Some other nice features are interrupt scheduling and better priority based message passing support (which may already be present, I'm still looking into this).
Finally, one very useful feature would be the ability to guarantee a real-time thread a certain percentage of the CPU, with the option of placing a hard limit if it tries to exceed that or temporarily lowering it's priority to non-realtime so as to not starve no
This post is encrypted twice with ROT-13. Documenting or attempting to crack this encryption is illegal.
The problem is that technology moves too quickly for it to get "NASA certified". When you send something up in space where making changes to it will be difficult, you need something that is known to be robust and reliable, that has several years of testing.
Last I read (maybe a year ago?), NASA still used 386 and 486 chips because they didn't generate a lot of heat (compared to todays machines) and could be made to withstand higher than normal forces (through extra padding on the device I imagine). They were more resiliant to the issues you might see in space than newer processors.
Simply put, if they put the latest CPU with tons of RAM in there, and it fails, how are they going to fix it?
-- Joe
Exactly!
The problem is that most /.ers are used to thinking of an OS as something that needs to run any arbitrary program under any arbitrary conditions and survive any arbitrary crash in those programs.
For a Rover, none of those are true. They know exactly what code is going to be run. They know exactly where it's going to sit in memory. And they test it. (This is the part that /.ers can't quite understand.) They test these programs far more rigorously than any bog-standard x86 Linux OSS program ever gets tested. Those programs have their problems, but they will be mistakes in logic (metric/imperial conversions, or thread priority inversions), not segfaults because of derefing a null pointer.
I wonder how many undergrand CS degree programs still teach correctness proofs? Not "yeah, I ran it lots of times and it didn't crash," but "I ran it 100,000 times with 100,000 different inputs, all random, and it didn't crash, but while it was running I also sat down and mathematically proved the code is correct."
Embedded programming is just plain different than "normal" progrmming. It's usually a mistake to try to generalize from one to the other.
(All that said, the next version of VxWorks is advertised to optionally support a "traditional Unix" process model, and I think protected memory boundaries are one of the features. In case your embedded app needs to run arbitrary third-party software which probably doesn't get stress-tested at JPL :-), you can turn all that stuff on and live with the overhead.)
You cannot apply a technological solution to a sociological problem. (Edwards' Law)
About five years ago, I worked for a major test equipment manufacturer who was contracted to deliver a test system for POTS lines (which could eventually do ADSL prequalification) to a national telco in a major European country. The idea was to test every POTS line in the system (millions of them) every night to detect early signs of degradation so repair crews could be dispatched before dialtone was completely lost.
As you can imagine, this involved a distributed system of test heads in each central office, networked back to a central command and control site. The sysem worked well, but had one flaw: downloading new firmware to the test heads was fraught with problems, and often led to the test head "locking up", even though a backup copy of firmware was always present, along with a hardware watchdog timer (though it was possible to lock out the watchdog interrupt, particularly when reprogramming flash, so it was a less than perfect watchdog). In these situations, one had to dispatch a "truck roll" to the affected central office, and replace EPROMs by hand.
Needless to say, the customer was pissed. More worrying was that even if we fixed the software download problem (which we were unable to reproduce in the lab), was that we'd be paying for truck rolls all over the country. This was a not insignificant amount of money.
Management frittered away time, instead of authorizing a root cause analysis, by requesting tweaks to TCP/IP operating parameters, and testing to see if the problem was getting better or worse. This did not prove illuminating, time was wasted, and the customer was getting royally angry.
Finally, a small team of us were permitted to undertake a root cause analysis to find and fix the problem: the engineer responsible for the embedded flash file system, the telecom engineer on the control side, and I: responsible for the embedded O/S, and TCP/IP stack (inherited from the supplier of the embedded O/S). We wanted a month. We got two weeks. Remember, deploying experimental software to live COs requires so many layers of approval, it isn't funny, and we were worried that would be our biggest bottleneck.
Finally, the controller telecom engineer was able to reproduce the problem, by attempting to download software from our controllers to deployed equipment in a single central office (getting permission was a feat in itself -- while there was little danger of affecting telephone service, this was a live CO).
The problem was clear: the data network was slow (9600 b/s over an X.25 PVC, carrying PPP-encapsulated TCP/IP), resulting in the use of large MTUs to minimize packetizing overhead (latency wasn't an issue - throughput was). Because of the way the controller's TCP/IP stack worked, it misestimated the packet/ack round trip time: it used a one byte payload for the first packet, and full MTUs after that. The resulting packet ACK timeout and retransmissions exposed an inconsistency between controller and embedded TCP/IP stacks that caused the embedded system to lock up.
Great. Now, how to fix it?
The fix wasn't a big deal (I implemented a fix in the embedded TCP/IP code since we didn't have source to the controller TCP/IP stack), but deploying it was: remember we couldn't download the code sucessfully, and we didn't want to pay for a truck roll.
At this point, I proposed something daring: download a small patch, in as few packets as possible (we could send three full MTUs safely). which would patch the existing code in place, which would be good enough to reliably download a complete replacement.
The thought of "self-modifying code" freaked management out to no end: it went against every rule in the book. But all three of us stood our ground: the only other alternative was a truck roll to each central office in the country. Reluctantly, we were allowed to proceed with that fix.
At this point, we had about ten days left. I had managed to get approval to pipeline the dev and tes
You could've hired me.
Shielding does not protect against single-event upsets (particle-induced bit flips), it only provides some mitigation against total ionizing dose (which causes long term cumulative degradation as a result of drift in transistor operating parameters). There are design techniques and fabrication processes that can reduce the likelihood that a circuit will suffer upsets, but it's still standard practice to provide either redundant memory, or error detection and correction coding. In the case of MER they had 3 physically separate PROMs carrying identical copies of the flight software, and the RAM was (IIRC) protected by an EDAC code implemented in a rad-hard FPGA.
Feature size. The smaller the feature (think gate level), the higher the chance it will be ruined by random radiation exposure. And that's the one-sentence summary of the "Radiation Effects on Microelectronics" class I took about 7 years ago.
Smaller memory capacity for a given surface area implies larger feature size.
By the way, the class I took was 1-on-1 with Prof. Stephen McGuire at Cornell. Extremely cool guy.