问题
I'm researching the development of safety-critical software, and in particular what effects the choice of programming language has on such development.
Please explain, in detail, which languages are commonly used, and why.
回答1:
Ada and SPARK (which is an Ada dialect with some hooks for static verification) are used in aerospace circles for building high reliability software such as avionics systems. There is something of an ecosystem of code verification tooling for these languages, although this technology also exists for more mainstream languages as well.
Erlang was designed from the ground up for writing high-reliability telecommunications code. It is designed to facilitate separation of concerns for error recovery (i.e. the subsystem generating the error is different from the subsystem that handles the error). It can also be subjected to formal proofs although this capability hasn't really moved far out of research circles.
Functional languages such as Haskell can be subjected to formal proofs by automated systems due to the declarative nature of the language. This allows code with side effects to be contained in monadic functions. For a formal correctness proof the rest of the code can be assumed to do nothing but what is specified.
However, these languages are garbage collected and the garbage collection is transparent to the code, so it cannot be reasoned about in this manner. Garbage collected languages are not normally predictable enough for hard realtime applications, although there is a body of ongoing research in time bounded incremental garbage collectors.
Eiffel and its descendants have built-in support for a technique called Design By Contract which provides a robust runtime mechanism for incorporating pre- and post- checks for invariants. While Eiffel never really caught on, developing high-reliability software tends to consist of writing checks and handlers for failure modes up-front before actually writing the functionality.
Although C and C++ were not specifically designed for this type of application, they are widely used for embedded and safety-critical software for several reasons. The main properties of note are control over memory management (which allows you to avoid having to garbage collect, for example), simple, well debugged core run-time libraries and mature tool support. A lot of the embedded development tool chains in use today were first developed in the 1980s and 1990s when this was current technology and come from the Unix culture that was prevalent at that time, so these tools remain popular for this sort of work.
While manual memory management code must be carefully checked to avoid errors, it allows a degree of control over application response times that is not available with languages that depend on garbage collection. The core run time libraries of C and C++ languages are relatively simple, mature and well understood, so they are amongst the most stable platforms available. Most if not all of the static analysis tools used for Ada also support C and C++, and there are many other such tools available for C. There are also several widely used C/C++ based tool chains; most tool chains used for Ada also come in versions that support C and/or C++.
Formal Methods such as Axiomatic Semantics (PDF), Z Notation or Communicating Sequential Processes allow program logic to be mathematically verified, and are often used in the design of safety critical software where the application is simple enough to apply them (typically embedded control systems). For example, one of my former lecturers did a formal correctness proof of a signaling system for the German railway network.
The main shortcoming of formal methods is their tendency to grow exponentially in complexity with respect to the underlying system being proved. This means that there is significant risk of errors in the proof, so they are practically limited to fairly simple applications. Formal methods are quite widely used for verifying hardware correctness as hardware bugs are very expensive to fix, particularly on mass-market products. Since the Pentium FDIV bug, formal methods have gained quite a lot of attention, and have been used to verify the correctness of the FPU on all Intel processors since the Pentium Pro.
Many other languages have been used to develop highly reliable software. A lot of research has been done on the subject. One can reasonably argue that methodology is more important than the platform although there are principles like simplicity and selection and control of dependencies that might preclude the use of certain platforms.
As various of the others have noted, certain O/S platforms have features to promote reliability and predictable behaviour, such as watchdog timers and guaranteed interrupt response times. Simplicity is also a strong driver of reliability, and many RT systems are deliberately kept very simple and compact. QNX (the only such O/S that I am familiar with, having once worked with a concrete batching system based on it) is very small, and will fit on a single floppy. For similar reasons, the people who make OpenBSD - which is known for its robust security and thorough code auditing - also go out of their way keep the system simple.
EDIT: This posting has some links to good articles about safety critical software, in particular Here and Here. Props to S.Lott and Adam Davis for the source. The story of the THERAC-25 is a bit of a classic work in this field.
回答2:
For C++, the Joint Strike Fighter (F-35) C++ Coding Standard is a good read:
http://www.stroustrup.com/JSF-AV-rules.pdf
回答3:
I believe Ada is still in use in some government projects that are safety and/or mission critical. I've never used the language, but it's on my list of "to learn", along with Eiffel. Eiffel offers Design By Contract, which is supposed to improve reliability and safety.
回答4:
Firstly, safety critical software adheres to the same principals that you would see in the classic mechanical and electrical engineering fields. Redundancy, fault tolerance and fail-safety.
As an aside, and as the previous poster alluded to (and was for some reason down-voted), the single most important factor in being able to achieve this is for your team to have a rock solid understanding of everything that is going on. It goes without saying that good software design on your part is key. But it also implies a language that is accessible, mature, well supported, for which there is a lot of communal knowledge and experienced developers available.
Many posters have already commented that the OS is a key element in this respect which is very true most because it must be deterministic (see QNX or VxWorks). This rules out most interpreted languages that do things behind the scenes for you.
ADA is a possibility but there is less tools and support out there, and more importantly the stellar people aren't as readily available.
C++ is a possibility but only if you strictly enforce a subset. In this respect it is devil's tool, promising to make our life easier, but often doing too much,
C is ideal. It is very mature, fast, has a diverse set of tools and support, many experienced developers out there, cross-platform, and extremely flexible, can work close to the hardware.
I've developed in everything from smalltalk to ruby and appreciate and enjoy everything that higher languages have to offer. But when I'm doing critical systems development I bite the bullet and stick with C. In my experience (defence and many class II and III medical devices) less is more.
回答5:
I'd pick up haskell if it's safety over everything else. I propose haskell because it has very rigid static type checking and it promotes programming where you build parts in a such manner that they are very easy to test.
But then I wouldn't care about language much. You can get much greater safety without compromising as much by having your project overall in condition and working without deadlines. Overall as in having all the basic project management in place. I'd perhaps concentrate on extensive testing to ensure everything works like it ought, tests that cover all the corner cases + more.
回答6:
The language and OS is important, but so is the design. Try to keep it bare-bones, drop-dead simple.
I would start by having the bare minimum of state information (run-time data), to minimize the chance of it getting inconsistent. Then, if you want to have redundancy for the purpose of fault-tolerance, make sure you have foolproof ways to recover from the data getting inconsistent. Redundancy without a way to recover from inconsistency is just asking for trouble.
Always have a fallback for when requested actions don't complete in a reasonable time. As they say in air traffic control, an unacknowledged clearance is no clearance.
Don't be afraid of polling methods. They are simple and reliable, even if they may waste a few cycles. Shy away from processing that relies solely on events or notifications, because they can be easily dropped or duplicated or misordered. As an adjunct to polling, they are fine.
A friend of mine on the APOLLO project once remarked that he knew they were getting serious when they decided to rely on polling, rather than events, even though the computer was horrendously slow.
P.S. I just read through the C++ Air Vehicle standards. They are OK, but they seem to assume that there will be lots of classes, data, pointers, and dynamic memory allocation. That is exactly what there should no more of than absolutely necessary. There should be a data structure czar with a big scythe.
回答7:
The OS is more important then the language. Use a real time kernel such as VxWorks or QNX. We looked at both for controlling industrial robots and decided to go with VxWorks. We use C for the actual robot control.
For truly critical software, such as aircraft autoland systems, you want multiple processors running independently to cross check results.
回答8:
Real-time environments usually have "safety-critical" requirements. For that sort of thing, you could look at VxWorks, a popular real-time operating system. It's currently in use in many diverse arenas such as Boeing aircraft, BMW iDrive internals, RAID controllers, and various space craft. (Check it out.)
Development for the VxWorks platform can be done with several tools, among them Eclipse, Workbench, SCORE, and others. C, C++, Ada, and Fortran (yes, Fortran) are supported, as well as some others.
回答9:
Since you don't give a platform, I would have to say C/C++. On most real-time platforms, you're relatively limited in options anyway.
The drawbacks of C's tendency to let you shoot yourself in the foot is offset by the number of tools to validate the code and the stability and direct mapping of the code to the hardware capabilities of the platform. Also, for anything critical, you will be unable to rely on third-party software which has not been extensively reviewed - this include most libraries - even many of those provided by hardware vendors.
Since everything will be your responsibility, you want a stable compiler, predictable behavior and a close mapping to the hardware.
回答10:
Here's a few updates for some tools that I had not seen discussed yet that I've been playing with lately which are fairly good.
The LLVM Compiler Infrastructure, a short blurb on their main page (includes front-ends for C and C++. Front-ends for Java, Scheme, and other languages are in development);
A compiler infrastructure - LLVM is also a collection of source code that implements the language and compilation strategy. The primary components of the LLVM infrastructure are a GCC-based C & C++ front-end, a link-time optimization framework with a growing set of global and interprocedural analyses and transformations, static back-ends for the X86, X86-64, PowerPC 32/64, ARM, Thumb, IA-64, Alpha, SPARC, MIPS and CellSPU architectures, a back-end which emits portable C code, and a Just-In-Time compiler for X86, X86-64, PowerPC 32/64 processors, and an emitter for MSIL.
VCC;
VCC is a tool that proves correctness of annotated concurrent C programs or finds problems in them. VCC extends C with design by contract features, like pre- and postcondition as well as type invariants. Annotated programs are translated to logical formulas using the Boogie tool, which passes them to an automated SMT solver Z3 to check their validity.
VCC uses the recently released Common Compiler Infrastructure.
Both of these tools, LLVM or VCC are designed to support multiple languages and architectures, I do think that their is a rise in coding by contract and other formal verification practices.
WPF (not the MS framework :), is a good place to start if you're trying to evaluate some of the recent research and tools in the program validation space.
WG23 is the primary resource however for fairly current and specific critical systems development language details. They cover everything from Ada, C, C++, Java, C#, Scripting, etc... and have at the very least a decent set of reference and guidance for direction to update information on language specific flaws and vulnerabilities.
回答11:
There are a lot of good references at http://www.dwheeler.com ("high-assurance software").
For automotive stuff, see the MISRA C standard. C but you can't use more than two levels of pointers, and some other stuff like that.
adahome.com has good info on Ada. I liked this C++ to Ada tutorial: http://adahome.com/Ammo/cpp2ada.html
For hard real-time, Tom Hawkins has done some interesting Haskell stuff. See: ImProve (language incorporates an SMT solver to check verification conditions) and Atom (EDSL for hard realtime concurrent programming without using actual threads or tasks).
回答12:
A language that imposes careful patterns may help, but you can impose careful patterns using any language, even assembler. Every assumption about every value needs code that tests the assumption. For example, always test divisor for zero before dividing.
The more you can trust reusable components, the easier the task, but reusable components are seldom certified for critical use and will not get you through regulatory safety processes. You should use a tiny OS kernel and then build tiny modules that are unit tested with random input. A language like Eiffel might help, but there is no silver bullet.
回答13:
Any software product can pass the DO-178b certification process using any tool but the questions is how difficult would it be. If the compiler isn't certified you may need to demonstrate your code is traceable at the assembly level. So it is helpful that your compiler is certified. We used C on our projects but had to verify at the assembly level and use a code standard that included turning off the optimizer, limited stack usage, limited interrupt usage, transparent certifiable libraries, etc. ADA isn't pixie dust but it makes the PSAC plan look more achievable.
As applicatons get larger, assembly code becomes less viable choice. The ARM processor just invites C++, but if you ask companies like Kiel it their tool is certified, they will return with a "huh?" And don't forget that verificaton tools also need to be certified. Try verifying a LabView test program.
回答14:
A new safety-critical standard for Java is currently in development - JSR 302: Safety Critical Java Technology.
The Safety-Critical Java (SCJ) is based on a subset of RTSJ. The goal is to have a framework suitable for the development and analysis of safety critical programs for safety critical certification (DO-178B, Level A and other safety-critical standards).
SCJ for example removes the heap, which is still present in RTSJ, it also defines 3 compliance levels to which both application and VM implementation may conform, the compliance levels are defined to ease certification of variously complex applications.
Resources:
- Java for Safety-Critical Applications.
- JSR 302: Safety Critical Java Technology
- Developing Safety-Critical Applications with Java
回答15:
I don't know what language I'd use, but I do know what language I wouldn't:
NOTE ON JAVA SUPPORT. THE SOFTWARE PRODUCT MAY CONTAIN SUPPORT FOR PROGRAMS WRITTEN IN JAVA. JAVA TECHNOLOGY IS NOT FAULT TOLERANT AND IS NOT DESIGNED, MANUFACTURED, OR INTENDED FOR USE OR RESALE AS ON-LINE CONTROL EQUIPMENT IN HAZARDOUS ENVIRONMENTS REQUIRING FAIL-SAFE PERFORMANCE, SUCH AS IN THE OPERATION OF NUCLEAR FACILITIES, AIRCRAFT NAVIGATION OR COMMUNICATION SYSTEMS, AIR TRAFFIC CONTROL, DIRECT LIFE SUPPORT MACHINES, OR WEAPONS SYSTEMS, IN WHICH THE FAILURE OF JAVA TECHNOLOGY COULD LEAD DIRECTLY TO DEATH, PERSONAL INJURY, OR SEVERE PHYSICAL OR ENVIRONMENTAL DAMAGE.
回答16:
HAL/S is used for the Space Shuttle.
来源：https://stackoverflow.com/questions/243387/which-languages-are-used-for-safety-critical-software