[SystemSafety] "Safe" and "Unsafe" Languages

mario.gleirscher at tum.de mario.gleirscher at tum.de
Mon Jan 23 07:23:24 CET 2017


Dear Peter,

Thanks for this nice encyclopedically concise summary. I have come across similar terminology problems when I was doing literature research for my thesis. Although, I decided to stick with a simple variant of Pnuelis safety/progress notions when formalizing my stuff.

Concerning your question: I'd very pragmatically start with improving the respective Wikipedia articles of which a view are not satisfyingly interlinked, maybe a category page would suffice. More idealistic, it would help to put a nice terminology chapter into a next version of one of the more popular course books around in the system safety community. Moreover, I don't think that a standard part or addendum is the rightest place to disseminate such important basic knowledge.

Best,
Mario

Von: Peter Bernard Ladkin
Gesendet: Montag, 23. Januar 2017 06:43
An: The System Safety List
Betreff: [SystemSafety] "Safe" and "Unsafe" Languages

There is an article in ACM TechNews from last week that talks about the programming language Nim,
which apparently has the language Rust's "safe-by-default" behaviour (it will be item 3 in Risks
30.10 when this edition goes up on the Risks Forum Digest WWW site). I think it was about a decade
or so that I read a short note from Andy Wellings which talked about a "safe" programming language.

Apparently, it is common parlance amongst the programming language community to talk about "safe"
languages. Is that right?

Here is an example from Rust. https://doc.rust-lang.org/nomicon/meet-safe-and-unsafe.html
Apparently, they seem to mean something such as: write your programs in Rust and they'll not have
run-time errors. I suspect a much weaker claim is justified, namely there are different kinds of
"X-safe", for example type-safe and memory-safe, and Rust is type-safe and maybe memory-safe. Pascal
and Modula were type-safe; I don't know what memory-safe means but presumably it includes absence of
memory leaks. Note that it doesn't make complete sense to talk about a programming language as being
type-safe simpliciter: whether a concrete program is type-safe also depends on whether your compiler
reliably preserves the semantics of types.

It is not clear to me what calling a language "safe" means in terms of concrete scientific
assertions about the language. For example, I doubt it is practically possible to have a
general-purpose language, all valid programs in which are guaranteed to be free of run-time error.
E.g., You'd have to guard against any expression used as the denominator in a division being
digitally-non-zero, and although there are industrially-practical ways of doing that with specific
programs using Hoare logic, I don't know any way of ruling that out a priori in a language
definition for a practical language.

I see no problems with the use of terms such as type-safe or (if it has a concrete meaning)
memory-safe. They clearly refer to specific properties and there is little or no chance of confusion
with the general term "safe" as used in functional safety (although many people such as myself do
not use that term unqualified; I prefer to speak of "adequately safe" rather than "safe"
simpliciter, because safety is a matter of degree, not of a binary property which a system has or
does not have).

There is another use of the term "safety", in "safety property" in concurrency theory. Here, there
is a theorem that any program property may be expressed as a conjunction of a "safety property" with
a "liveness property". A safety property is an invariant, "always X"; a liveness property is a
task-progression assertion "always eventually Y". They are often expressed as saying "nothing bad
[that would be NOT-X] ever happens" and "something good (Y) always eventually happens". Here, the
relevant term is "safety property", and people tend to talk about "fulfilling a safety property"
rather than saying "safe".

I have occasionally but regularly come across confusing use of the term "safe" by
concurrency/algorithm/logical-program-reliability specialists where the correctness of a program is
called "safety" if the program has some important function in some way. I often object that people
are talking about the reliability of a program, its ability to do what is specified of it, but not
safety, which is a system property partially and essentially defined by the environment in which the
program runs. Sometimes this is acknowledged graciously; sometimes not. The people who use the term
in this way are often distinguished computer scientists, so I do see it as important to say
something, and am disappointed when the response is a metaphorical shrug.

Standardised functional-safety conceptions don't necessarily help avoid this confusion between
safety and reliability. Functional safety is assured in IEC 61508 through reliability conditions on
specific functions introduced to mitigate a hazard with unacceptable risk.

The phenomenon is probably difficult socially to avoid. Back twenty years ago, in CERT's early days
when it was possible to count vulnerabilities, some 80% of vulnerabilities were said to be due to
buffer overflows, which are avoided if source code is written in strongly-typed languages and the
compiler enforces strong typing correctly. Calling such languages "safe" comes from knowing that
programs written in the language and compiled with a semantics-preserving compiler are *free from
such unwanted behaviour*. For functional-safety people, the "unwanted behaviour" is that causing
possible "harm", and "harm" is damage to people, property, environment, what have you, outside the
vocabulary of the system itself.

Do people see such memes, cases in which a word with an important technical meaning has three or
four such meanings, as well as a colloquial meaning, as problematic? If so, what is it reasonable to
do about it?

PBL

Prof. Peter Bernard Ladkin, Bielefeld, Germany
MoreInCommon
Je suis Charlie
Tel+msg +49 (0)521 880 7319  www.rvs-bi.de






-------------- next part --------------
An HTML attachment was scrubbed...
URL: <https://lists.techfak.uni-bielefeld.de/mailman/private/systemsafety/attachments/20170123/e8c4772b/attachment.html>


More information about the systemsafety mailing list