There is now a script
src/regression-tests/regression-test.py
that should in the future be the default for running regression tests
instead of the ad-hoc approach we are currently using. The goal is to
ultimately have more reliable and consistent regression testing.
The script takes as input "tests.txt" and tries to perform tests from
that. This is effectively a collection of inputs to the scyther-linux
binary. The results are writting to the 'results' directory, as
test-X.out and test-X.err, where those correspond to stdout and stderr,
respectively. Additionally, a measurement of wall-clock time in seconds
is written to test-X.time.
For now, we are using the timer to ensure all tests terminate. It would
be nicer to use a less environment-dependent way of enforcing
termination.
The use of 'strndup' in scanner.l caused problems for non-gnu modes of gcc, which
was being invoked for the mingw32 compilation. Replaced now by the more portable
strncpy + malloc version.
By specifying:
option "--X=Y";
in the Scyther input file, command-line options can be directly integrated.
For example, one can specify:
option "--one-role-per-agent";
The mechanism with the next pointers for tac's was working fine as long as all
tac's were unique by construction. The macro mechanism made it possible for
the same tac to occur twice in the tree. This could lead to an infinite loop.
Now we make explicit copies of the top-level tac. This should fix the problem
caused by the tuple parsing.
A more fundamental solution is to make a deep copy of the substituted terms.
It is now possible to declare syntactic macros at the global level.
macro ID = TERM;
After this definition, every occurrence of ID will be replaced by TERM.
For example, this can be used to avoid duplicating message definitions
among roles:
macro M1 = { nI, I}pk(R) ;
protocol X(I,R) {
role I {
send (I,R, M1);
}
role R {
recv (I,R, M1);
}
}
There is a new event:
not match(t1,t2)
where t1,t2 are terms.
They are implemented by using a special claim that simply stores the
intended inequality. The pruning theorems (prune_theorems.c) ensure that
these terms never become equal. If there are equal, the constraint is
violated. As long as they are not equal, there exists a solution using
groung terms such that their instantiation is not equal.
Currently not very efficient implemented and the graph out output is
also ugly for now.
Conflicts:
gui/Scyther/Trace.py
src/compiler.c
src/scanner.l
Introduced a new event:
match(pattern,groundterm)
This event can only be executed if pattern can be matched to groundterm.
Variable substitutions are persistent with respect to later events in
the same role.
Currently implemented as syntactic sugar, essentially unfolded in role R to:
fresh x;
send ( R,R, { groundterm }x );
recv ( R,R, { pattern }x );
This work is not complete yet in the send that the output still contains
the unfolding. Ideally, the graph rendered detects this syntactic sugar
and renders a simplified event. This should be possible on the basis of
the label name prefix.
Conflicts:
src/compiler.c
src/parser.y
src/scanner.l
src/tac.h
After we merged some concepts from the compromise branch, we forgot to add
for the hardcoded PKI that the adversary also should have access to (some) symmetric
keys.
This is not a full copy from the compromise branch. In particular,
some counts (in arachne.c) are missing, as well as the modified dot output (dotout.c).
The automatic mechanism to assign labels to claims was dependent on the
context. In practice, a claim could get a different label when analyzed in
isolation compared to when analyzed in parallel with some other protocols. This
caused problems for the multi-protocol analysis.
There are two new claims:
claim(X,Commit,t) : check for agreement on data
claim(X,Running,t) : signaling claim
The property checked is that each claim Commit needs to be preceded by a Running
with an identical term t.
Cherry-picked from commit 99a6be00e9d3d219ec73665607e8a3a7d65d04d1
Given that sk/pk/k are now hardcoded, we can exploit their occurrences with this
new heuristic.
The heuristic can now scan for the lowest term depth at which either sk or k occur.
This will cause the heuristic to favor looking for sk, then sk(x), and only later
other terms. In a small test this was twice as fast. For protocols based on pk only
the performance loss should be negligible.
The old heuristic was 162, now it is 162+512 = 674.
In the near future, the default exit code behavior should be made obsolete anyway,
as the exit codes are not a nice way to report status.
It used to be convenient for shell scripting in early times,
when the parallel tests were run using the forward model
checker, but no modern script should be relying on it.
When untyped variables occur, the encryption level depth pruning is for
now unjustified. Maybe we can get a proof later. Previously this was
hidden, which was a bad design decision. Now the output is much
clearer.
This cleans up some graphs rather nicely. There is only one potential
drawback (not observed in practive):
If two bindings have the same from/to, but different interpretations,
we might lose information. In particular the 'select' intermediate nodes might
pose a problem and we would be better off by not having any interpretation on
what is selected.
This switch was previously known as '--state-space', but the new name is
much better.
Backwards compatibility:
'-c' was previously used by '--check', so check is now abbreviated to
'-C'.
'-s,--state-space' still works but is from now on considered to be
deprecated.
I've added a marked for the Athena problem case, and now no more false 'complete proof' results are produced.
However, the tool reports, 'no attack within bounds', which is slightly inaccurate
depending on the interpretatio of 'bounds'.
