- Some obsolete bugs removed.
This commit is contained in:
parent
04b6ab4b36
commit
814fbf31cd
39
src/bugs.txt
39
src/bugs.txt
@ -1,13 +1,7 @@
|
||||
--+++ Crititcal Bugs
|
||||
|
||||
* soph segfaults at no switch or -r4 (-r3 is okay??) using non-debug version.
|
||||
* './scyther -a ../spdl/nsl3.spdl --increment-runs' segfaults. The main reason is that the Archne engine uses runs in a different way.
|
||||
Maybe it is best to disable --increment rules for non-ModelChecker verification.
|
||||
|
||||
---+++ Bugs
|
||||
|
||||
* Problem with goal bindings: instantiation of variable with a tuple might
|
||||
introduce a tuple goal, which is forbidden. We must find a way to deal with this. This typically occurs in type flaw searches.
|
||||
* Arachne seems to trip over claims with empty prec sets. Maybe we
|
||||
simply should not test these.
|
||||
* Splice/AS does not work well because priority key search stumbles over the
|
||||
@ -27,19 +21,9 @@
|
||||
|
||||
---++++ ArachneEngine
|
||||
|
||||
* There is no good test on the correct workings of
|
||||
add_goals/destruction of these. We can test this if after
|
||||
termination, we have 0 goals; for this we need to store the
|
||||
initially added goals as well. Furthermore, we can generate an
|
||||
error when <0 goals occur.
|
||||
* Consider where in Arachne dependency graph is used. If this is only for
|
||||
pruning states, we can construct it there only. However, the base 'role
|
||||
defs/bindings' graph might be re-used.
|
||||
* Add switch for arachne to prune encryption levels when using -m2.
|
||||
* To store attacks for arachne, maybe the following is needed:
|
||||
* The roles for each run
|
||||
* The variable bindings for all (local) variables
|
||||
* The goal bindings
|
||||
* Agent terms must have keylevel 0; enforce this!
|
||||
* Select_goal should consider, for singular variables, whether their
|
||||
type can be found in M_0. If so, the goal can be ignored.
|
||||
@ -51,19 +35,11 @@
|
||||
Note that there can be multiple solutions; for now, simply try to take the
|
||||
shortest one.
|
||||
|
||||
---++++ ModelChecker
|
||||
|
||||
* For secrecy, one trusted agent and one untrusted agent suffices.
|
||||
Implement this in the modelchecker.
|
||||
* Implement delayed protocol compiler (on run demand only) for the modelchecker?
|
||||
|
||||
---++++ Misc
|
||||
|
||||
* Make different error codes for compilation error/ other error. This can be
|
||||
useful for scripts. However, it might shift some constants for the Elegast
|
||||
scripts.
|
||||
* Rewrite termMguTerm such that it iterates and adapt all functions
|
||||
using it. This is to allow for associative tupling later.
|
||||
* Fix constants in intruder knowledge. Auto add single one of each type,
|
||||
when typed expl. Add single constant when untyped. Fix this also in
|
||||
semantics, and add proof to establish sufficiency.
|
||||
@ -102,18 +78,3 @@
|
||||
* How is % notation handled in Casper?
|
||||
* Vernam encryption?
|
||||
|
||||
---++++ ConstraintLogic (and thus obsolete)
|
||||
|
||||
* CLP: variables in keys must be branched: maybe even in three situations
|
||||
(have key and contents, have inverse key and content, nothing)
|
||||
* How should claims behave (trusted/untrusted) wrt uninstantiated
|
||||
agents? Branch again? That's what is causing the nsl3-var problem.
|
||||
* Constraints might be a part of a knowledge thing, because with we
|
||||
might now have a number of local knowledge sets, each with their own
|
||||
constraint sets. That doesn't make it easier though :( and will cause
|
||||
some performance loss I suppose. Each local set has to remain
|
||||
solveable as well.
|
||||
* Issue: how do untrusted claims work in the context of an intruder?
|
||||
Claim must be checked if it can be solved such that at least one of
|
||||
the agents is trusted.
|
||||
|
||||
|
Loading…
Reference in New Issue
Block a user