Summary: In some cases where a function is called directly on a formal (e.g, `def foo(o) { callSomething(o) }`, we were failing to propagate the footprint trace to the caller.
Reviewed By: jeremydubreil
Differential Revision: D4502404
fbshipit-source-id: d4d632f
Summary:
At one point I thought we'd want to have lots of different schedulers for things like exploring loops in different orders, but that hasn't materialized.
Let's make the common use-case simpler by hiding the `Scheduler` parameter inside the `AbstractInterpreter` module.
We can always expose `MakeWithScheduler` later if we want to.
Reviewed By: jberdine
Differential Revision: D4508095
fbshipit-source-id: 726e051
Summary:
When the receiver type and return type of an unknown call are the same, propagate taint to both the receiver and the return type.
This does the right thing for common "builder-style" methods that both update and return the receiver.
We already had custom models for a few such methods (e.g., `StringBuilder.append`), but we can remove them now.
Reviewed By: jeremydubreil
Differential Revision: D4490071
fbshipit-source-id: 325ea88
Summary: This will be useful in upcoming changes to the thread-safety analysis as well.
Reviewed By: dkgi
Differential Revision: D4402146
fbshipit-source-id: c750127
Summary:
A domain should not definite its initial state, since distinct users of the domain may want to choose different initial values.
For example, one user might want to bind all of the formals to some special values, and one user might want the initial domain to be an empty map
This diff makes this distinction clear in the types by (a) requiring the initial state to be passed to the abstract interpreter and (b) lifting the requirement that abstract domains define `initial`.
Reviewed By: jberdine
Differential Revision: D4359629
fbshipit-source-id: cbcee28
Summary: This more easily allow to switch between the different modes for handeling dynamic dispatch
Reviewed By: sblackshear
Differential Revision: D4367556
fbshipit-source-id: 795d2c4
Summary:
We currently can only model the return values of functions as sources.
In order to model inputs of endpoints as sources, we need the capability to treat the formals of certain functions as sources too.
This diff adds that capability by adding a function for getting the tainted sources to the source module, then using that info in the analysis.
Reviewed By: jeremydubreil
Differential Revision: D4314738
fbshipit-source-id: dd7d423
Summary: Different analyses need different preanalyses to run. It doesn't make sense for all of the pre-analyses to be bundled together into one package.
Reviewed By: jvillard
Differential Revision: D4348243
fbshipit-source-id: 46a8ebd
Summary:
Previously, summaries worked by flattening the access tree representing the post of the procedure into (in essence) a list of functions from caller input traces to callee output traces.
This is inefficient in many ways, and is also much more complex than just using the original access tree as the summary.
One big inefficiency of the old way is this: calling `Trace.append` is slow, and we want to do it as few times as possible.
Under the old summary system, we would do it at most once for each "function" in the summary list.
Now, we'll do it at most once for each node in the access tree summary.
This will be a smaller number of calls, since each node can summarize many input/output relationships.
Reviewed By: jeremydubreil
Differential Revision: D4271579
fbshipit-source-id: 34e407a
Summary:
Before, the Interprocedural functor was a bit inflexible. You couldn't do custom postprocessing like normalizing the post state or coverting the post from an astate type to a summary type.
Now, you can do whatever you want by passing a custom `~compute_post` function.
Since `AbstractInterpreter.compute_post` can be used by clients who don't care to do anything custom, this doesn't create too much boilerplate.
Reviewed By: jvillard
Differential Revision: D4309877
fbshipit-source-id: 8d1d85d
Summary:
Use In_channel and Out_channel operations instead of those in Pervasives. Don't
use physical equality on values that aren't heap-allocated since it doesn't help
the compiler generate faster code and the semantics is unspecified. Also use
phys_equal for physical equality.
Reviewed By: sblackshear
Differential Revision: D4232459
fbshipit-source-id: 36fcfa8
Summary:
Utils contains definitions intended to be in the global namespace for
all of the infer code-base, as well as pretty-printing functions, and
assorted utility functions mostly for dealing with files and processes.
This diff changes the module opened into the global namespace to
IStd (Std conflict with extlib), and moves the pretty-printing
definitions from Utils to Pp.
Reviewed By: jvillard
Differential Revision: D4232457
fbshipit-source-id: 1e070e0
Summary:
The Java frontend translates exceptions by assigning them to the return value.
This leads to weird behavior when the return type of the function is void.
Already handled one case of this in Quandary (ignoring assignments of exceptions to return value), but was missing the case where null is assigned to the return value.
The frontend does this to "clear" the value of previously assigned exceptions.
Reviewed By: jeremydubreil
Differential Revision: D4294060
fbshipit-source-id: 6bef5ef
Summary:
Functions related to source files were already namespaced by `source_file_` prefix. Make separate module for them.
In high level it replaces all `source_file_` with `SourceFile.` and then fixes all remaining compilation errors
Reviewed By: jvillard
Differential Revision: D4299053
fbshipit-source-id: 20b1d39
Summary:
Trying to stop other users of the trace domain from making the mistake that Quandary made before D4234766.
This should also improve the performance of Quandary, since the filtering of FP's is now done before building up the full interprocedural trace (which requires disk reads).
Reviewed By: jeremydubreil
Differential Revision: D4234770
fbshipit-source-id: e7e9291
Summary:
We only ought to report a source-sink flow at the call site where the sink is introduced.
Otherwise, we will report silly false positives.
Reviewed By: jeremydubreil
Differential Revision: D4234766
fbshipit-source-id: 118051f
Summary: Noticed this when I was writing the documentation for the abstract interpretation framework and was curious about why `Ondemand.analyze_proc` needs the type environment. It turns out that the type environment is only used to transform/normalize Infer bi-abduction specs before storing them to disk, but this can be done elsewhere. Doing this normalization elsewhere simplifies the on-demand API, which is a win for all of its clients.
Reviewed By: cristianoc
Differential Revision: D4241279
fbshipit-source-id: 957b243
Summary:
In Java, we handle unknown code by propagating behavior from the parameters of the unknown function call to the return value (or constructed object, in the case of a constructor). But we do this in a somewhat silly way--generating a new summary with these semantics at each unknown call site. Instead, this diff introduces these two options as predefined behaviors and adds specialized code for them.
As a side effect of this approach, unknown functions are no longer counted as passthroughs. This is ok; the original behavior was less of a reasoned decision and more of an unintended consequence of the way we decided to handle unknown code.
This new approach ought to be more efficient than the old one, and as a virtuous side effect it will be easier to specify how to handle unknown code in other languages like C++.
Reviewed By: jeremydubreil
Differential Revision: D4205624
fbshipit-source-id: bf97445
Summary:
Let's introduce some concepts. A "known unknown" function is one for which no Java code exists (e.g., `native`, `abstract`, and `interface methods`). An "unknown unknown" function is one for which Java code may or may not exist, but we don't have the code or we choose not to analyze it (e.g., non-modeled methods from the core Java or Android libraries).
Previously, Quandary handled both known unknowns and unknown unknowns by propagating taint from the parameters of the unknown function to its return value. It turns out that it is really expensive to do this for known unknown functions. D4142697 was the diff that starting handling known unknown functions in this way, and bisecting shows that it was the start of the recent performance problems for Quandary.
This diff essentially reverts D4142697 by handling known unknowns as skips instead. Pragmatically, doing the propagation trick for Java/Android library functions (e.g., `String` functions!) matters much more, so i'm not too worried about the missed behaviors from this. Ideally, we will go back to the old handling once performance has improved (have lots of ideas there). But I need this to unblock me in the meantime.
Reviewed By: jeremydubreil
Differential Revision: D4205507
fbshipit-source-id: 79cb9c8
Summary: Generalizing jvillard's awesome work to include passthroughs in traces, then calling it from Quandary.
Reviewed By: jvillard
Differential Revision: D4172108
fbshipit-source-id: 0296c59
Summary: If a procedure is both a source and a sink for the same value, and it's a sink first, you will get a false positive when applying the summary for the procedure.
Reviewed By: cristianoc
Differential Revision: D4145246
fbshipit-source-id: 97f0022
Summary:
Analyses should handle methods whose code is unknown and methods whose summary is a no-op differently.
Previously, this was done correctly for some kinds of methods (e.g., native methods, which were recognized as unknown), but not for others (interface and abstract methods).
This diff makes sure we correctly treat all three kinds as unknown.
Reviewed By: jeremydubreil
Differential Revision: D4142697
fbshipit-source-id: c88cff3
Summary:
Our default strategy for handling unknown code is to propagate taint from the actuals to the return value.
But for commonly-used methods like `StringBuilder.append` (used every time you do `+` with a string in Java), this doesn't work.
The taint should be propagated to both the receiver and the return value in these cases.
I'm considering a solution where we always propagate taint to the receiver of unknown functions in the future, but I am concerned about the performance.
So let's stick with a few special string cases for now.
Reviewed By: cristianoc
Differential Revision: D4124355
fbshipit-source-id: 5b2a232
Summary: A must-have for reporting taint errors and any other interprocedural error where the trace is sufficiently complex.
Reviewed By: jvillard
Differential Revision: D4124072
fbshipit-source-id: 26b3b2b
Summary: A must-have for reporting taint errors and any other interprocedural error where the trace is sufficiently complex.
Reviewed By: jvillard
Differential Revision: D4106352
fbshipit-source-id: b2677e6
Summary:
See code comment about `throw exn` being translated as `return exn`.
This problem was revealed by D4081279, which started grabbing access paths from exceptions.
Reviewed By: jvillard
Differential Revision: D4096391
fbshipit-source-id: 9d91513
Summary:
this makes frontends no longer depend on SymExec.ml. `ModelBuiltins` was split into two modules:
- `BuiltinDecl` with procnames for builtins (used to determine whether some function is a builtin)
- `BuiltinDefn` with implementations used by `SymExec`
- they both have similar type defined in `BUILTINS.S` which makes sure that new builtin gets added into both modules.
During the refactor I ran some scripts:
`BuiltinDecl.ml`:
let X = create_procname "X"
cat BuiltinDecl.ml | grep "create_procname" | tail -70 | awk ' { print $1,$2,$3,$4,"\42"$2"\42"} '
then manually confirm string match. Exceptions:
"__exit" -> "_exit"
"objc_cpp_throw" -> "__infer_objc_cpp_throw"
__objc_dictionary_literal
nsArray_arrayWithObjects
nsArray_arrayWithObjectsCount
`BuiltinDefn.ml`:
let X = Builtin.register BuiltinDecl.X execute_X
cat BuiltinDecl.ml | grep "create_procname" | tail -70 | awk ' { print $1,$2,$3,"Builtin.register BuiltinDecl."$2,"execute_"$2} '
then, fix all compilation problems
Reviewed By: jberdine
Differential Revision: D3951035
fbshipit-source-id: f059602
Summary:
Right now, taint gets lost if it flows into a constructor or procedure whose implementation is missing.
Since the core Java (e.g., String) and Android classes (e.g, Intent) are among these, this is bad.
We could handle this by writing a bunch of models instead, but that would be a lot of work (plus we may still miss cases).
Reviewed By: jvillard
Differential Revision: D4051591
fbshipit-source-id: 65851c8
Cristiano Calcagno
Sam Blackshear
Jeremy Dubreil
Josh Berdine
Andrzej Kotulski