[PATCH] D153131: [clang analysis][thread-safety] Handle return-by-reference...

[Aaron Puchert via Phabricator via cfe-commits]

aaronpuchert added a comment.

Tried this on our code base, and the number of new warnings seems acceptable. I'll still need to look through them in more detail, but there is one suspicious warning that boils down to this:

  > cat reference-bug.cpp
  struct __attribute__((capability("mutex"))) Mutex {} mu;
  int* p __attribute__((pt_guarded_by(mu)));
  
  int& f() {
    return *p;
  }
  > clang-16 -fsyntax-only -Wthread-safety-analysis reference-bug.cpp
  > clang-16-D153131 -fsyntax-only -Wthread-safety-analysis reference-bug.cpp
  reference-bug.cpp:5:11: warning: writing the value pointed to by 'p' requires holding mutex 'mu' exclusively [-Wthread-safety-analysis]
    return *p;
            ^
  1 warning generated.

That we're warning here is correct, but the warning message is a bit off (we're not quite writing here), and it's under `-Wthread-safety-analysis` instead of `-Wthread-safety-reference`.



================
Comment at: clang/include/clang/Basic/DiagnosticGroups.td:1046
 def ThreadSafetyPrecise    : DiagGroup<"thread-safety-precise">;
 def ThreadSafetyReference  : DiagGroup<"thread-safety-reference">;
+def ThreadSafetyReturn     : DiagGroup<"thread-safety-return">;
----------------
courbet wrote:
> aaronpuchert wrote:
> > courbet wrote:
> > > aaronpuchert wrote:
> > > > Why not under `-Wthread-safety-reference`, as it's return-by-reference that you're warning on? This seems too small for a separate flag to me.
> > > The main reason it so that we provide a soft transition period for users: If we put that in `-Wthread-safety-reference`, we'll start breaking compile for people who use `-Werror`, while a separate flag allows a transition period where people opt into the new feature.
> > Transition flags can end up resulting in more churn, assuming that we eventually want to put this under `-Wthread-safety-reference`, because then you have two categories of users:
> > * those that opted in will eventually have to remove the flag again, and
> > * those that didn't will get hard errors on updating the compiler at that point.
> > Of course you might argue that the latter case can be prevented by carefully reading the release notes, but we know how often that happens.
> > 
> > I'd argue that if you're using `-Wthread-safety-reference`, you're already opting into warnings on escaping references, and not warning on `return` is a false negative.
> > 
> > A separate flag would make sense to me if we want to keep it, for example because this produces a substantial amount of false positives under some circumstances. Did you try this on a larger code base that's using the annotations? I could try it on our code, and maybe we can get some Googler to test it on theirs, which is also heavily using Thread Safety Analysis. (Though I'm not sure whether they use `-Wthread-safety-reference`.)
> I don't have a strong opinion for where the warning should go.  We are indeed using `-Wthread-safety-reference`, though we're not enabling -Werror on these, so adding more warnings is fine for us.
> 
> I've run the check on a small sample of our codebase (which I don;t claim to be representative, I can do a larger analysis if needed). The warnings are more or less evenly split between missing annotations and actual bugs. I don't think any of the things I've seen qualify as false positives.
> 
> Among the missing annotations, most of the warnings are missing `ABSL_EXCLUSIVE_LOCKS_REQUIRED` on the function. In a small number of cases, the pattern is that a variable is lazily initialized under a lock and then returned by reference:
> 
> ```
> struct LazyImmutableThing {
>   const Thing& Get() {
>     {
>       MutexLock lock(&mutex_);
>       thing_->Initialize();
>     }
>     
>     return thing_;
>   }
>   
>   Mutex mutex_;
>   Thing thing_ GUARDED_BY(mutex_);
> };
> ```
> 
> I consider this to be a missing annotation as the returned value is dynamically immutable, the proper fix would be `return TS_UNCHECKED_READ(thing_)`.
> 
> 
> Most actual bugs are along the lines of:
> 
> ```
> struct S {
>   T& Get() const {
>     MutexLock lock(&mutex_);
>     return obj_;
>   }
> 
>   Mutex mutex_;
>   T obj_ GUARDED_BY(mutex_);
> };
> ```
> 
> though some are missing the guard altogether (`T& Get() const { return obj_; }`).
> 
> There are a few possible fixes. In rough order of occurrence:
>  - Return by value as the copy is not too expensive and memory ordering does not matter.
>  - Let the caller take the lock and annotate with `ABSL_EXCLUSIVE_LOCKS_REQUIRED` when the `Get` method is not called too often.
>  - Let `Get` take a callback and process the value under the lock instead of returning it (when ordering matters).
> 
> In a small number of cases, the pattern is that a variable is lazily initialized under a lock and then returned by reference:

I wonder why that's safe, is the initialization guarded to happen only once? Some kind of double-checked locking pattern perhaps? Otherwise it seems that reads could happen in parallel to writes. If it's a checked initialization, then I think the proper way to model this is:
* The initialization acquires a lock to exclude other initializations running in parallel. Reads cannot happen, because the reference has not yet escaped.
* After initialization, we essentially acquire an implicit shared lock. This is not tracked as a proper lock, but it doesn't need to: there are no more writes until the end of lifetime, so nobody will acquire another exclusive lock.
One could model this by creating a mutex wrapper that can be locked once in exclusive mode, and after that hands out shared locks to everybody who wants one without keeping track. (As this is slightly off-topic, we don't need to discuss this here though.)

Other than than, this matches what I'm seeing in our code.


Repository:
  rG LLVM Github Monorepo

CHANGES SINCE LAST ACTION
  https://reviews.llvm.org/D153131/new/

https://reviews.llvm.org/D153131