[llvm][Support] Call clear_error in LockFileManager to avoid report_fatal_error

[Bigcheese]

As per the comment in raw_fd_ostream's destructor, you must call clear_error() to prevent a call to report_fatal_error(). There's not really a way to test this, but we did encounter it in the wild.

rdar://117347895

[llvmbot]

@llvm/pr-subscribers-llvm-support

Author: Michael Spencer (Bigcheese)

Changes

As per the comment in raw_fd_ostream's destructor, you must call clear_error() to prevent a call to report_fatal_error(). There's not really a way to test this, but we did encounter it in the wild.

rdar://117347895

---

Full diff: https://github.com/llvm/llvm-project/pull/83655.diff

1 Files Affected:

• (modified) llvm/lib/Support/LockFileManager.cpp (+2)

diff --git a/llvm/lib/Support/LockFileManager.cpp b/llvm/lib/Support/LockFileManager.cpp
index 34c7a16b24be41..facdc5a0d7d421 100644
--- a/llvm/lib/Support/LockFileManager.cpp
+++ b/llvm/lib/Support/LockFileManager.cpp
@@ -205,6 +205,8 @@ LockFileManager::LockFileManager(StringRef FileName)
       S.append(std::string(UniqueLockFileName));
       setError(Out.error(), S);
       sys::fs::remove(UniqueLockFileName);
+      // Don't call report_fatal_error.
+      Out.clear_error();
       return;
     }
   }

[dwblaikie]

Hmm - this was sent for review but committed without review?
If you didn't mean for this to require pre-commit review, please add the tag skip-precommit-approval when creating the PR (per: https://llvm.org/docs/CodeReview.html#code-review-workflow - though I need to fix the link in the note there... )

As for the change itself - if this code was invoked by a unit test, would it fail the unit test due to report_fatal_error? (or would the unit test exit quietly? I'd think it'd fail) - so maybe could be tested that way? Or is it that it's hard to create the error condition that'd lead to the report fatal error in the first place?

[Bigcheese]

Yeah, this was just for CI, I'll add the tag in the future.

Getting into a situation that would have an error here is the hard part. I don't actually know how it happened, all I have is a stack trace from the call to abort after report_fatal_error, but not the actual error it failed with. To get here you have to successfully create and open the lock file with the correct permissions, but then fail to write to it. It's possible this happened because of a network failure or some other weird situation.