TempFile::keep() issue on WIndows

Hi All,

I’ve run into an issue trying to run an llvm-ar executable built on Windows 10 on a Windows 7 machine.

I get a permission denied error when the archiver tries to write the new archive file:

%> llvm-ar rv mylib.lib foo.o

llvm-ar.exe: warning: creating mylib.lib

llvm-ar.exe: error: mylib.lib: permission denied

I’ve tracked the issue down to a call to ::GetFinalPathByHandle() from realPathFromHandle() (defined in llvm/lib/Support/Windows/Path.inc).

realPathFromHandle() will set an error code when ::GetFinalPathByHandle() returns a value of 0 (as it does in the above case).

Is there a known incompatibility/limitation/bug with trying to run the Windows 10 kernel version of GetFinalPathByHandle() on a Windows 7 machine?

~ Todd Snider

It’s surprising (at least, to me) to get “permission denied” when attempting GetFilePathNameByHandle. It seems more likely that the function failed to resolve the path for some other reason, and then the archiver tried to do an operation to the file with the unresolved path, and that step caused the “permission denied.”

But if it is the GetFilePathNameByHandle, then I would guess that the file handle is corrupted thus the request is about an object the process doesn’t have rights to access…

I don’t have Windows 7 to try it out.

Does the file already exist? If not, is it being created? Is there anything unusual about the path (e.g., is it longer than 256-ish characters long or referencing a server over a network or contain characters outside of ASCII that are possibly in the wrong encoding)?

If you’re able to debug it at the point of the GetFilePathNameByHandle call, can you check that the handle looks reasonable? It should look like a valid pointer (possibly just the lower 32-bits if you’re on a 64-bit machine), so 0, 0xFFFFFFFF, and unaligned values would be suspicious.

Another common problem is that sometimes anti-malware scanners will briefly lock a newly created file to scan it, which can cause a sharing violation. Usually these are intermittent because they’re timing sensitive, but it might be worth temporarily disabling your scanners to see if that makes the problem go away. If so, we can look at the access pattern and see if there’s a less fragile approach.

I think Windows 7 support is broken currently.

The bug report sounds like https://github.com/rust-lang/rust/issues/81051

See discussion in https://reviews.llvm.org/D81803

My personal view is that LLVM should officially drop support for Windows 7. Windows 7 is end of life, as of February 2020, and AFAIK, we don’t have any build bots running Windows 7. As such, we can’t guarantee that any changes that are made will work on that version of Windows (D81803 being a case in point). Similarly, most LLVM developers won’t have access to such machines, so won’t be able to test fixes or reproduce bugs locally. As we move forward, it is likely that more of these issues will crop up, and usually won’t be noticed until late on when some project imports a release branch with a bug in.

Alexandre, thanks for the pointer to 81051. This is the same issue I am seeing.

If I remove the code that was added to fix https://reviews.llvm.org/D81803 from setDeleteDisposition(), my issue goes away, but that would just cause the D81803 issue to be resurrected.

James, it would be nice if we could drop support for Windows 7, but there is still significant usage of Windows 7 (especially outside the US).

~ Todd

There is also https://bugs.llvm.org/show_bug.cgi?id=48378

Hello Todd,

I’ve landed

https://github.com/llvm/llvm-project/commit/64ab2b6825c5aeae6e4afa7ef0829b89a6828102

which should fix the problem you’re seeing. If it doesn’t, please let me know.

Alex.

Hi Alex,

I was able to confirm that your fix addressed the issue I was seeing.

Thanks!!!

~ Todd