Current status, Voron on Linux
time to read 3 min | 416 words

We have been quite for a while on that front, but that is just because we were head down and working hard on it. The current status is that we are getting there Smile.

Recently we have been fighting with a SIGSEGV (segmentation fault) that appear out of nowhere and kind of ruin our fun. To be rather more exact, we have been on that bug for over a month. We have been getting great help from Sine Nomine about figure out the root cause of that, and today Neale Ferguson finally tracked down the final clues that gave up a simple reproduction.

Take the following code and run it on a Linux machine (in our case, Ubuntu 14.4 with Mono 3.10). It would crash with a really nasty error:

class MainClass
{
    public unsafe static void Main(string[] args)
    {
        var f = Syscall.mmap(IntPtr.Zero, 1024 * 1024, MmapProts.PROT_WRITE | MmapProts.PROT_READ, MmapFlags.MAP_SHARED | MmapFlags.MAP_ANONYMOUS, -1, 0);

        var p = new byte*[]{ (byte*)f.ToPointer() };

        Syscall.munmap(f,1024*1024);

        Console.WriteLine(p.Length);

        p = null;

        GC.Collect();
        GC.WaitForPendingFinalizers();
    }
}

The underlying reason, to the best of my understanding at the moment, is that the sgen GC implementation in Mono is going inspecting the p array for its values. But instead of just treating this as an opaque value (because I wouldn’t expect the GC to follow a pointer reference), it is trying to access the value. Since by this time, this value is no longer valid, it is accessing invalid memory location, and dying horribly.

The good news htat once we had a simple repro, the Mono team was really quick on their feet and this has already been fixed.

Tweet Share Share 2 comments
Tags: