SqlClient: SqlCommand.Cancel() does not work on Linux

SqlCommand.Cancel() doesn’t appear to work using System.Data.SqlClient on Linux (Ubuntu 19.04, .NET Core SDK 3.0.100-preview5-011568) - the command runs to completion, and only then throws SqlException. The issue does not seem to repro on Windows (with .NET Core).

Repro:

using (var cmd = new SqlCommand("WAITFOR DELAY '00:00:05'", conn))
{
    // sync cancellation
    Task.Delay(1000).ContinueWith(t => cmd.Cancel());
    var sw = Stopwatch.StartNew();
    try
    {
        cmd.ExecuteNonQuery();
    }
    catch (Exception e)
    {
        Console.WriteLine($"Sync cancellation: caught {e.GetType().Name} after {sw.ElapsedMilliseconds}ms");
        // Does not interrupt the call (completes after 5 seconds) but still generates SqlException at the end
    }

    // async cancellation
    sw.Restart();
    try
    {
        await cmd.ExecuteNonQueryAsync(new CancellationTokenSource(1000).Token);
    }
    catch (Exception e)
    {
        Console.WriteLine($"Async cancellation: caught {e.GetType().Name} after {sw.ElapsedMilliseconds}ms");
        // Interrupts and generates SqlException immediately
    }
}

/cc @divega @karinazhou

About this issue

Original URL
State: closed
Created 5 years ago
Comments: 15 (11 by maintainers)

Most upvoted comments

I’ve tried the idea above. It works as far as I can tell. The full manual test suite runs through without unusual problems (TVPMain really need fixing at some point).

It’s a bit annoying fix because of the current packet cloning which I’ve fixed in the managed networking rework PR. It will also be blocked by the same issue as that PR https://github.com/dotnet/corefx/pull/35363 where an unrelated test that fails anyway will continue to fail so it won’t get merged. So once things get unblocked I’ll be able to fix this.

Wraith2 on May 21, 2019

It’s a managed implementation issue. I’ve reproduced it on windows using your test script above and have added a new test.

There are lock(this) statements in the managed send and receive methods on SniTCPHandle. So what happens is you start a query which ultimately tracks down into Receive which takes the lock and then from another thread you call Cancel which tracks down into SendAttention and that tries to Send and blocks on the lock. When the receive times out the lock is released the send works, the attention is processed and everything unravels in the correct order just a lot later that you want it to.

I think send and receive aren’t mutually exclusive but that each one is not concurrent with itself. Changing the lock to be operation specific consumes a bit of object space but there shouldn’t be too many tcp handles around since they’re reused. It will need some careful testing because locking in this library is definitely “here be dragons” territory.

Wraith2 on May 20, 2019