Fix pipe hang (issue #319)

This issue turned out to be a kernel bug which has been fixed in:
torvalds/linux@46a525e

Reported here:
axboe/liburing#665 (comment)

We can workaround it by making sure pipes are non-blocking, uring considers
pipes unbounded work and relies on uring worker threads if they are blocking,
the bug is only triggered if this is the case, so force them to be non-blocking.

If we use splice, the splice call itself needs the worker threads and the bug
surfaces again (I verified this with perf probes), so disable splice for now.
Even with the fix, it's desirable to keep pipes as non-blocking to avoid thread
pooling.

The splice call can return EAGAIN in uring, this happens even with the kernel
patched, so handle it for the future.

We can tune this better by disabling splice only for the unpatched kernels.

lib_eio_linux/eio_linux.ml

@@ -953,16 +953,20 @@ let fast_copy src dst =
   with End_of_file -> ()
 
 (* Try a fast copy using splice. If the FDs don't support that, switch to copying. *)
-let fast_copy_try_splice src dst =
+let _fast_copy_try_splice src dst =
   try
     while true do
       let _ : int = Low_level.splice src ~dst ~len:max_int in
       ()
     done
   with
   | End_of_file -> ()
+  | Unix.Unix_error (Unix.EAGAIN, "splice", _) -> fast_copy src dst
   | Unix.Unix_error (Unix.EINVAL, "splice", _) -> fast_copy src dst
 
+(* XXX workaround for issue #319, PR #327 *)
+let fast_copy_try_splice src dst = fast_copy src dst
+
 (* Copy using the [Read_source_buffer] optimisation.
    Avoids a copy if the source already has the data. *)
 let copy_with_rsb rsb dst =
@@ -1290,6 +1294,9 @@ let stdenv ~run_event_loop =
 
 let pipe sw =
   let r, w = Unix.pipe () in
+  (* XXX workaround for issue #319, PR #327 *)
+  Unix.set_nonblock r;
+  Unix.set_nonblock w;
   let r = source (FD.of_unix ~sw ~seekable:false ~close_unix:true r) in
   let w = sink (FD.of_unix ~sw ~seekable:false ~close_unix:true w) in
   r, w