On 21 April 2005 17:57, Glynn Clements wrote:

Simon Marlow wrote:

...

...

The test program works after applying the following patch to GHC that disables nonblocking mode after the fork and before the exec.

...

If we did this, we would also need to re-enable non-blocking mode after the rawSystem completes. Also, this doesn't work well if runProcess is used and the Haskell program continues to run while the subprocess is executing, because the O_NONBLOCK flag is shared between parent and child.

Presumably the Haskell program won't like it very much if the spawned program disables O_NONBLOCK itself?

Is there a reason (other than performance) why the use of O_NONBLOCK can't be replaced with calling select/poll before accessing the descriptor?

That would work, as long as the current process is the only user of that file descriptor: otherwise there's a race condition between the select() and the read(). This is really a Unix design blunder, there ought to be a non-blocking version of read().

Is non-blocking I/O only necessary if you need explicit concurrency? Or do the runtime or critical libraries require it?

Right, it's only important if you need concurrency. Cheers, Simon