It seems that hPutBuffNonBlocking flush the buffer and blocks anyway when it has not enough space for the next message.

    -- else, we have to flush
        else do debugIO "hPutBuf: flushing first"
                old_buf' <- Buffered.flushWriteBuffer haDevice old_buf
                        -- TODO: we should do a non-blocking flush here

https://tldrify.com/bga

this should be a bug or a feature not implemented.

since the flush uses flushWriteBuffer, that blocks, hPutBuffNonBlocking does the same than hPutBuff and the buffer congestion can not be detected.

I will try to do a new version with flushWriteBuffer0 which do not blocks.

2015-09-17 15:41 GMT+02:00 james <james@mansionfamily.plus.com>:

On 17/09/2015 13:52, Brandon Allbery wrote:

...
It is possible that recent network package changed this, but in the past *all* socket operations had to be under the aegis of *one* withSocketsDo, otherwise any handles, buffers, etc. would become invalid when Winsock was deinitialized.

See http://neilmitchell.blogspot.co.uk/2015/02/making-withsocketsdo-unnecessary.html

Alberto.