Combining netcat with fifos results in stuck queues/sbwait

[Wayne Sierke]

I was hoping to establish a simple processing server using nc(1). After
finding numerous examples of combining netcat with fifos (named pipes) I
am unable to establish a reliable setup. E.g. following the example of
the canonical netcat server:

server:

        #mkfifo backpipe
        #nc -l 4242 < backpipe | tr -u "[:lower:]" "[:upper:]" >
        backpipe

client:

        #mkfifo frontpipe
        #nc localhost 4242 < frontpipe > testfile.out &
        #cat testfile.in > frontpipe

For any non-trivially small testfile I inevitably end up with the
processes at both ends stuck in sbwait.

I have managed to avoid the stuck condition by setting the TCP buffer
sizes at both ends:

nc -I 2048 -O 1024

with a resulting throughput of ~40kB/s (and a supplementary problem of
truncation).

There are other oddities, such as not using the fifo on the client end:

        #nc -I 2048 -O 1024 localhost 4242 < testfile.in > testfile.out
        &

also results in "stuck" processes.

So far I've tried on 8.2/8.3 i386 and amd64. The full implementation is
planned to use fifos for input and output along with fd redirection but
attempts using that have fared no better.

Can anyone provide a secret-sauce recipe for netcat/fifo success?


Thanks,

Wayne