This fixes a number of problems around the pcap-fifo handling

[pjsg]

The underlying issue is that writing to a fifo typically does not write the full amount of data. The code in out_flush tries to handle this, but gets it wrong. In particular, the existing code tends to go into a stack overflow after writing the first block of data to the fifo. This causes a memory overwrite, and then the file number used for subsequent writes is invalid, and so it returns EBADF. The is caught, adn the code then tries to shut down cleanly -- which calls out_close which calls out_flush which fails to write (with EBADF) and no we are doomed.

Also, the out_close function assumed that out_flush would always flush all the buffered data -- and it doesn't.

This PR fixes those issues.

[cmikk]

Thanks for investigating this issue and sharing your patches.

We will be pushing fixes to the size_t underflow error and out_flush/stop recursion shortly, and doing a further review of this code for more potential fixes.

Regarding the out_close leaving buffered data behind, this is by design: discarding any buffered data which cannot be immediately written reduces the time to re-establish the connection when restarting sratool, in turn reducing the overall amount of data lost or missed. The "flush all the buffered data" functionality will not be included in the upcoming change set, but we are open to making it configurable if this behavior is still desired.

[pjsg]

My use case is running in a kubernetes environment. When a new version of the container gets deployed, I want a graceful shutdown of the old container. I guess I can live with the not flushing of the output.

My actual script is:

    mkdir -p /pvc/ch14
    mkfifo /tmp/pcap
    START=`date +%Y%m%dT%H%M%S`
    (tcpdump -r /tmp/pcap -G 900 -w /pvc/ch14/%Y%m%dT%H%M%S_${START}.ch14.pcap -z /compress; /compress `ls -t /pvc/ch14/*${START}.ch14.pcap | head -n 1` ) &
    trap 'pkill -TERM sratunnel' TERM
    sratunnel -c 14 -s ssh:[email protected] -o pcap-fifo:/tmp/pcap -w ch=14 -n ~/axa_config &
    wait
    wait

This uses tcpdump to split the pcap stream into 15 minute chunks that are written to individually named files and then compressed. The container SIG_TERM is forwarded to sratunnel which shuts down, and the tcpdump drains the fifo and it too exits. Unfortunately, it doesn't call the -z command on termination, so that has to be done specially. Finally the script finishes and the container exits.