mjhjmtu

Intermittently, on a raspberry pi running raspbian, I see bizarre behaviour with (one of several) established TCP connections. The receiver has what he believes is a valid open socket gotten from accept. He's sitting in a read() and asking for one byte (in some cases more, but I've seen this in the one byte case).

The sender has sent data. /proc/net/tcp clearly shows the socket in the Established (01) state, with a nonzero count in the receive queue - and an inode of 0.

This happens rarely, and I think only on sockets that are open for days and weeks at a time. It has affected several different applications on the same raspberry pi 3B+; it is not always the same code that's affected. I am not certain if the socket starts out damaged or ends up that way over time; some of these sockets are used for traffic that happens very rarely.

I can't seem to attract attention to the issue on the rasp forums; but this has happened several times now to me. I'm looking for insight into how this can happen; ideally someone will say "oh, their TCP implementation forgot to snork the cache after globbing the snoid buffer" or some such thing. Does anyone have insight into how this is addressed?