ssh broken pipe problem

[James Walker]

On 1/23/20 2:04 PM, Paul Jackson wrote:
>>> Any idea what could be wrong?
> 
> Do you have an estimate as to how many bytes
> were successfully transferred over the ssh connection,
> prior to the abrupt failure?

As I said, this is happening on "hg incoming".  I don't know exactly how 
that's implemented, but I would expect it would just transfer a few 
changeset checksums.  (The entire repository is way less than a gigabyte.)

> 
> If it's on the order of 1 or 2 GBytes (I forget which)
> then I would suspect that the server  sending down
> those bytes is Linux and that it's kernel abruptly
> dropped the connection, exactly one transfer block
> short of sending exactly that 1 (or 2?) GBytes down
> the connection.
> 
> I've long noticed this problem, but have been unable
> to get enough of a test case and diagnosis to make
> a decent bug report, or to even know to what project
> to file the bug.
> 
> Rather I work around this by opening up a separate
> connection, using raw BSD socket calls (socket, bind,
> listen, accept, ...) to send down the bulk of the data
> whenever I might be sending a GByte or more in one
> ssh session.
>