IF YOU WOULD LIKE TO GET AN ACCOUNT, please write an
email to Administrator. User accounts are meant only to access repo
and report issues and/or generate pull requests.
This is a purpose-specific Git hosting for
BaseALT
projects. Thank you for your understanding!
Только зарегистрированные пользователи имеют доступ к сервису!
Для получения аккаунта, обратитесь к администратору.
When a connection is closed by the client before we have enough data
to determine if it contains a TLS Handshake or not, the socket stays
in a readable state.
While we setup a tokio backed timeout of 10s for the connection
build-up here, this timeout does not trigger on said early connection
abort from the client side, causing then the async_io loop to
endlessly loop around peeking into the client, which always returns
the last available bytes before the connection was closed. This in
turn causes 100% CPU usage for one of the PBS threads.
The timeout not triggering is rather odd, and does indicate some
potential for further improvement in tokio itself, but our
questionable use of the WouldBlock error does violate the API
contract, so this is not a clear cut.
Such an early connection abort is often triggered by monitoring
solutions, which use it to relatively cheaply check if TCP on a port
still works as "is service up" heuristic.
To fix this, save the amount of bytes peek returned and if they did
not change between invocations of the callback, we can assume that the
connection was closed and thus exit the connection attempt with an
error.
Signed-off-by: Dominik Csapak <d.csapak@proxmox.com>
[ TL: reword commit message and change error to ConnectionAborted ]
Signed-off-by: Thomas Lamprecht <t.lamprecht@proxmox.com>
dc9531d302