MDEV-39473: main.mdev375, main.connect fail on macOS

[DaveGosselin-MariaDB]

When the server rejects a connection because connection_count exceeds the maximum allowed, the server will send ER_CON_COUNT_ERROR and immediately close the socket. On macOS, the client's recv() of that error packet sometimes fails with errno EINVAL before the packet is actually delivered, surfacing as CR_SERVER_LOST instead of ER_CON_COUNT_ERROR. Update the tests to account for this possibility.

[svoj]

On macOS, the client's recv() of that error packet sometimes fails with errno EINVAL before the packet is actually delivered, surfacing as CR_SERVER_LOST instead of ER_CON_COUNT_ERROR.

@DaveGosselin-MariaDB and that's a bug, isn't it?

[svoj]

@svoj it seemed like a network stack related issue, especially where both the server and client are running on the same machine. Maybe there's a fast path that the operating system uses in this case? I don't see how we could change anything in our code to mitigate that.

@DaveGosselin-MariaDB missing vio_shutdown(SHUT_RDWR) call? There're not that many cases when the server closes connection. E.g. THD::awake_no_mutex() has this call.

[DaveGosselin-MariaDB]

@svoj it seemed like a network stack related issue, especially where both the server and client are running on the same machine. Maybe there's a fast path that the operating system uses in this case? I don't see how we could change anything in our code to mitigate that.

@DaveGosselin-MariaDB missing vio_shutdown(SHUT_RDWR) call? There're not that many cases when the server closes connection. E.g. THD::awake_no_mutex() has this call.

Thanks @svoj I'll take another look into it.

[svoj]

@DaveGosselin-MariaDB Thank you for taking care of it! It might not be the vio_shutdown() in the end, but the idea still applies: check how KILL CONNECTION works and compare to ER_CON_COUNT_ERROR handling. In the end both can be broken. :)