Linux Driver Verification

A way to do it without races is like this. Do not first check, then start server with the port found. Just start server at once, and if it fails with "already in use" error, attempt to start it on the next port. This way you'll start a server without any races.

Then the server somehow informs the calling program about the port it's managed to bind to (via standard output, or a pid-based temp file), and the calling program connects to that port.

To make a workaround, you might just sleep for a random time before checking if the port's in use.

I will check the STDERR for a message "address already in use" in case of crash CSD. And then try to run CSD with new address.

Alexandr Strakh wrote:

I will check the STDERR for a message "address already in use" in case of crash CSD. And then try to run CSD with new address.

Ok, but make sure you'll still show other error messages to user. Don't forget to add comments about that!

Just to not forget.

On my way home I realized one important thing, inspired by comment 4 to bug #974. This issue with race condition in address assignment would be much easier to solve if ldv-core was the server, and CSD was the client, not vice versa...

The right solution is proposed above:

Pavel Shved wrote:

if ldv-core was the server, and CSD was the client, not vice versa...

For now, trying to reduce probability of the race, I have committed a patch to select random ports from 8080 to 8099 instead of predefined sequence starting from 8080.

I don't know internals, but can't we "lock" a found address at the same time when it's found? In this case other ldv-core instances will find other addresses...