Upload lock held for longer than necessary about nix

[jkarni]

Yup - am investigating the impact of some changes on build speed!

[roberth]

The remote build feature could use some love. Would you be interested to contribute improvements?

(note that the remote build label is brand new; we used to have only protocol, and even that is fairly new) labels not helpful

[jkarni]

Reducing the timeout to 2 minutes dramatically improved the speed at which builds happen. The curious thing is that the wait was often quite long - much longer than the timeout. This wouldn't make much sense if the timeout was working as (I take it that was) intended. Moreoever, as I mentioned, the process holding the lock often seems to be past the copy stage. So though decreasing the timeout helps, I'm pretty sure there are other things to fix here - things which, if fixed, potentially would obviate the need to reduce the timeout.

In general, I'm not sure I understand yet how signals are working here. It seems a signal handling thread sometimes exist, and I presume it would be responsible for making sure these alarms only go to the right thread, but I also don't see the relevant sigmask.

Using signal (instead of sigaction) is also, it seems, a potential concern. Additionally, the existence of the second alarm is quite mysterious to me.

All of this is my quite untutored thoughts - just writing them here by way of update, and will continue investigating. Though for my own machine my inclination is to just disable this flock logic altogether.