This prevents a race where multiple threads see that machine X is
missing path P, and start sending it concurrently. Nix handles this
correctly, but it's still wasteful (especially for the case where P ==
GHC).
A more refined scheme would be to have per machine, per path locks.
Derivations with "preferLocalBuild = true" can now be executed on
specific machines (typically localhost) by setting the mandary system
features field to include "local". For example:
localhost x86_64-linux,i686-linux - 10 100 - local
says that "localhost" can *only* do builds with "preferLocalBuild =
true". The speed factor of 100 will make the machine almost always win
over other machines.
Otherwise we never recover from reset daemon connections, e.g.
hydra-queue-runner[16106]: while loading build 599369: cannot start daemon worker: reading from file: Connection reset by peer
hydra-queue-runner[16106]: while loading build 599236: writing to file: Broken pipe
...
The error is now handled queueMonitor(), causing the next call to
queueMonitorLoop() to create a new connection.
Having a hundred threads doing I/O at the same time is bad on magnetic
disks because of the excessive disk seeks. So allow only 4 threads to
copy closures in parallel.
While sorting machines by load, the load of a machine
(machine->currentJobs) can be changed by other threads. If that
happens, the comparator is no longer a proper ordering, in which case
std::sort() can segfault. So we now make a copy of currentJobs before
sorting.
There is a slight possibility that the queue monitor and a builder
thread simultaneously decide to mark a build as finished. That's fine,
as long as we ensure the DB update is idempotent (as ensured by doing
"update Builds set finished = 1 ... where finished = 0").
