while dependencies are missing
Following landry's remark, also take packages to build out of the race
if some RUN_DEPENDS are going to be ignored. Only do it right before we
put the package in the queue, so that the test is run exactly once per
package instead of during every scan.
I was also worried about multi-packages, but this only takes one fullpkgpath
out of the loop, and "make package" is going to fail half-way through anyways.
happening while some 'unjunk' ports are building.
this is a work-around for a known problem with cmake and qt4 include
dependency handlers...
Also, cache fullpkgpath while building a job, as this contributes for
a large part to the speed (not) of the display when building lots of
ports.
Yields more accurate stat output, where we can see a staircase effect on
the queue when the engine is run not that often, and Built actually go
up until it goes back down again.
duh. put "live" affinity markers while we're building stuff.
We don't re-read the on-disk markers outside of restart (should we ?)
so they HAVE to be in the pkgpath struct proper.
tweak the algorithm slightly (since we forget the old check_interval).
In particular, never keep old times in reserve.
Makes for simpler and clearer reading
Most things will move as a result of {install} changes.
Things that move because of EXTRA depends that are satisfied are unlikely to trigger
further changes.
So, stagger changes for "normal" tobuild -> queue first, and extra depends later.
ago, don't run it again right now.
This prevents the Engine from busy-checking the same data when it's building
lots of small packages very fast.
(50 and 30 may need some fine-tuning)
- affinity info is similar to locks, but with a completely different
lifetime.
- streamline the main loop of the engine, so that it can do two passes:
first pass shuns paths with the wrong affinity. If no good path is found,
those are considered during the second pass.
- make the Core factory aware of what hosts might be running, so that
affinity info for machines removed from a config file will be ignored.
thanks to landry@ for a few tests.
zap all subpackages from the queue.
need to distinguish the normal case (is_done) with a simpler case
(is_done_quick) that's used when scanning the ports tree. Otherwise, we
would check multi-packages loads of time.
was built already... so things will go straight to B/I instead of
getting through T/Q...
less noise in engine.log, and much more accurate information for choosing
queue items.
re-enable wait_timeout on localhost temporarily (should be done in another
way, most probably by checking whether repo is on nfs, we can steal code
from VStat.pm)
pass umask through ssh. This took us long enough to figure out, and it's
considerably simpler than tweaking every login class once again.
options on the command line now define *defaults* that host files can
override (for instance -j, stuck, -p, -J).
Add -p /n to mean "take number of jobs, if >1, divide by n, round up to 2,
and use that for parallel.
Document -p.
Make junk be 'by host' (and it's a prop, so you can tweak it).
concurrent log that records how many jobs are running each time it changes.
tag parallel builds *n in the time record.
