This was really shortly mentioned on undeadly, because it probably deserves
a separate announcement and article.
First, I want to really thank robert@ again for the organization, and putting
up with rude OpenBSD french developers as he has... plus the people who donated
enough to make these kind of events possible.
Also, my laptop died 3 weeks before the hackathon, and I got a new one,
thanks to project money. It's not really the most expensive laptop you've
ever seen, but it has a dual-core...
Now for some background. I've been maintaining OpenBSD's make for a long time
(over 8 years, I think). It started as simple bug-fixes, then speed-ups, then
more radical clean-ups.
About one year ago, I had a kind of epiphany (yeah, I'm a fan of Angel): I
realized that this code is really atrocious, and instead of fixing bugs,
I started replacing big chunks of it. Not to disparage the guy who wrote
pmake in the first place, as he had very different constraints and goals,
but it is painfully obvious this is a half-finished research project, and
not an industry-standard POSIX make.
So, I started cutting stuff that no-one uses, and options that simply don't
work, to try to make sense of the beast. And I changed algorithms. Most
specifically, I streamlined the suffix handling, and I killed all the
remote job handling.
To make sense of make, you've got to realize there are basically two beasts
folded into one: make in `compat' mode uses its own engine to figure out
which targets to compute first, and its own job runner. The engine is
rather simple, since it doesn't have to queue things up, and can just
run commands. The parallel engine is a bit more complex, since it tries
to explore more of the tree to start up several jobs at once. It also has
an interesting idea: it tries to create shell scripts that agregate commands
to minimize the number of processes created. Unfortunately, THIS is a bad
idea, in modern times, since POSIX mandates separate commands must be run
by separate jobs. As a result, things that work with standard sequential
make no longer run with parallel make.
A few months ago, I started designing a way to overcome those issues. Mostly,
I wanted to get rid of the shell script creation in the parallel make case.
I realized I could have a `tail-call optimization': if I fork() a job to
compute a target, and then fork/exec each command separately, I would not
need to fork() the last command, thus optimizing the really common case
that uses one command per target.
Enter p2k7, with a tall goal: try to make make -j usable.
This was an ideal setting: I had a week mostly empty of other contingencies,
and a few people motivated to give me feedback. So I started merging
the engines, and killing old make code (all the stuff that was building
shell scripts). Pretty soon, I ran into debug issues: the output was really
mangled, and unusable.
make -j uses pipes to separate the outputs from various jobs, and tries to
print stuff line by line. I realized it was keeping a lot of fds open!
whereas it should close about half of them (this explained why cnst@ had
run into the allocation bug that fast... make was gobbling file descriptors
like candy), and I also realized I could make things better by using
non-blocking fds and applying a greedy approach: try to get as many complete
lines/buffers from one fd before getting to the rest.
The result was immensely satisfying: instead of having chunks of intermixed
outputs, suddenly, very long linking lines were appearing as one single line
(since make's internal buffers are much smaller than a full pipe kernel buffer,
this means that, in most case, all the job output was already there).
The devil lies in the details, as usual. On tuesday, my src/ build was
stopping in make clean, in the middle of gnu/binutils. And my error
messages were worth shit: make was basically telling: `oh, btw, there
was an error in one job. Here, figure out what's going on'.
So I revamped the error messages on wednesday morning, and started polishing
other stuff, like duplicating pipes so that stdout and stderr do not get mixed
up.
I finally figured out what was going on: turns out make clean was running stuff
like:
distclean:
-rm somefile
and THIS rm was exiting with an error, as somefile didn't exit yet. And, yes,
thanks to my `optimization' I didn't catch the error: make was seeing the job
having an issue. So, back to the drawing board, let's optimize unless we
can't... and we cannot really decide beforehand, because this - can quite
well come from a make variable, in which case, we'll notice `just in time'.
So, I got a version that went past this problem in a few minutes, and got
it cleaned up properly the day after.
After that, I ran into quite a few more issues in src. Let's say that it's
not yet ready for parallel build. Together with miod@ (who was playing with
make remotely), we fixed a few of them... let's just say that what remains
is the hard ones... it will
