ghc threaded runtime gotchas
That's a seemingly simple question I started asking various people two
weeks ago. I didn't get many useful answers, but now I have experience
doing it myself, and so here's a blog post brain dump.
I have been trying to convert git-annex to use GHC's threaded runtime, for
a variety of reasons. Naively adding the
-threaded option resulted in a
git-annex command that seemed to randomly freeze, but only sometimes
(and, infuriatingly, never when I straced it), and a test suite that froze
at a random point almost every time I ran it. Not a good result, and
lacking any knowledge about gotchas with using the threaded runtime, I was
at a loss for a long time (most of DebConf) about how to fix it.
I now know of at least three classes of problems that enabling the threaded
runtime can turn up in programs that have been developed using the
accessing a MVar after forkprocess can hang
MissingH has some code similar to this, which works ok with the
forkProcess $ do
debugM "about to run some command"
In the above example,
debugM accesses a
MVar. Doing that after
forkProcess can result in a MVar deadlock, as it tries to access a MVar
value, that is, apparently, not accessible to the forked process.
(Bug report with test case)
System.Cmd.Utils from MissingH is asking for trouble.
I switched all my code to the newer and, apparently, threaded runtime
forkProcess is a massively bad idea
Even when not accessing a MVar after
forkProcess, it's very unsafe to
use. It's easy to miss the warning attached to forkProcess, when the code
seems to work. But with the threaded runtime, I've found that most
any call to
forkProcess will sometimes succeed, and sometimes freeze
the whole program. This might only happen around 1 time in 1000.
Then you'll find this warning and do a lot of head-scratching about what
it really means:
forkProcess comes with a giant warning: since any other running threads are
not copied into the child process, it's easy to go wrong: e.g. by accessing
some shared resource that was held by another thread in the parent.
The hangs I saw could be due to laziness issues deferring code to run
forkProcess that you'd expect to have run before it ... or
who knows what else.
It's not clear to me that it's possible to use
forkProcess safely in
Haskell code. I think it's notable that
System.Process runs the whole
fork/exec in C code instead.
unsafe FFI calls block
According to most of the documentation you'll find in eg, the Haskell wiki,
Real World Haskell, etc, the only difference between the
unsafe imports in the FFI is that
unsafe is faster, and shouldn't be
used for C code that calls back into Haskell code.
But the documentation is out of date. Actually, if you're using the FFI,
and the foreign function can block, you need to use
safe. When using
unsafe, a blocking foreign function can block all threads of the program.
In my case, I was using
kqueue to wait for changes to files, and this
indeed blocked my whole program when linked with
-threaded. Marking it
safe fixed this.
The details are well described in this paper:
Somewhat surprisingly, this blocking only happens when using the threaded
runtime. If you're using the non-threaded runtime with
FFI functions, your other pseudo-threads won't be blocked. This is because
the non-threaded runtime has an SIGALARM timer that interrupts (most)
blocking system calls. This leads to other troubles of its own (like
needing to restart interrupted FFI functions, or blocking the other
pseudo-threads from running if the C code ignores SIGALARM), but that's
offtopic for this post.
Converting a large Haskell code base from the default, non-threaded runtime
to the threaded runtime can be quite tricky. None of the problems are the
sort of thing that Haskell helps to manage either. I will be developing new
programs using the threaded runtime from the beginning from now on.
By the way, don't take this post to say that threading in Haskell sucks.
I've really been enjoying writing threaded Haskell code. The control
Haskell gives over isolation of state to threads, and the excellent and
wide-ranging suite of thread communications data types (
SampleVar, etc) have made developing a complex threaded
program something I feel comfortable doing for the first time, in any
Syndicated 2012-07-20 19:31:37 from see shy jo