Re: FreeBSD's Google Summer of Code 2016

[Rich Felker <dalias@libc.org>]

On Sat, Mar 05, 2016 at 05:41:25PM -0500, Pedro Giffuni wrote:
> First of all, great to hear there is interest on the musl side too.
> 
> I think the biggest precedent of porting linux-oriented C libraries
> came from Debian's kFreeBSD. We accomodated a little by for them
> by defining __FreeBSD_kernel__ in sys/param.h.
> 
> While using the optional linux-abi futex in FreeBSD could be an option,
> it is not really the cleanest option. The Debian guys did a port of
> NPTL using regular pthreads:
> 
> http://thread.gmane.org/gmane.linux.debian.ports.bsd/11702
> 
> I am certain this will require more research but it would be useful
> for other ports as well.

Glibc/NPTL has a lot of what I'd call "gratuitous abstraction" (like
the lll stuff) in their pthread primitives which makes this
"possible". I call it gratuitous because it's really really hard to
achieve correct implementations of the pthread sync primitives that
don't have serious corner-case bugs, and it's unlikely that their
abstractions actually suffice to make correct alternate
implementations.

musl does not have any such abstraction. We require a compare-and-swap
operation or equivalent on which arbitrary atomic operations can be
constructed, a futex or equivalent operation that's roughly
while(*addr==expected) sleep(), and implement all the sync primitives
just once on top of these.

In a worst case, futex can be emulated trivially by having wait and
wake both be nops; then everything just spins at 100% cpu load. Of
course this does not give very nice results. It's possible futex could
be emulated better; certainly it can be within a single process where
the futex wait queue could just be implemented in userspace, but
emulating process-shared futexes effectively this way sounds
nontrivial. Since the functionality is already there in the kernel,
though, just exposing it sounds like a much better solution to me.

Rich