Hi,
Alpha uses a generic futex implementation, which causes some problems [1].
I've read through the code, and it appears as if the implementation
could be done by using the ldq_l/stq_c instructions, relatively easy I
might add. I'm definitely interested in implementing this, but first...
I have only a few questions.
1) What are the benefits of a native futex implementation, other than
fixing the glibc test failures?
2) Is there a technical reason it hasn't been implemented on Alpha?
3) Am in correct that it could be done with ldq_l/stq_c instructions?
I'm not experienced with kernel development, but I'd like to learn, so
please bear with me.
Thanks,
Matt Turner
[1] http://bugs.gentoo.org/show_bug.cgi?id=205099
> Alpha uses a generic futex implementation, which causes some problems [1].
Summarizing this: the generic futex implementation leaves
futex_atomic_cmpxchg_inatomic() as a stub that just returns -ENOSYS, and
glibc robust futex code can't handle this return value, which leads to
glibc tests hanging.
> I've read through the code, and it appears as if the implementation
> could be done by using the ldq_l/stq_c instructions, relatively easy I
> might add. I'm definitely interested in implementing this, but first...
>
> I have only a few questions.
>
> 1) What are the benefits of a native futex implementation, other than
> fixing the glibc test failures?
Presumably native implementations can use optimized assembly, which
would be somewhat faster than a generic implementation in C.
> 2) Is there a technical reason it hasn't been implemented on Alpha?
I don't know for sure, but I would guess it's just that no one has cared
enough about optimizing futexes on alpha.
A native implementation for alpha is probably an amusing project, but
also figuring out a way of implementing the missing operations for
generic futexes would be nice too (given that glibc uses them now).
Although a generic futex_atomic_cmpxchg_inatomic() that works for SMP
might be tricky.
- R.
Matt Turner wrote:
> Hi,
>
> Alpha uses a generic futex implementation, which causes some problems [1].
>
> I've read through the code, and it appears as if the implementation
> could be done by using the ldq_l/stq_c instructions, relatively easy I
> might add. I'm definitely interested in implementing this, but first...
>
> I have only a few questions.
>
> 1) What are the benefits of a native futex implementation, other than
> fixing the glibc test failures?
> 2) Is there a technical reason it hasn't been implemented on Alpha?
> 3) Am in correct that it could be done with ldq_l/stq_c instructions?
I think all the futex stuff is 32-bits wide, so ldl_l/stl_c.
You can pretty much just copy the powerpc version, as that
platform also uses load-locked/store-conditional operations.
r~