On Wed, Dec 18, 2024 at 05:35:31PM +0100, Arnd Bergmann wrote:
> On Wed, Dec 18, 2024, at 16:46, Thomas Weißschuh wrote:
> > On Wed, Dec 18, 2024 at 03:08:28PM +0000, Conor Dooley wrote:
> >> On Mon, Dec 16, 2024 at 03:10:03PM +0100, Thomas Weißschuh wrote:
>
> >> Might be a clang thing, allmodconfig with clang doesn't build either.
> >
> > The proposed generic storage infrastructure currently expects that all
> > its users also use HAVE_GENERIC_VDSO.
> > I missed rv32 when checking this assumption.
> >
> > I can add a bunch of ifdefs into the storage code to handle this.
> >
> > Or we re-add the time vDSO functions which were removed in commit
> > d4c08b9776b3 ("riscv: Use latest system call ABI").
> > Today there are upstream ports of musl and glibc which can use them.
> > (currently musl even tries to use __vdso_clock_gettime() as 64-bit only
> > on rv32 due to a copy-and-paste error from its rv64 code)
>
> Adding back __vdso_clock_gettime() wouldn't work on rv32 because there
> is no fallback syscall for it, and it wouldn't really help since
> there is no existing userspace that uses time32 structures.
My original paragraph was worded confusingly.
It was about re-adding time-related vDSO function *in general*, not the
specific 32-bit ones.
The new ones should be 64-bit only, indeed.
> > There is precedence in providing 64bit only vDSO functions, for example
> > __vdso_clock_gettime64() in arm.
> > I do have a small, so far untested, proof-of-concept patch for it.
> > This would even be less code than the ifdefs.
> >
> > What do you think about it?
>
> Yes, simply exposing the normal time64 syscalls through vdso
> should be fine. I think this currently works on everything except
> rv32 and sparc32, probably because neither of them have actual
> users that are able to test.
Should it use the specific _vdso_clock_gettime64() naming or leave out
the 64 suffix?
General Note: I'll continue working on this next year.
Thanks,
Thomas
