On Wed, Jul 15, 2020 at 07:56:50AM +0200, Ahmed S. Darwish wrote:
> On Wed, Jul 15, 2020 at 10:05:07AM +0800, Leo Yan wrote:
> > From: Peter Zijlstra <pet...@infradead.org>
> >
> ...
> >
> > Provide struct clock_read_data and two (seqcount) helpers so that
> > architectures (arm64 in specific) can expose the numbers to userspace.
> >
> ...
> >
> > +struct clock_read_data *sched_clock_read_begin(unsigned int *seq)
> > +{
> > + *seq = raw_read_seqcount(&cd.seq);
> > + return cd.read_data + (*seq & 1);
> > +}
> > +
> ...
>
> Hmm, this seqcount_t is actually a latch seqcount. I know the original
> code also used raw_read_seqcount(), but while at it, let's use the
> proper read API for seqcount_t latchers: raw_read_seqcount_latch().
Good point. To be honest, I think myself cannot give a good judgement
for memory barrier related thing :)
I read a bit the document for the latch technique [1], comparing to
raw_read_seqcount_latch(), the function raw_read_seqcount() contains
smp_rmb(), IIUC, the *read* memory barrier is used to support for
kcsan.
The usage for smp_rmb() and kcsan flow is like below:
sched_clock_read_begin()
`-> raw_read_seqcount()
`-> smp_rmb()
`-> kcsan_atomic_next(KCSAN_SEQLOCK_REGION_MAX)
sched_clock_read_retry()
`-> read_seqcount_retry()
`-> smp_rmb()
`-> kcsan_atomic_next(0)
So the question is: should we support kcsan or not in this flow?
> raw_read_seqcount_latch() has no read memory barrier though, and a
> suspicious claim that READ_ONCE() pairs with an smp_wmb() (??). But if
> its implementation is wrong, let's fix it there instead.
I don't think we need pair with smp_wmb(), since it's mainly related
with data reading, so a smp_rmb() would be sufficient [2].
Thanks,
Leo
[1]
https://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git/tree/include/linux/seqlock.h?h=v5.8-rc5#n321
[2]
https://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git/tree/include/linux/seqlock.h?h=v5.8-rc5#n373
