Re: [PATCH v2 01/13] seqnum_ops: Introduce Sequence Number Ops

[Shuah Khan]

On 11/16/20 7:53 AM, Peter Zijlstra wrote:
> On Fri, Nov 13, 2020 at 10:46:03AM -0700, Shuah Khan wrote:
>
> > +Increment interface
> > +-------------------
> > +
> > +Increments sequence number and returns the new value. ::
> > +
> > +        seqnum32_inc_return() --> (u32) atomic_inc_return(seqnum)
> > +        seqnum64_inc_return() --> (u64) atomic64_inc_return(seqnum)
>
> Did you think about the ordering?

Looking at atomic_t.txt _inc_return() can be fully ordered without
loosing or making the intermediate state visible. This is good for
this sequence number use-case. Is there something I am overlooking?

> > +Fetch interface
> > +---------------
> > +
> > +Fetched and returns current sequence number value. ::
> > +
> > +        seqnum32_fetch() --> (u32) atomic_add_return(0, seqnum)
> > +        seqnum64_fetch() --> (u64) atomic64_add_return(0, seqnum)
>
> That's horrible. Please explain how that is not broken garbage.
>
> Per the fact that it is atomic, nothing prevents the counter being
> incremented concurrently. There is no such thing as a 'current' sequence
> number.

Correct. Some usages of this _fecth() in this patch series are for
printing sequence numbers in debug message and others are stats.

I will review the patches in this series and drop the ones that use
read/fetch - the reason being sequence numbers are strictly up counters
and don't need read/fetch.

thanks,
-- Shuah