Re: [PATCH] kcsan, seqlock: Support seqcount_latch_t

[Marco Elver <elver@google.com>]

On Tue, 29 Oct 2024 at 12:49, Peter Zijlstra <peterz@infradead.org> wrote:
>
> On Tue, Oct 29, 2024 at 09:36:29AM +0100, Marco Elver wrote:
> > Reviewing current raw_write_seqcount_latch() callers, the most common
> > patterns involve only few memory accesses, either a single plain C
> > assignment, or memcpy;
>
> Then I assume you've encountered latch_tree_{insert,erase}() in your
> travels, right?

Oops. That once certainly exceeds the "8 memory accesses".

> Also, I note that update_clock_read_data() seems to do things
> 'backwards' and will completely elide your proposed annotation.

Hmm, for the first access, yes. This particular oddity could be
"fixed" by surrounding the accesses by
kcsan_nestable_atomic_begin/end(). I don't know if it warrants adding
a raw_write_seqcount_latch_begin().

Preferences?

> > therefore, the value of 8 memory accesses after
> > raw_write_seqcount_latch() is chosen to (a) avoid most false positives,
> > and (b) avoid excessive number of false negatives (due to inadvertently
> > declaring most accesses in the proximity of update_fast_timekeeper() as
> > "atomic").
>
> The above latch'ed RB-trees can certainly exceed this magical number 8.
>
> > Reported-by: Alexander Potapenko <glider@google.com>
> > Tested-by: Alexander Potapenko <glider@google.com>
> > Fixes: 88ecd153be95 ("seqlock, kcsan: Add annotations for KCSAN")
> > Signed-off-by: Marco Elver <elver@google.com>
> > ---
> >  include/linux/seqlock.h | 9 +++++++++
> >  1 file changed, 9 insertions(+)
> >
> > diff --git a/include/linux/seqlock.h b/include/linux/seqlock.h
> > index fffeb754880f..e24cf144276e 100644
> > --- a/include/linux/seqlock.h
> > +++ b/include/linux/seqlock.h
> > @@ -614,6 +614,7 @@ typedef struct {
> >   */
> >  static __always_inline unsigned raw_read_seqcount_latch(const seqcount_latch_t *s)
> >  {
> > +     kcsan_atomic_next(KCSAN_SEQLOCK_REGION_MAX);
> >       /*
> >        * Pairs with the first smp_wmb() in raw_write_seqcount_latch().
> >        * Due to the dependent load, a full smp_rmb() is not needed.
> > @@ -631,6 +632,7 @@ static __always_inline unsigned raw_read_seqcount_latch(const seqcount_latch_t *
> >  static __always_inline int
> >  raw_read_seqcount_latch_retry(const seqcount_latch_t *s, unsigned start)
> >  {
> > +     kcsan_atomic_next(0);
> >       smp_rmb();
> >       return unlikely(READ_ONCE(s->seqcount.sequence) != start);
> >  }
> > @@ -721,6 +723,13 @@ static inline void raw_write_seqcount_latch(seqcount_latch_t *s)
> >       smp_wmb();      /* prior stores before incrementing "sequence" */
> >       s->seqcount.sequence++;
> >       smp_wmb();      /* increment "sequence" before following stores */
> > +
> > +     /*
> > +      * Latch writers do not have a well-defined critical section, but to
> > +      * avoid most false positives, at the cost of false negatives, assume
> > +      * the next few memory accesses belong to the latch writer.
> > +      */
> > +     kcsan_atomic_next(8);
> >  }
>
> Given there are so very few latch users, would it make sense to
> introduce a raw_write_seqcount_latch_end() callback that does
> kcsan_atomic_next(0) ? -- or something along those lines? Then you won't
> have to assume such a small number.

That's something I considered, but thought I'd try the unintrusive
version first. But since you proposed it here, I'd much prefer that,
too. ;-)
Let me try that.

Thanks,
-- Marco