[Bug 265974] SMR has several missing barriers

From: <bugzilla-noreply_at_freebsd.org>
Date: Sat, 27 Aug 2022 00:13:30 UTC

--- Comment #8 from Mark Johnston <markj@FreeBSD.org> ---
(In reply to Pierre Habouzit from comment #6)
I'm sorry, but I still don't quite see the problem.  I spent some time trying
to write litmus tests to explore interactions between readers, writers, and
smr_poll(), but didn't find anything surprising (yet).  For now I'm ignoring
the deferred and lazy modes of SMR since they're not used in FreeBSD currently.

I note that your litmus tests don't seem to take into account the fact that
smr_advance() issues a atomic_thread_fence_rel() before updating the write
sequence.  (On arm64 this expands to DMB SY, the same as
atomic_thread_fence_seq_cst().)  With smr_enter() patched, both readers and
writers issue a full memory barrier on arm64.

What I don't really understand is why readers need to synchronize with
smr_poll() at all.  Even with the barriers you proposed, it's possible for a
reader to store a snapshot of wr_seq that is older than the global rd_seq. 
This could happen if an interrupt or vmexit occurs between the load of wr_seq
and the store to per-CPU memory.  So long as smr_poll() does not advance rd_seq
beyond that of any CPU actively executing a read section, it's ok, and I don't
see how that can happen.  What am I missing?

Thanks for your patience and for taking the time to report these problems.  I
posted a first patch for review here: https://reviews.freebsd.org/D36370

You are receiving this mail because:
You are the assignee for the bug.