On Wed, 2019-01-09 at 11:51 -0500, bfields@xxxxxxxxxxxx wrote:
> On Tue, Jan 08, 2019 at 04:21:40PM +0000, Trond Myklebust wrote:
> > On Tue, 2019-01-08 at 10:01 -0500, bfields@xxxxxxxxxxxx wrote:
> > > On Mon, Jan 07, 2019 at 10:06:19PM +0000, Trond Myklebust wrote:
> > > > On Mon, 2019-01-07 at 16:32 -0500, bfields@xxxxxxxxxxxx wrote:
> > > > > So maybe we actually need
> > > > >
> > > > > static bool (struct svc_xprt *xprt)
> > > > > {
> > > > > + mb();
> > > >
> > > > You would at best need a 'smp_rmb()'. There is nothing to gain
> > > > from
> > > > adding a write barrier here,
> > >
> > > That's not my understanding.
> > >
> > > What we have is basically:
> > >
> > > 1 2
> > > ---- ----
> > > WRITE to A WRITE to B
> > >
> > > READ from A and B READ from A and B
> > >
> > > and we want to guarantee that at least one of those two reads
> > > will
> > > see
> > > both of the writes.
> > >
> > > A read barrier only orders reads with respect to the barrier, it
> > > doesn't
> > > do anything about writes, so doesn't guarantee anything here.
> >
> > In this context 'WRITE to A' and/or 'WRITE to B' are presumably the
> > operations of setting the flag bits in xprt->xpt_flags, no?
>
> Right, or I guess sk_sock->flags, or an atomic operation on
> xpt_reserved
> or xpt_nr_rqsts.
>
> > That's not occurring here, it is occurring elsewhere.
>
> Right. And I hadn't tried to verify whether there were corresponding
> (possibly implicit) write barriers in those places, thanks for doing
> that work:
>
> > The test_and_set_bit(XPT_DATA, &xprt->xpt_flags) in
> > svc_data_ready()
> > performs an explicit barrier, so we shouldn't really care.
>
> OK.
>
> > The other cases where we do set_bit(XPT_DATA) don't matter since
> > the
> > socket has its own locking, and so the XPT_DATA is really just a
> > test
> > for whether or not we need to enqueue the svc_xprt.
>
> I'm not following, apologies.
>
> In any case it's set only on initialization or in recvfrom, and in
> the
> recvfrom case I think the
>
> smp_mb__before_atomic();
> clear_bit(XPT_BUSY, &xprt->xpt_flags);
>
> in svc_xprt_received() provides the necessary write barrier.
>
> But there are some exceptions in the rdma code, in
> svc_rdma_wc_receive
> and svc_rdma_wc_done.
>
> > In the only place where XPT_DEFERRED is set, you have an implicit
> > write
> > barrier (due to a spin_unlock) between the call to set_bit() and
> > the
> > call to svc_xprt_enqueue(), so all data writes are guaranteed to be
> > complete before any attempt to enqueue the socket.
>
> OK.
>
> > I can't see that you really care for the case of XPT_CONN, since
> > the
> > just-created socket isn't going to be visible to other cpus until
> > you've added it to &pool->sp_sockets (which also has implicit write
> > barriers due to spin locks).
> >
> > I don't think you really care for the case of XPT_CLOSE either
> > since
> > svc_delete_xprt() doesn't depend on any other data writes that
> > aren't
> > already protected by spinlocks.
>
> OK. Yes, I'm not worried about XPT_CONN or XPT_CLOSE.
>
> > So the conclusion would be to add smp_rmb() in
> > svc_xprt_has_something_to_do(). No extra write barriers are needed
> > AFAICS.
> > You may still need the READ_ONCE() in order to add a data
> > dependency
> > barrier (i.e. to ensure that alpha processors don't reorder reads
> > of
> > the xpt_flags with other speculative reads). That should reduce to
> > a
> > standard read on all non-alpha architectures.
>
> That looks unnecessary; memory-barriers.txt say "Read memory barriers
> imply data dependency barriers", and later "As of v4.15 of the Linux
> kernel, an smp_read_barrier_depends() was added to READ_ONCE()".
>
The above is stating that
smp_rmb();
smp_read_barrier_depends();
if (xprt->xpt_flags & ....)
is redundant and can be replaced with just
smp_rmb();
if (xprt->xpt_flags & ....)
However that's not the case for smp_rmb() followed by READ_ONCE(). That
would expand to
smp_rmb();
if (xprt->xpt_flags & ...) {
smp_read_barrier_depends();
} else
smp_read_barrier_depends();
which is not redundant. It is ensuring (on alpha only) that the read of
xprt->xpt_flags is also not re-ordered w.r.t. other data reads that
follow.
See, for instance, kernel/events/core.c which has several examples, or
kernel/exit.c.
> I still wonder about:
>
> - the RDMA cases above.
> - svc_xprt_release_slot: no write barrier after writing to
> xprt->xpt_nr_rqsts.
> - svc_reserve: no barrier after writing to xpt_reserved
>
> Also svc_write_space is setting SOCK_NOSPACE and then calling
> svc_xprt_enqueue. I'm pretty sure the sk_write_space method has to
> have
> a write barrier after that, though, so this is OK.
>
> --b.
>
> > > --b.
> > >
> > >
> > >
> > > > and you don't even need a read barrier in
> > > > the non-smp case.
> > > >
> > > > > if (xprt->xpt_flags & ((1<<XPT_CONN)|(1<<XPT_CLOSE)))
> > > > > return true;
> > > > > if (xprt->xpt_flags &
> > > > > ((1<<XPT_DATA)|(1<<XPT_DEFERRED))) {
> > > > >
> > > > > Then whichever memory barrier executes second guarantees that
> > > > > the
> > > > > following check sees the result of both the XPT_DATA and
> > > > > xpt_nr_rqsts
> > > > > changes. I think....
--
Trond Myklebust
Linux NFS client maintainer, Hammerspace
trond.myklebust@xxxxxxxxxxxxxxx
