Strictly speaking, all that the rcu_read_lock()/rcu_read_unlock() does
is to ensure that any RCU-protected linked data objects referenced by
the corresponding critical section remain in existence for the full
duration of that critical section.
And even that overstates things a bit, as this describes not the
underlying RCU API itself, but rather some of that API's use cases.
(Though to be fair, these are the most popular of RCU's use cases.)
So even more strictly speaking, all the rcu_read_lock()/rcu_read_unlock()
does is to make any synchronize_rcu() or call_rcu() invocations starting
after a given rcu_read_lock() wait (synchronously or asynchronously,
respectively) until the corresponding rcu_read_unlock() is reached.
Returning back to the first paragraph, additional protections can be
arranged, depending on the RCU use case. For example, if the non-pointer
fields of objects added to an RCU-protected linked data structure remain
unchanged while that object is accessible to readers, then the protection
includes not just existence, but also value. And there are other use
cases where those values might change while accessible to readers.
So the most accurate answer to your question is "That is a design
choice, based on the RCU use case in question."
My guess is that you are thinking in terms of designs where the
non-pointer fields of an object are constant while that object is
accessible to readers. Is my guess correct?
Thanx, Paul
On Wed, Dec 22, 2021 at 04:35:24PM -0500, Elad Lahav wrote:
> I will try to come up with something, but first I wanted to get your
> opinion as to whether this design pattern, where you need to
> dereference a pointer to a data structure containing all of the data
> considered to be protected by the critical section, is inherent to
> RCU, or is it just an idiosyncrasy of the implementation used in the
> Linux kernel?
>
> --Elad
>
> On Wed, 22 Dec 2021 at 13:19, Paul E. McKenney <paulmck@xxxxxxxxxx> wrote:
> >
> > On Wed, Dec 22, 2021 at 11:35:02PM +0900, Akira Yokosawa wrote:
> > > Hi,
> > >
> > > On Wed, 22 Dec 2021 07:15:54 -0500, Elad Lahav wrote:
> > > > Hi Akira,
> > > >
> > > >> I think this is mostly covered in Section 9.5.2.1 "Publish-Subscribe
> > > >> Mechanism" and Figure 9.10 "Publication/Subscription Constraints".
> > > >
> > > > I don't believe it is. I think you can infer that from reading sections 9.5
> > > > and chapter 15, but it is never made explicit.
> > >
> > > I agree it is not explicit.
> > >
> > > >
> > > > Sub-section 9.5.2.1 talks about the use of rcu_dereference() to obtain a
> > > > pointer, but it doesn't go very deep into how that's implemented. You would
> > > > think that the writer's use of store-release semantics would necessitate
> > > > the reader's use of load-acquire semantics, but the discussion here, the
> > > > previous examples, and a quick inspection of how rcu_dereference() is
> > > > implemented, suggest that a READ_ONCE() is sufficient (or some less
> > > > Linux-kernel-specific equivalent).
> > > >
> > > > If I am a naive reader (and a lazy one, who haven't read chapter 15 and made
> > > > the necessary connection), I could write something like:
> > > >
> > > > struct foo_s {
> > > > int a;
> > > > } foo;
> > > > int b;
> > > > struct foo_s *g_foop = &foo;
> > > >
> > > > // Reader
> > > > rcu_read_lock();
> > > > struct foo *l_foop = rcu_dereference(g_foop);
> > > > use(l_foop->a);
> > > > use(b);
> > > > rcu_read_unlock();
> > > >
> > > > // Writer
> > > > struct foo *l_foop = malloc(sizeof(struct foo));
> > > > l_foop->a = 1;
> > > > b = 2;
> > > > // Release semantics ensure previous stores are observed
> > > > rcu_assign_pointer(g_foop, l_foop);
> > > >
> > > > But since b has no address dependency to g_foop and since neither
> > > > rcu_read_lock() nor rcu_dereference() impose acquire semantics, then
> > > > the reader may not observe the new value of b.
> > >
> > > No, it may not.
> > > So what you want is the mention of "address dependency" somewhere in
> > > Section 9.5.2.1?
> > >
> > > >
> > > > Again, I may be missing something, but this seems to be a major point
> > > > that needs to be explained and emphasized.
> > >
> > > I guess Paul is intentionally avoiding discussions on memory ordering
> > > here in Section 9.5.
> > >
> > > Such a discussion would easily frighten naive readers...
> > >
> > > Anyway, I guess Paul would have some good compromise to satisfy you.
> >
> > Actually, I am going to ask Elad to propose the location and wording of an
> > addition to Section 9.5 covering this, so that we can discuss and refine
> > it. This addition might be a new paragraph, a footnote, a quick quiz,
> > a citation, or what have you. And the refining might switch back and
> > forth among these options a few times. But we do have to start somewhere.
> >
> > I am thinking in terms of this going in after the release that I was
> > naively planning to do yesterday. (The objective universe had other
> > ideas.)
> >
> > > >> BTW, I couldn't figure out what you meant by "the Spanish
> > > >> Acquisition"...
> > > > It's a reference to an old Monty Python skit. Apologies for the silly pun...
> > >
> > > Ah, now I guess I see the pun of "acquire" and "acquisition". ;-)
> >
> > And here I was hoping that Elad was purchasing a house in Barcelona
> > or some such. ;-)
> >
> > (Sorry, couldn't resist!)
> >
> > Thanx, Paul
