Re: [PATCH memory-model] docs: memory-barriers: Add note on compiler transformation and address deps

[Paul E. McKenney]

On Fri, Oct 20, 2023 at 11:29:24AM +0200, Jonas Oberhauser wrote:
> 
> Am 10/19/2023 um 6:39 PM schrieb Paul E. McKenney:
> > On Wed, Oct 18, 2023 at 12:11:58PM +0200, Jonas Oberhauser wrote:
> > > Hi Paul,
> > > [...]
> > The compiler is forbidden from inventing pointer comparisons.
> 
> TIL :) Btw, do you remember a discussion where this is clarified? A quick
> search didn't turn up anything.

This was a verbal discussion with Richard Smith at the 2020 C++ Standards
Committee meeting in Prague.  I honestly do not know what standardese
supports this.

> > > Best wishes,
> > > 
> > > jonas
> > > 
> > > Am 10/6/2023 um 6:39 PM schrieb Jonas Oberhauser:
> > > > Hi Paul,
> > > > 
> > > > The "more up-to-date information" makes it sound like (some of) the
> > > > information in this section is out-of-date/no longer valid.
> > The old smp_read_barrier_depends() that these section cover really
> > does no longer exist.
> 
> You mean that they *intend to* cover? smp_read_barrier_depends never appears
> in the text, so anyone reading this section without prior knowledge has no
> way of realizing that this is what the sections are talking about.

It also doesn't appear in the kernel anymore.

> On the other hand the implicit address dependency barriers that do exist are
> mentioned in the text. And that part is still true.

And this relevant discussion is moving to rcu_dereference.rst, and the
current text is just for people who read memory-barriers.txt some time
back and are expecting to find the same information in the same place.

So if there are things that rcu_dereference.rst is missing, they do
need to be added.

> > > > But after reading the sections, it seems the information is valid, but
> > > > discusses mostly the history of address dependency barriers.
> > > > 
> > > > Given that the sepcond part  specifically already starts with a
> > > > disclaimer that this information is purely relevant to people interested
> > > > in history or working on alpha, I think it would make more sense to
> > > > modify things slightly differently.
> > > > 
> > > > Firstly I'd remove the "historical" part in the first section, and add
> > > > two short paragraphs explaining that
> > > > 
> > > > - every marked access implies a address dependency barrier
> > This is covered in rcu_dereference.rst.
> 
> Let me quote a much wiser man than myself here: "
> 
> The problem is that people insist on diving into the middle of documents,
> so sometimes repetition is a necessary form of self defense.  ;-)
> 
> "

;-) ;-) ;-)

> The main reason I would like to add this here at the very top is that
> 
> - this section serves to frigthen children about the dangers of address
> dependencies,
> 
> - never mentions a way to add them - I need to happen to read another
> section of the manual to find that out

Both are now the job of rcu_dereference.rst.

> - and says this information is historical without specifying which parts are
> still relevant

Readers not interested in history should just go to rcu_dereference.rst,
and if pieces are missing from rcu_dereference.rst, they should be
added there.  (Except of course not the historical points that are not
relevant to the current kernel.)

> (and the parts that are still there are all still relevant, while the parts
> that only the authors know was intended to be there and is out-of-date is
> already gone).

The question is instead what parts that are still relevant are missing
from rcu_dereference.rst.

> So I would add a disclaimer specifying that (since 4.15) *all* marked
> accesses imply read dependency barriers which resolve most of the issues
> mentioned in the remainder of the article.
> However, some issues remain because the dependencies that are preserved by
> such barriers are just *semantic* dependencies, and readers should check
> rcu_dereference.rst for examples of what that implies.

Or maybe it is now time to remove those sections from memory-barriers.txt,
leaving only the first section's pointer to rcu_dereference.rst.
It still feels a bit early to me, and I am still trying to figure out
why you care so much about these sections.  ;-)

> > [...]
> > most situations would be better served by an _acquire() suffix than by
> > a relaxed version of [...] an atomic [...]
> 
> I completely agree. I even considered removing address dependencies
> altogether from the company-internal memory models.
> But people sometimes get a little bit angry and start asking many questions.
> The valuable time of the model maintainer should be considered when
> designing memory models.

Yeah, that is always a tough tradeoff, to be sure!

> > > > - address dependencies considered by the model are *semantic*
> > > > dependencies, meaning that a *syntactic* dependency is not sufficient to
> > > > imply ordering; see the rcu file for some examples where compilers can
> > > > elide syntactic dependencies
> > There is a bunch of text in rcu_dereference.rst to this effect.  Or
> > is there some aspect that is missing from that document?
> 
> That's what I meant by "see the rcu file" --- include a link to
> rcu_dereference.rst in that paragraph.
> So that people know to check out rcu_dereference.rst for more explanations
> to this effect.

You mean this paragraph?

 (2) Address-dependency barriers (historical).
     [!] This section is marked as HISTORICAL: For more up-to-date
     information, including how compiler transformations related to pointer
     comparisons can sometimes cause problems, see
     Documentation/RCU/rcu_dereference.rst.

If so, that last line is intended to be the required link.

Or am I looking in the wrong place?

> > The longer-term direction, perhaps a few years from now, is for the
> > first section to simply reference rcu_dereference.rst and for the second
> > section to be removed completely.
> 
> Sounds good to me, but that doesn't mean we need to compromise the
> readability in the interim :)

Some compromise is needed for people that read the document some time
back and are looking for something specific.

> > [...]
> > The problem is that people insist on diving into the middle of documents,
> > so sometimes repetition is a necessary form of self defense.  ;-)
> > 
> > But I very much appreciate your review and feedback, and I also apologize
> > for my slowness.
> 
> Thanks for the response, I started thinking my mails aren't getting through
> again.

Again, apologies!

							Thanx, Paul