Date: Fri, 24 May 2024 10:14:25 -0400
From: Alan Stern <stern@...land.harvard.edu>
To: Boqun Feng <boqun.feng@...il.com>
Cc: Hernan Ponce de Leon <hernan.poncedeleon@...weicloud.com>,
  Jonas Oberhauser <jonas.oberhauser@...weicloud.com>,
  "Paul E. McKenney" <paulmck@...nel.org>, linux-kernel@...r.kernel.org,
  linux-arch@...r.kernel.org, kernel-team@...a.com, parri.andrea@...il.com,
  j.alglave@....ac.uk, luc.maranget@...ia.fr,
  Joel Fernandes <joel@...lfernandes.org>
Subject: Re: LKMM: Making RMW barriers explicit

On Thu, May 23, 2024 at 07:50:11PM -0700, Boqun Feng wrote:
> On Thu, May 23, 2024 at 09:38:05PM -0400, Alan Stern wrote:
> > On Thu, May 23, 2024 at 08:14:38AM -0700, Boqun Feng wrote:
> > > Besides, I'm not sure this is a good idea. Because the "{mb}, {once},
> > > etc" part is a syntax thing, you write a cmpxchg(), it should be
> > > translated to a cmpxchg event with MB tag on. As to failed cmpxchg()
> > > doesn't provide ordering, it's a semantics thing, as Jonas showed that
> > > it can be represent in cat file. As long as it's a semanitc thing and we
> > > can represent in cat file, I don't think we want herd to give a special
> > > treatment.
> > 
> > I don't really understand the distinction you're making between 
> > syntactic things and semantic things.  For most instructions there's no 
> 
> Syntax is how the code is written, and semantic is how the code is
> executed (in each execution candidate). So if we write a cmpxchg{mb}(),
> and in execution candiates, it could generates a read{MB} event and a
> write{MB} event (succeed case), or a read{MB} event (fail case), "{MB}"
> here doesn't mean it's a full barrier, it only means the event comes
> from a no suffix API. Here "{MB}" only has syntactic meaning (no
> semantic meaning).

Okay, I get it.  Then you might agree that it probably would be better 
to use a different tag here, because the mb tag is already in use with 
other instructions (like smp_mb()) where it does always mean there's a 
full barrier.

> Not really, RMW events contains all events generated from
> read-modify-write primitives, if there is an R event without a rmw
> relation (i.e there is no corresponding write event), it's a failed RWM
> by definition: it cannot be anything else.

Not true.  An R event without an rmw relation could be a READ_ONCE().  
Or a plain read.  The memory model uses the tag to distinguish these 
cases.

> > that it would work is merely an artifact of herd7's internal actions.  I 
> > think it would be much cleaner if herd7 represented these events in some 
> > other way, particularly if we can tell it how.
> > 
> > After all, herd7 already does generate different sets of events for 
> > successful (both R and W) and failed (only R) RMWs.  It's not such a big 
> > stretch to make the R events it generates different in the two cases.
> > 
> 
> I thought we want to simplify the herd internal processing by avoid
> adding mb events in cmpxchg() cases, in the same spirit, if syntactic
> tagging is already good enough, why do we want to make herd complicate?

Herd7 already is complicated by the fact that it needs to handle 
cmpxchg() instructions in two ways: success and failure.  This 
complication is unavoidable.  Adding one extra layer (different tags for 
the different ways) is an insignificant increase in the complication, 
IMO. And it's a net reduction when you compare it to the amount of 
complication currently in the herd7 code.

Also what about cmpxchg_acquire()?  If it fails, it will generate an R 
event with an acquire tag not in the rmw relation.  There is no way to 
tell such events apart from a normal smp_load_acquire(), and so the .cat 
file would have no way to know that the event should not have acquire 
semantics.  I guess we would have to rename this tag, as well.

Alan

Powered by blists - more mailing lists