Re: [RFC PATCH] fs: elide the smp_rmb fence in fd_install()

[Paul E. McKenney]

On Thu, Dec 05, 2024 at 11:26:35AM -0800, Linus Torvalds wrote:
> On Thu, 5 Dec 2024 at 10:41, Paul E. McKenney <paulmck@xxxxxxxxxx> wrote:
> >
> > > To my understanding this is the idiomatic way of spelling out the
> > > non-existent in Linux smp_consume_load, for the resize_in_progress
> > > flag.
> >
> > In Linus, "smp_consume_load()" is named rcu_dereference().
> 
> Linux.

One of those days...  ;-)

> But yes and no.
> 
> It's worth making it really really clear that "rcu_dereference()" is
> *not* just a different name for some "smp_consume_load()" operation.
> 
> Why? Because a true smp_consume_load() would work with any random kind
> of flags etc. And rcu_dereference() works only because it's a pointer,
> and there's an inherent data dependency to what the result points to.
> 
> Paul obviously knows this, but let's make it very clear in this
> discussion, because if somebody decided "I want a smp_consume_load(),
> and I'll use rcu_dereference() to do that", the end result would
> simply not work for arbitrary data, like a flags field or something,
> where comparing it against a value will only result in a control
> dependency, not an actual data dependency.

Fair points!

And Linus (and Linux, for that matter) equally obviously already knows
this, but please note also that an smp_load_consume() would still order
only later dereferences of the thing returned from smp_load_consume(),
which means that it pretty much needs to be a pointer.  (Yes, in theory,
it could be an array index, but in practice compilers know way too much
about integer arithmetic for this to be advisable.)

							Thanx, Paul