Re: [RFC PATCH 1/1] hpref: Hazard Pointers with Reference Counter

On Wed, Jan 07, 2026 at 09:40:56AM -0500, Mathieu Desnoyers wrote:
> On 2026-01-07 04:38, Jonas Oberhauser wrote:
> > 
> > > liburcu uatomics and "cmm_*" API initially aimed to replicate the LKMM in userspace (note: this was done circa 2009). However, starting from liburcu 0.15, we extended the uatomics API to support the C11 model
> > 
> > I expect the C11-style-semantics will have a lot of friction with RCU (at the least because of missing dependency ordering).
> 
> * RCU publication guarantee:
> 
> C11 has the "consume" MO (which we call CMM_CONSUME in liburcu) for the
> dependency required by rcu_dereference. AFAIK it is implemented as the
> stronger "acquire" MO on most toolchains today though. For those who
> really want to keep using a volatile access for rcu_dereference in
> liburcu for performance reasons, we introduced the
> URCU_DEREFERENCE_USE_VOLATILE define in liburcu. This pairs with a C11
> store-release on rcu_assign_pointer.
> 
> Note that the CONSUME MO with store-release pairing is required to
> let threadsanitizer know about RCU publication guarantee ordering.

Sadly, memory_order_consume is being deprecated in both C++ and C.
Given that (as you say) it was always promoted to memory_order_acquire,
this is not as much a tragedy as it might be.

Yes, I am thinking of replacements, as I have been for more than ten
years, but this is not an easy one.  Lifetime-end pointer zap and OOTA
have higher priority.  So if you have ideas, please do not keep them
secret!

> * RCU grace period guarantee:
> 
> For the read-side critical section (rcu_read_{lock,unlock}) pairing with
> synchronize_rcu, we've updated the liburcu read lock/unlock
> implementation to use either atomic thread fence as full barrier or
> SEQ_CST stores. We've also introduced a urcu/annotate.h header to help
> annotating memory access groups. This helps threadsanitizer track the
> dependency ordering inherent to RCU grace periods.

The atomic thread fence is for threadsanitizer's benefit?  In the MB
rcu_read_lock() and rcu_read_unlock() case, is barrier() or a light-weight
asymmetric fence still used?

> > Things like rmb and wmb may also not have a clean formalization in a C11-style semantics.
> 
> Right. We've been moving away from cmm_smp_rmb() and cmm_smp_wmb() in
> favor of explicit C11 acquire/release dependency chains within the
> liburcu implementation as of liburcu 0.15. This facilitates integration
> with tools like threadsanitizer. The rmb/wmb APIs are still there for
> legacy reasons, but users are encouraged to move to a acquire/release
> model when it makes sense.

Sounds good to me!

							Thanx, Paul