Re: [PATCH 01/28] hazptr: Implement Hazard Pointers

[Boqun Feng <boqun@kernel.org>]

On Sat, Sep 19, 2026 at 01:09:15PM -0400, Mathieu Desnoyers wrote:
> On 2026-09-19 13:00, Linus Torvalds wrote:
> > On Sat, 19 Sept 2026 at 09:35, Bradley Morgan <brads@mainlining.org> wrote:
> > > 
> > > I think hazard pointers are good, what test do YOU suggest we do here?
> > 
> > I want to see a single real-world example of "look, this speeds this
> > real load up by 10%, and the kernel code was actually cleaned up in
> > the process because hazard pointers are great".
> > 
> > Not a microbenchmark that tests just the hazard pointers themselves,
> > but a real kernel feature that has been converted to hazard pointers,
> > and in the process actually shows improvement.
> > 
> > The ONLY reason for hazard pointers to ever be merged is if they
> > actually buy us something real.
> > 
> > So I want to see that 'real" thing.
> AFAIR, Boqun wanted to use hazard pointers to cleanup/speed up an

Right, that's why I send this series:

	https://lore.kernel.org/lkml/20250625031101.12555-1-boqun.feng@gmail.com/

The gist of that work is basically:

	On my system (a 96-cpu VMs), the results of:

		time /usr/sbin/tc qdisc replace dev eth0 root handle 0x1: mq

	are (with lockdep enabled):

		(without the patchset, i.e. using RCU)
		real    0m1.039s
		user    0m0.001s
		sys     0m0.069s

		(with the patchset, i.e. using hazptr)
		real    0m0.053s
		user    0m0.000s
		sys     0m0.051s

i.e. almost 20x speed-up.

One important thing that I want to point out is in that series, I
avoided the busy-waiting in hazptr_synchronize() and made multiple
hazptr_synchronize()s share the same scan. And I do want to see this in
the new code. But unfortunately with the new implementation, we don't
have that part yet. And that's what holds me from trying lockdep
integration for this new implementation.

I could have improved my skill of time management, because I know at
certain point I said "I will finish the scan thread work for your
implementation", but it'll be helpful if you or someone can help get
that done.

> hot lockdep reclaim path. Boqun, Paul, how is this effort going ?
> 
> I suspect we should wait until that lockdep user of hazptr is ready for
> upstreaming and propose both at the same time, because a synchronization
> infrastructure without any significant in tree user is not really
> relevant, right ?
> 

The other thing we could also do is what I did in shazptr, getting the
numbers with rcuscale, that can tell use the actual waiting time for a
hazptr_synchronize().

Regards,
Boqun

> Thanks,
> 
> Mathieu
> 
> -- 
> Mathieu Desnoyers
> EfficiOS Inc.
> https://www.efficios.com