Le Wed, Jun 25, 2025 at 08:24:53AM -0700, Boqun Feng a écrit :
> On Wed, Jun 25, 2025 at 03:56:05PM +0200, Frederic Weisbecker wrote:
> > Le Tue, Jun 24, 2025 at 08:10:57PM -0700, Boqun Feng a écrit :
> > > +static void synchronize_shazptr_normal(void *ptr)
> > > +{
> > > + int cpu;
> > > + unsigned long blocking_grp_mask = 0;
> > > +
> > > + smp_mb(); /* Synchronize with the smp_mb() in shazptr_acquire(). */
> > > +
> > > + for_each_possible_cpu(cpu) {
> > > + void **slot = per_cpu_ptr(&shazptr_slots, cpu);
> > > + void *val;
> > > +
> > > + /* Pair with smp_store_release() in shazptr_clear(). */
> > > + val = smp_load_acquire(slot);
> > > +
> > > + if (val == ptr || val == SHAZPTR_WILDCARD)
> > > + blocking_grp_mask |= 1UL << (cpu /
> > > shazptr_scan.cpu_grp_size);
> > > + }
> > > +
> > > + /* Found blocking slots, prepare to wait. */
> > > + if (blocking_grp_mask) {
> >
> > synchronize_rcu() here would be enough since all users have preemption
> > disabled.
> > But I guess this defeats the performance purpose? (If so this might need a
> > comment somewhere).
> >
>
> synchronize_shazptr_normal() cannot wait for a whole grace period,
> because the point of hazard pointers is to avoid waiting for unrelated
> readers.
Fair enough!
>
> > I guess blocking_grp_mask is to avoid allocating a cpumask (again for
> > performance purpose? So I guess synchronize_shazptr_normal() has some perf
>
> If we are talking about {k,v}malloc allocation:
> synchronize_shazptr_normal() would mostly be used in cleanup/free path
> similar to synchronize_rcu(), therefor I would like to avoid "allocating
> memory to free memory".
Good point!
>
> > expectations?)
> >
> > One possibility is to have the ptr contained in:
> >
> > struct hazptr {
> > void *ptr;
> > struct cpumask scan_mask
> > };
> >
>
> You mean updaters passing a `struct hazptr *` into
> synchronize_shazptr_normal()? That may be a good idea, if multiple
> updaters can share the same `struct hazptr *`, we can add that later,
> but...
>
> > And then the caller could simply scan itself those remaining CPUs without
> > relying on the kthread.
>
> .. this is a bad idea, sure, we can always burn some CPU time to scan,
> but local optimization doesn't mean global optimization, if in the
> future, we have a lots of synchronize_shazptr_normal()s happening at
> the same time, the self busy-waiting scan would become problematic.
Ok.
Thanks.
--
Frederic Weisbecker
SUSE Labs
