Re: [RFC PATCH v1] mm: improve call_controls_lock

[JaeJoon Jung]

On Wed, 31 Dec 2025 at 13:59, SeongJae Park <sj@xxxxxxxxxx> wrote:
>
> On Wed, 31 Dec 2025 11:15:00 +0900 JaeJoon Jung <rgbi3307@xxxxxxxxx> wrote:
>
> > On Tue, 30 Dec 2025 at 00:23, SeongJae Park <sj@xxxxxxxxxx> wrote:
> > >
> > > Hello Asier,
> > >
> > >
> > > Thank you for sending this patch!
> > >
> > > On Mon, 29 Dec 2025 14:55:32 +0000 Asier Gutierrez <gutierrez.asier@xxxxxxxxxxxxxxxxxxx> wrote:
> > >
> > > > This is a minor patch set for a call_controls_lock synchronization improvement.
> > >
> > > Please break description lines to not exceed 75 characters per line.
> > >
> > > >
> > > > Spinlocks are faster than mutexes, even when the mutex takes the fast
> > > > path. Hence, this patch replaces the mutex call_controls_lock with a spinlock.
> > >
> > > But call_controls_lock is not being used on performance critical part.
> > > Actually, most of DAMON code is not performance critical.  I really appreciate
> > > your patch, but I have to say I don't think this change is really needed now.
> > > Please let me know if I'm missing something.
> >
> > Paradoxically, when it comes to locking, spin_lock is better than
> > mutex_lock
> > because "most of DAMON code is not performance critical."
> >
> > DAMON code only accesses the ctx belonging to kdamond itself. For
> > example:
> > kdamond.0 --> ctx.0
> > kdamond.1 --> ctx.1
> > kdamond.2 --> ctx.2
> > kdamond.# --> ctx.#
> >
> > There is no cross-approach as shown below:
> > kdamond.0 --> ctx.1
> > kdamond.1 --> ctx.2
> > kdamond.2 --> ctx.0
> >
> > Only the data belonging to kdamond needs to be resolved for concurrent access.
> > most DAMON code needs to lock/unlock briefly when add/del linked
> > lists,
> > so spin_lock is effective.
>
> I don't disagree this.  Both spinlock and mutex effectively work for DAMON's
> locking usages.
>
> > If you handle it with a mutex, it becomes
> > more
> > complicated because the rescheduling occurs as a context switch occurs
> > inside the kernel.
>
> Can you please elaborate what kind of complexities you are saying about?
> Adding some examples would be nice.

You probably know better than I do.  What I'm saying is too general.
spin_lock is good for short collision waits, while mutex is more efficient
for longer ones.  However, I'm saying that mutexes place a burden on the
kernel because they schedule internally.

>
> > Moreover, since the call_controls_lock that is
> > currently
> > being raised as a problem only occurs in two places, the kdamon_call()
> > loop
> > and the damon_call() function, it is effective to handle it with a
> > spin_lock
> > as shown below.
> >
> > @@ -1502,14 +1501,15 @@ int damon_call(struct damon_ctx *ctx, struct
> > damon_call_control *control)
> >         control->canceled = false;
> >         INIT_LIST_HEAD(&control->list);
> >
> > -       mutex_lock(&ctx->call_controls_lock);
> > +       spin_lock(&ctx->call_controls_lock);
> > +       /* damon_is_running */
> >         if (ctx->kdamond) {
> >                 list_add_tail(&control->list, &ctx->call_controls);
> >         } else {
> > -               mutex_unlock(&ctx->call_controls_lock);
> > +               spin_unlock(&ctx->call_controls_lock);
> >                 return -EINVAL;
> >         }
> > -       mutex_unlock(&ctx->call_controls_lock);
> > +       spin_unlock(&ctx->call_controls_lock);
> >
> >         if (control->repeat)
> >                 return 0;
>
> Are you saying the above diff can fix the damon_call() use-after-free bug [1]?
> Can you please elaborate why you think so?
>
> [1] https://lore.kernel.org/20251231012315.75835-1-sj@xxxxxxxxxx
>
>
> Thanks,
> SJ
>
> [...]