Re: [PATCH v3] cpufreq: fix race on cpufreq online

[Rafael J. Wysocki]

On Tue, May 24, 2022 at 1:48 PM Rafael J. Wysocki <rafael@xxxxxxxxxx> wrote:
>
> On Tue, May 24, 2022 at 1:29 PM Viresh Kumar <viresh.kumar@xxxxxxxxxx> wrote:
> >
> > On 24-05-22, 13:22, Rafael J. Wysocki wrote:
> > > On Tue, May 24, 2022 at 1:15 PM Viresh Kumar <viresh.kumar@xxxxxxxxxx> wrote:
> > > >
> > > > On 13-05-22, 09:57, Viresh Kumar wrote:
> > > > > On 12-05-22, 12:49, Rafael J. Wysocki wrote:
> > > > > > > > Moreover, I'm not sure why the locking dance in store() is necessary.
> > > > > > >
> > > > > > > commit fdd320da84c6 ("cpufreq: Lock CPU online/offline in cpufreq_register_driver()")
> > > > > >
> > > > > > I get that, but I'm wondering if locking CPU hotplug from store() is
> > > > > > needed at all.  I mean, if we are in store(), we are holding an active
> > > > > > reference to the policy kobject, so the policy cannot go away until we
> > > > > > are done anyway.  Thus it should be sufficient to use the policy rwsem
> > > > > > for synchronization.
> > > > >
> > > > > I think after the current patchset is applied and we have the inactive
> > > > > policy check in store(), we won't required the dance after all.
> > > >
> > > > I was writing a patch for this and then thought maybe look at mails
> > > > around this time, when you sent the patch, and found the reason why we
> > > > need the locking dance :)
> > > >
> > > > https://lore.kernel.org/lkml/20150729091136.GN7557@xxxxxxxxxxxxxxxxxxxxxx/
> >
> > Actually no, this is for the lock in cpufreq_driver_register().
> >
> > > Well, again, if we are in store(), we are holding a reference to the
> > > policy kobject, so this is not initialization time.
> >
> > This is the commit which made the change.
> >
> > commit 4f750c930822 ("cpufreq: Synchronize the cpufreq store_*() routines with CPU hotplug")
>
> So this was done before the entire CPU hotplug rework and it was
> useful at that time.
>
> The current code always runs cpufreq_set_policy() under policy->rwsem
> and governors are stopped under policy->rwsem, so this particular race
> cannot happen AFAICS.
>
> Locking CPU hotplug prevents CPUs from going away while store() is
> running, but in order to run store(), the caller must hold an active
> reference to the policy kobject.  That prevents the policy from being
> freed and so policy->rwsem can be acquired.  After policy->rwsem has
> been acquired, policy->cpus can be checked to determine whether or not
> there are any online CPUs for the given policy (there may be none),
> because policy->cpus is only manipulated under policy->rwsem.
>
> If a CPU that belongs to the given policy is going away,
> cpufreq_offline() has to remove it from policy->cpus under
> policy->rwsem, so either it has to wait for store() to release
> policy->rwsem, or store() will acquire policy->rwsem after it and will
> find that policy->cpus is empty.

Moreover, locking CPU hotplug doesn't actually prevent
cpufreq_remove_dev() from running which can happen when the cpufreq
driver is unregistered, for example.