Re: [PATCH 0/3] x86/resctrl: Fix a few issues in moving a task to a resource group

[Reinette Chatre]

Hi Valentin,

On 12/16/2020 9:41 AM, Valentin Schneider wrote:
> On 14/12/20 18:38, Reinette Chatre wrote:
> > > Thinking a bit more (too much?) about it, we could limit ourselves to
> > > wrapping only reads not protected by the rdtgroup_mutex: the only two
> > > task_struct {closid, rmid} writers are
> > > - rdtgroup_move_task()
> > > - rdt_move_group_tasks()
> > > and they are both invoked while holding said mutex. Thus, a reader holding
> > > the mutex cannot race with a write, so load tearing ought to be safe.
> >
> > The reads that are not protected by the rdtgroup_mutex can be found in
> > __resctrl_sched_in(). It thus sounds to me like your proposed changes to
> > this function found in your patch [1] is what is needed?
>
> Right.
>
> > It is not clear
> > to me how the pairing would work in this case though. If I understand
> > correctly the goal is for the  write to the closid/rmid in the functions
> > you mention above to be paired with the reads in resctrl_sched_in() and
> > it is not clear how adding a single READ_ONCE would accomplish this
> > pairing by itself.
>
> So all the writes would need WRITE_ONCE(), but not all reads would require
> a READ_ONCE() (those that can't race with writes shouldn't need them).

Got it. I misunderstood your previous response, mistakenly thinking that 
it stated that there would be only READ_ONCE() reads without being 
paired with WRITE_ONCE() writes. Thanks for clearing that up.

> I'll go and update that patch so that you can bundle it with v2 of this
> series.

Thank you so much.

> > It is also not entirely clear to me what the problematic scenario could
> > be. If I understand correctly, the risk is (as you explained in your
> > commit message), that a CPU could have its {closid, rmid} fields read
> > locally (resctrl_sched_in()) while they are concurrently being written
> > to from another CPU (in rdtgroup_move_task() and rdt_move_group_tasks()
> > as you state above). If this happens then a task being moved may be
> > scheduled in with its old closid/rmid.
>
> Worse, it may be scheduled with a mangled closid/rmid if the read in
> resctrl_sched_in() is torn (i.e. compiled as a sequence of multiple
> smaller-sized loads). This one of the things READ_ONCE() / WRITE_ONCE()
> try to address.

I see. This area is unfamiliar to me, thank you very much for catching 
this and helping to get it right.

> > The update of closid/rmid in
> > rdtgroup_move_task()/rdt_move_group_tasks() is followed by
> > smp_call_function_xx() where the registers are updated with preemption
> > disabled and thus protected against __switch_to. If a task was thus
> > incorrectly scheduled in with old closid/rmid, would it not be corrected
> > at this point?
>
> Excluding load/store tearing, then yes, the above works fine.

Thank you for this sanity check.

Reinette