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

[Reinette Chatre]

Hi Valentin,

On 12/11/2020 12:46 PM, Valentin Schneider wrote:
> On 03/12/20 23:25, Reinette Chatre wrote:
> > Valentin's series in [2] ends by adding memory barriers to support the
> > updating of the task_struct from one CPU and the usage of the task_struct data
> > from another CPU. This work is still needed and as discussed with Valentin in
> > that thread the work would be re-evaluated by him after seeing how this series
> > turns out.

Thank you very much for taking a look.

> So the "problematic" pattern is still there: a context switch can happen
> concurrently with a write to the switching-to-tasks's {closid, rmid}.
> Accesses to these fields would thus need to be wrapped by READ_ONCE() &
> WRITE_ONCE().

ok.

> 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? 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.

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. 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?

Thank you

Reinette


[1] 
https://lore.kernel.org/lkml/20201123022433.17905-4-valentin.schneider@xxxxxxx/