Re: [PATCH -next] power/qos: fix a data race in pm_qos_*_value

[Rafael J. Wysocki]

On Mon, Feb 24, 2020 at 2:01 AM Qian Cai <cai@xxxxxx> wrote:
>
>
>
> > On Feb 23, 2020, at 7:12 PM, Rafael J. Wysocki <rafael@xxxxxxxxxx> wrote:
> >
> > It may be a bug under certain conditions, but you don't mention what
> > conditions they are.  Reporting it as a general bug is not accurate at
> > the very least.
>
> Could we rule out load tearing, store tearing and reload of global_req in cpuidle_governor_latency() for all compilers and architectures which could introduce logic bugs?
>
>         int global_req = cpu_latency_qos_limit();
>
>         if (device_req > global_req)
>                 device_req = global_req;
>
> If under register pressure, the compiler might get ride of the tmp variable, i.e.,
>
> If (device_req > cpu_latency_qos_limit())
> â-> race with the writer.
>          device_req = cpu_latency_qos_limit();

Yes, there is a race here with or without the WRITE_ONCE()/READ_ONCE()
annotations (note that these annotations don't prevent CPUs from
reordering things, so device_req may be set before global_req
regardless).

However, worst-case it may cause an old value to be used and that can
happen anyway if the entire cpuidle_governor_latency_req() runs
between the curr_value update and pm_qos_set_value() in
pm_qos_update_target(), for example.

IOW, there is no guarantee that the new value will be used immediately
after updating a QoS request anyway.

I agree with adding the annotations (I was considering posting a patch
doing that myself), but just as a matter of making the intention
clear.