Re: [PATCH RFC] time: drop do_sys_times spinlock

[Rik van Riel <riel@redhat.com>]

-----BEGIN PGP SIGNED MESSAGE-----
Hash: SHA1

On 08/12/2014 03:12 PM, Oleg Nesterov wrote:

> Afaics, the problem is that a single thread can observe the
> decreasing (say) sum_exec_runtime if it calls do_sys_times() twice
> without the lock.
> 
> This is because it can account the exiting sub-thread twice if it
> races with __exit_signal() which increments sig->sum_sched_runtime,
> but this exiting thread can still be visible to
> thread_group_cputime().
> 
> IOW, it is not actually about decreasing, the problem is that the
> lockless thread_group_cputime() can return the wrong result, and
> the next ys_times() can show the right value.

You are right, changing the test case to call times() many
times in a row in each thread can result in the wrong value
being returned.

Not entirely sure what I can do there...

Replacing the spinlock with a seqlock, and taking it for
write in most places is pretty gross, and may lead to other
issues like reader livelock when there is a lot of write
activity.

Having a seqlock just for the stats?  Not sure the calls
to times() are a big enough issue for most workloads to
justify that...

Any other ideas?

- -- 
All rights reversed
-----BEGIN PGP SIGNATURE-----
Version: GnuPG v1

iQEcBAEBAgAGBQJT6pTdAAoJEM553pKExN6DMrIIAKFFHD8luyqgVUAm0jbV8JHm
O5PD81kot95POV7ZAl6crKmPi0OoeSdZIzcmuLFIvRJWqrbgWY6h4rQH9va5B830
F7TC2PRzWwUVwcuEoaUkuZMbUWkWqzUwXcwwFl1blYmkVJVRF27VcUB4S0jia1eq
l2TlQyC1HgXa3E7rbQ6vuKsOq50jB08MWwxEfhAEMNvndhos/fvZlsxL39UO3/X7
AVk+V/leE5tfAfyr6uPrWDR7/u9sJkqmi/dGJ/xjfWNU2swEPvMXk6UhspSIY+mg
KAMa+JWTPANeUSRM9HRA9YUpo0rqvy0Azmg84tIYr4nXsIyvzHuRgCUNQkOmEDQ=
=5ap3
-----END PGP SIGNATURE-----