Messages in this thread | | | Date | Tue, 19 Jun 2018 14:38:45 +0100 | From | Quentin Perret <> | Subject | Re: [RFC PATCH v3 03/10] PM: Introduce an Energy Model management framework |
| |
On Tuesday 19 Jun 2018 at 15:23:38 (+0200), Peter Zijlstra wrote: > On Tue, Jun 19, 2018 at 01:58:58PM +0100, Quentin Perret wrote: > > On Tuesday 19 Jun 2018 at 13:34:08 (+0200), Peter Zijlstra wrote: > > > On Mon, May 21, 2018 at 03:24:58PM +0100, Quentin Perret wrote: > > > > +struct em_freq_domain *em_cpu_get(int cpu) > > > > +{ > > > > + struct em_freq_domain *fd; > > > > + unsigned long flags; > > > > + > > > > + read_lock_irqsave(&em_data_lock, flags); > > > > + fd = per_cpu(em_data, cpu); > > > > + read_unlock_irqrestore(&em_data_lock, flags); > > > > > > Why can't this use RCU? This is the exact thing read_locks are terrible > > > at and RCU excells at. > > > > So the idea was that clients (the scheduler for ex) can get a reference > > to a frequency domain object once, and they're guaranteed it always > > exists without asking for it again. > > > > For example, my proposal was to have the scheduler (patch 05) build its > > own private list of frequency domains on which it can iterate efficiently > > in the wake-up path. If we protect this per_cpu variable with RCU, then > > this isn't possible any-more. The scheduler will have to re-ask > > em_cpu_get() at every wake-up, and that makes iterating over frequency > > domains a whole lot more complex. > > > > Does that make any sense ? > > None what so ever... The lock doesn't guarantee stability any more than > RCU does. > > If you hand out the pointer and then drop the read-lock, the write-lock > can proceed and change the pointer right after you. > > The very easiest solution is to never change the data, as I think was > suggested elsewhere in the thread. Construct the thing once and then > never mutate.
This is what is done actually. We will never write twice in the per_cpu array itself. One of the fields (the table) in the structure pointed from the per_cpu array can change, but not the pointer on the structure itself.
The only reason this lock is here is to ensure the atomicity of the write happening in em_register_freq_domain. But that write can happen only once, the first time the frequency domain is registered.
But maybe I could use something simpler than a lock in this case ? Would WRITE_ONCE/READ_ONCE be enough to ensure that atomicity for example ?
Thanks, Quentin
| |