On Fri, 2017-04-21 at 16:55 +0200, Peter Rosin wrote:
> On 2017-04-21 16:41, Philipp Zabel wrote:
> > On Fri, 2017-04-21 at 16:32 +0200, Peter Rosin wrote:
> >> On 2017-04-21 16:23, Philipp Zabel wrote:
> >>> On Thu, 2017-04-13 at 18:43 +0200, Peter Rosin wrote:
> >>> [...]
> >>>> +int mux_chip_register(struct mux_chip *mux_chip)
> >>>> +{
> >>>> + int i;
> >>>> + int ret;
> >>>> +
> >>>> + for (i = 0; i < mux_chip->controllers; ++i) {
> >>>> + struct mux_control *mux = &mux_chip->mux[i];
> >>>> +
> >>>> + if (mux->idle_state == mux->cached_state)
> >>>> + continue;
> >>>
> >>> I think this should be changed to
> >>>
> >>> - if (mux->idle_state == mux->cached_state)
> >>> + if (mux->idle_state == mux->cached_state ||
> >>> + mux->idle_state == MUX_IDLE_AS_IS)
> >>> continue;
> >>>
> >>> or the following mux_control_set will be called with state ==
> >>> MUX_IDLE_AS_IS. Alternatively, mux_control_set should return when passed
> >>> this value.
> >>
> >> That cannot happen because ->cached_state is initialized to -1
> >> in mux_chip_alloc, so should always be == MUX_IDLE_AS_IS when
> >> registering. And drivers are not supposed to touch ->cached_state.
> >> I.e., ->cached_state is "owned" by the core.
> >
> > So this was caused by me filling cached_state from register reads in the
> > mmio driver. Makes me wonder why I am not allowed to do this, though, if
> > I am able to read back the initial state?
>
> You gain fairly little by reading back the original state. If the mux
> should idle-as-is, you can avoid a maximum of one mux update if the first
> consumer happens to starts by requesting the previously active state.
> Similarly, if the mux should idle in a specific state, you can avoid a
> maximum of one mux update.
>
> In both cases it costs one unconditional read of the mux state.
>
> Sure, in some cases reads are cheaper than writes, but I didn't think
> support for seeding the cache was worth it. Is it worth it?
Probably not, I'll just drop the cached_state initialization. It should
be documented in the mux.h that this field is framework internal and not
to be touched by the drivers. At least I was surprised.
regards
Philipp
--
To unsubscribe from this list: send the line "unsubscribe devicetree" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at http://vger.kernel.org/majordomo-info.html
