On Mon, Oct 18, 2021 at 06:35:10PM +0200, Emil Renner Berthing wrote:
> On Mon, 18 Oct 2021 at 18:24, Andy Shevchenko <andy.shevchenko@xxxxxxxxx> wrote:
> > On Mon, Oct 18, 2021 at 6:56 PM Emil Renner Berthing <kernel@xxxxxxxx> wrote:
> > > On Mon, 18 Oct 2021 at 17:48, Andy Shevchenko <andy.shevchenko@xxxxxxxxx> wrote:
> > > > On Mon, Oct 18, 2021 at 6:35 PM Emil Renner Berthing <kernel@xxxxxxxx> wrote:
> > > > > On Tue, 12 Oct 2021 at 19:03, Andy Shevchenko <andy.shevchenko@xxxxxxxxx> wrote:
> > > > > > On Tue, Oct 12, 2021 at 4:43 PM Emil Renner Berthing <kernel@xxxxxxxx> wrote:
...
> > > > > > > + case PIN_CONFIG_BIAS_DISABLE:
> > > > > >
> > > > > > > + mask |= PAD_BIAS_MASK;
> > > > > >
> > > > > > Use it...
> > > > > >
> > > > > > > + value = (value & ~PAD_BIAS_MASK) | PAD_BIAS_DISABLE;
> > > > > >
> > > > > > ...here. Ditto for the similar cases in this function and elsewhere.
> > > > >
> > > > > I don't follow. How do you want me to use mask? If I did value =
> > > > > (value & ~mask) | PAD_BIAS_DISABLE; then I'd wipe the previous
> > > > > configuration. Eg. suppose the first config is the drive strength and
> > > > > second disables bias. Then on the 2nd loop mask =
> > > > > PAD_DRIVE_STRENGTH_MASK | PAD_BIAS_MASK and the drive strength value
> > > > > would be wiped.
> > > >
> > > > Collect masks and new values in temporary variables and apply them
> > > > once after the loop is done, no?
> > >
> > > But that's exactly what the code does. It merges all the config
> > > options into a single mask and value so we only need to do rmw on the
> > > register once.
> >
> > Then masking the value makes no sense.
> > What you should have is simply as
> >
> > mask |= FOO;
> > value |= BAR;
>
> Yeah, but then we could get into weird states if the device tree
> specifies both bias-disable and bias-pull-up by mistake. This code is
> written so that only the last valid state is chosen.
But shouldn't it be disallowed by:
1) DTC validator (Rob?)
2) GPIO / pin control (Linus, Bart?)
?
...
> > > > > > > + ret = clk_prepare_enable(clk);
> > > > > > > + if (ret) {
> > > > > >
> > > > > > > + reset_control_deassert(rst);
> > > > > >
> > > > > > Use devm_add_action_or_reset().
> > > > >
> > > > > I don't see how that is better.
> > > >
> > > > Pity. The rule of thumb is to either try to use devm_*() everywhere in
> > > > the probe, or don't use it at all. Above is the more-or-less standard
> > > > pattern where devn_add_action_or_reset() is being used in the entire
> > > > kernel.
> > > >
> > > > > Then I'd first need to call that and
> > > > > check for errors, but just on the line below enabling the clock the
> > > > > reset line is deasserted anyway, so then the action isn't needed any
> > > > > longer. So that 3 lines of code for devm_add_action_or_reset +
> > > > > lingering unneeded action or code to remove it again vs. just the line
> > > > > above.
> > > >
> > > > Then don't use devm_*() at all. What's the point?
> > >
> > > I'm confused. So you wan't an unneeded action to linger because the
> > > probe function temporarily asserts reset for 3 lines of code?
> >
> > I;m talking about clk_prepare_enable().
>
> Ok, you wrote your comment under the reset_control_deassert call. How
> would devm_add_action_or_reset for clk_prepare_enable work?
It seems both are needed to be converted, otherwise _everything_ after
reset_assert() should not be devm_*().
TL;DR: the rule is
Allowed: devm_*() followed by non-devm_*()
NOT allowed: devm_*() followed by non-devm_*() followed by devm_*()
Of course, you may try to work the latter one, but it diminishes the whole
idea behind it, that's why I told that may be not using devm_*() is the
correct approach here and that what you meant (?).
The example how to use above mentioned API, just grep for it.
# See [1] for the sources of the used script
$ gl4func.sh devm_add_action_or_reset clk_prepare_enable | wc -l
101
[1]: https://github.com/andy-shev/home-bin-tools/blob/master/gl4func.sh
--
With Best Regards,
Andy Shevchenko
