Re: [PATCH v2] reset: Exclusive resets must be dedicated to a single hardware block

[Geert Uytterhoeven]

Hi Philipp,

On Mon, Oct 8, 2018 at 12:57 PM Philipp Zabel <p.zabel@xxxxxxxxxxxxxx> wrote:
> On Mon, 2018-10-08 at 11:59 +0200, Geert Uytterhoeven wrote:
> > On Fri, Oct 5, 2018 at 5:16 PM Philipp Zabel <p.zabel@xxxxxxxxxxxxxx> wrote:
> > > On Fri, 2018-10-05 at 14:31 +0200, Geert Uytterhoeven wrote:
> > > [...]
> > > > > > > +                     eq = (args2.np == args.np &&
> > > > > > > +                           args2.args_count == args.args_count &&
> > > > > > > +                           !memcmp(args2.args, args.args,
> > > > > > > +                                   args.args_count * sizeof(args.args[0])));
> > > > >
> > > > > As there's at least one other function in -next that compares of_phandle_args,
> > > > > I will add a helper of_phandle_args_eq().
> > > > >
> > > > > > > +                     of_node_put(args2.np);
> > > > > > > +                     if (eq)
> > > > > >
> > > > > > Emitting a loud warning here could be very helpful if it contains
> > > > > > both the reset controller node and the reset index, as well as the
> > > >
> > > > Actually on DT-based systems, the index is a driver-specific
> > > > implementation detail, and may differ from the actual reset specifier in DT.
> > > > E.g. on R-Car systems, DT uses a human-readable representation matching
> > > > the datasheet, while internally, the driver uses a packed representation.
> > > > Hence printing the index may confuse the user.
> > > >
> > > > For lookup-based systems, this is different.
> > >
> > > Correct. I'm so used to #reset-cells = <1>, it's hard to remember the
> > > exceptions. So let's not try to print indices or args.
> > >
> > > > > > consumer nodes: node and node2.
> > > > >
> > > > > Indeed, will do, also for lookup resets.
> > > > >
> > > > > We already have of_print_phandle_args(), but that is a bit inflexible.
> > > > > Adding support for "%pOFa" looks like the modern thing to do.
> > > >
> > > > Scrap that: of_phandle_args is not derived from a device_node, so %pOFa
> > > > is not appropriate (and would crash instead of fall back to a pointer before
> > > > %pOFa support is implemented). And without more users, it doesn't make much
> > > > sense to go for a new type (e.g. "%pOA")...
> > > >
> > > > Actually, printing the full reset specifier is not needed. A message like
> > > >
> > > >     /soc/pwm@e6e31000 and /soc/pwm@e6e30000 share a reset on
> > > > /soc/clock-controller@e6150000
> > > >
> > > > should give sufficient clue to the user.
> > >
> > > Yes. You could also pass con_id into __of_reset_is_exclusive and print
> > > that. It would be nice to indicate which consumer requested exclusive
> > > access.
> >
> > con_id is used for lookup-based resets only?
> >
> > But the value passed there is the "id" parameter of
> > reset_control_get_exclusive().
>
> Sorry, I did mean the id parameter in the __of_reset_control_get case.

OK.

> > However, that is not the consumer name,
>
> It is the name of the reset signal from point of view of the consumer.
> It specifies, via its position in the reset-names property, which of
> potentially multiple reset phandles in the resets property is the one
> causing the conflict.
>
> > and usually NULL.
>
> In which case the resets property usually only contains one phandle, so
> it is not needed to determine the conflicting reset control.

For a device-specific driver knowing about all resets, that is indeed the case.
For the generic case (e.g. my vfio-platform use case), it just wants to reset
the device, and thus passes NULL.

Perhaps I should use devm_reset_control_array_get_exclusive() instead of
reset_control_get_exclusive(), to assert all resets?

However, that makes the detection of shared resets more tricky, as it needs
to consider all combinations. Currently e.g.
    resets = <&rst1> <&rst2>;
and
    resets = <&rst2> <&rst1>;
is not detected as reset sharing...

As device_reset() also uses the first reset only, I'll keep it that way...

> > I'm afraid the only way to know the consumer is to print a backtrace with
> > WARN()?
>
> I'm just suggesting to augment the warning message with the reset id, if
> available. For example:
>
>     /soc/pwm@e6e31000 requests exclusive control over reset "pwm"
>     shared with /soc/pwm@e6e30000 on /soc/clock-controller@e6150000

Thanks, consider it done for next version.

Gr{oetje,eeting}s,

                        Geert

-- 
Geert Uytterhoeven -- There's lots of Linux beyond ia32 -- geert@xxxxxxxxxxxxxx

In personal conversations with technical people, I call myself a hacker. But
when I'm talking to journalists I just say "programmer" or something like that.
                                -- Linus Torvalds