Re: [PATCH] of: Make of_node_name_eq() case insensitive

[Rob Herring]

On Thu, Jan 24, 2019 at 11:00 PM Florian Fainelli <f.fainelli@xxxxxxxxx> wrote:
>
>
>
> On 1/24/19 6:06 PM, Frank Rowand wrote:
> > On 1/24/19 5:20 PM, Florian Fainelli wrote:
> >>
> >>
> >> On 1/24/19 3:45 PM, Frank Rowand wrote:
> >>> On 1/24/19 12:08 PM, Florian Fainelli wrote:
> >>>> Since c32569e358ad ("regulator: Use of_node_name_eq for node name
> >>>> comparisons") Vivien reported the mc13892-regulator complaining about
> >>>> not being able to find regulators.
> >>>>
> >>>> This is because prior to that commit we used of_node_cmp() to compare
> >>>> the regulator array passed from mc13892_regulators down to
> >>>> mc13xxx_parse_regulators_dt() and they are all defined in uppercase
> >>>> letters by the MC13892_*_DEFINE* macros, whereas they are defined as
> >>>> lowercase in the DTS.
> >>>>
> >>>> Fix this by use strncasecmp() since that makes sure the comparison is
> >>>> case insensitive like what of_node_cmp() did.
> >>>>
> >>>> Reported-by: Vivien Didelot <vivien.didelot@xxxxxxxxx>
> >>>> Fixes: c32569e358ad ("regulator: Use of_node_name_eq for node name comparisons")
> >>>> Signed-off-by: Florian Fainelli <f.fainelli@xxxxxxxxx>
> >>>> ---
> >>>>  drivers/of/base.c | 3 ++-
> >>>>  1 file changed, 2 insertions(+), 1 deletion(-)
> >>>>
> >>>> diff --git a/drivers/of/base.c b/drivers/of/base.c
> >>>> index 5226e898476e..ff47c86277cb 100644
> >>>> --- a/drivers/of/base.c
> >>>> +++ b/drivers/of/base.c
> >>>> @@ -66,7 +66,8 @@ bool of_node_name_eq(const struct device_node *np, const char *name)
> >>>>    node_name = kbasename(np->full_name);
> >>>>    len = strchrnul(node_name, '@') - node_name;
> >>>>
> >>>> -  return (strlen(name) == len) && (strncmp(node_name, name, len) == 0);
> >>>> +  return (strlen(name) == len) &&
> >>>> +          (strncasecmp(node_name, name, len) == 0);
> >>>>  }
> >>>>  EXPORT_SYMBOL(of_node_name_eq);
> >>>>
> >>>>
> >>>
> >>> Node names are case sensitive.  Please fix mc13xxx_parse_regulators_dt() to
> >>> properly handle case instead of changing of_node_name_eq().
> >>
> >> Fair enough, should we issue a warning if np->full_name contains upper
> >> case while name does not (and vice versa) to help troubleshoot cases
> >> like the one we found with Vivien?
> >
> > It seems like a lot of work to detect that specific case.  If anything,
> > maybe just add some text to the existing "Unknown regulator: ..." warning
> > in mc13xxx_parse_regulators_dt() to mention that case matters.
>
> I was not thinking about something very clever, just issue a warning
> like this, completely untested and likely flawed:
>
> diff --git a/drivers/of/base.c b/drivers/of/base.c
> index 5226e898476e..2505286c875b 100644
> --- a/drivers/of/base.c
> +++ b/drivers/of/base.c
> @@ -58,6 +58,7 @@ DEFINE_RAW_SPINLOCK(devtree_lock);
>  bool of_node_name_eq(const struct device_node *np, const char *name)
>  {
>         const char *node_name;
> +       int ret1, ret2;
>         size_t len;
>
>         if (!np)
> @@ -66,7 +67,12 @@ bool of_node_name_eq(const struct device_node *np,
> const char *name)
>         node_name = kbasename(np->full_name);
>         len = strchrnul(node_name, '@') - node_name;
>
> -       return (strlen(name) == len) && (strncmp(node_name, name, len)
> == 0);
> +       ret1 = strncmp(node_name, name, len);
> +       ret2 = strncasecmp(node_name, len);
> +
> +       WARN(ret1 ^ ret2, "Comparing case sensitive names!");
> +
> +       return (strlen(name) == len && (strncmp(node_name, name, len) == 0);

If it comes to this, I'd rather not always do 3 string compares.

>  }
>  EXPORT_SYMBOL(of_node_name_eq);
>
>
> My concern is that we have identified one place here where the
> conversion to of_node_name_eq() broke that particular driver in fact,
> all other regulator drivers but qcom-rpmh-regulator.c that use
> of_node_name_eq() are broken after that change, but presumably this is
> not the only place in the kernel where things could break, so having a
> warning could potentially help (also adding the backtrace which is neat).

Most modern DT's don't really rely on the node names. Regulators are
the oddball.

I checked all the other affected regulator drivers. da9052 is also
affected. IMO, we should fix the drivers and I'm sending out patches
to do that.

> What should the fix look like though? Add an of_node_casename_eq() and
> use it, revert Rob's commit that changes these regulators to use
> of_node_name_eq()?

No, we don't want to add of_node_casename_eq.

We may have to internally go back to case insensitive compare, but I'd
like to avoid that and fix any specific cases if possible.

> I don't know the regulator framework enough to know whether forcibly
> making the names lowercase is not breaking sysfs/debugfs...

Userspace relying on specific regulator names seems questionable to
me. Certainly, debugfs is not an ABI.

Rob