Re: [PATCH v3 1/7] clkdev: Hold clocks_mutex while iterating clocks list

[Russell King - ARM Linux admin]

On Mon, Apr 08, 2019 at 10:00:02AM -0700, Stephen Boyd wrote:
> Quoting Matti Vaittinen (2019-04-08 03:49:41)
> > On Fri, Apr 05, 2019 at 01:37:24PM -0700, Stephen Boyd wrote:
> > > Quoting Vaittinen, Matti (2019-04-04 23:51:43)
> > > > On Thu, 2019-04-04 at 14:53 -0700, Stephen Boyd wrote:
> > > > > We recently introduced a change to support devm clk lookups. That
> > > > > change
> > > > > introduced a code-path that used clk_find() without holding the
> > > > > 'clocks_mutex'. Unfortunately, clk_find() iterates over the 'clocks'
> > > > > list and so we need to prevent the list from being modified while
> > > > > iterating over it by holding the mutex. Similarly, we don't need to
> > > > > hold
> > > > > the 'clocks_mutex' besides when we're dereferencing the clk_lookup
> > > > > pointer
> > > > 
> > > > /// Snip
> > > > 
> > > > > -out:
> > > > > +static struct clk_lookup *clk_find(const char *dev_id, const char
> > > > > *con_id)
> > > > > +{
> > > > > +     struct clk_lookup *cl;
> > > > > +
> > > > > +     mutex_lock(&clocks_mutex);
> > > > > +     cl = __clk_find(dev_id, con_id);
> > > > >       mutex_unlock(&clocks_mutex);
> > > > >  
> > > > > -     return cl ? clk : ERR_PTR(-ENOENT);
> > > > > +     return cl;
> > > > > +}
> > > > 
> > > > I am not an expert on this but reading commit message abowe and seeing
> > > > the code for clk_find() looks a bit scary. If I understand it
> > > > correctly, the clocks_mutex should be held when dereferencing the
> > > > clk_lookup returned by clk_find. The clk_find implementation drops the
> > > > lock before returning - which makes me think I miss something here. How
> > > > can the caller ever safely dereference returned clk_lookup pointer?
> > > > Just reading abowe makes me think that lock should be taken by whoever
> > > > is calling the clk_find, and dropped only after caller has used the
> > > > found clk_lookup for whatever caller intends to use it. Maybe I am
> > > > missing something?
> > > > 
> > > 
> > > The only user after this patch (devm) is doing a pointer comparison so
> > > it looks OK. But yes, in general there shouldn't be users of clk_find()
> > > that dereference the pointer because there isn't any protection besides
> > > the mutex.
> > 
> > If the only (intended) user for clk_find is devm_clk_release_clkdev,
> > then we might want to write it in devm_clk_release_clkdev - just to
> > avoid similar errors (as I did with devm) in the future? I might even
> > consider renaming __clk_find to clk_find or to clk_find_unsafe - but
> > that's all just nitpicking :) Go with what you like to maintain :D
> > 
> 
> Sure. I was thinking along the same lines after you asked.

This is rubbish.  The reason clk_find() doesn't take the lock is that
you _need_ to hold the lock while you dereference the clk_lookup data.
The lock isn't protecting just the lookup, it protects what you do with
the result of the lookup as well.

So, as I say, adding locking inside clk_find() is completely
misunderstanding the locking here.

-- 
RMK's Patch system: https://www.armlinux.org.uk/developer/patches/
FTTC broadband for 0.8mile line in suburbia: sync at 12.1Mbps down 622kbps up
According to speedtest.net: 11.9Mbps down 500kbps up