Re: [PATCH v3 2/2] i2c: Add GPIO-based hotplug gate

[Michał Mirosław]

On Sat, Aug 05, 2023 at 01:45:53AM +0200, Andi Shyti wrote:
> On Tue, Aug 01, 2023 at 01:01:43AM +0200, Michał Mirosław wrote:
> > On Mon, Jul 31, 2023 at 12:11:47AM +0200, Michał Mirosław wrote:
> > > On Sun, Jul 30, 2023 at 10:25:07PM +0200, Andi Shyti wrote:
> > > > On Sat, Jul 29, 2023 at 07:08:57PM +0300, Svyatoslav Ryhel wrote:
> > > > > +static int i2c_hotplug_activate(struct i2c_hotplug_priv *priv)
> > > [...]
> > > > > +{
> > > > > +	int ret;
> > > > > +
> > > > > +	if (priv->adap.algo_data)
> > > > > +		return 0;
> > > [...]
> > > > > +	ret = i2c_add_adapter(&priv->adap);
> > > > > +	if (!ret)
> > > > > +		priv->adap.algo_data = (void *)1;
> > > > 
> > > > You want to set algo_data to "1" in order to keep the
> > > > activate/deactivate ordering.
> > > > 
> > > > But if we fail to add the adapter, what's the point to keep it
> > > > active?
> > > 
> > > The code above does "if we added the adapter, remember we did so".
> > > IOW, if we failed to add the adapter we don't set the mark so that
> > > the next interrupt edge can trigger another try. Also we prevent
> > > trying to remove an adapter we didn't successfully add.
> > 
> > Maybe the function's name is misleading? We could find a better one.
> > Activation/deactivation in this driver means "initialize/shutdown the
> > hotplugged bus" and is done in response to an edge (triggering an
> > interrupt) of the hotplug-detect signal.
> 
> So that algo_data is randomly chosen as a boolean value given the
> fact that this particular driver doesn't have its own algorithms
> but it's using the ones from the parent. Right?
[...]

Not exactly.  There is an 'algorithm for this driver just forwards the calls to
the parent bus and has no use of the algo_data field.  The field is thus
used to store a flag noting whether the child bus was registered or not.

> And... thinking aloud... are there race conditions here? I
> mean... you can't attach two docking stations, but are there
> other scenarios?

The driver depends on I2C core code synchronization (e.g. i2c_del_adapter()
waiting for ongoing transfers). Outside of probe/remove there is only
a single thread used by the driver: the interrupt handler.

While reading to answer your question I noticed that IRQF_ONESHOT can be
removed: if the thread picks up the signal then it atomically clears the
trigger flag; if another signal arrives before the handler is done,
handler will be called again.

Best Regards,
Michał Mirosław