Re: [PATCH V1 1/2] soc: qcom: aoss: Expose send for generic usecase

[Bjorn Andersson <bjorn.andersson@linaro.org>]

On Fri 09 Apr 02:31 CDT 2021, Manivannan Sadhasivam wrote:

> On Sun, Apr 04, 2021 at 12:17:52PM -0500, Bjorn Andersson wrote:
> > On Fri 02 Apr 01:17 CDT 2021, Deepak Kumar Singh wrote:
> > 
> > > Not all upcoming usecases will have an interface to allow the aoss
> > > driver to hook onto. Expose the send api and create a get function to
> > > enable drivers to send their own messages to aoss.
> > > 
> > > Signed-off-by: Chris Lew <clew@codeaurora.org>
> > > Signed-off-by: Deepak Kumar Singh <deesin@codeaurora.org>
> > > ---
> > >  drivers/soc/qcom/qcom_aoss.c | 36 +++++++++++++++++++++++++++++++++++-
> > >  1 file changed, 35 insertions(+), 1 deletion(-)
> > > 
> > > diff --git a/drivers/soc/qcom/qcom_aoss.c b/drivers/soc/qcom/qcom_aoss.c
> > > index 53acb94..5c643f0 100644
> > > --- a/drivers/soc/qcom/qcom_aoss.c
> > > +++ b/drivers/soc/qcom/qcom_aoss.c
> > > @@ -8,10 +8,12 @@
> 
> [...]
> 
> > > +	pdev = of_find_device_by_node(np);
> > 
> > of_find_device_by_node() will increment the refcount of the underlying
> > struct device of pdev, so you need to platform_device_put() once you're
> > done with it.
> > 
> > As a side effect of not putting the struct device, the devm_kzalloc'ed
> > qmp pointer will remain valid. So care is needed to make sure that the
> > client doesn't end up with a dangling pointer if the qmp device is
> > removed.
> > 
> > My suggestion is that you add a "qmp_put()" function, which invokes
> > platform_device_put() and that you add some sort of tracking ("bool
> > orphan"?) to the struct qmp and make qmp_send() fail if this is set.
> > 
> 
> I think this is a duplication of what the struct device offers. Why
> can't we use the generic infrastructure for this usecase?
> 
> Like using device_initialize() in qmp_probe() along with a release
> callback for "struct device", then using get_device() in qmp_get().
> Then there should also be a qmp_put() API which calls put_device() to
> decrease the refcount.
> 
> Ideally, the final refcount should be dropped in qmp_remove() and then
> the release callback will be called automatically to free "struct qmp".
> 
> > That way if someone unbinds the aoss device, the client will still have
> > a "valid" pointer, but won't be able to qmp_send() after qmp_close() has
> > been called in the aoss remove function.
> > 
> 
> How can someone remove the qmp device if a client is holding its reference?
> 

The device could be unbound using sysfs, in which case remove() is
called and I assumed that devres wouldn't be released until the struct
device's refcount hit 0.

Apparently this does not seems to be how it works, following the unbind
path I see that devres is shot down regardless of the struct device's
refcount.

So we would need to ensure that struct qmp is refcounted on its own.
For this we don't need a separate struct device, we can simply add a
kref to the struct and avoid using devres to keep track of its lifetime.

Regards,
Bjorn

> Thanks,
> Mani
> 
> > Regards,
> > Bjorn
> > 
> > > +	if (!pdev)
> > > +		return ERR_PTR(-EINVAL);
> > > +
> > > +	qmp = platform_get_drvdata(pdev);
> > > +	return qmp ? qmp : ERR_PTR(-EPROBE_DEFER);
> > > +}
> > > +EXPORT_SYMBOL(qmp_get);
> > > +
> > >  static int qmp_probe(struct platform_device *pdev)
> > >  {
> > >  	struct resource *res;
> > > -- 
> > > The Qualcomm Innovation Center, Inc. is a member of the Code Aurora Forum,
> > > a Linux Foundation Collaborative Project
> > >