Re: [PATCH v2 14/16] net: qrtr: Add MHI transport layer

[Chris Lew]

On 2/10/2020 7:50 PM, Manivannan Sadhasivam wrote:
> Hi Chris,
>
> On Thu, Feb 06, 2020 at 04:14:19PM -0800, Chris Lew wrote:
> > On 2/4/2020 12:19 AM, Manivannan Sadhasivam wrote:
> > > Hi Jakub,
> > >
> > > On Mon, Feb 03, 2020 at 10:12:25AM -0800, Jakub Kicinski wrote:
> > > > On Fri, 31 Jan 2020 19:20:07 +0530, Manivannan Sadhasivam wrote:
> > > > > +/* From QRTR to MHI */
> > > > > +static void qcom_mhi_qrtr_ul_callback(struct mhi_device *mhi_dev,
> > > > > +				      struct mhi_result *mhi_res)
> > > > > +{
> > > > > +	struct qrtr_mhi_dev *qdev = dev_get_drvdata(&mhi_dev->dev);
> > > > > +	struct qrtr_mhi_pkt *pkt;
> > > > > +	unsigned long flags;
> > > > > +
> > > > > +	spin_lock_irqsave(&qdev->ul_lock, flags);
> > > > > +	pkt = list_first_entry(&qdev->ul_pkts, struct qrtr_mhi_pkt, node);
> > > > > +	list_del(&pkt->node);
> > > > > +	complete_all(&pkt->done);
> > > > > +
> > > > > +	kref_put(&pkt->refcount, qrtr_mhi_pkt_release);
> > > > Which kref_get() does this pair with?
> > > >
> > > > Looks like qcom_mhi_qrtr_send() will release a reference after
> > > > completion, too.
> > > Yikes, there is some issue here...
> > >
> > > Acutally the issue is not in what you referred above but the overall kref
> > > handling itself. Please see below.
> > >
> > > kref_put() should be present in qcom_mhi_qrtr_ul_callback() as it will
> > > decrement the refcount which got incremented in qcom_mhi_qrtr_send(). It
> > > should be noted that kref_init() will fix the refcount to 1 and kref_get() will
> > > increment to 2. So for properly releasing the refcount to 0, we need to call
> > > kref_put() twice.
> > >
> > > So if all goes well, the refcount will get decremented twice in
> > > qcom_mhi_qrtr_ul_callback() as well as in qcom_mhi_qrtr_send() and we are good.
> > >
> > > But, if the transfer has failed ie., when qcom_mhi_qrtr_ul_callback() doesn't
> > > get called, then we are leaking the refcount. I need to rework the kref handling
> > > code in next iteration.
> > >
> > > Thanks for triggering this!
> > >
> > > Regards,
> > > Mani
> > >
> > > > > +	spin_unlock_irqrestore(&qdev->ul_lock, flags);
> > > > > +}
> > Hi Mani,
> >
> > I'm not sure if this was changed in your patches but MHI is supposed to give a
> > ul_callback() for any packet that is successfully queued. In the case of the
> > transfer failing, the ul_callback() should still be called so there should
> > be no refcount leaking. It is an essential assumption I made, if that no longer
> > holds true then the entire driver needs to be reworked.
> Your assumption is correct. Only when the packet gets queued into the transfer
> ring, the ul_xfer_cb will be called irrespective of the transfer state (success
> or failure). But when the mhi_queue_transfer() returns even before queuing any
> packet, then we need to decrease the refcount in the error path.
>
> Please correct me if I'm wrong.

The error path for mhi_queue_transfer directly frees the packet structure since no
other context has a reference to those structs. If you wanted to clean it up and converge
using kref release to free, I think that would work. There are some things you'll have to
re-arrange like at what point the packet is added to the ul pkts list.

> Thanks,
> Mani
>
> > Thanks,
> > Chris
> >
> > --
> >
> > Qualcomm Innovation Center, Inc. is a member of Code Aurora Forum, a Linux Foundation Collaborative Project

--
Qualcomm Innovation Center, Inc. is a member of Code Aurora Forum, a Linux Foundation Collaborative Project