Re: [PATCH v5 net-next 06/12] net: ethernet: ti: introduce cpsw switchdev based driver part 1 - dual-emac

[Andrew Lunn]

On Fri, Nov 01, 2019 at 10:34:57PM +0200, Grygorii Strashko wrote:
> 
> 
> On 29/10/2019 14:32, Andrew Lunn wrote:
> > > +static int cpsw_probe(struct platform_device *pdev)
> > > +{
> > > +	const struct soc_device_attribute *soc;
> > > +	struct device *dev = &pdev->dev;
> > > +	struct resource *ss_res;
> > > +	struct cpsw_common *cpsw;
> > > +	struct gpio_descs *mode;
> > > +	void __iomem *ss_regs;
> > > +	int ret = 0, ch;
> > > +	struct clk *clk;
> > > +	int irq;
> > > +
> > 
> > ...
> > 
> > > +
> > > +	/* setup netdevs */
> > > +	ret = cpsw_create_ports(cpsw);
> > > +	if (ret)
> > > +		goto clean_unregister_netdev;
> > 
> > At this point, the slave ports go live. If the kernel is configured
> > with NFS root etc, it will start using the interfaces.
> > 
> > +
> > > +	/* Grab RX and TX IRQs. Note that we also have RX_THRESHOLD and
> > > +	 * MISC IRQs which are always kept disabled with this driver so
> > > +	 * we will not request them.
> > > +	 *
> > > +	 * If anyone wants to implement support for those, make sure to
> > > +	 * first request and append them to irqs_table array.
> > > +	 */
> > > +
> > > +	ret = devm_request_irq(dev, cpsw->irqs_table[0], cpsw_rx_interrupt,
> > > +			       0, dev_name(dev), cpsw);
> > > +	if (ret < 0) {
> > > +		dev_err(dev, "error attaching irq (%d)\n", ret);
> > > +		goto clean_unregister_netdev;
> > > +	}
> > > +
> > > +	ret = devm_request_irq(dev, cpsw->irqs_table[1], cpsw_tx_interrupt,
> > > +			       0, dev_name(dev), cpsw);
> > > +	if (ret < 0) {
> > > +		dev_err(dev, "error attaching irq (%d)\n", ret);
> > > +		goto clean_unregister_netdev;
> > > +	}
> > 
> > Are there any race conditions if the network starts using the devices
> > before interrupts are requested? To be safe, maybe this should be done
> > before the slaves are created?
> 
> Usually during boot - there is no parallel probing (as opposite to modules loading by
> udev, for example). Also, there is barrier init call deferred_probe_initcall() to ensure all
> drivers probed before going to mount rootfs.
> 
> So, i do not think this could cause any issue - max few packets will be delayed
> until kernel will switch back here, but the chances that ndo_open will be finished before probe ->0.

I helped track down a crash recently, along these lines. ndo_open()
was getting called before the probe function finished, when kernel ip
address auto config was in action. This is not to do with parallel
probing, i think there is something in register_netdev() which is
triggered each time an interface is added to do the ip
configuration. And the first thing that does is open the interface.

	  Andrew