Re: [PATCH net v3 2/2] net: fec: free the PPS interrupt before tearing down the PHC and netdev

[netdev-bot+sashiko]

Thank you for your contribution! Sashiko AI review found 1 potential
issue(s) to consider:

Critical: 0 · High: 1 (1) · Medium: 0 · Low: 0

- [High] Incomplete fix: only the PPS interrupt is converted to explicit
  request_irq()/free_irq().

Pre-existing issues:
- [High] Teardown ordering race in fec_ptp_stop()
  (drivers/net/ethernet/freescale/fec_ptp.c:856-874):…

--

net: fec: free the PPS interrupt before tearing down the PHC and netdev

This converts the PPS interrupt in fec_ptp_init() from devm_request_irq()
to request_irq(), records the irq number in fep->pps_irq, and releases it
with free_irq() in fec_ptp_stop() before ptp_clock_unregister().  The goal
is to make the handler go away before the PHC and the netdev are freed.

> diff --git a/drivers/net/ethernet/freescale/fec.h b/drivers/net/ethernet/freescale/fec.h
> index 7176803146f3..960b9f01c531 100644
> --- a/drivers/net/ethernet/freescale/fec.h
> +++ b/drivers/net/ethernet/freescale/fec.h
> @@ -670,6 +670,7 @@ struct fec_enet_private {
>  
>  	/* pps  */
>  	int pps_channel;
> +	int pps_irq;
>  	unsigned int reload_period;
>  	int pps_enable;
>  	unsigned int next_counter;

> diff --git a/drivers/net/ethernet/freescale/fec_ptp.c b/drivers/net/ethernet/freescale/fec_ptp.c
> index 56801c2009d5..598cd8024123 100644
> --- a/drivers/net/ethernet/freescale/fec_ptp.c
> +++ b/drivers/net/ethernet/freescale/fec_ptp.c

[ ... ]

> @@ -785,11 +786,12 @@ void fec_ptp_init(struct platform_device *pdev, int irq_idx)
>  	 * only the PTP_CLOCK_PPS clock events should stop
>  	 */
>  	if (irq >= 0) {
> -		ret = devm_request_irq(&pdev->dev, irq, fec_pps_interrupt,
> -				       0, pdev->name, ndev);
> +		ret = request_irq(irq, fec_pps_interrupt, 0, pdev->name, ndev);
>  		if (ret < 0)
>  			dev_warn(&pdev->dev, "request for pps irq failed(%d)\n",
>  				 ret);
> +		else
> +			fep->pps_irq = irq;
>  	}

[Severity: High]
The commit message says the change closes the case where "a later interrupt
can dereference the freed netdev".  Does that hold once the main FEC
interrupts are considered?

fec_probe() still registers them with devres, using the same ndev cookie:

drivers/net/ethernet/freescale/fec_main.c:fec_probe() {
	...
		ret = devm_request_irq(&pdev->dev, irq, fec_enet_interrupt,
				       0, pdev->name, ndev);
	...
		fep->irq[i] = irq;
}

devres_release_all() runs only after ->remove() returns, but
fec_drv_remove() already frees the netdev inside the callback:

drivers/net/ethernet/freescale/fec_main.c:fec_drv_remove() {
	...
	unregister_netdev(ndev);
	fec_ptp_stop(pdev);
	...
	fec_enet_deinit(ndev);
	free_netdev(ndev);
}

If the line is asserted in that window, fec_enet_interrupt() dereferences
the cookie immediately:

drivers/net/ethernet/freescale/fec_main.c:fec_enet_interrupt() {
	struct net_device *ndev = dev_id;
	struct fec_enet_private *fep = netdev_priv(ndev);
	...
}

The same ordering applies to the probe unwind, where failed_irq,
failed_mii_init and failed_register all fall through to free_netdev(ndev)
with the handlers still installed.

Should fep->irq[] be released with free_irq() in fec_drv_remove() (or the
netdev allocation switched to a devm-managed one) so the lifetime rule is
the same for every irq registered with ndev as dev_id?  As it stands the
driver ends up with two different conventions for identically-cookied
interrupts, and the failure mode described in the changelog is still
reachable through the main irq lines.

> @@ -861,6 +863,12 @@ void fec_ptp_stop(struct platform_device *pdev)
>  
>  	cancel_delayed_work_sync(&fep->time_keep);
>  	hrtimer_cancel(&fep->perout_timer);
> +
> +	if (fep->pps_irq >= 0) {
> +		free_irq(fep->pps_irq, ndev);
> +		fep->pps_irq = -1;
> +	}
> +
>  	if (fep->ptp_clock)
>  		ptp_clock_unregister(fep->ptp_clock);
>  }

[Severity: High]
This isn't a bug introduced by this patch, but the ordering in
fec_ptp_stop() means both hrtimer_cancel() and the new free_irq() run while
the PHC character device is still registered, since ptp_clock_unregister()
comes last.

Can an in-flight ioctl re-arm perout_timer after the cancel?
fec_ptp_enable() drops ptp_clk_mutex and tmreg_lock before arming:

drivers/net/ethernet/freescale/fec_ptp.c:fec_ptp_enable() {
	...
			if (delta > FEC_PTP_MAX_NSEC_COUNTER) {
				timeout = ns_to_ktime(delta - NSEC_PER_SEC);
				hrtimer_start(&fep->perout_timer, timeout, HRTIMER_MODE_REL);
	...
}

posix_clock_ioctl() holds only down_read(&clk->rwsem) across that callback,
and fec_ptp_stop() takes none of those locks.  ptp_clock_unregister() ->
posix_clock_unregister() then waits for the ioctl to finish, so the arm
completes and nothing cancels the timer again.

fec_drv_remove() afterwards calls free_netdev(ndev), which frees the
fec_enet_private embedding perout_timer, and the later expiry runs:

drivers/net/ethernet/freescale/fec_ptp.c:fec_ptp_pps_perout_handler() {
	struct fec_enet_private *fep = container_of(timer,
					struct fec_enet_private, perout_timer);

	fec_ptp_pps_perout(fep);
	...
}

The ptp_clk_on check in fec_ptp_enable() only rejects a fresh request; a
request already past that check is unaffected.  The window also exists on
the probe error path, fec_probe() failed_init -> fec_ptp_stop(), where
ptp_clk_on is true for the whole of probe.

Symmetrically, can a PTP_ENABLE_PPS ioctl reach fec_ptp_enable_pps(fep, 1)
after free_irq() and re-enable the hardware compare interrupt with no
handler installed?  fec_ptp_enable_pps() has no ptp_clk_on gate.

Would it be better to unregister (or defunct) the PHC first, or to re-cancel
the hrtimer after ptp_clock_unregister() returns?

-- 
Sashiko AI review · https://netdev-ai.bots.linux.dev/sashiko/#/patchset/20260909-fec-ptp-pps-event-uaf-v3-0-bead12c29a16%40cherr.cc