Re: [PATCH 3/3] can: rcar_canfd: Fix interrupt registration order

[sashiko-bot@kernel.org]

Thank you for your contribution! Sashiko AI review found 3 potential issue(s) to consider:

New issues:
- [High] Premature UAPI exposure causes unhandled hardware interrupts (IRQ storm).

Pre-existing issues:
- [High] Use-After-Free in global interrupt handlers during driver removal and error paths.
- [High] NULL pointer dereference due to early channel IRQ registration.
--

commit a0e1233eee22a1c4cd2d28a4473c39b527927759
Author: Biju Das <biju.das.jz@bp.renesas.com>

can: rcar_canfd: Fix interrupt registration order

In rcar_canfd_probe(), global interrupts are requested via
devm_request_irq() before rcar_canfd_global_init() enables the clocks
and before the per-channel structures in gpriv->ch[] are allocated.

> diff --git a/drivers/net/can/rcar/rcar_canfd.c b/drivers/net/can/rcar/rcar_canfd.c
> index 1f3dc4b354f87..48f624c1eda9a 100644
> --- a/drivers/net/can/rcar/rcar_canfd.c
> +++ b/drivers/net/can/rcar/rcar_canfd.c
> @@ -2216,6 +2216,17 @@ static int rcar_canfd_probe(struct platform_device *pdev)
>  	gpriv->base = addr;
>  	gpriv->fcbase = addr + gpriv->info->regs->coffset;
>  
> +	err = rcar_canfd_global_init(gpriv);
> +	if (err)
> +		goto fail_dev;
> +
> +	for_each_set_bit(ch, &gpriv->channels_mask, info->max_channels) {
> +		err = rcar_canfd_channel_probe(gpriv, ch, fcan_freq,
> +					       transceivers[ch]);

[Severity: High]
Does moving rcar_canfd_channel_probe() before the global interrupt
registration expose the device to userspace prematurely?

rcar_canfd_channel_probe() calls register_candev(), which makes the network
device visible. If a userspace daemon immediately brings the interface up, it
can transition to operational mode and receive CAN frames, asserting the
global receive interrupt before devm_request_irq() has registered the global
handler.

Could this lead to an unhandled hardware interrupt storm that permanently
disables the CAN interrupt line?

[Severity: High]
This is a pre-existing issue, but does rcar_canfd_channel_probe() suffer
from the exact same early interrupt registration bug being fixed here?

Inside rcar_canfd_channel_probe(), when multi_channel_irqs is enabled,
devm_request_irq() is called for the channel error interrupt before the
channel pointer is stored:

drivers/net/can/rcar/rcar_canfd.c:rcar_canfd_channel_probe() {
    ...
    err = devm_request_irq(dev, err_irq, rcar_canfd_channel_err_interrupt,
                           ...);
    ...
    gpriv->ch[priv->channel] = priv;
}

If a pending interrupt fires immediately, the handler calls
rcar_canfd_handle_channel_err(), which dereferences gpriv->ch[ch].

Since gpriv->ch[ch] is still NULL, will this result in a kernel panic?

> +		if (err)
> +			goto fail_channel;
> +	}
> +
>  	/* Request IRQ that's common for both channels */
>  	if (info->shared_global_irqs) {
>  		err = devm_request_irq(dev, ch_irq,
> @@ -2224,7 +2235,7 @@ static int rcar_canfd_probe(struct platform_device *pdev)
>  		if (err) {
>  			dev_err(dev, "devm_request_irq %d failed: %pe\n",
>  				ch_irq, ERR_PTR(err));
> -			goto fail_dev;
> +			goto fail_channel;

[Severity: High]
This is a pre-existing issue, but does the fail_channel error path have a
use-after-free race condition?

When tearing down the channels, rcar_canfd_channel_remove() calls
free_candev() but does not clear the gpriv->ch[ch] pointer:

drivers/net/can/rcar/rcar_canfd.c:rcar_canfd_channel_remove() {
    ...
    free_candev(priv->ndev);
    // gpriv->ch[ch] is left dangling
}

Since global interrupts remain enabled until rcar_canfd_global_deinit() runs
after the channel loop, if a global receive interrupt fires for an active
channel, the handler rcar_canfd_global_receive_fifo_interrupt() will iterate
over all channels and dereference the dangling gpriv->ch[ch] pointer.

Can this cause a kernel crash if CAN traffic is present during the error
path or driver removal?

[ ... ]

-- 
Sashiko AI review · https://sashiko.dev/#/patchset/20260630135119.104774-1-biju.das.jz@bp.renesas.com?part=3