Re: [PATCH] net: netcp: Fix crash in error path when DMA channel open fails

[Simon Horman <horms@kernel.org>]

On Fri, Sep 26, 2025 at 10:28:47PM +0530, Siddharth Vadapalli wrote:
> On 26/09/25 10:12 PM, Simon Horman wrote:
> > On Fri, Sep 26, 2025 at 09:57:02PM +0530, Siddharth Vadapalli wrote:
> > > On 26/09/25 9:43 PM, Simon Horman wrote:
> > > > On Fri, Sep 26, 2025 at 10:08:53AM -0500, Nishanth Menon wrote:
> > > > > When knav_dma_open_channel() fails in netcp_setup_navigator_resources(),
> > > > > the rx_channel field is set to an ERR_PTR value. Later, when
> > > > > netcp_free_navigator_resources() is called in the error path, it attempts
> > > > > to close this invalid channel pointer, causing a crash.
> > > > > 
> > > > > Add a check for ERR values to handle the failure scenario.
> > > > > 
> > > > > Fixes: 84640e27f230 ("net: netcp: Add Keystone NetCP core driver")
> > > > > Signed-off-by: Nishanth Menon <nm@ti.com>
> > > > > ---
> > > > > 
> > > > > Seen on kci log for k2hk: https://dashboard.kernelci.org/log-viewer?itemId=ti%3A2eb55ed935eb42c292e02f59&org=ti&type=test&url=http%3A%2F%2Ffiles.kernelci.org%2F%2Fti%2Fmainline%2Fmaster%2Fv6.17-rc7-59-gbf40f4b87761%2Farm%2Fmulti_v7_defconfig%2BCONFIG_EFI%3Dy%2BCONFIG_ARM_LPAE%3Dy%2Bdebug%2Bkselftest%2Btinyconfig%2Fgcc-12%2Fbaseline-nfs-boot.nfs-k2hk-evm.txt.gz
> > > > > 
> > > > >    drivers/net/ethernet/ti/netcp_core.c | 2 +-
> > > > >    1 file changed, 1 insertion(+), 1 deletion(-)
> > > > > 
> > > > > diff --git a/drivers/net/ethernet/ti/netcp_core.c b/drivers/net/ethernet/ti/netcp_core.c
> > > > > index 857820657bac..4ff17fd6caae 100644
> > > > > --- a/drivers/net/ethernet/ti/netcp_core.c
> > > > > +++ b/drivers/net/ethernet/ti/netcp_core.c
> > > > > @@ -1549,7 +1549,7 @@ static void netcp_free_navigator_resources(struct netcp_intf *netcp)
> > > > >    {
> > > > >    	int i;
> > > > > -	if (netcp->rx_channel) {
> > > > > +	if (!IS_ERR(netcp->rx_channel)) {
> > > > >    		knav_dma_close_channel(netcp->rx_channel);
> > > > >    		netcp->rx_channel = NULL;
> > > > >    	}
> > > > 
> > > > Hi Nishanth,
> > > > 
> > > > Thanks for your patch.
> > > > 
> > > > I expect that netcp_txpipe_close() has a similar problem too.
> > > > 
> > > > But I also think that using IS_ERR is not correct, because it seems to me
> > > > that there are also cases where rx_channel can be NULL.
> > > 
> > > Could you please clarify where rx_channel is NULL? rx_channel is set by
> > > invoking knav_dma_open_channel().
> > 
> > Hi Siddharth,
> > 
> > I am assuming that when netcp_setup_navigator_resources() is called, at
> > least for the first time, that netcp->rx_channel is NULL. So any of the
> > occurrence of 'goto fail' in that function before the call to
> > knav_dma_open_channel().
> 
> I missed this. Thank you for pointing this out.

No problem. These error paths are tricking things.

> > > Also, please refer to:
> > > https://github.com/torvalds/linux/commit/5b6cb43b4d62
> > > which specifically points out that knav_dma_open_channel() will not return
> > > NULL so the check for NULL isn't required.
> > > > 
> > > > I see that on error knav_dma_open_channel() always returns ERR_PTR(-EINVAL)
> > > > (open coded as (void *)-EINVAL) on error. So I think a better approach
> > > > would be to change knav_dma_open_channel() to return NULL, and update callers
> > > > accordingly.
> > > 
> > > The commit referred to above made changes to the driver specifically due to
> > > the observation that knav_dma_open_channel() never returns NULL. Modifying
> > > knav_dma_open_channel() to return NULL will effectively result in having to
> > > undo the changes made by the commit.
> > 
> > I wasn't aware of that patch. But my observation is that the return value
> > of knav_dma_open_channel() is still not handled correctly. E.g. the bug
> > your patch is fixing.  And I'm proposing an alternate approach which I feel
> > will be less error-prone.
> 
> Ok. If I understand correctly, you are proposing that the 'error codes'
> returned by knav_dma_open_channel() should be turned into a dev_err() print
> for the user and knav_dma_open_channel() should always return NULL in case
> of failure and a pointer to the channel in case of success. Is that right?

I'm ambivalent regarding the dev_err() part. Because the error is always
the same. And I'm not really sure that logging it adds anything. But if you
do go that way, please consider using %pe.  consider using

Regarding knav_dma_open_channel(0 always returning NULL, yes, that is my
suggestion. Of course the callers and anything else that uses that
return value need to be audited and updated as appropriate.

Thanks!