Re: [EXTERNAL] Re: [PATCH net v2 1/4] octeon_ep: fix race conditions in ndo_get_stats64

[Larysa Zaremba <larysa.zaremba@intel.com>]

On Tue, Dec 17, 2024 at 05:49:13PM +0000, Shinas Rasheed wrote:
> 
> 
> > > > On Mon, Dec 16, 2024 at 03:30:12PM +0100, Larysa Zaremba wrote:
> > > > > On Sun, Dec 15, 2024 at 11:58:39PM -0800, Shinas Rasheed wrote:
> > > > > > ndo_get_stats64() can race with ndo_stop(), which frees input and
> > > > > > output queue resources. Call synchronize_net() to avoid such races.
> > > > > >
> > > > > > Fixes: 6a610a46bad1 ("octeon_ep: add support for ndo ops")
> > > > > > Signed-off-by: Shinas Rasheed <srasheed@marvell.com>
> > > > > > ---
> > > > > > V2:
> > > > > >   - Changed sync mechanism to fix race conditions from using an
> > atomic
> > > > > >     set_bit ops to a much simpler synchronize_net()
> > > > > >
> > > > > >  drivers/net/ethernet/marvell/octeon_ep/octep_main.c | 1 +
> > > > > >  1 file changed, 1 insertion(+)
> > > > > >
> > > > > > diff --git a/drivers/net/ethernet/marvell/octeon_ep/octep_main.c
> > > > b/drivers/net/ethernet/marvell/octeon_ep/octep_main.c
> > > > > > index 549436efc204..941bbaaa67b5 100644
> > > > > > --- a/drivers/net/ethernet/marvell/octeon_ep/octep_main.c
> > > > > > +++ b/drivers/net/ethernet/marvell/octeon_ep/octep_main.c
> > > > > > @@ -757,6 +757,7 @@ static int octep_stop(struct net_device
> > *netdev)
> > > > > >  {
> > > > > >  	struct octep_device *oct = netdev_priv(netdev);
> > > > > >
> > > > > > +	synchronize_net();
> > > > >
> > > > > You should have elaborated on the fact that this synchronize_net() is for
> > > > > __LINK_STATE_START flag in the commit message, this is not obvious.
> > Also,
> > > > is
> > > > > octep_get_stats64() called from RCU-safe context?
> > > > >
> > > >
> > > > Now I see that in case !netif_running(), you do not bail out of
> > > > octep_get_stats64() fully (or at all after the second patch). So, could you
> > > > explain, how are you utilizing RCU here?
> > > >
> > >
> > > The understanding is that octep_get_stats64() (.ndo_get_stats64() in turn) is
> > called from RCU safe contexts, and
> > > that the netdev op is never called after the ndo_stop().
> > 
> > As I now see, in net/core/net-sysfs.c, yes there is an rcu read lock around the
> > thing, but there are a lot more callers and for example veth_get_stats64()
> > explicitly calls rcu_read_lock().
> > 
> > Also, even with RCU-protected section, I am not sure prevents the
> > octep_get_stats64() to be called after synchronize_net() finishes. Again, the
> > callers seem too diverse to definitely say that we can rely on built-in flags
> > for this to not happen :/
> 
> Usually, the understanding is that ndo_get_stats won't be called by the network stack after the interface is put down. As long as that is the case, I don't think we should keep adding checks until there is a strong reason to do so. What do you think?
>

It is hard to know without testing (but testing should not be hard). I think the 
phrase "Statistics must persist across routine operations like bringing the 
interface down and up." [0] implies that bringing the interface down may not 
necessarily prevent stats calls.

[0] https://docs.kernel.org/networking/statistics.html
 
> > > Thanks for the comments