On Wed, 2017-04-26 at 20:52 +0800, Honggang LI wrote:
> On Wed, Apr 26, 2017 at 09:46:34AM +0200, Paolo Abeni wrote:
> > On Wed, 2017-04-26 at 00:50 +0300, Or Gerlitz wrote:
> > > so maybe @ least for the time being, we should be picking Hong's patch
> > > with proper change log and without the giant stack dump till we have
> > > something better. If you agree, can you do the re-write of the change
> > > log?
> >
> > I think that Hong's patch is following the correct way to fix the
> > issue: ipoib_hard_header() can't assume that the skb headroom is at
> > least IPOIB_HARD_LEN bytes, as wrongly implied by commit fc791b633515
> > (my fault, I'm sorry).
> >
> > Perhaps we can make the code a little more robust with something
> > alongside the following (only compile tested):
> > ---
> > diff --git a/drivers/infiniband/ulp/ipoib/ipoib_main.c b/drivers/infiniband/ulp/ipoib/ipoib_main.c
> > index d1d3fb7..d53d2e1 100644
> > --- a/drivers/infiniband/ulp/ipoib/ipoib_main.c
> > +++ b/drivers/infiniband/ulp/ipoib/ipoib_main.c
> > @@ -1160,6 +1160,11 @@ static int ipoib_hard_header(struct sk_buff *skb,
> > const void *daddr, const void *saddr, unsigned len)
> > {
> > struct ipoib_header *header;
> > + int ret;
> > +
> > + ret = skb_cow_head(skb, IPOIB_HARD_LEN);
>
> I don't think so. When this skb_under_panic arise, all slaves had been
> removed from a busy bonding interface, so it is better to immediately
> give up and return error.
I fear we can hit the same condition even with other, more convoluted,
setup (e.g. some kind of ip tunnel on top of ipoib) and the
skb_cow_head() check should be cheap, in the common case.
Anyway I'm fine even with the original fix, which has the advantage to
minimize the side effects.
Paolo
--
To unsubscribe from this list: send the line "unsubscribe linux-rdma" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at http://vger.kernel.org/majordomo-info.html
