Re: [PATCH net] net/sched: act_pedit: really ensure the skb is writable

[Paolo Abeni <pabeni@redhat.com>]

On Wed, 2022-05-04 at 20:04 -0700, Jakub Kicinski wrote:
> On Tue,  3 May 2022 16:05:42 +0200 Paolo Abeni wrote:
> > Currently pedit tries to ensure that the accessed skb offset
> > is writeble via skb_unclone(). The action potentially allows
> > touching any skb bytes, so it may end-up modifying shared data.
> > 
> > The above causes some sporadic MPTCP self-test failures.
> > 
> > Address the issue keeping track of a rough over-estimate highest skb
> > offset accessed by the action and ensure such offset is really
> > writable.
> > 
> > Note that this may cause performance regressions in some scenario,
> > but hopefully pedit is not critical path.
> > 
> > Fixes: db2c24175d14 ("act_pedit: access skb->data safely")
> > Acked-by: Mat Martineau <mathew.j.martineau@linux.intel.com>
> > Tested-by: Geliang Tang <geliang.tang@suse.com>
> > Signed-off-by: Paolo Abeni <pabeni@redhat.com>
> > ---
> > Note: AFAICS the issue is present since 1da177e4c3f4
> > ("Linux-2.6.12-rc2"), but before the "Fixes" commit this change
> > is irrelevant, because accessing any data out of the skb head
> > will cause an oops.
> > ---
> >  include/net/tc_act/tc_pedit.h |  1 +
> >  net/sched/act_pedit.c         | 23 +++++++++++++++++++++--
> >  2 files changed, 22 insertions(+), 2 deletions(-)
> > 
> > diff --git a/include/net/tc_act/tc_pedit.h b/include/net/tc_act/tc_pedit.h
> > index 748cf87a4d7e..3e02709a1df6 100644
> > --- a/include/net/tc_act/tc_pedit.h
> > +++ b/include/net/tc_act/tc_pedit.h
> > @@ -14,6 +14,7 @@ struct tcf_pedit {
> >  	struct tc_action	common;
> >  	unsigned char		tcfp_nkeys;
> >  	unsigned char		tcfp_flags;
> > +	u32			tcfp_off_max_hint;
> >  	struct tc_pedit_key	*tcfp_keys;
> >  	struct tcf_pedit_key_ex	*tcfp_keys_ex;
> >  };
> > diff --git a/net/sched/act_pedit.c b/net/sched/act_pedit.c
> > index 31fcd279c177..a8ab6c3f1ea2 100644
> > --- a/net/sched/act_pedit.c
> > +++ b/net/sched/act_pedit.c
> > @@ -149,7 +149,7 @@ static int tcf_pedit_init(struct net *net, struct nlattr *nla,
> >  	struct nlattr *pattr;
> >  	struct tcf_pedit *p;
> >  	int ret = 0, err;
> > -	int ksize;
> > +	int i, ksize;
> >  	u32 index;
> >  
> >  	if (!nla) {
> > @@ -228,6 +228,20 @@ static int tcf_pedit_init(struct net *net, struct nlattr *nla,
> >  		p->tcfp_nkeys = parm->nkeys;
> >  	}
> >  	memcpy(p->tcfp_keys, parm->keys, ksize);
> > +	p->tcfp_off_max_hint = 0;
> 
> This gets zeroed here... [1]
> 
> > +	for (i = 0; i < p->tcfp_nkeys; ++i) {
> > +		u32 cur = p->tcfp_keys[i].off;
> > +
> > +		/* The AT option can read a single byte, we can bound the actual
> > +		 * value with uchar max. Each key touches 4 bytes starting from
> > +		 * the computed offset
> > +		 */
> > +		if (p->tcfp_keys[i].offmask) {
> > +			cur += 255 >> p->tcfp_keys[i].shift;
> 
> Could be written as:
> 
> 		cur += (0xff & p->tcfp_keys[i].offmask) >>
> 			p->tcfp_keys[i].shift;
> 
> without the if? That would be closer to the:
> 
> 		offset += (*d & tkey->offmask) >> tkey->shift;
> 
> which ends up getting executed.
> 
> > +			cur = max(p->tcfp_keys[i].at, cur);
> 
> We never write under ->at, tho, so this shouldn't be needed?

Every thing you mentioned looks correct, I'll take care in v2.

> 
> > +		}
> > +		p->tcfp_off_max_hint = max(p->tcfp_off_max_hint, cur + 4);
> > +	}
> >  
> >  	p->tcfp_flags = parm->flags;
> >  	goto_ch = tcf_action_set_ctrlact(*a, parm->action, goto_ch);
> > @@ -308,9 +322,14 @@ static int tcf_pedit_act(struct sk_buff *skb, const struct tc_action *a,
> >  			 struct tcf_result *res)
> >  {
> >  	struct tcf_pedit *p = to_pedit(a);
> > +	u32 max_offset;
> >  	int i;
> >  
> > -	if (skb_unclone(skb, GFP_ATOMIC))
> > +	max_offset = (skb_transport_header_was_set(skb) ?
> > +		      skb_transport_offset(skb) :
> > +		      skb_network_offset(skb)) +
> > +		     p->tcfp_off_max_hint;
> 
> [1] ... and used here outside of the lock. Isn't it racy?

Indeed it is. I'll fix in v2 extending the lock over here (including
the allocation, sigh).

Later, for net-next I think we could refactor the code to avoid the
lock in the data path with something alike c749cdda9089 ("net/sched:
act_skbedit: don't use spinlock in the data path")


Thanks!

Paolo