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

[Jakub Kicinski <kuba@kernel.org>]

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?

> +		}
> +		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?

> +	if (skb_ensure_writable(skb, min(skb->len, max_offset)))
>  		return p->tcf_action;
>  
>  	spin_lock(&p->tcf_lock);