Re: [PATCH NET-NEXT] ipv6: skb_expand_head() adjust skb->truesize incorrectly

[Vasily Averin]

On 8/24/21 1:23 AM, Eric Dumazet wrote:
> On 8/23/21 2:51 PM, Eric Dumazet wrote:
>> On 8/23/21 2:45 PM, Eric Dumazet wrote:
>>> On 8/23/21 10:25 AM, Christoph Paasch wrote:
>>>> Hello,
>>>>
>>>> On Mon, Aug 23, 2021 at 12:56 AM Vasily Averin <vvs@xxxxxxxxxxxxx> wrote:
>>>>>
>>>>> Christoph Paasch reports [1] about incorrect skb->truesize
>>>>> after skb_expand_head() call in ip6_xmit.
>>>>> This happen because skb_set_owner_w() for newly clone skb is called
>>>>> too early, before pskb_expand_head() where truesize is adjusted for
>>>>> (!skb-sk) case.
>>>>>
>>>>> [1] https://lkml.org/lkml/2021/8/20/1082
>>>>>
>>>>> Reported-by: Christoph Paasch <christoph.paasch@xxxxxxxxx>
>>>>> Signed-off-by: Vasily Averin <vvs@xxxxxxxxxxxxx>
>>>>> ---
>>>>>  net/core/skbuff.c | 24 +++++++++++++-----------
>>>>>  1 file changed, 13 insertions(+), 11 deletions(-)
>>>>>
>>>>> diff --git a/net/core/skbuff.c b/net/core/skbuff.c
>>>>> index f931176..508d5c4 100644
>>>>> --- a/net/core/skbuff.c
>>>>> +++ b/net/core/skbuff.c
>>>>> @@ -1803,6 +1803,8 @@ struct sk_buff *skb_realloc_headroom(struct sk_buff *skb, unsigned int headroom)
>>>>>
>>>>>  struct sk_buff *skb_expand_head(struct sk_buff *skb, unsigned int headroom)
>>>>>  {
>>>>> +       struct sk_buff *oskb = skb;
>>>>> +       struct sk_buff *nskb = NULL;
>>>>>         int delta = headroom - skb_headroom(skb);
>>>>>
>>>>>         if (WARN_ONCE(delta <= 0,
>>>>> @@ -1811,21 +1813,21 @@ struct sk_buff *skb_expand_head(struct sk_buff *skb, unsigned int headroom)
>>>>>
>>>>>         /* pskb_expand_head() might crash, if skb is shared */
>>>>>         if (skb_shared(skb)) {
>>>>> -               struct sk_buff *nskb = skb_clone(skb, GFP_ATOMIC);
>>>>> -
>>>>> -               if (likely(nskb)) {
>>>>> -                       if (skb->sk)
>>>>> -                               skb_set_owner_w(nskb, skb->sk);
>>>>> -                       consume_skb(skb);
>>>>> -               } else {
>>>>> -                       kfree_skb(skb);
>>>>> -               }
>>>>> +               nskb = skb_clone(skb, GFP_ATOMIC);
>>>>>                 skb = nskb;
>>>>>         }
>>>>>         if (skb &&
>>>>> -           pskb_expand_head(skb, SKB_DATA_ALIGN(delta), 0, GFP_ATOMIC)) {
>>>>> -               kfree_skb(skb);
>>>>> +           pskb_expand_head(skb, SKB_DATA_ALIGN(delta), 0, GFP_ATOMIC))
>>>>>                 skb = NULL;
>>>>> +
>>>>> +       if (!skb) {
>>>>> +               kfree_skb(oskb);
>>>>> +               if (nskb)
>>>>> +                       kfree_skb(nskb);
>>>>> +       } else if (nskb) {
>>>>> +               if (oskb->sk)
>>>>> +                       skb_set_owner_w(nskb, oskb->sk);
>>>>> +               consume_skb(oskb);
>>>>
>>>> sorry, this does not fix the problem. The syzkaller repro still
>>>> triggers the WARN.
>>>>
>>>> When it happens, the skb in ip6_xmit() is not shared as it comes from
>>>> __tcp_transmit_skb, where it is skb_clone()'d.
>>>>
>>>>
>>>
>>> Old code (in skb_realloc_headroom())
>>> was first calling skb2 = skb_clone(skb, GFP_ATOMIC); 
>>>
>>> At this point, skb2->sk was NULL
>>> So pskb_expand_head(skb2, SKB_DATA_ALIGN(delta), 0, ...) was able to tweak skb2->truesize
>>>
>>> I would try :
>>>
>>> diff --git a/net/core/skbuff.c b/net/core/skbuff.c
>>> index f9311762cc475bd38d87c33e988d7c983b902e56..326749a8938637b044a616cc33b6a19ed191ac41 100644
>>> --- a/net/core/skbuff.c
>>> +++ b/net/core/skbuff.c
>>> @@ -1804,6 +1804,7 @@ EXPORT_SYMBOL(skb_realloc_headroom);
>>>  struct sk_buff *skb_expand_head(struct sk_buff *skb, unsigned int headroom)
>>>  {
>>>         int delta = headroom - skb_headroom(skb);
>>> +       struct sk_buff *oskb = NULL;
>>>  
>>>         if (WARN_ONCE(delta <= 0,
>>>                       "%s is expecting an increase in the headroom", __func__))
>>> @@ -1813,19 +1814,21 @@ struct sk_buff *skb_expand_head(struct sk_buff *skb, unsigned int headroom)
>>>         if (skb_shared(skb)) {
>>>                 struct sk_buff *nskb = skb_clone(skb, GFP_ATOMIC);
>>>  
>>> -               if (likely(nskb)) {
>>> -                       if (skb->sk)
>>> -                               skb_set_owner_w(nskb, skb->sk);
>>> -                       consume_skb(skb);
>>> -               } else {
>>> +               if (unlikely(!nskb)) {
>>>                         kfree_skb(skb);
>>> +                       return NULL;
>>>                 }
>>> +               oskb = skb;
>>>                 skb = nskb;
>>>         }
>>> -       if (skb &&
>>> -           pskb_expand_head(skb, SKB_DATA_ALIGN(delta), 0, GFP_ATOMIC)) {
>>> +       if (pskb_expand_head(skb, SKB_DATA_ALIGN(delta), 0, GFP_ATOMIC)) {
>>>                 kfree_skb(skb);
>>> -               skb = NULL;
>>> +               kfree_skb(oskb);
>>> +               return NULL;
>>> +       }
>>> +       if (oskb) {
>>> +               skb_set_owner_w(skb, oskb->sk);
>>> +               consume_skb(oskb);
>>>         }
>>>         return skb;
>>>  }
>> Oh well, probably not going to work.
>>
>> We have to find a way to properly increase skb->truesize, even if skb_clone() is _not_ called.
> 
> I also note that current use of skb_set_owner_w(), forcing skb->destructor to sock_wfree()
> is probably breaking TCP Small queues, since original skb->destructor would be tcp_wfree() or __sock_wfree()

I asked Alexey Kuznetsov to look at this problem. Below is his answer:
"I think the current scheme is obsolete. It was created
when we had only two kinds of skb accounting (rmem & wmem)
and with more kinds of accounting it just does not work.
Even there we had ignored problems with adjusting accounting.

Logically the best solution would be replacing ->destructor,
set_owner* etc with skb_ops. Something like:

struct skb_ops
{
        void init(struct sk_buff * skb, struct skb_ops * ops, struct
sock * owner);
        void fini(struct sk_buff * skb);
        void update(struct sk_buff * skb, int adjust);
        void inherit(struct sk_buff * skb2, struct sk_buff * skb);
};

init - is replacement for skb_set_owner_r|w
fini - is replacement for skb_orphan
update - is new operation to be used in places where skb->truesize changes,
       instead of awful constructions like:

       if (!skb->sk || skb->destructor == sock_edemux)
            skb->truesize += size - osize;

       Now it will look like:

       if (skb->ops)
            skb->ops->update(skb, size - osize);

inherit - is replacement for also awful constructs like:

      if (skb->sk)
            skb_set_owner_w(skb2, skb->sk);

      Now it will be:

      if (skb->ops)
            skb->ops->inherit(skb2, skb);

The implementation looks mostly obvious.
Some troubles can be only with new functionality:
update of accounting was never done before.


More efficient, functionally equivalent, but uglier and less flexible
alternative would be removal of ->destructor, replaced with
a small numeric indicator of ownership:

enum
{
        SKB_OWNER_NONE,  /* aka destructor == NULL */
        SKB_OWNER_WMEM,  /* aka destructor == sk_wfree */
        SKB_OWNER_RMEM,  /* aka destructor == sk_rfree */
        SKB_OWNER_SK,    /* aka destructor == sk_edemux */
        SKB_OWNER_TCP,   /* aka destructor == tcp_wfree */
}

And the same init,fini,inherit,update become functions
w/o any inidirect calls. Not sure it is really more efficient though."

Thank you,
	Vasily Averin