Re: [PATCH] flow_dissector: introduce skb_get_hash_symmetric()

[Jon Kohler <jon@nutanix.com>]

> On May 31, 2023, at 2:00 PM, Eric Dumazet <edumazet@google.com> wrote:
> 
> On Wed, May 31, 2023 at 7:47 PM Jon Kohler <jon@nutanix.com> wrote:
>> 
>> 
>> 
>>> On May 31, 2023, at 1:33 PM, Eric Dumazet <edumazet@google.com> wrote:
>>> 
>>> On Wed, May 31, 2023 at 7:22 PM Jon Kohler <jon@nutanix.com> wrote:
>>>> 
>>>> tun.c changed from skb_get_hash() to __skb_get_hash_symmetric() on
>>>> commit feec084a7cf4 ("tun: use symmetric hash"), which exposes an
>>>> overhead for OVS datapath, where ovs_dp_process_packet() has to
>>>> calculate the hash again because __skb_get_hash_symmetric() does not
>>>> retain the hash that it calculates.
>>>> 
>>>> Introduce skb_get_hash_symmetric(), which will get and save the hash
>>>> in one go, so that calcuation work does not go to waste, and plumb it
>>>> into tun.c.
>>>> 
>>>> Fixes: feec084a7cf4 ("tun: use symmetric hash")
>>> 
>>> 
>>>> Signed-off-by: Jon Kohler <jon@nutanix.com>
>>>> CC: Jason Wang <jasowang@redhat.com>
>>>> CC: David S. Miller <davem@davemloft.net>
>>>> ---
>>>> 
>>> 
>>>> diff --git a/include/linux/skbuff.h b/include/linux/skbuff.h
>>>> index 0b40417457cd..8112b1ab5735 100644
>>>> --- a/include/linux/skbuff.h
>>>> +++ b/include/linux/skbuff.h
>>>> @@ -1474,6 +1474,7 @@ __skb_set_sw_hash(struct sk_buff *skb, __u32 hash, bool is_l4)
>>>> 
>>>> void __skb_get_hash(struct sk_buff *skb);
>>>> u32 __skb_get_hash_symmetric(const struct sk_buff *skb);
>>>> +u32 skb_get_hash_symmetric(struct sk_buff *skb);
>>>> u32 skb_get_poff(const struct sk_buff *skb);
>>>> u32 __skb_get_poff(const struct sk_buff *skb, const void *data,
>>>>                  const struct flow_keys_basic *keys, int hlen);
>>>> diff --git a/net/core/flow_dissector.c b/net/core/flow_dissector.c
>>>> index 25fb0bbc310f..d8c0e804bbfe 100644
>>>> --- a/net/core/flow_dissector.c
>>>> +++ b/net/core/flow_dissector.c
>>>> @@ -1747,6 +1747,35 @@ u32 __skb_get_hash_symmetric(const struct sk_buff *skb)
>>>> }
>>>> EXPORT_SYMBOL_GPL(__skb_get_hash_symmetric);
>>>> 
>>>> +/**
>>>> + * skb_get_hash_symmetric: calculate and set a flow hash in @skb, using
>>>> + * flow_keys_dissector_symmetric.
>>>> + * @skb: sk_buff to calculate flow hash from
>>>> + *
>>>> + * This function is similar to __skb_get_hash_symmetric except that it
>>>> + * retains the hash within the skb, such that it can be reused without
>>>> + * being recalculated later.
>>>> + */
>>>> +u32 skb_get_hash_symmetric(struct sk_buff *skb)
>>>> +{
>>>> +       struct flow_keys keys;
>>>> +       u32 hash;
>>>> +
>>>> +       __flow_hash_secret_init();
>>>> +
>>>> +       memset(&keys, 0, sizeof(keys));
>>>> +       __skb_flow_dissect(NULL, skb, &flow_keys_dissector_symmetric,
>>>> +                          &keys, NULL, 0, 0, 0,
>>>> +                          FLOW_DISSECTOR_F_STOP_AT_FLOW_LABEL);
>>>> +
>>>> +       hash = __flow_hash_from_keys(&keys, &hashrnd);
>>>> +
>>>> +       __skb_set_sw_hash(skb, hash, flow_keys_have_l4(&keys));
>>>> +
>>>> +       return hash;
>>>> +}
>>>> +EXPORT_SYMBOL_GPL(skb_get_hash_symmetric);
>>>> +
>>> 
>>> Why copy/pasting __skb_get_hash_symmetric() ?
>>> 
>>> Can you reuse it ?
>> 
>> Not directly, because to use __skb_set_sw_hash requires struct flow_keys
>> when using flow_keys_have_l4(). __skb_get_hash_symmetric() does not
>> take or return that struct, so we’d either have to refactor that (and its callers)
>> or introduce yet another function and consolidate down to that “new one”.
>> 
>> I played around with that exact thought by taking the functional guts out of
>> __skb_get_hash_symmetric, making it a new static function, plumbing that
>> into __skb_get_hash_symmetric and this new skb_get_hash_symmetric, but
>> the LOC churn was basically the same and it felt a bit worse than just a
>> copy/paste.
>> 
>> Alternatively, if it turned out that flow_keys_have_l4() wasn’t important, we
>> Could simply set that to false and then reuse __skb_get_hash_symmetric
>> in a trivial manner. I couldn’t quite figure out if L4 flag was necessary, so I
>> went the safe(maybe?) route and copy/paste instead.
>> 
>> Happy to take suggestions either way!
> 
> There are 6 callers of __skb_get_hash_symmetric()
> 
> I would convert __skb_get_hash_symmetric()  to
> 
> skb_get_hash_symmetric(struct sk_buff *skb, bool record_hash)

Ok, thank you for the suggestion. I’ll work that up as a v2 to this patch.