Re: [PATCH net-next] openvswitch: reduce stack usage in do_execute_actions

[Ilya Maximets]

On 9/21/23 21:08, Eric Dumazet wrote:
> On Thu, Sep 21, 2023 at 9:03 PM Ilya Maximets <i.maximets@xxxxxxx> wrote:
>>
>> do_execute_actions() function can be called recursively multiple
>> times while executing actions that require pipeline forking or
>> recirculations.  It may also be re-entered multiple times if the packet
>> leaves openvswitch module and re-enters it through a different port.
>>
>> Currently, there is a 256-byte array allocated on stack in this
>> function that is supposed to hold NSH header.  Compilers tend to
>> pre-allocate that space right at the beginning of the function:
>>
>>      a88:       48 81 ec b0 01 00 00    sub    $0x1b0,%rsp
>>
>> NSH is not a very common protocol, but the space is allocated on every
>> recursive call or re-entry multiplying the wasted stack space.
>>
>> Move the stack allocation to push_nsh() function that is only used
>> if NSH actions are actually present.  push_nsh() is also a simple
>> function without a possibility for re-entry, so the stack is returned
>> right away.
>>
>> With this change the preallocated space is reduced by 256 B per call:
>>
>>      b18:       48 81 ec b0 00 00 00    sub    $0xb0,%rsp
>>
>> Signed-off-by: Ilya Maximets <i.maximets@xxxxxxx>
>> ---
>>  net/openvswitch/actions.c | 20 +++++++++-----------
>>  1 file changed, 9 insertions(+), 11 deletions(-)
>>
>> diff --git a/net/openvswitch/actions.c b/net/openvswitch/actions.c
>> index 5f8094acd056..80cc5c512d7b 100644
>> --- a/net/openvswitch/actions.c
>> +++ b/net/openvswitch/actions.c
>> @@ -312,10 +312,16 @@ static int push_eth(struct sk_buff *skb, struct sw_flow_key *key,
>>  }
>>
>>  static int push_nsh(struct sk_buff *skb, struct sw_flow_key *key,
>> -                   const struct nshhdr *nh)
>> +                   const struct nlattr *a)
> 
> Presumably this function should be inlined. (one caller only)
> 
> I would add noinline_for_stack to make sure the compiler will not play
> games with this attempt.

Yeah, good point!  I didn't see it being inlined in my testing, but it's
better to be sure.  I'll post v2 with a flag.

> 
>>  {
>> +       u8 buffer[NSH_HDR_MAX_LEN];
>> +       struct nshhdr *nh = (struct nshhdr *)buffer;
>>         int err;
>>
>> +       err = nsh_hdr_from_nlattr(a, nh, NSH_HDR_MAX_LEN);
>> +       if (err)
>> +               return err;
>> +
>>         err = nsh_push(skb, nh);
>>         if (err)
>>                 return err;
>> @@ -1439,17 +1445,9 @@ static int do_execute_actions(struct datapath *dp, struct sk_buff *skb,
>>                         err = pop_eth(skb, key);
>>                         break;
>>
>> -               case OVS_ACTION_ATTR_PUSH_NSH: {
>> -                       u8 buffer[NSH_HDR_MAX_LEN];
>> -                       struct nshhdr *nh = (struct nshhdr *)buffer;
>> -
>> -                       err = nsh_hdr_from_nlattr(nla_data(a), nh,
>> -                                                 NSH_HDR_MAX_LEN);
>> -                       if (unlikely(err))
>> -                               break;
>> -                       err = push_nsh(skb, key, nh);
>> +               case OVS_ACTION_ATTR_PUSH_NSH:
>> +                       err = push_nsh(skb, key, nla_data(a));
>>                         break;
>> -               }
>>
>>                 case OVS_ACTION_ATTR_POP_NSH:
>>                         err = pop_nsh(skb, key);
>> --
>> 2.41.0
>>