Re: Does ACL support field size of 8 bytes?

[Konstantin Ananyev <konstantin.v.ananyev@yandex.ru>]

11/05/2022 15:28, Ido Goshen пишет:
> 
> 
>> -----Original Message-----
>> From: Ananyev, Konstantin <konstantin.ananyev@intel.com>
>> Sent: Tuesday, 26 April 2022 20:57
>> To: Ido Goshen <Ido@cgstowernetworks.com>; users@dpdk.org;
>> dev@dpdk.org
>> Cc: Konstantin Ananyev <konstantin.v.ananyev@yandex.ru>
>> Subject: RE: Does ACL support field size of 8 bytes?
>>
>>
>> Hi Ido,
>>
>>> I've lots of good experience with ACL but can't make it work with u64
>>> values I know it can be split to 2xu32 fields, but it makes it more
>>> complex to use and a wastes double  number of fields (we hit the
>>> RTE_ACL_MAX_FIELDS 64 limit)
>>
>> Wow, that's a lot of fields...
> 
> [idog]
> We provide a general purpose packet-broker that covers wide range of
> l2-l4 protocols + tunnels + some app level metadata.
> Though in most cases they won't be used simultaneously and many fields
> may end up being don't-care (e.g. mask=0) it's easier to code a static
> rte_acl_field_def struct that covers all the options then constructing it
> dynamically on each user configuration change
>   
>>> According to the documentation and rte_acl.h fields size can be 8[idog]
>>> ...
>>> Should it work?
>>> Did anyone try it successfully and/or can share an example?
>>
>> You are right: though it is formally supported, we do not test it, and AFAIK no-
>> one used it till now.
>> As we do group fields by 4B long chunks anyway, 8B field is sort of awkward and
>> confusing.
>> To be honest, I don't even remember what was the rationale beyond introducing
>> it at first place.
>> Anyway, just submitted patches that should fix 8B field support (at least it works
>> for me now):
>> https://patches.dpdk.org/project/dpdk/list/?series=22676
>> Please give it a try.
> 
> [idog] The patch works great for me. Thanx!
> 
>> In long term it probably would be good to hear from you and other users, should
>> we keep 8B support at all, or might be it would be easier just to abandon it.
>> Thanks
>> Konstantin
> 
> [idog] I find the 8B option very useful:
> 1. It's easier and more natural to use for long size fields
> e.g. part of how it simplifies our MAC rules code
> 
> @@ -231,48 +231,34 @@ struct rte_acl_field_def acl_fields[] = {
>                  {
>                                  .type = RTE_ACL_FIELD_TYPE_BITMASK,
> -                               .size = sizeof(uint32_t),
> -                               .field_index = FIELD_MAC_SRC_4MSB,
> +                               .size = sizeof(uint64_t),
> +                               .field_index = FIELD_MAC_SRC,
>                                  .input_index = INPUT_INDEX_GROUP_2,
>                                  .offset = offsetof(struct acl_data, mac_src),
>                  },
> -              {
> -                                .type = RTE_ACL_FIELD_TYPE_BITMASK,
> -                               .size = sizeof(uint16_t),
> -                               .field_index = FIELD_MAC_SRC_2LSB,
> -                               .input_index = INPUT_INDEX_GROUP_3,
> -                               .offset = offsetof(struct acl_data, mac_src) + sizeof(uint32_t),
> -               },
> .
> .
> .
> +static int get_mac_val(const char *in, uint64_t *mac)
>   {
> -       static const size_t MAC_4MSB_SIZE = sizeof(uint32_t);
> -       static const size_t MAC_2LSB_SIZE = sizeof(uint16_t);
>          uint32_t i = 0;
>          uint8_t octet = 0;
>          char dlm = ':';
> -
> -       for (i = 0; i < MAC_4MSB_SIZE; i++)
> -       {
> -               GET_CB_FIELD(in, octet, 16, UINT8_MAX, dlm);
> -               ((uint8_t*)mac4msb)[MAC_4MSB_SIZE - 1 - i] = octet;
> -       }
> -       for (i = 0; i < MAC_2LSB_SIZE; i++)
> +       *mac = 0;
> +       for (i = 0; i < RTE_ETHER_ADDR_LEN; i++)
>          {
> -               if (i ==  MAC_2LSB_SIZE - 1)
> +               if (i == RTE_ETHER_ADDR_LEN - 1)
>                          dlm = 0;
>                  GET_CB_FIELD(in, octet, 16, UINT8_MAX, dlm);
> -               ((uint8_t*)mac2lsb)[MAC_2LSB_SIZE - 1 - i] = octet;
> +               ((uint8_t*)mac)[RTE_ETHER_ADDR_LEN + 1 - i] = octet;
>          }
>          return 0;
>   }
> 
> It' even much more significant for RTE_ACL_FIELD_TYPE_RANGE that may require
> breaking a single U64 range to 3 U32 based rules

My concern was it is sort of awkward in terms of input_field
value for rules with 8B long.
But sure, if you believe it is useful, then let's try to keep it.
I submitted v2, there is no change in the library itself,
just updated the test script to cover new case.
If you'll have a chance, please add 'tested-by:' tag to it.

> 
> 2. It may save acl fields
> Alternative is to increase RTE_ACL_MAX_FIELDS (maybe expose it to rte_config.h)
> As long as the "64" doesn't stand for some algorithmic/performance reason

I kept RTE_ACL_MAX_FIELDS, but internally had to increase max number of 
input fields up to 2 * RTE_ACL_MAX_FIELDS, to cover the situation
when all fields are 8B long.

Thanks
Konstantin