[ovs-dev] [PATCH V9 04/31] tc: Add tc flower functions

[Joe Stringer]

On 3 June 2017 at 22:22, Roi Dayan <roid at mellanox.com> wrote:
>
>
> On 01/06/2017 20:53, Joe Stringer wrote:
>>
>> On 1 June 2017 at 07:39, Roi Dayan <roid at mellanox.com> wrote:
>>>
>>>
>>>
>>> On 31/05/2017 03:50, Joe Stringer wrote:
>>>>
>>>>
>>>> On 28 May 2017 at 04:59, Roi Dayan <roid at mellanox.com> wrote:
>>>>>
>>>>>
>>>>> Add tc helper functions to query and manipulate the flower classifier.
>>>>>
>>>>> Signed-off-by: Paul Blakey <paulb at mellanox.com>
>>>>> Signed-off-by: Roi Dayan <roid at mellanox.com>
>>>>> ---
>>>>
>>>>
>>>>
>>>> Again is this co-authored? utilities/checkpatch.py checks for stuff like
>>>> this.
>>>>
>>>> I didn't go through all of the enums and make sure they're all
>>>> covered, correct types, etc.
>>>>
>>>> <snip>
>>>>
>>>>> +    [TCA_FLOWER_KEY_ENC_IPV4_SRC_MASK] = { .type = NL_A_U32, .optional
>>>>> =
>>>>> true, },
>>>>> +    [TCA_FLOWER_KEY_ENC_IPV4_DST_MASK] = { .type = NL_A_U32, .optional
>>>>> =
>>>>> true, },
>>>>
>>>>
>>>>
>>>> Line lengths.
>>>>
>>>
>>> ok
>>>
>>>> <snip>
>>>>
>>>>> +static void
>>>>> +nl_parse_tcf(const struct tcf_t *tm, struct tc_flower *flower)
>>>>> +{
>>>>> +    unsigned long long int lastuse = tm->lastuse * 10;
>>>>
>>>>
>>>>
>>>> Where does the 10 come from? What does it mean? Should we have a #define
>>>> for it?
>>>>
>>>
>>> we want to convert here jiffies to ms and used some default value.
>>> as Flavio suggested maybe do it in a macro?
>>> later I saw in netdev-linux.c the functions read_psched() and
>>> tc_ticks_to_bytes(). can do a followup commit to refactor those
>>> somewhere else and use them in tc.c as well.
>>
>>
>> Please do.
>
>
> just to be clear here, is it ok to use a macro for this series
> and do the followup commit after this series?

That sounds reasonable.

>>>>> +int
>>>>> +tc_replace_flower(int ifindex, uint16_t prio, uint32_t handle,
>>>>> +                  struct tc_flower *flower)
>>>>> +{
>>>>> +    struct ofpbuf request;
>>>>> +    struct tcmsg *tcmsg;
>>>>> +    struct ofpbuf *reply;
>>>>> +    int error = 0;
>>>>> +    size_t basic_offset;
>>>>> +    uint16_t eth_type = (OVS_FORCE uint16_t) flower->key.eth_type;
>>>>
>>>>
>>>>
>>>> Why does this need forcing?
>>>>
>>>
>>> as eth_type is ove_be16 but tc_make_handle wants unsigned int.
>>> we get a compilation warning from sparse about incorrect type.
>>> this is to avoid sparse from failing.
>>
>>
>> I see. Should it be unsigned int, then? Rather than casting a
>> big-endian value to store a host-endian variable of the same size?
>>
>
> it's not unsigned int to follow up with the rest of the user space code
> that use ovs_be16 for eth_type.
> also we are using match_set_dl_type() that expect ovs_be16.

The rest of the code stores a big-endian eth_type in an ovs_be16, or
host byte-order version in uint16_t. The thing I found surprising in
this code was that the big endian version of the ethertype was being
placed into a uint16_t without converting to host byte order. In the
end, I think what we're trying to achieve is that the second parameter
to tc_make_handle() should be a big-endian ethertype but sparse would
complain if we passed it directly (because tc_make_handle()'s second
argument type isn't ovs_be16). So OVS_FORCE is necessary, though I
would have expected it to be used inside the arguments of
tc_make_handle().