Re: [dpdk-dev] [PATCH] ethdev: fix queue mapping documentation

[Andrew Rybchenko <arybchenko@solarflare.com>]

On 10.07.2018 09:20, Jerin Jacob wrote:
> -----Original Message-----
>> Date: Mon, 2 Jul 2018 16:45:33 +0100
>> From: Ferruh Yigit <ferruh.yigit@intel.com>
>> To: Andrew Rybchenko <arybchenko@solarflare.com>, Jerin Jacob
>>   <jerin.jacob@caviumnetworks.com>, dev@dpdk.org
>> CC: thomas@monjalon.net, stable@dpdk.org
>> Subject: Re: [dpdk-dev] [PATCH] ethdev: fix queue mapping documentation
>> User-Agent: Mozilla/5.0 (Windows NT 10.0; WOW64; rv:52.0) Gecko/20100101
>>   Thunderbird/52.8.0
>>
>>
>> On 7/2/2018 4:32 PM, Andrew Rybchenko wrote:
>>> On 07/02/2018 06:08 PM, Ferruh Yigit wrote:
>>>> On 6/29/2018 10:44 AM, Jerin Jacob wrote:
>>>>> The RTE_MAX_ETHPORT_QUEUE_STATS_MAPS does not exists, change
>>>>> to the correct definition(RTE_ETHDEV_QUEUE_STAT_CNTRS)
>>>>>
>>>>> Fixes: 5de201df8927 ("ethdev: add stats per queue")
>>>>> Cc: stable@dpdk.org
>>>>>
>>>>> Signed-off-by: Jerin Jacob <jerin.jacob@caviumnetworks.com>
>>>>> ---
>>>>>   lib/librte_ethdev/rte_ethdev.h | 4 ++--
>>>>>   1 file changed, 2 insertions(+), 2 deletions(-)
>>>>>
>>>>> diff --git a/lib/librte_ethdev/rte_ethdev.h b/lib/librte_ethdev/rte_ethdev.h
>>>>> index 36e3984ea..375ea24ce 100644
>>>>> --- a/lib/librte_ethdev/rte_ethdev.h
>>>>> +++ b/lib/librte_ethdev/rte_ethdev.h
>>>>> @@ -2144,7 +2144,7 @@ void rte_eth_xstats_reset(uint16_t port_id);
>>>>>    * @param stat_idx
>>>>>    *   The per-queue packet statistics functionality number that the transmit
>>>>>    *   queue is to be assigned.
>>>>> - *   The value must be in the range [0, RTE_MAX_ETHPORT_QUEUE_STATS_MAPS - 1].
>>>>> + *   The value must be in the range [0, RTE_ETHDEV_QUEUE_STAT_CNTRS - 1].
>>>> Yes RTE_MAX_ETHPORT_QUEUE_STATS_MAPS doesn't exits and comment is wrong, but
>>>> RTE_ETHDEV_QUEUE_STAT_CNTRS also slightly not correct.
>>>>
>>>> I think how testpmd uses it increase the confusion.
>>>>
>>>> In ixgbe there is no stats registers per queue, 128 queues are represented by 16
>>>> register set. stat_idx here is the index of that 16 registers. You map queue to
>>>> stats requester to get queue stats.
>>>>
>>>> Also there is RTE_ETHDEV_QUEUE_STAT_CNTRS config in the ethdev API, which is the
>>>> hardcoded size of the queue stats, its default value is 16. This limits number
>>>> of the queues we can get stats from but saves allocated space. (Why not dynamic?)
>>>>
>>>> You can increase the RTE_ETHDEV_QUEUE_STAT_CNTRS to the max supported number of
>>>> queue and ethdev code will be all valid. But "stat_idx" can't go beyond 16 (for
>>>> ixgbe) because it is hardware limitation and it may change from hw to hw.
>>>>
>>>> Also technically it should be possible to reduce RTE_ETHDEV_QUEUE_STAT_CNTRS to
>>>> a low number, like 2, but in ixgbe map two queues into stat registers 14 & 15
>>>> and display those two set as queue stat 0 and 1. It seems current implementation
>>>> prevents this and forces the queues mapped should be less than
>>>> RTE_ETHDEV_QUEUE_STAT_CNTRS. Overall it seems there is a mixed used of
>>>> RTE_ETHDEV_QUEUE_STAT_CNTRS and stats queue index values, I assume because both
>>>> are same values.
>>>>
>>>> I suggest updating it as:
>>>> "
>>>> The value must be in the range:
>>>>   [0 - MIN(HW Stat Registers Size, RTE_ETHDEV_QUEUE_STAT_CNTRS) - 1]
>>>> "
>>> Technically I think it is not a problem to specify more than HW supports.
>>> The function should simply return error.  RTE_ETHDEV_QUEUE_STAT_CNTRS is
>>> a hard limit which should be checked by ethdev.
>>> The reasonable next question is how to find out what is the maximum for PMD/HW.
>>> I think it deserves entry in dev_info. May be not now.
>> Yes there is not a way to find out that limit by application, setting
>> RTE_ETHDEV_QUEUE_STAT_CNTRS to 16 and using it as limit solving the issue for now :)
>
> If I understand it correctly, in the documentation, we are specify the
> limits to avoid the segfault etc and if the specific PMD does not
> support the range up to RTE_ETHDEV_QUEUE_STAT_CNTRS, It can simply return
> error which makes the documentation semantically correct.
>
> Considering the above point, I think this patch is correct considering
> there is no way currently to detect the limit supported by PMD. So,
> 1) Should we keep the patch as is?
>
> --
> The value must be in the range [0, RTE_ETHDEV_QUEUE_STAT_CNTRS - 1].
> --
>
> OR
>
> 2) Change to
>
> --
> The value must be in the range:
> [0 - MIN(Device max per Tx queue stats, RTE_ETHDEV_QUEUE_STAT_CNTRS) - 1]
> --
>
> I prefer option 1. But I am okay send v2 if any ethdev maintainers prefer
> option 2.
>
> Let me know.

I would prefer option 1 as well since right now we have no way to
advertise/find out "Device max per Tx queue stats".

Andrew.