[dpdk-dev] [RFC v2 1/3] ethdev: add the API for getting trace information

Ferruh Yigit ferruh.yigit at intel.com
Tue Sep 10 17:06:32 CEST 2019


On 9/10/2019 5:36 AM, Wang, Haiyue wrote:
> Thanks Ferruh, Bruce.
> 
>> -----Original Message-----
>> From: Yigit, Ferruh
>> Sent: Monday, September 9, 2019 21:18
>> To: Richardson, Bruce <bruce.richardson at intel.com>
>> Cc: Wang, Haiyue <haiyue.wang at intel.com>; Ray Kinsella <mdr at ashroe.eu>; dev at dpdk.org; Sun, Chenmin
>> <chenmin.sun at intel.com>
>> Subject: Re: [dpdk-dev] [RFC v2 1/3] ethdev: add the API for getting trace information
>>
>> On 9/9/2019 1:50 PM, Ferruh Yigit wrote:
>>> On 9/9/2019 1:40 PM, Bruce Richardson wrote:
>>>> On Mon, Sep 09, 2019 at 12:23:36PM +0100, Ferruh Yigit wrote:
>>>>> On 9/7/2019 3:42 AM, Wang, Haiyue wrote:
>>>>>>> -----Original Message-----
>>>>>>> From: Yigit, Ferruh
>>>>>>> Sent: Friday, September 6, 2019 22:22
>>>>>>> To: Ray Kinsella <mdr at ashroe.eu>; Wang, Haiyue <haiyue.wang at intel.com>
>>>>>>> Cc: dev at dpdk.org
>>>>>>> Subject: Re: [dpdk-dev] [RFC v2 1/3] ethdev: add the API for getting trace information
>>>>>>>
>>>>>>> On 8/13/2019 1:51 PM, Ray Kinsella wrote:
>>>>>>>>
>>>>>>>>
>>>>>>>> On 13/08/2019 04:24, Stephen Hemminger wrote:
>>>>>>>>> On Tue, 13 Aug 2019 11:06:10 +0800
>>>>>>>>> Haiyue Wang <haiyue.wang at intel.com> wrote:
>>>>>>>>>
>>>>>>>>>> Enhance the PMD to support retrieving trace information like
>>>>>>>>>> Rx/Tx burst selection etc.
>>>>>>>>>>
>>>>>>>>>> Signed-off-by: Haiyue Wang <haiyue.wang at intel.com>
>>>>>>>>>> ---
>>>>>>>>>>  lib/librte_ethdev/rte_ethdev.c      | 18 ++++++++++++++++++
>>>>>>>>>>  lib/librte_ethdev/rte_ethdev.h      |  9 +++++++++
>>>>>>>>>>  lib/librte_ethdev/rte_ethdev_core.h |  4 ++++
>>>>>>>>>>  3 files changed, 31 insertions(+)
>>>>>>>>>>
>>>>>>>>>> diff --git a/lib/librte_ethdev/rte_ethdev.c b/lib/librte_ethdev/rte_ethdev.c
>>>>>>>>>> index 17d183e..6098fad 100644
>>>>>>>>>> --- a/lib/librte_ethdev/rte_ethdev.c
>>>>>>>>>> +++ b/lib/librte_ethdev/rte_ethdev.c
>>>>>>>>>> @@ -4083,6 +4083,24 @@ rte_eth_tx_queue_info_get(uint16_t port_id, uint16_t queue_id,
>>>>>>>>>>  }
>>>>>>>>>>
>>>>>>>>>>  int
>>>>>>>>>> +rte_eth_trace_info_get(uint16_t port_id, uint16_t queue_id,
>>>>>>>>>> +		       enum rte_eth_trace type, char *buf, int sz)
>>>>>>>
>>>>>>> Better to use struct as argument instead of individual variables because it is
>>>>>>> easier to extend the struct later if needed.
>>>>>>>
>>>>>>>>>> +{
>>>>>>>>>> +	struct rte_eth_dev *dev;
>>>>>>>>>> +
>>>>>>>>>> +	RTE_ETH_VALID_PORTID_OR_ERR_RET(port_id, -ENODEV);
>>>>>>>>>> +
>>>>>>>>>> +	if (buf == NULL)
>>>>>>>>>> +		return -EINVAL;
>>>>>>>>>> +
>>>>>>>>>> +	dev = &rte_eth_devices[port_id];
>>>>>>>>>> +
>>>>>>>>>> +	RTE_FUNC_PTR_OR_ERR_RET(*dev->dev_ops->trace_info_get, -ENOTSUP);
>>>>>>>>>> +
>>>>>>>>>> +	return dev->dev_ops->trace_info_get(dev, queue_id, type, buf, sz);
>>>>>>>>>
>>>>>>>>> What if queueid is out of bounds?
>>>>>>>>>
>>>>>>>>> The bigger problem is that this information is like a log message
>>>>>>>>> and unstructured, which makes it device specific and useless for automation.
>>>>>>>>
>>>>>>>> IMHO - this is much better implemented as a capability bitfield, that
>>>>>>>> can be queried.
>>>>>>>
>>>>>>> +1 to return the datapath capability as bitfield.
>>>>>>>
>>>>>>> Also +1 to have a new API,
>>>>>>> - I am not sure about the API name, 'rte_eth_trace_info_get()', can we find
>>>>>>> something better instead of 'trace' there.
>>>>>>> - I think we should limit this API only to get current datapath configuration,
>>>>>>> for clarity of the API don't return capability or not datapath related config.
>>>>>>>
>>>>>>> Also this information not always supported in queue level, what do you think
>>>>>>> having ability to get this information in port level,
>>>>>>> like this API can return a struct, which may have a field that says if the
>>>>>>> output is for queue or port, or this can be another bitfield, what do you think?
>>>>>>>
>>>>>>
>>>>>> #define RX_SCALAR	(1ULL < 0)
>>>>>> #define RX_VECTOR_AVX2  ...
>>>>>
>>>>> What about having RX_VECTOR value, later another bit group for the details of
>>>>> the vectorization:
>>>>> SSE
>>>>> AVX2
>>>>> AVX512
>>>>> NEON
>>>>> ALTIVEC
>>>>>
>>>>> Since above options can exist together, what about using values for them instead
>>>>> of bitfields? Reserving 4 bits, 2^4 = 16, can be enough I think for long term.
>>>>>
>>>> Rather than having named vector types, we just need to worry about the ones
>>>> for the current architecture. Therefore I'd suggest just using vector
>>>> widths, one bit each for 16B, 32B and 64B vector support. For supporting
>>>> multiple values, 16 combinations is not enough for all the possibilities.
>>>>
>>>
>>> vector width can be an option too, no objection there. But this is only for
>>> current configuration, so it can be a combination, we have now 5 types and
>>> allocating space for 16.
>>>
>>
>> correction: it can *not* be a combination
> 
> I think we can merge the RX_VECTOR and TX_VECTOR together, use 6 bits for vector
> mode detail. And for vector width, the SSE, NEON name should indicates it ?
> 
> I renamed the definitions to try to make things clear.
> 
> enum rte_eth_burst_mode_option {
> 	BURST_SCALAR = (1 << 0),
> 	BURST_VECTOR = (1 << 1),
> 
> 	BURST_VECTOR_MODE_MASK = (0x3F << 2),
> 	BURST_ALTIVEC          = (1 << 2),
> 	BURST_NEON             = (2 << 2),
> 	BURST_SSE              = (3 << 2),
> 	BURST_AVX2             = (4 << 2),
> 	BURST_AVX512           = (5 << 2),
> 
> 	BURST_SCATTERED = (1 << 8),
> 	BURST_BULK_ALLOC = (1 << 9),
> 	BURST_NORMAL = (1 << 10),
> 	BURST_SIMPLE = (1 << 11),
> };
> 
> /**
>  * Ethernet device RX/TX queue packet burst mode information structure.
>  * Used to retrieve information about packet burst mode setting.
>  */
> struct rte_eth_burst_mode {
> 	uint32_t per_queue_support:1; /**< Support to set per queue burst */
> 
> 	uint64_t options;

We are using first 32bits just to detect the queue level support, what do you
think converting this into a field in 'rte_eth_burst_mode_option' and use
'options' fields, so we will fit into 64 bit.


More information about the dev mailing list