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

["Wang, Haiyue" <haiyue.wang@intel.com>]

> -----Original Message-----
> From: Yigit, Ferruh
> Sent: Tuesday, September 10, 2019 17:15
> To: Wang, Haiyue <haiyue.wang@intel.com>; Richardson, Bruce <bruce.richardson@intel.com>
> Cc: Ray Kinsella <mdr@ashroe.eu>; dev@dpdk.org; Sun, Chenmin <chenmin.sun@intel.com>
> Subject: Re: [dpdk-dev] [RFC v2 1/3] ethdev: add the API for getting trace information
> 
> On 9/10/2019 9:37 AM, Wang, Haiyue wrote:
> >> -----Original Message-----
> >> From: Yigit, Ferruh
> >> Sent: Tuesday, September 10, 2019 16:07
> >> To: Wang, Haiyue <haiyue.wang@intel.com>; Richardson, Bruce <bruce.richardson@intel.com>
> >> Cc: Ray Kinsella <mdr@ashroe.eu>; dev@dpdk.org; Sun, Chenmin <chenmin.sun@intel.com>
> >> Subject: Re: [dpdk-dev] [RFC v2 1/3] ethdev: add the API for getting trace information
> >>
> >> 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@intel.com>
> >>>> Cc: Wang, Haiyue <haiyue.wang@intel.com>; Ray Kinsella <mdr@ashroe.eu>; dev@dpdk.org; Sun,
> Chenmin
> >>>> <chenmin.sun@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@ashroe.eu>; Wang, Haiyue <haiyue.wang@intel.com>
> >>>>>>>>> Cc: dev@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@intel.com> wrote:
> >>>>>>>>>>>
> >>>>>>>>>>>> Enhance the PMD to support retrieving trace information like
> >>>>>>>>>>>> Rx/Tx burst selection etc.
> >>>>>>>>>>>>
> >>>>>>>>>>>> Signed-off-by: Haiyue Wang <haiyue.wang@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),
> >>
> >> Do we need to have bitfields for this, I was suggesting reserve 4 bits, bit 2-5
> >> (inclusive) and use their value:
> >>
> >> BURST_VECTOR_MODE_IDX  = 2
> >> BURST_VECTOR_MODE_SIZE = 4
> >> BURST_VECTOR_MODE_MASK =
> >> 	((1 << BURST_VECTOR_MODE_SIZE) - 1) << BURST_VECTOR_MODE_IDX
> >>
> >> vector_mode = (options & BURST_VECTOR_MODE_MASK) >> BURST_VECTOR_MODE_IDX
> >>
> >> if (vector_mode == 0) // BURST_SSE
> >> if (vector_mode == 1) // BURST_AVX2
> >> if (vector_mode == 2) // BURST_AVX512
> >> if (vector_mode == 3) // BURST_NEON
> >> ....
> >>
> >> Can any vector mode be combination of above, if not why use bitfields?
> >>
> >
> > I use it as this to *set* ...
> >
> > 	else if (pkt_burst == i40e_recv_scattered_pkts_vec_avx2)
> > 		options = BURST_VECTOR | BURST_AVX2 | BURST_SCATTERED;
> > 	else if (pkt_burst == i40e_recv_pkts_vec_avx2)
> > 		options = BURST_VECTOR | BURST_AVX2;
> > 	else if (pkt_burst == i40e_recv_scattered_pkts_vec)
> > 		options = BURST_VECTOR | BURST_SSE | BURST_SCATTERED;
> > 	else if (pkt_burst == i40e_recv_pkts_vec)
> > 		options = BURST_VECTOR | BURST_SSE;
> >
> > Then *get* like this, since we reserve the bit group.
> >
> > static void
> > burst_mode_options_display(uint64_t options)
> > {
> > 	uint64_t vec_mode = options & BURST_VECTOR_MODE_MASK;
> > 	uint64_t opt;
> >
> > 	options &= ~BURST_VECTOR_MODE_MASK;
> >
> > 	for (opt = 1; options != 0; opt <<= 1, options >>= 1) {
> > 		if (!(options & 1))
> > 			continue;
> >
> > 		printf(" %s", rte_eth_burst_mode_option_name(opt));
> >
> > 		if (opt == BURST_VECTOR)
> > 			printf("(%s)",
> > 			       rte_eth_burst_mode_option_name(vec_mode));
> > 	}
> > }
> >
> 
> I can see how you intended use it, only they don't need to be bitfield and using
> with value saves bits.
> Also I think good to reserve some bits for future modes.
> 

"BURST_VECTOR_MODE_MASK = (0x3F << 2)" has reserved 63 non-zero bits on position 2 ~ 7.
Then from bit 8, a new definition: BURST_SCATTERED = (1 << 8).

"using with value saves bits" -- Sorry, I didn't get the point. :-(
vector_mode = (options & BURST_VECTOR_MODE_MASK) >> BURST_VECTOR_MODE_IDX

From above, 'vector_mode's bits are from 'options' bits stream, how to save bits ?
In my understanding, this is some kind of more-bit-field, not each-bit-field.

I defined them together, so can quick check the vector type, like
(options & BURST_VECTOR_MODE_MASK) == BURST_NEON.

> >>
> >>>
> >>> 	BURST_SCATTERED = (1 << 8),
> >>> 	BURST_BULK_ALLOC = (1 << 9),
> >>> 	BURST_NORMAL = (1 << 10),
> >>
> >> Not sure about this one, what is the difference between scalar?
> >>
> >
> > Extract it from the function name and the debug message.
> >
> > 	if (pkt_burst == i40e_recv_scattered_pkts)
> > 		options = BURST_SCALAR | BURST_SCATTERED;
> > 	else if (pkt_burst == i40e_recv_pkts_bulk_alloc)
> > 		options = BURST_SCALAR | BURST_BULK_ALLOC;
> > 	else if (pkt_burst == i40e_recv_pkts)
> > 		options = BURST_SCALAR | BURST_NORMAL;
> 
> What is the difference between 'BURST_SCALAR' & "BURST_SCALAR | BURST_NORMAL" ?

IMO, "SCALAR" should be "non-Vector" ? Like "BURST_VECTOR" will append with
"SSE/AVX2" etc, "SCALAR" will append with other option bits. "Normal" is just
handing the Descriptor one by one as *normal*. As I said, I got this name idea
from the original log to try cover the right burst behaviors. :)

> 
> btw, for actual implementation please add 'RTE_ETH_' prefix.
> 
Got it, will add them.

> >
> >>> 	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;
> >>> };
> >>>
> >>> And three APIs:
> >>>
> >>> 1.
> >>> __rte_experimental
> >>> int rte_eth_rx_burst_mode_get(uint16_t port_id, uint16_t queue_id,
> >>> 	struct rte_eth_burst_mode *mode);
> >>>
> >>>
> >>> 2.
> >>> __rte_experimental
> >>> int rte_eth_tx_burst_mode_get(uint16_t port_id, uint16_t queue_id,
> >>> 	struct rte_eth_burst_mode *mode);
> >>>
> >>> 3.
> >>> __rte_experimental
> >>> const char *
> >>> rte_eth_burst_mode_option_name(uint64_t option);
> >>
> >> What about 'rte_eth_burst_mode_name()' ?
> >>
> >
> > The "mode" scope is bigger than "mode_option", so I defined it as "_mode_option_name()".
> >
> > struct rte_eth_burst_mode {
> > 	uint32_t per_queue_support:1; /**< Support to set per queue burst */
> >
> > 	uint64_t options;
> > };
> 
> Agreed the scope is bigger in implementation, but "burst mode option name" is
> same as "burst mode name" for user, so removing it may make easier for user.
> But since the API is generating name from 'options' variable, instead of
> directly from the port, OK to keep API name as you suggested.
> 
> >
> >>>
> >>>
> >>> PMD two ops:
> >>>
> >>> typedef void (*eth_burst_mode_get_t)(struct rte_eth_dev *dev,
> >>> 	uint16_t queue_id, struct rte_eth_burst_mode *mode);
> >>>
> >>> struct eth_dev_ops {
> >>> 	...
> >>> 	eth_burst_mode_get_t       rx_burst_mode_get; /**< Get RX burst mode */
> >>> 	eth_burst_mode_get_t       tx_burst_mode_get; /**< Get TX burst mode */
> >>> 	...
> >>> };
> >>>
> >