[EXT] Re: [PATCH v6 2/6] ethdev: add trace points for ethdev (part one)

[Ferruh Yigit]

On 2/1/2023 3:42 PM, Ankur Dwivedi wrote:
> 
> 
>> -----Original Message-----
>> From: Ferruh Yigit <ferruh.yigit at amd.com>
>> Sent: Wednesday, February 1, 2023 12:08 AM
>> To: Ankur Dwivedi <adwivedi at marvell.com>; dev at dpdk.org; David
>> Marchand <david.marchand at redhat.com>; Jerin Jacob Kollanukkaran
>> <jerinj at marvell.com>; Andrew Rybchenko
>> <andrew.rybchenko at oktetlabs.ru>
>> Cc: thomas at monjalon.net; mdr at ashroe.eu; orika at nvidia.com;
>> chas3 at att.com; humin29 at huawei.com; linville at tuxdriver.com;
>> ciara.loftus at intel.com; qi.z.zhang at intel.com; mw at semihalf.com;
>> mk at semihalf.com; shaibran at amazon.com; evgenys at amazon.com;
>> igorch at amazon.com; chandu at amd.com; Igor Russkikh
>> <irusskikh at marvell.com>; shepard.siegel at atomicrules.com;
>> ed.czeck at atomicrules.com; john.miller at atomicrules.com;
>> ajit.khaparde at broadcom.com; somnath.kotur at broadcom.com; Jerin Jacob
>> Kollanukkaran <jerinj at marvell.com>; Maciej Czekaj [C]
>> <mczekaj at marvell.com>; Shijith Thotton <sthotton at marvell.com>;
>> Srisivasubramanian Srinivasan <srinivasan at marvell.com>; Harman Kalra
>> <hkalra at marvell.com>; rahul.lakkireddy at chelsio.com; johndale at cisco.com;
>> hyonkim at cisco.com; liudongdong3 at huawei.com;
>> yisen.zhuang at huawei.com; xuanziyang2 at huawei.com;
>> cloud.wangxiaoyun at huawei.com; zhouguoyang at huawei.com;
>> simei.su at intel.com; wenjun1.wu at intel.com; qiming.yang at intel.com;
>> Yuying.Zhang at intel.com; beilei.xing at intel.com; xiao.w.wang at intel.com;
>> jingjing.wu at intel.com; junfeng.guo at intel.com; rosen.xu at intel.com; Nithin
>> Kumar Dabilpuram <ndabilpuram at marvell.com>; Kiran Kumar Kokkilagadda
>> <kirankumark at marvell.com>; Sunil Kumar Kori <skori at marvell.com>; Satha
>> Koteswara Rao Kottidi <skoteshwar at marvell.com>; Liron Himi
>> <lironh at marvell.com>; zr at semihalf.com; Radha Chintakuntla
>> <radhac at marvell.com>; Veerasenareddy Burru <vburru at marvell.com>;
>> Sathesh B Edara <sedara at marvell.com>; matan at nvidia.com;
>> viacheslavo at nvidia.com; longli at microsoft.com; spinler at cesnet.cz;
>> chaoyong.he at corigine.com; niklas.soderlund at corigine.com;
>> hemant.agrawal at nxp.com; sachin.saxena at oss.nxp.com; g.singh at nxp.com;
>> apeksha.gupta at nxp.com; sachin.saxena at nxp.com; aboyer at pensando.io;
>> Rasesh Mody <rmody at marvell.com>; Shahed Shaikh
>> <shshaikh at marvell.com>; Devendra Singh Rawat
>> <dsinghrawat at marvell.com>; andrew.rybchenko at oktetlabs.ru;
>> jiawenwu at trustnetic.com; jianwang at trustnetic.com;
>> jbehrens at vmware.com; maxime.coquelin at redhat.com;
>> chenbo.xia at intel.com; steven.webster at windriver.com;
>> matt.peters at windriver.com; bruce.richardson at intel.com;
>> mtetsuyah at gmail.com; grive at u256.net; jasvinder.singh at intel.com;
>> cristian.dumitrescu at intel.com; jgrajcia at cisco.com;
>> mb at smartsharesystems.com
>> Subject: Re: [EXT] Re: [PATCH v6 2/6] ethdev: add trace points for ethdev (part
>> one)
>>
>> On 1/30/2023 4:01 PM, Ankur Dwivedi wrote:
>>
>> <...>
>>
>>>>> diff --git a/lib/ethdev/meson.build b/lib/ethdev/meson.build index
>>>>> 39250b5da1..f5c0865023 100644
>>>>> --- a/lib/ethdev/meson.build
>>>>> +++ b/lib/ethdev/meson.build
>>>>> @@ -24,6 +24,7 @@ headers = files(
>>>>>          'rte_ethdev.h',
>>>>>          'rte_ethdev_trace.h',
>>>>>          'rte_ethdev_trace_fp.h',
>>>>> +        'rte_ethdev_trace_fp_burst.h',
>>>>
>>>> Why these trace headers are public?
>>>> Aren't trace points only used by the APIs, so I expect them to be
>>>> internal, so applications shouldn't need them. Why they are expsed to
>> user.
>>> 'rte_ethdev_trace.h' can be made as internal. Not sure about
>> 'rte_ethdev_trace_fp.h' and 'rte_ethdev_trace_fp_burst.h' as the tracepoints
>> in fast path may be called from public inline functions.
>>
>> Trace calls used by inline functions needs to be public, in this case at least
>> 'rte_ethdev_trace_fp_burst.h' needs to be public.
>>
>> Can you please at least move all trace points that are called by inline functions
>> to the same file, 'rte_ethdev_trace_fp_burst.h', to reduce number of the
>> header files to make public? Feel free to rename header if required.
>>
>> Meanwhile not sure about adding new header as dependency to end user.
>> @Jerin, @Andrew, what do you think
>> 1) to move these trace points to 'rte_ethdev_core.h'
>> OR
>> 2) disable trace calls in inline functions on compile time, possibly with existing
>> 'RTE_ETHDEV_DEBUG_*' macro
>>
>> <...>
>>
>>>>> -	if (port_id >= RTE_MAX_ETHPORTS)
>>>>> +	if (port_id >= RTE_MAX_ETHPORTS) {
>>>>> +		rte_eth_trace_find_next(RTE_MAX_ETHPORTS);
>>>>
>>>> Is there a specific reason to trace all paths, why not just keep the last one?
>>> This can be kept as the function returns with RTE_MAX_ETHPORTS here.
>>
>> 'RTE_MAX_ETHPORTS' more like error path, that is returned when no more
>> valid port left, since most of the trace doesn't record error return values, I
>> thought this can be dropped as well unless there is an explicit need for it.
> Ok.
>>
>> <...>
>>
>>>>> +RTE_TRACE_POINT(
>>>>> +	rte_eth_trace_rx_hairpin_queue_setup,
>>>>> +	RTE_TRACE_POINT_ARGS(uint16_t port_id, uint16_t rx_queue_id,
>>>>> +		uint16_t nb_rx_desc, const struct rte_eth_hairpin_conf *conf,
>>>>> +		int ret),
>>>>> +	uint32_t peer_count = conf->peer_count;
>>>>> +	uint32_t tx_explicit = conf->tx_explicit;
>>>>> +	uint32_t manual_bind = conf->manual_bind;
>>>>> +	uint32_t use_locked_device_memory = conf-
>>>>> use_locked_device_memory;
>>>>> +	uint32_t use_rte_memory = conf->use_rte_memory;
>>>>> +	uint32_t force_memory = conf->force_memory;
>>>>> +
>>>>> +	rte_trace_point_emit_u16(port_id);
>>>>> +	rte_trace_point_emit_u16(rx_queue_id);
>>>>> +	rte_trace_point_emit_u16(nb_rx_desc);
>>>>> +	rte_trace_point_emit_ptr(conf);
>>>>> +	rte_trace_point_emit_u32(peer_count);
>>>>> +	rte_trace_point_emit_u32(tx_explicit);
>>>>> +	rte_trace_point_emit_u32(manual_bind);
>>>>> +	rte_trace_point_emit_u32(use_locked_device_memory);
>>>>> +	rte_trace_point_emit_u32(use_rte_memory);
>>>>> +	rte_trace_point_emit_u32(force_memory);
>>>>> +	rte_trace_point_emit_int(ret);
>>>>
>>>> Do we need temporary variables like 'peer_count', why not directly use
>> them:
>>> Temporary variables are needed where the actual variable is a bitfield.
>>> For example in struct rte_eth_hairpin_conf:
>>> struct rte_eth_hairpin_conf {
>>> uint32_t peer_count:16;
>>> ....
>>> uint32_t tx_explicit:1
>>> ....
>>> };
>>>
>>> Otherwise there is a compilation error.
>>> ../lib/ethdev/rte_ethdev_trace.h:253:9: note: in expansion of macro
>> ‘rte_trace_point_emit_u16’
>>>   253 |         rte_trace_point_emit_u16(conf->peer_count);
>>>       |         ^~~~~~~~~~~~~~~~~~~~~~~~
>>> ../lib/eal/include/rte_trace_point.h:369:38: error: ‘sizeof’ applied to a bit-
>> field
>>>   369 |         mem = RTE_PTR_ADD(mem, sizeof(in)); \
>>>
>>
>> Got it, that is because of bitfields. But as I look the 'rte_trace_point_emit_u16'
>> macro, it is like:
>>
>> ```
>> #define rte_trace_point_emit_u16(in) __rte_trace_point_emit(in, uint16_t)
>>
>> #define __rte_trace_point_emit(in, type) \ do { \
>>        memcpy(mem, &(in), sizeof(in)); \
>>        mem = RTE_PTR_ADD(mem, sizeof(in)); \ } while (0) ```
>>
>> Here, in '__rte_trace_point_emit' macro 'type' is not used at all, if so what is
>> the point of passing type?
>>
>> If 'sizeof(type)' used, instead of 'sizeof(in)', it would be possible to use
>> bitfields directly, what do you think?
> 
> After using 'sizeof(type)', the following compile time error is observed. 
> In file included from ../lib/ethdev/rte_ethdev_trace_fp_burst.h:18,
>                  from ../lib/ethdev/rte_ethdev.h:175,
>                  from ../lib/ethdev/rte_tm.c:8:
> ../lib/ethdev/rte_ethdev_trace.h: In function ‘rte_eth_trace_rx_hairpin_queue_setup’:
> ../lib/eal/include/rte_trace_point.h:378:21: error: cannot take address of bit-field ‘peer_count’
>   378 |         memcpy(mem, &(in), sizeof(type)); \
>           |                                     ^

Right, you can't memcpy to bitfiled, so temporary variables are
required, OK.

Still, is it OK to ignore 'type' in the '__rte_trace_point_emit()'?
If variable (in) is uint8_t, it won't matter if
'rte_trace_point_emit_u8', 'rte_trace_point_emit_u16' or
'rte_trace_point_emit_u32' used, they will all give same result, right?


>>
>>
>>
>> Also, as for the "struct rte_eth_hairpin_conf" sample, some of the bitfields are
>> 1 bit length, does 'uint32_t' really needed for them?
> 
> uint8_t should suffice.

ack

>>
>> <...>
>>
>>>>> +RTE_TRACE_POINT_FP(
>>>>> +	rte_eth_trace_find_next,
>>>>> +	RTE_TRACE_POINT_ARGS(uint16_t port_id),
>>>>> +	rte_trace_point_emit_u16(port_id);
>>>>> +)
>>>>> +
>>>>
>>>> Why 'rte_eth_trace_find_next' added as fast path?
>>>> Can you please add comment for all why it is added as fast path, this
>>>> help to evaluate/review this later.
>>>
>>> There were many functions for which I was not sure about whether they
>> should be slow path or fast path. I made the following assumption:
>>>
>>> For slow path I have chosen the function which do some setup, configure or
>> write some configuration. For an example
>> rte_eth_trace_tx_hairpin_queue_setup,
>> rte_eth_trace_tx_buffer_set_err_callback,
>> rte_eth_trace_promiscuous_enable are slow path functions.
>>>
>>> The functions which read data are made as fastpath functions. Also for
>> functions for which I was not sure I made it as fastpath.
>>>
>>> For an example rte_ethdev_trace_owner_get,
>> rte_eth_trace_hairpin_get_peer_port, rte_eth_trace_macaddr_get are made
>> as fastpath.
>>>
>>> But there are few exceptions. Function like *_get_capability are made as
>> slowpath. Also rte_ethdev_trace_info_get is slowpath.
>>>
>>> The slowpath and fastpath functions are in separate files.
>> rte_ethdev_trace.h (slowpath) and rte_ethdev_trace_fp.h (fastpath).
>>>
>>> Please let me  know if any function needs to be swapped. I will make that
>> change.
>>>
>>
>> Got it, I think most of the trace points in the 'rte_ethdev_trace_fp.h'
>> are for control/helper functions like: 'rte_ethdev_trace_count_avail',
>> 'rte_ethdev_trace_get_port_by_name', 'rte_eth_trace_promiscuous_get' ...
>>
>> I thought you did based on some analysis, that is why I asked to add that
>> reasoning as code comment.
>>
>>
>> I think we can generalize as:
>>
>> 1) Anything called by ethdev static inline functions are datapath, and must be
>> 'RTE_TRACE_POINT_FP', like 'rte_eth_trace_call_[rt]x_callbacks',
>> 'rte_ethdev_trace_[rt]x_burst',
> 
> In this category the following functions come:
> rte_eth_rx_burst
> rte_eth_tx_burst
> rte_eth_call_rx_callbacks (called from rte_eth_rx_burst)
> rte_eth_call_tx_callbacks (called from rte_eth_tx_burst)
> rte_eth_tx_buffer_count_callback (registered as error callback, called from rte_eth_tx_buffer_flush)
> rte_eth_tx_buffer_drop_callback (registered as error callback)

ack

>>
>> 2) Anything that is called in endless loop in application/sample that has
>> potential impact although it may not really be datapath
> 
> Apart from functions in category [1], I have observed the following functions in ethdev library, called in some while loop in app/examples.
> rte_eth_stats_get (called in while loop in examples/qos_sched and examples/distributor)
> rte_eth_macaddr_get (called in while loop in examples/bond and examples/ethtool)
> rte_eth_link_get (called in for loop in examples/ip_pipeline)
> rte_eth_dev_get_mtu (called in for loop in examples/ip_pipeline)
> rte_eth_link_speed_to_str (called in for loop in examples/ip_pipeline)
> rte_eth_dev_rx_intr_enable ( called in loop in examples/l3fwd-power)
> rte_eth_dev_rx_intr_disable ( called in loop in examples/l3fwd-power)
> rte_eth_timesync_read_rx_timestamp (called in loop in examples/ptpclient)
> rte_eth_timesync_read_tx_timestamp (called in loop in examples/ptpclient)
> rte_eth_timesync_adjust_time (called in loop in examples/ptpclient)
> rte_eth_timesync_read_time (called in loop in examples/ptpclient)
> rte_flow_classifier_query (called in examples/flow_classify)
> rte_mtr_create (in app/test-flow-perf loop)
> rte_mtr_destroy (in app/test-flow-perf loop)
> rte_mtr_meter_policy_delete ((in app/test-flow-perf loop)
> rte_flow_create (in app/test-flow-perf)
> rte_flow_destroy (in app/test-flow-perf)
> 

Ack, and can you please add the note within the parenthesis as a
comment, whoever visits these later knows why there trace points added
as fast path trace point?

> Apart from the above can all other functions be moved to slowpath tracepoints?

I think yes, we can start with this.
At least this gives us a logic to follow.

And does trace point and fast path trace points needs to be in separate
header files?
Can you please put first group into a header an expose it to the user,
rest (as a mixture) to another header and keep it internal to ethdev?

Thanks.

> I think it was discussed in tech board meeting on 2022-11-30 that the functions for which the slowpath or fastpath
> distinction is not clear, they should be fastpath tracepoints.
> 
>>
>> 3) Rest are regular trace points, RTE_TRACE_POINT.
>>
>>
>> Item (2) requires some analysis and whenever you found one of them please
>> comment the reasoning in the code where trace point is defined.
>