Re: [PATCH] ethdev: add dump regs for telemetry

[fengchengwen <fengchengwen@huawei.com>]

Hi Ferruh,

On 2024/1/10 20:15, Ferruh Yigit wrote:
> On 1/10/2024 1:38 AM, fengchengwen wrote:
>> Hi Ferruh,
>>
>> On 2024/1/10 2:06, Ferruh Yigit wrote:
>>> On 1/9/2024 2:19 AM, Jie Hai wrote:
>>>> On 2023/12/14 20:49, Ferruh Yigit wrote:
>>>>> On 12/14/2023 1:56 AM, Jie Hai wrote:
>>>>>> The ethdev library now registers a telemetry command for
>>>>>> dump regs.
>>>>>>
>>>>>> An example usage is shown below:
>>>>>> --> /ethdev/regs,test
>>>>>> {
>>>>>>    "/ethdev/regs": {
>>>>>>      "regs_offset": 0,
>>>>>>      "regs_length": 3192,
>>>>>>      "regs_width": 4,
>>>>>>      "device_version": "0x1080f00",
>>>>>>      "regs_file": "port_0_regs_test"
>>>>>>    }
>>>>>> }
>>>>
>>>>>
>>>>> Above code writes register data to a file.
>>>>>
>>>>> I am not sure about this kind of usage of telemetry command, that it
>>>>> cause data to be written to a file.
>>>>>
>>>>> My understanding is, telemetry usage is based on what telemetry client
>>>>> receives.
>>>>> What do you think just keep the 'reg_info' fields excluding data to the
>>>>> file?
>>>>>
>>>>> .Hi, Ferruh
>>>>
>>>> I tried to write all register information to telemetry data,
>>>> but gave up because some drivers had too many registers (eg.ixgbe)
>>>> to carry. Therefore, the writing data to file approach is selected.
>>>>
>>>> When we query a register, the register content is the key.
>>>> The information such as the width and length is only auxiliary
>>>> information. If the register data cannot be obtained, the auxiliary
>>>> information is optional. So I don't think register data should be removed.
>>>>
>>>> In my opinion, writing a file is a more appropriate way to do it.
>>>> I wonder if there's a better way.
>>>>
>>>>
>>>
>>> Is there a usecase to get register information from telemetry interface?
>>
>> Among the available tools:
>> 1, ethtool/proc-info: should use multi-process mechanism to connect to the main process
>> 2, telemetry: easier, lighter load, and it don't need re-probe the ethdev in the secondary process,
>>               and also cost more resource, like hugepage, cores.
>>
>> From our users, they prefer use the second 'telemetry', so I think we should move
>> more status-query-points to telemetry.
>>
>> As for this question, I think it's okay to get register info from telemetry.
>>
>>
>>
>> Another question, we have some internal registers, which:
>> 1. Is not suitable expose by xstats, because they may includes configuration
>> 2. Is not suitable expose by dumps, because this dumps is hard to understand (because it only has value).
>>
>> So we plan to add some telemetry points in the driver itself, so we could display them like xstats:
>> "xxxx" : 0x1234
>> "yyyy" : 0x100
>>
>> Will the community accept this kind of telemetry points which limit one driver ?
>>
> 
> Hi Chengwen,
> 
> I see there is a usecase/requirement.
> 
> With this patch, even using file, only register values are dumped and
> isn't it hard to find value of specific register?
> 
> ("xxxx" : 0x1234) approach looks better, but instead of making this
> telemetry support for specific driver, what about making it in two steps.
> 
> First add new dev_ops, (or update existing one), to get registers with
> "name: value" format, (in a way to allow empty name), or even perhaps
> "name: offset, value" format.
> And in second stage add telemetry support around it.
> (Name being optional lets us wrap exiting 'get_reg' dev_ops with new one)
> 
> When adding dev_ops, it may get an additional 'filter' parameter, to get
> only subset of regs, like "mac*" to get regs name staring with "mac",
> this may help for the cases there are too many registers you mentioned.
> 
> Anyway, we can discuss more about its design, but what do you think
> about first having a dev_ops for this?

I prefer extend struct rte_dev_reg_info, like this:

struct rte_eth_reg_name {
	char name[RTE_ETH_REG_NAME_SIZE];
};

struct rte_dev_reg_info {
	void *data; /**< Buffer for return registers */
	uint32_t offset; /**< Start register table location for access */
	uint32_t length; /**< Number of registers to fetch */
	uint32_t width; /**< Size of device register */
	uint32_t version; /**< Device version */
/* Note: below two fields are new added. */
	char *filter; /**< Filter for target subset of registers. This field could affects register selection for data/length/name.  */
	struct rte_eth_reg_name *names; /**< Registers name saver. */
};

For driver which don't identify the new filter and names fields:
  1. .get_reg return the all registers value.
  2. and driver will not touch the name fields.
  3. rte_eth_dev_get_reg_info() could detect name fileds not filled, and then it fill with default names, e.g. offset-1/offset-2/...

For driver which identify the new filter and names fields:
  1. rte_eth_dev_get_reg_info() will return filtered register's value and also their names.

So that those which invoke rte_eth_dev_get_reg_info() could extra prepare names, and it call the same API will get data and name.


Add one new .get_reg_name ops and corresponding API like: rte_eth_dev_get_reg_name() could also feasible.
But I think the rte_eth_dev_get_reg_info()'s name is too broad, the info could includes value and also it's name.
So I prefer not add one new ops.


Another question? what are the supported values of filters ?
I prefer report by dev_info ops, something like a string array end with NULL.
Use could query from rte_eth_dev_info_get API.

Thanks.

> 
> .
>