On 7/28/2020 10:09 AM, Jakub Kicinski wrote:
> On Tue, 28 Jul 2020 09:58:44 -0700 Jacob Keller wrote:
>> On 7/28/2020 4:19 AM, Jiri Pirko wrote:
>>> Yes. Documentation is very easy to ignore unfortunatelly. The driver
>>> developer has to be tight up by the core code and api, I believe.
>>
>> So I'm not sure what the best proposal here is. We do have a list of
>> generic components, but given that each piece of HW has different
>> elements, it's not always feasible to have fully generic names. Some of
>> the names are driver specific.
>>
>> I guess we could use some system where components are "registered" when
>> loading the devlink, so that they can be verified by the stack when used
>> as a parameter for flash update? Perhaps take something like the
>> table-driven approach used for infos and extend that into devlink core
>> so that drivers basically register a table of the components which
>> includes both a function callback that gets the version string as well
>> as an indication of whether that component can be updated via flash_update?
>>
>> I know it would also be useful for ice to have a sort of "pre-info"
>> callback that generates a context structure that is passed to each of
>> the info callbacks. (that way a single up-front step could be to lookup
>> the relevant information, and this is then forwarded to each of the
>> formatter functions for each component).
>>
>> Am I on the right track here or just over-engineering?
>
> I don't understand why we're having this conversation.
>
> No driver right now uses the component name.
>
> AFAIU we agreed not to use the component name for config vs code.
>
> So you may as well remove the component name from the devlink op and
> add a todo there saying "when adding component back, make sure it's
> tightly coupled to info".
>
Another simpler option would be to just call .info_get and verify that
the requested component appeared in that list. Ofcourse we'd be making
this call every flash_update...
