On 7/31/19 4:28 PM, Jakub Kicinski wrote:
> On Wed, 31 Jul 2019 16:07:31 -0600, David Ahern wrote:
>> On 7/31/19 4:02 PM, Jakub Kicinski wrote:
>>> Can you elaborate further? Ports for most purposes are represented by
>>> netdevices. Devlink port instances expose global topological view of
>>> the ports which is primarily relevant if you can see the entire ASIC.
>>> I think the global configuration and global view of resources is still
>>> the most relevant need, so in your diagram you must account for some
>>> "all-seeing" instance, e.g.:
>>>
>>> namespace 1 | namespace 2 | ... | namespace N
>>> | | |
>>> { ports 1 } | { ports 2 } | ... | { ports N }
>>> | | |
>>> subdevlink 1 | subdevlink 2 | ... | subdevlink N
>>> \______ | _______/
>>> master ASIC devlink
>>> =================================================
>>> driver
>>>
>>> No?
>>
>> sure, there could be a master devlink visible to the user if that makes
>> sense or the driver can account for it behind the scenes as the sum of
>> the devlink instances.
>>
>> The goal is to allow ports within an asic [1] to be divided across
>> network namespace where each namespace sees a subset of the ports. This
>> allows creating multiple logical switches from a single physical asic.
>>
>> [1] within constraints imposed by the driver/hardware - for example to
>> account for resources shared by a set of ports. e.g., front panel ports
>> 1 - 4 have shared resources and must always be in the same devlink instance.
>
> So the ASIC would start out all partitioned? Presumably some would
> still like to use it non-partitioned? What follows there must be a top
> level instance to decide on partitioning, and moving resources between
> sub-instances.
>
> Right now I don't think there is much info in devlink ports which would
> be relevant without full view of the ASIC..
>
not sure how it would play out. really just 'thinking out loud' about
the above use case to make sure devlink with proper namespace support
allows it - or does not prevent it.
