Date:	Mon, 12 Jan 2015 17:54:06 +0530
From:	B Viswanath <marichika4@...il.com>
To:	"Arad, Ronen" <ronen.arad@...el.com>
Cc:	Scott Feldman <sfeldma@...il.com>, Netdev <netdev@...r.kernel.org>,
	Jiri Pirko <jiri@...nulli.us>,
	Siva Mannem <siva.mannem.lnx@...il.com>
Subject: Re: Clarification regarding IFLA_BRPORT_LEARNING_SYNC and aging of
 fdb entries learnt via br_fdb_external_learn_add()

On 12 January 2015 at 16:30, Arad, Ronen <ronen.arad@...el.com> wrote:
>
>
>>-----Original Message-----
>>From: Scott Feldman [mailto:sfeldma@...il.com]
>>Sent: Saturday, January 10, 2015 10:29 PM
>>To: Arad, Ronen
>>Cc: Netdev; Jiri Pirko; Siva Mannem; marichika4@...il.com
>>Subject: Re: Clarification regarding IFLA_BRPORT_LEARNING_SYNC and aging of
>>fdb entries learnt via br_fdb_external_learn_add()
>>
>>Perfect, I think we have a good working set of use-cases.  Thanks for
>>adding those.  Your case 3 seems do-able without too much work since
>>we already know which ones where externally added, just need another
>>per-bridge-port flag to turn off bridging aging of externally learned
>>entries.  This will address the performance issue you (and B
>>Viswanath) raised.
>> What about the entry stats, from the user's
>>'bridge -s fdb show" perspective for the bridge fdb entries?  Will
>>these numbers match expectations?  I think case 1 and case 4 provide a
>>coherent stats view.  Case 3 seems to be lacking in this regard.
>>
> I think that statistics accuracy is orthogonal to the mechanism used for aging
> in all the cases where learning sync is enabled (i.e. cases 1, 3, 4).
> Accuracy only depends on how frequent the switch port driver notifies the
> bridge FDB. The best accuracy achievable is with updates once per-second. At
> the other extreme the switch port driver does not refresh entries. It only
> notifies the bridge after entries are removed from the HW. In this extreme case
> the statistics will really show the time since an entry was first learned and
> not the time since it was last re-learned in HW.
> Switch port driver could pick some
> acceptable rate to update the bridge module. Within a typical 5 minutes aging
> interval, updates every 10 seconds or every 30 seconds could be a reasonable
> tradeoff between statistics accuracy and system overhead.
>

I would also like to add that 'whether the end user will really be
interested in the FDB, (and the hit notifications from silicon)'
depends on the actual deployment. In a L3 dominant deployment, people
may not care much about FDB at all. So any time spent inside the
driver trying to update the FDB for statistical reasons will be a
waste of CPU time. On  the other hand for an Access switch (or a
Controller) deployment, FDB may really be useful. Unfortunately, the
driver will not be able to guess the deployment.

I am thinking aloud here, but may be we can let this update-timer to
be a configurable item with default being current behaviour
(determined by driver). It serves both purposes.

Thanks
Vissu

>>On Fri, Jan 9, 2015 at 10:51 PM, Arad, Ronen <ronen.arad@...el.com> wrote:
>>> [...]
>>>>> It is indeed simpler. However, if the overhead of reading hit bits from
>>the
>>>>HW
>>>>> and updating freshness of entries using br_fdb_external_learn_add() is too
>>>>> expensive, it would force such platforms to disable learning sync
>>>>altogether.
>>>>> Therefore, I believe aging offload flag (could be sufficient at bridge
>>>>level)
>>>>> and external aging interval (possibly longer than the software aging
>>>>interval)
>>>>> will encourage drivers to use leaning sync.
>>>>>>-scott
>>>>
>>>>I'm not sure I follow that last part.
>>>>
>>>>Can we list out the use-cases to see what's missing?
>>>>
>>>>Case 1: bridge ages out learned_sync'ed entries
>>>>
>>>>bridge port learning: off
>>>>offload port learning: on
>>>>offload port learning_sync: on
>>>>
>>>>Driver calls br_fdb_external_learn_add() periodically to refresh
>>>>bridge fdb entry
>>>>to keep it from going stale.
>>>>Bridge removes aged out fdb entries (and indirectly tells offload
>>>>device to do the same).
>>>>
>>>>Case 2: offload device/bridge age out entries independently
>>>>
>>>>bridge port learning: on
>>>>offload port learning: on
>>>>offload port learning_sync: off
>>>>
>>>>Bridge ages out its stale learned entries, independent of offload device.
>>>>Offload device ages out its stale learned entries, independent of bridge.
>>>>
>>>>Case 3: ?
>>>>
>>>>Please help me finish the use-case list so we can see what's missing.
>>>
>>>
>>> Case 3: offload device ages out external entries and notifies bridge
>>>
>>> bridge port learning: on or off (Bridge only learns from packets seen
>>(Rx/Tx))
>>> offload port learning: on
>>> offload port learning_sync: on
>>> bridge aging of external learn: off
>>> offload device aging: on
>>>
>>> Switch port/device driver ages entries (could be by HW aging or soft aging
>>in
>>> driver/firmware),
>>> notifies bridge about aged entries using br_fdb_externallearn_del().
>>> This allows efficient HW aging and batched notification at a pace
>>independent
>>> of bridge aging interval.
>>> User still enjoys a single VLAN-aware FDB within the bridge module and
>>having
>>> all entries in one place. Externally learned entries are identified as such
>>by
>>> iproute2 "bridge fdb show" command. Device does not have to implement
>>> ndo_bridge_fdb_dump() for each offload port as the bridge module provides it
>>> for the common FDB.
>>>
>>> Case 4: bridge ages owned and external entries at different intervals
>>>
>>> bridge port learning: on (Effectively only for Rx/Tx traffic seen by
>>>                           software bridge)
>>> offload port learning: on (transient traffic and RxTx, overlap with bridge
>>>                            learned entries possible)
>>> offload port learning_sync: on
>>> bridge aging of external learn: on
>>> offload device aging: off
>>> bridge aging interval for owned entries: T1
>>> bridge aging interval for external entries: T2 (Typically T2 > T1)
>>>
>>> This allows for fine-tuning the overhead of periodic updates of entries
>>> freshness from offload port device.
>>>
>>> The bottom line of cases 3-4 is that it is desirable to use the common
>>bridge
>>> FDB as long as bridge aging of externally learned entries could be
>>controlled
>>> by the offload device: Either be at a longer interval or disabled.
>>>
>>>>
>>>>-scott
>>>>--
>>>>To unsubscribe from this list: send the line "unsubscribe netdev" in
>>>>the body of a message to majordomo@...r.kernel.org
>>>>More majordomo info at  http://vger.kernel.org/majordomo-info.html
--
To unsubscribe from this list: send the line "unsubscribe netdev" in
the body of a message to majordomo@...r.kernel.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html

Powered by blists - more mailing lists