On 10/17/2017 2:32 PM, Alexander Duyck wrote:
So the select_queue function being needed is the deal breaker on all
of this as far as I am concerned. We aren't allowed to use it under
other cases so why should macvlan be an exception to the rule?
I realize that the stack is pretty good at chosing the "right" queue,
which is my understanding as to why we shouldn't use select_queue(), but
it doesn't know how to use the accel_priv context associated with the
macvlan offload.
I saw DaveM's guidance to the HiNIC folks when they tried to add
select_queue(): "do not implement this function unless you absolutely
need to do something custom in your driver". I can see where this might
be the exception.
When originally thinking about how to do this, I wanted to use the
accel_priv as a pointer to the VSI to be used for the offload, then we
could have multiple queues and use all the VSI specific tuning
operations that XL710 has available. It can work when selecting the
queue, but by the time you get to start_xmit(), you no longer have that
context and only have the queue number. You can't do any fancy encoding
in the queue number because the value has to be within
dev->num_tx_queues. Maybe we can add accel_priv to the start_xmit
interface? (I can hear the groans already...)
However... for our case, you might be right anyway. If the stack is
doing its job at keeping the conversation on the one queue/irq/cpu
combination, any Tx following the offloaded Rx might already be headed
for the right Tx queue. I'll check on that.
I think we should probably look at a different approach for this. For
example why is it we need to use a different transmit path for a
macvlan packet vs any other packet? On the Rx side we get the
advantage of avoiding the software hashing and demux. What do we get
for reserving queues for transmit?
There are a couple of reasons I can think of to keep the Tx on the
specific queue pair:
- Keep the Tx traffic on the same CPU and irq as the Rx traffic
- Don't let the flow get interrupted, slowed, or otherwise perturbed by
other traffic flows.
- Allow for adding hardware assisted bandwidth constraints to the
offloaded flow without bothering the rest of the NIC's traffic
Are these enough to want to guarantee the Tx queue?
My plan for this is to go back and "fix" ixgbe so we can get it away
from having to use the select_queue call for the macvlan offload and
then maybe look at proving a few select NDO operations for allowing
macvlans that are being offloaded to make specific calls into the
hardware to perform tasks as needed.
The ixgbe implementation can certainly be improved. I think its biggest
failing is that the rest of the general traffic gets constrained to a
single queue - no more RSS for load balancing.
sln
