Hi Jerin,
On 5/30/2018 12:56 PM, Jerin Jacob wrote:
-----Original Message-----
Hi Nikhil,
I think, it is reasonable to have Tx adapter.
Some top level comments to starts with,
1) Slow path API looks fine. The only comment is, How about changing
rte_event_eth_tx_adapter_queue_add() instead of
rte_event_eth_tx_adapter_queue_start()
to get align with Rx adapter?
OK.
2) This my understanding of fastpath
a) Common code will have a extra port(given through adapter create)
where all workers invoke rte_event_enqueue_burst() to that port and then common
code
dequeue from that port and send packet
using rte_eth_tx_burst() and/or using tx buffering APIs
The workers invoke rte_event_enqueue_burst() to their local port not to
the extra port as you described. The queue ID specified when
enqueuing is linked to the the adapter's port, the adapter reads these
events and transmits mbufs on the
ethernet port and queue specified in these mbufs. The diagram below
illustrates what I just described.
+------+
| | +----+
|Worker+-->+port+--+
| | +----+ | +----+
+------+ | +-->+eth0|
| +---------+ +-------+ | +----+
+--+ | +----+ | +---+ +----+
| Queue +-->+port+-->+Adapter|------>+eth1|
+--+ | +----+ | +---+ +----+
+------+ | +---------+ +-------+ | +----+
| | +----+ | +-->+eth2|
|Worker+-->+port+--+ +----+
| | +----+
+------+
b) If the source queue or sched_type is ORDERED, When it enqueue to the extra
port it
will change to atomic/direct to maintain the the ingress order if need.
Couple of issues and a minor change in RFC to fix the issues as a proposal.
Issue 1) I think, Using mbuf private data for such purpose will be a problem as
exiting application already may using for it own needs and there will be
additional cache
miss etc on fastpath to get private data.
Instead of using mbuf private data, the eth port and queue can be
specified in
the mbuf itself using mbuf:port and mbuf:hash respectively.
Issue 2) At least on Cavium, We can do so optimization by introducing
the same RTE_EVENT_ETH_TX_ADAPTER_CAP_INTERNAL_PORT schematics of Rx
adapter. i.e we don't need extra port to converge all traffic from
different workers and transmit.
OK.
Considering above two issues2, We would like propose an fastpath API,
which semantics is almost same rte_ethdev_tx_burst().
a) It will get ride of mbuf metadata need
b) New fastpath API gives driver enough possibilities of optimization if
possible.(address the issue (2))
The API can be like,
uint16_t rte_event_eth_tx_adapter_enqueue(uint16_t eth_port_id, uint16_t
eth_queue_id, uint8_t event_port_id, const struct rte_event ev[], uint16_t
nb_events);
The worker core will receive events pointing to mbufs that need to be
transmitted to different
ports/queues, as described above. The port and the queue will be
populated in the mbuf and the
API can be as below
uint16_t rte_event_eth_tx_adapter_enqueue(uint8_t instance_id, uint8_t
event_port_id, const struct rte_event ev[], uint16_t nb_events);
Let me know if that works for you.
This API would land in driver as function pointer if device has
RTE_EVENT_ETH_TX_ADAPTER_CAP_INTERNAL_PORT capability.
Agreed.
If device does not have !RTE_EVENT_ETH_TX_ADAPTER_CAP_INTERNAL_PORT then
common code can choose to ignore event_port_id in
rte_event_eth_tx_adapter_enqueue() and select the port given in adapter
create in common code.
The common code will still use the local worker port.
So from the top-level, each worker lcore will look like
while(1)
{
rte_event_dequeue_burst(..,worker[port],..);
(process the packets)
rte_event_eth_tx_adapter_enqueue(... worker[port],...);
}
Agreed, with the API modification above
Nikhil
