From: Maciej Fijalkowski <maciej.fijalkow...@intel.com>
Date: Tue, 11 Jun 2024 16:21:27 +0200
> On Tue, Jun 11, 2024 at 01:59:37PM +0200, Alexander Lobakin wrote:
>> From: Maciej Fijalkowski <maciej.fijalkow...@intel.com>
>> Date: Tue, 4 Jun 2024 15:21:48 +0200
>>
>>> From: Michal Kubiak <michal.kub...@intel.com>
>>>
>>> Address a scenario in which XSK ZC Tx produces descriptors to XDP Tx
>>> ring when link is either not yet fully initialized or process of
>>> stopping the netdev has already started. To avoid this, add checks
>>> against carrier readiness in ice_xsk_wakeup() and in ice_xmit_zc().
>>> One could argue that bailing out early in ice_xsk_wakeup() would be
>>> sufficient but given the fact that we produce Tx descriptors on behalf
>>> of NAPI that is triggered for Rx traffic, the latter is also needed.
>>>
>>> Bringing link up is an asynchronous event executed within
>>> ice_service_task so even though interface has been brought up there is
>>> still a time frame where link is not yet ok.
>>>
>>> Without this patch, when AF_XDP ZC Tx is used simultaneously with stack
>>> Tx, Tx timeouts occur after going through link flap (admin brings
>>> interface down then up again). HW seem to be unable to transmit
>>> descriptor to the wire after HW tail register bump which in turn causes
>>> bit __QUEUE_STATE_STACK_XOFF to be set forever as
>>> netdev_tx_completed_queue() sees no cleaned bytes on the input.
>>>
>>> Fixes: 126cdfe1007a ("ice: xsk: Improve AF_XDP ZC Tx and use batching API")
>>> Fixes: 2d4238f55697 ("ice: Add support for AF_XDP")
>>> Reviewed-by: Shannon Nelson <shannon.nel...@amd.com>
>>> Tested-by: Chandan Kumar Rout <chandanx.r...@intel.com> (A Contingent
>>> Worker at Intel)
>>> Signed-off-by: Michal Kubiak <michal.kub...@intel.com>
>>> Signed-off-by: Maciej Fijalkowski <maciej.fijalkow...@intel.com>
>>> ---
>>> drivers/net/ethernet/intel/ice/ice_xsk.c | 6 +++++-
>>> 1 file changed, 5 insertions(+), 1 deletion(-)
>>>
>>> diff --git a/drivers/net/ethernet/intel/ice/ice_xsk.c
>>> b/drivers/net/ethernet/intel/ice/ice_xsk.c
>>> index 2015f66b0cf9..1bd4b054dd80 100644
>>> --- a/drivers/net/ethernet/intel/ice/ice_xsk.c
>>> +++ b/drivers/net/ethernet/intel/ice/ice_xsk.c
>>> @@ -1048,6 +1048,10 @@ bool ice_xmit_zc(struct ice_tx_ring *xdp_ring)
>>>
>>> ice_clean_xdp_irq_zc(xdp_ring);
>>>
>>> + if (!netif_carrier_ok(xdp_ring->vsi->netdev) ||
>>> + !netif_running(xdp_ring->vsi->netdev))
>>> + return true;
>>
>> Why is it checked after clean_xdp_irq_zc()?
>
> There's nothing wrong with cleaning descriptors that have been sent
> previously. We don't touch anything HW nor netstack related there, just
> bumping ntc and producing CQ descriptors, both ops are pure SW things.
Sure, but do we need to do that if we don't send anything this time?
Lazy cleaning and all that :p
Thanks,
Olek
