On Thu, Aug 22, 2019 at 5:30 AM Ilya Maximets <[email protected]> wrote:
>
> Tx code doesn't clear the descriptors' status after cleaning.
> So, if the budget is larger than number of used elems in a ring, some
> descriptors will be accounted twice and xsk_umem_complete_tx will move
> prod_tail far beyond the prod_head breaking the comletion queue ring.
s/comletion/completion/
>
> Fix that by limiting the number of descriptors to clean by the number
> of used descriptors in the tx ring.
>
> 'ixgbe_clean_xdp_tx_irq()' function refactored to look more like
> 'ixgbe_xsk_clean_tx_ring()' since we don't need most of the
> complications implemented in the regular 'ixgbe_clean_tx_irq()'
> and we're allowed to directly use 'next_to_clean' and 'next_to_use'
> indexes.
>
> Fixes: 8221c5eba8c1 ("ixgbe: add AF_XDP zero-copy Tx support")
> Signed-off-by: Ilya Maximets <[email protected]>
> ---
Tested-by: William Tu <[email protected]>
Instead of measuring tx performance at the tx machine, I measured the TX
performance at the other side (the traffic generating machine). This time it
is more consistent and showing not much difference with (5.9Mpps) and
without this patch (6.1Mpps).
>
> Version 2:
> * 'ixgbe_clean_xdp_tx_irq()' refactored to look more like
> 'ixgbe_xsk_clean_tx_ring()'.
>
> drivers/net/ethernet/intel/ixgbe/ixgbe_xsk.c | 34 ++++++++------------
> 1 file changed, 13 insertions(+), 21 deletions(-)
>
<snip>