On 2020-09-04 17:35, Jesper Dangaard Brouer wrote:
On Fri, 4 Sep 2020 15:53:31 +0200
Björn Töpel <bjorn.topel@xxxxxxxxx> wrote:
From: Björn Töpel <bjorn.topel@xxxxxxxxx>
Make the AF_XDP zero-copy path aware that the reason for redirect
failure was due to full Rx queue. If so, exit the napi loop as soon as
possible (exit the softirq processing), so that the userspace AF_XDP
process can hopefully empty the Rx queue. This mainly helps the "one
core scenario", where the userland process and Rx softirq processing
is on the same core.
Note that the early exit can only be performed if the "need wakeup"
feature is enabled, because otherwise there is no notification
mechanism available from the kernel side.
This requires that the driver starts using the newly introduced
xdp_do_redirect_ext() and xsk_do_redirect_rx_full() functions.
Signed-off-by: Björn Töpel <bjorn.topel@xxxxxxxxx>
---
drivers/net/ethernet/intel/ixgbe/ixgbe_xsk.c | 23 ++++++++++++++------
1 file changed, 16 insertions(+), 7 deletions(-)
diff --git a/drivers/net/ethernet/intel/ixgbe/ixgbe_xsk.c b/drivers/net/ethernet/intel/ixgbe/ixgbe_xsk.c
index 3771857cf887..a4aebfd986b3 100644
--- a/drivers/net/ethernet/intel/ixgbe/ixgbe_xsk.c
+++ b/drivers/net/ethernet/intel/ixgbe/ixgbe_xsk.c
@@ -93,9 +93,11 @@ int ixgbe_xsk_pool_setup(struct ixgbe_adapter *adapter,
static int ixgbe_run_xdp_zc(struct ixgbe_adapter *adapter,
struct ixgbe_ring *rx_ring,
- struct xdp_buff *xdp)
+ struct xdp_buff *xdp,
+ bool *early_exit)
{
int err, result = IXGBE_XDP_PASS;
+ enum bpf_map_type map_type;
struct bpf_prog *xdp_prog;
struct xdp_frame *xdpf;
u32 act;
@@ -116,8 +118,13 @@ static int ixgbe_run_xdp_zc(struct ixgbe_adapter *adapter,
result = ixgbe_xmit_xdp_ring(adapter, xdpf);
break;
case XDP_REDIRECT:
- err = xdp_do_redirect(rx_ring->netdev, xdp, xdp_prog);
- result = !err ? IXGBE_XDP_REDIR : IXGBE_XDP_CONSUMED;
+ err = xdp_do_redirect_ext(rx_ring->netdev, xdp, xdp_prog, &map_type);
+ if (err) {
+ *early_exit = xsk_do_redirect_rx_full(err, map_type);
Have you tried calling xdp_do_flush (that calls __xsk_map_flush()) and
(I guess) xsk_set_rx_need_wakeup() here, instead of stopping the loop?
(Or doing this in xsk core).
Moving the need_wake logic to the xsk core/flush would be a very nice
cleanup. The driver would still need to pass information from the driver
though. Still, much cleaner. I'll take a stab at that. Thanks!
Looking at the code, the AF_XDP frames are "published" in the queue
rather late for AF_XDP. Maybe in an orthogonal optimization, have you
considered "publishing" the ring producer when e.g. the queue is
half-full?
Hmm, I haven't. You mean instead of yielding, you publish/submit? I
*think* I still prefer stopping the processing.
I'll play with this a bit!
Very nice suggestions, Jesper! Thanks!
Björn
