Hi Wei,
On Mon, Sep 23, 2024 at 04:59:56AM +0300, Wei Fang wrote:
Okay, I have tested this solution (see changes below), and from what I observed, the xdp_tx_in_flight can naturally drop to 0 in every test. So if there are no other risks, the next version will use this solution.
Sorry for the delay. I have tested this variant and it works. Just one thing below.
@@ -2467,10 +2469,6 @@ void enetc_start(struct net_device *ndev) struct enetc_ndev_priv *priv = netdev_priv(ndev); int i;
enetc_setup_interrupts(priv);enetc_enable_tx_bdrs(priv);for (i = 0; i < priv->bdr_int_num; i++) { int irq = pci_irq_vector(priv->si->pdev, ENETC_BDR_INT_BASE_IDX + i);@@ -2479,6 +2477,10 @@ void enetc_start(struct net_device *ndev) enable_irq(irq); }
enetc_setup_interrupts(priv);enetc_enable_tx_bdrs(priv);enetc_enable_rx_bdrs(priv); netif_tx_start_all_queues(ndev);@@ -2547,6 +2549,12 @@ void enetc_stop(struct net_device *ndev)
enetc_disable_rx_bdrs(priv);
enetc_wait_bdrs(priv);enetc_disable_tx_bdrs(priv);enetc_clear_interrupts(priv);
Here, NAPI may still be scheduled. So if you clear interrupts, enetc_poll() on another CPU might still have time to re-enable them. This makes the call pointless.
Please move the enetc_clear_interrupts() call after the for() loop below (AKA leave it where it is).
for (i = 0; i < priv->bdr_int_num; i++) { int irq = pci_irq_vector(priv->si->pdev, ENETC_BDR_INT_BASE_IDX + i);@@ -2555,12 +2563,6 @@ void enetc_stop(struct net_device *ndev) napi_synchronize(&priv->int_vector[i]->napi); napi_disable(&priv->int_vector[i]->napi); }
enetc_wait_bdrs(priv);enetc_disable_tx_bdrs(priv);enetc_clear_interrupts(priv);} EXPORT_SYMBOL_GPL(enetc_stop);
FWIW, there are at least 2 other valid ways of solving this problem. One has already been mentioned (reset the counter in enetc_free_rx_ring()):
@@ -2014,6 +2015,8 @@ static void enetc_free_rx_ring(struct enetc_bdr *rx_ring) __free_page(rx_swbd->page); rx_swbd->page = NULL; } + + rx_ring->xdp.xdp_tx_in_flight = 0; }
static void enetc_free_rxtx_rings(struct enetc_ndev_priv *priv)
And the other would be to keep rescheduling NAPI until there are no more pending XDP_TX frames.
diff --git a/drivers/net/ethernet/freescale/enetc/enetc.c b/drivers/net/ethernet/freescale/enetc/enetc.c index 3cff76923ab9..36520f8c49a6 100644 --- a/drivers/net/ethernet/freescale/enetc/enetc.c +++ b/drivers/net/ethernet/freescale/enetc/enetc.c @@ -1689,6 +1689,7 @@ static int enetc_poll(struct napi_struct *napi, int budget) work_done = enetc_clean_rx_ring_xdp(rx_ring, napi, budget, prog); else work_done = enetc_clean_rx_ring(rx_ring, napi, budget); - if (work_done == budget) + if (work_done == budget || rx_ring->xdp.xdp_tx_in_flight) complete = false; if (work_done)
But I like your second proposal the best. It doesn't involve adding an unnecessary extra test in the fast path.