r8169: improve rtl8169_start_xmit
Heiner Kallweit
maintainerPATCH 1/1 r8169: improve rtl8169_start_xmit
Improve the following in rtl8169_start_xmit:
- tp->cur_tx can be accessed in parallel by rtl_tx(), therefore annotate the race by using WRITE_ONCE
- avoid checking stop_queue a second time by moving the doorbell check
- netif_stop_queue() uses atomic operation set_bit() that includes a full memory barrier on some platforms, therefore use smp_mb__after_atomic to avoid overhead
--- drivers/net/ethernet/realtek/r8169_main.c | 15 ++++++--------- 1 file changed, 6 insertions(+), 9 deletions(-)
drivers/net/ethernet/realtek/r8169_main.c +6 -9
@@ -4226,7 +4226,7 @@ static netdev_tx_t rtl8169_start_xmit(struct sk_buff *skb,/* rtl_tx needs to see descriptor changes before updated tp->cur_tx */ smp_wmb (); - tp->cur_tx += frags +1 ; +WRITE_ONCE (tp->cur_tx, tp->cur_tx + frags +1 ); stop_queue = !rtl_tx_slots_avail (tp,MAX_SKB_FRAGS );if (unlikely (stop_queue)) {
@@ -4235,13 +4235,6 @@ static netdev_tx_t rtl8169_start_xmit(struct sk_buff *skb,*/ smp_wmb ();netif_stop_queue (dev); - door_bell = true; - } - -if (door_bell) -rtl8169_doorbell (tp); - -if (unlikely (stop_queue)) {/* Sync with rtl_tx: * - publish queue status and cur_tx ring index (write barrier) * - refresh dirty_tx ring index (read barrier).
@@ -4249,11 +4242,15 @@ static netdev_tx_t rtl8169_start_xmit(struct sk_buff *skb,* status and forget to wake up queue, a racing rtl_tx thread * can't. */ -smp_mb (); +smp_mb__after_atomic ();if (rtl_tx_slots_avail (tp,MAX_SKB_FRAGS ))netif_start_queue (dev); + door_bell = true; } +if (door_bell) +rtl8169_doorbell (tp); +return NETDEV_TX_OK ; err_dma_1:
Signature · 1 line
2.29.2
Quoted 9 lines · On Sat, 14 Nov 2020 21:49:53 +0100 Heiner Kallweit wrote:
Improve the following in
rtl8169_start_xmit:
-tp->cur_txcan be accessed in parallel byrtl_tx(), therefore annotate the race by usingWRITE_ONCE
- avoid checkingstop_queuea second time by moving the doorbell check
-netif_stop_queue()uses atomic operationset_bit()that includes a full memory barrier on some platforms, therefore usesmp_mb__after_atomicto avoid overheadSigned-off-by Heiner Kallweit <[email protected]>
Applied, thanks.