Thread (15 messages) 15 messages, 4 authors, 2023-09-19

Re: [PATCH net] net: stmmac: Use hrtimer for TX coalescing

From: Maxim Mikityanskiy <hidden>
Date: 2023-08-30 18:05:22

On Wed, 30 Aug 2023 at 14:55:37 +0000, Vincent Whitchurch wrote:
Any test results with this patch on the hardware with the performance
problems would be appreciated.
TL/DR: it's definitely better than without the patch, but still worse
than fully reverting hrtimer [1].

OpenWrt on Netgear R7800, iperf3 test in both directions (LAN->WAN and
WAN->LAN), 3 runs for the duration of 1 minute each. OpenWrt options
packet_steering (RPS + XPS) and flow_offloading (flowtables) enabled,
irqbalance disabled.

Numbers in Mbit/s, I had to fit the table into 72 characters, so:
U is the original kernel 6.1.46 from OpenWrt, R is reverted hrtimer
(patch [1]), P is patched with the new patch from the previous email,
^ is upload (LAN->WAN), v is download (WAN->LAN).

  |  ^  |  v  |  ^  |  v  |  ^  |  v  | avg ^ | avg v | std ^ | std v
- | --- | --- | --- | --- | --- | --- | ----- | ----- | ----- | -----
U | 742 | 709 | 740 | 715 | 556 | 750 |  679  |  725  |  107  |   22
R | 931 | 938 | 935 | 939 | 935 | 939 |  934  |  939  |    2  |    1
P | 845 | 939 | 934 | 939 | 845 | 909 |  875  |  929  |   51  |   17

Full revert allows to get really close to the theoretical maximum
goodput (~949 Mbit/s) with minimal deviation. The new patch, however,
gives less stable numbers, while sometimes hits the maximum as well.

More numbers for the [U]npatched and [R]everted kernels are in the
OpenWrt thread [2].

[1]: https://github.com/openwrt/openwrt/files/12422351/revert-stmmac.patch.txt
[2]: https://github.com/openwrt/openwrt/issues/11676
quoted hunk ↗ jump to hunk
diff --git a/drivers/net/ethernet/stmicro/stmmac/stmmac_main.c b/drivers/net/ethernet/stmicro/stmmac/stmmac_main.c
index 4727f7be4f86..4b6e5061b5a6 100644
--- a/drivers/net/ethernet/stmicro/stmmac/stmmac_main.c
+++ b/drivers/net/ethernet/stmicro/stmmac/stmmac_main.c
@@ -2703,9 +2703,7 @@ static int stmmac_tx_clean(struct stmmac_priv *priv, int budget, u32 queue)
 
 	/* We still have pending packets, let's call for a new scheduling */
 	if (tx_q->dirty_tx != tx_q->cur_tx)
-		hrtimer_start(&tx_q->txtimer,
-			      STMMAC_COAL_TIMER(priv->tx_coal_timer[queue]),
-			      HRTIMER_MODE_REL);
+		stmmac_tx_timer_arm(priv, queue);
 
 	__netif_tx_unlock_bh(netdev_get_tx_queue(priv->dev, queue));
 
@@ -2987,6 +2985,20 @@ static void stmmac_tx_timer_arm(struct stmmac_priv *priv, u32 queue)
 {
 	struct stmmac_tx_queue *tx_q = &priv->dma_conf.tx_queue[queue];
 
+	/*
+	 * Note that the hrtimer could expire immediately after we check this,
+	 * and the hrtimer and the callers of this function do not share a
+	 * lock.
+	 *
+	 * This should however be safe since the only thing the hrtimer does is
+	 * schedule napi (or ask for it run again if it's already running), and
+	 * stmmac_tx_clean(), called from the napi poll function, also calls
+	 * stmmac_tx_timer_arm() at the end if it sees that there are any TX
+	 * packets which have not yet been cleaned.
+	 */
+	if (hrtimer_is_queued(&tx_q->txtimer))
+		return;
+
 	hrtimer_start(&tx_q->txtimer,
 		      STMMAC_COAL_TIMER(priv->tx_coal_timer[queue]),
 		      HRTIMER_MODE_REL);
Keyboard shortcuts
hback out one level
jnext message in thread
kprevious message in thread
ldrill in
Escclose help / fold thread tree
?toggle this help