[PATCH net-next v4 6/6] net: mvneta: serve the TX queues in weighted round robin
From: Josef Schlehofer
Date: Wed Nov 18 2026 - 04:31:00 EST
Under load, some TX queues get no service at all. With four TCP
streams on a 100 Mbit/s link, up to three streams sent nothing for 20
seconds. The TX queue arbiter is left at its reset default, which
serves all queues in fixed priority, and the stack spreads the flows
over the queues.
Give the TX queues the same WRR weight, the smallest one that is
larger than the largest packet, and clear the fixed priority. The
Armada 370 and XP manuals list separate registers for the version 1
arbiter, which the driver runs unless a shaper is set, so write both
register sets. On the Armada 385 they are the same registers, as
writing one set and reading the other shows.
Keep the fixed priority while an mqprio offload is set up, as its
traffic classes may rely on it, and clear it when the offload is
removed. A rejected setup leaves it as it was, and after a rejected
replace it applies to all flows until the qdisc is deleted. A class
that fills the link starves the lower classes, and as the fixed
priority is set per queue, the queues of one class still starve each
other.
Observed on machines in real use. OpenWrt users reported it as a
multi-flow lockout on a Linksys WRT1200AC, and OpenWrt has worked
around it since 2017 by sending all TX traffic of its 32-bit boards
through queue 0. Reproduced on a Turris Omnia (Armada 385). With this
patch each busy TX queue gets the same share and no stream stalls,
4 x 23.5 Mbit/s on a 100 Mbit/s link and 4 x 232 Mbit/s on a 1 Gbit/s
link. With an mqprio offload a higher queue still has priority over a
lower one. Only the Armada 385 was tested. No kernel selftest covers
the TX queue arbiter.
Fixes: ee40a116ebf1 ("net: mvneta: fix improper tx queue usage in mvneta_tx()")
Link: https://github.com/openwrt/openwrt/issues/5411
Suggested-by: Maxime Chevallier <maxime.chevallier@xxxxxxxxxxx>
Signed-off-by: Josef Schlehofer <pepe.schlehofer@xxxxxxxxx>
---
v4: The fixed priority and the weights are also written to the version
1 registers. A setup that the driver rejects no longer clears the fixed
priority. Discovery statement added.
---
drivers/net/ethernet/marvell/mvneta.c | 56 ++++++++++++++++++++++++++-
1 file changed, 55 insertions(+), 1 deletion(-)
diff --git a/drivers/net/ethernet/marvell/mvneta.c b/drivers/net/ethernet/marvell/mvneta.c
index 89ade2ed7ce4..35409695d3d7 100644
--- a/drivers/net/ethernet/marvell/mvneta.c
+++ b/drivers/net/ethernet/marvell/mvneta.c
@@ -118,6 +118,7 @@
#define MVNETA_HSGMII_SERDES_PROTO 0x1107
#define MVNETA_TYPE_PRIO 0x24bc
#define MVNETA_FORCE_UNI BIT(21)
+#define MVETH_TXQ_FIXED_PRIO_REG 0x24dc
#define MVNETA_TXQ_CMD_1 0x24e4
#define MVNETA_TXQ_CMD 0x2448
#define MVNETA_TXQ_DISABLE_SHIFT 8
@@ -196,6 +197,7 @@
#define MVNETA_RXQ_ENABLE_MASK 0x000000ff
#define MVETH_TXQ_TOKEN_COUNT_REG(q) (0x2700 + ((q) << 4))
#define MVETH_TXQ_TOKEN_CFG_REG(q) (0x2704 + ((q) << 4))
+#define MVETH_TXQ_WRR_ARB_REG(q) (0x2708 + ((q) << 4))
#define MVNETA_GMAC_CTRL_0 0x2c00
#define MVNETA_GMAC_MAX_RX_SIZE_SHIFT 2
#define MVNETA_GMAC_MAX_RX_SIZE_MASK 0x7ffc
@@ -253,6 +255,8 @@
#define MVNETA_TXQ_CMD1_REG 0x3e00
#define MVNETA_TXQ_CMD1_BW_LIM_SEL_V1 BIT(3)
#define MVNETA_TXQ_CMD1_BW_LIM_EN BIT(0)
+#define MVNETA_TXQ_FIXED_PRIO_REG 0x3e04
+#define MVNETA_TXQ_FIXED_PRIO_ALL 0xff
#define MVNETA_REFILL_NUM_CLK_REG 0x3e08
#define MVNETA_REFILL_MAX_NUM_CLK 0x0000ffff
#define MVNETA_TX_MTU 0x3e0c
@@ -264,6 +268,8 @@
#define MVNETA_TXQ_BUCKET_REFILL_VALUE_MAX 0x0007ffff
#define MVNETA_TXQ_TOKEN_SIZE_REG(q) (0x3e40 + ((q) << 2))
#define MVNETA_TXQ_TOKEN_SIZE_MAX 0x7fffffff
+#define MVNETA_TXQ_WRR_ARB_REG(q) (0x3e80 + ((q) << 2))
+#define MVNETA_TXQ_WRR_WEIGHT_MASK 0xff
/* The values of the bucket refill base period and refill period are taken from
* the reference manual, and adds up to a base resolution of 10Kbps. This allows
@@ -518,6 +524,8 @@ struct mvneta_port {
*/
spinlock_t lock;
bool is_stopped;
+ /* An mqprio offload keeps the TX queues in fixed priority */
+ bool txq_fixed_prio;
u32 cause_rx_tx;
struct napi_struct napi;
@@ -1632,6 +1640,39 @@ static void mvneta_txq_max_tx_size_set(struct mvneta_port *pp, int max_tx_size)
}
}
+static void mvneta_txq_wrr_weight(struct mvneta_port *pp, u32 reg, u32 weight)
+{
+ u32 val = mvreg_read(pp, reg);
+
+ val &= ~MVNETA_TXQ_WRR_WEIGHT_MASK;
+ mvreg_write(pp, reg, val | weight);
+}
+
+/* With the default fixed priority, a busy queue starves the others.
+ * The Armada 370 and XP manuals list separate registers for the
+ * version 1 arbiter, which runs without a shaper, so set both.
+ */
+static void mvneta_txq_wrr_set(struct mvneta_port *pp, int max_tx_size)
+{
+ /* In units of 256 bytes, must exceed the largest packet */
+ u32 weight = max_tx_size / 256 + 1;
+ int queue;
+
+ for (queue = 0; queue < txq_number; queue++) {
+ mvneta_txq_wrr_weight(pp, MVNETA_TXQ_WRR_ARB_REG(queue), weight);
+ mvneta_txq_wrr_weight(pp, MVETH_TXQ_WRR_ARB_REG(queue), weight);
+ }
+}
+
+/* Only the traffic classes of an mqprio offload keep the fixed priority */
+static void mvneta_txq_fixed_prio_set(struct mvneta_port *pp)
+{
+ u32 val = pp->txq_fixed_prio ? MVNETA_TXQ_FIXED_PRIO_ALL : 0;
+
+ mvreg_write(pp, MVNETA_TXQ_FIXED_PRIO_REG, val);
+ mvreg_write(pp, MVETH_TXQ_FIXED_PRIO_REG, val);
+}
+
/* Set unicast address */
static void mvneta_set_ucast_addr(struct mvneta_port *pp, u8 last_nibble,
int queue)
@@ -3765,6 +3806,8 @@ static void mvneta_start_dev(struct mvneta_port *pp)
mvneta_max_rx_size_set(pp, pp->pkt_size);
mvneta_txq_max_tx_size_set(pp, pp->pkt_size);
+ mvneta_txq_wrr_set(pp, pp->pkt_size);
+ mvneta_txq_fixed_prio_set(pp);
mvneta_restore_per_queue_rate_limit(pp);
/* start the Rx/Tx activity */
@@ -5345,9 +5388,20 @@ static int mvneta_setup_mqprio(struct net_device *dev,
static int mvneta_setup_tc(struct net_device *dev, enum tc_setup_type type,
void *type_data)
{
+ struct mvneta_port *pp = netdev_priv(dev);
+ struct tc_mqprio_qopt_offload *mqprio;
+ int ret;
+
switch (type) {
case TC_SETUP_QDISC_MQPRIO:
- return mvneta_setup_mqprio(dev, type_data);
+ mqprio = type_data;
+ ret = mvneta_setup_mqprio(dev, mqprio);
+ if (ret)
+ return ret;
+
+ pp->txq_fixed_prio = mqprio->qopt.num_tc;
+ mvneta_txq_fixed_prio_set(pp);
+ return 0;
default:
return -EOPNOTSUPP;
}
--
2.54.0 (Apple Git-157)