Some DSA tagging protocols change the EtherType field in the MAC header e.g. DSA_TAG_PROTO_(DSA/EDSA/BRCM/MTK/RTL4C_A/SJA1105). On TX these tagged frames are ignored by the checksum offload engine and IP header checker of some stmmac cores.
On RX, the stmmac driver wrongly assumes that checksums have been computed for these tagged packets, and sets CHECKSUM_UNNECESSARY.
Add an additional check in the stmmac TX and RX hotpaths so that COE is deactivated for packets with ethertypes that will not trigger the COE and IP header checks.
Fixes: 6b2c6e4a938f ("net: stmmac: propagate feature flags to vlan") Cc: stable@vger.kernel.org Reported-by: Richard Tresidder rtresidd@electromag.com.au Link: https://lore.kernel.org/netdev/e5c6c75f-2dfa-4e50-a1fb-6bf4cdb617c2@electrom... Reported-by: Romain Gantois romain.gantois@bootlin.com Link: https://lore.kernel.org/netdev/c57283ed-6b9b-b0e6-ee12-5655c1c54495@bootlin.... Signed-off-by: Romain Gantois romain.gantois@bootlin.com --- .../net/ethernet/stmicro/stmmac/stmmac_main.c | 23 ++++++++++++++++--- 1 file changed, 20 insertions(+), 3 deletions(-)
diff --git a/drivers/net/ethernet/stmicro/stmmac/stmmac_main.c b/drivers/net/ethernet/stmicro/stmmac/stmmac_main.c index a9b6b383e863..6797c944a2ac 100644 --- a/drivers/net/ethernet/stmicro/stmmac/stmmac_main.c +++ b/drivers/net/ethernet/stmicro/stmmac/stmmac_main.c @@ -4371,6 +4371,19 @@ static netdev_tx_t stmmac_tso_xmit(struct sk_buff *skb, struct net_device *dev) return NETDEV_TX_OK; }
+/* Check if ethertype will trigger IP + * header checks/COE in hardware + */ +static inline bool stmmac_has_ip_ethertype(struct sk_buff *skb) +{ + int depth = 0; + __be16 proto; + + proto = __vlan_get_protocol(skb, eth_header_parse_protocol(skb), &depth); + + return depth <= ETH_HLEN && (proto == htons(ETH_P_IP) || proto == htons(ETH_P_IPV6)); +} + /** * stmmac_xmit - Tx entry point of the driver * @skb : the socket buffer @@ -4435,9 +4448,13 @@ static netdev_tx_t stmmac_xmit(struct sk_buff *skb, struct net_device *dev) /* DWMAC IPs can be synthesized to support tx coe only for a few tx * queues. In that case, checksum offloading for those queues that don't * support tx coe needs to fallback to software checksum calculation. + * + * Packets that won't trigger the COE e.g. most DSA-tagged packets will + * also have to be checksummed in software. */ if (csum_insertion && - priv->plat->tx_queues_cfg[queue].coe_unsupported) { + (priv->plat->tx_queues_cfg[queue].coe_unsupported || + !stmmac_has_ip_ethertype(skb))) { if (unlikely(skb_checksum_help(skb))) goto dma_map_err; csum_insertion = !csum_insertion; @@ -4997,7 +5014,7 @@ static void stmmac_dispatch_skb_zc(struct stmmac_priv *priv, u32 queue, stmmac_rx_vlan(priv->dev, skb); skb->protocol = eth_type_trans(skb, priv->dev);
- if (unlikely(!coe)) + if (unlikely(!coe) || !stmmac_has_ip_ethertype(skb)) skb_checksum_none_assert(skb); else skb->ip_summed = CHECKSUM_UNNECESSARY; @@ -5513,7 +5530,7 @@ static int stmmac_rx(struct stmmac_priv *priv, int limit, u32 queue) stmmac_rx_vlan(priv->dev, skb); skb->protocol = eth_type_trans(skb, priv->dev);
- if (unlikely(!coe)) + if (unlikely(!coe) || !stmmac_has_ip_ethertype(skb)) skb_checksum_none_assert(skb); else skb->ip_summed = CHECKSUM_UNNECESSARY;
On Mon, Jan 08, 2024 at 12:17:45PM +0100, Romain Gantois wrote:
Some DSA tagging protocols change the EtherType field in the MAC header e.g. DSA_TAG_PROTO_(DSA/EDSA/BRCM/MTK/RTL4C_A/SJA1105). On TX these tagged frames are ignored by the checksum offload engine and IP header checker of some stmmac cores.
On RX, the stmmac driver wrongly assumes that checksums have been computed for these tagged packets, and sets CHECKSUM_UNNECESSARY.
Add an additional check in the stmmac TX and RX hotpaths so that COE is deactivated for packets with ethertypes that will not trigger the COE and IP header checks.
Fixes: 6b2c6e4a938f ("net: stmmac: propagate feature flags to vlan") Cc: stable@vger.kernel.org Reported-by: Richard Tresidder rtresidd@electromag.com.au Link: https://lore.kernel.org/netdev/e5c6c75f-2dfa-4e50-a1fb-6bf4cdb617c2@electrom... Reported-by: Romain Gantois romain.gantois@bootlin.com Link: https://lore.kernel.org/netdev/c57283ed-6b9b-b0e6-ee12-5655c1c54495@bootlin.... Signed-off-by: Romain Gantois romain.gantois@bootlin.com
Reviewed-by: Vladimir Oltean vladimir.oltean@nxp.com
.../net/ethernet/stmicro/stmmac/stmmac_main.c | 23 ++++++++++++++++--- 1 file changed, 20 insertions(+), 3 deletions(-)
diff --git a/drivers/net/ethernet/stmicro/stmmac/stmmac_main.c b/drivers/net/ethernet/stmicro/stmmac/stmmac_main.c index a9b6b383e863..6797c944a2ac 100644 --- a/drivers/net/ethernet/stmicro/stmmac/stmmac_main.c +++ b/drivers/net/ethernet/stmicro/stmmac/stmmac_main.c @@ -4371,6 +4371,19 @@ static netdev_tx_t stmmac_tso_xmit(struct sk_buff *skb, struct net_device *dev) return NETDEV_TX_OK; } +/* Check if ethertype will trigger IP
- header checks/COE in hardware
- */
Nitpick: you could render this in kernel-doc format. https://docs.kernel.org/doc-guide/kernel-doc.html
+static inline bool stmmac_has_ip_ethertype(struct sk_buff *skb)
Nitpick: in netdev it is preferred not to use the "inline" keyword at all in C files, only "static inline" in headers, and to let the compiler decide by itself when it is appropriate to inline the code (which it does by itself even without the "inline" keyword). For a bit more background why, you can view Documentation/process/4.Coding.rst, section "Inline functions".
+{
- int depth = 0;
- __be16 proto;
- proto = __vlan_get_protocol(skb, eth_header_parse_protocol(skb), &depth);
- return depth <= ETH_HLEN && (proto == htons(ETH_P_IP) || proto == htons(ETH_P_IPV6));
+}
Hi Romain,
+/* Check if ethertype will trigger IP
- header checks/COE in hardware
- */
Nitpick: you could render this in kernel-doc format. https://docs.kernel.org/doc-guide/kernel-doc.html
+static inline bool stmmac_has_ip_ethertype(struct sk_buff *skb)
Nitpick: in netdev it is preferred not to use the "inline" keyword at all in C files, only "static inline" in headers, and to let the compiler decide by itself when it is appropriate to inline the code (which it does by itself even without the "inline" keyword). For a bit more background why, you can view Documentation/process/4.Coding.rst, section "Inline functions".
+{
- int depth = 0;
- __be16 proto;
- proto = __vlan_get_protocol(skb, eth_header_parse_protocol(skb), &depth);
- return depth <= ETH_HLEN && (proto == htons(ETH_P_IP) || proto == htons(ETH_P_IPV6));
I also want to nitpick a bit :) If you are to send a v4, maybe you can enclose the first condition within parenthesis to further clarify the return logic.
Cheers, Miquèl
On Mon, 8 Jan 2024, Vladimir Oltean wrote:
...
Nitpick: you could render this in kernel-doc format. https://docs.kernel.org/doc-guide/kernel-doc.html
+static inline bool stmmac_has_ip_ethertype(struct sk_buff *skb)
Nitpick: in netdev it is preferred not to use the "inline" keyword at all in C files, only "static inline" in headers, and to let the compiler decide by itself when it is appropriate to inline the code (which it does by itself even without the "inline" keyword). For a bit more background why, you can view Documentation/process/4.Coding.rst, section "Inline functions".
I see, the kernel docs were indeed enlightening on this point. As a side note, I've just benchmarked both the "with-inline" and "without-inline" versions. First of all, objdump seems to confirm that GCC does indeed follow this pragma in this particular case. Also, RX perfs are better with stmmac_has_ip_ethertype inlined, but TX perfs are actually consistently worse with this function inlined, which could very well be caused by cache effects.
In any case, I think it is better to remove the "inline" pragma as you said. I'll do that in v4.
Best Regards,
On Mon, Jan 08, 2024 at 03:23:38PM +0100, Romain Gantois wrote:
I see, the kernel docs were indeed enlightening on this point. As a side note, I've just benchmarked both the "with-inline" and "without-inline" versions. First of all, objdump seems to confirm that GCC does indeed follow this pragma in this particular case. Also, RX perfs are better with stmmac_has_ip_ethertype inlined, but TX perfs are actually consistently worse with this function inlined, which could very well be caused by cache effects.
In any case, I think it is better to remove the "inline" pragma as you said. I'll do that in v4.
Are you doing any code instrumentation, or just measuring the results and deducing what might cause them?
It might be worth looking at the perf events and seeing what function consumes the most amount of time.
CPU_CORE=0 perf record -e cycles -C $CPU_CORE sleep 10 && perf report perf record -e cache-misses -C $CPU_CORE sleep 10 && perf report
On Mon, 8 Jan 2024, Vladimir Oltean wrote:
On Mon, Jan 08, 2024 at 03:23:38PM +0100, Romain Gantois wrote:
I see, the kernel docs were indeed enlightening on this point. As a side note, I've just benchmarked both the "with-inline" and "without-inline" versions. First of all, objdump seems to confirm that GCC does indeed follow this pragma in this particular case. Also, RX perfs are better with stmmac_has_ip_ethertype inlined, but TX perfs are actually consistently worse with this function inlined, which could very well be caused by cache effects.
In any case, I think it is better to remove the "inline" pragma as you said. I'll do that in v4.
Are you doing any code instrumentation, or just measuring the results and deducing what might cause them?
It might be worth looking at the perf events and seeing what function consumes the most amount of time.
CPU_CORE=0 perf record -e cycles -C $CPU_CORE sleep 10 && perf report perf record -e cache-misses -C $CPU_CORE sleep 10 && perf report
Unfortunately my hardware doesn't support these performance metrics, but I did manage to do some instrumentation with the ftrace profiler:
Same test conditions as before, 10 second iperf3 runs with unfragmented UDP packets.
no inline TX average time per call for stmmac_xmit(): 85us average time per call for stmmac_has_ip_ethertype(): 2us
no inline RX average time per call for stmmac_napi_poll_rx(): 8142us average time per call for stmmac_has_ip_ethertype(): 2us
inline TX: average time per call for stmmac_xmit(): 85us
inline RX: average time per call for stmmac_napi_poll_rx(): 8410us
It seems like this time, RX performed slightly worse with the function inline. To be honest, I'm starting to doubt the reproducibility of these tests. In any case it seems better to just remove the "inline" and let gcc do the optimizing.
Best Regards,
linux-stable-mirror@lists.linaro.org
-
Miquel Raynal -
Romain Gantois -
Vladimir Oltean