[PATCH bpf-next V2] igc: enable and fix RX hash usage by netstack

[PATCH bpf-next V2] igc: enable and fix RX hash usage by netstack

[Jesper Dangaard Brouer]

When function igc_rx_hash() was introduced in v4.20 via commit 0507ef8a0372
("igc: Add transmit and receive fastpath and interrupt handlers"), the
hardware wasn't configured to provide RSS hash, thus it made sense to not
enable net_device NETIF_F_RXHASH feature bit.

The NIC hardware was configured to enable RSS hash info in v5.2 via commit
2121c2712f82 ("igc: Add multiple receive queues control supporting"), but
forgot to set the NETIF_F_RXHASH feature bit.

The original implementation of igc_rx_hash() didn't extract the associated
pkt_hash_type, but statically set PKT_HASH_TYPE_L3. The largest portions of
this patch are about extracting the RSS Type from the hardware and mapping
this to enum pkt_hash_types. This was based on Foxville i225 software user
manual rev-1.3.1 and tested on Intel Ethernet Controller I225-LM (rev 03).

For UDP it's worth noting that RSS (type) hashing have been disabled both for
IPv4 and IPv6 (see IGC_MRQC_RSS_FIELD_IPV4_UDP + IGC_MRQC_RSS_FIELD_IPV6_UDP)
because hardware RSS doesn't handle fragmented pkts well when enabled (can
cause out-of-order). This results in PKT_HASH_TYPE_L3 for UDP packets, and
hash value doesn't include UDP port numbers. Not being PKT_HASH_TYPE_L4, have
the effect that netstack will do a software based hash calc calling into
flow_dissect, but only when code calls skb_get_hash(), which doesn't
necessary happen for local delivery.

Fixes: 2121c2712f82 ("igc: Add multiple receive queues control supporting")
Signed-off-by: Jesper Dangaard Brouer <brouer@redhat.com>
---
 drivers/net/ethernet/intel/igc/igc.h      |   56 +++++++++++++++++++++++++++++
 drivers/net/ethernet/intel/igc/igc_main.c |   35 ++++++++++++++++--
 2 files changed, 87 insertions(+), 4 deletions(-)

diff --git a/drivers/net/ethernet/intel/igc/igc.h b/drivers/net/ethernet/intel/igc/igc.h
index df3e26c0cf01..97f629362d20 100644
--- a/drivers/net/ethernet/intel/igc/igc.h
+++ b/drivers/net/ethernet/intel/igc/igc.h
@@ -311,6 +311,62 @@ extern char igc_driver_name[];
 #define IGC_MRQC_RSS_FIELD_IPV4_UDP	0x00400000
 #define IGC_MRQC_RSS_FIELD_IPV6_UDP	0x00800000
 
+/* RX-desc Write-Back format RSS Type's */
+enum igc_rss_type_num {
+	IGC_RSS_TYPE_NO_HASH		= 0,
+	IGC_RSS_TYPE_HASH_TCP_IPV4	= 1,
+	IGC_RSS_TYPE_HASH_IPV4		= 2,
+	IGC_RSS_TYPE_HASH_TCP_IPV6	= 3,
+	IGC_RSS_TYPE_HASH_IPV6_EX	= 4,
+	IGC_RSS_TYPE_HASH_IPV6		= 5,
+	IGC_RSS_TYPE_HASH_TCP_IPV6_EX	= 6,
+	IGC_RSS_TYPE_HASH_UDP_IPV4	= 7,
+	IGC_RSS_TYPE_HASH_UDP_IPV6	= 8,
+	IGC_RSS_TYPE_HASH_UDP_IPV6_EX	= 9,
+	IGC_RSS_TYPE_MAX		= 10,
+};
+#define IGC_RSS_TYPE_MAX_TABLE		16
+#define IGC_RSS_TYPE_MASK		0xF
+
+/* igc_rss_type - Rx descriptor RSS type field */
+static inline u32 igc_rss_type(const union igc_adv_rx_desc *rx_desc)
+{
+	/* RSS Type 4-bit number: 0-9 (above 9 is reserved)
+	 * Accessing the same bits via u16 (wb.lower.lo_dword.hs_rss.pkt_info)
+	 * is slightly slower than via u32 (wb.lower.lo_dword.data)
+	 */
+	return le32_to_cpu(rx_desc->wb.lower.lo_dword.data) & IGC_RSS_TYPE_MASK;
+}
+
+/* Packet header type identified by hardware (when BIT(11) is zero).
+ * Even when UDP ports are not part of RSS hash HW still parse and mark UDP bits
+ */
+enum igc_pkt_type_bits {
+	IGC_PKT_TYPE_HDR_IPV4	=	BIT(0),
+	IGC_PKT_TYPE_HDR_IPV4_WITH_OPT=	BIT(1), /* IPv4 Hdr includes IP options */
+	IGC_PKT_TYPE_HDR_IPV6	=	BIT(2),
+	IGC_PKT_TYPE_HDR_IPV6_WITH_EXT=	BIT(3), /* IPv6 Hdr includes extensions */
+	IGC_PKT_TYPE_HDR_L4_TCP	=	BIT(4),
+	IGC_PKT_TYPE_HDR_L4_UDP	=	BIT(5),
+	IGC_PKT_TYPE_HDR_L4_SCTP=	BIT(6),
+	IGC_PKT_TYPE_HDR_NFS	=	BIT(7),
+	/* Above only valid when BIT(11) is zero */
+	IGC_PKT_TYPE_L2		=	BIT(11),
+	IGC_PKT_TYPE_VLAN	=	BIT(12),
+	IGC_PKT_TYPE_MASK	=	GENMASK(16,4), /* 13-bits (16:4) */
+	IGC_PKT_TYPE_SHIFT	=	4
+};
+
+/* igc_pkt_type - Rx descriptor Packet type field */
+static inline u32 igc_pkt_type(const union igc_adv_rx_desc *rx_desc)
+{
+	u32 data = le32_to_cpu(rx_desc->wb.lower.lo_dword.data);
+	/* Packet type is 13-bits - as bits (16:4) in lower.lo_dword*/
+	u32 pkt_type = (data & IGC_PKT_TYPE_MASK) >> IGC_PKT_TYPE_SHIFT;
+
+	return pkt_type;
+}
+
 /* Interrupt defines */
 #define IGC_START_ITR			648 /* ~6000 ints/sec */
 #define IGC_4K_ITR			980
diff --git a/drivers/net/ethernet/intel/igc/igc_main.c b/drivers/net/ethernet/intel/igc/igc_main.c
index 8b572cd2c350..27c028f86a9e 100644
--- a/drivers/net/ethernet/intel/igc/igc_main.c
+++ b/drivers/net/ethernet/intel/igc/igc_main.c
@@ -1677,14 +1677,40 @@ static void igc_rx_checksum(struct igc_ring *ring,
 		   le32_to_cpu(rx_desc->wb.upper.status_error));
 }
 
+/* Mapping HW RSS Type to enum pkt_hash_types */
+struct igc_rss_type {
+	u32 hash_type; /* can contain enum pkt_hash_types */
+} igc_rss_type_table[IGC_RSS_TYPE_MAX_TABLE] = {
+	[IGC_RSS_TYPE_NO_HASH].hash_type	  = PKT_HASH_TYPE_L2,
+	[IGC_RSS_TYPE_HASH_TCP_IPV4].hash_type	  = PKT_HASH_TYPE_L4,
+	[IGC_RSS_TYPE_HASH_IPV4].hash_type	  = PKT_HASH_TYPE_L3,
+	[IGC_RSS_TYPE_HASH_TCP_IPV6].hash_type	  = PKT_HASH_TYPE_L4,
+	[IGC_RSS_TYPE_HASH_IPV6_EX].hash_type	  = PKT_HASH_TYPE_L3,
+	[IGC_RSS_TYPE_HASH_IPV6].hash_type	  = PKT_HASH_TYPE_L3,
+	[IGC_RSS_TYPE_HASH_TCP_IPV6_EX].hash_type = PKT_HASH_TYPE_L4,
+	[IGC_RSS_TYPE_HASH_UDP_IPV4].hash_type	  = PKT_HASH_TYPE_L4,
+	[IGC_RSS_TYPE_HASH_UDP_IPV6].hash_type	  = PKT_HASH_TYPE_L4,
+	[IGC_RSS_TYPE_HASH_UDP_IPV6_EX].hash_type = PKT_HASH_TYPE_L4,
+	[10].hash_type = PKT_HASH_TYPE_L2, /* RSS Type above 9 "Reserved" by HW */
+	[11].hash_type = PKT_HASH_TYPE_L2,
+	[12].hash_type = PKT_HASH_TYPE_L2,
+	[13].hash_type = PKT_HASH_TYPE_L2,
+	[14].hash_type = PKT_HASH_TYPE_L2,
+	[15].hash_type = PKT_HASH_TYPE_L2,
+};
+
 static inline void igc_rx_hash(struct igc_ring *ring,
 			       union igc_adv_rx_desc *rx_desc,
 			       struct sk_buff *skb)
 {
-	if (ring->netdev->features & NETIF_F_RXHASH)
-		skb_set_hash(skb,
-			     le32_to_cpu(rx_desc->wb.lower.hi_dword.rss),
-			     PKT_HASH_TYPE_L3);
+	if (ring->netdev->features & NETIF_F_RXHASH) {
+		u32 rss_hash = le32_to_cpu(rx_desc->wb.lower.hi_dword.rss);
+		u32 rss_type = igc_rss_type(rx_desc);
+		enum pkt_hash_types hash_type;
+
+		hash_type = igc_rss_type_table[rss_type].hash_type;
+		skb_set_hash(skb, rss_hash, hash_type);
+	}
 }
 
 static void igc_rx_vlan(struct igc_ring *rx_ring,
@@ -6501,6 +6527,7 @@ static int igc_probe(struct pci_dev *pdev,
 	netdev->features |= NETIF_F_TSO;
 	netdev->features |= NETIF_F_TSO6;
 	netdev->features |= NETIF_F_TSO_ECN;
+	netdev->features |= NETIF_F_RXHASH;
 	netdev->features |= NETIF_F_RXCSUM;
 	netdev->features |= NETIF_F_HW_CSUM;
 	netdev->features |= NETIF_F_SCTP_CRC;

Re: [PATCH bpf-next V2] igc: enable and fix RX hash usage by netstack

[Daniel Borkmann]

On 2/16/23 5:52 PM, Jesper Dangaard Brouer wrote:
> When function igc_rx_hash() was introduced in v4.20 via commit 0507ef8a0372
> ("igc: Add transmit and receive fastpath and interrupt handlers"), the
> hardware wasn't configured to provide RSS hash, thus it made sense to not
> enable net_device NETIF_F_RXHASH feature bit.
> 
> The NIC hardware was configured to enable RSS hash info in v5.2 via commit
> 2121c2712f82 ("igc: Add multiple receive queues control supporting"), but
> forgot to set the NETIF_F_RXHASH feature bit.
> 
> The original implementation of igc_rx_hash() didn't extract the associated
> pkt_hash_type, but statically set PKT_HASH_TYPE_L3. The largest portions of
> this patch are about extracting the RSS Type from the hardware and mapping
> this to enum pkt_hash_types. This was based on Foxville i225 software user
> manual rev-1.3.1 and tested on Intel Ethernet Controller I225-LM (rev 03).
> 
> For UDP it's worth noting that RSS (type) hashing have been disabled both for
> IPv4 and IPv6 (see IGC_MRQC_RSS_FIELD_IPV4_UDP + IGC_MRQC_RSS_FIELD_IPV6_UDP)
> because hardware RSS doesn't handle fragmented pkts well when enabled (can
> cause out-of-order). This results in PKT_HASH_TYPE_L3 for UDP packets, and
> hash value doesn't include UDP port numbers. Not being PKT_HASH_TYPE_L4, have
> the effect that netstack will do a software based hash calc calling into
> flow_dissect, but only when code calls skb_get_hash(), which doesn't
> necessary happen for local delivery.
> 
> Fixes: 2121c2712f82 ("igc: Add multiple receive queues control supporting")
> Signed-off-by: Jesper Dangaard Brouer <brouer@redhat.com>

I presume this should go via net-next, not bpf-next? (Didn't find specific
dependencies, so moved to patchwork netdev bucket..)

Re: [PATCH bpf-next V2] igc: enable and fix RX hash usage by netstack

[Jesper Dangaard Brouer]

On 17/02/2023 21.59, Daniel Borkmann wrote:
> On 2/16/23 5:52 PM, Jesper Dangaard Brouer wrote:
>> When function igc_rx_hash() was introduced in v4.20 via commit 0507ef8a0372
>> ("igc: Add transmit and receive fastpath and interrupt handlers"), the
>> hardware wasn't configured to provide RSS hash, thus it made sense to not
>> enable net_device NETIF_F_RXHASH feature bit.
>>
>> The NIC hardware was configured to enable RSS hash info in v5.2 via commit
>> 2121c2712f82 ("igc: Add multiple receive queues control supporting"), but
>> forgot to set the NETIF_F_RXHASH feature bit.
>>
>> The original implementation of igc_rx_hash() didn't extract the associated
>> pkt_hash_type, but statically set PKT_HASH_TYPE_L3. The largest portions of
>> this patch are about extracting the RSS Type from the hardware and mapping
>> this to enum pkt_hash_types. This was based on Foxville i225 software user
>> manual rev-1.3.1 and tested on Intel Ethernet Controller I225-LM (rev 03).
>>
>> For UDP it's worth noting that RSS (type) hashing have been disabled both for
>> IPv4 and IPv6 (see IGC_MRQC_RSS_FIELD_IPV4_UDP + IGC_MRQC_RSS_FIELD_IPV6_UDP)
>> because hardware RSS doesn't handle fragmented pkts well when enabled (can
>> cause out-of-order). This results in PKT_HASH_TYPE_L3 for UDP packets, and
>> hash value doesn't include UDP port numbers. Not being PKT_HASH_TYPE_L4, have
>> the effect that netstack will do a software based hash calc calling into
>> flow_dissect, but only when code calls skb_get_hash(), which doesn't
>> necessary happen for local delivery.
>>
>> Fixes: 2121c2712f82 ("igc: Add multiple receive queues control supporting")
>> Signed-off-by: Jesper Dangaard Brouer <brouer@redhat.com>
> 
> I presume this should go via net-next, not bpf-next? (Didn't find specific
> dependencies, so moved to patchwork netdev bucket..)

Thanks for letting me/us know.

I posted against bpf-next, because I have patches (for kfunc XDP-hints)
that depend on this patch.  As we are at rc8, netdev maintainers feel
free to simply drop the patch as the patch isn't critical.  I will just
include this patch as part of my kfunc XDP-hints series later.

--Jesper