Skip to contentSkip to conversation

Rework ice Tx checksum

PATCH In review Michal Swiatkowski 6 participants 14 messages Apr 28 – Oct 30

Subsystems (3)
lore
Share & tools
lore
#0 lore raw

cover Rework ice Tx checksum

Hi,

Jakub find the problem with how ice handles Tx checksum offload [1].

The issue is that IP_CSUM/IPV6_CSUM should support only basic packet types. Kernel assumes that if there are extensions headers and there is no HW_CSUM software fallback should be applied. Here is a patch that introduced that [2].

However, software fallback can't be applied when TSO offload happens. That is the situation that this patchset is fixing.

When TSO is enabled and IP_CSUM/IPV6_CSUM is on, kernel is assuming hardware will be able to calculate the checksum, but later on the extension headers are checked and if there is no HW_CSUM the skb_bad_offload() is called.

Following the documentation ice driver is capable of HW_CSUM support, so switch from IP_CSUM/IPV6_CSUM to HW_CSUM. Other Intel drivers (even previous like i40e, ixgbe) also supports HW_CSUM.

The HW_CSUM was used to indicate that GCS (Intel name for checksumming offload using additional descriptor) is support, which is variation of HW_CSUM. Even without GCS th HW_CSUM is supported. First patch is switching to use custom Tx ring flags to allow HW_CSUM usage.

[1] https://lore.kernel.org/netdev/[email protected]/
[2] https://lore.kernel.org/all/5fbeecfc311ea182aa1d1c771725ab8b4cac515e.1729778144.git.benoit.monin@gmx.fr/

Michal Swiatkowski (2):
ice: always do GCS if hardware supports it
ice: use NETIF_F_HW_CSUM instead of IP/IPV6

 drivers/net/ethernet/intel/ice/ice_lib.c  |  4 ++++
 drivers/net/ethernet/intel/ice/ice_main.c | 21 +--------------------
 drivers/net/ethernet/intel/ice/ice_txrx.c |  2 +-
 drivers/net/ethernet/intel/ice/ice_txrx.h |  1 +
 4 files changed, 7 insertions(+), 21 deletions(-)
Signature · 1 line
2.49.0
#1 lore raw

PATCH 1/2 ice: always do GCS if hardware supports it

There is no need to check for NETIF_HW_CSUM. If the code reach calculating checksum it means that correct checksum flags are set, because kernel is checking that when setting ip->summed.

Instead of netdev feature flag use Tx ring flag to check if the hardware can use special descriptor for checksum calculating.

---
 drivers/net/ethernet/intel/ice/ice_lib.c  | 4 ++++
 drivers/net/ethernet/intel/ice/ice_txrx.c | 2 +-
 drivers/net/ethernet/intel/ice/ice_txrx.h | 1 +
 3 files changed, 6 insertions(+), 1 deletion(-)
drivers/net/ethernet/intel/ice/ice_lib.c +4 -0
index 837b71b7b2b7..033fabc22f58 100644
@@ -1415,6 +1415,10 @@ static int ice_vsi_alloc_rings(struct ice_vsi *vsi)
 			set_bit(ICE_TX_RING_FLAGS_VLAN_L2TAG2, ring->flags);
 		else
 			set_bit(ICE_TX_RING_FLAGS_VLAN_L2TAG1, ring->flags);
+
+		if (ice_is_feature_supported(pf, ICE_F_GCS))
+			set_bit(ICE_TX_RING_FLAGS_GCS, ring->flags);
+
 		WRITE_ONCE(vsi->tx_rings[i], ring);
 	}
 
drivers/net/ethernet/intel/ice/ice_txrx.c +1 -1
index 4ca1a0602307..25b19a5d817e 100644
@@ -1743,7 +1743,7 @@ int ice_tx_csum(struct ice_tx_buf *first, struct ice_tx_offload_params *off)
 	l3_len = l4.hdr - ip.hdr;
 	offset |= (l3_len / 4) << ICE_TX_DESC_LEN_IPLEN_S;
 
-	if ((tx_ring->netdev->features & NETIF_F_HW_CSUM) &&
+	if (test_bit(ICE_TX_RING_FLAGS_GCS, tx_ring->flags) &&
 	    !(first->tx_flags & ICE_TX_FLAGS_TSO) &&
 	    !skb_csum_is_sctp(skb)) {
 		/* Set GCS */
drivers/net/ethernet/intel/ice/ice_txrx.h +1 -0
index 5e517f219379..15dbd5100912 100644
@@ -217,6 +217,7 @@ enum ice_tx_ring_flags {
 	ICE_TX_RING_FLAGS_VLAN_L2TAG1,
 	ICE_TX_RING_FLAGS_VLAN_L2TAG2,
 	ICE_TX_RING_FLAGS_TXTIME,
+	ICE_TX_RING_FLAGS_GCS,
 	ICE_TX_RING_FLAGS_NBITS,
 };
 
Signature · 1 line
2.49.0
Tested-by on v1 1/2
#2 lore raw

RE: [Intel-wired-lan] [PATCH iwl-net v1 1/2] ice: always do GCS if hardware supports it

Quoted 15 lines

-----Original Message-----
From: Intel-wired-lan <[email protected]> On Behalf Of Michal Swiatkowski
Sent: Tuesday, April 28, 2026 12:07 AM
To: [email protected]
Cc: [email protected]; [email protected]; Nguyen, Anthony L <[email protected]>; Kitszel, Przemyslaw <[email protected]>; Loktionov, Aleksandr <[email protected]>; Michal Swiatkowski <[email protected]>
Subject: [Intel-wired-lan] [PATCH iwl-net v1 1/2] ice: always do GCS if hardware supports it

There is no need to check for NETIF_HW_CSUM. If the code reach calculating checksum it means that correct checksum flags are set, because kernel is checking that when setting ip->summed.

Instead of netdev feature flag use Tx ring flag to check if the hardware can use special descriptor for checksum calculating.

Quoted 7 lines
---
 drivers/net/ethernet/intel/ice/ice_lib.c  | 4 ++++  drivers/net/ethernet/intel/ice/ice_txrx.c | 2 +-  
drivers/net/ethernet/intel/ice/ice_txrx.h | 1 +
 3 files changed, 6 insertions(+), 1 deletion(-)
#3 lore raw

PATCH 2/2 ice: use NETIF_F_HW_CSUM instead of IP/IPV6

In reply to Michal Swiatkowski

The hardware is capable of calculating checksum for IPV6 packets with extension header. To not drop such packets switch from IP/IPV6 checksum to HW_CSUM.

HW_CSUM is also used in previous generation (i40e).

Previously HW_CSUM was used to indicate that hardware supports general checksum. Drop it assuming that if the hardware supports it, it is used.

Disabling offload for E830 in case of TSO isn't needed anymore as the check for TSO is done in Tx path just before preparation of the special GCS descriptor.

The commit from Fixes didn't introduce a bug, it just shown that the driver is doing sth wrong with the checksum features.

Fixes 04c20a9356f2 ("net: skip offload for NETIF_F_IPV6_CSUM if ipv6 header contains extension")
---
 drivers/net/ethernet/intel/ice/ice_main.c | 21 +--------------------
 1 file changed, 1 insertion(+), 20 deletions(-)
drivers/net/ethernet/intel/ice/ice_main.c +1 -20
index 15550216fbf0..0f2f949af536 100644
@@ -3500,9 +3500,8 @@ void ice_set_netdev_features(struct net_device *netdev)
 			NETIF_F_RXHASH;
 
 	csumo_features = NETIF_F_RXCSUM	  |
-			 NETIF_F_IP_CSUM  |
 			 NETIF_F_SCTP_CRC |
-			 NETIF_F_IPV6_CSUM;
+			 NETIF_F_HW_CSUM;
 
 	vlano_features = NETIF_F_HW_VLAN_CTAG_FILTER |
 			 NETIF_F_HW_VLAN_CTAG_TX     |
@@ -3564,12 +3563,6 @@ void ice_set_netdev_features(struct net_device *netdev)
 	/* Allow core to manage IRQs affinity */
 	netif_set_affinity_auto(netdev);
 
-	/* Mutual exclusivity for TSO and GCS is enforced by the set features
-	 * ndo callback.
-	 */
-	if (ice_is_feature_supported(pf, ICE_F_GCS))
-		netdev->hw_features |= NETIF_F_HW_CSUM;
-
 	netif_set_tso_max_size(netdev, ICE_MAX_TSO_SIZE);
 }
 
@@ -6489,18 +6482,6 @@ ice_set_features(struct net_device *netdev, netdev_features_t features)
 	if (changed & NETIF_F_LOOPBACK)
 		ret = ice_set_loopback(vsi, !!(features & NETIF_F_LOOPBACK));
 
-	/* Due to E830 hardware limitations, TSO (NETIF_F_ALL_TSO) with GCS
-	 * (NETIF_F_HW_CSUM) is not supported.
-	 */
-	if (ice_is_feature_supported(pf, ICE_F_GCS) &&
-	    ((features & NETIF_F_HW_CSUM) && (features & NETIF_F_ALL_TSO))) {
-		if (netdev->features & NETIF_F_HW_CSUM)
-			dev_err(ice_pf_to_dev(pf), "To enable TSO, you must first disable HW checksum.\n");
-		else
-			dev_err(ice_pf_to_dev(pf), "To enable HW checksum, you must first disable TSO.\n");
-		return -EIO;
-	}
-
 	return ret;
 }
 
Signature · 1 line
2.49.0
Reviewed-by on v1 2/2
#4 lore raw

RE: [PATCH iwl-net v1 2/2] ice: use NETIF_F_HW_CSUM instead of IP/IPV6

Quoted 14 lines

-----Original Message-----
From: Michal Swiatkowski <[email protected]>
Sent: Tuesday, April 28, 2026 9:07 AM
To: [email protected]
Cc: [email protected]; [email protected]; Nguyen, Anthony L <[email protected]>; Kitszel, Przemyslaw <[email protected]>; Loktionov, Aleksandr <[email protected]>; Michal Swiatkowski <[email protected]>
Subject: [PATCH iwl-net v1 2/2] ice: use NETIF_F_HW_CSUM instead of IP/IPV6

Quoted 3 lines

The hardware is capable of calculating checksum for IPV6 packets with extension header. To not drop such packets switch from IP/IPV6 checksum to HW_CSUM.

I'd recommend "To not drop" -> "To avoid dropping"

Quoted 7 lines

HW_CSUM is also used in previous generation (i40e).

Previously HW_CSUM was used to indicate that hardware supports general checksum. Drop it assuming that if the hardware supports it, it is used.

Quoted 7 lines

Disabling offload for E830 in case of TSO isn't needed anymore as the check for TSO is done in Tx path just before preparation of the special GCS descriptor.

The commit from Fixes didn't introduce a bug, it just shown that the driver is doing sth wrong with the checksum features.

Except commit message nits

Quoted 67 lines
Fixes 04c20a9356f2 ("net: skip offload for NETIF_F_IPV6_CSUM if ipv6

header contains extension")

---
 drivers/net/ethernet/intel/ice/ice_main.c | 21 +--------------------
 1 file changed, 1 insertion(+), 20 deletions(-)
drivers/net/ethernet/intel/ice/ice_main.c +1 -2
index 15550216fbf0..0f2f949af536 100644
@@ -3500,9 +3500,8 @@ void ice_set_netdev_features(struct net_device
 *netdev)
 			NETIF_F_RXHASH;
 
 	csumo_features = NETIF_F_RXCSUM	  |
-			 NETIF_F_IP_CSUM  |
 			 NETIF_F_SCTP_CRC |
-			 NETIF_F_IPV6_CSUM;
+			 NETIF_F_HW_CSUM;
 
 	vlano_features = NETIF_F_HW_VLAN_CTAG_FILTER |
drivers/net/ethernet/intel/ice/ice_main.c +0 -1
 			 NETIF_F_HW_VLAN_CTAG_TX     |
@@ -3564,12 +3563,6 @@ void ice_set_netdev_features(struct net_device
 *netdev)
 	/* Allow core to manage IRQs affinity */
 	netif_set_affinity_auto(netdev);
 
-	/* Mutual exclusivity for TSO and GCS is enforced by the set

features

drivers/net/ethernet/intel/ice/ice_main.c +0 -6
-	 * ndo callback.
-	 */
-	if (ice_is_feature_supported(pf, ICE_F_GCS))
-		netdev->hw_features |= NETIF_F_HW_CSUM;
-
 	netif_set_tso_max_size(netdev, ICE_MAX_TSO_SIZE);  }
@@ -6489,18 +6482,6 @@ ice_set_features(struct net_device *netdev,
 netdev_features_t features)
 	if (changed & NETIF_F_LOOPBACK)
 		ret = ice_set_loopback(vsi, !!(features &
 NETIF_F_LOOPBACK));
 
-	/* Due to E830 hardware limitations, TSO (NETIF_F_ALL_TSO) with

GCS

drivers/net/ethernet/intel/ice/ice_main.c +0 -4
-	 * (NETIF_F_HW_CSUM) is not supported.
-	 */
-	if (ice_is_feature_supported(pf, ICE_F_GCS) &&
-	    ((features & NETIF_F_HW_CSUM) && (features &

NETIF_F_ALL_TSO))) {

drivers/net/ethernet/intel/ice/ice_main.c +0 -2
-		if (netdev->features & NETIF_F_HW_CSUM)
-			dev_err(ice_pf_to_dev(pf), "To enable TSO, you

must first disable HW checksum.\n");

drivers/net/ethernet/intel/ice/ice_main.c +0 -2
-		else
-			dev_err(ice_pf_to_dev(pf), "To enable HW

checksum, you must first disable TSO.\n");

drivers/net/ethernet/intel/ice/ice_main.c +0 -4
-		return -EIO;
-	}
-
 	return ret;
 }
 
--

2.49.0

reviewer
#5 lore raw

Re: [PATCH iwl-net v1 2/2] ice: use NETIF_F_HW_CSUM instead of IP/IPV6

In reply to Michal Swiatkowski

From: 'Simon Horman' <[email protected]>

This is an AI-generated review of your patch. The human sending this email has considered the AI review valid, or at least plausible. Full review at: https://sashiko.dev
---
ice: use NETIF_F_HW_CSUM instead of IP/IPV6

This commit replaces the IP and IPV6 checksum offload flags with HW_CSUM in the ice driver to allow hardware checksumming for IPv6 packets with extension headers. It also removes the mutual exclusivity check between TSO and GCS.

drivers/net/ethernet/intel/ice/ice_main.c void ice_set_netdev_features(struct net_device *netdev) · line 3495

Quoted 12 lines
drivers/net/ethernet/intel/ice/ice_main.c +1 -2
index 5f92377d4dfc..33f252a6a22f 100644
@@ -3491,9 +3491,8 @@ void ice_set_netdev_features(struct net_device *netdev)
 			NETIF_F_RXHASH;
 
 	csumo_features = NETIF_F_RXCSUM	  |
-			 NETIF_F_IP_CSUM  |
 			 NETIF_F_SCTP_CRC |
-			 NETIF_F_IPV6_CSUM;
+			 NETIF_F_HW_CSUM;

By advertising NETIF_F_HW_CSUM, the network stack assumes the driver can handle checksums for any protocol and passes non-IP packets (e.g., from AF_PACKET injected frames) to the driver with CHECKSUM_PARTIAL.

Since ice_tx_csum() strictly parses the L3 header and expects IPv4 or IPv6, returning -1 for other protocols, will this regression cause non-IP packets to be silently dropped instead of falling back to software checksums in the kernel?

Additionally, skb_csum_hwoffload_help() relies on NETIF_F_IPV6_CSUM to safely fall back to software checksums when IPv6 extension headers are present.

Does using NETIF_F_HW_CSUM bypass this safeguard, potentially causing legacy hardware (like E810) or E830 without GCS descriptors to compute incorrect pseudo-headers and silently corrupt IPv6 packets with extension headers?

Finally, for protocols other than TCP, UDP, or SCTP, the driver falls back to calling skb_checksum_help(skb) in the default case of ice_tx_csum().

However, the driver appears to ignore the return value of skb_checksum_help().

If skb_checksum_help() fails (e.g., due to -ENOMEM during pskb_expand_head()), could this result in the packet being transmitted with an uncomputed checksum?

reviewer
#6 lore raw

Re: [PATCH iwl-net v1 2/2] ice: use NETIF_F_HW_CSUM instead of IP/IPV6

Quoted 5 lines · Simon Horman

From: 'Simon Horman' <[email protected]>

This is an AI-generated review of your patch. The human sending this email has considered the AI review valid, or at least plausible. Full review at: https://sashiko.dev

Sorry, the email I am responding to was supposed to come with some commentary - it seems my method for adding it is broken :( So I'm adding it below (as best as I can remember what it was).

Quoted 25 lines

---
ice: use NETIF_F_HW_CSUM instead of IP/IPV6

This commit replaces the IP and IPV6 checksum offload flags with HW_CSUM in the ice driver to allow hardware checksumming for IPv6 packets with extension headers. It also removes the mutual exclusivity check between TSO and GCS.

drivers/net/ethernet/intel/ice/ice_main.c void ice_set_netdev_features(struct net_device *netdev) · line 3495

Quoted 12 lines
drivers/net/ethernet/intel/ice/ice_main.c +1 -2
index 5f92377d4dfc..33f252a6a22f 100644
@@ -3491,9 +3491,8 @@ void ice_set_netdev_features(struct net_device *netdev)
 			NETIF_F_RXHASH;
 
 	csumo_features = NETIF_F_RXCSUM	  |
-			 NETIF_F_IP_CSUM  |
 			 NETIF_F_SCTP_CRC |
-			 NETIF_F_IPV6_CSUM;
+			 NETIF_F_HW_CSUM;

By advertising NETIF_F_HW_CSUM, the network stack assumes the driver can handle checksums for any protocol and passes non-IP packets (e.g., from AF_PACKET injected frames) to the driver with CHECKSUM_PARTIAL.

Quoted 3 lines

Since ice_tx_csum() strictly parses the L3 header and expects IPv4 or IPv6, returning -1 for other protocols, will this regression cause non-IP packets to be silently dropped instead of falling back to software checksums in the kernel?

The problem above seems to be a legitimate concern, if a corner case. It seems to also exist in i40e.

Quoted 8 lines

Additionally, skb_csum_hwoffload_help() relies on NETIF_F_IPV6_CSUM to safely fall back to software checksums when IPv6 extension headers are present.

Does using NETIF_F_HW_CSUM bypass this safeguard, potentially causing legacy hardware (like E810) or E830 without GCS descriptors to compute incorrect pseudo-headers and silently corrupt IPv6 packets with extension headers?

The issue above seems to be a false positive as the ice driver doesn't rely on the behaviour described. Rather ice_tx_csum() uses ipv6_skip_exthdr() to correctly handle extension headers.

Quoted 8 lines

Finally, for protocols other than TCP, UDP, or SCTP, the driver falls back to calling skb_checksum_help(skb) in the default case of ice_tx_csum().

However, the driver appears to ignore the return value of skb_checksum_help().

If skb_checksum_help() fails (e.g., due to -ENOMEM during pskb_expand_head()), could this result in the packet being transmitted with an uncomputed checksum?

The problem above seems real to me, but it also seems to be pre-existing. So I don't think it should delay progress of this patch.

If it is a problem, it also seems to be present in i40e.

#7 lore raw

Re: [PATCH iwl-net v1 2/2] ice: use NETIF_F_HW_CSUM instead of IP/IPV6

Quoted 11 lines · Simon Horman
Quoted 5 lines · Simon Horman

From: 'Simon Horman' <[email protected]>

This is an AI-generated review of your patch. The human sending this email has considered the AI review valid, or at least plausible. Full review at: https://sashiko.dev

Sorry, the email I am responding to was supposed to come with some commentary - it seems my method for adding it is broken :( So I'm adding it below (as best as I can remember what it was).

Thanks for adding the comments.

Quoted 25 lines
Quoted 25 lines

---
ice: use NETIF_F_HW_CSUM instead of IP/IPV6

This commit replaces the IP and IPV6 checksum offload flags with HW_CSUM in the ice driver to allow hardware checksumming for IPv6 packets with extension headers. It also removes the mutual exclusivity check between TSO and GCS.

drivers/net/ethernet/intel/ice/ice_main.c void ice_set_netdev_features(struct net_device *netdev) · line 3495

Quoted 12 lines
drivers/net/ethernet/intel/ice/ice_main.c +1 -2
index 5f92377d4dfc..33f252a6a22f 100644
@@ -3491,9 +3491,8 @@ void ice_set_netdev_features(struct net_device *netdev)
 			NETIF_F_RXHASH;
 
 	csumo_features = NETIF_F_RXCSUM	  |
-			 NETIF_F_IP_CSUM  |
 			 NETIF_F_SCTP_CRC |
-			 NETIF_F_IPV6_CSUM;
+			 NETIF_F_HW_CSUM;

By advertising NETIF_F_HW_CSUM, the network stack assumes the driver can handle checksums for any protocol and passes non-IP packets (e.g., from AF_PACKET injected frames) to the driver with CHECKSUM_PARTIAL.

Quoted 7 lines
Quoted 3 lines

Since ice_tx_csum() strictly parses the L3 header and expects IPv4 or IPv6, returning -1 for other protocols, will this regression cause non-IP packets to be silently dropped instead of falling back to software checksums in the kernel?

The problem above seems to be a legitimate concern, if a corner case. It seems to also exist in i40e.

I think there is a lack of skb_checksum_help() before one error return from ice_tx_csum(). I will add it.

Quoted 4 lines
Quoted 4 lines

Additionally, skb_csum_hwoffload_help() relies on NETIF_F_IPV6_CSUM to safely fall back to software checksums when IPv6 extension headers are present.

Quoted 9 lines
Quoted 4 lines

Does using NETIF_F_HW_CSUM bypass this safeguard, potentially causing legacy hardware (like E810) or E830 without GCS descriptors to compute incorrect pseudo-headers and silently corrupt IPv6 packets with extension headers?

The issue above seems to be a false positive as the ice driver doesn't rely on the behaviour described. Rather ice_tx_csum() uses ipv6_skip_exthdr() to correctly handle extension headers.

Yeah, righ.

Quoted 6 lines
Quoted 6 lines

Finally, for protocols other than TCP, UDP, or SCTP, the driver falls back to calling skb_checksum_help(skb) in the default case of ice_tx_csum().

However, the driver appears to ignore the return value of skb_checksum_help().

Quoted 7 lines
Quoted 2 lines

If skb_checksum_help() fails (e.g., due to -ENOMEM during pskb_expand_head()), could this result in the packet being transmitted with an uncomputed checksum?

The problem above seems real to me, but it also seems to be pre-existing. So I don't think it should delay progress of this patch.

If it is a problem, it also seems to be present in i40e.

idpf, igbe too. It should return error in such case and drop the packet. What do you think, should I add the patch for that here or send it separately? To be honest I prefer sending it separately as this is in already exsisting code and touch more than one driver.

Thanks

#8 lore raw

Re: [Intel-wired-lan] [PATCH iwl-net v1 2/2] ice: use NETIF_F_HW_CSUM instead of IP/IPV6

In reply to Michal Swiatkowski
Quoted 13 lines · Michal Swiatkowski

The hardware is capable of calculating checksum for IPV6 packets with extension header. To not drop such packets switch from IP/IPV6 checksum to HW_CSUM.

HW_CSUM is also used in previous generation (i40e).

Previously HW_CSUM was used to indicate that hardware supports general checksum. Drop it assuming that if the hardware supports it, it is used.

Disabling offload for E830 in case of TSO isn't needed anymore as the check for TSO is done in Tx path just before preparation of the special GCS descriptor.

Quoted 8 lines

The commit from Fixes didn't introduce a bug, it just shown that the driver is doing sth wrong with the checksum features.

Fixes 04c20a9356f2 ("net: skip offload for NETIF_F_IPV6_CSUM if ipv6 header contains extension")

---

Am I correct in thinking that this supersedes (really, properly fixes) the patch "ice: enable NETIF_F_HW_CSUM for GSO packets" at
https://patchwork.ozlabs.org/project/intel-wired-lan/patch/[email protected]/
?

Thanks,
Jake

#9 lore raw

Re: [Intel-wired-lan] [PATCH iwl-net v1 2/2] ice: use NETIF_F_HW_CSUM instead of IP/IPV6

Quoted 22 lines · Jacob Keller
Quoted 13 lines · Michal Swiatkowski

The hardware is capable of calculating checksum for IPV6 packets with extension header. To not drop such packets switch from IP/IPV6 checksum to HW_CSUM.

HW_CSUM is also used in previous generation (i40e).

Previously HW_CSUM was used to indicate that hardware supports general checksum. Drop it assuming that if the hardware supports it, it is used.

Disabling offload for E830 in case of TSO isn't needed anymore as the check for TSO is done in Tx path just before preparation of the special GCS descriptor.

Quoted 8 lines

The commit from Fixes didn't introduce a bug, it just shown that the driver is doing sth wrong with the checksum features.

Fixes 04c20a9356f2 ("net: skip offload for NETIF_F_IPV6_CSUM if ipv6 header contains extension")

---

Quoted 7 lines

Am I correct in thinking that this supersedes (really, properly fixes) the patch "ice: enable NETIF_F_HW_CSUM for GSO packets" at
https://patchwork.ozlabs.org/project/intel-wired-lan/patch/[email protected]/
?

Thanks,
Jake

Yes, exactly. I think I linked it in cover letter, but maybe I should do it also here.

Thanks

#10 lore raw

Re: [Intel-wired-lan] [PATCH iwl-net v1 2/2] ice: use NETIF_F_HW_CSUM instead of IP/IPV6

Quoted 28 lines · Michal Swiatkowski
Quoted 27 lines · Jacob Keller
Quoted 13 lines · Michal Swiatkowski

The hardware is capable of calculating checksum for IPV6 packets with extension header. To not drop such packets switch from IP/IPV6 checksum to HW_CSUM.

HW_CSUM is also used in previous generation (i40e).

Previously HW_CSUM was used to indicate that hardware supports general checksum. Drop it assuming that if the hardware supports it, it is used.

Disabling offload for E830 in case of TSO isn't needed anymore as the check for TSO is done in Tx path just before preparation of the special GCS descriptor.

Quoted 8 lines

The commit from Fixes didn't introduce a bug, it just shown that the driver is doing sth wrong with the checksum features.

Fixes 04c20a9356f2 ("net: skip offload for NETIF_F_IPV6_CSUM if ipv6 header contains extension")

---

Am I correct in thinking that this supersedes (really, properly fixes) the patch "ice: enable NETIF_F_HW_CSUM for GSO packets" at
https://patchwork.ozlabs.org/project/intel-wired-lan/patch/[email protected]/
?

Quoted 7 lines
Quoted 2 lines

Thanks,
Jake

Yes, exactly. I think I linked it in cover letter, but maybe I should do it also here.

Thanks

I think its fine, I just wanted to make sure I was correct in marking that patch as rejected in patchwork.

Thanks,
Jake

Tested-by on v1 2/2
#11 lore raw

RE: [Intel-wired-lan] [PATCH iwl-net v1 2/2] ice: use NETIF_F_HW_CSUM instead of IP/IPV6

In reply to Michal Swiatkowski
Quoted 23 lines

-----Original Message-----
From: Intel-wired-lan <[email protected]> On Behalf Of Michal Swiatkowski
Sent: Tuesday, April 28, 2026 12:07 AM
To: [email protected]
Cc: [email protected]; [email protected]; Nguyen, Anthony L <[email protected]>; Kitszel, Przemyslaw <[email protected]>; Loktionov, Aleksandr <[email protected]>; Michal Swiatkowski <[email protected]>
Subject: [Intel-wired-lan] [PATCH iwl-net v1 2/2] ice: use NETIF_F_HW_CSUM instead of IP/IPV6

The hardware is capable of calculating checksum for IPV6 packets with extension header. To not drop such packets switch from IP/IPV6 checksum to HW_CSUM.

HW_CSUM is also used in previous generation (i40e).

Previously HW_CSUM was used to indicate that hardware supports general checksum. Drop it assuming that if the hardware supports it, it is used.

Disabling offload for E830 in case of TSO isn't needed anymore as the check for TSO is done in Tx path just before preparation of the special GCS descriptor.

The commit from Fixes didn't introduce a bug, it just shown that the driver is doing sth wrong with the checksum features.

Quoted 7 lines
Fixes 04c20a9356f2 ("net: skip offload for NETIF_F_IPV6_CSUM if ipv6 header contains extension")
---
 drivers/net/ethernet/intel/ice/ice_main.c | 21 +--------------------
 1 file changed, 1 insertion(+), 20 deletions(-)
Reviewed-by on v1 2/2
#12 lore raw

RE: [PATCH iwl-net v1 2/2] ice: use NETIF_F_HW_CSUM instead of IP/IPV6

In reply to Michal Swiatkowski
Quoted 97 lines

-----Original Message-----
From: Michal Swiatkowski <[email protected]>
Sent: Tuesday, April 28, 2026 9:07 AM
To: [email protected]
Cc: [email protected]; [email protected]; Nguyen, Anthony L <[email protected]>; Kitszel, Przemyslaw <[email protected]>; Loktionov, Aleksandr <[email protected]>; Michal Swiatkowski <[email protected]>
Subject: [PATCH iwl-net v1 2/2] ice: use NETIF_F_HW_CSUM instead of IP/IPV6

The hardware is capable of calculating checksum for IPV6 packets with extension header. To not drop such packets switch from IP/IPV6 checksum to HW_CSUM.

HW_CSUM is also used in previous generation (i40e).

Previously HW_CSUM was used to indicate that hardware supports general checksum. Drop it assuming that if the hardware supports it, it is used.

Disabling offload for E830 in case of TSO isn't needed anymore as the check for TSO is done in Tx path just before preparation of the special GCS descriptor.

The commit from Fixes didn't introduce a bug, it just shown that the driver is doing sth wrong with the checksum features.

Fixes 04c20a9356f2 ("net: skip offload for NETIF_F_IPV6_CSUM if ipv6

header contains extension")

---
 drivers/net/ethernet/intel/ice/ice_main.c | 21 +--------------------
 1 file changed, 1 insertion(+), 20 deletions(-)
drivers/net/ethernet/intel/ice/ice_main.c +1 -2
index 15550216fbf0..0f2f949af536 100644
@@ -3500,9 +3500,8 @@ void ice_set_netdev_features(struct net_device
 *netdev)
 			NETIF_F_RXHASH;
 
 	csumo_features = NETIF_F_RXCSUM	  |
-			 NETIF_F_IP_CSUM  |
 			 NETIF_F_SCTP_CRC |
-			 NETIF_F_IPV6_CSUM;
+			 NETIF_F_HW_CSUM;
 
 	vlano_features = NETIF_F_HW_VLAN_CTAG_FILTER |
drivers/net/ethernet/intel/ice/ice_main.c +0 -1
 			 NETIF_F_HW_VLAN_CTAG_TX     |
@@ -3564,12 +3563,6 @@ void ice_set_netdev_features(struct net_device
 *netdev)
 	/* Allow core to manage IRQs affinity */
 	netif_set_affinity_auto(netdev);
 
-	/* Mutual exclusivity for TSO and GCS is enforced by the set

features

drivers/net/ethernet/intel/ice/ice_main.c +0 -6
-	 * ndo callback.
-	 */
-	if (ice_is_feature_supported(pf, ICE_F_GCS))
-		netdev->hw_features |= NETIF_F_HW_CSUM;
-
 	netif_set_tso_max_size(netdev, ICE_MAX_TSO_SIZE);  }
@@ -6489,18 +6482,6 @@ ice_set_features(struct net_device *netdev,
 netdev_features_t features)
 	if (changed & NETIF_F_LOOPBACK)
 		ret = ice_set_loopback(vsi, !!(features &
 NETIF_F_LOOPBACK));
 
-	/* Due to E830 hardware limitations, TSO (NETIF_F_ALL_TSO) with

GCS

drivers/net/ethernet/intel/ice/ice_main.c +0 -4
-	 * (NETIF_F_HW_CSUM) is not supported.
-	 */
-	if (ice_is_feature_supported(pf, ICE_F_GCS) &&
-	    ((features & NETIF_F_HW_CSUM) && (features &

NETIF_F_ALL_TSO))) {

drivers/net/ethernet/intel/ice/ice_main.c +0 -2
-		if (netdev->features & NETIF_F_HW_CSUM)
-			dev_err(ice_pf_to_dev(pf), "To enable TSO, you

must first disable HW checksum.\n");

drivers/net/ethernet/intel/ice/ice_main.c +0 -2
-		else
-			dev_err(ice_pf_to_dev(pf), "To enable HW

checksum, you must first disable TSO.\n");

drivers/net/ethernet/intel/ice/ice_main.c +0 -4
-		return -EIO;
-	}
-
 	return ret;
 }
 
--
Quoted 1 line

2.49.0

#13 lore raw

Re: [PATCH iwl-net v1 2/2] ice: use NETIF_F_HW_CSUM instead of IP/IPV6

In reply to Michal Swiatkowski
Thank you for your contribution! Sashiko AI review found 3 potential issue(s) to consider: - [Critical] Replacing NETIF_F_IPV6_CSUM with NETIF_F_HW_CSUM bypasses the kernel's software checksum…

Thank you for your contribution! Sashiko AI review found 3 potential issue(s) to consider:
- [Critical] Replacing NETIF_F_IPV6_CSUM with NETIF_F_HW_CSUM bypasses the kernel's software checksum fallback for IPv6 extension headers, causing hardware to silently corrupt checksums.
- [High] Advertising NETIF_F_HW_CSUM forces the kernel to offload checksums for non-IP protocols, which the driver's Tx path unconditionally drops.
- [High] Unconditionally advertising NETIF_F_HW_CSUM forces non-TCP/UDP/SCTP packets into a buggy software fallback in the driver's Tx hot path that ignores errors, leading to silently corrupted packets. --

commit f1566dd32ecb941ad20b193c17ecae7df05ca815
Author: Michal Swiatkowski <[email protected]>

ice: use NETIF_F_HW_CSUM instead of IP/IPV6

This commit replaces the IP and IPV6 checksum offload flags with HW_CSUM in the ice driver to allow hardware checksumming for IPv6 packets with extension headers. It also removes the mutual exclusivity check between TSO and GCS.

drivers/net/ethernet/intel/ice/ice_main.c void ice_set_netdev_features(struct net_device *netdev) · line 3495

Quoted 12 lines
drivers/net/ethernet/intel/ice/ice_main.c +1 -2
index 5f92377d4dfc..33f252a6a22f 100644
@@ -3491,9 +3491,8 @@ void ice_set_netdev_features(struct net_device *netdev)
 			NETIF_F_RXHASH;
 
 	csumo_features = NETIF_F_RXCSUM	  |
-			 NETIF_F_IP_CSUM  |
 			 NETIF_F_SCTP_CRC |
-			 NETIF_F_IPV6_CSUM;
+			 NETIF_F_HW_CSUM;

By advertising NETIF_F_HW_CSUM, the network stack assumes the driver can handle checksums for any protocol and passes non-IP packets (e.g., from AF_PACKET injected frames) to the driver with CHECKSUM_PARTIAL.

Since ice_tx_csum() strictly parses the L3 header and expects IPv4 or IPv6, returning -1 for other protocols, will this regression cause non-IP packets to be silently dropped instead of falling back to software checksums in the kernel?

Additionally, skb_csum_hwoffload_help() relies on NETIF_F_IPV6_CSUM to safely fall back to software checksums when IPv6 extension headers are present.

Does using NETIF_F_HW_CSUM bypass this safeguard, potentially causing legacy hardware (like E810) or E830 without GCS descriptors to compute incorrect pseudo-headers and silently corrupt IPv6 packets with extension headers?

Finally, for protocols other than TCP, UDP, or SCTP, the driver falls back to calling skb_checksum_help(skb) in the default case of ice_tx_csum().

However, the driver appears to ignore the return value of skb_checksum_help().

If skb_checksum_help() fails (e.g., due to -ENOMEM during pskb_expand_head()), could this result in the packet being transmitted with an uncomputed checksum?

Signature · 1 line
Sashiko AI review · https://sashiko.dev/#/patchset/[email protected]?part=2