Currently, passing HW hash from descriptors to skb is broken two times.
The first bug effectively disables hash from being filled at all, unless
%NETIF_F_RXHASH is *disabled* via Ethtool. The second incorrectly says
that IPv6 UDP packets are L3, which also triggers CPU hashing when
needed (the networking core treats only L4 HW hash as "true").
The very same problems were fixed in i40e and ice, but not in iavf,
although each of the original commits bugged at least two drivers.
It's never too late (I hope), so fix iavf this time.
Alexander Lobakin (2):
iavf: fix inverted Rx hash condition leading to disabled hash
iavf: fix non-tunneled IPv6 UDP packet type and hashing
drivers/net/ethernet/intel/iavf/iavf_common.c | 2 +-
drivers/net/ethernet/intel/iavf/iavf_txrx.c | 2 +-
2 files changed, 2 insertions(+), 2 deletions(-)
---
To Tony: this is very trivial and tested for a while already, I hope it
could hit one of the first couple RCs :p
--
2.39.2
Condition, which checks whether the netdev has hashing enabled is
inverted. Basically, the tagged commit effectively disabled passing flow
hash from descriptor to skb, unless user *disables* it via Ethtool.
Commit a876c3ba59a6 ("i40e/i40evf: properly report Rx packet hash")
fixed this problem, but only for i40e.
Invert the condition now in iavf and unblock passing hash to skbs again.
Fixes: 857942fd1aa1 ("i40e: Fix Rx hash reported to the stack by our driver")
Reviewed-by: Larysa Zaremba <[email protected]>
Reviewed-by: Michal Kubiak <[email protected]>
Signed-off-by: Alexander Lobakin <[email protected]>
---
drivers/net/ethernet/intel/iavf/iavf_txrx.c | 2 +-
1 file changed, 1 insertion(+), 1 deletion(-)
diff --git a/drivers/net/ethernet/intel/iavf/iavf_txrx.c b/drivers/net/ethernet/intel/iavf/iavf_txrx.c
index 18b6a702a1d6..e989feda133c 100644
--- a/drivers/net/ethernet/intel/iavf/iavf_txrx.c
+++ b/drivers/net/ethernet/intel/iavf/iavf_txrx.c
@@ -1096,7 +1096,7 @@ static inline void iavf_rx_hash(struct iavf_ring *ring,
cpu_to_le64((u64)IAVF_RX_DESC_FLTSTAT_RSS_HASH <<
IAVF_RX_DESC_STATUS_FLTSTAT_SHIFT);
- if (ring->netdev->features & NETIF_F_RXHASH)
+ if (!(ring->netdev->features & NETIF_F_RXHASH))
return;
if ((rx_desc->wb.qword1.status_error_len & rss_mask) == rss_mask) {
--
2.39.2
Currently, IAVF's decode_rx_desc_ptype() correctly reports payload type
of L4 for IPv4 UDP packets and IPv{4,6} TCP, but only L3 for IPv6 UDP.
Originally, i40e, ice and iavf were affected.
Commit 73df8c9e3e3d ("i40e: Correct UDP packet header for non_tunnel-ipv6")
fixed that in i40e, then
commit 638a0c8c8861 ("ice: fix incorrect payload indicator on PTYPE")
fixed that for ice.
IPv6 UDP is L4 obviously. Fix it and make iavf report correct L4 hash
type for such packets, so that the stack won't calculate it on CPU when
needs it.
Fixes: 206812b5fccb ("i40e/i40evf: i40e implementation for skb_set_hash")
Reviewed-by: Larysa Zaremba <[email protected]>
Reviewed-by: Michal Kubiak <[email protected]>
Signed-off-by: Alexander Lobakin <[email protected]>
---
drivers/net/ethernet/intel/iavf/iavf_common.c | 2 +-
1 file changed, 1 insertion(+), 1 deletion(-)
diff --git a/drivers/net/ethernet/intel/iavf/iavf_common.c b/drivers/net/ethernet/intel/iavf/iavf_common.c
index 16c490965b61..dd11dbbd5551 100644
--- a/drivers/net/ethernet/intel/iavf/iavf_common.c
+++ b/drivers/net/ethernet/intel/iavf/iavf_common.c
@@ -661,7 +661,7 @@ struct iavf_rx_ptype_decoded iavf_ptype_lookup[BIT(8)] = {
/* Non Tunneled IPv6 */
IAVF_PTT(88, IP, IPV6, FRG, NONE, NONE, NOF, NONE, PAY3),
IAVF_PTT(89, IP, IPV6, NOF, NONE, NONE, NOF, NONE, PAY3),
- IAVF_PTT(90, IP, IPV6, NOF, NONE, NONE, NOF, UDP, PAY3),
+ IAVF_PTT(90, IP, IPV6, NOF, NONE, NONE, NOF, UDP, PAY4),
IAVF_PTT_UNUSED_ENTRY(91),
IAVF_PTT(92, IP, IPV6, NOF, NONE, NONE, NOF, TCP, PAY4),
IAVF_PTT(93, IP, IPV6, NOF, NONE, NONE, NOF, SCTP, PAY4),
--
2.39.2
On Wed, Mar 01, 2023 at 12:59:06PM +0100, Alexander Lobakin wrote:
> Currently, passing HW hash from descriptors to skb is broken two times.
> The first bug effectively disables hash from being filled at all, unless
> %NETIF_F_RXHASH is *disabled* via Ethtool. The second incorrectly says
> that IPv6 UDP packets are L3, which also triggers CPU hashing when
> needed (the networking core treats only L4 HW hash as "true").
> The very same problems were fixed in i40e and ice, but not in iavf,
> although each of the original commits bugged at least two drivers.
> It's never too late (I hope), so fix iavf this time.
>
> Alexander Lobakin (2):
> iavf: fix inverted Rx hash condition leading to disabled hash
> iavf: fix non-tunneled IPv6 UDP packet type and hashing
>
Thanks,
Reviewed-by: Leon Romanovsky <[email protected]>