[PATCH 1/2] ixgbe: Avoid to write the RETA table when unnecessary.

[PATCH 1/2] ixgbe: Avoid to write the RETA table when unnecessary.

[Tonghao Zhang]

If indir == 0 in the ixgbe_set_rxfh(), it is unnecessary
to write the HW. Because redirection table is not changed.

Signed-off-by: Tonghao Zhang <xiangxia.m.yue@gmail.com>
---
 drivers/net/ethernet/intel/ixgbe/ixgbe_ethtool.c | 3 ++-
 1 file changed, 2 insertions(+), 1 deletion(-)

diff --git a/drivers/net/ethernet/intel/ixgbe/ixgbe_ethtool.c b/drivers/net/ethernet/intel/ixgbe/ixgbe_ethtool.c
index 0aaf70b..0f1d642 100644
--- a/drivers/net/ethernet/intel/ixgbe/ixgbe_ethtool.c
+++ b/drivers/net/ethernet/intel/ixgbe/ixgbe_ethtool.c
@@ -3060,6 +3060,8 @@ static int ixgbe_set_rxfh(struct net_device *netdev, const u32 *indir,
 
 		for (i = 0; i < reta_entries; i++)
 			adapter->rss_indir_tbl[i] = indir[i];
+
+		ixgbe_store_reta(adapter);
 	}
 
 	/* Fill out the rss hash key */
@@ -3068,7 +3070,6 @@ static int ixgbe_set_rxfh(struct net_device *netdev, const u32 *indir,
 		ixgbe_store_key(adapter);
 	}
 
-	ixgbe_store_reta(adapter);
 
 	return 0;
 }
-- 
1.8.3.1

[PATCH 2/2] ixgbe: Set rss key in order.

[Tonghao Zhang]

If we use ethtool to set rss key, for example,
we except 0x6d is K[0], and 0x5a is K[1]. But
the ixgbe driver set the 0xda is K[0].

ethtool -X eth0 hkey 6d:5a:56:da:25:5b:0e:c2:41:67:25:3d
	:43:a3:8f:b0:d0:ca:2b:cb:ae:7b:30:b4:77:cb:2d:a3
	:80:30:f2:0c:6a:42:b7:3b:be:ac:01:fa

The key is the original Microsoft's key.

7.1.2.8.1 RSS Hash Function: (82599 datasheet)
Given an array K with k bytes, our nomenclature assumes
that the array is laid out as follows:

K[0] K[1] K[2] ... K[k-1]

K[0] is the left-most byte, and the MSB of K[0] is the
left-most bit. K[k-1] is the right-most byte, and the
LSB of K[k-1] is the right-most bit.

8.2.3.7.14 RSS Random Key Register:
The K[0] is the 7:0 bit of RSSRK (RSS Random Key Register).

Signed-off-by: Tonghao Zhang <xiangxia.m.yue@gmail.com>
---
 drivers/net/ethernet/intel/ixgbe/ixgbe_main.c | 14 +++++++++++---
 1 file changed, 11 insertions(+), 3 deletions(-)

diff --git a/drivers/net/ethernet/intel/ixgbe/ixgbe_main.c b/drivers/net/ethernet/intel/ixgbe/ixgbe_main.c
index 95aba97..0ca2e0e 100644
--- a/drivers/net/ethernet/intel/ixgbe/ixgbe_main.c
+++ b/drivers/net/ethernet/intel/ixgbe/ixgbe_main.c
@@ -3777,10 +3777,18 @@ u32 ixgbe_rss_indir_tbl_entries(struct ixgbe_adapter *adapter)
 void ixgbe_store_key(struct ixgbe_adapter *adapter)
 {
 	struct ixgbe_hw *hw = &adapter->hw;
-	int i;
+	u8 *rss_key = (u8 *)adapter->rss_key;
+	u32 key, i;
 
-	for (i = 0; i < 10; i++)
-		IXGBE_WRITE_REG(hw, IXGBE_RSSRK(i), adapter->rss_key[i]);
+	/* Fill out hash function seeds */
+	for (i = 0; i < 10; i++) {
+		key  = rss_key[(i * 4)];
+		key |= rss_key[(i * 4) + 1] << 8;
+		key |= rss_key[(i * 4) + 2] << 16;
+		key |= rss_key[(i * 4) + 3] << 24;
+
+		IXGBE_WRITE_REG(hw, IXGBE_RSSRK(i), key);
+	}
 }
 
 /**
-- 
1.8.3.1

Re: [PATCH 1/2] ixgbe: Avoid to write the RETA table when unnecessary.

[Jeff Kirsher]

[-- Attachment #1: Type: text/plain, Size: 441 bytes --]

On Tue, 2018-01-23 at 07:24 -0800, Tonghao Zhang wrote:
> If indir == 0 in the ixgbe_set_rxfh(), it is unnecessary
> to write the HW. Because redirection table is not changed.
> 
> Signed-off-by: Tonghao Zhang <xiangxia.m.yue@gmail.com>
> ---
>  drivers/net/ethernet/intel/ixgbe/ixgbe_ethtool.c | 3 ++-
>  1 file changed, 2 insertions(+), 1 deletion(-)

Please send this series to the intel-wired-lan@lists.osuosl.org mailing
list.

[-- Attachment #2: This is a digitally signed message part --]
[-- Type: application/pgp-signature, Size: 833 bytes --]

Re: [PATCH 2/2] ixgbe: Set rss key in order.

[Alexander Duyck]

On Tue, Jan 23, 2018 at 7:24 AM, Tonghao Zhang <xiangxia.m.yue@gmail.com> wrote:
> If we use ethtool to set rss key, for example,
> we except 0x6d is K[0], and 0x5a is K[1]. But
> the ixgbe driver set the 0xda is K[0].
>
> ethtool -X eth0 hkey 6d:5a:56:da:25:5b:0e:c2:41:67:25:3d
>         :43:a3:8f:b0:d0:ca:2b:cb:ae:7b:30:b4:77:cb:2d:a3
>         :80:30:f2:0c:6a:42:b7:3b:be:ac:01:fa
>
> The key is the original Microsoft's key.
>
> 7.1.2.8.1 RSS Hash Function: (82599 datasheet)
> Given an array K with k bytes, our nomenclature assumes
> that the array is laid out as follows:
>
> K[0] K[1] K[2] ... K[k-1]
>
> K[0] is the left-most byte, and the MSB of K[0] is the
> left-most bit. K[k-1] is the right-most byte, and the
> LSB of K[k-1] is the right-most bit.
>
> 8.2.3.7.14 RSS Random Key Register:
> The K[0] is the 7:0 bit of RSSRK (RSS Random Key Register).
>
> Signed-off-by: Tonghao Zhang <xiangxia.m.yue@gmail.com>
> ---
>  drivers/net/ethernet/intel/ixgbe/ixgbe_main.c | 14 +++++++++++---
>  1 file changed, 11 insertions(+), 3 deletions(-)
>
> diff --git a/drivers/net/ethernet/intel/ixgbe/ixgbe_main.c b/drivers/net/ethernet/intel/ixgbe/ixgbe_main.c
> index 95aba97..0ca2e0e 100644
> --- a/drivers/net/ethernet/intel/ixgbe/ixgbe_main.c
> +++ b/drivers/net/ethernet/intel/ixgbe/ixgbe_main.c
> @@ -3777,10 +3777,18 @@ u32 ixgbe_rss_indir_tbl_entries(struct ixgbe_adapter *adapter)
>  void ixgbe_store_key(struct ixgbe_adapter *adapter)
>  {
>         struct ixgbe_hw *hw = &adapter->hw;
> -       int i;
> +       u8 *rss_key = (u8 *)adapter->rss_key;
> +       u32 key, i;
>
> -       for (i = 0; i < 10; i++)
> -               IXGBE_WRITE_REG(hw, IXGBE_RSSRK(i), adapter->rss_key[i]);
> +       /* Fill out hash function seeds */
> +       for (i = 0; i < 10; i++) {
> +               key  = rss_key[(i * 4)];
> +               key |= rss_key[(i * 4) + 1] << 8;
> +               key |= rss_key[(i * 4) + 2] << 16;
> +               key |= rss_key[(i * 4) + 3] << 24;
> +
> +               IXGBE_WRITE_REG(hw, IXGBE_RSSRK(i), key);
> +       }
>  }
>
>  /**

I would argue the issue here isn't in ixgbe_store_key. The problem
appears to be in ixgbe_set_rxfh and ixgbe_get_rxfh. It seems the key
provided should be populated as if it were a __be32 array, not a u32
array. So doing a memcpy from the u8 to u32 isn't correct in this
case. Really what should probably be happening is the population of
rss_key should be done using a for loop with be32_to_cpu calls, and
when we display it we should be using cpu_to_be32 to translate it
back.

An alternative would be to convert the rss_key pointer in the adapter
struct to be a be32 array pointer in the first place, and then change
the key population code to use be32_to_cpup in order to do the correct
byte swapping before writing the the value to the register.

Also as pointed out by Jeff please submit this to the intel-wired-lan
mailing list next time.

Thanks.

- Alex