[PATCH net-next] r8169: use netif_receive_skb_list batching

[PATCH net-next] r8169: use netif_receive_skb_list batching

[Heiner Kallweit]

Use netif_receive_skb_list() instead of napi_gro_receive() to benefit
from batched skb processing.

Signed-off-by: Heiner Kallweit <hkallweit1@gmail.com>
---
 drivers/net/ethernet/realtek/r8169.c | 5 ++++-
 1 file changed, 4 insertions(+), 1 deletion(-)

diff --git a/drivers/net/ethernet/realtek/r8169.c b/drivers/net/ethernet/realtek/r8169.c
index a8ca26c2a..c9ee1c8eb 100644
--- a/drivers/net/ethernet/realtek/r8169.c
+++ b/drivers/net/ethernet/realtek/r8169.c
@@ -6426,6 +6426,7 @@ static int rtl_rx(struct net_device *dev, struct rtl8169_private *tp, u32 budget
 {
 	unsigned int cur_rx, rx_left;
 	unsigned int count;
+	LIST_HEAD(rx_list);
 
 	cur_rx = tp->cur_rx;
 
@@ -6501,7 +6502,7 @@ static int rtl_rx(struct net_device *dev, struct rtl8169_private *tp, u32 budget
 			if (skb->pkt_type == PACKET_MULTICAST)
 				dev->stats.multicast++;
 
-			napi_gro_receive(&tp->napi, skb);
+			list_add_tail(&skb->list, &rx_list);
 
 			u64_stats_update_begin(&tp->rx_stats.syncp);
 			tp->rx_stats.packets++;
@@ -6516,6 +6517,8 @@ static int rtl_rx(struct net_device *dev, struct rtl8169_private *tp, u32 budget
 	count = cur_rx - tp->cur_rx;
 	tp->cur_rx = cur_rx;
 
+	netif_receive_skb_list(&rx_list);
+
 	return count;
 }
 
-- 
2.21.0

Re: [PATCH net-next] r8169: use netif_receive_skb_list batching

[David Miller]

From: Heiner Kallweit <hkallweit1@gmail.com>
Date: Sun, 31 Mar 2019 15:18:48 +0200

> Use netif_receive_skb_list() instead of napi_gro_receive() to benefit
> from batched skb processing.
> 
> Signed-off-by: Heiner Kallweit <hkallweit1@gmail.com>

Applied.

Re: [PATCH net-next] r8169: use netif_receive_skb_list batching

[Eric Dumazet]

On 03/31/2019 11:11 AM, David Miller wrote:
> From: Heiner Kallweit <hkallweit1@gmail.com>
> Date: Sun, 31 Mar 2019 15:18:48 +0200
> 
>> Use netif_receive_skb_list() instead of napi_gro_receive() to benefit
>> from batched skb processing.
>>
>> Signed-off-by: Heiner Kallweit <hkallweit1@gmail.com>
> 
> Applied.
> 

This means no GRO at all, and reduced performance in many useful cases.

Re: [PATCH net-next] r8169: use netif_receive_skb_list batching

[Jesper Dangaard Brouer]

On Mon, 1 Apr 2019 02:17:12 -0700
Eric Dumazet <eric.dumazet@gmail.com> wrote:

> On 03/31/2019 11:11 AM, David Miller wrote:
> > From: Heiner Kallweit <hkallweit1@gmail.com>
> > Date: Sun, 31 Mar 2019 15:18:48 +0200
> >   
> >> Use netif_receive_skb_list() instead of napi_gro_receive() to benefit
> >> from batched skb processing.
> >>
> >> Signed-off-by: Heiner Kallweit <hkallweit1@gmail.com>  
> > 
> > Applied.
> >   
> 
> This means no GRO at all, and reduced performance in many useful
> cases.

As Eric says, this in-effect disables GRO for this driver.  This will
very likely give a slowdown for local TCP delivery (while it might
helps for UDP and forwarding). 

That is why I was saying we need a GRO version of this...
(which need to be combined with bulk SKB alloc, to compete with GRO
fast-recycle of same SKB).

-- 
Best regards,
  Jesper Dangaard Brouer
  MSc.CS, Principal Kernel Engineer at Red Hat
  LinkedIn: http://www.linkedin.com/in/brouer

Re: [PATCH net-next] r8169: use netif_receive_skb_list batching

[David Miller]

From: Eric Dumazet <eric.dumazet@gmail.com>
Date: Mon, 1 Apr 2019 02:17:12 -0700

> This means no GRO at all

I do not think that is true as the SKB list layer does queue up to
GRO.

Re: [PATCH net-next] r8169: use netif_receive_skb_list batching

[Edward Cree]

On 01/04/2019 18:14, David Miller wrote:
> From: Eric Dumazet <eric.dumazet@gmail.com>
> Date: Mon, 1 Apr 2019 02:17:12 -0700
>
>> This means no GRO at all
> I do not think that is true as the SKB list layer does queue up to
> GRO.
No, Eric is right; the current list layer bypasses GRO completely.

netif_receive_skb_list() ends up doing the same things netif_receive_skb()
would do on each SKB in the list, and that does not include GRO.
(For this reason the sfc driver only uses netif_receive_skb_list() for
non-TCP packets; TCP packets go to napi_gro_frags().)
I had a patch series to add napi_gro_receive_list() which would use the
SKB list layer to handle the packets GRO didn't coalesce ([1]) but the
performance tests I ran were inconclusive and it never got applied.

-Ed

[1]: https://marc.info/?l=linux-netdev&m=154221888012410&w=2

Re: [PATCH net-next] r8169: use netif_receive_skb_list batching

[Heiner Kallweit]

On 01.04.2019 19:31, Edward Cree wrote:
> On 01/04/2019 18:14, David Miller wrote:
>> From: Eric Dumazet <eric.dumazet@gmail.com>
>> Date: Mon, 1 Apr 2019 02:17:12 -0700
>>
>>> This means no GRO at all
>> I do not think that is true as the SKB list layer does queue up to
>> GRO.
> No, Eric is right; the current list layer bypasses GRO completely.
> 
> netif_receive_skb_list() ends up doing the same things netif_receive_skb()
> would do on each SKB in the list, and that does not include GRO.
> (For this reason the sfc driver only uses netif_receive_skb_list() for
> non-TCP packets; TCP packets go to napi_gro_frags().)
> I had a patch series to add napi_gro_receive_list() which would use the
> SKB list layer to handle the packets GRO didn't coalesce ([1]) but the
> performance tests I ran were inconclusive and it never got applied.
> 
> -Ed
> 
> [1]: https://marc.info/?l=linux-netdev&m=154221888012410&w=2
> 
Patch 0a25d92c6f4f ("dpaa2-eth: use netif_receive_skb_list") was well
perceived [1], therefore I was under the assumption that
netif_receive_skb_list is kind of successor for napi_gro_receive().
Does what you say apply to that patch too?

Based on feedback so far it seems to best if I revert the patch.

Heiner

[1] https://patchwork.ozlabs.org/patch/1064334/

Re: [PATCH net-next] r8169: use netif_receive_skb_list batching

[David Miller]

From: Jesper Dangaard Brouer <brouer@redhat.com>
Date: Mon, 1 Apr 2019 14:53:26 +0200

> On Mon, 1 Apr 2019 02:17:12 -0700
> Eric Dumazet <eric.dumazet@gmail.com> wrote:
> 
>> On 03/31/2019 11:11 AM, David Miller wrote:
>> > From: Heiner Kallweit <hkallweit1@gmail.com>
>> > Date: Sun, 31 Mar 2019 15:18:48 +0200
>> >   
>> >> Use netif_receive_skb_list() instead of napi_gro_receive() to benefit
>> >> from batched skb processing.
>> >>
>> >> Signed-off-by: Heiner Kallweit <hkallweit1@gmail.com>  
>> > 
>> > Applied.
>> >   
>> 
>> This means no GRO at all, and reduced performance in many useful
>> cases.
> 
> As Eric says, this in-effect disables GRO for this driver.

Really?  __netif_receive_skb_list_core() does GRO.

Re: [PATCH net-next] r8169: use netif_receive_skb_list batching

[Eric Dumazet]

On 04/01/2019 10:41 AM, David Miller wrote:
> From: Jesper Dangaard Brouer <brouer@redhat.com>
>>
>> As Eric says, this in-effect disables GRO for this driver.
> 
> Really?  __netif_receive_skb_list_core() does GRO.

At least on current public trees, it can not do GRO,
since we do not pass a napi context.