[PATCH net-next 0/2] igc: driver change to support XDP metadata

Changes to fix and enable XDP metadata to a specific Intel driver igc.
Tested with hardware i225 that uses driver igc, while testing AF_XDP
access to metadata area.

---

Jesper Dangaard Brouer (2):
      igc: AF_XDP zero-copy metadata adjust breaks SKBs on XDP_PASS
      igc: enable XDP metadata in driver


 drivers/net/ethernet/intel/igc/igc_main.c |   33 +++++++++++++++++++----------
 1 file changed, 22 insertions(+), 11 deletions(-)

--

Enabling the XDP bpf_prog access to data_meta area is a very small
change. Hint passing 'true' to xdp_prepare_buff().

The SKB layers can also access data_meta area, which required more
driver changes to support. Reviewers, notice the igc driver have two
different functions that can create SKBs, depending on driver config.

Hint for testers, ethtool priv-flags legacy-rx enables
the function igc_construct_skb()

 ethtool --set-priv-flags DEV legacy-rx on

Signed-off-by: Jesper Dangaard Brouer <redacted>
---
 drivers/net/ethernet/intel/igc/igc_main.c |   29 +++++++++++++++++++----------
 1 file changed, 19 insertions(+), 10 deletions(-)

diff --git a/drivers/net/ethernet/intel/igc/igc_main.c b/drivers/net/ethernet/intel/igc/igc_main.c
index 76b0a7311369..b516f1b301b4 100644
--- a/drivers/net/ethernet/intel/igc/igc_main.c
+++ b/drivers/net/ethernet/intel/igc/igc_main.c

@@ -1718,24 +1718,26 @@ static void igc_add_rx_frag(struct igc_ring *rx_ring,
 
 static struct sk_buff *igc_build_skb(struct igc_ring *rx_ring,
 				     struct igc_rx_buffer *rx_buffer,
-				     union igc_adv_rx_desc *rx_desc,
-				     unsigned int size)
+				     struct xdp_buff *xdp)
 {
-	void *va = page_address(rx_buffer->page) + rx_buffer->page_offset;
+	unsigned int size = xdp->data_end - xdp->data;
 	unsigned int truesize = igc_get_rx_frame_truesize(rx_ring, size);
+	unsigned int metasize = xdp->data - xdp->data_meta;
 	struct sk_buff *skb;
 
 	/* prefetch first cache line of first page */
-	net_prefetch(va);
+	net_prefetch(xdp->data);
 
 	/* build an skb around the page buffer */
-	skb = build_skb(va - IGC_SKB_PAD, truesize);
+	skb = build_skb(xdp->data_hard_start, truesize);
 	if (unlikely(!skb))
 		return NULL;
 
 	/* update pointers within the skb to store the data */
-	skb_reserve(skb, IGC_SKB_PAD);
+	skb_reserve(skb, xdp->data - xdp->data_hard_start);
 	__skb_put(skb, size);
+	if (metasize)
+		skb_metadata_set(skb, metasize);
 
 	igc_rx_buffer_flip(rx_buffer, truesize);
 	return skb;

@@ -1746,6 +1748,7 @@ static struct sk_buff *igc_construct_skb(struct igc_ring *rx_ring,
 					 struct xdp_buff *xdp,
 					 ktime_t timestamp)
 {
+	unsigned int metasize = xdp->data - xdp->data_meta;
 	unsigned int size = xdp->data_end - xdp->data;
 	unsigned int truesize = igc_get_rx_frame_truesize(rx_ring, size);
 	void *va = xdp->data;

@@ -1756,7 +1759,7 @@ static struct sk_buff *igc_construct_skb(struct igc_ring *rx_ring,
 	net_prefetch(va);
 
 	/* allocate a skb to store the frags */
-	skb = napi_alloc_skb(&rx_ring->q_vector->napi, IGC_RX_HDR_LEN);
+	skb = napi_alloc_skb(&rx_ring->q_vector->napi, IGC_RX_HDR_LEN + metasize);
 	if (unlikely(!skb))
 		return NULL;
 

@@ -1769,7 +1772,13 @@ static struct sk_buff *igc_construct_skb(struct igc_ring *rx_ring,
 		headlen = eth_get_headlen(skb->dev, va, IGC_RX_HDR_LEN);
 
 	/* align pull length to size of long to optimize memcpy performance */
-	memcpy(__skb_put(skb, headlen), va, ALIGN(headlen, sizeof(long)));
+	memcpy(__skb_put(skb, headlen + metasize), xdp->data_meta,
+	       ALIGN(headlen + metasize, sizeof(long)));
+
+	if (metasize) {
+		skb_metadata_set(skb, metasize);
+		__skb_pull(skb, metasize);
+	}
 
 	/* update all of the pointers */
 	size -= headlen;

@@ -2354,7 +2363,7 @@ static int igc_clean_rx_irq(struct igc_q_vector *q_vector, const int budget)
 		if (!skb) {
 			xdp_init_buff(&xdp, truesize, &rx_ring->xdp_rxq);
 			xdp_prepare_buff(&xdp, pktbuf - igc_rx_offset(rx_ring),
-					 igc_rx_offset(rx_ring) + pkt_offset, size, false);
+					 igc_rx_offset(rx_ring) + pkt_offset, size, true);
 
 			skb = igc_xdp_run_prog(adapter, &xdp);
 		}

@@ -2378,7 +2387,7 @@ static int igc_clean_rx_irq(struct igc_q_vector *q_vector, const int budget)
 		} else if (skb)
 			igc_add_rx_frag(rx_ring, rx_buffer, skb, size);
 		else if (ring_uses_build_skb(rx_ring))
-			skb = igc_build_skb(rx_ring, rx_buffer, rx_desc, size);
+			skb = igc_build_skb(rx_ring, rx_buffer, &xdp);
 		else
 			skb = igc_construct_skb(rx_ring, rx_buffer, &xdp,
 						timestamp);

Driver already implicitly supports XDP metadata access in AF_XDP
zero-copy mode, as xsk_buff_pool's xp_alloc() naturally set xdp_buff
data_meta equal data.

This works fine for XDP and AF_XDP, but if a BPF-prog adjust via
bpf_xdp_adjust_meta() and choose to call XDP_PASS, then igc function
igc_construct_skb_zc() will construct an invalid SKB packet. The
function correctly include the xdp->data_meta area in the memcpy, but
forgot to pull header to take metasize into account.

Fixes: fc9df2a0b520 ("igc: Enable RX via AF_XDP zero-copy")
Signed-off-by: Jesper Dangaard Brouer <redacted>
---
 drivers/net/ethernet/intel/igc/igc_main.c |    4 +++-
 1 file changed, 3 insertions(+), 1 deletion(-)

diff --git a/drivers/net/ethernet/intel/igc/igc_main.c b/drivers/net/ethernet/intel/igc/igc_main.c
index 8e448288ee26..76b0a7311369 100644
--- a/drivers/net/ethernet/intel/igc/igc_main.c
+++ b/drivers/net/ethernet/intel/igc/igc_main.c

@@ -2448,8 +2448,10 @@ static struct sk_buff *igc_construct_skb_zc(struct igc_ring *ring,
 
 	skb_reserve(skb, xdp->data_meta - xdp->data_hard_start);
 	memcpy(__skb_put(skb, totalsize), xdp->data_meta, totalsize);
-	if (metasize)
+	if (metasize) {
 		skb_metadata_set(skb, metasize);
+		__skb_pull(skb, metasize);
+	}
 
 	return skb;
 }

On 11/15/2021 22:36, Jesper Dangaard Brouer wrote:

Driver already implicitly supports XDP metadata access in AF_XDP
zero-copy mode, as xsk_buff_pool's xp_alloc() naturally set xdp_buff
data_meta equal data.

This works fine for XDP and AF_XDP, but if a BPF-prog adjust via
bpf_xdp_adjust_meta() and choose to call XDP_PASS, then igc function
igc_construct_skb_zc() will construct an invalid SKB packet. The
function correctly include the xdp->data_meta area in the memcpy, but
forgot to pull header to take metasize into account.

Fixes: fc9df2a0b520 ("igc: Enable RX via AF_XDP zero-copy")
Signed-off-by: Jesper Dangaard Brouer <redacted>
---
  drivers/net/ethernet/intel/igc/igc_main.c |    4 +++-
  1 file changed, 3 insertions(+), 1 deletion(-)

Tested-by: Nechama Kraus <redacted>

On 11/15/2021 22:36, Jesper Dangaard Brouer wrote:

Enabling the XDP bpf_prog access to data_meta area is a very small
change. Hint passing 'true' to xdp_prepare_buff().

The SKB layers can also access data_meta area, which required more
driver changes to support. Reviewers, notice the igc driver have two
different functions that can create SKBs, depending on driver config.

Hint for testers, ethtool priv-flags legacy-rx enables
the function igc_construct_skb()

  ethtool --set-priv-flags DEV legacy-rx on

Signed-off-by: Jesper Dangaard Brouer <redacted>
---
  drivers/net/ethernet/intel/igc/igc_main.c |   29 +++++++++++++++++++----------
  1 file changed, 19 insertions(+), 10 deletions(-)

Tested-by: Nechama Kraus <redacted>

On Mon, Nov 15, 2021 at 09:36:25PM +0100, Jesper Dangaard Brouer wrote:

Driver already implicitly supports XDP metadata access in AF_XDP
zero-copy mode, as xsk_buff_pool's xp_alloc() naturally set xdp_buff
data_meta equal data.

This works fine for XDP and AF_XDP, but if a BPF-prog adjust via
bpf_xdp_adjust_meta() and choose to call XDP_PASS, then igc function
igc_construct_skb_zc() will construct an invalid SKB packet. The
function correctly include the xdp->data_meta area in the memcpy, but
forgot to pull header to take metasize into account.

Fixes: fc9df2a0b520 ("igc: Enable RX via AF_XDP zero-copy")
Signed-off-by: Jesper Dangaard Brouer <redacted>

Acked-by: Maciej Fijalkowski <maciej.fijalkowski@intel.com>

Great catch. Will take a look at other ZC enabled Intel drivers if they
are affected as well.

Thanks!

On 26/11/2021 16.25, Maciej Fijalkowski wrote:

On Mon, Nov 15, 2021 at 09:36:25PM +0100, Jesper Dangaard Brouer wrote:

quoted

Driver already implicitly supports XDP metadata access in AF_XDP
zero-copy mode, as xsk_buff_pool's xp_alloc() naturally set xdp_buff
data_meta equal data.

This works fine for XDP and AF_XDP, but if a BPF-prog adjust via
bpf_xdp_adjust_meta() and choose to call XDP_PASS, then igc function
igc_construct_skb_zc() will construct an invalid SKB packet. The
function correctly include the xdp->data_meta area in the memcpy, but
forgot to pull header to take metasize into account.

Fixes: fc9df2a0b520 ("igc: Enable RX via AF_XDP zero-copy")
Signed-off-by: Jesper Dangaard Brouer<redacted>

Acked-by: Maciej Fijalkowski<maciej.fijalkowski@intel.com>

Great catch. Will take a look at other ZC enabled Intel drivers if they
are affected as well.

Thanks a lot for taking this task!!! :-)
--Jesper

From: Jesper Dangaard Brouer <redacted>
Date: Fri, 26 Nov 2021 16:32:47 +0100

On 26/11/2021 16.25, Maciej Fijalkowski wrote:

quoted

On Mon, Nov 15, 2021 at 09:36:25PM +0100, Jesper Dangaard Brouer wrote:

quoted

Driver already implicitly supports XDP metadata access in AF_XDP
zero-copy mode, as xsk_buff_pool's xp_alloc() naturally set xdp_buff
data_meta equal data.

This works fine for XDP and AF_XDP, but if a BPF-prog adjust via
bpf_xdp_adjust_meta() and choose to call XDP_PASS, then igc function
igc_construct_skb_zc() will construct an invalid SKB packet. The
function correctly include the xdp->data_meta area in the memcpy, but
forgot to pull header to take metasize into account.

Fixes: fc9df2a0b520 ("igc: Enable RX via AF_XDP zero-copy")
Signed-off-by: Jesper Dangaard Brouer<redacted>

Acked-by: Maciej Fijalkowski<maciej.fijalkowski@intel.com>

Great catch. Will take a look at other ZC enabled Intel drivers if they
are affected as well.

They are. We'll cover them in a separate series, much thanks for
revealing that (:

Thanks a lot for taking this task!!! :-)
--Jesper

Al

From: Jesper Dangaard Brouer <redacted>
Date: Mon, 15 Nov 2021 21:36:30 +0100

I'd prefer prefetching xdp->data_meta here. GRO layer accesses it.
Maximum meta size for now is 32, so at least 96 bytes of the frame
will stil be prefetched.

...here as well.

Thanks!
Al

From: Jesper Dangaard Brouer <redacted>
Date: Mon, 15 Nov 2021 21:36:20 +0100

Changes to fix and enable XDP metadata to a specific Intel driver igc.
Tested with hardware i225 that uses driver igc, while testing AF_XDP
access to metadata area.

Would you mind if I take this your series into my bigger one that
takes care of it throughout all the Intel drivers?

---

Jesper Dangaard Brouer (2):
      igc: AF_XDP zero-copy metadata adjust breaks SKBs on XDP_PASS
      igc: enable XDP metadata in driver


 drivers/net/ethernet/intel/igc/igc_main.c |   33 +++++++++++++++++++----------
 1 file changed, 22 insertions(+), 11 deletions(-)

--

Thanks,
Al

On 29/11/2021 15.10, Alexander Lobakin wrote:

From: Jesper Dangaard Brouer <redacted>
Date: Mon, 15 Nov 2021 21:36:20 +0100

quoted

Changes to fix and enable XDP metadata to a specific Intel driver igc.
Tested with hardware i225 that uses driver igc, while testing AF_XDP
access to metadata area.

Would you mind if I take this your series into my bigger one that
takes care of it throughout all the Intel drivers?

I have a customer that depend on this fix.  They will have to do the 
backport anyway (to v5.13), but it would bring confidence on their side 
if the commits appear in an official git-tree before doing the backport 
(and optimally with a SHA they can refer to).

Tony Nguyen have these landed in your git-tree?

--JEsper

On 26/11/2021 17.16, Alexander Lobakin wrote:

From: Jesper Dangaard Brouer <redacted>
Date: Mon, 15 Nov 2021 21:36:30 +0100

quoted

Enabling the XDP bpf_prog access to data_meta area is a very small
change. Hint passing 'true' to xdp_prepare_buff().

The SKB layers can also access data_meta area, which required more
driver changes to support. Reviewers, notice the igc driver have two
different functions that can create SKBs, depending on driver config.

Hint for testers, ethtool priv-flags legacy-rx enables
the function igc_construct_skb()

  ethtool --set-priv-flags DEV legacy-rx on

Signed-off-by: Jesper Dangaard Brouer <redacted>
---
  drivers/net/ethernet/intel/igc/igc_main.c |   29 +++++++++++++++++++----------
  1 file changed, 19 insertions(+), 10 deletions(-)

diff --git a/drivers/net/ethernet/intel/igc/igc_main.c b/drivers/net/ethernet/intel/igc/igc_main.c
index 76b0a7311369..b516f1b301b4 100644
--- a/drivers/net/ethernet/intel/igc/igc_main.c
+++ b/drivers/net/ethernet/intel/igc/igc_main.c

@@ -1718,24 +1718,26 @@ static void igc_add_rx_frag(struct igc_ring *rx_ring,
  
  static struct sk_buff *igc_build_skb(struct igc_ring *rx_ring,
  				     struct igc_rx_buffer *rx_buffer,
-				     union igc_adv_rx_desc *rx_desc,
-				     unsigned int size)
+				     struct xdp_buff *xdp)
  {
-	void *va = page_address(rx_buffer->page) + rx_buffer->page_offset;
+	unsigned int size = xdp->data_end - xdp->data;
  	unsigned int truesize = igc_get_rx_frame_truesize(rx_ring, size);
+	unsigned int metasize = xdp->data - xdp->data_meta;
  	struct sk_buff *skb;
  
  	/* prefetch first cache line of first page */
-	net_prefetch(va);
+	net_prefetch(xdp->data);

I'd prefer prefetching xdp->data_meta here. GRO layer accesses it.
Maximum meta size for now is 32, so at least 96 bytes of the frame
will stil be prefetched.

Prefetch works for "full" cachelines. Intel CPUs often prefect two 
cache-lines, when doing this, thus I guess we still get xdp->data.

I don't mind prefetching xdp->data_meta, but (1) I tried to keep the 
change minimal as current behavior was data area I kept that. (2) 
xdp->data starts on a cacheline and we know NIC hardware have touched 
that, it is not a full-cache-miss due to DDIO/DCA it is known to be in 
L3 cache (gain is around 2-3 ns in my machine for data prefetch).
Given this is only a 2.5 Gbit/s driver/HW I doubt this make any difference.

Tony is it worth resending a V2 of this patch?

quoted

  
  	/* build an skb around the page buffer */
-	skb = build_skb(va - IGC_SKB_PAD, truesize);
+	skb = build_skb(xdp->data_hard_start, truesize);
  	if (unlikely(!skb))
  		return NULL;
  
  	/* update pointers within the skb to store the data */
-	skb_reserve(skb, IGC_SKB_PAD);
+	skb_reserve(skb, xdp->data - xdp->data_hard_start);
  	__skb_put(skb, size);
+	if (metasize)
+		skb_metadata_set(skb, metasize);
  
  	igc_rx_buffer_flip(rx_buffer, truesize);
  	return skb;

@@ -1746,6 +1748,7 @@ static struct sk_buff *igc_construct_skb(struct igc_ring *rx_ring,
  					 struct xdp_buff *xdp,
  					 ktime_t timestamp)
  {
+	unsigned int metasize = xdp->data - xdp->data_meta;
  	unsigned int size = xdp->data_end - xdp->data;
  	unsigned int truesize = igc_get_rx_frame_truesize(rx_ring, size);
  	void *va = xdp->data;

@@ -1756,7 +1759,7 @@ static struct sk_buff *igc_construct_skb(struct igc_ring *rx_ring,
  	net_prefetch(va);

...here as well.

From: Jesper Dangaard Brouer <redacted>
Date: Mon, 29 Nov 2021 15:29:07 +0100

On 29/11/2021 15.10, Alexander Lobakin wrote:

quoted

From: Jesper Dangaard Brouer <redacted>
Date: Mon, 15 Nov 2021 21:36:20 +0100

quoted

Changes to fix and enable XDP metadata to a specific Intel driver igc.
Tested with hardware i225 that uses driver igc, while testing AF_XDP
access to metadata area.

Would you mind if I take this your series into my bigger one that
takes care of it throughout all the Intel drivers?

I have a customer that depend on this fix.  They will have to do the 
backport anyway (to v5.13), but it would bring confidence on their side 
if the commits appear in an official git-tree before doing the backport 
(and optimally with a SHA they can refer to).

Yeah, sure, it's totally fine to get them accepted separately, I'll
just refer to them in my series.

Tony Nguyen have these landed in your git-tree?

Doesn't seem like. The reason might be that you responded to my
patch 2/2 comments only now.

--JEsper

Al

From: Jesper Dangaard Brouer <redacted>
Date: Mon, 29 Nov 2021 15:39:04 +0100

On 26/11/2021 17.16, Alexander Lobakin wrote:

quoted

From: Jesper Dangaard Brouer <redacted>
Date: Mon, 15 Nov 2021 21:36:30 +0100

quoted

Enabling the XDP bpf_prog access to data_meta area is a very small
change. Hint passing 'true' to xdp_prepare_buff().

The SKB layers can also access data_meta area, which required more
driver changes to support. Reviewers, notice the igc driver have two
different functions that can create SKBs, depending on driver config.

Hint for testers, ethtool priv-flags legacy-rx enables
the function igc_construct_skb()

  ethtool --set-priv-flags DEV legacy-rx on

Signed-off-by: Jesper Dangaard Brouer <redacted>
---
  drivers/net/ethernet/intel/igc/igc_main.c |   29 +++++++++++++++++++----------
  1 file changed, 19 insertions(+), 10 deletions(-)

diff --git a/drivers/net/ethernet/intel/igc/igc_main.c b/drivers/net/ethernet/intel/igc/igc_main.c
index 76b0a7311369..b516f1b301b4 100644
--- a/drivers/net/ethernet/intel/igc/igc_main.c
+++ b/drivers/net/ethernet/intel/igc/igc_main.c

@@ -1718,24 +1718,26 @@ static void igc_add_rx_frag(struct igc_ring *rx_ring,
  
  static struct sk_buff *igc_build_skb(struct igc_ring *rx_ring,
  				     struct igc_rx_buffer *rx_buffer,
-				     union igc_adv_rx_desc *rx_desc,
-				     unsigned int size)
+				     struct xdp_buff *xdp)
  {
-	void *va = page_address(rx_buffer->page) + rx_buffer->page_offset;
+	unsigned int size = xdp->data_end - xdp->data;
  	unsigned int truesize = igc_get_rx_frame_truesize(rx_ring, size);
+	unsigned int metasize = xdp->data - xdp->data_meta;
  	struct sk_buff *skb;
  
  	/* prefetch first cache line of first page */
-	net_prefetch(va);
+	net_prefetch(xdp->data);

I'd prefer prefetching xdp->data_meta here. GRO layer accesses it.
Maximum meta size for now is 32, so at least 96 bytes of the frame
will stil be prefetched.

Prefetch works for "full" cachelines. Intel CPUs often prefect two 
cache-lines, when doing this, thus I guess we still get xdp->data.

Sure. I mean, net_prefetch() prefetches 128 bytes in a row.
xdp->data is usually aligned to XDP_PACKET_HEADROOM (or two bytes
to the right). If our CL is 64 and the meta is present, then... ah
right, 64 to the left and 64 starting from data to the right.

I don't mind prefetching xdp->data_meta, but (1) I tried to keep the 
change minimal as current behavior was data area I kept that. (2) 
xdp->data starts on a cacheline and we know NIC hardware have touched 
that, it is not a full-cache-miss due to DDIO/DCA it is known to be in 
L3 cache (gain is around 2-3 ns in my machine for data prefetch).
Given this is only a 2.5 Gbit/s driver/HW I doubt this make any difference.

Code constistency at least. On 10+ Gbps we prefetch meta, and I plan
to continue doing this in my series.

Tony is it worth resending a V2 of this patch?

Tony, you can take it as it is if you want, I'll correct it later in
mine. Up to you.

Reviewed-by: Alexander Lobakin <redacted>

quoted

quoted

  
  	/* build an skb around the page buffer */
-	skb = build_skb(va - IGC_SKB_PAD, truesize);
+	skb = build_skb(xdp->data_hard_start, truesize);
  	if (unlikely(!skb))
  		return NULL;
  
  	/* update pointers within the skb to store the data */
-	skb_reserve(skb, IGC_SKB_PAD);
+	skb_reserve(skb, xdp->data - xdp->data_hard_start);
  	__skb_put(skb, size);
+	if (metasize)
+		skb_metadata_set(skb, metasize);
  
  	igc_rx_buffer_flip(rx_buffer, truesize);
  	return skb;

@@ -1746,6 +1748,7 @@ static struct sk_buff *igc_construct_skb(struct igc_ring *rx_ring,
  					 struct xdp_buff *xdp,
  					 ktime_t timestamp)
  {
+	unsigned int metasize = xdp->data - xdp->data_meta;
  	unsigned int size = xdp->data_end - xdp->data;
  	unsigned int truesize = igc_get_rx_frame_truesize(rx_ring, size);
  	void *va = xdp->data;

@@ -1756,7 +1759,7 @@ static struct sk_buff *igc_construct_skb(struct igc_ring *rx_ring,
  	net_prefetch(va);

...here as well.

Thanks,
Al

From: Alexander Lobakin <redacted>
Date: Mon, 29 Nov 2021 15:53:03 +0100

From: Jesper Dangaard Brouer <redacted>
Date: Mon, 29 Nov 2021 15:39:04 +0100

quoted

On 26/11/2021 17.16, Alexander Lobakin wrote:

quoted

From: Jesper Dangaard Brouer <redacted>
Date: Mon, 15 Nov 2021 21:36:30 +0100

quoted

Enabling the XDP bpf_prog access to data_meta area is a very small
change. Hint passing 'true' to xdp_prepare_buff().

[ snip ]

quoted

Prefetch works for "full" cachelines. Intel CPUs often prefect two 
cache-lines, when doing this, thus I guess we still get xdp->data.

Sure. I mean, net_prefetch() prefetches 128 bytes in a row.
xdp->data is usually aligned to XDP_PACKET_HEADROOM (or two bytes
to the right). If our CL is 64 and the meta is present, then... ah
right, 64 to the left and 64 starting from data to the right.

quoted

I don't mind prefetching xdp->data_meta, but (1) I tried to keep the 
change minimal as current behavior was data area I kept that. (2) 
xdp->data starts on a cacheline and we know NIC hardware have touched 
that, it is not a full-cache-miss due to DDIO/DCA it is known to be in 
L3 cache (gain is around 2-3 ns in my machine for data prefetch).
Given this is only a 2.5 Gbit/s driver/HW I doubt this make any difference.

Code constistency at least. On 10+ Gbps we prefetch meta, and I plan
to continue doing this in my series.

quoted

Tony is it worth resending a V2 of this patch?

Tony, you can take it as it is if you want, I'll correct it later in
mine. Up to you.

My "fixup" looks like (in case of v2 needed or so):

diff --git a/drivers/net/ethernet/intel/igc/igc_main.c b/drivers/net/ethernet/intel/igc/igc_main.c
index b516f1b301b4..142c57b7a451 100644
--- a/drivers/net/ethernet/intel/igc/igc_main.c
+++ b/drivers/net/ethernet/intel/igc/igc_main.c

@@ -1726,7 +1726,7 @@ static struct sk_buff *igc_build_skb(struct igc_ring *rx_ring,
 	struct sk_buff *skb;
 
 	/* prefetch first cache line of first page */
-	net_prefetch(xdp->data);
+	net_prefetch(xdp->data_meta);
 
 	/* build an skb around the page buffer */
 	skb = build_skb(xdp->data_hard_start, truesize);

@@ -1756,10 +1756,11 @@ static struct sk_buff *igc_construct_skb(struct igc_ring *rx_ring,
 	struct sk_buff *skb;
 
 	/* prefetch first cache line of first page */
-	net_prefetch(va);
+	net_prefetch(xdp->data_meta);
 
 	/* allocate a skb to store the frags */
-	skb = napi_alloc_skb(&rx_ring->q_vector->napi, IGC_RX_HDR_LEN + metasize);
+	skb = napi_alloc_skb(&rx_ring->q_vector->napi,
+			     IGC_RX_HDR_LEN + metasize);
 	if (unlikely(!skb))
 		return NULL;
 

@@ -2363,7 +2364,8 @@ static int igc_clean_rx_irq(struct igc_q_vector *q_vector, const int budget)
 		if (!skb) {
 			xdp_init_buff(&xdp, truesize, &rx_ring->xdp_rxq);
 			xdp_prepare_buff(&xdp, pktbuf - igc_rx_offset(rx_ring),
-					 igc_rx_offset(rx_ring) + pkt_offset, size, true);
+					 igc_rx_offset(rx_ring) + pkt_offset,
+					 size, true);
 
 			skb = igc_xdp_run_prog(adapter, &xdp);
 		}

Reviewed-by: Alexander Lobakin <redacted>

quoted

quoted

quoted

  
  	/* build an skb around the page buffer */
-	skb = build_skb(va - IGC_SKB_PAD, truesize);
+	skb = build_skb(xdp->data_hard_start, truesize);
  	if (unlikely(!skb))
  		return NULL;
  
  	/* update pointers within the skb to store the data */
-	skb_reserve(skb, IGC_SKB_PAD);
+	skb_reserve(skb, xdp->data - xdp->data_hard_start);
  	__skb_put(skb, size);
+	if (metasize)
+		skb_metadata_set(skb, metasize);
  
  	igc_rx_buffer_flip(rx_buffer, truesize);
  	return skb;

@@ -1746,6 +1748,7 @@ static struct sk_buff *igc_construct_skb(struct igc_ring *rx_ring,
  					 struct xdp_buff *xdp,
  					 ktime_t timestamp)
  {
+	unsigned int metasize = xdp->data - xdp->data_meta;
  	unsigned int size = xdp->data_end - xdp->data;
  	unsigned int truesize = igc_get_rx_frame_truesize(rx_ring, size);
  	void *va = xdp->data;

@@ -1756,7 +1759,7 @@ static struct sk_buff *igc_construct_skb(struct igc_ring *rx_ring,
  	net_prefetch(va);

...here as well.

Thanks,
Al

Al

On Mon, 2021-11-29 at 19:13 +0100, Alexander Lobakin wrote:

From: Alexander Lobakin <redacted>
Date: Mon, 29 Nov 2021 15:53:03 +0100

quoted

From: Jesper Dangaard Brouer <redacted>
Date: Mon, 29 Nov 2021 15:39:04 +0100

quoted

On 26/11/2021 17.16, Alexander Lobakin wrote:

quoted

From: Jesper Dangaard Brouer <redacted>
Date: Mon, 15 Nov 2021 21:36:30 +0100

quoted

Enabling the XDP bpf_prog access to data_meta area is a very
small
change. Hint passing 'true' to xdp_prepare_buff().

[ snip ]

quoted

quoted

Prefetch works for "full" cachelines. Intel CPUs often prefect
two 
cache-lines, when doing this, thus I guess we still get xdp-

quoted

data.

Sure. I mean, net_prefetch() prefetches 128 bytes in a row.
xdp->data is usually aligned to XDP_PACKET_HEADROOM (or two bytes
to the right). If our CL is 64 and the meta is present, then... ah
right, 64 to the left and 64 starting from data to the right.

quoted

I don't mind prefetching xdp->data_meta, but (1) I tried to keep
the 
xdp->data starts on a cacheline and we know NIC hardware have
touched 
that, it is not a full-cache-miss due to DDIO/DCA it is known to
be in 
L3 cache (gain is around 2-3 ns in my machine for data prefetch).
Given this is only a 2.5 Gbit/s driver/HW I doubt this make any
difference.

Code constistency at least. On 10+ Gbps we prefetch meta, and I
plan
to continue doing this in my series.

quoted

Tony is it worth resending a V2 of this patch?

Tony, you can take it as it is if you want, I'll correct it later
in
mine. Up to you.

My "fixup" looks like (in case of v2 needed or so):

Thanks Al. If Jesper is ok with this, I'll incorporate it in before
sending the pull request to netdev. Otherwise, you can do it as follow
on in the other series you previously referenced.

Thanks,
Tony

On 29/11/2021 20.03, Nguyen, Anthony L wrote:

On Mon, 2021-11-29 at 19:13 +0100, Alexander Lobakin wrote:

quoted

From: Alexander Lobakin <redacted>
Date: Mon, 29 Nov 2021 15:53:03 +0100

quoted

From: Jesper Dangaard Brouer <redacted>
Date: Mon, 29 Nov 2021 15:39:04 +0100

quoted

On 26/11/2021 17.16, Alexander Lobakin wrote:

quoted

From: Jesper Dangaard Brouer <redacted>
Date: Mon, 15 Nov 2021 21:36:30 +0100

quoted

Enabling the XDP bpf_prog access to data_meta area is a very
small
change. Hint passing 'true' to xdp_prepare_buff().

[ snip ]

[ snip ]

quoted

quoted

quoted

Tony is it worth resending a V2 of this patch?

Tony, you can take it as it is if you want, I'll correct it later
in
mine. Up to you.

My "fixup" looks like (in case of v2 needed or so):

Thanks Al. If Jesper is ok with this, I'll incorporate it in before
sending the pull request to netdev. Otherwise, you can do it as follow
on in the other series you previously referenced.

I'm fine with you incorporating this change. Thanks! :-)
--Jesper