Skip to contentSkip to conversation

net: ethernet: mtk_eth_soc: improve RX performance

PATCH In review Elad Yifee 7 participants 16 messages Jul 29, 2024 – Jan 21, 2025

Subsystems (4)
lore
Share & tools
lore

Elad Yifee

author
#0 lore raw

v2 cover net: ethernet: mtk_eth_soc: improve RX performance

This small series includes two short and simple patches to improve RX performance on this driver.

iperf3 result without these patches:

	[ ID] Interval           Transfer     Bandwidth
	[  4]   0.00-1.00   sec   563 MBytes  4.72 Gbits/sec
	[  4]   1.00-2.00   sec   563 MBytes  4.73 Gbits/sec
	[  4]   2.00-3.00   sec   552 MBytes  4.63 Gbits/sec
	[  4]   3.00-4.00   sec   561 MBytes  4.70 Gbits/sec
	[  4]   4.00-5.00   sec   562 MBytes  4.71 Gbits/sec
	[  4]   5.00-6.00   sec   565 MBytes  4.74 Gbits/sec
	[  4]   6.00-7.00   sec   563 MBytes  4.72 Gbits/sec
	[  4]   7.00-8.00   sec   565 MBytes  4.74 Gbits/sec
	[  4]   8.00-9.00   sec   562 MBytes  4.71 Gbits/sec
	[  4]   9.00-10.00  sec   558 MBytes  4.68 Gbits/sec
	- - - - - - - - - - - - - - - - - - - - - - - - -
	[ ID] Interval           Transfer     Bandwidth
	[  4]   0.00-10.00  sec  5.48 GBytes  4.71 Gbits/sec                  sender
	[  4]   0.00-10.00  sec  5.48 GBytes  4.71 Gbits/sec                  receiver

iperf3 result with "use prefetch methods" patch:

	[ ID] Interval           Transfer     Bandwidth
	[  4]   0.00-1.00   sec   598 MBytes  5.02 Gbits/sec
	[  4]   1.00-2.00   sec   588 MBytes  4.94 Gbits/sec
	[  4]   2.00-3.00   sec   592 MBytes  4.97 Gbits/sec
	[  4]   3.00-4.00   sec   594 MBytes  4.98 Gbits/sec
	[  4]   4.00-5.00   sec   590 MBytes  4.95 Gbits/sec
	[  4]   5.00-6.00   sec   594 MBytes  4.98 Gbits/sec
	[  4]   6.00-7.00   sec   594 MBytes  4.98 Gbits/sec
	[  4]   7.00-8.00   sec   593 MBytes  4.98 Gbits/sec
	[  4]   8.00-9.00   sec   593 MBytes  4.98 Gbits/sec
	[  4]   9.00-10.00  sec   594 MBytes  4.98 Gbits/sec
	- - - - - - - - - - - - - - - - - - - - - - - - -
	[ ID] Interval           Transfer     Bandwidth
	[  4]   0.00-10.00  sec  5.79 GBytes  4.98 Gbits/sec                  sender
	[  4]   0.00-10.00  sec  5.79 GBytes  4.98 Gbits/sec                  receiver

iperf3 result with "use PP exclusively for XDP programs" patch:

	[ ID] Interval           Transfer     Bandwidth
	[  4]   0.00-1.00   sec   635 MBytes  5.33 Gbits/sec
	[  4]   1.00-2.00   sec   636 MBytes  5.33 Gbits/sec
	[  4]   2.00-3.00   sec   637 MBytes  5.34 Gbits/sec
	[  4]   3.00-4.00   sec   636 MBytes  5.34 Gbits/sec
	[  4]   4.00-5.00   sec   637 MBytes  5.34 Gbits/sec
	[  4]   5.00-6.00   sec   637 MBytes  5.35 Gbits/sec
	[  4]   6.00-7.00   sec   637 MBytes  5.34 Gbits/sec
	[  4]   7.00-8.00   sec   636 MBytes  5.33 Gbits/sec
	[  4]   8.00-9.00   sec   634 MBytes  5.32 Gbits/sec
	[  4]   9.00-10.00  sec   637 MBytes  5.34 Gbits/sec
	- - - - - - - - - - - - - - - - - - - - - - - - -
	[ ID] Interval           Transfer     Bandwidth
	[  4]   0.00-10.00  sec  6.21 GBytes  5.34 Gbits/sec                  sender
	[  4]   0.00-10.00  sec  6.21 GBytes  5.34 Gbits/sec                  receiver

iperf3 result with both patches:

	[ ID] Interval           Transfer     Bandwidth
	[  4]   0.00-1.00   sec   652 MBytes  5.47 Gbits/sec
	[  4]   1.00-2.00   sec   653 MBytes  5.47 Gbits/sec
	[  4]   2.00-3.00   sec   654 MBytes  5.48 Gbits/sec
	[  4]   3.00-4.00   sec   654 MBytes  5.49 Gbits/sec
	[  4]   4.00-5.00   sec   653 MBytes  5.48 Gbits/sec
	[  4]   5.00-6.00   sec   653 MBytes  5.48 Gbits/sec
	[  4]   6.00-7.00   sec   653 MBytes  5.48 Gbits/sec
	[  4]   7.00-8.00   sec   653 MBytes  5.48 Gbits/sec
	[  4]   8.00-9.00   sec   653 MBytes  5.48 Gbits/sec
	[  4]   9.00-10.00  sec   654 MBytes  5.48 Gbits/sec
	- - - - - - - - - - - - - - - - - - - - - - - - -
	[ ID] Interval           Transfer     Bandwidth
	[  4]   0.00-10.00  sec  6.38 GBytes  5.48 Gbits/sec                  sender
	[  4]   0.00-10.00  sec  6.38 GBytes  5.48 Gbits/sec                  receiver

About 16% more packets/sec without XDP program loaded, and about 5% more packets/sec when using PP. Tested on Banana Pi BPI-R4 (MT7988A)

---
Technically, this is version 2 of the “use prefetch methods” patch. Initially, I submitted it as a single patch for review (RFC),
but later I decided to include a second patch, resulting in this series Changes in v2:
- Add "use PP exclusively for XDP programs" patch and create this series
---
Elad Yifee (2):
net: ethernet: mtk_eth_soc: use prefetch methods
net: ethernet: mtk_eth_soc: use PP exclusively for XDP programs

 drivers/net/ethernet/mediatek/mtk_eth_soc.c | 10 ++++++++--
 1 file changed, 8 insertions(+), 2 deletions(-)
Signature · 1 line
2.45.2
author
#1 lore raw

PATCH v2 1/2 net: ethernet: mtk_eth_soc: use prefetch methods

Utilize kernel prefetch methods for faster cache line access. This change boosts driver performance,
allowing the CPU to handle about 5% more packets/sec.

---
Changes in v2:
- use net_prefetchw as suggested by Joe Damato
- add (NET_SKB_PAD + eth->ip_align) offset to prefetched data
- use eth->ip_align instead of NET_IP_ALIGN as it could be 0, depending on the platform

---
 drivers/net/ethernet/mediatek/mtk_eth_soc.c | 8 +++++++-
 1 file changed, 7 insertions(+), 1 deletion(-)
drivers/net/ethernet/mediatek/mtk_eth_soc.c +7 -1
index 16ca427cf4c3..4d0052dbe3f4 100644
@@ -1963,6 +1963,7 @@ static u32 mtk_xdp_run(struct mtk_eth *eth, struct mtk_rx_ring *ring,
 	if (!prog)
 		goto out;
 
+	net_prefetchw(xdp->data_hard_start);
 	act = bpf_prog_run_xdp(prog, xdp);
 	switch (act) {
 	case XDP_PASS:
@@ -2038,6 +2039,7 @@ static int mtk_poll_rx(struct napi_struct *napi, int budget,
 
 		idx = NEXT_DESP_IDX(ring->calc_idx, ring->dma_size);
 		rxd = ring->dma + idx * eth->soc->rx.desc_size;
+		prefetch(rxd);
 		data = ring->data[idx];
 
 		if (!mtk_rx_get_desc(eth, &trxd, rxd))
@@ -2105,6 +2107,7 @@ static int mtk_poll_rx(struct napi_struct *napi, int budget,
 			if (ret != XDP_PASS)
 				goto skip_rx;
 
+			net_prefetch(xdp.data_meta);
 			skb = build_skb(data, PAGE_SIZE);
 			if (unlikely(!skb)) {
 				page_pool_put_full_page(ring->page_pool,
@@ -2113,6 +2116,7 @@ static int mtk_poll_rx(struct napi_struct *napi, int budget,
 				goto skip_rx;
 			}
 
+			net_prefetchw(skb->data);
 			skb_reserve(skb, xdp.data - xdp.data_hard_start);
 			skb_put(skb, xdp.data_end - xdp.data);
 			skb_mark_for_recycle(skb);
@@ -2143,6 +2147,7 @@ static int mtk_poll_rx(struct napi_struct *napi, int budget,
 			dma_unmap_single(eth->dma_dev, ((u64)trxd.rxd1 | addr64),
 					 ring->buf_size, DMA_FROM_DEVICE);
 
+			net_prefetch(data + NET_SKB_PAD + eth->ip_align);
 			skb = build_skb(data, ring->frag_size);
 			if (unlikely(!skb)) {
 				netdev->stats.rx_dropped++;
@@ -2150,7 +2155,8 @@ static int mtk_poll_rx(struct napi_struct *napi, int budget,
 				goto skip_rx;
 			}
 
-			skb_reserve(skb, NET_SKB_PAD + NET_IP_ALIGN);
+			net_prefetchw(skb->data);
+			skb_reserve(skb, NET_SKB_PAD + eth->ip_align);
 			skb_put(skb, pktlen);
 		}
 
Signature · 1 line
2.45.2
#2 lore raw

Re: [PATCH net-next v2 1/2] net: ethernet: mtk_eth_soc: use prefetch methods

Quoted 19 lines · Elad Yifee

Utilize kernel prefetch methods for faster cache line access. This change boosts driver performance,
allowing the CPU to handle about 5% more packets/sec.

---
Changes in v2:
- use net_prefetchw as suggested by Joe Damato
- add (NET_SKB_PAD + eth->ip_align) offset to prefetched data
- use eth->ip_align instead of NET_IP_ALIGN as it could be 0, depending on the platform

---
 drivers/net/ethernet/mediatek/mtk_eth_soc.c | 8 +++++++-
 1 file changed, 7 insertions(+), 1 deletion(-)
drivers/net/ethernet/mediatek/mtk_eth_soc.c +0 -0
index 16ca427cf4c3..4d0052dbe3f4 100644

[...]

Quoted 9 lines
drivers/net/ethernet/mediatek/mtk_eth_soc.c +1 -0
@@ -2143,6 +2147,7 @@ static int mtk_poll_rx(struct napi_struct *napi, int budget,
 			dma_unmap_single(eth->dma_dev, ((u64)trxd.rxd1 | addr64),
 					 ring->buf_size, DMA_FROM_DEVICE);
 
+			net_prefetch(data + NET_SKB_PAD + eth->ip_align);
 			skb = build_skb(data, ring->frag_size);
 			if (unlikely(!skb)) {
 				netdev->stats.rx_dropped++;

drivers/net/ethernet/mediatek/mtk_eth_soc.c static int mtk_poll_rx(struct napi_struct *napi, int budget, · line 2159

Quoted 6 lines
drivers/net/ethernet/mediatek/mtk_eth_soc.c +2 -1
@@ -2150,7 +2155,8 @@ static int mtk_poll_rx(struct napi_struct *napi, int budget,
 				goto skip_rx;
 			}
 
-			skb_reserve(skb, NET_SKB_PAD + NET_IP_ALIGN);
+			net_prefetchw(skb->data);
+			skb_reserve(skb, NET_SKB_PAD + eth->ip_align);

Based on the code in mtk_probe, I am guessing that only MTK_SOC_MT7628 can DMA to unaligned addresses, because for everything else eth->ip_align would be 0.

Is that right?

I am asking because the documentation in Documentation/core-api/unaligned-memory-access.rst refers to the case you mention, NET_IP_ALIGN = 0, suggesting that this is intentional for performance reasons on powerpc:

One notable exception here is powerpc which defines NET_IP_ALIGN to 0 because DMA to unaligned addresses can be very expensive and dwarf the cost of unaligned loads.

It goes on to explain that some devices cannot DMA to unaligned addresses and I assume that for your driver that is everything which is not MTK_SOC_MT7628 ?

author
#3 lore raw

Re: [PATCH net-next v2 1/2] net: ethernet: mtk_eth_soc: use prefetch methods

Quoted 12 lines · Joe Damato

Based on the code in mtk_probe, I am guessing that only MTK_SOC_MT7628 can DMA to unaligned addresses, because for everything else eth->ip_align would be 0.

Is that right?

I am asking because the documentation in Documentation/core-api/unaligned-memory-access.rst refers to the case you mention, NET_IP_ALIGN = 0, suggesting that this is intentional for performance reasons on powerpc:

Quoted 7 lines

One notable exception here is powerpc which defines NET_IP_ALIGN to 0 because DMA to unaligned addresses can be very expensive and dwarf the cost of unaligned loads.

It goes on to explain that some devices cannot DMA to unaligned addresses and I assume that for your driver that is everything which is not MTK_SOC_MT7628 ?

I have no explanation for this partial use of 'eth->ip_align', it could be a mistake
or maybe I'm missing something.
Perhaps Stefan Roese, who wrote this part, has an explanation. (adding Stefan to CC)

#4 lore raw

Re: [PATCH net-next v2 1/2] net: ethernet: mtk_eth_soc: use prefetch methods

Quoted 21 lines · Elad Yifee
Quoted 12 lines · Joe Damato

Based on the code in mtk_probe, I am guessing that only MTK_SOC_MT7628 can DMA to unaligned addresses, because for everything else eth->ip_align would be 0.

Is that right?

I am asking because the documentation in Documentation/core-api/unaligned-memory-access.rst refers to the case you mention, NET_IP_ALIGN = 0, suggesting that this is intentional for performance reasons on powerpc:

Quoted 7 lines

One notable exception here is powerpc which defines NET_IP_ALIGN to 0 because DMA to unaligned addresses can be very expensive and dwarf the cost of unaligned loads.

It goes on to explain that some devices cannot DMA to unaligned addresses and I assume that for your driver that is everything which is not MTK_SOC_MT7628 ?

Quoted 5 lines

I have no explanation for this partial use of 'eth->ip_align', it could be a mistake
or maybe I'm missing something.
Perhaps Stefan Roese, who wrote this part, has an explanation. (adding Stefan to CC)

Sorry, I can't answer this w/o digging deeper into this driver and SoC again. And I didn't use it for a few years now. It might be a mistake.

Thanks,
Stefan

#5 lore raw

Re: [PATCH net-next v2 1/2] net: ethernet: mtk_eth_soc: use prefetch methods

Quoted 28 lines · Stefan Roese
Quoted 21 lines · Elad Yifee
Quoted 12 lines · Joe Damato

Based on the code in mtk_probe, I am guessing that only MTK_SOC_MT7628 can DMA to unaligned addresses, because for everything else eth->ip_align would be 0.

Is that right?

I am asking because the documentation in Documentation/core-api/unaligned-memory-access.rst refers to the case you mention, NET_IP_ALIGN = 0, suggesting that this is intentional for performance reasons on powerpc:

Quoted 7 lines

One notable exception here is powerpc which defines NET_IP_ALIGN to 0 because DMA to unaligned addresses can be very expensive and dwarf the cost of unaligned loads.

It goes on to explain that some devices cannot DMA to unaligned addresses and I assume that for your driver that is everything which is not MTK_SOC_MT7628 ?

Quoted 5 lines

I have no explanation for this partial use of 'eth->ip_align', it could be a mistake
or maybe I'm missing something.
Perhaps Stefan Roese, who wrote this part, has an explanation. (adding Stefan to CC)

Quoted 3 lines

Sorry, I can't answer this w/o digging deeper into this driver and SoC again. And I didn't use it for a few years now. It might be a mistake.

I asked about it because it was added in v2 of the patch, see the changelog from the patch:

- use eth->ip_align instead of NET_IP_ALIGN as it could be 0, depending on the platform

It seemed like from the changelog some one decided adding that made sense and I was just confirming the reasoning above.

#6 lore raw

Re: [PATCH net-next v2 1/2] net: ethernet: mtk_eth_soc: use prefetch methods

In reply to Elad Yifee

Hello,

Sorry to bother, but what happened to this patch? Is it given up or something?

Best regards,
Shengyu

Quoted 69 lines · Elad Yifee

Utilize kernel prefetch methods for faster cache line access. This change boosts driver performance,
allowing the CPU to handle about 5% more packets/sec.

---
Changes in v2:
- use net_prefetchw as suggested by Joe Damato
- add (NET_SKB_PAD + eth->ip_align) offset to prefetched data
- use eth->ip_align instead of NET_IP_ALIGN as it could be 0, depending on the platform

---
  drivers/net/ethernet/mediatek/mtk_eth_soc.c | 8 +++++++-
  1 file changed, 7 insertions(+), 1 deletion(-)
drivers/net/ethernet/mediatek/mtk_eth_soc.c +7 -1
index 16ca427cf4c3..4d0052dbe3f4 100644
@@ -1963,6 +1963,7 @@ static u32 mtk_xdp_run(struct mtk_eth *eth, struct mtk_rx_ring *ring,
  	if (!prog)
  		goto out;
 
+	net_prefetchw(xdp->data_hard_start);
  	act = bpf_prog_run_xdp(prog, xdp);
  	switch (act) {
  	case XDP_PASS:
@@ -2038,6 +2039,7 @@ static int mtk_poll_rx(struct napi_struct *napi, int budget,
 
  		idx = NEXT_DESP_IDX(ring->calc_idx, ring->dma_size);
  		rxd = ring->dma + idx * eth->soc->rx.desc_size;
+		prefetch(rxd);
  		data = ring->data[idx];
 
  		if (!mtk_rx_get_desc(eth, &trxd, rxd))
@@ -2105,6 +2107,7 @@ static int mtk_poll_rx(struct napi_struct *napi, int budget,
  			if (ret != XDP_PASS)
  				goto skip_rx;
 
+			net_prefetch(xdp.data_meta);
  			skb = build_skb(data, PAGE_SIZE);
  			if (unlikely(!skb)) {
  				page_pool_put_full_page(ring->page_pool,
@@ -2113,6 +2116,7 @@ static int mtk_poll_rx(struct napi_struct *napi, int budget,
  				goto skip_rx;
  			}
 
+			net_prefetchw(skb->data);
  			skb_reserve(skb, xdp.data - xdp.data_hard_start);
  			skb_put(skb, xdp.data_end - xdp.data);
  			skb_mark_for_recycle(skb);
@@ -2143,6 +2147,7 @@ static int mtk_poll_rx(struct napi_struct *napi, int budget,
  			dma_unmap_single(eth->dma_dev, ((u64)trxd.rxd1 | addr64),
  					 ring->buf_size, DMA_FROM_DEVICE);
 
+			net_prefetch(data + NET_SKB_PAD + eth->ip_align);
  			skb = build_skb(data, ring->frag_size);
  			if (unlikely(!skb)) {
  				netdev->stats.rx_dropped++;
@@ -2150,7 +2155,8 @@ static int mtk_poll_rx(struct napi_struct *napi, int budget,
  				goto skip_rx;
  			}
 
-			skb_reserve(skb, NET_SKB_PAD + NET_IP_ALIGN);
+			net_prefetchw(skb->data);
+			skb_reserve(skb, NET_SKB_PAD + eth->ip_align);
  			skb_put(skb, pktlen);
  		}
 
#7 lore raw

Re: [PATCH net-next v2 1/2] net: ethernet: mtk_eth_soc: use prefetch methods

Quoted 4 lines · Shengyu Qu

Hello,

Sorry to bother, but what happened to this patch? Is it given up or something?

There appear to be open questions about it. Those questions need answering. Also, the net-next is closed at the moment for the merge window. So a new version will need to be submitted once net-next reopens in two weeks time.

Andrew

---
pw-bot: cr

author
#8 lore raw

PATCH v2 2/2 net: ethernet: mtk_eth_soc: use PP exclusively for XDP programs

In reply to Elad Yifee

PP allocations and XDP code path traversal are unnecessary when no XDP program is loaded.
Prevent that by simply not creating the pool. This change boosts driver performance for this use case, allowing the CPU to handle about 13% more packets/sec.

---
 drivers/net/ethernet/mediatek/mtk_eth_soc.c | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)
drivers/net/ethernet/mediatek/mtk_eth_soc.c +1 -1
index 4d0052dbe3f4..2d1a48287c73 100644
@@ -2644,7 +2644,7 @@ static int mtk_rx_alloc(struct mtk_eth *eth, int ring_no, int rx_flag)
 	if (!ring->data)
 		return -ENOMEM;
 
-	if (mtk_page_pool_enabled(eth)) {
+	if (mtk_page_pool_enabled(eth) && rcu_access_pointer(eth->prog)) {
 		struct page_pool *pp;
 
 		pp = mtk_create_page_pool(eth, &ring->xdp_q, ring_no,
Signature · 1 line
2.45.2
maintainer
#9 lore raw
In reply to Elad Yifee
Quoted 2 lines

This small series includes two short and simple patches to improve RX performance on this driver.

Hi Elad,

What is the chip revision you are running?
If you are using a device that does not support HW-LRO (e.g. MT7986 or MT7988), I guess we can try to use page_pool_dev_alloc_frag() APIs and request a 2048B buffer. Doing so, we can use use a single page for two rx buffers improving recycling with page_pool. What do you think?

Regards,
Lorenzo

Quoted 90 lines

iperf3 result without these patches:

	[ ID] Interval           Transfer     Bandwidth
	[  4]   0.00-1.00   sec   563 MBytes  4.72 Gbits/sec
	[  4]   1.00-2.00   sec   563 MBytes  4.73 Gbits/sec
	[  4]   2.00-3.00   sec   552 MBytes  4.63 Gbits/sec
	[  4]   3.00-4.00   sec   561 MBytes  4.70 Gbits/sec
	[  4]   4.00-5.00   sec   562 MBytes  4.71 Gbits/sec
	[  4]   5.00-6.00   sec   565 MBytes  4.74 Gbits/sec
	[  4]   6.00-7.00   sec   563 MBytes  4.72 Gbits/sec
	[  4]   7.00-8.00   sec   565 MBytes  4.74 Gbits/sec
	[  4]   8.00-9.00   sec   562 MBytes  4.71 Gbits/sec
	[  4]   9.00-10.00  sec   558 MBytes  4.68 Gbits/sec
	- - - - - - - - - - - - - - - - - - - - - - - - -
	[ ID] Interval           Transfer     Bandwidth
	[  4]   0.00-10.00  sec  5.48 GBytes  4.71 Gbits/sec                  sender
	[  4]   0.00-10.00  sec  5.48 GBytes  4.71 Gbits/sec                  receiver

iperf3 result with "use prefetch methods" patch:

	[ ID] Interval           Transfer     Bandwidth
	[  4]   0.00-1.00   sec   598 MBytes  5.02 Gbits/sec
	[  4]   1.00-2.00   sec   588 MBytes  4.94 Gbits/sec
	[  4]   2.00-3.00   sec   592 MBytes  4.97 Gbits/sec
	[  4]   3.00-4.00   sec   594 MBytes  4.98 Gbits/sec
	[  4]   4.00-5.00   sec   590 MBytes  4.95 Gbits/sec
	[  4]   5.00-6.00   sec   594 MBytes  4.98 Gbits/sec
	[  4]   6.00-7.00   sec   594 MBytes  4.98 Gbits/sec
	[  4]   7.00-8.00   sec   593 MBytes  4.98 Gbits/sec
	[  4]   8.00-9.00   sec   593 MBytes  4.98 Gbits/sec
	[  4]   9.00-10.00  sec   594 MBytes  4.98 Gbits/sec
	- - - - - - - - - - - - - - - - - - - - - - - - -
	[ ID] Interval           Transfer     Bandwidth
	[  4]   0.00-10.00  sec  5.79 GBytes  4.98 Gbits/sec                  sender
	[  4]   0.00-10.00  sec  5.79 GBytes  4.98 Gbits/sec                  receiver

iperf3 result with "use PP exclusively for XDP programs" patch:

	[ ID] Interval           Transfer     Bandwidth
	[  4]   0.00-1.00   sec   635 MBytes  5.33 Gbits/sec
	[  4]   1.00-2.00   sec   636 MBytes  5.33 Gbits/sec
	[  4]   2.00-3.00   sec   637 MBytes  5.34 Gbits/sec
	[  4]   3.00-4.00   sec   636 MBytes  5.34 Gbits/sec
	[  4]   4.00-5.00   sec   637 MBytes  5.34 Gbits/sec
	[  4]   5.00-6.00   sec   637 MBytes  5.35 Gbits/sec
	[  4]   6.00-7.00   sec   637 MBytes  5.34 Gbits/sec
	[  4]   7.00-8.00   sec   636 MBytes  5.33 Gbits/sec
	[  4]   8.00-9.00   sec   634 MBytes  5.32 Gbits/sec
	[  4]   9.00-10.00  sec   637 MBytes  5.34 Gbits/sec
	- - - - - - - - - - - - - - - - - - - - - - - - -
	[ ID] Interval           Transfer     Bandwidth
	[  4]   0.00-10.00  sec  6.21 GBytes  5.34 Gbits/sec                  sender
	[  4]   0.00-10.00  sec  6.21 GBytes  5.34 Gbits/sec                  receiver

iperf3 result with both patches:

	[ ID] Interval           Transfer     Bandwidth
	[  4]   0.00-1.00   sec   652 MBytes  5.47 Gbits/sec
	[  4]   1.00-2.00   sec   653 MBytes  5.47 Gbits/sec
	[  4]   2.00-3.00   sec   654 MBytes  5.48 Gbits/sec
	[  4]   3.00-4.00   sec   654 MBytes  5.49 Gbits/sec
	[  4]   4.00-5.00   sec   653 MBytes  5.48 Gbits/sec
	[  4]   5.00-6.00   sec   653 MBytes  5.48 Gbits/sec
	[  4]   6.00-7.00   sec   653 MBytes  5.48 Gbits/sec
	[  4]   7.00-8.00   sec   653 MBytes  5.48 Gbits/sec
	[  4]   8.00-9.00   sec   653 MBytes  5.48 Gbits/sec
	[  4]   9.00-10.00  sec   654 MBytes  5.48 Gbits/sec
	- - - - - - - - - - - - - - - - - - - - - - - - -
	[ ID] Interval           Transfer     Bandwidth
	[  4]   0.00-10.00  sec  6.38 GBytes  5.48 Gbits/sec                  sender
	[  4]   0.00-10.00  sec  6.38 GBytes  5.48 Gbits/sec                  receiver

About 16% more packets/sec without XDP program loaded, and about 5% more packets/sec when using PP. Tested on Banana Pi BPI-R4 (MT7988A)

---
Technically, this is version 2 of the “use prefetch methods” patch. Initially, I submitted it as a single patch for review (RFC),
but later I decided to include a second patch, resulting in this series Changes in v2:
- Add "use PP exclusively for XDP programs" patch and create this series
---
Elad Yifee (2):
net: ethernet: mtk_eth_soc: use prefetch methods
net: ethernet: mtk_eth_soc: use PP exclusively for XDP programs

 drivers/net/ethernet/mediatek/mtk_eth_soc.c | 10 ++++++++--
 1 file changed, 8 insertions(+), 2 deletions(-)
Signature · 1 line
2.45.2
author
#10 lore raw
Quoted 6 lines · Lorenzo Bianconi
Quoted 3 lines

This small series includes two short and simple patches to improve RX performance on this driver.

Hi Elad,

Quoted 9 lines

What is the chip revision you are running?
If you are using a device that does not support HW-LRO (e.g. MT7986 or MT7988), I guess we can try to use page_pool_dev_alloc_frag() APIs and request a 2048B buffer. Doing so, we can use use a single page for two rx buffers improving recycling with page_pool. What do you think?

Regards,
Lorenzo

Hey Lorenzo,
It's Rev0. why, do you have any info on the revisions?
Since it's probably the reason for the performance hit, allocating full pages every time, I think your suggestion would improve the performance and probably match it with the napi_alloc_frag path. I'll give it a try when I have time.
You also mentioned HW-LRO, which makes me think we also need the second patch if we want to allow HW-LRO to co-exists with XDP on NETSYS2/3 devices.

maintainer
#11 lore raw
Quoted 4 lines · On Tue, 30 Jul 2024 08:29:58 +0300 Elad Yifee wrote:

Since it's probably the reason for the performance hit, allocating full pages every time, I think your suggestion would improve the performance and probably match it with the napi_alloc_frag path. I'll give it a try when I have time.

This is a better direction than disabling PP.
Feel free to repost patch 1 separately.

Signature · 1 line
pw-bot: cr
author
#12 lore raw
Quoted 11 lines · Jakub Kicinski
Quoted 4 lines · On Tue, 30 Jul 2024 08:29:58 +0300 Elad Yifee wrote:

Since it's probably the reason for the performance hit, allocating full pages every time, I think your suggestion would improve the performance and probably match it with the napi_alloc_frag path. I'll give it a try when I have time.

This is a better direction than disabling PP.
Feel free to repost patch 1 separately.
--
pw-bot: cr

In this driver, the existence of PP is the condition to execute all XDP-related operations which aren't necessary
on this hot path, so we anyway wouldn't want that. on XDP program setup the rings are reallocated and the PP
would be created.
Other than that, for HWLRO we need contiguous pages of different order than the PP, so the creation of PP
basically prevents the use of HWLRO.
So we solve this LRO problem and get a performance boost with this simple change.

Lorenzo's suggestion would probably improve the performance of the XDP path and we should try that nonetheless.

maintainer
#13 lore raw
Quoted 12 lines · Jakub Kicinski
Quoted 11 lines · On Tue, 30 Jul 2024 08:29:58 +0300 Elad Yifee wrote:
Quoted 4 lines

Since it's probably the reason for the performance hit, allocating full pages every time, I think your suggestion would improve the performance and probably match it with the napi_alloc_frag path. I'll give it a try when I have time.

This is a better direction than disabling PP.
Feel free to repost patch 1 separately.
--
pw-bot: cr

Quoted 5 lines

In this driver, the existence of PP is the condition to execute all XDP-related operations which aren't necessary
on this hot path, so we anyway wouldn't want that. on XDP program setup the rings are reallocated and the PP
would be created.

nope, I added page_pool support even for non-XDP mode for hw that does not support HW-LRO. I guess mtk folks can correct me if I am wrong but IIRC there were some hw limirations on mt7986/mt7988 for HW-LRO, so I am not sure if it can be supported.

Quoted 8 lines

Other than that, for HWLRO we need contiguous pages of different order than the PP, so the creation of PP
basically prevents the use of HWLRO.
So we solve this LRO problem and get a performance boost with this simple change.

Lorenzo's suggestion would probably improve the performance of the XDP path and we should try that nonetheless.

nope, I mean to improve peformances even for non-XDP case with page_pool frag APIs.

Regards,
Lorenzo

author
#14 lore raw
Quoted 5 lines · Lorenzo Bianconi

nope, I added page_pool support even for non-XDP mode for hw that does not support HW-LRO. I guess mtk folks can correct me if I am wrong but IIRC there were some hw limirations on mt7986/mt7988 for HW-LRO, so I am not sure if it can be supported.

I know, but if we want to add support for HWLRO alongside XDP on NETSYS2/3, we need to prevent the PP use (for HWLRO allocations) and enable it only when there's
an XDP program.
I've been told HWLRO works on the MTK SDK version.

Quoted 6 lines
Quoted 6 lines

Other than that, for HWLRO we need contiguous pages of different order than the PP, so the creation of PP
basically prevents the use of HWLRO.
So we solve this LRO problem and get a performance boost with this simple change.

Quoted 8 lines
Quoted 2 lines

Lorenzo's suggestion would probably improve the performance of the XDP path and we should try that nonetheless.

nope, I mean to improve peformances even for non-XDP case with page_pool frag APIs.

Regards,
Lorenzo

Yes of course it would improve it for non-XDP case if we still use PP for non-XDP,
but my point is we shouldn't, mainly because of HWLRO, but also the extra unnecessary code.

maintainer
#15 lore raw
Quoted 11 lines · Lorenzo Bianconi
Quoted 5 lines

nope, I added page_pool support even for non-XDP mode for hw that does not support HW-LRO. I guess mtk folks can correct me if I am wrong but IIRC there were some hw limirations on mt7986/mt7988 for HW-LRO, so I am not sure if it can be supported.

I know, but if we want to add support for HWLRO alongside XDP on NETSYS2/3, we need to prevent the PP use (for HWLRO allocations) and enable it only when there's
an XDP program.
I've been told HWLRO works on the MTK SDK version.

ack, but in this case, please provide even the HW-LRO support in the same series. Moreover, I am not sure if it is performant enough or not, we could increase the page_pool order.
Moreover I guess we should be sure the HW-LRO works on all NETSYS2/3 hws revisions.

Regards,
Lorenzo

Quoted 19 lines
Quoted 7 lines
Quoted 7 lines

Other than that, for HWLRO we need contiguous pages of different order than the PP, so the creation of PP
basically prevents the use of HWLRO.
So we solve this LRO problem and get a performance boost with this simple change.

Quoted 8 lines
Quoted 2 lines

Lorenzo's suggestion would probably improve the performance of the XDP path and we should try that nonetheless.

nope, I mean to improve peformances even for non-XDP case with page_pool frag APIs.

Regards,
Lorenzo

Yes of course it would improve it for non-XDP case if we still use PP for non-XDP,
but my point is we shouldn't, mainly because of HWLRO, but also the extra unnecessary code.