Skip to contentSkip to conversation

stmmac: align RX buffers

PATCH Applied Matteo Croce 8 participants 37 messages Jun 14, 2021 – Aug 20, 2021

Subsystems (5)
lore
Share & tools
lore
#0 lore raw

PATCH 1/1 stmmac: align RX buffers

From: Matteo Croce <[email protected]>

On RX an SKB is allocated and the received buffer is copied into it. But on some architectures, the memcpy() needs the source and destination buffers to have the same alignment to be efficient.

This is not our case, because SKB data pointer is misaligned by two bytes to compensate the ethernet header.

Align the RX buffer the same way as the SKB one, so the copy is faster. An iperf3 RX test gives a decent improvement on a RISC-V machine:

before:
[ ID] Interval Transfer Bitrate Retr

[  5]   0.00-10.00  sec   733 MBytes   615 Mbits/sec   88             sender
[  5]   0.00-10.01  sec   730 MBytes   612 Mbits/sec                  receiver

after:

[ ID] Interval           Transfer     Bitrate         Retr
[  5]   0.00-10.00  sec  1.10 GBytes   942 Mbits/sec    0             sender
[  5]   0.00-10.00  sec  1.09 GBytes   940 Mbits/sec                  receiver

And the memcpy() overhead during the RX drops dramatically.

before:

Overhead  Shared O  Symbol
  43.35%  [kernel]  [k] memcpy
  33.77%  [kernel]  [k] __asm_copy_to_user
   3.64%  [kernel]  [k] sifive_l2_flush64_range

after:

Overhead  Shared O  Symbol
  45.40%  [kernel]  [k] __asm_copy_to_user
  28.09%  [kernel]  [k] memcpy
   4.27%  [kernel]  [k] sifive_l2_flush64_range
---
 drivers/net/ethernet/stmicro/stmmac/stmmac.h | 4 ++--
 1 file changed, 2 insertions(+), 2 deletions(-)
drivers/net/ethernet/stmicro/stmmac/stmmac.h +2 -2
index b6cd43eda7ac..04bdb3950d63 100644
@@ -338,9 +338,9 @@ static inline bool stmmac_xdp_is_enabled(struct stmmac_priv *priv)
 static inline unsigned int stmmac_rx_offset(struct stmmac_priv *priv)
 {
 	if (stmmac_xdp_is_enabled(priv))
-		return XDP_PACKET_HEADROOM;
+		return XDP_PACKET_HEADROOM + NET_IP_ALIGN;
 
-	return 0;
+	return NET_SKB_PAD + NET_IP_ALIGN;
 }
 
 void stmmac_disable_rx_queue(struct stmmac_priv *priv, u32 queue);
Signature · 7 lines
2.31.1


_______________________________________________
linux-riscv mailing list
[email protected]
http://lists.infradead.org/mailman/listinfo/linux-riscv
maintainer
#1 lore raw

But thois means the ethernet header will be misaliugned and this will kill performance on some cpus as misaligned accessed are resolved wioth a trap handler.

Even on cpus that don't trap, the access will be slower.

Thanks.

#2 lore raw
Quoted 8 lines · On Mon, 14 Jun 2021 12:51:11 -0700 (PDT) David Miller <[email protected]>…

But thois means the ethernet header will be misaliugned and this will kill performance on some cpus as misaligned accessed are resolved wioth a trap handler.

Even on cpus that don't trap, the access will be slower.

Thanks.

Isn't the IP header which should be aligned to avoid expensive traps? From include/linux/skbuff.h:

 * Since an ethernet header is 14 bytes network drivers often end up with
 * the IP header at an unaligned offset. The IP header can be aligned by
 * shifting the start of the packet by 2 bytes. Drivers should do this
 * with:
 *
 * skb_reserve(skb, NET_IP_ALIGN);

But the problem here really is not the header alignment, the problem is that the rx buffer is copied into an skb, and the two buffers have different alignments.
If I add this print, I get this for every packet:

--- a/drivers/net/ethernet/stmicro/stmmac/stmmac_main.c
@@ -5460,6 +5460,8 @@ static int stmmac_rx(struct stmmac_priv *priv, int limit, u32 queue)
+ printk("skb->data alignment: %lu\n", (uintptr_t)skb->data & 7);

+               printk("xdp.data alignment: %lu\n" , (uintptr_t)xdp.data & 7);
                skb_copy_to_linear_data(skb, xdp.data, buf1_len);
[ 1060.967768] skb->data alignment: 2
[ 1060.971174] xdp.data alignment: 0
[ 1061.967589] skb->data alignment: 2
[ 1061.970994] xdp.data alignment: 0

And many architectures do an optimized memcpy when the low order bits of the two pointers match, to name a few:

arch/alpha/lib/memcpy.c:

	/* If both source and dest are word aligned copy words */
	if (!((unsigned int)dest_w & 3) && !((unsigned int)src_w & 3)) {

arch/xtensa/lib/memcopy.S:

	/*
	 * Destination and source are word-aligned, use word copy.
	 */
	# copy 16 bytes per iteration for word-aligned dst and word-aligned src

arch/openrisc/lib/memcpy.c:

	/* If both source and dest are word aligned copy words */
	if (!((unsigned int)dest_w & 3) && !((unsigned int)src_w & 3)) {

And so on. With my patch I (mis)align the two buffer at an offset 2 (NET_IP_ALIGN) so the data can be copied faster:

[   16.648485] skb->data alignment: 2
[   16.651894] xdp.data alignment: 2
[   16.714260] skb->data alignment: 2
[   16.717688] xdp.data alignment: 2

Does this make sense?

Regards,

Signature · 1 line
per aspera ad upstream
maintainer
#3 lore raw

Thank you for the explanation, I have appplied this patch.

#4 lore raw
In reply to Matteo Croce
This patch was applied to netdev/net-next.git (refs/heads/master):

Hello:

This patch was applied to netdev/net-next.git (refs/heads/master):

Quoted 10 lines · On Mon, 14 Jun 2021 04:25:04 +0200 you wrote:

From: Matteo Croce <[email protected]>

On RX an SKB is allocated and the received buffer is copied into it. But on some architectures, the memcpy() needs the source and destination buffers to have the same alignment to be efficient.

This is not our case, because SKB data pointer is misaligned by two bytes to compensate the ethernet header.

[...]

Here is the summary with links:
- [net-next] stmmac: align RX buffers

    https://git.kernel.org/netdev/net-next/c/a955318fe67e

You are awesome, thank you!
--
Deet-doot-dot, I am a bot.
https://korg.docs.kernel.org/patchwork/pwbot.html

#5 lore raw
In reply to Matteo Croce

Hi all,

[adding Thierry, Jon and Will to the fun]

Quoted 31 lines · Matteo Croce

From: Matteo Croce <[email protected]>

On RX an SKB is allocated and the received buffer is copied into it. But on some architectures, the memcpy() needs the source and destination buffers to have the same alignment to be efficient.

This is not our case, because SKB data pointer is misaligned by two bytes to compensate the ethernet header.

Align the RX buffer the same way as the SKB one, so the copy is faster. An iperf3 RX test gives a decent improvement on a RISC-V machine:

before:
[ ID] Interval Transfer Bitrate Retr

[  5]   0.00-10.00  sec   733 MBytes   615 Mbits/sec   88             sender
[  5]   0.00-10.01  sec   730 MBytes   612 Mbits/sec                  receiver

after:

[ ID] Interval           Transfer     Bitrate         Retr
[  5]   0.00-10.00  sec  1.10 GBytes   942 Mbits/sec    0             sender
[  5]   0.00-10.00  sec  1.09 GBytes   940 Mbits/sec                  receiver

And the memcpy() overhead during the RX drops dramatically.

before:

Overhead  Shared O  Symbol
  43.35%  [kernel]  [k] memcpy
  33.77%  [kernel]  [k] __asm_copy_to_user
   3.64%  [kernel]  [k] sifive_l2_flush64_range
Quoted 7 lines

after:

Overhead  Shared O  Symbol
  45.40%  [kernel]  [k] __asm_copy_to_user
  28.09%  [kernel]  [k] memcpy
   4.27%  [kernel]  [k] sifive_l2_flush64_range

This patch completely breaks my Jetson TX2 system, composed of 2 Nvidia Denver and 4 Cortex-A57, in a very "funny" way.

Any significant amount of traffic result in all sort of corruption (ssh connections get dropped, Debian packages downloaded have the wrong checksums) if any Denver core is involved in any significant way (packet processing, interrupt handling). And it is all triggered by this very change.

The only way I have to make it work on a Denver core is to route the interrupt to that particular core and taskset the workload to it. Any other configuration involving a Denver CPU results in some sort of corruption. On their own, the A57s are fine.

This smells of memory ordering going really wrong, which this change would expose. I haven't had a chance to dig into the driver yet (it took me long enough to bisect it), but if someone points me at what is supposed to synchronise the DMA when receiving an interrupt, I'll have a look.

Thanks,

	M.
Quoted 24 lines
---
 drivers/net/ethernet/stmicro/stmmac/stmmac.h | 4 ++--
 1 file changed, 2 insertions(+), 2 deletions(-)
drivers/net/ethernet/stmicro/stmmac/stmmac.h +2 -2
index b6cd43eda7ac..04bdb3950d63 100644
@@ -338,9 +338,9 @@ static inline bool stmmac_xdp_is_enabled(struct stmmac_priv *priv)
 static inline unsigned int stmmac_rx_offset(struct stmmac_priv *priv)
 {
 	if (stmmac_xdp_is_enabled(priv))
-		return XDP_PACKET_HEADROOM;
+		return XDP_PACKET_HEADROOM + NET_IP_ALIGN;
 
-	return 0;
+	return NET_SKB_PAD + NET_IP_ALIGN;
 }
 
 void stmmac_disable_rx_queue(struct stmmac_priv *priv, u32 queue);
Signature · 1 line
2.31.1
Signature · 1 line
Without deviation from the norm, progress is not possible.
#6 lore raw
Quoted 59 lines · Marc Zyngier

Hi all,

[adding Thierry, Jon and Will to the fun]

Quoted 31 lines · Matteo Croce

From: Matteo Croce <[email protected]>

On RX an SKB is allocated and the received buffer is copied into it. But on some architectures, the memcpy() needs the source and destination buffers to have the same alignment to be efficient.

This is not our case, because SKB data pointer is misaligned by two bytes to compensate the ethernet header.

Align the RX buffer the same way as the SKB one, so the copy is faster. An iperf3 RX test gives a decent improvement on a RISC-V machine:

before:
[ ID] Interval Transfer Bitrate Retr

[  5]   0.00-10.00  sec   733 MBytes   615 Mbits/sec   88             sender
[  5]   0.00-10.01  sec   730 MBytes   612 Mbits/sec                  receiver

after:

[ ID] Interval           Transfer     Bitrate         Retr
[  5]   0.00-10.00  sec  1.10 GBytes   942 Mbits/sec    0             sender
[  5]   0.00-10.00  sec  1.09 GBytes   940 Mbits/sec                  receiver

And the memcpy() overhead during the RX drops dramatically.

before:

Overhead  Shared O  Symbol
  43.35%  [kernel]  [k] memcpy
  33.77%  [kernel]  [k] __asm_copy_to_user
   3.64%  [kernel]  [k] sifive_l2_flush64_range
Quoted 7 lines

after:

Overhead  Shared O  Symbol
  45.40%  [kernel]  [k] __asm_copy_to_user
  28.09%  [kernel]  [k] memcpy
   4.27%  [kernel]  [k] sifive_l2_flush64_range

This patch completely breaks my Jetson TX2 system, composed of 2 Nvidia Denver and 4 Cortex-A57, in a very "funny" way.

Any significant amount of traffic result in all sort of corruption (ssh connections get dropped, Debian packages downloaded have the wrong checksums) if any Denver core is involved in any significant way (packet processing, interrupt handling). And it is all triggered by this very change.

The only way I have to make it work on a Denver core is to route the interrupt to that particular core and taskset the workload to it. Any other configuration involving a Denver CPU results in some sort of corruption. On their own, the A57s are fine.

Quoted 5 lines

This smells of memory ordering going really wrong, which this change would expose. I haven't had a chance to dig into the driver yet (it took me long enough to bisect it), but if someone points me at what is supposed to synchronise the DMA when receiving an interrupt, I'll have a look.

I recall that Jon was looking into a similar issue recently, though I think the failure mode was slightly different. I also vaguely recall that CPU frequency was impacting this to some degree (lower CPU frequencies would increase the chances of this happening).

Jon's currently out of office, but let me try and dig up the details on this.

Thierry

Quoted 32 lines

Thanks,

	M.
Quoted 24 lines
---
 drivers/net/ethernet/stmicro/stmmac/stmmac.h | 4 ++--
 1 file changed, 2 insertions(+), 2 deletions(-)
drivers/net/ethernet/stmicro/stmmac/stmmac.h +2 -2
index b6cd43eda7ac..04bdb3950d63 100644
@@ -338,9 +338,9 @@ static inline bool stmmac_xdp_is_enabled(struct stmmac_priv *priv)
 static inline unsigned int stmmac_rx_offset(struct stmmac_priv *priv)
 {
 	if (stmmac_xdp_is_enabled(priv))
-		return XDP_PACKET_HEADROOM;
+		return XDP_PACKET_HEADROOM + NET_IP_ALIGN;
 
-	return 0;
+	return NET_SKB_PAD + NET_IP_ALIGN;
 }
 
 void stmmac_disable_rx_queue(struct stmmac_priv *priv, u32 queue);
Signature · 1 line
2.31.1
Signature · 1 line
Without deviation from the norm, progress is not possible.
#31 lore raw
In reply to Marc Zyngier
Quoted 59 lines · Marc Zyngier

Hi all,

[adding Thierry, Jon and Will to the fun]

Quoted 31 lines · Matteo Croce

From: Matteo Croce <[email protected]>

On RX an SKB is allocated and the received buffer is copied into it. But on some architectures, the memcpy() needs the source and destination buffers to have the same alignment to be efficient.

This is not our case, because SKB data pointer is misaligned by two bytes to compensate the ethernet header.

Align the RX buffer the same way as the SKB one, so the copy is faster. An iperf3 RX test gives a decent improvement on a RISC-V machine:

before:
[ ID] Interval Transfer Bitrate Retr

[  5]   0.00-10.00  sec   733 MBytes   615 Mbits/sec   88             sender
[  5]   0.00-10.01  sec   730 MBytes   612 Mbits/sec                  receiver

after:

[ ID] Interval           Transfer     Bitrate         Retr
[  5]   0.00-10.00  sec  1.10 GBytes   942 Mbits/sec    0             sender
[  5]   0.00-10.00  sec  1.09 GBytes   940 Mbits/sec                  receiver

And the memcpy() overhead during the RX drops dramatically.

before:

Overhead  Shared O  Symbol
  43.35%  [kernel]  [k] memcpy
  33.77%  [kernel]  [k] __asm_copy_to_user
   3.64%  [kernel]  [k] sifive_l2_flush64_range
Quoted 7 lines

after:

Overhead  Shared O  Symbol
  45.40%  [kernel]  [k] __asm_copy_to_user
  28.09%  [kernel]  [k] memcpy
   4.27%  [kernel]  [k] sifive_l2_flush64_range

This patch completely breaks my Jetson TX2 system, composed of 2 Nvidia Denver and 4 Cortex-A57, in a very "funny" way.

Any significant amount of traffic result in all sort of corruption (ssh connections get dropped, Debian packages downloaded have the wrong checksums) if any Denver core is involved in any significant way (packet processing, interrupt handling). And it is all triggered by this very change.

The only way I have to make it work on a Denver core is to route the interrupt to that particular core and taskset the workload to it. Any other configuration involving a Denver CPU results in some sort of corruption. On their own, the A57s are fine.

Quoted 5 lines

This smells of memory ordering going really wrong, which this change would expose. I haven't had a chance to dig into the driver yet (it took me long enough to bisect it), but if someone points me at what is supposed to synchronise the DMA when receiving an interrupt, I'll have a look.

One other thing that kind of rings a bell when reading DMA and interrupts is a recent report (and attempt to fix this) where upon resume from system suspend, the DMA descriptors would get corrupted.

I don't think we ever figured out what exactly the problem was, but interestingly the fix for the issue immediately caused things to go haywire on... Jetson TX2.

I recall looking at this a bit and couldn't find where exactly the DMA was being synchronized on suspend/resume, or what the mechanism was to ensure that (in transit) packets were not received after the suspension of the Ethernet device. Some information about this can be found here:

	https://lore.kernel.org/netdev/[email protected]/

It's interesting that this happens only on Jetson TX2. Apparently on the newer Jetson AGX Xavier this problem does not occur. I think Jon also narrowed this down to being related to the IOMMU being enabled on Jetson TX2, whereas Jetson AGX Xavier didn't have it enabled. I wasn't able to find any notes on whether disabling the IOMMU on Jetson TX2 did anything to improve on this, so perhaps that's something worth trying.

We have since enabled the IOMMU on Jetson AGX Xavier, and I haven't seen any test reports indicating that this is causing issues. So I don't think this has anything directly to do with the IOMMU support.

That said, if these problems are all exclusive to Jetson TX2, or rather Tegra186, that could indicate that we're missing something at a more fundamental level (maybe some cache maintenance quirk?).

Thierry

Quoted 27 lines
Quoted 24 lines
---
 drivers/net/ethernet/stmicro/stmmac/stmmac.h | 4 ++--
 1 file changed, 2 insertions(+), 2 deletions(-)
drivers/net/ethernet/stmicro/stmmac/stmmac.h +2 -2
index b6cd43eda7ac..04bdb3950d63 100644
@@ -338,9 +338,9 @@ static inline bool stmmac_xdp_is_enabled(struct stmmac_priv *priv)
 static inline unsigned int stmmac_rx_offset(struct stmmac_priv *priv)
 {
 	if (stmmac_xdp_is_enabled(priv))
-		return XDP_PACKET_HEADROOM;
+		return XDP_PACKET_HEADROOM + NET_IP_ALIGN;
 
-	return 0;
+	return NET_SKB_PAD + NET_IP_ALIGN;
 }
 
 void stmmac_disable_rx_queue(struct stmmac_priv *priv, u32 queue);
Signature · 1 line
2.31.1
Signature · 1 line
Without deviation from the norm, progress is not possible.
#32 lore raw
Quoted 108 lines

-----Original Message-----
From: Thierry Reding <[email protected]>
Sent: 2021年8月11日 18:42
To: Marc Zyngier <[email protected]>
Cc: Matteo Croce <[email protected]>; [email protected]; [email protected]; [email protected]; Giuseppe Cavallaro <[email protected]>; Alexandre Torgue <[email protected]>; David S. Miller <[email protected]>; Jakub Kicinski <[email protected]>; Palmer Dabbelt <[email protected]>; Paul Walmsley <[email protected]>; Drew Fustini <[email protected]>; Emil Renner Berthing <[email protected]>; Jon Hunter <[email protected]>; Will Deacon <[email protected]>
Subject: Re: [PATCH net-next] stmmac: align RX buffers

Quoted 22 lines · Marc Zyngier

Hi all,

[adding Thierry, Jon and Will to the fun]

Quoted 16 lines · Matteo Croce

From: Matteo Croce <[email protected]>

On RX an SKB is allocated and the received buffer is copied into it. But on some architectures, the memcpy() needs the source and destination buffers to have the same alignment to be efficient.

This is not our case, because SKB data pointer is misaligned by two bytes to compensate the ethernet header.

Align the RX buffer the same way as the SKB one, so the copy is faster. An iperf3 RX test gives a decent improvement on a RISC-V machine:

before:
[ ID] Interval Transfer Bitrate Retr
[ 5] 0.00-10.00 sec 733 MBytes 615 Mbits/sec 88

sender

Quoted 1 line
Quoted 1 line

[ 5] 0.00-10.01 sec 730 MBytes 612 Mbits/sec

receiver

Quoted 4 lines
Quoted 4 lines

after:
[ ID] Interval Transfer Bitrate Retr
[ 5] 0.00-10.00 sec 1.10 GBytes 942 Mbits/sec 0

sender

Quoted 1 line
Quoted 1 line

[ 5] 0.00-10.00 sec 1.09 GBytes 940 Mbits/sec

receiver

Quoted 31 lines
Quoted 9 lines

And the memcpy() overhead during the RX drops dramatically.

before:

Overhead  Shared O  Symbol
  43.35%  [kernel]  [k] memcpy
  33.77%  [kernel]  [k] __asm_copy_to_user
   3.64%  [kernel]  [k] sifive_l2_flush64_range
Quoted 7 lines

after:

Overhead  Shared O  Symbol
  45.40%  [kernel]  [k] __asm_copy_to_user
  28.09%  [kernel]  [k] memcpy
   4.27%  [kernel]  [k] sifive_l2_flush64_range

This patch completely breaks my Jetson TX2 system, composed of 2 Nvidia Denver and 4 Cortex-A57, in a very "funny" way.

Any significant amount of traffic result in all sort of corruption (ssh connections get dropped, Debian packages downloaded have the wrong checksums) if any Denver core is involved in any significant way (packet processing, interrupt handling). And it is all triggered by this very change.

The only way I have to make it work on a Denver core is to route the interrupt to that particular core and taskset the workload to it. Any other configuration involving a Denver CPU results in some sort of corruption. On their own, the A57s are fine.

Quoted 5 lines

This smells of memory ordering going really wrong, which this change would expose. I haven't had a chance to dig into the driver yet (it took me long enough to bisect it), but if someone points me at what is supposed to synchronise the DMA when receiving an interrupt, I'll have a look.

One other thing that kind of rings a bell when reading DMA and interrupts is a recent report (and attempt to fix this) where upon resume from system suspend, the DMA descriptors would get corrupted.

I don't think we ever figured out what exactly the problem was, but interestingly the fix for the issue immediately caused things to go haywire on... Jetson TX2.

I recall looking at this a bit and couldn't find where exactly the DMA was being synchronized on suspend/resume, or what the mechanism was to ensure that (in transit) packets were not received after the suspension of the Ethernet device. Some information about this can be found here:

	https://lore.kernel.org/netdev/708edb92-a5df-ecc4-3126-5ab36707e275

@nvidia.com/

It's interesting that this happens only on Jetson TX2. Apparently on the newer Jetson AGX Xavier this problem does not occur. I think Jon also narrowed this down to being related to the IOMMU being enabled on Jetson TX2, whereas Jetson AGX Xavier didn't have it enabled. I wasn't able to find any notes on whether disabling the IOMMU on Jetson TX2 did anything to improve on this, so perhaps that's something worth trying.

Quoted 7 lines

We have since enabled the IOMMU on Jetson AGX Xavier, and I haven't seen any test reports indicating that this is causing issues. So I don't think this has anything directly to do with the IOMMU support.

That said, if these problems are all exclusive to Jetson TX2, or rather Tegra186, that could indicate that we're missing something at a more fundamental level (maybe some cache maintenance quirk?).

Hey Thierry,

Please also notice me if you found the root cause, that would be appreciated! I have not upstream the fix you mentioned yet since your continuous NACK.

Thanks in advance 😊

Best Regards,
Joakim Zhang

Quoted 30 lines

Thierry

Quoted 28 lines
Quoted 22 lines
---
 drivers/net/ethernet/stmicro/stmmac/stmmac.h | 4 ++--
 1 file changed, 2 insertions(+), 2 deletions(-)
drivers/net/ethernet/stmicro/stmmac/stmmac.h +0 -0
index b6cd43eda7ac..04bdb3950d63 100644
@@ -338,9 +338,9 @@ static inline bool stmmac_xdp_is_enabled(struct

stmmac_priv *priv) static inline unsigned int
stmmac_rx_offset(struct stmmac_priv *priv) {

drivers/net/ethernet/stmicro/stmmac/stmmac.h +2 -3
 	if (stmmac_xdp_is_enabled(priv))
-		return XDP_PACKET_HEADROOM;
+		return XDP_PACKET_HEADROOM + NET_IP_ALIGN;
 
-	return 0;
+	return NET_SKB_PAD + NET_IP_ALIGN;
 }
 
 void stmmac_disable_rx_queue(struct stmmac_priv *priv, u32 queue);
--
Quoted 3 lines

2.31.1

--
Without deviation from the norm, progress is not possible.

#7 lore raw
In reply to Thierry Reding
Quoted 71 lines · Thierry Reding
Quoted 59 lines · Marc Zyngier

Hi all,

[adding Thierry, Jon and Will to the fun]

Quoted 31 lines · Matteo Croce

From: Matteo Croce <[email protected]>

On RX an SKB is allocated and the received buffer is copied into it. But on some architectures, the memcpy() needs the source and destination buffers to have the same alignment to be efficient.

This is not our case, because SKB data pointer is misaligned by two bytes to compensate the ethernet header.

Align the RX buffer the same way as the SKB one, so the copy is faster. An iperf3 RX test gives a decent improvement on a RISC-V machine:

before:
[ ID] Interval Transfer Bitrate Retr

[  5]   0.00-10.00  sec   733 MBytes   615 Mbits/sec   88             sender
[  5]   0.00-10.01  sec   730 MBytes   612 Mbits/sec                  receiver

after:

[ ID] Interval           Transfer     Bitrate         Retr
[  5]   0.00-10.00  sec  1.10 GBytes   942 Mbits/sec    0             sender
[  5]   0.00-10.00  sec  1.09 GBytes   940 Mbits/sec                  receiver

And the memcpy() overhead during the RX drops dramatically.

before:

Overhead  Shared O  Symbol
  43.35%  [kernel]  [k] memcpy
  33.77%  [kernel]  [k] __asm_copy_to_user
   3.64%  [kernel]  [k] sifive_l2_flush64_range
Quoted 7 lines

after:

Overhead  Shared O  Symbol
  45.40%  [kernel]  [k] __asm_copy_to_user
  28.09%  [kernel]  [k] memcpy
   4.27%  [kernel]  [k] sifive_l2_flush64_range

This patch completely breaks my Jetson TX2 system, composed of 2 Nvidia Denver and 4 Cortex-A57, in a very "funny" way.

Any significant amount of traffic result in all sort of corruption (ssh connections get dropped, Debian packages downloaded have the wrong checksums) if any Denver core is involved in any significant way (packet processing, interrupt handling). And it is all triggered by this very change.

The only way I have to make it work on a Denver core is to route the interrupt to that particular core and taskset the workload to it. Any other configuration involving a Denver CPU results in some sort of corruption. On their own, the A57s are fine.

Quoted 5 lines

This smells of memory ordering going really wrong, which this change would expose. I haven't had a chance to dig into the driver yet (it took me long enough to bisect it), but if someone points me at what is supposed to synchronise the DMA when receiving an interrupt, I'll have a look.

I recall that Jon was looking into a similar issue recently, though I think the failure mode was slightly different. I also vaguely recall that CPU frequency was impacting this to some degree (lower CPU frequencies would increase the chances of this happening).

Quoted 37 lines

Jon's currently out of office, but let me try and dig up the details on this.

Thierry

Quoted 32 lines

Thanks,

	M.
Quoted 24 lines
---
 drivers/net/ethernet/stmicro/stmmac/stmmac.h | 4 ++--
 1 file changed, 2 insertions(+), 2 deletions(-)
drivers/net/ethernet/stmicro/stmmac/stmmac.h +2 -2
index b6cd43eda7ac..04bdb3950d63 100644
@@ -338,9 +338,9 @@ static inline bool stmmac_xdp_is_enabled(struct stmmac_priv *priv)
 static inline unsigned int stmmac_rx_offset(struct stmmac_priv *priv)
 {
 	if (stmmac_xdp_is_enabled(priv))
-		return XDP_PACKET_HEADROOM;
+		return XDP_PACKET_HEADROOM + NET_IP_ALIGN;
 
-	return 0;
+	return NET_SKB_PAD + NET_IP_ALIGN;
 }
 
 void stmmac_disable_rx_queue(struct stmmac_priv *priv, u32 queue);
Signature · 1 line
2.31.1
Signature · 1 line
Without deviation from the norm, progress is not possible.

Are you sure you do not need to adjust stmmac_set_bfsize(), stmmac_rx_buf1_len() and stmmac_rx_buf2_len() ?

Presumably DEFAULT_BUFSIZE also want to be increased by NET_SKB_PAD

Patch for stmmac_rx_buf1_len() :

drivers/net/ethernet/stmicro/stmmac/stmmac_main.c +2 -2
index 7b8404a21544cf29668e8a14240c3971e6bce0c3..041a74e7efca3436bfe3e17f972dd156173957a9 100644
@@ -4508,12 +4508,12 @@ static unsigned int stmmac_rx_buf1_len(struct stmmac_priv *priv,
 
        /* First descriptor, not last descriptor and not split header */
        if (status & rx_not_ls)
-               return priv->dma_buf_sz;
+               return priv->dma_buf_sz - NET_SKB_PAD - NET_IP_ALIGN;
 
        plen = stmmac_get_rx_frame_len(priv, p, coe);
 
        /* First descriptor and last descriptor and not split header */
-       return min_t(unsigned int, priv->dma_buf_sz, plen);
+       return min_t(unsigned int, priv->dma_buf_sz - NET_SKB_PAD - NET_IP_ALIGN, plen);
 }
 
 static unsigned int stmmac_rx_buf2_len(struct stmmac_priv *priv,
#33 lore raw
In reply to Thierry Reding
Quoted 67 lines · Thierry Reding
Quoted 59 lines · Marc Zyngier

Hi all,

[adding Thierry, Jon and Will to the fun]

Quoted 31 lines · Matteo Croce

From: Matteo Croce <[email protected]>

On RX an SKB is allocated and the received buffer is copied into it. But on some architectures, the memcpy() needs the source and destination buffers to have the same alignment to be efficient.

This is not our case, because SKB data pointer is misaligned by two bytes to compensate the ethernet header.

Align the RX buffer the same way as the SKB one, so the copy is faster. An iperf3 RX test gives a decent improvement on a RISC-V machine:

before:
[ ID] Interval Transfer Bitrate Retr

[  5]   0.00-10.00  sec   733 MBytes   615 Mbits/sec   88             sender
[  5]   0.00-10.01  sec   730 MBytes   612 Mbits/sec                  receiver

after:

[ ID] Interval           Transfer     Bitrate         Retr
[  5]   0.00-10.00  sec  1.10 GBytes   942 Mbits/sec    0             sender
[  5]   0.00-10.00  sec  1.09 GBytes   940 Mbits/sec                  receiver

And the memcpy() overhead during the RX drops dramatically.

before:

Overhead  Shared O  Symbol
  43.35%  [kernel]  [k] memcpy
  33.77%  [kernel]  [k] __asm_copy_to_user
   3.64%  [kernel]  [k] sifive_l2_flush64_range
Quoted 7 lines

after:

Overhead  Shared O  Symbol
  45.40%  [kernel]  [k] __asm_copy_to_user
  28.09%  [kernel]  [k] memcpy
   4.27%  [kernel]  [k] sifive_l2_flush64_range

This patch completely breaks my Jetson TX2 system, composed of 2 Nvidia Denver and 4 Cortex-A57, in a very "funny" way.

Any significant amount of traffic result in all sort of corruption (ssh connections get dropped, Debian packages downloaded have the wrong checksums) if any Denver core is involved in any significant way (packet processing, interrupt handling). And it is all triggered by this very change.

The only way I have to make it work on a Denver core is to route the interrupt to that particular core and taskset the workload to it. Any other configuration involving a Denver CPU results in some sort of corruption. On their own, the A57s are fine.

Quoted 5 lines

This smells of memory ordering going really wrong, which this change would expose. I haven't had a chance to dig into the driver yet (it took me long enough to bisect it), but if someone points me at what is supposed to synchronise the DMA when receiving an interrupt, I'll have a look.

Quoted 7 lines

One other thing that kind of rings a bell when reading DMA and interrupts is a recent report (and attempt to fix this) where upon resume from system suspend, the DMA descriptors would get corrupted.

I don't think we ever figured out what exactly the problem was, but interestingly the fix for the issue immediately caused things to go haywire on... Jetson TX2.

I love this machine... Did this issue occur with the Denver CPUs disabled?

Quoted 5 lines

I recall looking at this a bit and couldn't find where exactly the DMA was being synchronized on suspend/resume, or what the mechanism was to ensure that (in transit) packets were not received after the suspension of the Ethernet device. Some information about this can be found here:

Quoted 8 lines
	https://lore.kernel.org/netdev/[email protected]/

It's interesting that this happens only on Jetson TX2. Apparently on the newer Jetson AGX Xavier this problem does not occur. I think Jon also narrowed this down to being related to the IOMMU being enabled on Jetson TX2, whereas Jetson AGX Xavier didn't have it enabled. I wasn't able to find any notes on whether disabling the IOMMU on Jetson TX2 did anything to improve on this, so perhaps that's something worth trying.

Actually, I was running with the SMMU disabled, as I use the upstream u-boot provided DT. Switching to the kernel one didn't change a thing (with passthough or not).

Quoted 3 lines

We have since enabled the IOMMU on Jetson AGX Xavier, and I haven't seen any test reports indicating that this is causing issues. So I don't think this has anything directly to do with the IOMMU support.

No, it looks more like either ordering or cache management. The fact that this patch messes with the buffer alignment makes me favour the latter...

Quoted 3 lines

That said, if these problems are all exclusive to Jetson TX2, or rather Tegra186, that could indicate that we're missing something at a more fundamental level (maybe some cache maintenance quirk?).

That'd be pretty annoying. Do you know if the Ethernet is a coherent device on this machine? or does it need active cache maintenance?

Thanks,

	M.
Signature · 6 lines
Without deviation from the norm, progress is not possible.

_______________________________________________
linux-riscv mailing list
[email protected]
http://lists.infradead.org/mailman/listinfo/linux-riscv
#8 lore raw
In reply to Eric Dumazet
Quoted 129 lines · Eric Dumazet
Quoted 71 lines · Thierry Reding
Quoted 59 lines · Marc Zyngier

Hi all,

[adding Thierry, Jon and Will to the fun]

Quoted 31 lines · Matteo Croce

From: Matteo Croce <[email protected]>

On RX an SKB is allocated and the received buffer is copied into it. But on some architectures, the memcpy() needs the source and destination buffers to have the same alignment to be efficient.

This is not our case, because SKB data pointer is misaligned by two bytes to compensate the ethernet header.

Align the RX buffer the same way as the SKB one, so the copy is faster. An iperf3 RX test gives a decent improvement on a RISC-V machine:

before:
[ ID] Interval Transfer Bitrate Retr

[  5]   0.00-10.00  sec   733 MBytes   615 Mbits/sec   88             sender
[  5]   0.00-10.01  sec   730 MBytes   612 Mbits/sec                  receiver

after:

[ ID] Interval           Transfer     Bitrate         Retr
[  5]   0.00-10.00  sec  1.10 GBytes   942 Mbits/sec    0             sender
[  5]   0.00-10.00  sec  1.09 GBytes   940 Mbits/sec                  receiver

And the memcpy() overhead during the RX drops dramatically.

before:

Overhead  Shared O  Symbol
  43.35%  [kernel]  [k] memcpy
  33.77%  [kernel]  [k] __asm_copy_to_user
   3.64%  [kernel]  [k] sifive_l2_flush64_range
Quoted 7 lines

after:

Overhead  Shared O  Symbol
  45.40%  [kernel]  [k] __asm_copy_to_user
  28.09%  [kernel]  [k] memcpy
   4.27%  [kernel]  [k] sifive_l2_flush64_range

This patch completely breaks my Jetson TX2 system, composed of 2 Nvidia Denver and 4 Cortex-A57, in a very "funny" way.

Any significant amount of traffic result in all sort of corruption (ssh connections get dropped, Debian packages downloaded have the wrong checksums) if any Denver core is involved in any significant way (packet processing, interrupt handling). And it is all triggered by this very change.

The only way I have to make it work on a Denver core is to route the interrupt to that particular core and taskset the workload to it. Any other configuration involving a Denver CPU results in some sort of corruption. On their own, the A57s are fine.

Quoted 5 lines

This smells of memory ordering going really wrong, which this change would expose. I haven't had a chance to dig into the driver yet (it took me long enough to bisect it), but if someone points me at what is supposed to synchronise the DMA when receiving an interrupt, I'll have a look.

I recall that Jon was looking into a similar issue recently, though I think the failure mode was slightly different. I also vaguely recall that CPU frequency was impacting this to some degree (lower CPU frequencies would increase the chances of this happening).

Quoted 37 lines

Jon's currently out of office, but let me try and dig up the details on this.

Thierry

Quoted 32 lines

Thanks,

	M.
Quoted 24 lines
---
 drivers/net/ethernet/stmicro/stmmac/stmmac.h | 4 ++--
 1 file changed, 2 insertions(+), 2 deletions(-)
drivers/net/ethernet/stmicro/stmmac/stmmac.h +2 -2
index b6cd43eda7ac..04bdb3950d63 100644
@@ -338,9 +338,9 @@ static inline bool stmmac_xdp_is_enabled(struct stmmac_priv *priv)
 static inline unsigned int stmmac_rx_offset(struct stmmac_priv *priv)
 {
 	if (stmmac_xdp_is_enabled(priv))
-		return XDP_PACKET_HEADROOM;
+		return XDP_PACKET_HEADROOM + NET_IP_ALIGN;
 
-	return 0;
+	return NET_SKB_PAD + NET_IP_ALIGN;
 }
 
 void stmmac_disable_rx_queue(struct stmmac_priv *priv, u32 queue);
Signature · 1 line
2.31.1
Signature · 1 line
Without deviation from the norm, progress is not possible.

Are you sure you do not need to adjust stmmac_set_bfsize(), stmmac_rx_buf1_len() and stmmac_rx_buf2_len() ?

Presumably DEFAULT_BUFSIZE also want to be increased by NET_SKB_PAD

Patch for stmmac_rx_buf1_len() :

drivers/net/ethernet/stmicro/stmmac/stmmac_main.c +0 -1
index 7b8404a21544cf29668e8a14240c3971e6bce0c3..041a74e7efca3436bfe3e17f972dd156173957a9 100644
@@ -4508,12 +4508,12 @@ static unsigned int stmmac_rx_buf1_len(struct stmmac_priv *priv,
 
        /* First descriptor, not last descriptor and not split header */
        if (status & rx_not_ls)
-               return priv->dma_buf_sz;

drivers/net/ethernet/stmicro/stmmac/stmmac_main.c static unsigned int stmmac_rx_buf1_len(struct stmmac_priv *priv, · line 4519

Quoted 10 lines
drivers/net/ethernet/stmicro/stmmac/stmmac_main.c +2 -1
+               return priv->dma_buf_sz - NET_SKB_PAD - NET_IP_ALIGN;
 
        plen = stmmac_get_rx_frame_len(priv, p, coe);
 
        /* First descriptor and last descriptor and not split header */
-       return min_t(unsigned int, priv->dma_buf_sz, plen);
+       return min_t(unsigned int, priv->dma_buf_sz - NET_SKB_PAD - NET_IP_ALIGN, plen);
 }
 
 static unsigned int stmmac_rx_buf2_len(struct stmmac_priv *priv,

Feels like a major deficiency of the original patch. Happy to test a more complete patch if/when you have one.

Thanks,

	M.
Signature · 1 line
Without deviation from the norm, progress is not possible.
#9 lore raw
Quoted 28 lines · Marc Zyngier
Quoted 25 lines · Eric Dumazet

Are you sure you do not need to adjust stmmac_set_bfsize(), stmmac_rx_buf1_len() and stmmac_rx_buf2_len() ?

Presumably DEFAULT_BUFSIZE also want to be increased by NET_SKB_PAD

Patch for stmmac_rx_buf1_len() :

drivers/net/ethernet/stmicro/stmmac/stmmac_main.c +2 -2
index 7b8404a21544cf29668e8a14240c3971e6bce0c3..041a74e7efca3436bfe3e17f972dd156173957a9 100644
@@ -4508,12 +4508,12 @@ static unsigned int stmmac_rx_buf1_len(struct stmmac_priv *priv,
 
        /* First descriptor, not last descriptor and not split header */
        if (status & rx_not_ls)
-               return priv->dma_buf_sz;
+               return priv->dma_buf_sz - NET_SKB_PAD - NET_IP_ALIGN;
 
        plen = stmmac_get_rx_frame_len(priv, p, coe);
 
        /* First descriptor and last descriptor and not split header */
-       return min_t(unsigned int, priv->dma_buf_sz, plen);
+       return min_t(unsigned int, priv->dma_buf_sz - NET_SKB_PAD - NET_IP_ALIGN, plen);
 }
 
Quoted 4 lines

drivers/net/ethernet/stmicro/stmmac/stmmac_main.c static unsigned int stmmac_rx_buf1_len(struct stmmac_priv *priv, · line 4519

Quoted 1 line
drivers/net/ethernet/stmicro/stmmac/stmmac_main.c +0 -0
 static unsigned int stmmac_rx_buf2_len(struct stmmac_priv *priv,

Feels like a major deficiency of the original patch. Happy to test a more complete patch if/when you have one.

I wont have time in the immediate future.

Matteo, if you do not work on a fix, I suggest we revert
a955318fe67ec0d962760b5ee58e74bffaf649b8 stmmac: align RX buffers

before a more polished version can be submitted.

#10 lore raw
Quoted 37 lines · On Thu, 12 Aug 2021 10:48:03 +0200 Eric Dumazet <[email protected]> wrote:
Quoted 29 lines · Marc Zyngier
Quoted 26 lines · Eric Dumazet

Are you sure you do not need to adjust stmmac_set_bfsize(), stmmac_rx_buf1_len() and stmmac_rx_buf2_len() ?

Presumably DEFAULT_BUFSIZE also want to be increased by NET_SKB_PAD

Patch for stmmac_rx_buf1_len() :

(quoted diff) +0 -0

b/drivers/net/ethernet/stmicro/stmmac/stmmac_main.c index 7b8404a21544cf29668e8a14240c3971e6bce0c3..041a74e7efca3436bfe3e17f972dd156173957a9 100644 --- a/drivers/net/ethernet/stmicro/stmmac/stmmac_main.c +++ b/drivers/net/ethernet/stmicro/stmmac/stmmac_main.c @@ -4508,12 +4508,12 @@ static unsigned int stmmac_rx_buf1_len(struct
stmmac_priv *priv, /* First descriptor, not last descriptor and
not split header */ if (status & rx_not_ls)

(quoted diff) +1 -1
-               return priv->dma_buf_sz;
+               return priv->dma_buf_sz - NET_SKB_PAD -
NET_IP_ALIGN; 
        plen = stmmac_get_rx_frame_len(priv, p, coe);
 
        /* First descriptor and last descriptor and not split

header */

(quoted diff) +1 -2
-       return min_t(unsigned int, priv->dma_buf_sz, plen);
+       return min_t(unsigned int, priv->dma_buf_sz - NET_SKB_PAD
- NET_IP_ALIGN, plen); }
Quoted 4 lines
Quoted 1 line

static unsigned int stmmac_rx_buf2_len(struct stmmac_priv *priv,

Feels like a major deficiency of the original patch. Happy to test a more complete patch if/when you have one.

Quoted 7 lines

I wont have time in the immediate future.

Matteo, if you do not work on a fix, I suggest we revert
a955318fe67ec0d962760b5ee58e74bffaf649b8 stmmac: align RX buffers

before a more polished version can be submitted.

Better to use stmmac_rx_offset() so to have the correct length when using XDP. Also, when XDP is enabled, the offset was XDP_PACKET_HEADROOM (i.e. 256 bytes) even before the change, so it could be already broken. Mark, can you try on the Jetson TX2 by attaching an XDP program and see if it works without my change?

A possible fix, which takes in account also the XDP headroom for stmmac_rx_buf1_len() only could be (only compile tested, I don't have the hardware now):

drivers/net/ethernet/stmicro/stmmac/stmmac_main.c +3 -3
index 7b8404a21544..b205f43f849a 100644
@@ -93,7 +93,7 @@ static int tc = TC_DEFAULT;
 module_param(tc, int, 0644);
 MODULE_PARM_DESC(tc, "DMA threshold control value");
 
-#define	DEFAULT_BUFSIZE	1536
+#define	DEFAULT_BUFSIZE	1536 + XDP_PACKET_HEADROOM + NET_IP_ALIGN
 static int buf_sz = DEFAULT_BUFSIZE;
 module_param(buf_sz, int, 0644);
 MODULE_PARM_DESC(buf_sz, "DMA buffer size");
@@ -4508,12 +4508,12 @@ static unsigned int stmmac_rx_buf1_len(struct stmmac_priv *priv,
 
 	/* First descriptor, not last descriptor and not split header */
 	if (status & rx_not_ls)
-		return priv->dma_buf_sz;
+		return priv->dma_buf_sz - stmmac_rx_offset(priv);
 
 	plen = stmmac_get_rx_frame_len(priv, p, coe);
 
 	/* First descriptor and last descriptor and not split header */
-	return min_t(unsigned int, priv->dma_buf_sz, plen);
+	return min_t(unsigned int, priv->dma_buf_sz - stmmac_rx_offset(priv), plen);
 }
 
 static unsigned int stmmac_rx_buf2_len(struct stmmac_priv *priv,
Signature · 1 line
per aspera ad upstream
#11 lore raw
Quoted 46 lines · Matteo Croce
Quoted 42 lines · On Thu, 12 Aug 2021 10:48:03 +0200 Eric Dumazet <[email protected]> wrote:
Quoted 29 lines · Marc Zyngier
Quoted 26 lines · Eric Dumazet

Are you sure you do not need to adjust stmmac_set_bfsize(), stmmac_rx_buf1_len() and stmmac_rx_buf2_len() ?

Presumably DEFAULT_BUFSIZE also want to be increased by NET_SKB_PAD

Patch for stmmac_rx_buf1_len() :

(quoted diff) +0 -0

b/drivers/net/ethernet/stmicro/stmmac/stmmac_main.c index 7b8404a21544cf29668e8a14240c3971e6bce0c3..041a74e7efca3436bfe3e17f972dd156173957a9 100644 --- a/drivers/net/ethernet/stmicro/stmmac/stmmac_main.c +++ b/drivers/net/ethernet/stmicro/stmmac/stmmac_main.c @@ -4508,12 +4508,12 @@ static unsigned int stmmac_rx_buf1_len(struct
stmmac_priv *priv, /* First descriptor, not last descriptor and
not split header */ if (status & rx_not_ls)

(quoted diff) +1 -1
-               return priv->dma_buf_sz;
+               return priv->dma_buf_sz - NET_SKB_PAD -
NET_IP_ALIGN; 
        plen = stmmac_get_rx_frame_len(priv, p, coe);
 
        /* First descriptor and last descriptor and not split

header */

(quoted diff) +1 -2
-       return min_t(unsigned int, priv->dma_buf_sz, plen);
+       return min_t(unsigned int, priv->dma_buf_sz - NET_SKB_PAD
- NET_IP_ALIGN, plen); }
Quoted 4 lines
Quoted 1 line

static unsigned int stmmac_rx_buf2_len(struct stmmac_priv *priv,

Feels like a major deficiency of the original patch. Happy to test a more complete patch if/when you have one.

I wont have time in the immediate future.

Matteo, if you do not work on a fix, I suggest we revert
a955318fe67ec0d962760b5ee58e74bffaf649b8 stmmac: align RX buffers

Quoted 8 lines
Quoted 2 lines

before a more polished version can be submitted.

Better to use stmmac_rx_offset() so to have the correct length when using XDP. Also, when XDP is enabled, the offset was XDP_PACKET_HEADROOM (i.e. 256 bytes) even before the change, so it could be already broken. Mark, can you try on the Jetson TX2 by attaching an XDP program and see if it works without my change?

Sorry, you'll have to hold my hand here, as I know exactly nothing about XDP....

Quoted 3 lines

A possible fix, which takes in account also the XDP headroom for stmmac_rx_buf1_len() only could be (only compile tested, I don't have the hardware now):

However, this doesn't fix my issue. I still get all sort of corruption. Probably stmmac_rx_buf2_len() also need adjusting (it has a similar logic as its buf1 counterpart...)

Unless you can fix it very quickly, and given that we're towards the end of the cycle, I'd be more comfortable if we reverted this patch.

Thanks,

	M.
Signature · 1 line
Without deviation from the norm, progress is not possible.
#12 lore raw
Quoted 52 lines · Marc Zyngier
Quoted 49 lines · Matteo Croce
Quoted 44 lines · On Thu, 12 Aug 2021 10:48:03 +0200 Eric Dumazet <[email protected]> wrote:
Quoted 29 lines · Marc Zyngier
Quoted 26 lines · Eric Dumazet

Are you sure you do not need to adjust stmmac_set_bfsize(), stmmac_rx_buf1_len() and stmmac_rx_buf2_len() ?

Presumably DEFAULT_BUFSIZE also want to be increased by NET_SKB_PAD

Patch for stmmac_rx_buf1_len() :

(quoted diff) +0 -0

b/drivers/net/ethernet/stmicro/stmmac/stmmac_main.c index 7b8404a21544cf29668e8a14240c3971e6bce0c3..041a74e7efca3436bfe3e17f972dd156173957a9 100644 --- a/drivers/net/ethernet/stmicro/stmmac/stmmac_main.c +++ b/drivers/net/ethernet/stmicro/stmmac/stmmac_main.c @@ -4508,12 +4508,12 @@ static unsigned int stmmac_rx_buf1_len(struct
stmmac_priv *priv, /* First descriptor, not last descriptor and
not split header */ if (status & rx_not_ls)

(quoted diff) +1 -1
-               return priv->dma_buf_sz;
+               return priv->dma_buf_sz - NET_SKB_PAD -
NET_IP_ALIGN;
        plen = stmmac_get_rx_frame_len(priv, p, coe);

        /* First descriptor and last descriptor and not split

header */

(quoted diff) +1 -2
-       return min_t(unsigned int, priv->dma_buf_sz, plen);
+       return min_t(unsigned int, priv->dma_buf_sz - NET_SKB_PAD
- NET_IP_ALIGN, plen); }
Quoted 4 lines
Quoted 1 line

static unsigned int stmmac_rx_buf2_len(struct stmmac_priv *priv,

Feels like a major deficiency of the original patch. Happy to test a more complete patch if/when you have one.

I wont have time in the immediate future.

Matteo, if you do not work on a fix, I suggest we revert
a955318fe67ec0d962760b5ee58e74bffaf649b8 stmmac: align RX buffers

before a more polished version can be submitted.

Quoted 9 lines
Quoted 5 lines

Better to use stmmac_rx_offset() so to have the correct length when using XDP. Also, when XDP is enabled, the offset was XDP_PACKET_HEADROOM (i.e. 256 bytes) even before the change, so it could be already broken. Mark, can you try on the Jetson TX2 by attaching an XDP program and see if it works without my change?

Sorry, you'll have to hold my hand here, as I know exactly nothing about XDP....

Attach the attached object with:

ip link set eth0 xdp object passall.o

This is an empty XDP program, its source:

__attribute__((section("prog"), used))
int xdp_main(struct xdp_md *ctx)

{
       return XDP_PASS;
}

Every packet will pass untouched, but the offset will be shifted from 0 to 256 bytes, which could trigger the problem anyway:

Quoted 11 lines
Quoted 3 lines

A possible fix, which takes in account also the XDP headroom for stmmac_rx_buf1_len() only could be (only compile tested, I don't have the hardware now):

However, this doesn't fix my issue. I still get all sort of corruption. Probably stmmac_rx_buf2_len() also need adjusting (it has a similar logic as its buf1 counterpart...)

Unless you can fix it very quickly, and given that we're towards the end of the cycle, I'd be more comfortable if we reverted this patch.

Can it be that the HW can't do DMA on an address which is not word aligned? What if you replace NET_SKB_PAD with, let's say, 8?

Regards,

Signature · 1 line
per aspera ad upstream
#34 lore raw
In reply to Marc Zyngier
Quoted 73 lines · Marc Zyngier
Quoted 71 lines · Thierry Reding
Quoted 59 lines · Marc Zyngier

Hi all,

[adding Thierry, Jon and Will to the fun]

Quoted 31 lines · Matteo Croce

From: Matteo Croce <[email protected]>

On RX an SKB is allocated and the received buffer is copied into it. But on some architectures, the memcpy() needs the source and destination buffers to have the same alignment to be efficient.

This is not our case, because SKB data pointer is misaligned by two bytes to compensate the ethernet header.

Align the RX buffer the same way as the SKB one, so the copy is faster. An iperf3 RX test gives a decent improvement on a RISC-V machine:

before:
[ ID] Interval Transfer Bitrate Retr

[  5]   0.00-10.00  sec   733 MBytes   615 Mbits/sec   88             sender
[  5]   0.00-10.01  sec   730 MBytes   612 Mbits/sec                  receiver

after:

[ ID] Interval           Transfer     Bitrate         Retr
[  5]   0.00-10.00  sec  1.10 GBytes   942 Mbits/sec    0             sender
[  5]   0.00-10.00  sec  1.09 GBytes   940 Mbits/sec                  receiver

And the memcpy() overhead during the RX drops dramatically.

before:

Overhead  Shared O  Symbol
  43.35%  [kernel]  [k] memcpy
  33.77%  [kernel]  [k] __asm_copy_to_user
   3.64%  [kernel]  [k] sifive_l2_flush64_range
Quoted 7 lines

after:

Overhead  Shared O  Symbol
  45.40%  [kernel]  [k] __asm_copy_to_user
  28.09%  [kernel]  [k] memcpy
   4.27%  [kernel]  [k] sifive_l2_flush64_range

This patch completely breaks my Jetson TX2 system, composed of 2 Nvidia Denver and 4 Cortex-A57, in a very "funny" way.

Any significant amount of traffic result in all sort of corruption (ssh connections get dropped, Debian packages downloaded have the wrong checksums) if any Denver core is involved in any significant way (packet processing, interrupt handling). And it is all triggered by this very change.

The only way I have to make it work on a Denver core is to route the interrupt to that particular core and taskset the workload to it. Any other configuration involving a Denver CPU results in some sort of corruption. On their own, the A57s are fine.

Quoted 5 lines

This smells of memory ordering going really wrong, which this change would expose. I haven't had a chance to dig into the driver yet (it took me long enough to bisect it), but if someone points me at what is supposed to synchronise the DMA when receiving an interrupt, I'll have a look.

One other thing that kind of rings a bell when reading DMA and interrupts is a recent report (and attempt to fix this) where upon resume from system suspend, the DMA descriptors would get corrupted.

Quoted 6 lines
Quoted 3 lines

I don't think we ever figured out what exactly the problem was, but interestingly the fix for the issue immediately caused things to go haywire on... Jetson TX2.

I love this machine... Did this issue occur with the Denver CPUs disabled?

Interestingly I've been doing some work on a newer device called Jetson TX2 NX (which is kind of a trimmed-down version of Jetson TX2, in the spirit of the Jetson Nano) and I can't seem to reproduce these failures there (tested on next-20210812).

I'll go dig out my Jetson TX2 to run the same tests there, because I've also been using a development version of the bootloader stack and flashing tools and all that, so it's possible that something was fixed at that level. I don't think I've ever tried disabling the Denver CPUs, but then I've also never seen these issues myself.

Just out of curiosity, what version of the BSP have you been using to flash?

One other thing that I ran into: there's a known issue with the PHY configuration. We mark the PHY on most devices as "rgmii-id" on most devices and then the Marvell PHY driver needs to be enabled. Jetson TX2 has phy-mode = "rgmii", so it /should/ work okay.

Typically what we're seeing with that misconfiguration is that the device fails to get an IP address, but it might still be worth trying to switch Jetson TX2 to rgmii-id and using the Marvell PHY, to see if that improves anything.

Quoted 26 lines
Quoted 5 lines

I recall looking at this a bit and couldn't find where exactly the DMA was being synchronized on suspend/resume, or what the mechanism was to ensure that (in transit) packets were not received after the suspension of the Ethernet device. Some information about this can be found here:

Quoted 8 lines
	https://lore.kernel.org/netdev/[email protected]/

It's interesting that this happens only on Jetson TX2. Apparently on the newer Jetson AGX Xavier this problem does not occur. I think Jon also narrowed this down to being related to the IOMMU being enabled on Jetson TX2, whereas Jetson AGX Xavier didn't have it enabled. I wasn't able to find any notes on whether disabling the IOMMU on Jetson TX2 did anything to improve on this, so perhaps that's something worth trying.

Actually, I was running with the SMMU disabled, as I use the upstream u-boot provided DT. Switching to the kernel one didn't change a thing (with passthough or not).

Quoted 3 lines

We have since enabled the IOMMU on Jetson AGX Xavier, and I haven't seen any test reports indicating that this is causing issues. So I don't think this has anything directly to do with the IOMMU support.

No, it looks more like either ordering or cache management. The fact that this patch messes with the buffer alignment makes me favour the latter...

Quoted 6 lines
Quoted 3 lines

That said, if these problems are all exclusive to Jetson TX2, or rather Tegra186, that could indicate that we're missing something at a more fundamental level (maybe some cache maintenance quirk?).

That'd be pretty annoying. Do you know if the Ethernet is a coherent device on this machine? or does it need active cache maintenance?

I don't think Ethernet is a coherent device on Tegra186. I think Tegra194 had various improvements with regard to coherency, but most devices on Tegra186 do need active cache maintenance.

Let me dig through some old patches and mailing list threads. I vaguely recall prototyping a patch that did something special for outer cache flushing, but that may have been Tegra132, not Tegra186. I also don't think we ended up merging that because it turned out to not be needed.

Thierry

#35 lore raw
Quoted 2 lines · Thierry Reding

On Wed, Aug 11, 2021 at 02:23:10PM +0100, Marc Zyngier wrote:

[...]

Quoted 8 lines
Quoted 2 lines

I love this machine... Did this issue occur with the Denver CPUs disabled?

Interestingly I've been doing some work on a newer device called Jetson TX2 NX (which is kind of a trimmed-down version of Jetson TX2, in the spirit of the Jetson Nano) and I can't seem to reproduce these failures there (tested on next-20210812).

Quoted 8 lines

I'll go dig out my Jetson TX2 to run the same tests there, because I've also been using a development version of the bootloader stack and flashing tools and all that, so it's possible that something was fixed at that level. I don't think I've ever tried disabling the Denver CPUs, but then I've also never seen these issues myself.

Just out of curiosity, what version of the BSP have you been using to flash?

I've only used the BSP for a few weeks when I got the board last year. The only thing I use from it is u-boot to chainload an upstream u-boot, and boot Debian from there.

Quoted 9 lines

One other thing that I ran into: there's a known issue with the PHY configuration. We mark the PHY on most devices as "rgmii-id" on most devices and then the Marvell PHY driver needs to be enabled. Jetson TX2 has phy-mode = "rgmii", so it /should/ work okay.

Typically what we're seeing with that misconfiguration is that the device fails to get an IP address, but it might still be worth trying to switch Jetson TX2 to rgmii-id and using the Marvell PHY, to see if that improves anything.

I never failed to get an IP address. Overall, networking has been solid on this machine until this patch. I'll try and mess with this when I get time, but that's probably going to be next week now.

[...]

Quoted 11 lines
Quoted 2 lines

That'd be pretty annoying. Do you know if the Ethernet is a coherent device on this machine? or does it need active cache maintenance?

I don't think Ethernet is a coherent device on Tegra186. I think Tegra194 had various improvements with regard to coherency, but most devices on Tegra186 do need active cache maintenance.

Let me dig through some old patches and mailing list threads. I vaguely recall prototyping a patch that did something special for outer cache flushing, but that may have been Tegra132, not Tegra186. I also don't think we ended up merging that because it turned out to not be needed.

ARMv8 forbid any sort of *visible* outer cache, so I really hope this is not required. We wouldn't be able to support it.

Thanks,

	M.
Signature · 1 line
Without deviation from the norm, progress is not possible.
#36 lore raw
Quoted 21 lines · Marc Zyngier
Quoted 2 lines · Thierry Reding

On Wed, Aug 11, 2021 at 02:23:10PM +0100, Marc Zyngier wrote:

[...]

Quoted 14 lines
Quoted 2 lines

I love this machine... Did this issue occur with the Denver CPUs disabled?

Interestingly I've been doing some work on a newer device called Jetson TX2 NX (which is kind of a trimmed-down version of Jetson TX2, in the spirit of the Jetson Nano) and I can't seem to reproduce these failures there (tested on next-20210812).

I'll go dig out my Jetson TX2 to run the same tests there, because I've also been using a development version of the bootloader stack and flashing tools and all that, so it's possible that something was fixed at that level. I don't think I've ever tried disabling the Denver CPUs, but then I've also never seen these issues myself.

Quoted 6 lines
Quoted 2 lines

Just out of curiosity, what version of the BSP have you been using to flash?

I've only used the BSP for a few weeks when I got the board last year. The only thing I use from it is u-boot to chainload an upstream u-boot, and boot Debian from there.

That's interesting... have you ever tried to inject a version of upstream U-Boot into the BSP and have it flash that instead? That should allow you to drop the chainloading step.

Not that that's likely to have anything to do with this.

Quoted 5 lines
Quoted 5 lines

One other thing that I ran into: there's a known issue with the PHY configuration. We mark the PHY on most devices as "rgmii-id" on most devices and then the Marvell PHY driver needs to be enabled. Jetson TX2 has phy-mode = "rgmii", so it /should/ work okay.

Quoted 8 lines
Quoted 4 lines

Typically what we're seeing with that misconfiguration is that the device fails to get an IP address, but it might still be worth trying to switch Jetson TX2 to rgmii-id and using the Marvell PHY, to see if that improves anything.

I never failed to get an IP address. Overall, networking has been solid on this machine until this patch. I'll try and mess with this when I get time, but that's probably going to be next week now.

So I've hooked up my Jetson TX2 and tried various workloads. I wasn't able to reproduce this on next-20210813. I've tried both the L4T 32.6.1 release and a local development build.

Perhaps one thing to try would be to upgrade your L4T BSP to something newer. I know that there have occasionally been bugs in the MTS firmware, which is what's running on the Denver cores, and newer BSPs can fix those kinds of issues.

If that doesn't help, perhaps try to read out the SoC version numbers so that we can compare. I know that some newer Tegra186 chips behave slightly differently, so that's perhaps a difference that would explain why it's not happening on all devices.

You can read the version and revision from sysfs using something like:

	# cat /sys/devices/soc0/{major,minor,revision}
Quoted 9 lines

[...]

Quoted 7 lines
Quoted 2 lines

That'd be pretty annoying. Do you know if the Ethernet is a coherent device on this machine? or does it need active cache maintenance?

I don't think Ethernet is a coherent device on Tegra186. I think Tegra194 had various improvements with regard to coherency, but most devices on Tegra186 do need active cache maintenance.

Quoted 7 lines
Quoted 4 lines

Let me dig through some old patches and mailing list threads. I vaguely recall prototyping a patch that did something special for outer cache flushing, but that may have been Tegra132, not Tegra186. I also don't think we ended up merging that because it turned out to not be needed.

ARMv8 forbid any sort of *visible* outer cache, so I really hope this is not required. We wouldn't be able to support it.

I couldn't find any trace of this anywhere. So I'm possibly misremembering. It's also more likely that this was on an earlier SoC generation, otherwise I'd probably remember more clearly.

Thierry

maintainer
#28 lore raw
In reply to Marc Zyngier
Quoted 10 lines · On Thu, 12 Aug 2021 12:05:38 +0100 Marc Zyngier wrote:
Quoted 3 lines

A possible fix, which takes in account also the XDP headroom for stmmac_rx_buf1_len() only could be (only compile tested, I don't have the hardware now):

However, this doesn't fix my issue. I still get all sort of corruption. Probably stmmac_rx_buf2_len() also need adjusting (it has a similar logic as its buf1 counterpart...)

Unless you can fix it very quickly, and given that we're towards the end of the cycle, I'd be more comfortable if we reverted this patch.

Any luck investigating this one? The rc6 announcement sounds like there may not be that many more rc releases for 5.14.

#29 lore raw
Quoted 9 lines · On Mon, 16 Aug 2021 08:12:08 -0700 Jakub Kicinski <[email protected]> wrote:
Quoted 8 lines · On Thu, 12 Aug 2021 12:05:38 +0100 Marc Zyngier wrote:
Quoted 3 lines

A possible fix, which takes in account also the XDP headroom for stmmac_rx_buf1_len() only could be (only compile tested, I don't have the hardware now):

However, this doesn't fix my issue. I still get all sort of corruption. Probably stmmac_rx_buf2_len() also need adjusting (it has a similar logic as its buf1 counterpart...)

Quoted 5 lines
Quoted 2 lines

Unless you can fix it very quickly, and given that we're towards the end of the cycle, I'd be more comfortable if we reverted this patch.

Any luck investigating this one? The rc6 announcement sounds like there may not be that many more rc releases for 5.14.

Hi Jackub.

Unfortunately I have only a device with stmmac, and it works fine with the patch. It seems that not all hardware suffers from this issue.

Also, using NET_IP_ALIGN on RX is a common pattern, I think that any ethernet device is doing the same to align the IPv4 header.

Anyway, I asked for two tests on the affected device:
1. Change NET_IP_ALIGN with 8, to see if the DMA has problems in receiving to a non word aligned address
2. load a nop XDP program (I provided one), to see if the problem is already there when XDP is used

I doubt that changing also stmmac_rx_buf2_len would help, but it's worth a try, here is a patch:

drivers/net/ethernet/stmicro/stmmac/stmmac_main.c +5 -5
index 7b8404a21544..73d1f0ec66ff 100644
@@ -93,7 +93,7 @@ static int tc = TC_DEFAULT;
 module_param(tc, int, 0644);
 MODULE_PARM_DESC(tc, "DMA threshold control value");
 
-#define	DEFAULT_BUFSIZE	1536
+#define	DEFAULT_BUFSIZE	1536 + XDP_PACKET_HEADROOM + NET_IP_ALIGN
 static int buf_sz = DEFAULT_BUFSIZE;
 module_param(buf_sz, int, 0644);
 MODULE_PARM_DESC(buf_sz, "DMA buffer size");
@@ -4508,12 +4508,12 @@ static unsigned int stmmac_rx_buf1_len(struct stmmac_priv *priv,
 
 	/* First descriptor, not last descriptor and not split header */
 	if (status & rx_not_ls)
-		return priv->dma_buf_sz;
+		return priv->dma_buf_sz - stmmac_rx_offset(priv);
 
 	plen = stmmac_get_rx_frame_len(priv, p, coe);
 
 	/* First descriptor and last descriptor and not split header */
-	return min_t(unsigned int, priv->dma_buf_sz, plen);
+	return min_t(unsigned int, priv->dma_buf_sz - stmmac_rx_offset(priv), plen);
 }
 
 static unsigned int stmmac_rx_buf2_len(struct stmmac_priv *priv,
@@ -4529,12 +4529,12 @@ static unsigned int stmmac_rx_buf2_len(struct stmmac_priv *priv,
 
 	/* Not last descriptor */
 	if (status & rx_not_ls)
-		return priv->dma_buf_sz;
+		return priv->dma_buf_sz - stmmac_rx_offset(priv);
 
 	plen = stmmac_get_rx_frame_len(priv, p, coe);
 
 	/* Last descriptor */
-	return plen - len;
+	return plen - len - stmmac_rx_offset(priv);
 }
 
 static int stmmac_xdp_xmit_xdpf(struct stmmac_priv *priv, int queue,

Regards,

Signature · 1 line
per aspera ad upstream
#30 lore raw
Quoted 24 lines · Matteo Croce
Quoted 9 lines · On Mon, 16 Aug 2021 08:12:08 -0700 Jakub Kicinski <[email protected]> wrote:
Quoted 8 lines · On Thu, 12 Aug 2021 12:05:38 +0100 Marc Zyngier wrote:
Quoted 3 lines

A possible fix, which takes in account also the XDP headroom for stmmac_rx_buf1_len() only could be (only compile tested, I don't have the hardware now):

However, this doesn't fix my issue. I still get all sort of corruption. Probably stmmac_rx_buf2_len() also need adjusting (it has a similar logic as its buf1 counterpart...)

Quoted 5 lines
Quoted 2 lines

Unless you can fix it very quickly, and given that we're towards the end of the cycle, I'd be more comfortable if we reverted this patch.

Any luck investigating this one? The rc6 announcement sounds like there may not be that many more rc releases for 5.14.

Hi Jackub.

Unfortunately I have only a device with stmmac, and it works fine with the patch. It seems that not all hardware suffers from this issue.

Quoted 8 lines

Also, using NET_IP_ALIGN on RX is a common pattern, I think that any ethernet device is doing the same to align the IPv4 header.

Anyway, I asked for two tests on the affected device:
1. Change NET_IP_ALIGN with 8, to see if the DMA has problems in receiving to a non word aligned address
2. load a nop XDP program (I provided one), to see if the problem is already there when XDP is used

Catching up on email, as I was away earlier this week. I'm yet to test these two things (hopefully by the end of the day), but I just gave the patch below a go, just in case...

Quoted 37 lines

I doubt that changing also stmmac_rx_buf2_len would help, but it's worth a try, here is a patch:

drivers/net/ethernet/stmicro/stmmac/stmmac_main.c +3 -4
index 7b8404a21544..73d1f0ec66ff 100644
@@ -93,7 +93,7 @@ static int tc = TC_DEFAULT;
 module_param(tc, int, 0644);
 MODULE_PARM_DESC(tc, "DMA threshold control value");
 
-#define	DEFAULT_BUFSIZE	1536
+#define	DEFAULT_BUFSIZE	1536 + XDP_PACKET_HEADROOM + NET_IP_ALIGN
 static int buf_sz = DEFAULT_BUFSIZE;
 module_param(buf_sz, int, 0644);
 MODULE_PARM_DESC(buf_sz, "DMA buffer size");
@@ -4508,12 +4508,12 @@ static unsigned int stmmac_rx_buf1_len(struct stmmac_priv *priv,
 
 	/* First descriptor, not last descriptor and not split header */
 	if (status & rx_not_ls)
-		return priv->dma_buf_sz;
+		return priv->dma_buf_sz - stmmac_rx_offset(priv);
 
 	plen = stmmac_get_rx_frame_len(priv, p, coe);
 
 	/* First descriptor and last descriptor and not split header */
-	return min_t(unsigned int, priv->dma_buf_sz, plen);
+	return min_t(unsigned int, priv->dma_buf_sz - stmmac_rx_offset(priv), plen);
 }
 
 static unsigned int stmmac_rx_buf2_len(struct stmmac_priv *priv,
@@ -4529,12 +4529,12 @@ static unsigned int stmmac_rx_buf2_len(struct stmmac_priv *priv,
 
 	/* Not last descriptor */
 	if (status & rx_not_ls)
-		return priv->dma_buf_sz;

drivers/net/ethernet/stmicro/stmmac/stmmac_main.c static unsigned int stmmac_rx_buf2_len(struct stmmac_priv *priv, · line 4540

Quoted 10 lines
drivers/net/ethernet/stmicro/stmmac/stmmac_main.c +2 -1
+		return priv->dma_buf_sz - stmmac_rx_offset(priv);
 
 	plen = stmmac_get_rx_frame_len(priv, p, coe);
 
 	/* Last descriptor */
-	return plen - len;
+	return plen - len - stmmac_rx_offset(priv);
 }
 
 static int stmmac_xdp_xmit_xdpf(struct stmmac_priv *priv, int queue,

This patch results in a cosmic explosion, as it tries to invalidate a range of VAs that cannot possibly be allocated to the DMA. Not really what was expected.

Thanks,

	M.
[   19.587499] Unable to handle kernel write to read-only memory at virtual address ffff000088f89000
[   19.596375] Mem abort info:
[   19.599165]   ESR = 0x9600014f
[   19.602215]   EC = 0x25: DABT (current EL), IL = 32 bits
[   19.607519]   SET = 0, FnV = 0
[   19.610568]   EA = 0, S1PTW = 0
[   19.613703]   FSC = 0x0f: level 3 permission fault
[   19.618487] Data abort info:
[   19.621362]   ISV = 0, ISS = 0x0000014f
[   19.625192]   CM = 1, WnR = 1
[   19.628154] swapper pgtable: 4k pages, 48-bit VAs, pgdp=00000000fa7b2000
[   19.634844] [ffff000088f89000] pgd=18000002771f7003, p4d=18000002771f7003, pud=1800000275ffd003, pmd=1800000275fb5003, pte=0060000108f89f87
[   19.647358] Internal error: Oops: 9600014f [#1] PREEMPT SMP
[   19.652920] Modules linked in: nls_ascii(E) nls_cp437(E) vfat(E) fat(E) tegra_drm(E) cec(E) drm_kms_helper(E) evdev(E) snd_hda_codec_hdmi(E) snd_hda_tegra(E) snd_hda_codec(E) aes_ce_blk(E) crypto_simd(E) snd_hda_core(E) cryptd(E) snd_hwdep(E) at24(E) aes_ce_cipher(E) snd_pcm(E) ghash_ce(E) gf128mul(E) snd_timer(E) sha2_ce(E) sha256_arm64(E) sha1_ce(E) snd(E) soundcore(E) tegra_bpmp_thermal(E) efi_pstore(E) tegra_xudc(E) udc_core(E) host1x(E) ina3221(E) fuse(E) drm(E) configfs(E) efivarfs(E) ip_tables(E) x_tables(E) autofs4(E) nvme(E) nvme_core(E) t10_pi(E) broadcom(E) bcm_phy_lib(E) ahci_tegra(E) libahci_platform(E) gpio_keys(E) libahci(E) sdhci_tegra(E) dwmac_dwc_qos_eth(E) stmmac_platform(E) libata(E) stmmac(E) sdhci_pltfm(E) pcs_xpcs(E) max77620_regulator(E) xhci_tegra(E) phylink(E) of_mdio(E) cqhci(E) scsi_mod(E) fixed_phy(E) sdhci(E) fwnode_mdio(E) libphy(E)
[   19.729311] CPU: 0 PID: 0 Comm: swapper/0 Tainted: G S          E     5.14.0-rc6-dirty #4155
[   19.737734] Hardware name:  , BIOS 2021.04-rc2-00042-g2dddc1bb29 02/18/2021
[   19.744679] pstate: 80000005 (Nzcv daif -PAN -UAO -TCO BTYPE=--)
[   19.750673] pc : dcache_inval_poc+0x40/0x58
[   19.754852] lr : arch_sync_dma_for_cpu+0x2c/0x40
[   19.759460] sp : ffff800010003be0
[   19.762765] x29: ffff800010003be0 x28: ffff000085338000 x27: 0000000000000028
[   19.769889] x26: ffff000080a8fc00 x25: ffff000085338028 x24: 0000000000000000
[   19.777012] x23: 0000000000000002 x22: 0000000000000002 x21: 00000000ffffffbc
[   19.784135] x20: 0000000108f43000 x19: ffff000080223010 x18: 0000000000000000
[   19.791258] x17: ffff8001e3815000 x16: ffff800010004000 x15: 0000000000004000
[   19.798380] x14: 0000000000000000 x13: 0000000080000000 x12: 0000000000000001
[   19.805504] x11: 0000000000000004 x10: 0000000000000008 x9 : ffff8000109b9a2c
[   19.812626] x8 : 0000000000000000 x7 : 0000000000000001 x6 : ffff000088246500
[   19.819748] x5 : fffffffffffff000 x4 : 0003000108f40000 x3 : 000000000000003f
[   19.826869] x2 : 0000000000000040 x1 : ffff000188f42f80 x0 : ffff000088f89000
[   19.833994] Call trace:
[   19.836434]  dcache_inval_poc+0x40/0x58
[   19.840262]  iommu_dma_sync_single_for_cpu+0xdc/0xe0
[   19.845219]  dma_sync_single_for_cpu+0x3c/0x124
[   19.849742]  stmmac_rx+0x448/0x930 [stmmac]
[   19.853937]  stmmac_napi_poll_rx+0x4c/0xec [stmmac]
[   19.858818]  __napi_poll+0x40/0x210
[   19.862300]  net_rx_action+0x304/0x370
[   19.866041]  __do_softirq+0x130/0x3d8
[   19.869695]  __irq_exit_rcu+0x100/0x110
[   19.873525]  irq_exit+0x1c/0x30
[   19.876658]  handle_domain_irq+0x70/0x9c
[   19.880573]  gic_handle_irq+0x58/0xe0
[   19.884225]  call_on_irq_stack+0x2c/0x54
[   19.888138]  do_interrupt_handler+0x5c/0x70
[   19.892313]  el1_interrupt+0x30/0x70
[   19.895881]  el1h_64_irq_handler+0x18/0x24
[   19.899970]  el1h_64_irq+0x78/0x7c
[   19.903361]  arch_cpu_idle+0x18/0x3c
[   19.906930]  default_idle_call+0x4c/0x1d4
[   19.910931]  cpuidle_idle_call+0x168/0x1e0
[   19.915019]  do_idle+0xb8/0x110
[   19.918152]  cpu_startup_entry+0x30/0x8c
[   19.922065]  rest_init+0xf0/0x100
[   19.925372]  arch_call_rest_init+0x1c/0x28
[   19.929463]  start_kernel+0x4a0/0x4d8
[   19.933114]  __primary_switched+0xc0/0xc8
[   19.937118] Code: 8a230000 54000060 d50b7e20 14000002 (d5087620) 
[   19.943204] ---[ end trace f26fd0d14eea9a7d ]---
[   19.947811] Kernel panic - not syncing: Oops: Fatal exception in interrupt
[   19.954670] SMP: stopping secondary CPUs
[   19.958592] Kernel Offset: 0x1a0000 from 0xffff800010000000
[   19.964151] PHYS_OFFSET: 0x80000000
[   19.967630] CPU features: 0x10000231,00000846
[   19.971976] Memory Limit: none
[   19.975026] ---[ end Kernel panic - not syncing: Oops: Fatal exception in interrupt ]---
Signature · 1 line
Without deviation from the norm, progress is not possible.
#13 lore raw
In reply to Matteo Croce
Quoted 71 lines · Matteo Croce

[1 <text/plain; UTF-8 (7bit)>]

Quoted 52 lines · Marc Zyngier
Quoted 49 lines · Matteo Croce
Quoted 44 lines · On Thu, 12 Aug 2021 10:48:03 +0200 Eric Dumazet <[email protected]> wrote:
Quoted 29 lines · Marc Zyngier
Quoted 26 lines · Eric Dumazet

Are you sure you do not need to adjust stmmac_set_bfsize(), stmmac_rx_buf1_len() and stmmac_rx_buf2_len() ?

Presumably DEFAULT_BUFSIZE also want to be increased by NET_SKB_PAD

Patch for stmmac_rx_buf1_len() :

(quoted diff) +0 -0

b/drivers/net/ethernet/stmicro/stmmac/stmmac_main.c index 7b8404a21544cf29668e8a14240c3971e6bce0c3..041a74e7efca3436bfe3e17f972dd156173957a9 100644 --- a/drivers/net/ethernet/stmicro/stmmac/stmmac_main.c +++ b/drivers/net/ethernet/stmicro/stmmac/stmmac_main.c @@ -4508,12 +4508,12 @@ static unsigned int stmmac_rx_buf1_len(struct
stmmac_priv *priv, /* First descriptor, not last descriptor and
not split header */ if (status & rx_not_ls)

(quoted diff) +1 -1
-               return priv->dma_buf_sz;
+               return priv->dma_buf_sz - NET_SKB_PAD -
NET_IP_ALIGN;
        plen = stmmac_get_rx_frame_len(priv, p, coe);

        /* First descriptor and last descriptor and not split

header */

(quoted diff) +1 -2
-       return min_t(unsigned int, priv->dma_buf_sz, plen);
+       return min_t(unsigned int, priv->dma_buf_sz - NET_SKB_PAD
- NET_IP_ALIGN, plen); }
Quoted 4 lines
Quoted 1 line

static unsigned int stmmac_rx_buf2_len(struct stmmac_priv *priv,

Feels like a major deficiency of the original patch. Happy to test a more complete patch if/when you have one.

I wont have time in the immediate future.

Matteo, if you do not work on a fix, I suggest we revert
a955318fe67ec0d962760b5ee58e74bffaf649b8 stmmac: align RX buffers

before a more polished version can be submitted.

Quoted 9 lines
Quoted 5 lines

Better to use stmmac_rx_offset() so to have the correct length when using XDP. Also, when XDP is enabled, the offset was XDP_PACKET_HEADROOM (i.e. 256 bytes) even before the change, so it could be already broken. Mark, can you try on the Jetson TX2 by attaching an XDP program and see if it works without my change?

Sorry, you'll have to hold my hand here, as I know exactly nothing about XDP....

Attach the attached object with:

ip link set eth0 xdp object passall.o

This is an empty XDP program, its source:

Quoted 8 lines

__attribute__((section("prog"), used))
int xdp_main(struct xdp_md *ctx)

{
       return XDP_PASS;
}

Every packet will pass untouched, but the offset will be shifted from 0 to 256 bytes, which could trigger the problem anyway:

Nope. On 5.13, which doesn't have the issue, adding this payload doesn't result in any problem and the whole thing is rock solid.

Quoted 9 lines
Quoted 9 lines
Quoted 4 lines

A possible fix, which takes in account also the XDP headroom for stmmac_rx_buf1_len() only could be (only compile tested, I don't have the hardware now):

However, this doesn't fix my issue. I still get all sort of corruption. Probably stmmac_rx_buf2_len() also need adjusting (it has a similar logic as its buf1 counterpart...)

Quoted 6 lines
Quoted 3 lines

Unless you can fix it very quickly, and given that we're towards the end of the cycle, I'd be more comfortable if we reverted this patch.

Can it be that the HW can't do DMA on an address which is not word aligned? What if you replace NET_SKB_PAD with, let's say, 8?

With this:

drivers/net/ethernet/stmicro/stmmac/stmmac.h +1 -1
index fcdb1d20389b..244aa6579ef4 100644
@@ -341,7 +341,7 @@ static inline unsigned int stmmac_rx_offset(struct stmmac_priv *priv)
 	if (stmmac_xdp_is_enabled(priv))
 		return XDP_PACKET_HEADROOM + NET_IP_ALIGN;
 
-	return NET_SKB_PAD + NET_IP_ALIGN;
+	return 8 + NET_IP_ALIGN;
 }
 
 void stmmac_disable_rx_queue(struct stmmac_priv *priv, u32 queue);

I don't see the system corrupting packets anymore. Is that exactly what you had in mind? This really seems to point to a basic buffer overflow.

Thanks,

	M.
Signature · 1 line
Without deviation from the norm, progress is not possible.
#14 lore raw
Quoted 122 lines · Marc Zyngier
Quoted 71 lines · Matteo Croce

[1 <text/plain; UTF-8 (7bit)>]

Quoted 52 lines · Marc Zyngier
Quoted 49 lines · Matteo Croce
Quoted 44 lines · On Thu, 12 Aug 2021 10:48:03 +0200 Eric Dumazet <[email protected]> wrote:
Quoted 29 lines · Marc Zyngier
Quoted 26 lines · Eric Dumazet

Are you sure you do not need to adjust stmmac_set_bfsize(), stmmac_rx_buf1_len() and stmmac_rx_buf2_len() ?

Presumably DEFAULT_BUFSIZE also want to be increased by NET_SKB_PAD

Patch for stmmac_rx_buf1_len() :

(quoted diff) +0 -0

b/drivers/net/ethernet/stmicro/stmmac/stmmac_main.c index 7b8404a21544cf29668e8a14240c3971e6bce0c3..041a74e7efca3436bfe3e17f972dd156173957a9 100644 --- a/drivers/net/ethernet/stmicro/stmmac/stmmac_main.c +++ b/drivers/net/ethernet/stmicro/stmmac/stmmac_main.c @@ -4508,12 +4508,12 @@ static unsigned int stmmac_rx_buf1_len(struct
stmmac_priv *priv, /* First descriptor, not last descriptor and
not split header */ if (status & rx_not_ls)

(quoted diff) +1 -1
-               return priv->dma_buf_sz;
+               return priv->dma_buf_sz - NET_SKB_PAD -
NET_IP_ALIGN;
        plen = stmmac_get_rx_frame_len(priv, p, coe);

        /* First descriptor and last descriptor and not split

header */

(quoted diff) +1 -2
-       return min_t(unsigned int, priv->dma_buf_sz, plen);
+       return min_t(unsigned int, priv->dma_buf_sz - NET_SKB_PAD
- NET_IP_ALIGN, plen); }
Quoted 4 lines
Quoted 1 line

static unsigned int stmmac_rx_buf2_len(struct stmmac_priv *priv,

Feels like a major deficiency of the original patch. Happy to test a more complete patch if/when you have one.

I wont have time in the immediate future.

Matteo, if you do not work on a fix, I suggest we revert
a955318fe67ec0d962760b5ee58e74bffaf649b8 stmmac: align RX buffers

before a more polished version can be submitted.

Quoted 9 lines
Quoted 5 lines

Better to use stmmac_rx_offset() so to have the correct length when using XDP. Also, when XDP is enabled, the offset was XDP_PACKET_HEADROOM (i.e. 256 bytes) even before the change, so it could be already broken. Mark, can you try on the Jetson TX2 by attaching an XDP program and see if it works without my change?

Sorry, you'll have to hold my hand here, as I know exactly nothing about XDP....

Attach the attached object with:

ip link set eth0 xdp object passall.o

This is an empty XDP program, its source:

Quoted 8 lines

__attribute__((section("prog"), used))
int xdp_main(struct xdp_md *ctx)

{
       return XDP_PASS;
}

Every packet will pass untouched, but the offset will be shifted from 0 to 256 bytes, which could trigger the problem anyway:

Nope. On 5.13, which doesn't have the issue, adding this payload doesn't result in any problem and the whole thing is rock solid.

Quoted 9 lines
Quoted 9 lines
Quoted 4 lines

A possible fix, which takes in account also the XDP headroom for stmmac_rx_buf1_len() only could be (only compile tested, I don't have the hardware now):

However, this doesn't fix my issue. I still get all sort of corruption. Probably stmmac_rx_buf2_len() also need adjusting (it has a similar logic as its buf1 counterpart...)

Quoted 6 lines
Quoted 3 lines

Unless you can fix it very quickly, and given that we're towards the end of the cycle, I'd be more comfortable if we reverted this patch.

Can it be that the HW can't do DMA on an address which is not word aligned? What if you replace NET_SKB_PAD with, let's say, 8?

With this:

drivers/net/ethernet/stmicro/stmmac/stmmac.h +1 -1
index fcdb1d20389b..244aa6579ef4 100644
@@ -341,7 +341,7 @@ static inline unsigned int stmmac_rx_offset(struct stmmac_priv *priv)
        if (stmmac_xdp_is_enabled(priv))
                return XDP_PACKET_HEADROOM + NET_IP_ALIGN;
 
-       return NET_SKB_PAD + NET_IP_ALIGN;
+       return 8 + NET_IP_ALIGN;
 }
 
 void stmmac_disable_rx_queue(struct stmmac_priv *priv, u32 queue);

I don't see the system corrupting packets anymore. Is that exactly what you had in mind? This really seems to point to a basic buffer overflow.

Quoted 6 lines

Thanks,

        M.

--
Without deviation from the norm, progress is not possible.

Sorry, I meant something like:

- return NET_SKB_PAD + NET_IP_ALIGN;
+ return 8;

I had some hardware which DMA fails if the receive buffer was not word aligned, but this seems not the case, as 8 + NET_IP_ALIGN = 10, and it's not aligned too.

Signature · 1 line
per aspera ad upstream
#15 lore raw
Quoted 2 lines · Matteo Croce

On Thu, Aug 19, 2021 at 6:29 PM Marc Zyngier <[email protected]> wrote:

[...]

Quoted 17 lines
Quoted 17 lines
drivers/net/ethernet/stmicro/stmmac/stmmac.h +1 -1
index fcdb1d20389b..244aa6579ef4 100644
@@ -341,7 +341,7 @@ static inline unsigned int stmmac_rx_offset(struct stmmac_priv *priv)
        if (stmmac_xdp_is_enabled(priv))
                return XDP_PACKET_HEADROOM + NET_IP_ALIGN;
 
-       return NET_SKB_PAD + NET_IP_ALIGN;
+       return 8 + NET_IP_ALIGN;
 }
 
 void stmmac_disable_rx_queue(struct stmmac_priv *priv, u32 queue);

I don't see the system corrupting packets anymore. Is that exactly what you had in mind? This really seems to point to a basic buffer overflow.

[...]

Quoted 8 lines

Sorry, I meant something like:

(quoted diff) +1 -1
-       return NET_SKB_PAD + NET_IP_ALIGN;
+       return 8;

I had some hardware which DMA fails if the receive buffer was not word aligned, but this seems not the case, as 8 + NET_IP_ALIGN = 10, and it's not aligned too.

No error in that case either, as expected. Given that NET_SKB_PAD is likely to expand to 64, it is likely a DMA buffer overflow which probably only triggers for large-ish packets.

Now, we're almost at -rc7, and we don't have a solution in sight.

Can we please revert this until we have an understanding of what is happening? I'll hopefully have more cycles to work on the issue once 5.14 is out, and hopefully the maintainers of this driver can chime in (they have been pretty quiet so far).

Thanks,

	M.
Signature · 1 line
Without deviation from the norm, progress is not possible.
#16 lore raw
Quoted 48 lines · Marc Zyngier
Quoted 2 lines · Matteo Croce

On Thu, Aug 19, 2021 at 6:29 PM Marc Zyngier <[email protected]> wrote:

[...]

Quoted 17 lines
Quoted 17 lines
drivers/net/ethernet/stmicro/stmmac/stmmac.h +1 -1
index fcdb1d20389b..244aa6579ef4 100644
@@ -341,7 +341,7 @@ static inline unsigned int stmmac_rx_offset(struct stmmac_priv *priv)
        if (stmmac_xdp_is_enabled(priv))
                return XDP_PACKET_HEADROOM + NET_IP_ALIGN;
 
-       return NET_SKB_PAD + NET_IP_ALIGN;
+       return 8 + NET_IP_ALIGN;
 }
 
 void stmmac_disable_rx_queue(struct stmmac_priv *priv, u32 queue);

I don't see the system corrupting packets anymore. Is that exactly what you had in mind? This really seems to point to a basic buffer overflow.

[...]

Quoted 8 lines

Sorry, I meant something like:

(quoted diff) +1 -1
-       return NET_SKB_PAD + NET_IP_ALIGN;
+       return 8;

I had some hardware which DMA fails if the receive buffer was not word aligned, but this seems not the case, as 8 + NET_IP_ALIGN = 10, and it's not aligned too.

No error in that case either, as expected. Given that NET_SKB_PAD is likely to expand to 64, it is likely a DMA buffer overflow which probably only triggers for large-ish packets.

Now, we're almost at -rc7, and we don't have a solution in sight.

Can we please revert this until we have an understanding of what is happening? I'll hopefully have more cycles to work on the issue once 5.14 is out, and hopefully the maintainers of this driver can chime in (they have been pretty quiet so far).

Quoted 6 lines

Thanks,

        M.

--
Without deviation from the norm, progress is not possible.

Last try, what about adding only NET_IP_ALIGN and leaving NET_SKB_PAD?

- return NET_SKB_PAD + NET_IP_ALIGN;
+ return NET_IP_ALIGN;

I think that alloc_skb adds another NET_SKB_PAD anyway.

Signature · 1 line
per aspera ad upstream
#17 lore raw
Quoted 62 lines · Matteo Croce
Quoted 48 lines · Marc Zyngier
Quoted 2 lines · Matteo Croce

On Thu, Aug 19, 2021 at 6:29 PM Marc Zyngier <[email protected]> wrote:

[...]

Quoted 17 lines
Quoted 17 lines
drivers/net/ethernet/stmicro/stmmac/stmmac.h +1 -1
index fcdb1d20389b..244aa6579ef4 100644
@@ -341,7 +341,7 @@ static inline unsigned int stmmac_rx_offset(struct stmmac_priv *priv)
        if (stmmac_xdp_is_enabled(priv))
                return XDP_PACKET_HEADROOM + NET_IP_ALIGN;
 
-       return NET_SKB_PAD + NET_IP_ALIGN;
+       return 8 + NET_IP_ALIGN;
 }
 
 void stmmac_disable_rx_queue(struct stmmac_priv *priv, u32 queue);

I don't see the system corrupting packets anymore. Is that exactly what you had in mind? This really seems to point to a basic buffer overflow.

[...]

Quoted 8 lines

Sorry, I meant something like:

(quoted diff) +1 -1
-       return NET_SKB_PAD + NET_IP_ALIGN;
+       return 8;

I had some hardware which DMA fails if the receive buffer was not word aligned, but this seems not the case, as 8 + NET_IP_ALIGN = 10, and it's not aligned too.

No error in that case either, as expected. Given that NET_SKB_PAD is likely to expand to 64, it is likely a DMA buffer overflow which probably only triggers for large-ish packets.

Now, we're almost at -rc7, and we don't have a solution in sight.

Can we please revert this until we have an understanding of what is happening? I'll hopefully have more cycles to work on the issue once 5.14 is out, and hopefully the maintainers of this driver can chime in (they have been pretty quiet so far).

Quoted 6 lines

Thanks,

        M.

--
Without deviation from the norm, progress is not possible.

Last try, what about adding only NET_IP_ALIGN and leaving NET_SKB_PAD?

(quoted diff) +1 -1
-       return NET_SKB_PAD + NET_IP_ALIGN;
+       return NET_IP_ALIGN;
Quoted 1 line

I think that alloc_skb adds another NET_SKB_PAD anyway.

I don't see any packet corruption with this. However, this doesn't prove that this is correct either. What was the rational for adding NET_SKB_PAD the first place?

	M.
Signature · 1 line
Without deviation from the norm, progress is not possible.
#18 lore raw
Quoted 62 lines · Marc Zyngier
Quoted 59 lines · Matteo Croce
Quoted 48 lines · Marc Zyngier
Quoted 2 lines · Matteo Croce

On Thu, Aug 19, 2021 at 6:29 PM Marc Zyngier <[email protected]> wrote:

[...]

Quoted 17 lines
Quoted 17 lines
drivers/net/ethernet/stmicro/stmmac/stmmac.h +1 -1
index fcdb1d20389b..244aa6579ef4 100644
@@ -341,7 +341,7 @@ static inline unsigned int stmmac_rx_offset(struct stmmac_priv *priv)
        if (stmmac_xdp_is_enabled(priv))
                return XDP_PACKET_HEADROOM + NET_IP_ALIGN;
 
-       return NET_SKB_PAD + NET_IP_ALIGN;
+       return 8 + NET_IP_ALIGN;
 }
 
 void stmmac_disable_rx_queue(struct stmmac_priv *priv, u32 queue);

I don't see the system corrupting packets anymore. Is that exactly what you had in mind? This really seems to point to a basic buffer overflow.

[...]

Quoted 8 lines

Sorry, I meant something like:

(quoted diff) +1 -1
-       return NET_SKB_PAD + NET_IP_ALIGN;
+       return 8;

I had some hardware which DMA fails if the receive buffer was not word aligned, but this seems not the case, as 8 + NET_IP_ALIGN = 10, and it's not aligned too.

No error in that case either, as expected. Given that NET_SKB_PAD is likely to expand to 64, it is likely a DMA buffer overflow which probably only triggers for large-ish packets.

Now, we're almost at -rc7, and we don't have a solution in sight.

Can we please revert this until we have an understanding of what is happening? I'll hopefully have more cycles to work on the issue once 5.14 is out, and hopefully the maintainers of this driver can chime in (they have been pretty quiet so far).

Quoted 6 lines

Thanks,

        M.

--
Without deviation from the norm, progress is not possible.

Last try, what about adding only NET_IP_ALIGN and leaving NET_SKB_PAD?

Quoted 9 lines
Quoted 4 lines
(quoted diff) +1 -1
-       return NET_SKB_PAD + NET_IP_ALIGN;
+       return NET_IP_ALIGN;

I think that alloc_skb adds another NET_SKB_PAD anyway.

I don't see any packet corruption with this. However, this doesn't prove that this is correct either. What was the rational for adding NET_SKB_PAD the first place?

I think it's wrong. The original offset was 0, and to align it to the boundary we need to add just NET_IP_ALIGN, which is two. NET_SKB_PAD is a much bigger value, (I think 64), which is used to reserve space to prepend an header, e.g. with tunnels.

Signature · 1 line
per aspera ad upstream
#19 lore raw
Quoted 74 lines · Matteo Croce
Quoted 62 lines · Marc Zyngier
Quoted 59 lines · Matteo Croce
Quoted 48 lines · Marc Zyngier
Quoted 2 lines · Matteo Croce

On Thu, Aug 19, 2021 at 6:29 PM Marc Zyngier <[email protected]> wrote:

[...]

Quoted 17 lines
Quoted 17 lines
drivers/net/ethernet/stmicro/stmmac/stmmac.h +1 -1
index fcdb1d20389b..244aa6579ef4 100644
@@ -341,7 +341,7 @@ static inline unsigned int stmmac_rx_offset(struct stmmac_priv *priv)
        if (stmmac_xdp_is_enabled(priv))
                return XDP_PACKET_HEADROOM + NET_IP_ALIGN;
 
-       return NET_SKB_PAD + NET_IP_ALIGN;
+       return 8 + NET_IP_ALIGN;
 }
 
 void stmmac_disable_rx_queue(struct stmmac_priv *priv, u32 queue);

I don't see the system corrupting packets anymore. Is that exactly what you had in mind? This really seems to point to a basic buffer overflow.

[...]

Quoted 8 lines

Sorry, I meant something like:

(quoted diff) +1 -1
-       return NET_SKB_PAD + NET_IP_ALIGN;
+       return 8;

I had some hardware which DMA fails if the receive buffer was not word aligned, but this seems not the case, as 8 + NET_IP_ALIGN = 10, and it's not aligned too.

No error in that case either, as expected. Given that NET_SKB_PAD is likely to expand to 64, it is likely a DMA buffer overflow which probably only triggers for large-ish packets.

Now, we're almost at -rc7, and we don't have a solution in sight.

Can we please revert this until we have an understanding of what is happening? I'll hopefully have more cycles to work on the issue once 5.14 is out, and hopefully the maintainers of this driver can chime in (they have been pretty quiet so far).

Quoted 6 lines

Thanks,

        M.

--
Without deviation from the norm, progress is not possible.

Last try, what about adding only NET_IP_ALIGN and leaving NET_SKB_PAD?

Quoted 9 lines
Quoted 4 lines
(quoted diff) +1 -1
-       return NET_SKB_PAD + NET_IP_ALIGN;
+       return NET_IP_ALIGN;

I think that alloc_skb adds another NET_SKB_PAD anyway.

I don't see any packet corruption with this. However, this doesn't prove that this is correct either. What was the rational for adding NET_SKB_PAD the first place?

Quoted 4 lines

I think it's wrong. The original offset was 0, and to align it to the boundary we need to add just NET_IP_ALIGN, which is two. NET_SKB_PAD is a much bigger value, (I think 64), which is used to reserve space to prepend an header, e.g. with tunnels.

How about the other adjustments that Eric mentioned regarding the size of the buffer? Aren't they required?

	M.
Signature · 1 line
Without deviation from the norm, progress is not possible.
#20 lore raw
Quoted 77 lines · Marc Zyngier
Quoted 74 lines · Matteo Croce
Quoted 71 lines · Marc Zyngier
Quoted 62 lines · Matteo Croce
Quoted 48 lines · Marc Zyngier
Quoted 2 lines · Matteo Croce

On Thu, Aug 19, 2021 at 6:29 PM Marc Zyngier <[email protected]> wrote:

[...]

Quoted 17 lines
Quoted 17 lines
drivers/net/ethernet/stmicro/stmmac/stmmac.h +1 -1
index fcdb1d20389b..244aa6579ef4 100644
@@ -341,7 +341,7 @@ static inline unsigned int stmmac_rx_offset(struct stmmac_priv *priv)
        if (stmmac_xdp_is_enabled(priv))
                return XDP_PACKET_HEADROOM + NET_IP_ALIGN;
 
-       return NET_SKB_PAD + NET_IP_ALIGN;
+       return 8 + NET_IP_ALIGN;
 }
 
 void stmmac_disable_rx_queue(struct stmmac_priv *priv, u32 queue);

I don't see the system corrupting packets anymore. Is that exactly what you had in mind? This really seems to point to a basic buffer overflow.

[...]

Quoted 8 lines

Sorry, I meant something like:

(quoted diff) +1 -1
-       return NET_SKB_PAD + NET_IP_ALIGN;
+       return 8;

I had some hardware which DMA fails if the receive buffer was not word aligned, but this seems not the case, as 8 + NET_IP_ALIGN = 10, and it's not aligned too.

No error in that case either, as expected. Given that NET_SKB_PAD is likely to expand to 64, it is likely a DMA buffer overflow which probably only triggers for large-ish packets.

Now, we're almost at -rc7, and we don't have a solution in sight.

Can we please revert this until we have an understanding of what is happening? I'll hopefully have more cycles to work on the issue once 5.14 is out, and hopefully the maintainers of this driver can chime in (they have been pretty quiet so far).

Quoted 6 lines

Thanks,

        M.

--
Without deviation from the norm, progress is not possible.

Last try, what about adding only NET_IP_ALIGN and leaving NET_SKB_PAD?

(quoted diff) +1 -1
-       return NET_SKB_PAD + NET_IP_ALIGN;
+       return NET_IP_ALIGN;
Quoted 1 line

I think that alloc_skb adds another NET_SKB_PAD anyway.

I don't see any packet corruption with this. However, this doesn't prove that this is correct either. What was the rational for adding NET_SKB_PAD the first place?

Quoted 8 lines
Quoted 4 lines

I think it's wrong. The original offset was 0, and to align it to the boundary we need to add just NET_IP_ALIGN, which is two. NET_SKB_PAD is a much bigger value, (I think 64), which is used to reserve space to prepend an header, e.g. with tunnels.

How about the other adjustments that Eric mentioned regarding the size of the buffer? Aren't they required?

I guess that if stmmac_rx_buf1_len() needed such adjustment, it would be already broken when XDP is in use.
When you use XDP, stmmac_rx_offset() adds a pretty big headroom of 256 byte, which would easily trigger an overflow if not accounted. Did you try attaching a simple XDP program on a stock 5.13 kernel?

Signature · 1 line
per aspera ad upstream
#21 lore raw
Quoted 88 lines · Matteo Croce
Quoted 77 lines · Marc Zyngier
Quoted 74 lines · Matteo Croce
Quoted 71 lines · Marc Zyngier
Quoted 62 lines · Matteo Croce
Quoted 48 lines · Marc Zyngier
Quoted 2 lines · Matteo Croce

On Thu, Aug 19, 2021 at 6:29 PM Marc Zyngier <[email protected]> wrote:

[...]

Quoted 17 lines
Quoted 17 lines
drivers/net/ethernet/stmicro/stmmac/stmmac.h +1 -1
index fcdb1d20389b..244aa6579ef4 100644
@@ -341,7 +341,7 @@ static inline unsigned int stmmac_rx_offset(struct stmmac_priv *priv)
        if (stmmac_xdp_is_enabled(priv))
                return XDP_PACKET_HEADROOM + NET_IP_ALIGN;
 
-       return NET_SKB_PAD + NET_IP_ALIGN;
+       return 8 + NET_IP_ALIGN;
 }
 
 void stmmac_disable_rx_queue(struct stmmac_priv *priv, u32 queue);

I don't see the system corrupting packets anymore. Is that exactly what you had in mind? This really seems to point to a basic buffer overflow.

[...]

Quoted 8 lines

Sorry, I meant something like:

(quoted diff) +1 -1
-       return NET_SKB_PAD + NET_IP_ALIGN;
+       return 8;

I had some hardware which DMA fails if the receive buffer was not word aligned, but this seems not the case, as 8 + NET_IP_ALIGN = 10, and it's not aligned too.

No error in that case either, as expected. Given that NET_SKB_PAD is likely to expand to 64, it is likely a DMA buffer overflow which probably only triggers for large-ish packets.

Now, we're almost at -rc7, and we don't have a solution in sight.

Can we please revert this until we have an understanding of what is happening? I'll hopefully have more cycles to work on the issue once 5.14 is out, and hopefully the maintainers of this driver can chime in (they have been pretty quiet so far).

Quoted 6 lines

Thanks,

        M.

--
Without deviation from the norm, progress is not possible.

Last try, what about adding only NET_IP_ALIGN and leaving NET_SKB_PAD?

(quoted diff) +1 -1
-       return NET_SKB_PAD + NET_IP_ALIGN;
+       return NET_IP_ALIGN;
Quoted 1 line

I think that alloc_skb adds another NET_SKB_PAD anyway.

I don't see any packet corruption with this. However, this doesn't prove that this is correct either. What was the rational for adding NET_SKB_PAD the first place?

Quoted 8 lines
Quoted 4 lines

I think it's wrong. The original offset was 0, and to align it to the boundary we need to add just NET_IP_ALIGN, which is two. NET_SKB_PAD is a much bigger value, (I think 64), which is used to reserve space to prepend an header, e.g. with tunnels.

How about the other adjustments that Eric mentioned regarding the size of the buffer? Aren't they required?

Quoted 5 lines

I guess that if stmmac_rx_buf1_len() needed such adjustment, it would be already broken when XDP is in use.
When you use XDP, stmmac_rx_offset() adds a pretty big headroom of 256 byte, which would easily trigger an overflow if not accounted. Did you try attaching a simple XDP program on a stock 5.13 kernel?

Yes, as mentioned in [1], to which you replied...

	M.

[1] https://lore.kernel.org/r/[email protected]

Signature · 1 line
Without deviation from the norm, progress is not possible.
#22 lore raw
Quoted 97 lines · Marc Zyngier
Quoted 88 lines · Matteo Croce
Quoted 77 lines · Marc Zyngier
Quoted 74 lines · Matteo Croce
Quoted 71 lines · Marc Zyngier
Quoted 62 lines · Matteo Croce
Quoted 48 lines · Marc Zyngier
Quoted 2 lines · Matteo Croce

On Thu, Aug 19, 2021 at 6:29 PM Marc Zyngier <[email protected]> wrote:

[...]

Quoted 17 lines
Quoted 17 lines
drivers/net/ethernet/stmicro/stmmac/stmmac.h +1 -1
index fcdb1d20389b..244aa6579ef4 100644
@@ -341,7 +341,7 @@ static inline unsigned int stmmac_rx_offset(struct stmmac_priv *priv)
        if (stmmac_xdp_is_enabled(priv))
                return XDP_PACKET_HEADROOM + NET_IP_ALIGN;
 
-       return NET_SKB_PAD + NET_IP_ALIGN;
+       return 8 + NET_IP_ALIGN;
 }
 
 void stmmac_disable_rx_queue(struct stmmac_priv *priv, u32 queue);

I don't see the system corrupting packets anymore. Is that exactly what you had in mind? This really seems to point to a basic buffer overflow.

[...]

Quoted 8 lines

Sorry, I meant something like:

(quoted diff) +1 -1
-       return NET_SKB_PAD + NET_IP_ALIGN;
+       return 8;

I had some hardware which DMA fails if the receive buffer was not word aligned, but this seems not the case, as 8 + NET_IP_ALIGN = 10, and it's not aligned too.

No error in that case either, as expected. Given that NET_SKB_PAD is likely to expand to 64, it is likely a DMA buffer overflow which probably only triggers for large-ish packets.

Now, we're almost at -rc7, and we don't have a solution in sight.

Can we please revert this until we have an understanding of what is happening? I'll hopefully have more cycles to work on the issue once 5.14 is out, and hopefully the maintainers of this driver can chime in (they have been pretty quiet so far).

Quoted 6 lines

Thanks,

        M.

--
Without deviation from the norm, progress is not possible.

Last try, what about adding only NET_IP_ALIGN and leaving NET_SKB_PAD?

(quoted diff) +1 -1
-       return NET_SKB_PAD + NET_IP_ALIGN;
+       return NET_IP_ALIGN;
Quoted 1 line

I think that alloc_skb adds another NET_SKB_PAD anyway.

I don't see any packet corruption with this. However, this doesn't prove that this is correct either. What was the rational for adding NET_SKB_PAD the first place?

Quoted 8 lines
Quoted 4 lines

I think it's wrong. The original offset was 0, and to align it to the boundary we need to add just NET_IP_ALIGN, which is two. NET_SKB_PAD is a much bigger value, (I think 64), which is used to reserve space to prepend an header, e.g. with tunnels.

How about the other adjustments that Eric mentioned regarding the size of the buffer? Aren't they required?

Quoted 5 lines

I guess that if stmmac_rx_buf1_len() needed such adjustment, it would be already broken when XDP is in use.
When you use XDP, stmmac_rx_offset() adds a pretty big headroom of 256 byte, which would easily trigger an overflow if not accounted. Did you try attaching a simple XDP program on a stock 5.13 kernel?

Quoted 6 lines

Yes, as mentioned in [1], to which you replied...

        M.

[1] https://lore.kernel.org/r/[email protected]

Great.
So I doubt that the adjustment is needed.
Does it work with all the frame size?

Signature · 1 line
per aspera ad upstream
#23 lore raw
Quoted 101 lines · Matteo Croce
Quoted 99 lines · Marc Zyngier
Quoted 88 lines · Matteo Croce
Quoted 77 lines · Marc Zyngier
Quoted 74 lines · Matteo Croce
Quoted 71 lines · Marc Zyngier
Quoted 62 lines · Matteo Croce
Quoted 48 lines · Marc Zyngier
Quoted 2 lines · Matteo Croce

On Thu, Aug 19, 2021 at 6:29 PM Marc Zyngier <[email protected]> wrote:

[...]

Quoted 17 lines
Quoted 17 lines
drivers/net/ethernet/stmicro/stmmac/stmmac.h +1 -1
index fcdb1d20389b..244aa6579ef4 100644
@@ -341,7 +341,7 @@ static inline unsigned int stmmac_rx_offset(struct stmmac_priv *priv)
        if (stmmac_xdp_is_enabled(priv))
                return XDP_PACKET_HEADROOM + NET_IP_ALIGN;
 
-       return NET_SKB_PAD + NET_IP_ALIGN;
+       return 8 + NET_IP_ALIGN;
 }
 
 void stmmac_disable_rx_queue(struct stmmac_priv *priv, u32 queue);

I don't see the system corrupting packets anymore. Is that exactly what you had in mind? This really seems to point to a basic buffer overflow.

[...]

Quoted 8 lines

Sorry, I meant something like:

(quoted diff) +1 -1
-       return NET_SKB_PAD + NET_IP_ALIGN;
+       return 8;

I had some hardware which DMA fails if the receive buffer was not word aligned, but this seems not the case, as 8 + NET_IP_ALIGN = 10, and it's not aligned too.

No error in that case either, as expected. Given that NET_SKB_PAD is likely to expand to 64, it is likely a DMA buffer overflow which probably only triggers for large-ish packets.

Now, we're almost at -rc7, and we don't have a solution in sight.

Can we please revert this until we have an understanding of what is happening? I'll hopefully have more cycles to work on the issue once 5.14 is out, and hopefully the maintainers of this driver can chime in (they have been pretty quiet so far).

Quoted 6 lines

Thanks,

        M.

--
Without deviation from the norm, progress is not possible.

Last try, what about adding only NET_IP_ALIGN and leaving NET_SKB_PAD?

(quoted diff) +1 -1
-       return NET_SKB_PAD + NET_IP_ALIGN;
+       return NET_IP_ALIGN;
Quoted 1 line

I think that alloc_skb adds another NET_SKB_PAD anyway.

I don't see any packet corruption with this. However, this doesn't prove that this is correct either. What was the rational for adding NET_SKB_PAD the first place?

Quoted 8 lines
Quoted 4 lines

I think it's wrong. The original offset was 0, and to align it to the boundary we need to add just NET_IP_ALIGN, which is two. NET_SKB_PAD is a much bigger value, (I think 64), which is used to reserve space to prepend an header, e.g. with tunnels.

How about the other adjustments that Eric mentioned regarding the size of the buffer? Aren't they required?

Quoted 5 lines

I guess that if stmmac_rx_buf1_len() needed such adjustment, it would be already broken when XDP is in use.
When you use XDP, stmmac_rx_offset() adds a pretty big headroom of 256 byte, which would easily trigger an overflow if not accounted. Did you try attaching a simple XDP program on a stock 5.13 kernel?

Yes, as mentioned in [1], to which you replied...

Quoted 9 lines
Quoted 4 lines
        M.

[1] https://lore.kernel.org/r/[email protected]

Great.
So I doubt that the adjustment is needed.
Does it work with all the frame size?

Last check, are you sure that the bpf program was loaded in the driver and not as generic XDP?
You can force it as native with "xdpdrv":

ip link set eth xdpdrv object kernel_passall.o

Signature · 1 line
per aspera ad upstream
#25 lore raw
In reply to Matteo Croce
Quoted 23 lines · Matteo Croce
Quoted 21 lines · Marc Zyngier
Quoted 10 lines · Matteo Croce
Quoted 9 lines
Quoted 5 lines

I think it's wrong. The original offset was 0, and to align it to the boundary we need to add just NET_IP_ALIGN, which is two. NET_SKB_PAD is a much bigger value, (I think 64), which is used to reserve space to prepend an header, e.g. with tunnels.

How about the other adjustments that Eric mentioned regarding the size of the buffer? Aren't they required?

Quoted 5 lines

I guess that if stmmac_rx_buf1_len() needed such adjustment, it would be already broken when XDP is in use.
When you use XDP, stmmac_rx_offset() adds a pretty big headroom of 256 byte, which would easily trigger an overflow if not accounted. Did you try attaching a simple XDP program on a stock 5.13 kernel?

Yes, as mentioned in [1], to which you replied...

Quoted 8 lines
Quoted 4 lines
        M.

[1] https://lore.kernel.org/r/[email protected]

Great.
So I doubt that the adjustment is needed.
Does it work with all the frame size?

I have no idea. Honestly, you are the one who should be able to answer these questions, given that you should have worked out how the buffer allocations work in this particular driver.

This whole "let's try another random set of values until something sticks" is not how things ought to be done, and doesn't fill me with the utmost confidence that 5.14 (which apparently may well be cut in *two days*) is going to have a solid stmmac driver.

I re-re-request that this patch gets reverted until you figure out what is wrong with the initial patch.

Thanks,

	M.
Signature · 1 line
Without deviation from the norm, progress is not possible.
#24 lore raw
In reply to Matteo Croce
Quoted 12 lines · Matteo Croce
Quoted 9 lines · Matteo Croce
Quoted 3 lines · Marc Zyngier

On Fri, 20 Aug 2021 18:35:45 +0100,
Matteo Croce <[email protected]> wrote:

Great.
So I doubt that the adjustment is needed.
Does it work with all the frame size?

Quoted 5 lines

Last check, are you sure that the bpf program was loaded in the driver and not as generic XDP?
You can force it as native with "xdpdrv":

ip link set eth xdpdrv object kernel_passall.o

No issue whatsoever with 5.13.

	M.
Signature · 1 line
Without deviation from the norm, progress is not possible.
#26 lore raw
In reply to Marc Zyngier
Quoted 44 lines · Marc Zyngier
Quoted 23 lines · Matteo Croce
Quoted 21 lines · Marc Zyngier
Quoted 10 lines · Matteo Croce
Quoted 9 lines
Quoted 5 lines

I think it's wrong. The original offset was 0, and to align it to the boundary we need to add just NET_IP_ALIGN, which is two. NET_SKB_PAD is a much bigger value, (I think 64), which is used to reserve space to prepend an header, e.g. with tunnels.

How about the other adjustments that Eric mentioned regarding the size of the buffer? Aren't they required?

Quoted 5 lines

I guess that if stmmac_rx_buf1_len() needed such adjustment, it would be already broken when XDP is in use.
When you use XDP, stmmac_rx_offset() adds a pretty big headroom of 256 byte, which would easily trigger an overflow if not accounted. Did you try attaching a simple XDP program on a stock 5.13 kernel?

Yes, as mentioned in [1], to which you replied...

Quoted 8 lines
Quoted 4 lines
        M.

[1] https://lore.kernel.org/r/[email protected]

Great.
So I doubt that the adjustment is needed.
Does it work with all the frame size?

I have no idea. Honestly, you are the one who should be able to answer these questions, given that you should have worked out how the buffer allocations work in this particular driver.

This whole "let's try another random set of values until something sticks" is not how things ought to be done, and doesn't fill me with the utmost confidence that 5.14 (which apparently may well be cut in *two days*) is going to have a solid stmmac driver.

Quoted 5 lines

I re-re-request that this patch gets reverted until you figure out what is wrong with the initial patch.

Thanks,

I would have done it, but I'll not have the hardware until next week at least, otherwise I'd have tried all these tests myself.

I'm sure that NET_SKB_PAD doesn't need to be there, if just removing it fixes the problem, consider applying it and put a Fixes tag.

Signature · 1 line
per aspera ad upstream
#27 lore raw
Quoted 49 lines · Matteo Croce
Quoted 47 lines · Marc Zyngier
Quoted 23 lines · Matteo Croce
Quoted 21 lines · Marc Zyngier
Quoted 10 lines · Matteo Croce
Quoted 9 lines
Quoted 5 lines

I think it's wrong. The original offset was 0, and to align it to the boundary we need to add just NET_IP_ALIGN, which is two. NET_SKB_PAD is a much bigger value, (I think 64), which is used to reserve space to prepend an header, e.g. with tunnels.

How about the other adjustments that Eric mentioned regarding the size of the buffer? Aren't they required?

Quoted 5 lines

I guess that if stmmac_rx_buf1_len() needed such adjustment, it would be already broken when XDP is in use.
When you use XDP, stmmac_rx_offset() adds a pretty big headroom of 256 byte, which would easily trigger an overflow if not accounted. Did you try attaching a simple XDP program on a stock 5.13 kernel?

Yes, as mentioned in [1], to which you replied...

Quoted 8 lines
Quoted 4 lines
        M.

[1] https://lore.kernel.org/r/[email protected]

Great.
So I doubt that the adjustment is needed.
Does it work with all the frame size?

I have no idea. Honestly, you are the one who should be able to answer these questions, given that you should have worked out how the buffer allocations work in this particular driver.

This whole "let's try another random set of values until something sticks" is not how things ought to be done, and doesn't fill me with the utmost confidence that 5.14 (which apparently may well be cut in *two days*) is going to have a solid stmmac driver.

I re-re-request that this patch gets reverted until you figure out what is wrong with the initial patch.

Quoted 8 lines
Quoted 2 lines

Thanks,

I would have done it, but I'll not have the hardware until next week at least, otherwise I'd have tried all these tests myself.

I'm sure that NET_SKB_PAD doesn't need to be there, if just removing it fixes the problem, consider applying it and put a Fixes tag.

No, I don't think that's the right thing to do. A patch breaks a driver, and the author of the patch is not in a position to fix it. That's OK, these things happen, it's just bad timing.

But I don't understand this part of the kernel well enough to submit a patch based on a sample of *one*, at the last minute, just because "it works for me", and have the confidence that it doesn't break anything else.

I have now posted a revert of the original patch[1]. I'll be happy to work with you, with a less pressure, in order to have something that works for everyone in the next cycle.

Thanks,

	M.

[1] https://lore.kernel.org/netdev/[email protected]

Signature · 1 line
Without deviation from the norm, progress is not possible.