Skip to contentSkip to conversation

net: xilinx: axienet: Query about checksum partial implementation

DISCUSSION Simon Horman 2 participants 3 messages Jul 26, 2024 – Jul 31, 2024

Subsystems (2)
lore
Share & tools
lore

Simon Horman

author reviewer
#0 lore raw

Hi Radhey, all,

I am wondering if you could shed some light on the following checksum partial handling in the axienet_rx_poll():

                        /* if we're doing Rx csum offload, set it up */
                        if (lp->features & XAE_FEATURE_FULL_RX_CSUM) {
				...
                        } else if ((lp->features & XAE_FEATURE_PARTIAL_RX_CSUM) != 0 &&
                                   skb->protocol == htons(ETH_P_IP) &&
                                   skb->len > 64) {
                                skb->csum = be32_to_cpu(cur_p->app3 & 0xFFFF);
                                ...
                        }

In particluar the "skb->csum =" line.

The type of cur_p->app3 is u32, and 0xFFFF is also host byte order. So far so good. But after the bitwise operation it is treated as a big-endian value by passing it to be32_to_cpu.

Perhaps I am missing something obvious, but my question is how does that work?

* Was it only tested on big endian sysgtems where be32_to_cpu() is a no-op

* Was it only tested on little endian systems where be32_to_cpu() is a byteswap and somehow that works (how?).

* Is the code unecessised because the XAE_FEATURE_FULL_RX_CSUM branch is always taken?

A grep of dts files shows up arch/microblaze/boot/dts/system.dts which sets sets xlnx,rxcsum to 0, which corresponds to XAE_NO_CSUM_OFFLOAD.

* Something else

Flagged by Sparse

The in quesoitn code seems to have been introduced by 8a3b7a252dca ("drivers/net/ethernet/xilinx: added Xilinx AXI Ethernet driver")

#1 lore raw
Quoted 39 lines

-----Original Message-----
From: Simon Horman <[email protected]>
Sent: Friday, July 26, 2024 5:37 PM
To: Pandey, Radhey Shyam <[email protected]>
Cc: Daniel Borkmann <[email protected]>; Ariane Keller <[email protected]>; Simek, Michal <[email protected]>; [email protected]; [email protected]
Subject: net: xilinx: axienet: Query about checksum partial implementation

Hi Radhey, all,

I am wondering if you could shed some light on the following checksum partial handling in the axienet_rx_poll():

                        /* if we're doing Rx csum offload, set it up */
                        if (lp->features & XAE_FEATURE_FULL_RX_CSUM) {
				...
                        } else if ((lp->features & XAE_FEATURE_PARTIAL_RX_CSUM) != 0
&&
                                   skb->protocol == htons(ETH_P_IP) &&
                                   skb->len > 64) {
                                skb->csum = be32_to_cpu(cur_p->app3 & 0xFFFF);
                                ...
                        }

In particluar the "skb->csum =" line.

The type of cur_p->app3 is u32, and 0xFFFF is also host byte order. So far so good. But after the bitwise operation it is treated as a big-endian value by passing it to be32_to_cpu.

Perhaps I am missing something obvious, but my question is how does that work?

* Was it only tested on big endian sysgtems where be32_to_cpu() is a no-op

* Was it only tested on little endian systems where be32_to_cpu() is a byteswap and somehow that works (how?).

Quoted 6 lines

* Is the code unecessised because the XAE_FEATURE_FULL_RX_CSUM branch is
always taken?

A grep of dts files shows up arch/microblaze/boot/dts/system.dts which sets sets xlnx,rxcsum to 0, which corresponds to XAE_NO_CSUM_OFFLOAD.

+ Harini

Yes, IIRC default AXI Ethernet IP RX checksum is set to "No checksum offload" so, it is default case and being set in most designs. Have added Harini to this thread to confirm on partial checksum verification results.

Assuming partial implementation is functional then likely DMA IP updates application field in big endian format and that is the reason we have this be32 to CPU conversion in place. will dig a bit more and get back on it.

Quoted 8 lines

* Something else

Flagged by Sparse

The in quesoitn code seems to have been introduced by 8a3b7a252dca ("drivers/net/ethernet/xilinx: added Xilinx AXI Ethernet driver")

author reviewer
#2 lore raw
Quoted 48 lines · Radhey Shyam
Quoted 39 lines

-----Original Message-----
From: Simon Horman <[email protected]>
Sent: Friday, July 26, 2024 5:37 PM
To: Pandey, Radhey Shyam <[email protected]>
Cc: Daniel Borkmann <[email protected]>; Ariane Keller <[email protected]>; Simek, Michal <[email protected]>; [email protected]; [email protected]
Subject: net: xilinx: axienet: Query about checksum partial implementation

Hi Radhey, all,

I am wondering if you could shed some light on the following checksum partial handling in the axienet_rx_poll():

                        /* if we're doing Rx csum offload, set it up */
                        if (lp->features & XAE_FEATURE_FULL_RX_CSUM) {
				...
                        } else if ((lp->features & XAE_FEATURE_PARTIAL_RX_CSUM) != 0
&&
                                   skb->protocol == htons(ETH_P_IP) &&
                                   skb->len > 64) {
                                skb->csum = be32_to_cpu(cur_p->app3 & 0xFFFF);
                                ...
                        }

In particluar the "skb->csum =" line.

The type of cur_p->app3 is u32, and 0xFFFF is also host byte order. So far so good. But after the bitwise operation it is treated as a big-endian value by passing it to be32_to_cpu.

Perhaps I am missing something obvious, but my question is how does that work?

* Was it only tested on big endian sysgtems where be32_to_cpu() is a no-op

* Was it only tested on little endian systems where be32_to_cpu() is a byteswap and somehow that works (how?).

Quoted 6 lines

* Is the code unecessised because the XAE_FEATURE_FULL_RX_CSUM branch is
always taken?

A grep of dts files shows up arch/microblaze/boot/dts/system.dts which sets sets xlnx,rxcsum to 0, which corresponds to XAE_NO_CSUM_OFFLOAD.

+ Harini

Quoted 7 lines

Yes, IIRC default AXI Ethernet IP RX checksum is set to "No checksum offload" so, it is default case and being set in most designs. Have added Harini to this thread to confirm on partial checksum verification results.

Assuming partial implementation is functional then likely DMA IP updates application field in big endian format and that is the reason we have this be32 to CPU conversion in place. will dig a bit more and get back on it.

Thanks, much appreciated.

FWIIW, I do agree that the scenario you describe would mostly explain things, although the mask with 0xFFFF still seems off.