On 7/21/25 8:34 PM, Gal Pressman wrote:
On 21/07/2025 18:40, Jakub Kicinski wrote:
On Mon, 21 Jul 2025 14:43:15 +0300 Nimrod Oren wrote:
+static struct udphdr *filter_udphdr(struct xdp_md *ctx, __u16 port) +{
- void *data_end = (void *)(long)ctx->data_end;
- void *data = (void *)(long)ctx->data;
- struct udphdr *udph = NULL;
- struct ethhdr *eth = data;
- if (data + sizeof(*eth) > data_end)
return NULL;This check assumes that the packet headers reside in the linear part of the xdp_buff. However, this assumption does not hold across all drivers. For example, in mlx5, the linear part is empty when using multi-buffer mode with striding rq configuration. This causes all multi-buffer test cases to fail over mlx5.
To ensure correctness across all drivers, all direct accesses to packet data should use these safer helper functions instead: bpf_xdp_load_bytes() and bpf_xdp_store_bytes().
Related discussion and context can be found here: https://github.com/xdp-project/xdp-tools/pull/409
That's a reasonable way to modify the test. But I'm not sure it's something that should be blocking merging the patches. Or for that matter whether it's Mohsin's responsibility to make the test cater to quirks of mlx5,
Definitely not a quirk, you cannot assume the headers are in the linear part, especially if you're going to put this program as reference in the kernel tree.
This issue has nothing to do with mlx5, but a buggy XDP program.
Note that with the self-tests we have a slightly different premise WRT the actual kernel code. We prefer on-boarding tests cases that work for some/most of the possible setup vs perfect ones, and eventually improve incrementally as needed: the goal is to increase the code coverage _and_ encourage people to contribute new tests upstream.
We try to avoid breaking existing tests (at least the ones actually reporting into the infrastructure), but for new ones the barriers are intentionally different than VS kernel code.
Cheers,
Paolo