I have managed to reproduce and root cause it. Will produce a patch. Thanks for finding this issue. /Magnus On Fri, Oct 18, 2019 at 1:23 PM Magnus Karlsson <magnus.karlsson@xxxxxxxxx> wrote: > > Thanks Kal. Will take a look at it and get back. > > /Magnus > > On Fri, Oct 18, 2019 at 12:18 PM Kal Cutter Conley > <kal.conley@xxxxxxxxxxx> wrote: > > > > Hello, > > I have managed to extract a reproducer that crashes the kernel. We > > observe the crash with all drivers that support XDP zerocopy (i40e, > > ixgbe, mlx5_core). The program source is attached (C++). Compile with > > a newer clang++ or g++ using flag -std=c++17. It must also be linked > > with libbpf. Use libbpf v0.0.5 and make sure to revert > > 5771dacd3dc2fdd041c51242819a9f212e04af55. > > > > The crash is a kernel NULL pointer dereference in xsk_umem_consume_tx. > > > > Hopefully this is useful, > > Kal > > > > On Wed, Aug 7, 2019 at 10:48 PM Kal Cutter Conley > > <kal.conley@xxxxxxxxxxx> wrote: > > > > > > Hello, > > > I am trying to get AF_XDP working with the i40e driver (Ethernet > > > Controller X710 for 10GbE SFP+). After bind() with XDP_ZEROCOPY the > > > kernel (machine) freezes hard. I have reproduced this on varying > > > kernel versions between 5.1 and 5.3-rc3 with 5.3 kernels freezing > > > also, but at a later stage. I tried replacing my XDP program with a > > > trivial one that simply returns XDP_PASS but it didn't help. On the > > > same system, the xdpsock sample does appear to work with the -z flag, > > > however. Are there any current known issues that could be causing > > > this? I will try to extract a minimal example that exercises the > > > freeze. > > > > > > Thanks, > > > Kal
|