[<prev] [next>] [<thread-prev] [thread-next>] [day] [month] [year] [list]
Message-ID: <CAMB2axPpaoDfFEBzNTaTjp4GnFKtWy0k-sTez56ap+FBZzLFeA@mail.gmail.com>
Date: Fri, 29 Aug 2025 00:26:29 -0700
From: Amery Hung <ameryhung@...il.com>
To: Nimrod Oren <noren@...dia.com>
Cc: bpf@...r.kernel.org, netdev@...r.kernel.org, alexei.starovoitov@...il.com,
andrii@...nel.org, daniel@...earbox.net, kuba@...nel.org,
martin.lau@...nel.org, mohsin.bashr@...il.com, saeedm@...dia.com,
tariqt@...dia.com, mbloch@...dia.com, maciej.fijalkowski@...el.com,
kernel-team@...a.com, Dragos Tatulea <dtatulea@...dia.com>
Subject: Re: [RFC bpf-next v1 0/7] Add kfunc bpf_xdp_pull_data
On Thu, Aug 28, 2025 at 6:39 AM Nimrod Oren <noren@...dia.com> wrote:
>
> On 25/08/2025 22:39, Amery Hung wrote:
> > This patchset introduces a new kfunc bpf_xdp_pull_data() to allow
> > pulling nonlinear xdp data. This may be useful when a driver places
> > headers in fragments. When an xdp program would like to keep parsing
> > packet headers using direct packet access, it can call
> > bpf_xdp_pull_data() to make the header available in the linear data
> > area. The kfunc can also be used to decapsulate the header in the
> > nonlinear data, as currently there is no easy way to do this.
>
> I'm currently working on a series that converts the xdp_native program
> to use dynptr for accessing header data. If accepted, it should provide
> better performance, since dynptr can access without copying the data.
>
I feel that bpf_xdp_pull_data() is a more generic approach, but yeah
dynptr may yield better performance. Looking forward to seeing the
numbers.
It will also be great if the dynptr approach doesn't require
xdp_native to make assumptions about the xdp_buff layout (headers are
in frags if linear data is empty), and creates two versions of header
parsing code.
> > This patchset also tries to fix an issue in the mlx5e driver. The driver
> > curretly assumes the packet layout to be unchanged after xdp program
> > runs and may generate packet with corrupted data or trigger kernel warning
> > if xdp programs calls layout-changing kfunc such as bpf_xdp_adjust_tail(),
> > bpf_xdp_adjust_head() or bpf_xdp_pull_data() introduced in this set.
>
> Thanks for working on this!
>
> > Tested with the added bpf selftest using bpf test_run and also on
> > mlx5e with the tools/testing/selftests/drivers/net/xdp.py. mlx5e with
> > striding RQ will produce xdp_buff with empty linear data.
> > xdp.test_xdp_native_pass_mb would fail to parse the header before this
> > patchset.
>
> I got a crash when testing this series with the xdp_dummy program from
> tools/testing/selftests/net/lib/. Need to make sure we're not breaking
> compatibility for programs that keep the linear part empty.
Appreciate the detailed review! I will make sure to test xdp_dummy as
well. I am taking some time off and will get back to this patchset
next week.
Powered by blists - more mailing lists