Date:   Tue, 31 Jan 2023 14:00:32 +0100
From:   Jesper Dangaard Brouer <jbrouer@...hat.com>
To:     Stanislav Fomichev <sdf@...gle.com>,
        Toke Høiland-Jørgensen <toke@...hat.com>
Cc:     brouer@...hat.com, bpf@...r.kernel.org, netdev@...r.kernel.org,
        martin.lau@...nel.org, ast@...nel.org, daniel@...earbox.net,
        andrii@...nel.org, martin.lau@...ux.dev, song@...nel.org,
        yhs@...com, john.fastabend@...il.com, dsahern@...il.com,
        willemb@...gle.com, void@...ifault.com, kuba@...nel.org,
        xdp-hints@...-project.net
Subject: Re: [xdp-hints] [PATCH bpf-next RFC V1] selftests/bpf:
 xdp_hw_metadata clear metadata when -EOPNOTSUPP



On 27/01/2023 18.18, Stanislav Fomichev wrote:
> On Fri, Jan 27, 2023 at 5:58 AM Toke Høiland-Jørgensen <toke@...hat.com> wrote:
>>
>> Jesper Dangaard Brouer <brouer@...hat.com> writes:
>>
>>> The AF_XDP userspace part of xdp_hw_metadata see non-zero as a signal of
>>> the availability of rx_timestamp and rx_hash in data_meta area. The
>>> kernel-side BPF-prog code doesn't initialize these members when kernel
>>> returns an error e.g. -EOPNOTSUPP.  This memory area is not guaranteed to
>>> be zeroed, and can contain garbage/previous values, which will be read
>>> and interpreted by AF_XDP userspace side.
>>>
>>> Tested this on different drivers. The experiences are that for most
>>> packets they will have zeroed this data_meta area, but occasionally it
>>> will contain garbage data.
>>>
>>> Example of failure tested on ixgbe:
>>>   poll: 1 (0)
>>>   xsk_ring_cons__peek: 1
>>>   0x18ec788: rx_desc[0]->addr=100000000008000 addr=8100 comp_addr=8000
>>>   rx_hash: 3697961069
>>>   rx_timestamp:  9024981991734834796 (sec:9024981991.7348)
>>>   0x18ec788: complete idx=8 addr=8000
>>>
>>> Converting to date:
>>>   date -d @9024981991
>>>   2255-12-28T20:26:31 CET
>>>
>>> I choose a simple fix in this patch. When kfunc fails or isn't supported
>>> assign zero to the corresponding struct meta value.
>>>
>>> It's up to the individual BPF-programmer to do something smarter e.g.
>>> that fits their use-case, like getting a software timestamp and marking
>>> a flag that gives the type of timestamp.
>>>
>>> Another possibility is for the behavior of kfunc's
>>> bpf_xdp_metadata_rx_timestamp and bpf_xdp_metadata_rx_hash to require
>>> clearing return value pointer.
>>
>> I definitely think we should leave it up to the BPF programmer to react
>> to failures; that's what the return code is there for, after all :)
> 
> +1

+1 I agree.
We should keep this default functions as simple as possible, for future
"unroll" of BPF-bytecode.

I the -EOPNOTSUPP case (default functions for drivers not implementing
kfunc), will likely be used runtime by BPF-prog to determine if the
hardware have this offload hint, but it comes with the overhead of a
function pointer call.

I hope we can somehow BPF-bytecode "unroll" these (default functions) at
BPF-load time, to remove this overhead, and perhaps even let BPF
bytecode do const propagation and code elimination?


> Maybe we can unconditionally memset(meta, sizeof(*meta), 0) in
> tools/testing/selftests/bpf/progs/xdp_hw_metadata.c?
> Since it's not a performance tool, it should be ok functionality-wise.

I know this isn't a performance test, but IMHO always memsetting
metadata area is a misleading example.  We know from experience that
developer simply copy-paste code examples, even quick-n-dirty testing
example code.

The specific issue in this example can lead to hard-to-find bugs, as my
testing shows it is only occasionally that data_meta area contains
garbage. We could do a memset, but it deserves a large code comment, why
this is needed, so people copy-pasting understand. I choose current
approach to keep code close to code people will copy-paste.

--Jesper

Powered by blists - more mailing lists