Re: BUG: unable to handle kernel paging request in bpf_dispatcher_xdp

[Hao Sun]

> On 12 Dec 2022, at 11:04 PM, Jiri Olsa <olsajiri@xxxxxxxxx> wrote:
> 
> On Sat, Dec 10, 2022 at 02:11:34PM +0100, Jiri Olsa wrote:
>> On Fri, Dec 09, 2022 at 05:12:03PM -0800, Alexei Starovoitov wrote:
>>> On Fri, Dec 9, 2022 at 4:06 PM Jiri Olsa <olsajiri@xxxxxxxxx> wrote:
>>>> 
>>>> On Fri, Dec 09, 2022 at 03:34:45PM -0800, Jakub Kicinski wrote:
>>>>> On Sat, 10 Dec 2022 00:32:07 +0100 Daniel Borkmann wrote:
>>>>>> fwiw, these should not be necessary, Documentation/RCU/checklist.rst :
>>>>>> 
>>>>>>   [...] One example of non-obvious pairing is the XDP feature in networking,
>>>>>>   which calls BPF programs from network-driver NAPI (softirq) context. BPF
>>>>>>   relies heavily on RCU protection for its data structures, but because the
>>>>>>   BPF program invocation happens entirely within a single local_bh_disable()
>>>>>>   section in a NAPI poll cycle, this usage is safe. The reason that this usage
>>>>>>   is safe is that readers can use anything that disables BH when updaters use
>>>>>>   call_rcu() or synchronize_rcu(). [...]
>>>>> 
>>>>> FWIW I sent a link to the thread to Paul and he confirmed
>>>>> the RCU will wait for just the BH.
>>>> 
>>>> so IIUC we can omit the rcu_read_lock/unlock on bpf_prog_run_xdp side
>>>> 
>>>> Paul,
>>>> any thoughts on what we can use in here to synchronize bpf_dispatcher_change_prog
>>>> with bpf_prog_run_xdp callers?
>>>> 
>>>> with synchronize_rcu_tasks I'm getting splats like:
>>>>  https://lore.kernel.org/bpf/20221209153445.22182ca5@xxxxxxxxxx/T/#m0a869f93404a2744884d922bc96d497ffe8f579f
>>>> 
>>>> synchronize_rcu_tasks_rude seems to work (patch below), but it also sounds special ;-)
>>> 
>>> Jiri,
>>> 
>>> I haven't tried to repro this yet, but I feel you're on
>>> the wrong path here. The splat has this:
>>> ? bpf_prog_run_xdp include/linux/filter.h:775 [inline]
>>> ? bpf_test_run+0x2ce/0x990 net/bpf/test_run.c:400
>>> that test_run logic takes rcu_read_lock.
>>> See bpf_test_timer_enter.
>>> I suspect the addition of synchronize_rcu_tasks_rude
>>> only slows down the race.
>>> The synchronize_rcu_tasks_trace also behaves like synchronize_rcu.
>>> See our new and fancy rcu_trace_implies_rcu_gp(),
>>> but I'm not sure it applies to synchronize_rcu_tasks_rude.
>>> Have you tried with just synchronize_rcu() ?
>>> If your theory about the race is correct then
>>> the vanila sync_rcu should help.
>>> If not, the issue is some place else.
>> 
>> synchronize_rcu seems to work as well, I'll keep the test
>> running for some time
> 
> looks good, Hao Sun, could you please test change below?

Hi,

Tested on a latest bpf-next build. The reproducer would trigger
the Oops in 5 mins without the patch. After applying the patch,
the reproducer cannot trigger any issue for more than 15 mins.
Seems working, tested on:

HEAD commit: ef3911a3e4d6 docs/bpf: Reword docs for BPF_MAP_TYPE_SK_STORAGE
git tree: bpf-next
kernel config: https://pastebin.com/raw/rZdWLcgK
C reproducer: https://pastebin.com/raw/GFfDn2Gk

> 
> ---
> diff --git a/kernel/bpf/dispatcher.c b/kernel/bpf/dispatcher.c
> index c19719f48ce0..4b0fa5b98137 100644
> --- a/kernel/bpf/dispatcher.c
> +++ b/kernel/bpf/dispatcher.c
> @@ -124,6 +124,7 @@ static void bpf_dispatcher_update(struct bpf_dispatcher *d, int prev_num_progs)
> }
> 
> __BPF_DISPATCHER_UPDATE(d, new ?: (void *)&bpf_dispatcher_nop_func);
> + synchronize_rcu();
> 
> if (new)
> d->image_off = noff;