Date:   Sat, 10 Dec 2022 00:32:07 +0100
From:   Daniel Borkmann <daniel@...earbox.net>
To:     Jiri Olsa <olsajiri@...il.com>
Cc:     Yonghong Song <yhs@...a.com>,
        Alexei Starovoitov <alexei.starovoitov@...il.com>,
        Song Liu <song@...nel.org>, Hao Sun <sunhao.th@...il.com>,
        Peter Zijlstra <peterz@...radead.org>,
        bpf <bpf@...r.kernel.org>, Alexei Starovoitov <ast@...nel.org>,
        John Fastabend <john.fastabend@...il.com>,
        Andrii Nakryiko <andrii@...nel.org>,
        Martin KaFai Lau <martin.lau@...ux.dev>,
        Yonghong Song <yhs@...com>, KP Singh <kpsingh@...nel.org>,
        Stanislav Fomichev <sdf@...gle.com>,
        Hao Luo <haoluo@...gle.com>,
        David Miller <davem@...emloft.net>,
        Jakub Kicinski <kuba@...nel.org>,
        Jesper Dangaard Brouer <hawk@...nel.org>,
        Linux Kernel Mailing List <linux-kernel@...r.kernel.org>,
        netdev <netdev@...r.kernel.org>,
        Thorsten Leemhuis <regressions@...mhuis.info>
Subject: Re: BUG: unable to handle kernel paging request in bpf_dispatcher_xdp

On 12/10/22 12:07 AM, Jiri Olsa wrote:
> On Fri, Dec 09, 2022 at 11:41:11PM +0100, Daniel Borkmann wrote:
>> On 12/9/22 10:53 PM, Jiri Olsa wrote:
>>> On Fri, Dec 09, 2022 at 12:31:06PM -0800, Yonghong Song wrote:
>>>>
>>>>
>>>> On 12/9/22 7:20 AM, Jiri Olsa wrote:
>>>>> On Fri, Dec 09, 2022 at 02:50:55PM +0100, Jiri Olsa wrote:
>>>>>> On Fri, Dec 09, 2022 at 12:22:37PM +0100, Jiri Olsa wrote:
>>>>>>
>>>>>> SBIP
>>>>>>
>>>>>>>>>>>>>
>>>>>>>>>>>>> I'm trying to understand the severity of the issues and
>>>>>>>>>>>>> whether we need to revert that commit asap since the merge window
>>>>>>>>>>>>> is about to start.
>>>>>>>>>>>>
>>>>>>>>>>>> Jiri, Peter,
>>>>>>>>>>>>
>>>>>>>>>>>> ping.
>>>>>>>>>>>>
>>>>>>>>>>>> cc-ing Thorsten, since he's tracking it now.
>>>>>>>>>>>>
>>>>>>>>>>>> The config has CONFIG_X86_KERNEL_IBT=y.
>>>>>>>>>>>> Is it related?
>>>>>>>>>>>
>>>>>>>>>>> sorry for late reply.. I still did not find the reason,
>>>>>>>>>>> but I did not try with IBT yet, will test now
>>>>>>>>>>
>>>>>>>>>> no difference with IBT enabled, can't reproduce the issue
>>>>>>>>>>
>>>>>>>>>
>>>>>>>>> ok, scratch that.. the reproducer got stuck on wifi init :-\
>>>>>>>>>
>>>>>>>>> after I fix that I can now reproduce on my local config with
>>>>>>>>> IBT enabled or disabled.. it's something else
>>>>>>>>
>>>>>>>> I'm getting the error also when reverting the static call change,
>>>>>>>> looking for good commit, bisecting
>>>>>>>>
>>>>>>>> I'm getting fail with:
>>>>>>>>       f0c4d9fc9cc9 (tag: v6.1-rc4) Linux 6.1-rc4
>>>>>>>>
>>>>>>>> v6.1-rc1 is ok
>>>>>>>
>>>>>>> so far I narrowed it down between rc1 and rc3.. bisect got me nowhere so far
>>>>>>>
>>>>>>> attaching some more logs
>>>>>>
>>>>>> looking at the code.. how do we ensure that code running through
>>>>>> bpf_prog_run_xdp will not get dispatcher image changed while
>>>>>> it's being exetuted
>>>>>>
>>>>>> we use 'the other half' of the image when we add/remove programs,
>>>>>> but could bpf_dispatcher_update race with bpf_prog_run_xdp like:
>>>>>>
>>>>>>
>>>>>> cpu 0:                                  cpu 1:
>>>>>>
>>>>>> bpf_prog_run_xdp
>>>>>>       ...
>>>>>>       bpf_dispatcher_xdp_func
>>>>>>          start exec image at offset 0x0
>>>>>>
>>>>>>                                            bpf_dispatcher_update
>>>>>>                                                    update image at offset 0x800
>>>>>>                                            bpf_dispatcher_update
>>>>>>                                                    update image at offset 0x0
>>>>>>
>>>>>>          still in image at offset 0x0
>>>>>>
>>>>>>
>>>>>> that might explain why I wasn't able to trigger that on
>>>>>> bare metal just in qemu
>>>>>
>>>>> I tried patch below and it fixes the issue for me and seems
>>>>> to confirm the race above.. but not sure it's the best fix
>>>>>
>>>>> jirka
>>>>>
>>>>>
>>>>> ---
>>>>> diff --git a/kernel/bpf/dispatcher.c b/kernel/bpf/dispatcher.c
>>>>> index c19719f48ce0..6a2ced102fc7 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_tasks();
>>>>>     	if (new)
>>>>>     		d->image_off = noff;
>>>>
>>>> This might work. In arch/x86/kernel/alternative.c, we have following
>>>> code and comments. For text_poke, synchronize_rcu_tasks() might be able
>>>> to avoid concurrent execution and update.
>>>
>>> so my idea was that we need to ensure all the current callers of
>>> bpf_dispatcher_xdp_func (which should have rcu read lock, based
>>> on the comment in bpf_prog_run_xdp) are gone before and new ones
>>> execute the new image, so the next call to the bpf_dispatcher_update
>>> will be safe to overwrite the other half of the image
>>
>> If v6.1-rc1 was indeed okay, then it looks like this may be related to
>> the trampoline patching for the static_call? Did it repro on v6.1-rc1
>> just with dbe69b299884 ("bpf: Fix dispatcher patchable function entry
>> to 5 bytes nop") cherry-picked?
> 
> I'll try that.. it looks to me like the problem was always there,
> maybe harder to trigger.. also to reproduce it you need to call
> bpf_dispatcher_update heavily, which is not probably the common
> use case
> 
> one other thing is that I think the fix might need rcu locking
> on the bpf_dispatcher_xdp_func side, because local_bh_disable
> seems not to be enough to make synchronize_rcu_tasks work
> 
> I'm now testing patch below
> 
> jirka
> 
> ---
> diff --git a/include/linux/filter.h b/include/linux/filter.h
> index efc42a6e3aed..a27245b96d6b 100644
> --- a/include/linux/filter.h
> +++ b/include/linux/filter.h
> @@ -772,7 +772,13 @@ static __always_inline u32 bpf_prog_run_xdp(const struct bpf_prog *prog,
>   	 * under local_bh_disable(), which provides the needed RCU protection
>   	 * for accessing map entries.
>   	 */
> -	u32 act = __bpf_prog_run(prog, xdp, BPF_DISPATCHER_FUNC(xdp));
> +	u32 act;
> +
> +	rcu_read_lock();
> +
> +	act = __bpf_prog_run(prog, xdp, BPF_DISPATCHER_FUNC(xdp));
> +
> +	rcu_read_unlock();

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(). [...]

>   	if (static_branch_unlikely(&bpf_master_redirect_enabled_key)) {
>   		if (act == XDP_TX && netif_is_bond_slave(xdp->rxq->dev))
> diff --git a/kernel/bpf/dispatcher.c b/kernel/bpf/dispatcher.c
> index c19719f48ce0..6a2ced102fc7 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_tasks();
>   
>   	if (new)
>   		d->image_off = noff;
> 

Powered by blists - more mailing lists