Re: [1/2 bpf-next] bpf: expose net_device from xdp for metadata

[Yonghong Song <yhs@meta.com>]

On 11/10/22 2:58 PM, John Fastabend wrote:
> Yonghong Song wrote:
>>
>>
>> On 11/9/22 6:17 PM, John Fastabend wrote:
>>> Yonghong Song wrote:
>>>>
>>>>
>>>> On 11/9/22 1:52 PM, John Fastabend wrote:
>>>>> Allow xdp progs to read the net_device structure. Its useful to extract
>>>>> info from the dev itself. Currently, our tracing tooling uses kprobes
>>>>> to capture statistics and information about running net devices. We use
>>>>> kprobes instead of other hooks tc/xdp because we need to collect
>>>>> information about the interface not exposed through the xdp_md structures.
>>>>> This has some down sides that we want to avoid by moving these into the
>>>>> XDP hook itself. First, placing the kprobes in a generic function in
>>>>> the kernel is after XDP so we miss redirects and such done by the
>>>>> XDP networking program. And its needless overhead because we are
>>>>> already paying the cost for calling the XDP program, calling yet
>>>>> another prog is a waste. Better to do everything in one hook from
>>>>> performance side.
>>>>>
>>>>> Of course we could one-off each one of these fields, but that would
>>>>> explode the xdp_md struct and then require writing convert_ctx_access
>>>>> writers for each field. By using BTF we avoid writing field specific
>>>>> convertion logic, BTF just knows how to read the fields, we don't
>>>>> have to add many fields to xdp_md, and I don't have to get every
>>>>> field we will use in the future correct.
>>>>>
>>>>> For reference current examples in our code base use the ifindex,
>>>>> ifname, qdisc stats, net_ns fields, among others. With this
>>>>> patch we can now do the following,
>>>>>
>>>>>            dev = ctx->rx_dev;
>>>>>            net = dev->nd_net.net;
>>>>>
>>>>> 	uid.ifindex = dev->ifindex;
>>>>> 	memcpy(uid.ifname, dev->ifname, NAME);
>>>>>            if (net)
>>>>> 		uid.inum = net->ns.inum;
>>>>>
>>>>> to report the name, index and ns.inum which identifies an
>>>>> interface in our system.
>>>>
>>>> In
>>>> https://lore.kernel.org/bpf/ad15b398-9069-4a0e-48cb-4bb651ec3088@meta.com/
>>>> Namhyung Kim wanted to access new perf data with a helper.
>>>> I proposed a helper bpf_get_kern_ctx() which will get
>>>> the kernel ctx struct from which the actual perf data
>>>> can be retrieved. The interface looks like
>>>> 	void *bpf_get_kern_ctx(void *)
>>>> the input parameter needs to be a PTR_TO_CTX and
>>>> the verifer is able to return the corresponding kernel
>>>> ctx struct based on program type.
>>>>
>>>> The following is really hacked demonstration with
>>>> some of change coming from my bpf_rcu_read_lock()
>>>> patch set https://lore.kernel.org/bpf/20221109211944.3213817-1-yhs@fb.com/
>>>>
>>>> I modified your test to utilize the
>>>> bpf_get_kern_ctx() helper in your test_xdp_md.c.
>>>>
>>>> With this single helper, we can cover the above perf
>>>> data use case and your use case and maybe others
>>>> to avoid new UAPI changes.
>>>
>>> hmm I like the idea of just accessing the xdp_buff directly
>>> instead of adding more fields. I'm less convinced of the
>>> kfunc approach. What about a terminating field *self in the
>>> xdp_md. Then we can use existing convert_ctx_access to make
>>> it BPF inlined and no verifier changes needed.
>>>
>>> Something like this quickly typed up and not compiled, but
>>> I think shows what I'm thinking.
>>>
>>> diff --git a/include/uapi/linux/bpf.h b/include/uapi/linux/bpf.h
>>> index 94659f6b3395..10ebd90d6677 100644
>>> --- a/include/uapi/linux/bpf.h
>>> +++ b/include/uapi/linux/bpf.h
>>> @@ -6123,6 +6123,10 @@ struct xdp_md {
>>>           __u32 rx_queue_index;  /* rxq->queue_index  */
>>>    
>>>           __u32 egress_ifindex;  /* txq->dev->ifindex */
>>> +       /* Last xdp_md entry, for new types add directly to xdp_buff and use
>>> +        * BTF access. Reading this gives BTF access to xdp_buff.
>>> +        */
>>> +       __bpf_md_ptr(struct xdp_buff *, self);
>>>    };
>>
>> This would be the first instance to have a kernel internal struct
>> in a uapi struct. Not sure whether this is a good idea or not.
> 
> We can use probe_read from some of the socket progs already but
> sure.
> 
>>
>>>    
>>>    /* DEVMAP map-value layout
>>> diff --git a/net/core/filter.c b/net/core/filter.c
>>> index bb0136e7a8e4..547e9576a918 100644
>>> --- a/net/core/filter.c
>>> +++ b/net/core/filter.c
>>> @@ -9808,6 +9808,11 @@ static u32 xdp_convert_ctx_access(enum bpf_access_type type,
>>>                   *insn++ = BPF_LDX_MEM(BPF_W, si->dst_reg, si->dst_reg,
>>>                                         offsetof(struct net_device, ifindex));
>>>                   break;
>>> +       case offsetof(struct xdp_md, self):
>>> +               *insn++ = BPF_LDX_MEM(BPF_FIELD_SIZEOF(struct xdp_buff, self),
>>> +                                     si->dst_reg, si->src_reg,
>>> +                                     offsetof(struct xdp_buff, 0));
>>> +               break;
>>>           }
>>>    
>>>           return insn - insn_buf;
>>>
>>> Actually even that single insn conversion is a bit unnessary because
>>> should be enough to just change the type to the correct BTF_ID in the
>>> verifier and omit any instructions. But it wwould be a bit confusing
>>> for C side. Might be a good use for passing 'cast' info through to
>>> the verifier as an annotation so it could just do the BTF_ID cast for
>>> us without any insns.
>>
>> We cannot change the context type to BTF_ID style which will be a
>> uapi violation.
> 
> I don't think it would be uapi violation if user asks for it
> by annotating the cast.
> 
>>
>> The helper I proposed can be rewritten by verifier as
>> 	r0 = r1
>> so we should not have overhead for this.
> 
> Agree other than reading the bpf asm where its a bit odd.
> 
>> It cover all program types with known uapi ctx -> kern ctx
>> conversions. So there is no need to change existing uapi structs.
>> Also I except that most people probably won't use this kfunc.
>> The existing uapi fields might already serve most needs.
> 
> Maybe not sure missing some things we need.
> 
>>
>> Internally we have another use case to access some 'struct sock' fields
>> but the uapi struct only has struct bpf_sock. Currently it is advised
>> to use bpf_probe_read_kernel(...) to get the needed information.
>> The proposed helper should help that too without uapi change.
> 
> Yep.
> 
> I'm fine doing it with bpf_get_kern_ctx() did you want me to code it
> the rest of the way up and test it?

I have an off-line discussion with Martin. Martin has a use case
to get the btf_id from a pointer casting, something like
     void *p = ...
     struct t *pt = (struct t *)p; // having a kfunc to do this.
So I would like to generate the above helper to be something like
    bpf_get_kern_btf_id(...)
which will cover ctx case as well.

I should be able to have a RFC patch ready next week.

> 
> .John