Re: [PATCH RFC v4 bpf-next 03/11] xdp: Add xdp_txq_info to xdp_buff

["Toke Høiland-Jørgensen" <toke@redhat.com>]

Jesper Dangaard Brouer <brouer@redhat.com> writes:

> On Wed, 26 Feb 2020 20:20:05 -0700
> David Ahern <dsahern@kernel.org> wrote:
>
>> From: David Ahern <dahern@digitalocean.com>
>> 
>> Add xdp_txq_info as the Tx counterpart to xdp_rxq_info. At the
>> moment only the device is added. Other fields (queue_index)
>> can be added as use cases arise.
>> 
>> From a UAPI perspective, egress_ifindex is a union with ingress_ifindex
>> since only one applies based on where the program is attached.
>> 
>> Signed-off-by: David Ahern <dahern@digitalocean.com>
>> ---
>>  include/net/xdp.h        |  5 +++++
>>  include/uapi/linux/bpf.h |  6 ++++--
>>  net/core/filter.c        | 27 +++++++++++++++++++--------
>>  3 files changed, 28 insertions(+), 10 deletions(-)
>> 
>> diff --git a/include/net/xdp.h b/include/net/xdp.h
>> index 40c6d3398458..5584b9db86fe 100644
>> --- a/include/net/xdp.h
>> +++ b/include/net/xdp.h
>> @@ -63,6 +63,10 @@ struct xdp_rxq_info {
>>  	struct xdp_mem_info mem;
>>  } ____cacheline_aligned; /* perf critical, avoid false-sharing */
>>  
>> +struct xdp_txq_info {
>> +	struct net_device *dev;
>> +};
>> +
>>  struct xdp_buff {
>>  	void *data;
>>  	void *data_end;
>> @@ -70,6 +74,7 @@ struct xdp_buff {
>>  	void *data_hard_start;
>>  	unsigned long handle;
>>  	struct xdp_rxq_info *rxq;
>> +	struct xdp_txq_info *txq;
>>  };
>>  
>>  struct xdp_frame {
>> diff --git a/include/uapi/linux/bpf.h b/include/uapi/linux/bpf.h
>> index 7850f8683b81..5e3f8aefad41 100644
>> --- a/include/uapi/linux/bpf.h
>> +++ b/include/uapi/linux/bpf.h
>> @@ -3334,8 +3334,10 @@ struct xdp_md {
>>  	__u32 data;
>>  	__u32 data_end;
>>  	__u32 data_meta;
>> -	/* Below access go through struct xdp_rxq_info */
>> -	__u32 ingress_ifindex; /* rxq->dev->ifindex */
>> +	union {
>> +		__u32 ingress_ifindex; /* rxq->dev->ifindex */
>> +		__u32 egress_ifindex;  /* txq->dev->ifindex */
>> +	};
>
> Are we sure it is wise to "union share" (struct) xdp_md as the
> XDP-context in the XDP programs, with different expected_attach_type?
> As this allows the XDP-programmer to code an EGRESS program that access
> ctx->ingress_ifindex, this will under the hood be translated to
> ctx->egress_ifindex, because from the compilers-PoV this will just be an
> offset.
>
> We are setting up the XDP-programmer for a long debugging session, as
> she will be expecting to read 'ingress_ifindex', but will be getting
> 'egress_ifindex'.  (As the compiler cannot warn her, and it is also
> correct seen from the verifier).

+1 on this; also, an egress program may want to actually know which
ingress iface the packet was first received on. So why not just keep
both fields? Since ifindex 0 is invalid anyway, the field could just be
0 when it isn't known (e.g., egress ifindex on RX, or ingress ifindex if
it comes from the stack)?

>>  	__u32 rx_queue_index;  /* rxq->queue_index  */
>
> So, the TX program can still read 'rx_queue_index', is this wise?

Why shouldn't it be able to (as well as ingress ifindex)?

-Toke