Re: [PATCH bpf v2] xsk: Fix out of order segment free in __xsk_generic_xmit()

[Stanislav Fomichev <stfomichev@gmail.com>]

On 05/30, e.kubanski wrote:
> Move xsk completion queue descriptor write-back to destructor.
> 
> Fix xsk descriptor management in completion queue. Descriptor
> management mechanism didn't take care of situations where
> completion queue submission can happen out-of-order to
> descriptor write-back.
> 
> __xsk_generic_xmit() was assigning descriptor to slot right
> after completion queue slot reservation. If multiple CPUs
> access the same completion queue after xmit, this can result
> in out-of-order submission of invalid descriptor batch.
> SKB destructor call can submit descriptor batch that is
> currently in use by other CPU, instead of correct transmitted
> ones. This could result in User-Space <-> Kernel-Space data race.
> 
> Forbid possible out-of-order submissions:
> CPU A: Reservation + Descriptor Write
> CPU B: Reservation + Descriptor Write
> CPU B: Submit (submitted first batch reserved by CPU A)
> CPU A: Submit (submitted second batch reserved by CPU B)
> 
> Move Descriptor Write to submission phase:
> CPU A: Reservation (only moves local writer)
> CPU B: Reservation (only moves local writer)
> CPU B: Descriptor Write + Submit
> CPU A: Descriptor Write + Submit
> 
> This solves potential out-of-order free of xsk buffers.

I'm not sure I understand what's the issue here. If you're using the
same XSK from different CPUs, you should take care of the ordering
yourself on the userspace side?

> Signed-off-by: Eryk Kubanski <e.kubanski@partner.samsung.com>
> Fixes: e6c4047f5122 ("xsk: Use xsk_buff_pool directly for cq functions")
> ---
>  include/linux/skbuff.h |  2 ++
>  net/xdp/xsk.c          | 17 +++++++++++------
>  net/xdp/xsk_queue.h    | 11 +++++++++++
>  3 files changed, 24 insertions(+), 6 deletions(-)
> 
> diff --git a/include/linux/skbuff.h b/include/linux/skbuff.h
> index 5520524c93bf..cc37b62638cd 100644
> --- a/include/linux/skbuff.h
> +++ b/include/linux/skbuff.h
> @@ -624,6 +624,8 @@ struct skb_shared_info {
>  		void		*destructor_arg;
>  	};
>  
> +	u64 xsk_descs[MAX_SKB_FRAGS];

This is definitely a no-go (sk_buff and skb_shared_info space is
precious).