Re: [PATCH 2/2 bpf v2] xsk: avoid data corruption on cq descriptor number

[Jakub Kicinski <kuba@kernel.org>]

On Wed, 29 Oct 2025 08:51:58 +0100 Fernando Fernandez Mancera wrote:
> On 10/29/25 12:01 AM, Jakub Kicinski wrote:
> > On Tue, 28 Oct 2025 19:30:32 +0100 Fernando Fernandez Mancera wrote:  
> >> Since commit 30f241fcf52a ("xsk: Fix immature cq descriptor
> >> production"), the descriptor number is stored in skb control block and
> >> xsk_cq_submit_addr_locked() relies on it to put the umem addrs onto
> >> pool's completion queue.  
> > 
> > Looking at the past discussion it sounds like you want to optimize
> > the single descriptor case? Can you not use a magic pointer for that?
> > 
> > 	#define XSK_DESTRUCT_SINGLE_BUF	(void *)1
> > 	destructor_arg = XSK_DESTRUCT_SINGLE_BUF
> > 
> > Let's target this fix at net, please, I think the complexity here is
> > all in skbs paths.  
> 
> I might be missing something here but if the destructor_arg pointer is 
> used to do this, where should we store the umem address associated with 
> it? In the proposed approach the skb extension should not be increased 
> for non-fragmented traffic as there is only a single descriptor and 
> therefore we can store the umem address in destructor_arg directly.

I see. Pointers are always aligned to 8B, you can stash the "pointer
type" there. If the bottom bit is 1 it's a umem and the skb was
single-chunk. If it's non-0 then it's a full kmalloc'ed struct.

> The size of the skb extension will only increase for fragmented traffic 
> (multiple descriptors).. but sure, if there is a fallback to the 
> slowpath, it will burden a bit the performance. Although, for that to 
> happen the must have tried to use AF_XDP family initially.. AFAICS, the 
> size of skb extension is only increased when skb_ext_add() is called.

To be clear by adding an skb extension you are de-facto allocating 
a bit in the skb struct. Just one of the bits of the active_extensions
field instead of a separate bitfield. If you can depend on the socket
association instead this is quite wasteful.