Re: Re: [PATCH bpf v2] xsk: Fix out of order segment free in __xsk_generic_xmit()
From: Stanislav Fomichev
Date: Mon Jun 02 2025 - 11:51:00 EST
On 06/02, Eryk Kubanski wrote:
> > 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?
>
> It's not a problem with user-space Completion Queue READER side.
> Im talking exclusively about kernel-space Completion Queue WRITE side.
>
> This problem can occur when multiple sockets are bound to the same
> umem, device, queue id. In this situation Completion Queue is shared.
> This means it can be accessed by multiple threads on kernel-side.
> Any use is indeed protected by spinlock, however any write sequence
> (Acquire write slot as writer, write to slot, submit write slot to reader)
> isn't atomic in any way and it's possible to submit not-yet-sent packet
> descriptors back to user-space as TX completed.
>
> Up untill now, all write-back operations had two phases, each phase
> locks the spinlock and unlocks it:
> 1) Acquire slot + Write descriptor (increase cached-writer by N + write values)
> 2) Submit slot to the reader (increase writer by N)
>
> Slot submission was solely based on the timing. Let's consider situation,
> where two different threads issue a syscall for two different AF_XDP sockets
> that are bound to the same umem, dev, queue-id.
>
> AF_XDP setup:
>
> kernel-space
>
> Write Read
> +--+ +--+
> | | | |
> | | | |
> | | | |
> Completion | | | | Fill
> Queue | | | | Queue
> | | | |
> | | | |
> | | | |
> | | | |
> +--+ +--+
> Read Write
> user-space
>
>
> +--------+ +--------+
> | AF_XDP | | AF_XDP |
> +--------+ +--------+
>
>
>
>
>
> Possible out-of-order scenario:
>
>
> writer cached_writer1 cached_writer2
> | | |
> | | |
> | | |
> | | |
> +--------------|--------|--------|--------|--------|--------|--------|----------------------------------------------+
> | | | | | | | | |
> Completion Queue | | | | | | | | |
> | | | | | | | | |
> +--------------|--------|--------|--------|--------|--------|--------|----------------------------------------------+
> | | |
> | | |
> |-----------------| |
> A) T1 syscall | |
> writes 2 | |
> descriptors |-----------------------------------|
> B) T2 syscall writes 4 descriptors
>
>
>
>
> Notes:
> 1) T1 and T2 AF_XDP sockets are two different sockets,
> __xsk_generic_xmit will obtain two different mutexes.
> 2) T1 and T2 can be executed simultaneously, there is no
> critical section whatsoever between them.
XSK represents a single queue and each queue is single producer single
consumer. The fact that you can dup a socket and call sendmsg from
different threads/processes does not lift that restriction. I think
if you add synchronization on the userspace (lock(); sendmsg();
unlock();), that should help, right?