Re: Sharing UMEM without XDP_SHARED_UMEM

On Thu, 5 Dec 2024 at 15:00, Richard Patel <[email protected]> wrote:
>
> Hello,
>
> Is it safe to share the same virtual memory region as UMEM between two ze=
ro-copy XSKs without setting XDP_SHARED_UMEM?
> The XSKs do not share the same (dev, queue) pair. They are backed by sepa=
rate receive queues and have separate fill and completion rings.

It is safe to share a umem by just registering the same umem region
for each socket. This was the method before XDP_SHARED_UMEM. The
drawback is that it consumes more memory as the xsk code in the kernel
does not understand that it is indeed shared and structures could be
shared too.

> I have an application receiving and sending packets via two such XSKs on =
the same thread.
> I=E2=80=99d like to reuse buffers across both XSKs while ensuring each ch=
unk is only submitted to one of the fill or TX rings.
>
> In principle, this seems safe but undocumented.
> If queue IDs or devices between two XSKs differ, the only notable differe=
nce in behavior when setting XDP_SHARED_UMEM is that xp_assign_dev_shared i=
s called instead of xp_assign_dev. However, xdp_assign_dev_shared only copi=
es XSK flags and then tail calls to xp_assign_dev.
>
> Is there any special behaviour introduced by XDP_SHARED_UMEM in this case=
 that I=E2=80=99m missing?
> (Perhaps something to do with ref counting of memory pinning, or support =
for simultaneous DMAs from different devices into the same page?)
> And is it possible that future changes to the kernel AF_XDP code or kerne=
l drivers might break such operation without XDP_SHARED_UMEM?
>
> Thanks
> Richard