Re: [PATCH net v4] xsk: Do not free ZC Rx buffer when the Rx ring is full
From: Marcin Knap
Date: Thu Nov 12 2026 - 07:08:51 EST
>Hmm. So it seems like currently, when we managed to properly copy from zc
>the 'normal' page and redirect failed, we would interpret this as (based
>on ice implementation, not sure about other zc drivers) ICE_XDP_CONSUMED
>which would do another xsk_buff_free() which would be a no-op on this
>xdp_buff at this point but 'normal' page would indeed leak?
Yes, that's my understanding too. xdp_convert_zc_to_xdp_frame() puts the
ZC buffer back on the free list right after the copy, and when
__xdp_do_redirect_frame() then fails nothing frees the copied frame,
so that page leaks. On ice the following xsk_buff_free() for
ICE_XDP_CONSUMED is a no-op, since xp_free() skips a buffer that is
already on the free list.
That only works because ice frees right away, before anything else is
allocated from the pool. Drivers that release the buffer later, like
mlx5 on Rx WQE refill, can still free a buffer that has been handed
out again in the meantime, the same way as on the XSKMAP path.
If the copy itself fails, the ZC buffer is not freed and the driver
frees it once, so that case is fine.
In short, for failed __xdp_do_redirect_frame():
- the copied page leaks, in every ZC driver
- the ZC buffer is freed twice, which is harmless when the driver
frees it right away (ice, i40e, ...) but can corrupt the pool when
it frees it later (e.g. mlx5).
The patch does not affect that path.