Re: [PATCH net] net: mana: fix XDP_REDIRECT of pre-allocated RX buffers

[Orgad Shaneh <orgads@gmail.com>]

On Wed, Nov 04, 2026 at 05:22:24PM +0100, Maciej Fijalkowski wrote:
> On Mon, Nov 02, 2026 at 06:43:24AM +0000, Orgad Shaneh wrote:
[...]
> > -	act = mana_run_xdp(ndev, rxq, &xdp, buf_va, pkt_len);
> > +	act = mana_run_xdp(ndev, rxq, &xdp, buf_va, from_pool, pkt_len);
>
> I wonder if it wouldn't be better to work with rxoob objects which seem to
> carry the from_pool setting?

Thanks for the review.

The rxoob no longer describes this buffer when XDP runs.
mana_process_one_rx_pkt() calls mana_refill_rx_oob() before
mana_rx_skb(), and the refill puts the new buffer into the same rxoob:

	*old_buf = rxoob->buf_va;
	*old_fp = rxoob->from_pool;

	rxoob->buf_va = va;
	...
	rxoob->from_pool = from_pool;

So by the time mana_run_xdp() runs, rxoob->from_pool is the flag of the
replacement buffer. The flag of the received buffer survives only as
old_fp, which is what mana_rx_skb() already gets as its from_pool
argument and uses to free the buffer on PASS/TX/DROP. The patch passes
that same value one call further down, to mana_run_xdp().

The two can differ. Refills come from the page pool (or from
xdp_save_va, which only holds non-pool buffers), so on the first pass
over a ring of pre-allocated buffers XDP gets a buffer with
from_pool == false while its rxoob already says true. That first pass
is exactly where the bug shows.

Using the rxoob would mean moving the refill after XDP, which changes
what happens when a refill fails: today the packet is dropped and its
buffer stays in the ring. I preferred not to touch that in a fix.

If you would rather mana_run_xdp() take something other than a bool,
I can pass the xdp_rxq_info to use instead, chosen in mana_rx_skb().

Orgad