Re: [PATCH iwl-next 09/15] idpf: psp: add support for psp key rotation
From: Daniel Zahka
Date: Tue Nov 17 2026 - 17:15:16 EST
On Tue Nov 17, 2026 at 9:55 AM EST, Stan Iliev wrote:
> Add support for PSP master key rotation:
> - Introduce VIRTCHNL2_OP_PSP_ROTATE_KEYGEN_REQUEST (563),
> struct virtchnl2_psp_rotate_keygen_request/_response,
> VIRTCHNL2_EVENT_PSP_ROTATE_KEY (4) and
> struct virtchnl2_psp_rotation_event/_data.
> - Implement the key_rotate PSP device op with
> idpf_psp_rotation_request(), which fails with -EOPNOTSUPP if
> VIRTCHNL2_CAP_PSP_REQUEST_ROTATION was not granted. The PSP core
> increments the PSP device generation before calling the op and
> doesn't restore it if the op fails: the op restores it then.
> - Track the master key of the allocated RX SPIs, to report their
> generation on receive in a later patch: bit 31 of an SPI selects one
> of the two master keys of the CP, and the CP master key generation
> of the allocation tells when a master key was replaced. The packets
> of a master key get the PSP device generation that was current when
> its first SPI was allocated, which is the generation of the
> associations using it. Associations then survive one rotation, as the
> PSP core expects.
> - Handle VIRTCHNL2_EVENT_PSP_ROTATE_KEY in idpf_recv_event_msg() with
> idpf_psp_rotation_event(), which logs the new master key generation.
> The PSP core can't be told about rotations it didn't request, like
> in mlx5: document the limitation and a possible solution.
>
> Signed-off-by: Stan Iliev <stani@xxxxxxxxxx>
> ---
[...]
> +/**
> + * idpf_psp_rx_gen_update - Track the master key of an allocated SPI
> + * @adapter: private data struct
> + * @psd: PSP device, locked
> + * @spi: SPI allocated by the CP
> + * @cp_gen: CP master key generation of @spi
> + *
> + * Bit 31 of an SPI selects one of the two master keys of the CP, a rotation
> + * replaces the older one. The HW doesn't tell which master key decrypted a
> + * packet: report the PSP device generation that was current when the first
> + * SPI of the master key was allocated, which is the generation of the
> + * associations using it, see idpf_psp_rx_process_skb(). The CP master key
> + * generation tells when a master key was replaced, including by a rotation
> + * the PSP core didn't request, see idpf_psp_rotation_event().
> + */
> +static void idpf_psp_rx_gen_update(struct idpf_adapter *adapter,
> + struct psp_dev *psd, __be32 spi, u32 cp_gen)
> +{
> + struct idpf_psp_rx_gen *rx_gen;
> +
> + lockdep_assert_held(&psd->lock);
> +
> + rx_gen = &adapter->psp.rx_gen[be32_to_cpu(spi) >> 31];
> + if (rx_gen->gen != IDPF_PSP_GEN_UNKNOWN && rx_gen->cp_gen == cp_gen)
> + return;
> +
> + rx_gen->cp_gen = cp_gen;
> + WRITE_ONCE(rx_gen->gen, psd->generation);
WRITE_ONCE isn't paired here. I'm guessing the Rx code in series 2 will
copy rx_gen->gen into the psp skb extension, but that would likely not
be correct. The generation field in the extension should be copied from
Rx completion metadata. If such a descriptor field doesn't exist, then
the key_rotate() implementation might as well pin psd->generation = 0
like mlx5 does.
> +}
> +