Re: [PATCH RFC net-next 0/4] bonding: support LAG IPsec offload with replicated SAs

[Jihong Min <hurryman2212@gmail.com>]

On 6/11/26 19:06, Leon Romanovsky wrote:
> On Thu, Jun 11, 2026 at 05:56:17PM +0900, Jihong Min wrote:
>> Hi,
>>
>> On 6/10/26 23:18, Leon Romanovsky wrote:
>>> On Wed, May 20, 2026 at 05:10:00PM +0900, Jihong Min wrote:
>>>> This RFC adds a bonding model for IPsec/XFRM hardware offload on
>>>> 802.3ad and balance-xor LAG devices when the transmit hash policy is
>>>> layer3+4. This is an intentional scope limit rather than a hard limit,
>>>> as this is the configuration I can test with my gear.
>>>>
>>>> The main idea is to leave the existing upstream single-lower-device XFRM
>>>> offload path for active-backup intentionally untouched, while adding a
>>>> replicated state model for LAG.
>>>>
>>>> For LAG bonds, the bonding driver installs the same XFRM state on every
>>>> eligible running slave and stores the per-slave hardware handles in
>>>> bonding-private state. Lower drivers that support this model can then
>>>> resolve the handle for the concrete lower netdev used by the datapath.
>>>>
>>>> LAG IPsec features are user controlled. Newly eligible LAG bonds start
>>>> with the ESP/XFRM features disabled, but advertise supported mutable
>>>> features when all running eligible slaves can support them. Users can
>>>> then opt in with ethtool. Feature enable is propagated to the lower
>>>> devices and rolled back if a lower device cannot enable the requested
>>>> features.
>>>>
>>>> The series also handles LAG membership and eligibility changes by adding
>>>> replicated SAs to newly usable slaves, removing the departing lower
>>>> instance on down/remove, and flushing bond-owned XFRM offload state when
>>>> the bond leaves the supported mode or hash-policy configuration.
>>>>
>>>> This series does not convert any physical NIC driver. A lower driver
>>>> must explicitly opt in to the replicated-upper-device model before it can
>>>> use these bond-owned states in its datapath.
>>>>
>>>> For example, a driver such as mlx5 would opt in by marking its
>>>> xfrmdev_ops and by resolving datapath handles through the helper:
>>>>
>>>>         static const struct xfrmdev_ops mlx5e_ipsec_xfrmdev_ops = {
>>>>                 ...
>>>>                 .xdo_dev_state_lower_handle = NULL,
>>>>                 .flags = XFRMDEV_OPS_F_LOWER_HANDLE,
>>>>         };
>>>>
>>>>         handle = xfrm_dev_state_lower_handle(x, netdev);
>>>>         if (!handle)
>>>>                 goto drop;
>>>>
>>>>         sa_entry = (struct mlx5e_ipsec_sa_entry *)handle;
>>>
>>> I’m curious how you replicate and maintain the hardware state across these
>>> devices. How are you handling the anti-replay window?
>>>
>>> Thanks
>>>
>>
>> The short answer is that the RFC I sent was not complete enough in this
>> area.
> 
> The issue is that you are adding support for crypto offload where the
> anti-replay window is managed in software. You must ensure that all SAs can
> safely share and update this state.
> 
> Thanks

I agree.

Generic XFRM crypto/packet offload with bonding enabled is not
impossible, but crypto offload needs a different contract from packet
offload. If the sequence number and replay window are still owned by the
single software xfrm_state, and the lower devices only hold crypto
context, then replicated lower handles may be valid. But the RFC did not
prove or document that, and it used the same weak opt-in for both CRYPTO
and PACKET.

For a respin I should either leave XFRM_DEV_OFFLOAD_CRYPTO out of the
replicated LAG path at first, or add a separate capability/contract for
crypto offload which states that the software xfrm_state safely owns and
updates the shared sequence/replay state across all lower instances.