Re: [PATCH net v3] net: tls: use sync AEAD for sk_msg BPF sockets

[Jiayuan Chen <jiayuan.chen@linux.dev>]

On 5/26/26 10:51 AM, Christopher Lusk wrote:
> The kTLS TX path can hand an open record to a sk_msg verdict
> program before encryption.  If the verdict applies fewer bytes
> than the open record contains, tls_push_record() splits
> ctx->open_rec into the record being encrypted and a remainder.
> The synchronous path reattaches that remainder before continuing.
>
> With an async AEAD provider, crypto_aead_encrypt() can return
> -EINPROGRESS after ctx->open_rec has been unhooked but before the
> split remainder is reattached.  The remainder is no longer
> reachable through ctx->open_rec or ctx->tx_list, silently dropping
> transmitted data and leaking the unreachable tls_rec.  The same
> composition also entangles the user-page zerocopy lifetime rules
> with an async completion path.
>
> A sockmap cannot be attached to a socket after an inet ULP is
> installed: sk_psock_init() returns -EINVAL when
> inet_csk_has_ulp() is true.  So the supported ordering for
> sockmap + kTLS TX is sockmap first, TLS_TX setup second.  When
> TLS_TX setup sees an existing sk_psock, allocate the AEAD with
> CRYPTO_ALG_ASYNC masked out and latch the TX zerocopy gate
> (sw_ctx_tx->async_capable) so the buggy composition becomes
> structurally unreachable.  Ordinary kTLS sockets without sk_msg
> BPF attached are unaffected and continue to use async-capable
> providers.
>
> Fixes: d3b18ad31f93 ("tls: add bpf support to sk_msg handling")
> Cc: stable@vger.kernel.org # 4.20+
> Signed-off-by: Christopher Lusk <clusk@northecho.dev>
> Assisted-by: Codex:gpt-5.5
> Assisted-by: Claude:claude-opus-4-7
> ---
>
> Changes since v2 [1]:
> - Per netdev maintainer guidance [2], replace the Option-C
>    drain-on-error fix with a setup-time surface narrowing in
>    tls_set_sw_offload(): when a sockmap is already attached at
>    TLS_TX setup, request a synchronous AEAD (CRYPTO_ALG_ASYNC in
>    the allocation mask) and set sw_ctx_tx->async_capable = 1.
>    Both moves are needed: latching async_capable alone disables
>    zerocopy but tls_do_encryption() can still return -EINPROGRESS
>    on the copy path; selecting a sync provider removes that return
>    path for sk_msg-attached sockets.
> - Drop the selftest from the series per Jakub's note that the
>    existing sockmap + TLS coverage at
>    tools/testing/selftests/bpf/prog_tests/sockmap_ktls.c exercises
>    this configuration [3].  That suite covers sockmap + kTLS
>    policy paths broadly; the specific async-pcrypt pass-then-drop
>    failure mode from the v2 reproducer was validated for v3 on
>    QEMU/KVM with a KASAN+LOCKDEP-instrumented kernel against net
>    base 2156a29aecff before send.
> - Single-patch series.
>
> Changes since v1:
> - v1's remainder-rooting fix was incomplete; Sashiko AI review
>    surfaced a real UAF in the v2 follow-up that John Fastabend
>    endorsed on the v1 thread [4].  The surface-narrowing approach
>    in v3 makes both failure modes unreachable by avoiding the
>    async + sk_msg composition entirely rather than patching each
>    continuation point.
>
> [1] https://lore.kernel.org/all/20260521025840.976378-1-clusk@northecho.dev/
> [2] https://lore.kernel.org/all/20260525133028.58494274@kernel.org/
> [3] https://lore.kernel.org/all/20260525133048.2dc6d8d3@kernel.org/
> [4] https://lore.kernel.org/all/huduxtn6parzgiaf5cyiyrrvjjvx6jsdedowvrd4nkwmuyeind@j6migjgofh2i/
>
>   net/tls/tls_sw.c | 15 ++++++++++++++-
>   1 file changed, 14 insertions(+), 1 deletion(-)
>
> diff --git a/net/tls/tls_sw.c b/net/tls/tls_sw.c
> index 964ebc268..0000000 100644
> --- a/net/tls/tls_sw.c
> +++ b/net/tls/tls_sw.c
> @@ -2867,7 +2867,20 @@ int tls_set_sw_offload(struct sock *sk, int tx,
>   	rec_seq = crypto_info_rec_seq(src_crypto_info, cipher_desc);
>
>   	if (!*aead) {
> -		*aead = crypto_alloc_aead(cipher_desc->cipher_name, 0, 0);
> +		u32 mask = 0;
> +
> +		if (tx) {
> +			struct sk_psock *psock;
> +
> +			psock = sk_psock_get(sk);
> +			if (psock) {
> +				mask = CRYPTO_ALG_ASYNC;
> +				sw_ctx_tx->async_capable = 1;
> +				sk_psock_put(sk, psock);
> +			}
> +		}
> +
> +		*aead = crypto_alloc_aead(cipher_desc->cipher_name, 0, mask);
>   		if (IS_ERR(*aead)) {
>   			rc = PTR_ERR(*aead);
>   			*aead = NULL;
> --
> 2.54.0

If async_capable is set to 1, the zerocopy path in tls_sw_sendmsg() is 
skipped.
Unfortunately ktls with bpf_msg_pop_data() does not work correctly under 
this
copy path.

tls_clone_plaintext_msg() aliases msg_pl onto msg_en's plaintext area 
(in-place encryption).

BPF runs bpf_msg_pop_data(msg, 0, 2). This shifts msg_pl's SG entry 
forward by 2 bytes.
The two SGs now point to the same page at different offsets. Physical 
memory overlaps but the start of
address differ.

I think selecting a sync provider via mask = CRYPTO_ALG_ASYNC is 
sufficient to
remove the -EINPROGRESS return path.

May be time to remove skmsg from ktls? (disable by default first, 
re-enable via a new ktls module_param?)