David Howells <dhowells@xxxxxxxxxx> writes:
>
> Fix af_alg_alloc_areq() to initialise areq->first_rsgl.sgl.sgt.sgl to point
> to the scatterlist array in areq->first_rsgl.sgl.sgl.
>
> Without this, the gcm-aes-s390 driver will oops when it tries to do
> gcm_walk_start() on req->dst because req->dst is set to the value of
> areq->first_rsgl.sgl.sgl by _aead_recvmsg() calling
> aead_request_set_crypt().
>
> The problem comes if an empty ciphertext is passed: the loop in
> af_alg_get_rsgl() just passes straight out and doesn't set areq->first_rsgl
> up.
>
> This isn't a problem on x86_64 using gcmaes_crypt_by_sg() because, as far
> as I can tell, that ignores req->dst and only uses req->src[*].
>
> [*] Is this a bug in aesni-intel_glue.c?
>
> The s390x oops looks something like:
>
> Unable to handle kernel pointer dereference in virtual kernel address space
> Failing address: 0000000a00000000 TEID: 0000000a00000803
> Fault in home space mode while using kernel ASCE.
> AS:00000000a43a0007 R3:0000000000000024
> Oops: 003b ilc:2 [#1] SMP
> ...
> Call Trace:
> [<000003ff7fc3d47e>] gcm_walk_start+0x16/0x28 [aes_s390]
> [<00000000a2a342f2>] crypto_aead_decrypt+0x9a/0xb8
> [<00000000a2a60888>] aead_recvmsg+0x478/0x698
> [<00000000a2e519a0>] sock_recvmsg+0x70/0xb0
> [<00000000a2e51a56>] sock_read_iter+0x76/0xa0
> [<00000000a273e066>] vfs_read+0x26e/0x2a8
> [<00000000a273e8c4>] ksys_read+0xbc/0x100
> [<00000000a311d808>] __do_syscall+0x1d0/0x1f8
> [<00000000a312ff30>] system_call+0x70/0x98
> Last Breaking-Event-Address:
> [<000003ff7fc3e6b4>] gcm_aes_crypt+0x104/0xa68 [aes_s390]
>
> Fixes: c1abe6f570af ("crypto: af_alg: Use extract_iter_to_sg() to create scatterlists")
> Reported-by: Ondrej Mosnáček <omosnacek@xxxxxxxxx>
> Link: https://lore.kernel.org/r/CAAUqJDuRkHE8fPgZJGaKjUjd3QfGwzfumuJBmStPqBhubxyk_A@xxxxxxxxxxxxxx/
> Signed-off-by: David Howells <dhowells@xxxxxxxxxx>
> cc: Herbert Xu <herbert@xxxxxxxxxxxxxxxxxxx>
> cc: Sven Schnelle <svens@xxxxxxxxxxxxx>
> cc: Harald Freudenberger <freude@xxxxxxxxxxxxxxxxxx>
> cc: "David S. Miller" <davem@xxxxxxxxxxxxx>
> cc: Paolo Abeni <pabeni@xxxxxxxxxx>
> cc: linux-crypto@xxxxxxxxxxxxxxx
> cc: linux-s390@xxxxxxxxxxxxxxx
> cc: regressions@xxxxxxxxxxxxxxx
> ---
> crypto/af_alg.c | 1 +
> 1 file changed, 1 insertion(+)
>
> diff --git a/crypto/af_alg.c b/crypto/af_alg.c
> index 06b15b9f661c..9ee8575d3b1a 100644
> --- a/crypto/af_alg.c
> +++ b/crypto/af_alg.c
> @@ -1192,6 +1192,7 @@ struct af_alg_async_req *af_alg_alloc_areq(struct sock *sk,
>
> areq->areqlen = areqlen;
> areq->sk = sk;
> + areq->first_rsgl.sgl.sgt.sgl = areq->first_rsgl.sgl.sgl;
> areq->last_rsgl = NULL;
> INIT_LIST_HEAD(&areq->rsgl_list);
> areq->tsgl = NULL;
Just tested, with this fix the kernel no longer crashes. Thanks!
Tested-by: Sven Schnelle <svens@xxxxxxxxxxxxx>
