tls: rx: strp: don't use GFP_KERNEL in softirq context

commit 74836ec upstream.

When receive buffer is small, or the TCP rx queue looks too
complicated to bother using it directly - we allocate a new
skb and copy data into it.

We already use sk->sk_allocation... but nothing actually
sets it to GFP_ATOMIC on the ->sk_data_ready() path.

Users of HW offload are far more likely to experience problems
due to scheduling while atomic. "Copy mode" is very rarely
triggered with SW crypto.

Fixes: 84c61fe ("tls: rx: do not use the standard strparser")
Tested-by: Shai Amiram <samiram@nvidia.com>
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
Reviewed-by: Simon Horman <simon.horman@corigine.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>

net/tls/tls_sw.c

@@ -2289,8 +2289,12 @@ static void tls_data_ready(struct sock *sk)
 	struct tls_context *tls_ctx = tls_get_ctx(sk);
 	struct tls_sw_context_rx *ctx = tls_sw_ctx_rx(tls_ctx);
 	struct sk_psock *psock;
+	gfp_t alloc_save;
 
+	alloc_save = sk->sk_allocation;
+	sk->sk_allocation = GFP_ATOMIC;
 	tls_strp_data_ready(&ctx->strp);
+	sk->sk_allocation = alloc_save;
 
 	psock = sk_psock_get(sk);
 	if (psock) {