Merge branch 'bpf-sockmap-tls-fixes'

John Fastabend says:

====================
To date our usage of sockmap/tls has been fairly simple, the BPF programs
did only well-defined pop, push, pull and apply/cork operations.

Now that we started to push more complex programs into sockmap we uncovered
a series of issues addressed here. Further OpenSSL3.0 version should be
released soon with kTLS support so its important to get any remaining
issues on BPF and kTLS support resolved.

Additionally, I have a patch under development to allow sockmap to be
enabled/disabled at runtime for Cilium endpoints. This allows us to stress
the map insert/delete with kTLS more than previously where Cilium only
added the socket to the map when it entered ESTABLISHED state and never
touched it from the control path side again relying on the sockets own
close() hook to remove it.

To test I have a set of test cases in test_sockmap.c that expose these
issues. Once we get fixes here merged and in bpf-next I'll submit the
tests to bpf-next tree to ensure we don't regress again. Also I've run
these patches in the Cilium CI with OpenSSL (master branch) this will
run tools such as netperf, ab, wrk2, curl, etc. to get a broad set of
testing.

I'm aware of two more issues that we are working to resolve in another
couple (probably two) patches. First we see an auth tag corruption in
kTLS when sending small 1byte chunks under stress. I've not pinned this
down yet. But, guessing because its under 1B stress tests it must be
some error path being triggered. And second we need to ensure BPF RX
programs are not skipped when kTLS ULP is loaded. This breaks some of the
sockmap selftests when running with kTLS. I'll send a follow up for this.

v2: I dropped a patch that added !0 size check in tls_push_record
    this originated from a panic I caught awhile ago with a trace
    in the crypto stack. But I can not reproduce it anymore so will
    dig into that and send another patch later if needed. Anyways
    after a bit of thought it would be nicer if tls/crypto/bpf didn't
    require special case handling for the !0 size.
====================

Signed-off-by: Daniel Borkmann <daniel@iogearbox.net>

2 files changed, 34 insertions, 7 deletions

diff --git a/net/tls/tls_main.c b/net/tls/tls_main.c
index dac24c7aa7d4..94774c0e5ff3 100644
--- a/net/tls/tls_main.c
+++ b/net/tls/tls_main.c
@@ -732,15 +732,19 @@ out:
 	return rc;
 }
 
-static void tls_update(struct sock *sk, struct proto *p)
+static void tls_update(struct sock *sk, struct proto *p,
+		       void (*write_space)(struct sock *sk))
 {
 	struct tls_context *ctx;
 
 	ctx = tls_get_ctx(sk);
-	if (likely(ctx))
+	if (likely(ctx)) {
+		ctx->sk_write_space = write_space;
 		ctx->sk_proto = p;
-	else
+	} else {
 		sk->sk_prot = p;
+		sk->sk_write_space = write_space;
+	}
 }
 
 static int tls_get_info(const struct sock *sk, struct sk_buff *skb)
diff --git a/net/tls/tls_sw.c b/net/tls/tls_sw.c
index c6803a82b769..159d49dab403 100644
--- a/net/tls/tls_sw.c
+++ b/net/tls/tls_sw.c
@@ -682,12 +682,32 @@ static int tls_push_record(struct sock *sk, int flags,
 
 	split_point = msg_pl->apply_bytes;
 	split = split_point && split_point < msg_pl->sg.size;
+	if (unlikely((!split &&
+		      msg_pl->sg.size +
+		      prot->overhead_size > msg_en->sg.size) ||
+		     (split &&
+		      split_point +
+		      prot->overhead_size > msg_en->sg.size))) {
+		split = true;
+		split_point = msg_en->sg.size;
+	}
 	if (split) {
 		rc = tls_split_open_record(sk, rec, &tmp, msg_pl, msg_en,
 					   split_point, prot->overhead_size,
 					   &orig_end);
 		if (rc < 0)
 			return rc;
+		/* This can happen if above tls_split_open_record allocates
+		 * a single large encryption buffer instead of two smaller
+		 * ones. In this case adjust pointers and continue without
+		 * split.
+		 */
+		if (!msg_pl->sg.size) {
+			tls_merge_open_record(sk, rec, tmp, orig_end);
+			msg_pl = &rec->msg_plaintext;
+			msg_en = &rec->msg_encrypted;
+			split = false;
+		}
 		sk_msg_trim(sk, msg_en, msg_pl->sg.size +
 			    prot->overhead_size);
 	}
@@ -709,6 +729,12 @@ static int tls_push_record(struct sock *sk, int flags,
 		sg_mark_end(sk_msg_elem(msg_pl, i));
 	}
 
+	if (msg_pl->sg.end < msg_pl->sg.start) {
+		sg_chain(&msg_pl->sg.data[msg_pl->sg.start],
+			 MAX_SKB_FRAGS - msg_pl->sg.start + 1,
+			 msg_pl->sg.data);
+	}
+
 	i = msg_pl->sg.start;
 	sg_chain(rec->sg_aead_in, 2, &msg_pl->sg.data[i]);
 
@@ -783,10 +809,7 @@ more_data:
 	if (psock->eval == __SK_NONE) {
 		delta = msg->sg.size;
 		psock->eval = sk_psock_msg_verdict(sk, psock, msg);
-		if (delta < msg->sg.size)
-			delta -= msg->sg.size;
-		else
-			delta = 0;
+		delta -= msg->sg.size;
 	}
 	if (msg->cork_bytes && msg->cork_bytes > msg->sg.size &&
 	    !enospc && !full_record) {