Merge branch 'nfp-tls-fixes-for-initial-TLS-support'

Jakub Kicinski says:

====================
nfp: tls: fixes for initial TLS support

This series brings various fixes to nfp tls offload recently added
to net-next.

First 4 patches revolve around device mailbox communication, trying
to make it more reliable. Next patch fixes statistical counter.
Patch 6 improves the TX resync if device communication failed.
Patch 7 makes sure we remove keys from memory after talking to FW.
Patch 8 adds missing tls context initialization, we fill in the
context information from various places based on the configuration
and looks like we missed the init in the case of where TX is
offloaded, but RX wasn't initialized yet. Patches 9 and 10 make
the nfp driver undo TLS state changes if we need to drop the
frame (e.g. due to DMA mapping error).

Last but not least TLS fallback should not adjust socket memory
after skb_orphan_partial(). This code will go away once we forbid
orphaning of skbs in need of crypto, but that's "real" -next
material, so lets do a quick fix.
====================

Signed-off-by: David S. Miller <[e-mail]>

drivers/net/ethernet/mellanox/mlx5/core/en_accel/tls.c

···
 				direction == TLS_OFFLOAD_CTX_DIR_TX);
 }
 
-static void mlx5e_tls_resync(struct net_device *netdev, struct sock *sk,
-			     u32 seq, u8 *rcd_sn_data,
-			     enum tls_offload_ctx_dir direction)
+static int mlx5e_tls_resync(struct net_device *netdev, struct sock *sk,
+			    u32 seq, u8 *rcd_sn_data,
+			    enum tls_offload_ctx_dir direction)
 {
 	struct tls_context *tls_ctx = tls_get_ctx(sk);
 	struct mlx5e_priv *priv = netdev_priv(netdev);
···
 		    be64_to_cpu(rcd_sn));
 	mlx5_accel_tls_resync_rx(priv->mdev, rx_ctx->handle, seq, rcd_sn);
 	atomic64_inc(&priv->tls->sw_stats.rx_tls_resync_reply);
+
+	return 0;
 }
 
 static const struct tlsdev_ops mlx5e_tls_ops = {

drivers/net/ethernet/netronome/nfp/ccm.h

···
 struct sk_buff *
 nfp_ccm_mbox_msg_alloc(struct nfp_net *nn, unsigned int req_size,
 		       unsigned int reply_size, gfp_t flags);
+int __nfp_ccm_mbox_communicate(struct nfp_net *nn, struct sk_buff *skb,
+			       enum nfp_ccm_type type,
+			       unsigned int reply_size,
+			       unsigned int max_reply_size, bool critical);
 int nfp_ccm_mbox_communicate(struct nfp_net *nn, struct sk_buff *skb,
 			     enum nfp_ccm_type type,
 			     unsigned int reply_size,

drivers/net/ethernet/netronome/nfp/ccm_mbox.c

···
  * form a batch.  Threads come in with CMSG formed in an skb, then
  * enqueue that skb onto the request queue.  If threads skb is first
  * in queue this thread will handle the mailbox operation.  It copies
- * up to 16 messages into the mailbox (making sure that both requests
+ * up to 64 messages into the mailbox (making sure that both requests
  * and replies will fit.  After FW is done processing the batch it
  * copies the data out and wakes waiting threads.
  * If a thread is waiting it either gets its the message completed
···
  * to limit potential cache line bounces.
  */
 
-#define NFP_CCM_MBOX_BATCH_LIMIT	16
+#define NFP_CCM_MBOX_BATCH_LIMIT	64
 #define NFP_CCM_TIMEOUT			(NFP_NET_POLL_TIMEOUT * 1000)
-#define NFP_CCM_MAX_QLEN		256
+#define NFP_CCM_MAX_QLEN		1024
 
 enum nfp_net_mbox_cmsg_state {
 	NFP_NET_MBOX_CMSG_STATE_QUEUED,
···
 
 static int
 nfp_ccm_mbox_msg_enqueue(struct nfp_net *nn, struct sk_buff *skb,
-			 enum nfp_ccm_type type)
+			 enum nfp_ccm_type type, bool critical)
 {
 	struct nfp_ccm_hdr *hdr;
 
 	assert_spin_locked(&nn->mbox_cmsg.queue.lock);
 
-	if (nn->mbox_cmsg.queue.qlen >= NFP_CCM_MAX_QLEN) {
+	if (!critical && nn->mbox_cmsg.queue.qlen >= NFP_CCM_MAX_QLEN) {
 		nn_dp_warn(&nn->dp, "mailbox request queue too long\n");
 		return -EBUSY;
 	}
···
 	return 0;
 }
 
-int nfp_ccm_mbox_communicate(struct nfp_net *nn, struct sk_buff *skb,
-			     enum nfp_ccm_type type,
-			     unsigned int reply_size,
-			     unsigned int max_reply_size)
+int __nfp_ccm_mbox_communicate(struct nfp_net *nn, struct sk_buff *skb,
+			       enum nfp_ccm_type type,
+			       unsigned int reply_size,
+			       unsigned int max_reply_size, bool critical)
 {
 	int err;
 
···
 
 	spin_lock_bh(&nn->mbox_cmsg.queue.lock);
 
-	err = nfp_ccm_mbox_msg_enqueue(nn, skb, type);
+	err = nfp_ccm_mbox_msg_enqueue(nn, skb, type, critical);
 	if (err)
 		goto err_unlock;
 
···
 err_free_skb:
 	dev_kfree_skb_any(skb);
 	return err;
+}
+
+int nfp_ccm_mbox_communicate(struct nfp_net *nn, struct sk_buff *skb,
+			     enum nfp_ccm_type type,
+			     unsigned int reply_size,
+			     unsigned int max_reply_size)
+{
+	return __nfp_ccm_mbox_communicate(nn, skb, type, reply_size,
+					  max_reply_size, false);
 }
 
 static void nfp_ccm_mbox_post_runq_work(struct work_struct *work)
···
 
 	spin_lock_bh(&nn->mbox_cmsg.queue.lock);
 
-	err = nfp_ccm_mbox_msg_enqueue(nn, skb, type);
+	err = nfp_ccm_mbox_msg_enqueue(nn, skb, type, false);
 	if (err)
 		goto err_unlock;
 

drivers/net/ethernet/netronome/nfp/crypto/fw.h

···
 	u8 key_len;
 	__be16 ipver_vlan __packed;
 	u8 l4_proto;
+#define NFP_NET_TLS_NON_ADDR_KEY_LEN	8
+	u8 l3_addrs[0];
 };
 
 struct nfp_crypto_req_add_back {

drivers/net/ethernet/netronome/nfp/crypto/tls.c

···
 #include <linux/bitfield.h>
 #include <linux/ipv6.h>
 #include <linux/skbuff.h>
+#include <linux/string.h>
 #include <net/tls.h>
 
 #include "../ccm.h"
···
 	struct nfp_crypto_reply_simple *reply;
 	int err;
 
-	err = nfp_ccm_mbox_communicate(nn, skb, type,
-				       sizeof(*reply), sizeof(*reply));
+	err = __nfp_ccm_mbox_communicate(nn, skb, type,
+					 sizeof(*reply), sizeof(*reply),
+					 type == NFP_CCM_TYPE_CRYPTO_DEL);
 	if (err) {
 		nn_dp_warn(&nn->dp, "failed to %s TLS: %d\n", name, err);
 		return err;
···
 				       NFP_CCM_TYPE_CRYPTO_DEL);
 }
 
+static void
+nfp_net_tls_set_ipver_vlan(struct nfp_crypto_req_add_front *front, u8 ipver)
+{
+	front->ipver_vlan = cpu_to_be16(FIELD_PREP(NFP_NET_TLS_IPVER, ipver) |
+					FIELD_PREP(NFP_NET_TLS_VLAN,
+						   NFP_NET_TLS_VLAN_UNUSED));
+}
+
+static void
+nfp_net_tls_assign_conn_id(struct nfp_net *nn,
+			   struct nfp_crypto_req_add_front *front)
+{
+	u32 len;
+	u64 id;
+
+	id = atomic64_inc_return(&nn->ktls_conn_id_gen);
+	len = front->key_len - NFP_NET_TLS_NON_ADDR_KEY_LEN;
+
+	memcpy(front->l3_addrs, &id, sizeof(id));
+	memset(front->l3_addrs + sizeof(id), 0, len - sizeof(id));
+}
+
 static struct nfp_crypto_req_add_back *
-nfp_net_tls_set_ipv4(struct nfp_crypto_req_add_v4 *req, struct sock *sk,
-		     int direction)
+nfp_net_tls_set_ipv4(struct nfp_net *nn, struct nfp_crypto_req_add_v4 *req,
+		     struct sock *sk, int direction)
 {
 	struct inet_sock *inet = inet_sk(sk);
 
 	req->front.key_len += sizeof(__be32) * 2;
-	req->front.ipver_vlan = cpu_to_be16(FIELD_PREP(NFP_NET_TLS_IPVER, 4) |
-					    FIELD_PREP(NFP_NET_TLS_VLAN,
-						       NFP_NET_TLS_VLAN_UNUSED));
 
 	if (direction == TLS_OFFLOAD_CTX_DIR_TX) {
-		req->src_ip = inet->inet_saddr;
-		req->dst_ip = inet->inet_daddr;
+		nfp_net_tls_assign_conn_id(nn, &req->front);
 	} else {
 		req->src_ip = inet->inet_daddr;
 		req->dst_ip = inet->inet_saddr;
···
 }
 
 static struct nfp_crypto_req_add_back *
-nfp_net_tls_set_ipv6(struct nfp_crypto_req_add_v6 *req, struct sock *sk,
-		     int direction)
+nfp_net_tls_set_ipv6(struct nfp_net *nn, struct nfp_crypto_req_add_v6 *req,
+		     struct sock *sk, int direction)
 {
 #if IS_ENABLED(CONFIG_IPV6)
 	struct ipv6_pinfo *np = inet6_sk(sk);
 
 	req->front.key_len += sizeof(struct in6_addr) * 2;
-	req->front.ipver_vlan = cpu_to_be16(FIELD_PREP(NFP_NET_TLS_IPVER, 6) |
-					    FIELD_PREP(NFP_NET_TLS_VLAN,
-						       NFP_NET_TLS_VLAN_UNUSED));
 
 	if (direction == TLS_OFFLOAD_CTX_DIR_TX) {
-		memcpy(req->src_ip, &np->saddr, sizeof(req->src_ip));
-		memcpy(req->dst_ip, &sk->sk_v6_daddr, sizeof(req->dst_ip));
+		nfp_net_tls_assign_conn_id(nn, &req->front);
 	} else {
 		memcpy(req->src_ip, &sk->sk_v6_daddr, sizeof(req->src_ip));
 		memcpy(req->dst_ip, &np->saddr, sizeof(req->dst_ip));
···
 	front->l4_proto = IPPROTO_TCP;
 
 	if (direction == TLS_OFFLOAD_CTX_DIR_TX) {
-		back->src_port = inet->inet_sport;
-		back->dst_port = inet->inet_dport;
+		back->src_port = 0;
+		back->dst_port = 0;
 	} else {
 		back->src_port = inet->inet_dport;
 		back->dst_port = inet->inet_sport;
···
 	struct nfp_crypto_reply_add *reply;
 	struct sk_buff *skb;
 	size_t req_sz;
+	void *req;
 	bool ipv6;
 	int err;
 
···
 
 	front = (void *)skb->data;
 	front->ep_id = 0;
-	front->key_len = 8;
+	front->key_len = NFP_NET_TLS_NON_ADDR_KEY_LEN;
 	front->opcode = nfp_tls_1_2_dir_to_opcode(direction);
 	memset(front->resv, 0, sizeof(front->resv));
 
+	nfp_net_tls_set_ipver_vlan(front, ipv6 ? 6 : 4);
+
+	req = (void *)skb->data;
 	if (ipv6)
-		back = nfp_net_tls_set_ipv6((void *)skb->data, sk, direction);
+		back = nfp_net_tls_set_ipv6(nn, req, sk, direction);
 	else
-		back = nfp_net_tls_set_ipv4((void *)skb->data, sk, direction);
+		back = nfp_net_tls_set_ipv4(nn, req, sk, direction);
 
 	nfp_net_tls_set_l4(front, back, sk, direction);
 
···
 	memcpy(&back->salt, tls_ci->salt, TLS_CIPHER_AES_GCM_128_SALT_SIZE);
 	memcpy(back->rec_no, tls_ci->rec_seq, sizeof(tls_ci->rec_seq));
 
+	/* Get an extra ref on the skb so we can wipe the key after */
+	skb_get(skb);
+
 	err = nfp_ccm_mbox_communicate(nn, skb, NFP_CCM_TYPE_CRYPTO_ADD,
 				       sizeof(*reply), sizeof(*reply));
+	reply = (void *)skb->data;
+
+	/* We depend on CCM MBOX code not reallocating skb we sent
+	 * so we can clear the key material out of the memory.
+	 */
+	if (!WARN_ON_ONCE((u8 *)back < skb->head ||
+			  (u8 *)back > skb_end_pointer(skb)) &&
+	    !WARN_ON_ONCE((u8 *)&reply[1] > (u8 *)back))
+		memzero_explicit(back, sizeof(*back));
+	dev_consume_skb_any(skb); /* the extra ref from skb_get() above */
+
 	if (err) {
-		nn_dp_warn(&nn->dp, "failed to add TLS: %d\n", err);
+		nn_dp_warn(&nn->dp, "failed to add TLS: %d (%d)\n",
+			   err, direction == TLS_OFFLOAD_CTX_DIR_TX);
 		/* communicate frees skb on error */
 		goto err_conn_remove;
 	}
 
-	reply = (void *)skb->data;
 	err = -be32_to_cpu(reply->error);
 	if (err) {
 		if (err == -ENOSPC) {
···
 	nfp_net_tls_del_fw(nn, ntls->fw_handle);
 }
 
-static void
+static int
 nfp_net_tls_resync(struct net_device *netdev, struct sock *sk, u32 seq,
 		   u8 *rcd_sn, enum tls_offload_ctx_dir direction)
 {
···
 	struct nfp_crypto_req_update *req;
 	struct sk_buff *skb;
 	gfp_t flags;
+	int err;
 
 	flags = direction == TLS_OFFLOAD_CTX_DIR_TX ? GFP_KERNEL : GFP_ATOMIC;
 	skb = nfp_net_tls_alloc_simple(nn, sizeof(*req), flags);
 	if (!skb)
-		return;
+		return -ENOMEM;
 
 	ntls = tls_driver_ctx(sk, direction);
 	req = (void *)skb->data;
···
 	memcpy(req->rec_no, rcd_sn, sizeof(req->rec_no));
 
 	if (direction == TLS_OFFLOAD_CTX_DIR_TX) {
-		nfp_net_tls_communicate_simple(nn, skb, "sync",
-					       NFP_CCM_TYPE_CRYPTO_UPDATE);
+		err = nfp_net_tls_communicate_simple(nn, skb, "sync",
+						     NFP_CCM_TYPE_CRYPTO_UPDATE);
+		if (err)
+			return err;
 		ntls->next_seq = seq;
 	} else {
 		nfp_ccm_mbox_post(nn, skb, NFP_CCM_TYPE_CRYPTO_UPDATE,
 				  sizeof(struct nfp_crypto_reply_simple));
 	}
+
+	return 0;
 }
 
 static const struct tlsdev_ops nfp_net_tls_ops = {

drivers/net/ethernet/netronome/nfp/nfp_net.h

···
  * @tlv_caps:		Parsed TLV capabilities
  * @ktls_tx_conn_cnt:	Number of offloaded kTLS TX connections
  * @ktls_rx_conn_cnt:	Number of offloaded kTLS RX connections
+ * @ktls_conn_id_gen:	Trivial generator for kTLS connection ids (for TX)
  * @ktls_no_space:	Counter of firmware rejecting kTLS connection due to
  *			lack of space
  * @mbox_cmsg:		Common Control Message via vNIC mailbox state
···
 
 	unsigned int ktls_tx_conn_cnt;
 	unsigned int ktls_rx_conn_cnt;
+
+	atomic64_t ktls_conn_id_gen;
 
 	atomic_t ktls_no_space;
 

drivers/net/ethernet/netronome/nfp/nfp_net_common.c

···
 	u64_stats_update_end(&r_vec->tx_sync);
 }
 
-#ifdef CONFIG_TLS_DEVICE
 static struct sk_buff *
 nfp_net_tls_tx(struct nfp_net_dp *dp, struct nfp_net_r_vector *r_vec,
 	       struct sk_buff *skb, u64 *tls_handle, int *nr_frags)
 {
+#ifdef CONFIG_TLS_DEVICE
 	struct nfp_net_tls_offload_ctx *ntls;
 	struct sk_buff *nskb;
 	bool resync_pending;
···
 
 	if (datalen) {
 		u64_stats_update_begin(&r_vec->tx_sync);
-		r_vec->hw_tls_tx++;
+		if (!skb_is_gso(skb))
+			r_vec->hw_tls_tx++;
+		else
+			r_vec->hw_tls_tx += skb_shinfo(skb)->gso_segs;
 		u64_stats_update_end(&r_vec->tx_sync);
 	}
 
 	memcpy(tls_handle, ntls->fw_handle, sizeof(ntls->fw_handle));
 	ntls->next_seq += datalen;
+#endif
 	return skb;
 }
+
+static void nfp_net_tls_tx_undo(struct sk_buff *skb, u64 tls_handle)
+{
+#ifdef CONFIG_TLS_DEVICE
+	struct nfp_net_tls_offload_ctx *ntls;
+	u32 datalen, seq;
+
+	if (!tls_handle)
+		return;
+	if (WARN_ON_ONCE(!skb->sk || !tls_is_sk_tx_device_offloaded(skb->sk)))
+		return;
+
+	datalen = skb->len - (skb_transport_offset(skb) + tcp_hdrlen(skb));
+	seq = ntohl(tcp_hdr(skb)->seq);
+
+	ntls = tls_driver_ctx(skb->sk, TLS_OFFLOAD_CTX_DIR_TX);
+	if (ntls->next_seq == seq + datalen)
+		ntls->next_seq = seq;
+	else
+		WARN_ON_ONCE(1);
 #endif
+}
 
 static void nfp_net_tx_xmit_more_flush(struct nfp_net_tx_ring *tx_ring)
 {
···
 		return NETDEV_TX_BUSY;
 	}
 
-#ifdef CONFIG_TLS_DEVICE
 	skb = nfp_net_tls_tx(dp, r_vec, skb, &tls_handle, &nr_frags);
 	if (unlikely(!skb)) {
 		nfp_net_tx_xmit_more_flush(tx_ring);
 		return NETDEV_TX_OK;
 	}
-#endif
 
 	md_bytes = nfp_net_prep_tx_meta(skb, tls_handle);
 	if (unlikely(md_bytes < 0))
···
 	u64_stats_update_begin(&r_vec->tx_sync);
 	r_vec->tx_errors++;
 	u64_stats_update_end(&r_vec->tx_sync);
+	nfp_net_tls_tx_undo(skb, tls_handle);
 	dev_kfree_skb_any(skb);
 	return NETDEV_TX_OK;
 }

include/net/tls.h

···
 	void (*tls_dev_del)(struct net_device *netdev,
 			    struct tls_context *ctx,
 			    enum tls_offload_ctx_dir direction);
-	void (*tls_dev_resync)(struct net_device *netdev,
-			       struct sock *sk, u32 seq, u8 *rcd_sn,
-			       enum tls_offload_ctx_dir direction);
+	int (*tls_dev_resync)(struct net_device *netdev,
+			      struct sock *sk, u32 seq, u8 *rcd_sn,
+			      enum tls_offload_ctx_dir direction);
 };
 
 enum tls_offload_sync_type {

net/tls/tls_device.c

···
 {
 	struct net_device *netdev;
 	struct sk_buff *skb;
+	int err = 0;
 	u8 *rcd_sn;
 
 	skb = tcp_write_queue_tail(sk);
···
 	down_read(&device_offload_lock);
 	netdev = tls_ctx->netdev;
 	if (netdev)
-		netdev->tlsdev_ops->tls_dev_resync(netdev, sk, seq, rcd_sn,
-						   TLS_OFFLOAD_CTX_DIR_TX);
+		err = netdev->tlsdev_ops->tls_dev_resync(netdev, sk, seq,
+							 rcd_sn,
+							 TLS_OFFLOAD_CTX_DIR_TX);
 	up_read(&device_offload_lock);
+	if (err)
+		return;
 
 	clear_bit_unlock(TLS_TX_SYNC_SCHED, &tls_ctx->flags);
 }
···
 		goto free_offload_ctx;
 	}
 
+	prot->version = crypto_info->version;
+	prot->cipher_type = crypto_info->cipher_type;
 	prot->prepend_size = TLS_HEADER_SIZE + nonce_size;
 	prot->tag_size = tag_size;
 	prot->overhead_size = prot->prepend_size + prot->tag_size;

net/tls/tls_device_fallback.c

···
 
 	update_chksum(nskb, headln);
 
+	/* sock_efree means skb must gone through skb_orphan_partial() */
+	if (nskb->destructor == sock_efree)
+		return;
+
 	delta = nskb->truesize - skb->truesize;
 	if (likely(delta < 0))
 		WARN_ON_ONCE(refcount_sub_and_test(-delta, &sk->sk_wmem_alloc));