Merge branch 'nfp-tls-implement-the-stream-sync-RX-resync'

Jakub Kicinski says:

====================
nfp: tls: implement the stream sync RX resync

This small series adds support for using the device
in stream scan RX resync mode which improves the RX
resync success rate. Without stream scan it's pretty
much impossible to successfully resync a continuous
stream.
====================
Signed-off-by: David S. Miller <davem@davemloft.net>

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

@@ -26,6 +26,7 @@ enum nfp_ccm_type {
 	NFP_CCM_TYPE_CRYPTO_ADD		= 10,
 	NFP_CCM_TYPE_CRYPTO_DEL		= 11,
 	NFP_CCM_TYPE_CRYPTO_UPDATE	= 12,
+	NFP_CCM_TYPE_CRYPTO_RESYNC	= 13,
 	__NFP_CCM_TYPE_MAX,
 };
 

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

@@ -4,6 +4,10 @@
 #ifndef NFP_CRYPTO_H
 #define NFP_CRYPTO_H 1
 
+struct net_device;
+struct nfp_net;
+struct nfp_net_tls_resync_req;
+
 struct nfp_net_tls_offload_ctx {
 	__be32 fw_handle[2];
 
@@ -17,11 +21,22 @@ struct nfp_net_tls_offload_ctx {
 
 #ifdef CONFIG_TLS_DEVICE
 int nfp_net_tls_init(struct nfp_net *nn);
+int nfp_net_tls_rx_resync_req(struct net_device *netdev,
+			      struct nfp_net_tls_resync_req *req,
+			      void *pkt, unsigned int pkt_len);
 #else
 static inline int nfp_net_tls_init(struct nfp_net *nn)
 {
 	return 0;
 }
+
+static inline int
+nfp_net_tls_rx_resync_req(struct net_device *netdev,
+			  struct nfp_net_tls_resync_req *req,
+			  void *pkt, unsigned int pkt_len)
+{
+	return -EOPNOTSUPP;
+}
 #endif
 
 #endif

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

@@ -9,6 +9,14 @@
 #define NFP_NET_CRYPTO_OP_TLS_1_2_AES_GCM_128_ENC	0
 #define NFP_NET_CRYPTO_OP_TLS_1_2_AES_GCM_128_DEC	1
 
+struct nfp_net_tls_resync_req {
+	__be32 fw_handle[2];
+	__be32 tcp_seq;
+	u8 l3_offset;
+	u8 l4_offset;
+	u8 resv[2];
+};
+
 struct nfp_crypto_reply_simple {
 	struct nfp_ccm_hdr hdr;
 	__be32 error;

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

@@ -5,6 +5,7 @@
 #include <linux/ipv6.h>
 #include <linux/skbuff.h>
 #include <linux/string.h>
+#include <net/inet6_hashtables.h>
 #include <net/tls.h>
 
 #include "../ccm.h"
@@ -391,8 +392,9 @@ nfp_net_tls_add(struct net_device *netdev, struct sock *sk,
 	if (direction == TLS_OFFLOAD_CTX_DIR_TX)
 		return 0;
 
-	tls_offload_rx_resync_set_type(sk,
-				       TLS_OFFLOAD_SYNC_TYPE_CORE_NEXT_HINT);
+	if (!nn->tlv_caps.tls_resync_ss)
+		tls_offload_rx_resync_set_type(sk, TLS_OFFLOAD_SYNC_TYPE_CORE_NEXT_HINT);
+
 	return 0;
 
 err_fw_remove:
@@ -424,6 +426,7 @@ nfp_net_tls_resync(struct net_device *netdev, struct sock *sk, u32 seq,
 	struct nfp_net *nn = netdev_priv(netdev);
 	struct nfp_net_tls_offload_ctx *ntls;
 	struct nfp_crypto_req_update *req;
+	enum nfp_ccm_type type;
 	struct sk_buff *skb;
 	gfp_t flags;
 	int err;
@@ -442,15 +445,18 @@ nfp_net_tls_resync(struct net_device *netdev, struct sock *sk, u32 seq,
 	req->tcp_seq = cpu_to_be32(seq);
 	memcpy(req->rec_no, rcd_sn, sizeof(req->rec_no));
 
+	type = NFP_CCM_TYPE_CRYPTO_UPDATE;
 	if (direction == TLS_OFFLOAD_CTX_DIR_TX) {
-		err = nfp_net_tls_communicate_simple(nn, skb, "sync",
-						     NFP_CCM_TYPE_CRYPTO_UPDATE);
+		err = nfp_net_tls_communicate_simple(nn, skb, "sync", type);
 		if (err)
 			return err;
 		ntls->next_seq = seq;
 	} else {
-		nfp_ccm_mbox_post(nn, skb, NFP_CCM_TYPE_CRYPTO_UPDATE,
+		if (nn->tlv_caps.tls_resync_ss)
+			type = NFP_CCM_TYPE_CRYPTO_RESYNC;
+		nfp_ccm_mbox_post(nn, skb, type,
 				  sizeof(struct nfp_crypto_reply_simple));
+		atomic_inc(&nn->ktls_rx_resync_sent);
 	}
 
 	return 0;
@@ -462,6 +468,79 @@ static const struct tlsdev_ops nfp_net_tls_ops = {
 	.tls_dev_resync = nfp_net_tls_resync,
 };
 
+int nfp_net_tls_rx_resync_req(struct net_device *netdev,
+			      struct nfp_net_tls_resync_req *req,
+			      void *pkt, unsigned int pkt_len)
+{
+	struct nfp_net *nn = netdev_priv(netdev);
+	struct nfp_net_tls_offload_ctx *ntls;
+	struct ipv6hdr *ipv6h;
+	struct tcphdr *th;
+	struct iphdr *iph;
+	struct sock *sk;
+	__be32 tcp_seq;
+	int err;
+
+	iph = pkt + req->l3_offset;
+	ipv6h = pkt + req->l3_offset;
+	th = pkt + req->l4_offset;
+
+	if ((u8 *)&th[1] > (u8 *)pkt + pkt_len) {
+		netdev_warn_once(netdev, "invalid TLS RX resync request (l3_off: %hhu l4_off: %hhu pkt_len: %u)\n",
+				 req->l3_offset, req->l4_offset, pkt_len);
+		err = -EINVAL;
+		goto err_cnt_ign;
+	}
+
+	switch (iph->version) {
+	case 4:
+		sk = inet_lookup_established(dev_net(netdev), &tcp_hashinfo,
+					     iph->saddr, th->source, iph->daddr,
+					     th->dest, netdev->ifindex);
+		break;
+#if IS_ENABLED(CONFIG_IPV6)
+	case 6:
+		sk = __inet6_lookup_established(dev_net(netdev), &tcp_hashinfo,
+						&ipv6h->saddr, th->source,
+						&ipv6h->daddr, ntohs(th->dest),
+						netdev->ifindex, 0);
+		break;
+#endif
+	default:
+		netdev_warn_once(netdev, "invalid TLS RX resync request (l3_off: %hhu l4_off: %hhu ipver: %u)\n",
+				 req->l3_offset, req->l4_offset, iph->version);
+		err = -EINVAL;
+		goto err_cnt_ign;
+	}
+
+	err = 0;
+	if (!sk)
+		goto err_cnt_ign;
+	if (!tls_is_sk_rx_device_offloaded(sk) ||
+	    sk->sk_shutdown & RCV_SHUTDOWN)
+		goto err_put_sock;
+
+	ntls = tls_driver_ctx(sk, TLS_OFFLOAD_CTX_DIR_RX);
+	/* some FW versions can't report the handle and report 0s */
+	if (memchr_inv(&req->fw_handle, 0, sizeof(req->fw_handle)) &&
+	    memcmp(&req->fw_handle, &ntls->fw_handle, sizeof(ntls->fw_handle)))
+		goto err_put_sock;
+
+	/* copy to ensure alignment */
+	memcpy(&tcp_seq, &req->tcp_seq, sizeof(tcp_seq));
+	tls_offload_rx_resync_request(sk, tcp_seq);
+	atomic_inc(&nn->ktls_rx_resync_req);
+
+	sock_gen_put(sk);
+	return 0;
+
+err_put_sock:
+	sock_gen_put(sk);
+err_cnt_ign:
+	atomic_inc(&nn->ktls_rx_resync_ign);
+	return err;
+}
+
 static int nfp_net_tls_reset(struct nfp_net *nn)
 {
 	struct nfp_crypto_req_reset *req;

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

@@ -586,6 +586,9 @@ struct nfp_net_dp {
  * @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
+ * @ktls_rx_resync_req:	Counter of TLS RX resync requested
+ * @ktls_rx_resync_ign:	Counter of TLS RX resync requests ignored
+ * @ktls_rx_resync_sent:    Counter of TLS RX resync completed
  * @mbox_cmsg:		Common Control Message via vNIC mailbox state
  * @mbox_cmsg.queue:	CCM mbox queue of pending messages
  * @mbox_cmsg.wq:	CCM mbox wait queue of waiting processes
@@ -674,6 +677,9 @@ struct nfp_net {
 	atomic64_t ktls_conn_id_gen;
 
 	atomic_t ktls_no_space;
+	atomic_t ktls_rx_resync_req;
+	atomic_t ktls_rx_resync_ign;
+	atomic_t ktls_rx_resync_sent;
 
 	struct {
 		struct sk_buff_head queue;

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

@@ -47,6 +47,7 @@
 #include "nfp_net_sriov.h"
 #include "nfp_port.h"
 #include "crypto/crypto.h"
+#include "crypto/fw.h"
 
 /**
  * nfp_net_get_fw_version() - Read and parse the FW version
@@ -1661,9 +1662,9 @@ nfp_net_set_hash_desc(struct net_device *netdev, struct nfp_meta_parsed *meta,
 			 &rx_hash->hash);
 }
 
-static void *
+static bool
 nfp_net_parse_meta(struct net_device *netdev, struct nfp_meta_parsed *meta,
-		   void *data, int meta_len)
+		   void *data, void *pkt, unsigned int pkt_len, int meta_len)
 {
 	u32 meta_info;
 
@@ -1693,14 +1694,20 @@ nfp_net_parse_meta(struct net_device *netdev, struct nfp_meta_parsed *meta,
 				(__force __wsum)__get_unaligned_cpu32(data);
 			data += 4;
 			break;
+		case NFP_NET_META_RESYNC_INFO:
+			if (nfp_net_tls_rx_resync_req(netdev, data, pkt,
+						      pkt_len))
+				return NULL;
+			data += sizeof(struct nfp_net_tls_resync_req);
+			break;
 		default:
-			return NULL;
+			return true;
 		}
 
 		meta_info >>= NFP_NET_META_FIELD_SIZE;
 	}
 
-	return data;
+	return data != pkt;
 }
 
 static void
@@ -1885,12 +1892,10 @@ static int nfp_net_rx(struct nfp_net_rx_ring *rx_ring, int budget)
 			nfp_net_set_hash_desc(dp->netdev, &meta,
 					      rxbuf->frag + meta_off, rxd);
 		} else if (meta_len) {
-			void *end;
-
-			end = nfp_net_parse_meta(dp->netdev, &meta,
-						 rxbuf->frag + meta_off,
-						 meta_len);
-			if (unlikely(end != rxbuf->frag + pkt_off)) {
+			if (unlikely(nfp_net_parse_meta(dp->netdev, &meta,
+							rxbuf->frag + meta_off,
+							rxbuf->frag + pkt_off,
+							pkt_len, meta_len))) {
 				nn_dp_warn(dp, "invalid RX packet metadata\n");
 				nfp_net_rx_drop(dp, r_vec, rx_ring, rxbuf,
 						NULL);

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

@@ -17,6 +17,30 @@ static void nfp_net_tlv_caps_reset(struct nfp_net_tlv_caps *caps)
 	caps->mbox_len = NFP_NET_CFG_MBOX_VAL_MAX_SZ;
 }
 
+static bool
+nfp_net_tls_parse_crypto_ops(struct device *dev, struct nfp_net_tlv_caps *caps,
+			     u8 __iomem *ctrl_mem, u8 __iomem *data,
+			     unsigned int length, unsigned int offset,
+			     bool rx_stream_scan)
+{
+	/* Ignore the legacy TLV if new one was already parsed */
+	if (caps->tls_resync_ss && !rx_stream_scan)
+		return true;
+
+	if (length < 32) {
+		dev_err(dev,
+			"CRYPTO OPS TLV should be at least 32B, is %dB offset:%u\n",
+			length, offset);
+		return false;
+	}
+
+	caps->crypto_ops = readl(data);
+	caps->crypto_enable_off = data - ctrl_mem + 16;
+	caps->tls_resync_ss = rx_stream_scan;
+
+	return true;
+}
+
 int nfp_net_tlv_caps_parse(struct device *dev, u8 __iomem *ctrl_mem,
 			   struct nfp_net_tlv_caps *caps)
 {
@@ -104,15 +128,10 @@ int nfp_net_tlv_caps_parse(struct device *dev, u8 __iomem *ctrl_mem,
 				caps->mbox_cmsg_types = readl(data);
 			break;
 		case NFP_NET_CFG_TLV_TYPE_CRYPTO_OPS:
-			if (length < 32) {
-				dev_err(dev,
-					"CRYPTO OPS TLV should be at least 32B, is %dB offset:%u\n",
-					length, offset);
+			if (!nfp_net_tls_parse_crypto_ops(dev, caps, ctrl_mem,
+							  data, length, offset,
+							  false))
 				return -EINVAL;
-			}
-
-			caps->crypto_ops = readl(data);
-			caps->crypto_enable_off = data - ctrl_mem + 16;
 			break;
 		case NFP_NET_CFG_TLV_TYPE_VNIC_STATS:
 			if ((data - ctrl_mem) % 8) {
@@ -123,6 +142,12 @@ int nfp_net_tlv_caps_parse(struct device *dev, u8 __iomem *ctrl_mem,
 			caps->vnic_stats_off = data - ctrl_mem;
 			caps->vnic_stats_cnt = length / 10;
 			break;
+		case NFP_NET_CFG_TLV_TYPE_CRYPTO_OPS_RX_SCAN:
+			if (!nfp_net_tls_parse_crypto_ops(dev, caps, ctrl_mem,
+							  data, length, offset,
+							  true))
+				return -EINVAL;
+			break;
 		default:
 			if (!FIELD_GET(NFP_NET_CFG_TLV_HEADER_REQUIRED, hdr))
 				break;

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

@@ -45,6 +45,7 @@
 #define NFP_NET_META_PORTID		5
 #define NFP_NET_META_CSUM		6 /* checksum complete type */
 #define NFP_NET_META_CONN_HANDLE	7
+#define NFP_NET_META_RESYNC_INFO	8 /* RX resync info request */
 
 #define NFP_META_PORT_ID_CTRL		~0U
 
@@ -479,6 +480,7 @@
  * 8 words, bitmaps of supported and enabled crypto operations.
  * First 16B (4 words) contains a bitmap of supported crypto operations,
  * and next 16B contain the enabled operations.
+ * This capability is made obsolete by ones with better sync methods.
  *
  * %NFP_NET_CFG_TLV_TYPE_VNIC_STATS:
  * Variable, per-vNIC statistics, data should be 8B aligned (FW should insert
@@ -490,6 +492,10 @@
  * This TLV overwrites %NFP_NET_CFG_STATS_* values (statistics in this TLV
  * duplicate the old ones, so driver should be careful not to unnecessarily
  * render both).
+ *
+ * %NFP_NET_CFG_TLV_TYPE_CRYPTO_OPS_RX_SCAN:
+ * Same as %NFP_NET_CFG_TLV_TYPE_CRYPTO_OPS, but crypto TLS does stream scan
+ * RX sync, rather than kernel-assisted sync.
  */
 #define NFP_NET_CFG_TLV_TYPE_UNKNOWN		0
 #define NFP_NET_CFG_TLV_TYPE_RESERVED		1
@@ -502,6 +508,7 @@
 #define NFP_NET_CFG_TLV_TYPE_MBOX_CMSG_TYPES	10
 #define NFP_NET_CFG_TLV_TYPE_CRYPTO_OPS		11 /* see crypto/fw.h */
 #define NFP_NET_CFG_TLV_TYPE_VNIC_STATS		12
+#define NFP_NET_CFG_TLV_TYPE_CRYPTO_OPS_RX_SCAN	13
 
 struct device;
 
@@ -516,6 +523,7 @@ struct device;
  * @crypto_enable_off:	offset of crypto ops enable region
  * @vnic_stats_off:	offset of vNIC stats area
  * @vnic_stats_cnt:	number of vNIC stats
+ * @tls_resync_ss:	TLS resync will be performed via stream scan
  */
 struct nfp_net_tlv_caps {
 	u32 me_freq_mhz;
@@ -527,6 +535,7 @@ struct nfp_net_tlv_caps {
 	unsigned int crypto_enable_off;
 	unsigned int vnic_stats_off;
 	unsigned int vnic_stats_cnt;
+	unsigned int tls_resync_ss:1;
 };
 
 int nfp_net_tlv_caps_parse(struct device *dev, u8 __iomem *ctrl_mem,

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

@@ -174,7 +174,7 @@ static const char nfp_tlv_stat_names[][ETH_GSTRING_LEN] = {
 #define NN_ET_SWITCH_STATS_LEN 9
 #define NN_RVEC_GATHER_STATS	13
 #define NN_RVEC_PER_Q_STATS	3
-#define NN_CTRL_PATH_STATS	1
+#define NN_CTRL_PATH_STATS	4
 
 #define SFP_SFF_REV_COMPLIANCE	1
 
@@ -476,6 +476,9 @@ static u8 *nfp_vnic_get_sw_stats_strings(struct net_device *netdev, u8 *data)
 	data = nfp_pr_et(data, "tx_tls_drop_no_sync_data");
 
 	data = nfp_pr_et(data, "hw_tls_no_space");
+	data = nfp_pr_et(data, "rx_tls_resync_req_ok");
+	data = nfp_pr_et(data, "rx_tls_resync_req_ign");
+	data = nfp_pr_et(data, "rx_tls_resync_sent");
 
 	return data;
 }
@@ -524,6 +527,9 @@ static u64 *nfp_vnic_get_sw_stats(struct net_device *netdev, u64 *data)
 		*data++ = gathered_stats[j];
 
 	*data++ = atomic_read(&nn->ktls_no_space);
+	*data++ = atomic_read(&nn->ktls_rx_resync_req);
+	*data++ = atomic_read(&nn->ktls_rx_resync_ign);
+	*data++ = atomic_read(&nn->ktls_rx_resync_sent);
 
 	return data;
 }

include/net/tls.h

@@ -641,6 +641,7 @@ int tls_sw_fallback_init(struct sock *sk,
 #ifdef CONFIG_TLS_DEVICE
 void tls_device_init(void);
 void tls_device_cleanup(void);
+void tls_device_sk_destruct(struct sock *sk);
 int tls_set_device_offload(struct sock *sk, struct tls_context *ctx);
 void tls_device_free_resources_tx(struct sock *sk);
 int tls_set_device_offload_rx(struct sock *sk, struct tls_context *ctx);
@@ -649,6 +650,14 @@ void tls_device_rx_resync_new_rec(struct sock *sk, u32 rcd_len, u32 seq);
 void tls_offload_tx_resync_request(struct sock *sk, u32 got_seq, u32 exp_seq);
 int tls_device_decrypted(struct sock *sk, struct tls_context *tls_ctx,
 			 struct sk_buff *skb, struct strp_msg *rxm);
+
+static inline bool tls_is_sk_rx_device_offloaded(struct sock *sk)
+{
+	if (!sk_fullsock(sk) ||
+	    smp_load_acquire(&sk->sk_destruct) != tls_device_sk_destruct)
+		return false;
+	return tls_get_ctx(sk)->rx_conf == TLS_HW;
+}
 #else
 static inline void tls_device_init(void) {}
 static inline void tls_device_cleanup(void) {}

net/tls/tls_device.c

@@ -178,7 +178,7 @@ static void tls_icsk_clean_acked(struct sock *sk, u32 acked_seq)
  * socket and no in-flight SKBs associated with this
  * socket, so it is safe to free all the resources.
  */
-static void tls_device_sk_destruct(struct sock *sk)
+void tls_device_sk_destruct(struct sock *sk)
 {
 	struct tls_context *tls_ctx = tls_get_ctx(sk);
 	struct tls_offload_context_tx *ctx = tls_offload_ctx_tx(tls_ctx);
@@ -196,6 +196,7 @@ static void tls_device_sk_destruct(struct sock *sk)
 	if (refcount_dec_and_test(&tls_ctx->refcount))
 		tls_device_queue_ctx_destruction(tls_ctx);
 }
+EXPORT_SYMBOL_GPL(tls_device_sk_destruct);
 
 void tls_device_free_resources_tx(struct sock *sk)
 {
@@ -903,7 +904,7 @@ static void tls_device_attach(struct tls_context *ctx, struct sock *sk,
 		spin_unlock_irq(&tls_device_lock);
 
 		ctx->sk_destruct = sk->sk_destruct;
-		sk->sk_destruct = tls_device_sk_destruct;
+		smp_store_release(&sk->sk_destruct, tls_device_sk_destruct);
 	}
 }