sfc: handle TX timestamps in the normal data path

Before this work, TX timestamping is done by sending each SKB to the MC.
On the 8000 series (Medford1) we have high speed timestamping via the
 MAC, which means we can use normal TX queues for this without a
 significant drop in bandwidth.  On the X2000 series (Medford2) support
 for transmitting via the MC is removed, so the new way must be used.

This patch enables timestamping on a TX queue, if requested.
It also enhances TX event handling to process the extra completion events,
 and puts the time in the SKB.
Signed-off-by: Martin Habets <mhabets@solarflare.com>
Signed-off-by: Edward Cree <ecree@solarflare.com>
Signed-off-by: David S. Miller <davem@davemloft.net>

drivers/net/ethernet/sfc/ef10.c

@@ -2411,9 +2411,11 @@ static void efx_ef10_tx_init(struct efx_tx_queue *tx_queue)
 	/* TSOv2 is a limited resource that can only be configured on a limited
 	 * number of queues. TSO without checksum offload is not really a thing,
 	 * so we only enable it for those queues.
+	 * TSOv2 cannot be used with Hardware timestamping.
 	 */
 	if (csum_offload && (nic_data->datapath_caps2 &
-			(1 << MC_CMD_GET_CAPABILITIES_V2_OUT_TX_TSO_V2_LBN))) {
+			(1 << MC_CMD_GET_CAPABILITIES_V2_OUT_TX_TSO_V2_LBN)) &&
+	    !tx_queue->timestamping) {
 		tso_v2 = true;
 		netif_dbg(efx, hw, efx->net_dev, "Using TSOv2 for channel %u\n",
 				channel->channel);
@@ -2439,14 +2441,16 @@ static void efx_ef10_tx_init(struct efx_tx_queue *tx_queue)
 	inlen = MC_CMD_INIT_TXQ_IN_LEN(entries);
 
 	do {
-		MCDI_POPULATE_DWORD_3(inbuf, INIT_TXQ_IN_FLAGS,
+		MCDI_POPULATE_DWORD_4(inbuf, INIT_TXQ_IN_FLAGS,
 				/* This flag was removed from mcdi_pcol.h for
 				 * the non-_EXT version of INIT_TXQ.  However,
 				 * firmware still honours it.
 				 */
 				INIT_TXQ_EXT_IN_FLAG_TSOV2_EN, tso_v2,
 				INIT_TXQ_IN_FLAG_IP_CSUM_DIS, !csum_offload,
-				INIT_TXQ_IN_FLAG_TCP_CSUM_DIS, !csum_offload);
+				INIT_TXQ_IN_FLAG_TCP_CSUM_DIS, !csum_offload,
+				INIT_TXQ_EXT_IN_FLAG_TIMESTAMP,
+						tx_queue->timestamping);
 
 		rc = efx_mcdi_rpc_quiet(efx, MC_CMD_INIT_TXQ, inbuf, inlen,
 					NULL, 0, NULL);
@@ -2472,12 +2476,13 @@ static void efx_ef10_tx_init(struct efx_tx_queue *tx_queue)
 	tx_queue->buffer[0].flags = EFX_TX_BUF_OPTION;
 	tx_queue->insert_count = 1;
 	txd = efx_tx_desc(tx_queue, 0);
-	EFX_POPULATE_QWORD_4(*txd,
+	EFX_POPULATE_QWORD_5(*txd,
 			     ESF_DZ_TX_DESC_IS_OPT, true,
 			     ESF_DZ_TX_OPTION_TYPE,
 			     ESE_DZ_TX_OPTION_DESC_CRC_CSUM,
 			     ESF_DZ_TX_OPTION_UDP_TCP_CSUM, csum_offload,
-			     ESF_DZ_TX_OPTION_IP_CSUM, csum_offload);
+			     ESF_DZ_TX_OPTION_IP_CSUM, csum_offload,
+			     ESF_DZ_TX_TIMESTAMP, tx_queue->timestamping);
 	tx_queue->write_count = 1;
 
 	if (tso_v2) {
@@ -3572,6 +3577,17 @@ static int efx_ef10_handle_rx_event(struct efx_channel *channel,
 	return n_packets;
 }
 
+static u32 efx_ef10_extract_event_ts(efx_qword_t *event)
+{
+	u32 tstamp;
+
+	tstamp = EFX_QWORD_FIELD(*event, TX_TIMESTAMP_EVENT_TSTAMP_DATA_HI);
+	tstamp <<= 16;
+	tstamp |= EFX_QWORD_FIELD(*event, TX_TIMESTAMP_EVENT_TSTAMP_DATA_LO);
+
+	return tstamp;
+}
+
 static void
 efx_ef10_handle_tx_event(struct efx_channel *channel, efx_qword_t *event)
 {
@@ -3579,6 +3595,8 @@ efx_ef10_handle_tx_event(struct efx_channel *channel, efx_qword_t *event)
 	struct efx_tx_queue *tx_queue;
 	unsigned int tx_ev_desc_ptr;
 	unsigned int tx_ev_q_label;
+	unsigned int tx_ev_type;
+	u64 ts_part;
 
 	if (unlikely(READ_ONCE(efx->reset_pending)))
 		return;
@@ -3586,12 +3604,65 @@ efx_ef10_handle_tx_event(struct efx_channel *channel, efx_qword_t *event)
 	if (unlikely(EFX_QWORD_FIELD(*event, ESF_DZ_TX_DROP_EVENT)))
 		return;
 
-	/* Transmit completion */
-	tx_ev_desc_ptr = EFX_QWORD_FIELD(*event, ESF_DZ_TX_DESCR_INDX);
+	/* Get the transmit queue */
 	tx_ev_q_label = EFX_QWORD_FIELD(*event, ESF_DZ_TX_QLABEL);
 	tx_queue = efx_channel_get_tx_queue(channel,
 					    tx_ev_q_label % EFX_TXQ_TYPES);
+
+	if (!tx_queue->timestamping) {
+		/* Transmit completion */
+		tx_ev_desc_ptr = EFX_QWORD_FIELD(*event, ESF_DZ_TX_DESCR_INDX);
 		efx_xmit_done(tx_queue, tx_ev_desc_ptr & tx_queue->ptr_mask);
+		return;
+	}
+
+	/* Transmit timestamps are only available for 8XXX series. They result
+	 * in three events per packet. These occur in order, and are:
+	 *  - the normal completion event
+	 *  - the low part of the timestamp
+	 *  - the high part of the timestamp
+	 *
+	 * Each part of the timestamp is itself split across two 16 bit
+	 * fields in the event.
+	 */
+	tx_ev_type = EFX_QWORD_FIELD(*event, ESF_EZ_TX_SOFT1);
+
+	switch (tx_ev_type) {
+	case TX_TIMESTAMP_EVENT_TX_EV_COMPLETION:
+		/* In case of Queue flush or FLR, we might have received
+		 * the previous TX completion event but not the Timestamp
+		 * events.
+		 */
+		if (tx_queue->completed_desc_ptr != tx_queue->ptr_mask)
+			efx_xmit_done(tx_queue, tx_queue->completed_desc_ptr);
+
+		tx_ev_desc_ptr = EFX_QWORD_FIELD(*event,
+						 ESF_DZ_TX_DESCR_INDX);
+		tx_queue->completed_desc_ptr =
+					tx_ev_desc_ptr & tx_queue->ptr_mask;
+		break;
+
+	case TX_TIMESTAMP_EVENT_TX_EV_TSTAMP_LO:
+		ts_part = efx_ef10_extract_event_ts(event);
+		tx_queue->completed_timestamp_minor = ts_part;
+		break;
+
+	case TX_TIMESTAMP_EVENT_TX_EV_TSTAMP_HI:
+		ts_part = efx_ef10_extract_event_ts(event);
+		tx_queue->completed_timestamp_major = ts_part;
+
+		efx_xmit_done(tx_queue, tx_queue->completed_desc_ptr);
+		tx_queue->completed_desc_ptr = tx_queue->ptr_mask;
+		break;
+
+	default:
+		netif_err(efx, hw, efx->net_dev,
+			  "channel %d unknown tx event type %d (data "
+			  EFX_QWORD_FMT ")\n",
+			  channel->channel, tx_ev_type,
+			  EFX_QWORD_VAL(*event));
+		break;
+	}
 }
 
 static void

drivers/net/ethernet/sfc/net_driver.h

@@ -191,6 +191,7 @@ struct efx_tx_buffer {
  *	Size of the region is efx_piobuf_size.
  * @piobuf_offset: Buffer offset to be specified in PIO descriptors
  * @initialised: Has hardware queue been initialised?
+ * @timestamping: Is timestamping enabled for this channel?
  * @handle_tso: TSO xmit preparation handler.  Sets up the TSO metadata and
  *	may also map tx data, depending on the nature of the TSO implementation.
  * @read_count: Current read pointer.
@@ -202,6 +203,10 @@ struct efx_tx_buffer {
  *	avoid cache-line ping-pong between the xmit path and the
  *	completion path.
  * @merge_events: Number of TX merged completion events
+ * @completed_desc_ptr: Most recent completed pointer - only used with
+ *      timestamping.
+ * @completed_timestamp_major: Top part of the most recent tx timestamp.
+ * @completed_timestamp_minor: Low part of the most recent tx timestamp.
  * @insert_count: Current insert pointer
  *	This is the number of buffers that have been added to the
  *	software ring.
@@ -247,6 +252,7 @@ struct efx_tx_queue {
 	void __iomem *piobuf;
 	unsigned int piobuf_offset;
 	bool initialised;
+	bool timestamping;
 
 	/* Function pointers used in the fast path. */
 	int (*handle_tso)(struct efx_tx_queue*, struct sk_buff*, bool *);
@@ -257,6 +263,9 @@ struct efx_tx_queue {
 	unsigned int merge_events;
 	unsigned int bytes_compl;
 	unsigned int pkts_compl;
+	unsigned int completed_desc_ptr;
+	u32 completed_timestamp_major;
+	u32 completed_timestamp_minor;
 
 	/* Members used only on the xmit path */
 	unsigned int insert_count ____cacheline_aligned_in_smp;

drivers/net/ethernet/sfc/nic.h

@@ -471,6 +471,7 @@ static inline void efx_rx_skb_attach_timestamp(struct efx_channel *channel,
 }
 void efx_ptp_start_datapath(struct efx_nic *efx);
 void efx_ptp_stop_datapath(struct efx_nic *efx);
+ktime_t efx_ptp_nic_to_kernel_time(struct efx_tx_queue *tx_queue);
 
 extern const struct efx_nic_type falcon_a1_nic_type;
 extern const struct efx_nic_type falcon_b0_nic_type;

drivers/net/ethernet/sfc/ptp.c

@@ -471,6 +471,17 @@ static ktime_t efx_ptp_s27_to_ktime_correction(u32 nic_major, u32 nic_minor,
 	return efx_ptp_s27_to_ktime(nic_major, nic_minor);
 }
 
+ktime_t efx_ptp_nic_to_kernel_time(struct efx_tx_queue *tx_queue)
+{
+	struct efx_nic *efx = tx_queue->efx;
+	struct efx_ptp_data *ptp = efx->ptp_data;
+	ktime_t kt;
+
+	kt = ptp->nic_to_kernel_time(tx_queue->completed_timestamp_major,
+				     tx_queue->completed_timestamp_minor, 0);
+	return kt;
+}
+
 /* Get PTP attributes and set up time conversions */
 static int efx_ptp_get_attributes(struct efx_nic *efx)
 {

drivers/net/ethernet/sfc/tx.c

@@ -77,9 +77,23 @@ static void efx_dequeue_buffer(struct efx_tx_queue *tx_queue,
 	}
 
 	if (buffer->flags & EFX_TX_BUF_SKB) {
+		struct sk_buff *skb = (struct sk_buff *)buffer->skb;
+
 		EFX_WARN_ON_PARANOID(!pkts_compl || !bytes_compl);
 		(*pkts_compl)++;
-		(*bytes_compl) += buffer->skb->len;
+		(*bytes_compl) += skb->len;
+		if (tx_queue->timestamping &&
+		    (tx_queue->completed_timestamp_major ||
+		     tx_queue->completed_timestamp_minor)) {
+			struct skb_shared_hwtstamps hwtstamp;
+
+			hwtstamp.hwtstamp =
+				efx_ptp_nic_to_kernel_time(tx_queue);
+			skb_tstamp_tx(skb, &hwtstamp);
+
+			tx_queue->completed_timestamp_major = 0;
+			tx_queue->completed_timestamp_minor = 0;
+		}
 		dev_consume_skb_any((struct sk_buff *)buffer->skb);
 		netif_vdbg(tx_queue->efx, tx_done, tx_queue->efx->net_dev,
 			   "TX queue %d transmission id %x complete\n",
@@ -828,6 +842,10 @@ void efx_init_tx_queue(struct efx_tx_queue *tx_queue)
 	tx_queue->old_read_count = 0;
 	tx_queue->empty_read_count = 0 | EFX_EMPTY_COUNT_VALID;
 	tx_queue->xmit_more_available = false;
+	tx_queue->timestamping = false;
+	tx_queue->completed_desc_ptr = tx_queue->ptr_mask;
+	tx_queue->completed_timestamp_major = 0;
+	tx_queue->completed_timestamp_minor = 0;
 
 	/* Set up default function pointers. These may get replaced by
 	 * efx_nic_init_tx() based off NIC/queue capabilities.