sfc: use budget for TX completions

When running workloads heavy unbalanced towards TX (high TX, low RX
traffic), sfc driver can retain the CPU during too long times. Although
in many cases this is not enough to be visible, it can affect
performance and system responsiveness.

A way to reproduce it is to use a debug kernel and run some parallel
netperf TX tests. In some systems, this will lead to this message being
logged:
  kernel:watchdog: BUG: soft lockup - CPU#12 stuck for 22s!

The reason is that sfc driver doesn't account any NAPI budget for the TX
completion events work. With high-TX/low-RX traffic, this makes that the
CPU is held for long time for NAPI poll.

Documentations says "drivers can process completions for any number of Tx
packets but should only process up to budget number of Rx packets".
However, many drivers do limit the amount of TX completions that they
process in a single NAPI poll.

In the same way, this patch adds a limit for the TX work in sfc. With
the patch applied, the watchdog warning never appears.

Tested with netperf in different combinations: single process / parallel
processes, TCP / UDP and different sizes of UDP messages. Repeated the
tests before and after the patch, without any noticeable difference in
network or CPU performance.

Test hardware:
Intel(R) Xeon(R) CPU E5-1620 v4 @ 3.50GHz (4 cores, 2 threads/core)
Solarflare Communications XtremeScale X2522-25G Network Adapter

Fixes: 5227ecc ("sfc: remove tx and MCDI handling from NAPI budget consideration")
Fixes: d19a537 ("sfc_ef100: TX path for EF100 NICs")
Reported-by: Fei Liu <[email protected]>
Signed-off-by: Íñigo Huguet <[email protected]>
Acked-by: Martin Habets <[email protected]>
Link: https://lore.kernel.org/r/[email protected]
Signed-off-by: Jakub Kicinski <[email protected]>

Author: ihuguet

drivers/net/ethernet/sfc/ef10.c

@@ -2950,21 +2950,22 @@ static u32 efx_ef10_extract_event_ts(efx_qword_t *event)
 	return tstamp;
 }
 
-static void
+static int
 efx_ef10_handle_tx_event(struct efx_channel *channel, efx_qword_t *event)
 {
 	struct efx_nic *efx = channel->efx;
 	struct efx_tx_queue *tx_queue;
 	unsigned int tx_ev_desc_ptr;
 	unsigned int tx_ev_q_label;
 	unsigned int tx_ev_type;
+	int work_done;
 	u64 ts_part;
 
 	if (unlikely(READ_ONCE(efx->reset_pending)))
-		return;
+		return 0;
 
 	if (unlikely(EFX_QWORD_FIELD(*event, ESF_DZ_TX_DROP_EVENT)))
-		return;
+		return 0;
 
 	/* Get the transmit queue */
 	tx_ev_q_label = EFX_QWORD_FIELD(*event, ESF_DZ_TX_QLABEL);
@@ -2973,8 +2974,7 @@ efx_ef10_handle_tx_event(struct efx_channel *channel, efx_qword_t *event)
 	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;
+		return efx_xmit_done(tx_queue, tx_ev_desc_ptr & tx_queue->ptr_mask);
 	}
 
 	/* Transmit timestamps are only available for 8XXX series. They result
@@ -3000,6 +3000,7 @@ efx_ef10_handle_tx_event(struct efx_channel *channel, efx_qword_t *event)
 	 * fields in the event.
 	 */
 	tx_ev_type = EFX_QWORD_FIELD(*event, ESF_EZ_TX_SOFT1);
+	work_done = 0;
 
 	switch (tx_ev_type) {
 	case TX_TIMESTAMP_EVENT_TX_EV_COMPLETION:
@@ -3016,6 +3017,7 @@ efx_ef10_handle_tx_event(struct efx_channel *channel, efx_qword_t *event)
 		tx_queue->completed_timestamp_major = ts_part;
 
 		efx_xmit_done_single(tx_queue);
+		work_done = 1;
 		break;
 
 	default:
@@ -3026,6 +3028,8 @@ efx_ef10_handle_tx_event(struct efx_channel *channel, efx_qword_t *event)
 			  EFX_QWORD_VAL(*event));
 		break;
 	}
+
+	return work_done;
 }
 
 static void
@@ -3081,13 +3085,16 @@ static void efx_ef10_handle_driver_generated_event(struct efx_channel *channel,
 	}
 }
 
+#define EFX_NAPI_MAX_TX 512
+
 static int efx_ef10_ev_process(struct efx_channel *channel, int quota)
 {
 	struct efx_nic *efx = channel->efx;
 	efx_qword_t event, *p_event;
 	unsigned int read_ptr;
-	int ev_code;
+	int spent_tx = 0;
 	int spent = 0;
+	int ev_code;
 
 	if (quota <= 0)
 		return spent;
@@ -3126,7 +3133,11 @@ static int efx_ef10_ev_process(struct efx_channel *channel, int quota)
 			}
 			break;
 		case ESE_DZ_EV_CODE_TX_EV:
-			efx_ef10_handle_tx_event(channel, &event);
+			spent_tx += efx_ef10_handle_tx_event(channel, &event);
+			if (spent_tx >= EFX_NAPI_MAX_TX) {
+				spent = quota;
+				goto out;
+			}
 			break;
 		case ESE_DZ_EV_CODE_DRIVER_EV:
 			efx_ef10_handle_driver_event(channel, &event);

drivers/net/ethernet/sfc/ef100_nic.c

@@ -253,13 +253,16 @@ static void ef100_ev_read_ack(struct efx_channel *channel)
 		   efx_reg(channel->efx, ER_GZ_EVQ_INT_PRIME));
 }
 
+#define EFX_NAPI_MAX_TX 512
+
 static int ef100_ev_process(struct efx_channel *channel, int quota)
 {
 	struct efx_nic *efx = channel->efx;
 	struct ef100_nic_data *nic_data;
 	bool evq_phase, old_evq_phase;
 	unsigned int read_ptr;
 	efx_qword_t *p_event;
+	int spent_tx = 0;
 	int spent = 0;
 	bool ev_phase;
 	int ev_type;
@@ -295,7 +298,9 @@ static int ef100_ev_process(struct efx_channel *channel, int quota)
 			efx_mcdi_process_event(channel, p_event);
 			break;
 		case ESE_GZ_EF100_EV_TX_COMPLETION:
-			ef100_ev_tx(channel, p_event);
+			spent_tx += ef100_ev_tx(channel, p_event);
+			if (spent_tx >= EFX_NAPI_MAX_TX)
+				spent = quota;
 			break;
 		case ESE_GZ_EF100_EV_DRIVER:
 			netif_info(efx, drv, efx->net_dev,

drivers/net/ethernet/sfc/ef100_tx.c

@@ -346,7 +346,7 @@ void ef100_tx_write(struct efx_tx_queue *tx_queue)
 	ef100_tx_push_buffers(tx_queue);
 }
 
-void ef100_ev_tx(struct efx_channel *channel, const efx_qword_t *p_event)
+int ef100_ev_tx(struct efx_channel *channel, const efx_qword_t *p_event)
 {
 	unsigned int tx_done =
 		EFX_QWORD_FIELD(*p_event, ESF_GZ_EV_TXCMPL_NUM_DESC);
@@ -357,7 +357,7 @@ void ef100_ev_tx(struct efx_channel *channel, const efx_qword_t *p_event)
 	unsigned int tx_index = (tx_queue->read_count + tx_done - 1) &
 				tx_queue->ptr_mask;
 
-	efx_xmit_done(tx_queue, tx_index);
+	return efx_xmit_done(tx_queue, tx_index);
 }
 
 /* Add a socket buffer to a TX queue

drivers/net/ethernet/sfc/ef100_tx.h

@@ -20,7 +20,7 @@ void ef100_tx_init(struct efx_tx_queue *tx_queue);
 void ef100_tx_write(struct efx_tx_queue *tx_queue);
 unsigned int ef100_tx_max_skb_descs(struct efx_nic *efx);
 
-void ef100_ev_tx(struct efx_channel *channel, const efx_qword_t *p_event);
+int ef100_ev_tx(struct efx_channel *channel, const efx_qword_t *p_event);
 
 netdev_tx_t ef100_enqueue_skb(struct efx_tx_queue *tx_queue, struct sk_buff *skb);
 int __ef100_enqueue_skb(struct efx_tx_queue *tx_queue, struct sk_buff *skb,

drivers/net/ethernet/sfc/tx_common.c

@@ -249,7 +249,7 @@ void efx_xmit_done_check_empty(struct efx_tx_queue *tx_queue)
 	}
 }
 
-void efx_xmit_done(struct efx_tx_queue *tx_queue, unsigned int index)
+int efx_xmit_done(struct efx_tx_queue *tx_queue, unsigned int index)
 {
 	unsigned int fill_level, pkts_compl = 0, bytes_compl = 0;
 	unsigned int efv_pkts_compl = 0;
@@ -279,6 +279,8 @@ void efx_xmit_done(struct efx_tx_queue *tx_queue, unsigned int index)
 	}
 
 	efx_xmit_done_check_empty(tx_queue);
+
+	return pkts_compl + efv_pkts_compl;
 }
 
 /* Remove buffers put into a tx_queue for the current packet.

drivers/net/ethernet/sfc/tx_common.h

@@ -28,7 +28,7 @@ static inline bool efx_tx_buffer_in_use(struct efx_tx_buffer *buffer)
 }
 
 void efx_xmit_done_check_empty(struct efx_tx_queue *tx_queue);
-void efx_xmit_done(struct efx_tx_queue *tx_queue, unsigned int index);
+int efx_xmit_done(struct efx_tx_queue *tx_queue, unsigned int index);
 
 void efx_enqueue_unwind(struct efx_tx_queue *tx_queue,
 			unsigned int insert_count);