Donate to e Foundation | Murena handsets with /e/OS | Own a part of Murena! Learn more

Commit 868fdb06 authored by Eric Dumazet's avatar Eric Dumazet Committed by David S. Miller
Browse files

mlx4: remove mlx4_en_low_latency_recv()



Busy polling can now be handled in generic NAPI poll infrastructure.
This removes complexity and fast path overhead :

mlx4 used two spin_lock()/spin_unlock() pair per napi->poll() call
in mlx4_en_cq_lock_napi()/mlx4_en_cq_unlock_napi()

Tested:

Without busy polling :

lpaa23:~# echo 0 >/proc/sys/net/core/busy_read
lpaa24:~# echo 0 >/proc/sys/net/core/busy_read
lpaa23:~# ./netperf -H lpaa24 -t TCP_RR
MIGRATED TCP REQUEST/RESPONSE TEST from 0.0.0.0 (0.0.0.0) port 0 AF_INET to lpaa24.prod.google.com () port 0 AF_INET : first burst 0
Local /Remote
Socket Size   Request  Resp.   Elapsed  Trans.
Send   Recv   Size     Size    Time     Rate
bytes  Bytes  bytes    bytes   secs.    per sec

16384  87380  1        1       10.00    47330.78

With busy polling :

lpaa23:~# echo 70 >/proc/sys/net/core/busy_read
lpaa24:~# echo 70 >/proc/sys/net/core/busy_read
lpaa23:~# ./netperf -H lpaa24 -t TCP_RR
MIGRATED TCP REQUEST/RESPONSE TEST from 0.0.0.0 (0.0.0.0) port 0 AF_INET to lpaa24.prod.google.com () port 0 AF_INET : first burst 0
Local /Remote
Socket Size   Request  Resp.   Elapsed  Trans.
Send   Recv   Size     Size    Time     Rate
bytes  Bytes  bytes    bytes   secs.    per sec

16384  87380  1        1       10.00    97643.55

Signed-off-by: default avatarEric Dumazet <edumazet@google.com>
Signed-off-by: default avatarDavid S. Miller <davem@davemloft.net>
parent b59768c6
Loading
Loading
Loading
Loading
+0 −17
Original line number Diff line number Diff line
@@ -337,11 +337,7 @@ static int mlx4_en_get_sset_count(struct net_device *dev, int sset)
	case ETH_SS_STATS:
		return bitmap_iterator_count(&it) +
			(priv->tx_ring_num * 2) +
#ifdef CONFIG_NET_RX_BUSY_POLL
			(priv->rx_ring_num * 5);
#else
			(priv->rx_ring_num * 2);
#endif
	case ETH_SS_TEST:
		return MLX4_EN_NUM_SELF_TEST - !(priv->mdev->dev->caps.flags
					& MLX4_DEV_CAP_FLAG_UC_LOOPBACK) * 2;
@@ -408,11 +404,6 @@ static void mlx4_en_get_ethtool_stats(struct net_device *dev,
	for (i = 0; i < priv->rx_ring_num; i++) {
		data[index++] = priv->rx_ring[i]->packets;
		data[index++] = priv->rx_ring[i]->bytes;
#ifdef CONFIG_NET_RX_BUSY_POLL
		data[index++] = priv->rx_ring[i]->yields;
		data[index++] = priv->rx_ring[i]->misses;
		data[index++] = priv->rx_ring[i]->cleaned;
#endif
	}
	spin_unlock_bh(&priv->stats_lock);

@@ -486,14 +477,6 @@ static void mlx4_en_get_strings(struct net_device *dev,
				"rx%d_packets", i);
			sprintf(data + (index++) * ETH_GSTRING_LEN,
				"rx%d_bytes", i);
#ifdef CONFIG_NET_RX_BUSY_POLL
			sprintf(data + (index++) * ETH_GSTRING_LEN,
				"rx%d_napi_yield", i);
			sprintf(data + (index++) * ETH_GSTRING_LEN,
				"rx%d_misses", i);
			sprintf(data + (index++) * ETH_GSTRING_LEN,
				"rx%d_cleaned", i);
#endif
		}
		break;
	case ETH_SS_PRIV_FLAGS:
+0 −40
Original line number Diff line number Diff line
@@ -69,34 +69,6 @@ int mlx4_en_setup_tc(struct net_device *dev, u8 up)
	return 0;
}

#ifdef CONFIG_NET_RX_BUSY_POLL
/* must be called with local_bh_disable()d */
static int mlx4_en_low_latency_recv(struct napi_struct *napi)
{
	struct mlx4_en_cq *cq = container_of(napi, struct mlx4_en_cq, napi);
	struct net_device *dev = cq->dev;
	struct mlx4_en_priv *priv = netdev_priv(dev);
	struct mlx4_en_rx_ring *rx_ring = priv->rx_ring[cq->ring];
	int done;

	if (!priv->port_up)
		return LL_FLUSH_FAILED;

	if (!mlx4_en_cq_lock_poll(cq))
		return LL_FLUSH_BUSY;

	done = mlx4_en_process_rx_cq(dev, cq, 4);
	if (likely(done))
		rx_ring->cleaned += done;
	else
		rx_ring->misses++;

	mlx4_en_cq_unlock_poll(cq);

	return done;
}
#endif	/* CONFIG_NET_RX_BUSY_POLL */

#ifdef CONFIG_RFS_ACCEL

struct mlx4_en_filter {
@@ -1561,8 +1533,6 @@ int mlx4_en_start_port(struct net_device *dev)
	for (i = 0; i < priv->rx_ring_num; i++) {
		cq = priv->rx_cq[i];

		mlx4_en_cq_init_lock(cq);

		err = mlx4_en_init_affinity_hint(priv, i);
		if (err) {
			en_err(priv, "Failed preparing IRQ affinity hint\n");
@@ -1859,13 +1829,6 @@ void mlx4_en_stop_port(struct net_device *dev, int detach)
	for (i = 0; i < priv->rx_ring_num; i++) {
		struct mlx4_en_cq *cq = priv->rx_cq[i];

		local_bh_disable();
		while (!mlx4_en_cq_lock_napi(cq)) {
			pr_info("CQ %d locked\n", i);
			mdelay(1);
		}
		local_bh_enable();

		napi_synchronize(&cq->napi);
		mlx4_en_deactivate_rx_ring(priv, priv->rx_ring[i]);
		mlx4_en_deactivate_cq(priv, cq);
@@ -2503,9 +2466,6 @@ static const struct net_device_ops mlx4_netdev_ops = {
	.ndo_setup_tc		= mlx4_en_setup_tc,
#ifdef CONFIG_RFS_ACCEL
	.ndo_rx_flow_steer	= mlx4_en_filter_rfs,
#endif
#ifdef CONFIG_NET_RX_BUSY_POLL
	.ndo_busy_poll		= mlx4_en_low_latency_recv,
#endif
	.ndo_get_phys_port_id	= mlx4_en_get_phys_port_id,
#ifdef CONFIG_MLX4_EN_VXLAN
+2 −13
Original line number Diff line number Diff line
@@ -873,10 +873,8 @@ int mlx4_en_process_rx_cq(struct net_device *dev, struct mlx4_en_cq *cq, int bud
		 * - TCP/IP (v4)
		 * - without IP options
		 * - not an IP fragment
		 * - no LLS polling in progress
		 */
		if (!mlx4_en_cq_busy_polling(cq) &&
		    (dev->features & NETIF_F_GRO)) {
		if (dev->features & NETIF_F_GRO) {
			struct sk_buff *gro_skb = napi_get_frags(&cq->napi);
			if (!gro_skb)
				goto next;
@@ -992,11 +990,7 @@ int mlx4_en_process_rx_cq(struct net_device *dev, struct mlx4_en_cq *cq, int bud

		skb_mark_napi_id(skb, &cq->napi);

		if (!mlx4_en_cq_busy_polling(cq))
		napi_gro_receive(&cq->napi, skb);
		else
			netif_receive_skb(skb);

next:
		for (nr = 0; nr < priv->num_frags; nr++)
			mlx4_en_free_frag(priv, frags, nr);
@@ -1038,13 +1032,8 @@ int mlx4_en_poll_rx_cq(struct napi_struct *napi, int budget)
	struct mlx4_en_priv *priv = netdev_priv(dev);
	int done;

	if (!mlx4_en_cq_lock_napi(cq))
		return budget;

	done = mlx4_en_process_rx_cq(dev, cq, budget);

	mlx4_en_cq_unlock_napi(cq);

	/* If we used up all the quota - we're probably not done yet... */
	if (done == budget) {
		const struct cpumask *aff;
+0 −128
Original line number Diff line number Diff line
@@ -320,11 +320,6 @@ struct mlx4_en_rx_ring {
	void *rx_info;
	unsigned long bytes;
	unsigned long packets;
#ifdef CONFIG_NET_RX_BUSY_POLL
	unsigned long yields;
	unsigned long misses;
	unsigned long cleaned;
#endif
	unsigned long csum_ok;
	unsigned long csum_none;
	unsigned long csum_complete;
@@ -347,18 +342,6 @@ struct mlx4_en_cq {
	struct mlx4_cqe *buf;
#define MLX4_EN_OPCODE_ERROR	0x1e

#ifdef CONFIG_NET_RX_BUSY_POLL
	unsigned int state;
#define MLX4_EN_CQ_STATE_IDLE        0
#define MLX4_EN_CQ_STATE_NAPI     1    /* NAPI owns this CQ */
#define MLX4_EN_CQ_STATE_POLL     2    /* poll owns this CQ */
#define MLX4_CQ_LOCKED (MLX4_EN_CQ_STATE_NAPI | MLX4_EN_CQ_STATE_POLL)
#define MLX4_EN_CQ_STATE_NAPI_YIELD  4    /* NAPI yielded this CQ */
#define MLX4_EN_CQ_STATE_POLL_YIELD  8    /* poll yielded this CQ */
#define CQ_YIELD (MLX4_EN_CQ_STATE_NAPI_YIELD | MLX4_EN_CQ_STATE_POLL_YIELD)
#define CQ_USER_PEND (MLX4_EN_CQ_STATE_POLL | MLX4_EN_CQ_STATE_POLL_YIELD)
	spinlock_t poll_lock; /* protects from LLS/napi conflicts */
#endif  /* CONFIG_NET_RX_BUSY_POLL */
	struct irq_desc *irq_desc;
};

@@ -622,117 +605,6 @@ static inline struct mlx4_cqe *mlx4_en_get_cqe(void *buf, int idx, int cqe_sz)
	return buf + idx * cqe_sz;
}

#ifdef CONFIG_NET_RX_BUSY_POLL
static inline void mlx4_en_cq_init_lock(struct mlx4_en_cq *cq)
{
	spin_lock_init(&cq->poll_lock);
	cq->state = MLX4_EN_CQ_STATE_IDLE;
}

/* called from the device poll rutine to get ownership of a cq */
static inline bool mlx4_en_cq_lock_napi(struct mlx4_en_cq *cq)
{
	int rc = true;
	spin_lock(&cq->poll_lock);
	if (cq->state & MLX4_CQ_LOCKED) {
		WARN_ON(cq->state & MLX4_EN_CQ_STATE_NAPI);
		cq->state |= MLX4_EN_CQ_STATE_NAPI_YIELD;
		rc = false;
	} else
		/* we don't care if someone yielded */
		cq->state = MLX4_EN_CQ_STATE_NAPI;
	spin_unlock(&cq->poll_lock);
	return rc;
}

/* returns true is someone tried to get the cq while napi had it */
static inline bool mlx4_en_cq_unlock_napi(struct mlx4_en_cq *cq)
{
	int rc = false;
	spin_lock(&cq->poll_lock);
	WARN_ON(cq->state & (MLX4_EN_CQ_STATE_POLL |
			       MLX4_EN_CQ_STATE_NAPI_YIELD));

	if (cq->state & MLX4_EN_CQ_STATE_POLL_YIELD)
		rc = true;
	cq->state = MLX4_EN_CQ_STATE_IDLE;
	spin_unlock(&cq->poll_lock);
	return rc;
}

/* called from mlx4_en_low_latency_recv(), BH are disabled */
static inline bool mlx4_en_cq_lock_poll(struct mlx4_en_cq *cq)
{
	int rc = true;

	spin_lock(&cq->poll_lock);
	if ((cq->state & MLX4_CQ_LOCKED)) {
		struct net_device *dev = cq->dev;
		struct mlx4_en_priv *priv = netdev_priv(dev);
		struct mlx4_en_rx_ring *rx_ring = priv->rx_ring[cq->ring];

		cq->state |= MLX4_EN_CQ_STATE_POLL_YIELD;
		rc = false;
		rx_ring->yields++;
	} else
		/* preserve yield marks */
		cq->state |= MLX4_EN_CQ_STATE_POLL;
	spin_unlock(&cq->poll_lock);
	return rc;
}

/* returns true if someone tried to get the cq while it was locked */
static inline bool mlx4_en_cq_unlock_poll(struct mlx4_en_cq *cq)
{
	int rc = false;

	spin_lock(&cq->poll_lock);
	WARN_ON(cq->state & (MLX4_EN_CQ_STATE_NAPI));

	if (cq->state & MLX4_EN_CQ_STATE_POLL_YIELD)
		rc = true;
	cq->state = MLX4_EN_CQ_STATE_IDLE;
	spin_unlock(&cq->poll_lock);
	return rc;
}

/* true if a socket is polling, even if it did not get the lock */
static inline bool mlx4_en_cq_busy_polling(struct mlx4_en_cq *cq)
{
	WARN_ON(!(cq->state & MLX4_CQ_LOCKED));
	return cq->state & CQ_USER_PEND;
}
#else
static inline void mlx4_en_cq_init_lock(struct mlx4_en_cq *cq)
{
}

static inline bool mlx4_en_cq_lock_napi(struct mlx4_en_cq *cq)
{
	return true;
}

static inline bool mlx4_en_cq_unlock_napi(struct mlx4_en_cq *cq)
{
	return false;
}

static inline bool mlx4_en_cq_lock_poll(struct mlx4_en_cq *cq)
{
	return false;
}

static inline bool mlx4_en_cq_unlock_poll(struct mlx4_en_cq *cq)
{
	return false;
}

static inline bool mlx4_en_cq_busy_polling(struct mlx4_en_cq *cq)
{
	return false;
}
#endif /* CONFIG_NET_RX_BUSY_POLL */

#define MLX4_EN_WOL_DO_MODIFY (1ULL << 63)

void mlx4_en_update_loopback_state(struct net_device *dev,