d/freebsd-dev
1
0
Fork 0
Code Issues Pull Requests Packages Projects Releases Wiki Activity

Allow transmit packet bufring in software to be disabled.

Browse Source 
- Add new sysctl node to control the transmit packet bufring.

- Add optimised version of the transmit routine which output packets
directly to the DMA ring instead of using bufring in case the transmit
lock is congested. This can reduce the number of taskswitches which in
turn influence the overall system CPU usage, depending on the
workload.

- Add " TX" suffix to debug name for transmit mutexes to silence some
witness warnings about aquiring duplicate locks having same name.

MFC after:		1 week
Sponsored by:		Mellanox Technologies
Suggested by:		gallatin @
This commit is contained in:
Hans Petter Selasky 2017-01-20 15:45:21 +00:00
parent f2be81e92c
commit b98ba64027
Notes: svn2git 2020-12-20 02:59:44 +00:00 
svn path=/head/; revision=312536
4 changed files with 99 additions and 32 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

1
sys/dev/mlx5/mlx5_en/en.h
View File

@ -402,6 +402,7 @@ struct mlx5e_params {
m(+1, u64 tx_coalesce_usecs, "tx_coalesce_usecs", "Limit in usec for joining tx packets") \
m(+1, u64 tx_coalesce_pkts, "tx_coalesce_pkts", "Maximum number of tx packets to join") \
m(+1, u64 tx_coalesce_mode, "tx_coalesce_mode", "0: EQE mode 1: CQE mode") \
m(+1, u64 tx_bufring_disable, "tx_bufring_disable", "0: Enable bufring 1: Disable bufring") \
m(+1, u64 tx_completion_fact, "tx_completion_fact", "1..MAX: Completion event ratio") \
m(+1, u64 tx_completion_fact_max, "tx_completion_fact_max", "Maximum completion event ratio") \
m(+1, u64 hw_lro, "hw_lro", "set to enable hw_lro") \

12
sys/dev/mlx5/mlx5_en/mlx5_en_ethtool.c
View File

@ -352,6 +352,18 @@ mlx5e_ethtool_handler(SYSCTL_HANDLER_ARGS)
mlx5e_open_locked(priv->ifp);
break;
case MLX5_PARAM_OFFSET(tx_bufring_disable):
/* rangecheck input value */
priv->params_ethtool.tx_bufring_disable =
priv->params_ethtool.tx_bufring_disable ? 1 : 0;
/* reconfigure the sendqueues, if any */
if (was_opened) {
mlx5e_close_locked(priv->ifp);
mlx5e_open_locked(priv->ifp);
}
break;
case MLX5_PARAM_OFFSET(tx_completion_fact):
/* network interface must be down */
if (was_opened)

68
sys/dev/mlx5/mlx5_en/mlx5_en_main.c
View File

@ -439,7 +439,8 @@ mlx5e_update_stats_work(struct work_struct *work)
tso_packets += sq_stats->tso_packets;
tso_bytes += sq_stats->tso_bytes;
tx_queue_dropped += sq_stats->dropped;
tx_queue_dropped += sq_br->br_drops;
if (sq_br != NULL)
tx_queue_dropped += sq_br->br_drops;
tx_defragged += sq_stats->defragged;
tx_offload_none += sq_stats->csum_offload_none;
}
@ -987,34 +988,37 @@ mlx5e_create_sq(struct mlx5e_channel *c,
sq->priv = priv;
sq->tc = tc;
sq->br = buf_ring_alloc(MLX5E_SQ_TX_QUEUE_SIZE, M_MLX5EN,
M_WAITOK, &sq->lock);
if (sq->br == NULL) {
if_printf(c->ifp, "%s: Failed allocating sq drbr buffer\n",
__func__);
err = -ENOMEM;
goto err_free_sq_db;
}
/* check if we should allocate a second packet buffer */
if (priv->params_ethtool.tx_bufring_disable == 0) {
sq->br = buf_ring_alloc(MLX5E_SQ_TX_QUEUE_SIZE, M_MLX5EN,
M_WAITOK, &sq->lock);
if (sq->br == NULL) {
if_printf(c->ifp, "%s: Failed allocating sq drbr buffer\n",
__func__);
err = -ENOMEM;
goto err_free_sq_db;
}
sq->sq_tq = taskqueue_create_fast("mlx5e_que", M_WAITOK,
taskqueue_thread_enqueue, &sq->sq_tq);
if (sq->sq_tq == NULL) {
if_printf(c->ifp, "%s: Failed allocating taskqueue\n",
__func__);
err = -ENOMEM;
goto err_free_drbr;
}
sq->sq_tq = taskqueue_create_fast("mlx5e_que", M_WAITOK,
taskqueue_thread_enqueue, &sq->sq_tq);
if (sq->sq_tq == NULL) {
if_printf(c->ifp, "%s: Failed allocating taskqueue\n",
__func__);
err = -ENOMEM;
goto err_free_drbr;
}
TASK_INIT(&sq->sq_task, 0, mlx5e_tx_que, sq);
TASK_INIT(&sq->sq_task, 0, mlx5e_tx_que, sq);
#ifdef RSS
cpu_id = rss_getcpu(c->ix % rss_getnumbuckets());
CPU_SETOF(cpu_id, &cpu_mask);
taskqueue_start_threads_cpuset(&sq->sq_tq, 1, PI_NET, &cpu_mask,
"%s TX SQ%d.%d CPU%d", c->ifp->if_xname, c->ix, tc, cpu_id);
cpu_id = rss_getcpu(c->ix % rss_getnumbuckets());
CPU_SETOF(cpu_id, &cpu_mask);
taskqueue_start_threads_cpuset(&sq->sq_tq, 1, PI_NET, &cpu_mask,
"%s TX SQ%d.%d CPU%d", c->ifp->if_xname, c->ix, tc, cpu_id);
#else
taskqueue_start_threads(&sq->sq_tq, 1, PI_NET,
"%s TX SQ%d.%d", c->ifp->if_xname, c->ix, tc);
taskqueue_start_threads(&sq->sq_tq, 1, PI_NET,
"%s TX SQ%d.%d", c->ifp->if_xname, c->ix, tc);
#endif
}
snprintf(buffer, sizeof(buffer), "txstat%dtc%d", c->ix, tc);
mlx5e_create_stats(&sq->stats.ctx, SYSCTL_CHILDREN(priv->sysctl_ifnet),
buffer, mlx5e_sq_stats_desc, MLX5E_SQ_STATS_NUM,
@ -1047,9 +1051,12 @@ mlx5e_destroy_sq(struct mlx5e_sq *sq)
mlx5e_free_sq_db(sq);
mlx5_wq_destroy(&sq->wq_ctrl);
mlx5_unmap_free_uar(sq->priv->mdev, &sq->uar);
taskqueue_drain(sq->sq_tq, &sq->sq_task);
taskqueue_free(sq->sq_tq);
buf_ring_free(sq->br, M_MLX5EN);
if (sq->sq_tq != NULL) {
taskqueue_drain(sq->sq_tq, &sq->sq_task);
taskqueue_free(sq->sq_tq);
}
if (sq->br != NULL)
buf_ring_free(sq->br, M_MLX5EN);
}
int
@ -1497,9 +1504,10 @@ mlx5e_chan_mtx_init(struct mlx5e_channel *c)
for (tc = 0; tc < c->num_tc; tc++) {
struct mlx5e_sq *sq = c->sq + tc;
mtx_init(&sq->lock, "mlx5tx", MTX_NETWORK_LOCK, MTX_DEF);
mtx_init(&sq->comp_lock, "mlx5comp", MTX_NETWORK_LOCK,
MTX_DEF);
mtx_init(&sq->lock, "mlx5tx",
MTX_NETWORK_LOCK " TX", MTX_DEF);
mtx_init(&sq->comp_lock, "mlx5comp",
MTX_NETWORK_LOCK " TX", MTX_DEF);
callout_init_mtx(&sq->cev_callout, &sq->lock, 0);

50
sys/dev/mlx5/mlx5_en/mlx5_en_tx.c
View File

@ -439,7 +439,8 @@ mlx5e_poll_tx_cq(struct mlx5e_sq *sq, int budget)
sq->cc = sqcc;
if (atomic_cmpset_int(&sq->queue_state, MLX5E_SQ_FULL, MLX5E_SQ_READY))
if (sq->sq_tq != NULL &&
atomic_cmpset_int(&sq->queue_state, MLX5E_SQ_FULL, MLX5E_SQ_READY))
taskqueue_enqueue(sq->sq_tq, &sq->sq_task);
}
@ -498,6 +499,45 @@ mlx5e_xmit_locked(struct ifnet *ifp, struct mlx5e_sq *sq, struct mbuf *mb)
return (err);
}
static int
mlx5e_xmit_locked_no_br(struct ifnet *ifp, struct mlx5e_sq *sq, struct mbuf *mb)
{
int err = 0;
if (unlikely((ifp->if_drv_flags & IFF_DRV_RUNNING) == 0 ||
sq->stopped != 0)) {
m_freem(mb);
return (ENETDOWN);
}
/* Do transmit */
if (mlx5e_sq_xmit(sq, &mb) != 0) {
/* NOTE: m_freem() is NULL safe */
m_freem(mb);
err = ENOBUFS;
}
/* Check if we need to write the doorbell */
if (likely(sq->doorbell.d64 != 0)) {
mlx5e_tx_notify_hw(sq, sq->doorbell.d32, 0);
sq->doorbell.d64 = 0;
}
/*
* Check if we need to start the event timer which flushes the
* transmit ring on timeout:
*/
if (unlikely(sq->cev_next_state == MLX5E_CEV_STATE_INITIAL &&
sq->cev_factor != 1)) {
/* start the timer */
mlx5e_sq_cev_timeout(sq);
} else {
/* don't send NOPs yet */
sq->cev_next_state = MLX5E_CEV_STATE_HOLD_NOPS;
}
return (err);
}
int
mlx5e_xmit(struct ifnet *ifp, struct mbuf *mb)
{
@ -510,7 +550,13 @@ mlx5e_xmit(struct ifnet *ifp, struct mbuf *mb)
m_freem(mb);
return (ENXIO);
}
if (mtx_trylock(&sq->lock)) {
if (unlikely(sq->br == NULL)) {
/* rate limited traffic */
mtx_lock(&sq->lock);
ret = mlx5e_xmit_locked_no_br(ifp, sq, mb);
mtx_unlock(&sq->lock);
} else if (mtx_trylock(&sq->lock)) {
ret = mlx5e_xmit_locked(ifp, sq, mb);
mtx_unlock(&sq->lock);
} else {