Re: [PATCH net-next v3] net/mlx5e: link NAPI instances to queues and IRQs

[Joe Damato <jdamato@fastly.com>]

On Thu, Feb 08, 2024 at 08:43:57PM -0800, Rahul Rameshbabu wrote:
> On Thu, 08 Feb, 2024 21:13:25 +0200 Tariq Toukan <ttoukan.linux@gmail.com> wrote:
> > On 08/02/2024 5:07, Joe Damato wrote:
> >> Make mlx5 compatible with the newly added netlink queue GET APIs.
> >> Signed-off-by: Joe Damato <jdamato@fastly.com>
> >> ---
> 
> Just came back from testing this code. Let me make one cosmetic point
> here. I noticed this patch has a line that is past 90 characters. Would
> be nice to get it wrapped in the next version. We use 90 instead of 80
> characters for the line wrap in the mlx5 driver because of firmware
> command interface related code would lead to very hard to read lines if
> wrapped at 80.
> 
>   https://patchwork.kernel.org/project/netdevbpf/patch/20240208030702.27296-1-jdamato@fastly.com/

OK, I had wrapped them in the next version I was going to send to 80, but
I'll adjust that to 90.

> >> v2 -> v3:
> >>    - Fix commit message subject
> >>    - call netif_queue_set_napi in mlx5e_ptp_activate_channel and
> >>      mlx5e_ptp_deactivate_channel to enable/disable NETDEV_QUEUE_TYPE_RX for
> >>      the PTP channel.
> >>    - Modify mlx5e_activate_txqsq and mlx5e_deactivate_txqsq to set
> >>      NETDEV_QUEUE_TYPE_TX which should take care of all TX queues including
> >>      QoS/HTB and PTP.
> >>    - Rearrange mlx5e_activate_channel and mlx5e_deactivate_channel for
> >>      better ordering when setting and unsetting NETDEV_QUEUE_TYPE_RX NAPI
> >>      structs
> >> v1 -> v2:
> >>    - Move netlink NULL code to mlx5e_deactivate_channel
> >>    - Move netif_napi_set_irq to mlx5e_open_channel and avoid storing the
> >>      irq, after netif_napi_add which itself sets the IRQ to -1
> >>   drivers/net/ethernet/mellanox/mlx5/core/en/ptp.c  | 3 +++
> >>   drivers/net/ethernet/mellanox/mlx5/core/en_main.c | 7 +++++++
> >>   2 files changed, 10 insertions(+)
> >> diff --git a/drivers/net/ethernet/mellanox/mlx5/core/en/ptp.c
> >> b/drivers/net/ethernet/mellanox/mlx5/core/en/ptp.c
> >> index 078f56a3cbb2..fbbc287d924d 100644
> >> --- a/drivers/net/ethernet/mellanox/mlx5/core/en/ptp.c
> >> +++ b/drivers/net/ethernet/mellanox/mlx5/core/en/ptp.c
> >> @@ -927,6 +927,8 @@ void mlx5e_ptp_activate_channel(struct mlx5e_ptp *c)
> >>   	int tc;
> >>     	napi_enable(&c->napi);
> >> +	netif_queue_set_napi(c->netdev, c->rq.ix, NETDEV_QUEUE_TYPE_RX,
> >> +			     &c->napi);
> 
> This should only be set if MLX5E_PTP_STATE_RX is set. Otherwise, the rq
> is not initialized. The following callgraph should help illustrate this.
> 
>    mlx5e_ptp_open
>     |_ mlx5e_ptp_open_queues
>        |_ mlx5e_ptp_open_rq
>           |_ mlx5e_open_rq

I had made the change that Tariq had suggested in the previous message of
using sq->netdev and sq->cq.napi, but I can also tie that into
MLX5E_PTP_STATE_RX.

I'll add this change as well, test locally again and resend shortly.

> >>     	if (test_bit(MLX5E_PTP_STATE_TX, c->state)) {
> >>   		for (tc = 0; tc < c->num_tc; tc++)
> >> @@ -951,6 +953,7 @@ void mlx5e_ptp_deactivate_channel(struct mlx5e_ptp *c)
> >>   			mlx5e_deactivate_txqsq(&c->ptpsq[tc].txqsq);
> >>   	}
> >>   +	netif_queue_set_napi(c->netdev, c->rq.ix, NETDEV_QUEUE_TYPE_RX, NULL);
> 
> I believe it would be best to tie this to whether MLX5E_PTP_STATE_RX is
> set or not.
> 
> >>   	napi_disable(&c->napi);
> >>   }
> >>   diff --git a/drivers/net/ethernet/mellanox/mlx5/core/en_main.c
> >> b/drivers/net/ethernet/mellanox/mlx5/core/en_main.c
> >> index c8e8f512803e..2f1792854dd5 100644
> >> --- a/drivers/net/ethernet/mellanox/mlx5/core/en_main.c
> >> +++ b/drivers/net/ethernet/mellanox/mlx5/core/en_main.c
> >> @@ -1806,6 +1806,7 @@ void mlx5e_activate_txqsq(struct mlx5e_txqsq *sq)
> >>   	set_bit(MLX5E_SQ_STATE_ENABLED, &sq->state);
> >>   	netdev_tx_reset_queue(sq->txq);
> >>   	netif_tx_start_queue(sq->txq);
> >> +	netif_queue_set_napi(sq->channel->netdev, sq->txq_ix, NETDEV_QUEUE_TYPE_TX, &sq->channel->napi);
> >
> > Might be called with channel==NULL.
> > For example for PTP.
> >
> > Prefer sq->netdev and sq->cq.napi.
> >
> >>   }
> >>     void mlx5e_tx_disable_queue(struct netdev_queue *txq)
> >> @@ -1819,6 +1820,7 @@ void mlx5e_deactivate_txqsq(struct mlx5e_txqsq *sq)
> >>   {
> >>   	struct mlx5_wq_cyc *wq = &sq->wq;
> >>   +	netif_queue_set_napi(sq->channel->netdev, sq->txq_ix,
> >> NETDEV_QUEUE_TYPE_TX, NULL);
> >
> > Same here.
> >
> >>   	clear_bit(MLX5E_SQ_STATE_ENABLED, &sq->state);
> >>   	synchronize_net(); /* Sync with NAPI to prevent netif_tx_wake_queue. */
> >>   @@ -2560,6 +2562,7 @@ static int mlx5e_open_channel(struct mlx5e_priv *priv,
> >> int ix,
> >>   	c->lag_port = mlx5e_enumerate_lag_port(priv->mdev, ix);
> >>     	netif_napi_add(netdev, &c->napi, mlx5e_napi_poll);
> >> +	netif_napi_set_irq(&c->napi, irq);
> >>     	err = mlx5e_open_queues(c, params, cparam);
> >>   	if (unlikely(err))
> >> @@ -2602,12 +2605,16 @@ static void mlx5e_activate_channel(struct mlx5e_channel *c)
> >>   		mlx5e_activate_xsk(c);
> >>   	else
> >>   		mlx5e_activate_rq(&c->rq);
> >> +
> >> +	netif_queue_set_napi(c->netdev, c->ix, NETDEV_QUEUE_TYPE_RX, &c->napi);
> >>   }
> >>     static void mlx5e_deactivate_channel(struct mlx5e_channel *c)
> >>   {
> >>   	int tc;
> >>   +	netif_queue_set_napi(c->netdev, c->ix, NETDEV_QUEUE_TYPE_RX, NULL);
> >> +
> >>   	if (test_bit(MLX5E_CHANNEL_STATE_XSK, c->state))
> >>   		mlx5e_deactivate_xsk(c);
> >>   	else
> 
> --
> Thanks,
> 
> Rahul Rameshbabu