Message-ID: <87zfwel5qi.fsf@nvidia.com>
Date: Mon, 05 Feb 2024 17:44:27 -0800
From: Rahul Rameshbabu <rrameshbabu@...dia.com>
To: Joe Damato <jdamato@...tly.com>
Cc: linux-kernel@...r.kernel.org, netdev@...r.kernel.org, tariqt@...dia.com,
 Saeed Mahameed <saeedm@...dia.com>, Leon Romanovsky <leon@...nel.org>,
 "David S. Miller" <davem@...emloft.net>, Eric Dumazet
 <edumazet@...gle.com>, Jakub Kicinski <kuba@...nel.org>, Paolo Abeni
 <pabeni@...hat.com>, "open list:MELLANOX MLX5 core VPI driver"
 <linux-rdma@...r.kernel.org>
Subject: Re: [PATCH net-next] eth: mlx5: link NAPI instances to queues and IRQs


On Mon, 05 Feb, 2024 17:41:52 -0800 Joe Damato <jdamato@...tly.com> wrote:
> On Mon, Feb 05, 2024 at 05:33:39PM -0800, Rahul Rameshbabu wrote:
>> 
>> On Mon, 05 Feb, 2024 17:32:47 -0800 Joe Damato <jdamato@...tly.com> wrote:
>> > On Mon, Feb 05, 2024 at 05:09:09PM -0800, Rahul Rameshbabu wrote:
>> >> On Tue, 06 Feb, 2024 01:03:11 +0000 Joe Damato <jdamato@...tly.com> wrote:
>> >> > Make mlx5 compatible with the newly added netlink queue GET APIs.
>> >> >
>> >> > Signed-off-by: Joe Damato <jdamato@...tly.com>
>> >> > ---
>> >> >  drivers/net/ethernet/mellanox/mlx5/core/en.h      | 1 +
>> >> >  drivers/net/ethernet/mellanox/mlx5/core/en_main.c | 8 ++++++++
>> >> >  2 files changed, 9 insertions(+)
>> >> >
>> >> > diff --git a/drivers/net/ethernet/mellanox/mlx5/core/en.h b/drivers/net/ethernet/mellanox/mlx5/core/en.h
>> >> > index 55c6ace0acd5..3f86ee1831a8 100644
>> >> > --- a/drivers/net/ethernet/mellanox/mlx5/core/en.h
>> >> > +++ b/drivers/net/ethernet/mellanox/mlx5/core/en.h
>> >> > @@ -768,6 +768,7 @@ struct mlx5e_channel {
>> >> >  	u16                        qos_sqs_size;
>> >> >  	u8                         num_tc;
>> >> >  	u8                         lag_port;
>> >> > +	unsigned int		   irq;
>> >> >  
>> >> >  	/* XDP_REDIRECT */
>> >> >  	struct mlx5e_xdpsq         xdpsq;
>> >> > diff --git a/drivers/net/ethernet/mellanox/mlx5/core/en_main.c b/drivers/net/ethernet/mellanox/mlx5/core/en_main.c
>> >> > index c8e8f512803e..e1bfff1fb328 100644
>> >> > --- a/drivers/net/ethernet/mellanox/mlx5/core/en_main.c
>> >> > +++ b/drivers/net/ethernet/mellanox/mlx5/core/en_main.c
>> >> > @@ -2473,6 +2473,9 @@ static void mlx5e_close_queues(struct mlx5e_channel *c)
>> >> >  	mlx5e_close_tx_cqs(c);
>> >> >  	mlx5e_close_cq(&c->icosq.cq);
>> >> >  	mlx5e_close_cq(&c->async_icosq.cq);
>> >> > +
>> >> > +	netif_queue_set_napi(c->netdev, c->ix, NETDEV_QUEUE_TYPE_TX, NULL);
>> >> > +	netif_queue_set_napi(c->netdev, c->ix, NETDEV_QUEUE_TYPE_RX, NULL);
>> >> 
>> >> This should be set to NULL *before* actually closing the rqs, sqs, and
>> >> related cqs right? I would expect these two lines to be the first ones
>> >> called in mlx5e_close_queues. Btw, I think this should be done in
>> >> mlx5e_deactivate_channel where the NAPI is disabled.
>> >> 
>> >> >  }
>> >> >  
>> >> >  static u8 mlx5e_enumerate_lag_port(struct mlx5_core_dev *mdev, int ix)
>> >> > @@ -2558,6 +2561,7 @@ static int mlx5e_open_channel(struct mlx5e_priv *priv, int ix,
>> >> >  	c->stats    = &priv->channel_stats[ix]->ch;
>> >> >  	c->aff_mask = irq_get_effective_affinity_mask(irq);
>> >> >  	c->lag_port = mlx5e_enumerate_lag_port(priv->mdev, ix);
>> >> > +	c->irq		= irq;
>> >> >  
>> >> >  	netif_napi_add(netdev, &c->napi, mlx5e_napi_poll);
>> >> >  
>> >> > @@ -2602,6 +2606,10 @@ static void mlx5e_activate_channel(struct mlx5e_channel *c)
>> >> >  		mlx5e_activate_xsk(c);
>> >> >  	else
>> >> >  		mlx5e_activate_rq(&c->rq);
>> >> > +
>> >> > +	netif_napi_set_irq(&c->napi, c->irq);
>> 
>> One small comment that I missed in my previous iteration. I think the
>> above should be moved to mlx5e_open_channel right after netif_napi_add.
>> This avoids needing to save the irq in struct mlx5e_channel.
>
> I couldn't move it to mlx5e_open_channel because of how safe_switch_params
> and the mechanics around that seem to work (at least as far as I could
> tell).
>
> mlx5 seems to create a new set of channels before closing the previous
> channel. So, moving this logic to open_channels and close_channels means
> you end up with a flow like this:
>
>   - Create new channels (NAPI netlink API is used to set NAPIs)
>   - Old channels are closed (NAPI netlink API sets NULL and overwrites the
>     previous NAPI netlink calls)
>
> Now, the associations are all NULL.
>
> I think moving the calls to active / deactivate fixes that problem, but
> requires that irq is stored, if I am understanding the driver correctly.

I believe moving the changes to activate / deactivate channels resolves
this problem because only one set of channels will be active, so you
will no longer have dangling association conflicts for the queue ->
napi. This is partially why I suggested the change in that iteration.

As for netif_napi_set_irq, that alone can be in mlx5e_open_channel (that
was the intention of my most recent comment. Not that all the other
associations should be moved as well). I agree that the other
association calls should be part of activate / deactivate channels.

>
>> >> > +	netif_queue_set_napi(c->netdev, c->ix, NETDEV_QUEUE_TYPE_TX, &c->napi);
>> >> > +	netif_queue_set_napi(c->netdev, c->ix, NETDEV_QUEUE_TYPE_RX, &c->napi);
>> >> 
>> >> It's weird that netlink queue API is being configured in
>> >> mlx5e_activate_channel and deconfigured in mlx5e_close_queues. This
>> >> leads to a problem where the napi will be falsely referred to even when
>> >> we deactivate the channels in mlx5e_switch_priv_channels and may not
>> >> necessarily get to closing the channels due to an error.
>> >> 
>> >> Typically, we use the following clean up patterns.
>> >> 
>> >> mlx5e_activate_channel -> mlx5e_deactivate_channel
>> >> mlx5e_open_queues -> mlx5e_close_queues
>> >
>> > OK, I'll move it to mlx5e_deactivate_channel before the NAPI is disabled.
>> > That makes sense to me.
>> 
>> Appreciated. Thank you for the patch btw.
>
> Sure, thanks for the review.

Powered by blists - more mailing lists