LKML Archive on lore.kernel.org
help / color / mirror / Atom feed
From: Saeed Mahameed <saeedm@dev.mellanox.co.il>
To: Qing Huang <qing.huang@oracle.com>
Cc: Eran Ben Elisha <eranbe@mellanox.com>,
	linux-kernel <linux-kernel@vger.kernel.org>,
	RDMA mailing list <linux-rdma@vger.kernel.org>,
	Linux Netdev List <netdev@vger.kernel.org>,
	Leon Romanovsky <leon@kernel.org>,
	Matan Barak <matanb@mellanox.com>,
	Saeed Mahameed <saeedm@mellanox.com>
Subject: Re: [PATCH] net/mlx5: report persistent netdev stats across ifdown/ifup commands
Date: Thu, 26 Apr 2018 16:43:32 -0700	[thread overview]
Message-ID: <CALzJLG_ROqv1Kby1FbGXFSgB5My4Gj=3cubw=DEts_9_3V3=3A@mail.gmail.com> (raw)
In-Reply-To: <CALzJLG8atGxoJNQj4F6wR3kRg7jrj4oQi3O-1i0=UOo+3HPq0w@mail.gmail.com>

On Thu, Apr 26, 2018 at 4:30 PM, Saeed Mahameed
<saeedm@dev.mellanox.co.il> wrote:
> On Thu, Apr 26, 2018 at 3:37 PM, Qing Huang <qing.huang@oracle.com> wrote:
>>
>>
>> On 04/26/2018 02:50 PM, Saeed Mahameed wrote:
>>>
>>> On Thu, Apr 26, 2018 at 1:37 PM, Qing Huang <qing.huang@oracle.com> wrote:
>>>>
>>>> Current stats collecting scheme in mlx5 driver is to periodically fetch
>>>> aggregated stats from all the active mlx5 software channels associated
>>>> with the device. However when a mlx5 interface is brought down(ifdown),
>>>> all the channels will be deactivated and closed. A new set of channels
>>>> will be created when next ifup command or a similar command is called.
>>>> Unfortunately the new channels will have all stats reset to 0. So you
>>>> lose the accumulated stats information. This behavior is different from
>>>> other netdev drivers including the mlx4 driver. In order to fix it, we
>>>> now save prior mlx5 software stats into netdev stats fields, so all the
>>>> accumulated stats will survive multiple runs of ifdown/ifup commands and
>>>> be shown correctly.
>>>>
>>>> Orabug: 27548610
>
> I don't think we need this internal bug reference, let's remove it.
>
>>>>
>>>> Signed-off-by: Qing Huang <qing.huang@oracle.com>
>>>> ---
>>>
>>> Hi Qing,
>>>
>>> I am adding Eran since he is currently working on a similar patch,
>>> He is also taking care of all cores/rings stats to make them
>>> persistent, so you won't have discrepancy  between
>>> ethtool and ifconfig stats.
>>>
>>> I am ok with this patch, but this means Eran has to work his way around
>>> it.
>>>
>>> so we have two options:
>>>
>>> 1. Temporary accept this patch, and change it later with Eran's work.
>>> 2. Wait for Eran's work.
>>>
>>> I am ok with either one of them, please let me know.
>>>
>>> Thanks !
>>
>> Hi Saeed,
>>
>> Any idea on rough ETA of Eran's stats work to be in upstream? If it will be
>
> All i know that it is planned for v4.18.
>
>> available soon, I think
>> we can wait a bit. If it will take a while to redesign the whole stats
>> scheme (for both ethtool and ifconfig),
>> maybe we can go with this incremental fix first?
>>

Qing,

Before you address my comments, it looks like Eran's work is
converging and we will finalize the internal
review next week, in which case submission will take 2-3 weeks from today.

So i suggest to wait.

please checkout:
https://git.kernel.org/pub/scm/linux/kernel/git/saeed/linux.git/commit/?h=testing/mlx5e-stats&id=15ffb7f87432d073e8ac0e6b895188d40fdda4d4

>
> Yes we can go with incremental fix,
> but let's fix the commit message as requested above and make the patch
> title prefix "net/mlx5e:"
> since this is a mlx5 netdev related change, not mlx5/core.
> also please see a comments below.
>
> Thanks,
> Saeed.
>
>> Thanks!
>>
>>
>>>
>>>>   drivers/net/ethernet/mellanox/mlx5/core/en_main.c | 30
>>>> +++++++++++++++++++----
>>>>   1 file changed, 25 insertions(+), 5 deletions(-)
>>>>
>>>> diff --git a/drivers/net/ethernet/mellanox/mlx5/core/en_main.c
>>>> b/drivers/net/ethernet/mellanox/mlx5/core/en_main.c
>>>> index f1fe490..5d50e69 100644
>>>> --- a/drivers/net/ethernet/mellanox/mlx5/core/en_main.c
>>>> +++ b/drivers/net/ethernet/mellanox/mlx5/core/en_main.c
>>>> @@ -2621,6 +2621,23 @@ static void mlx5e_netdev_set_tcs(struct net_device
>>>> *netdev)
>>>>                  netdev_set_tc_queue(netdev, tc, nch, 0);
>>>>   }
>>>>
>>>> +static void mlx5e_netdev_save_stats(struct mlx5e_priv *priv)
>>>> +{
>>>> +       struct net_device *netdev = priv->netdev;
>>>> +
>>>> +       netdev->stats.rx_packets += priv->stats.sw.rx_packets;
>>>> +       netdev->stats.rx_bytes   += priv->stats.sw.rx_bytes;
>>>> +       netdev->stats.tx_packets += priv->stats.sw.tx_packets;
>>>> +       netdev->stats.tx_bytes   += priv->stats.sw.tx_bytes;
>>>> +       netdev->stats.tx_dropped += priv->stats.sw.tx_queue_dropped;
>>>> +
>>>> +       priv->stats.sw.rx_packets       = 0;
>>>> +       priv->stats.sw.rx_bytes         = 0;
>>>> +       priv->stats.sw.tx_packets       = 0;
>>>> +       priv->stats.sw.tx_bytes         = 0;
>>>> +       priv->stats.sw.tx_queue_dropped = 0;
>
> Note: it is safe to do this here since writing to priv->stats.sw is
> currently done under priv->state_lock,
> Please add a comment to this function that it must be called under
> priv->state_lock
>
>>>> +}
>>>> +
>>>
>>> This means that we are now explicitly clearing channels stats on
>>> ifconfig down or switch_channels.
>>> and now after ifconfing down, ethtool will always show 0, before this
>>> patch it didn't.
>>> Anyway update sw stats function will always override them with the new
>>> channels stats next time we load new channels.
>>> so it is not that big of a deal.
>>>
>>>
>>>>   static void mlx5e_build_channels_tx_maps(struct mlx5e_priv *priv)
>>>>   {
>>>>          struct mlx5e_channel *c;
>>>> @@ -2691,6 +2708,7 @@ void mlx5e_switch_priv_channels(struct mlx5e_priv
>>>> *priv,
>>>>                  netif_set_real_num_tx_queues(netdev, new_num_txqs);
>>>>
>>>>          mlx5e_deactivate_priv_channels(priv);
>>>> +       mlx5e_netdev_save_stats(priv);
>
> This function call doesn't have to be between deactivate and close,
> lets move it to after
> mlx5e_close_channels here and below to avoid confusion.
>
> also you need to accumulate what is left in the rings/channels stats,
> before calling
> mlx5e_close_channels, see mlx5e_grp_sw_update_stats.
>
>>>>          mlx5e_close_channels(&priv->channels);
>>>>
>>>>          priv->channels = *new_chs;
>>>> @@ -2770,6 +2788,7 @@ int mlx5e_close_locked(struct net_device *netdev)
>>>>
>>>>          netif_carrier_off(priv->netdev);
>>>>          mlx5e_deactivate_priv_channels(priv);
>>>> +       mlx5e_netdev_save_stats(priv);
>>>>          mlx5e_close_channels(&priv->channels);
>>>>
>>>>          return 0;
>>>> @@ -3215,11 +3234,12 @@ static int mlx5e_setup_tc(struct net_device *dev,
>>>> enum tc_setup_type type,
>>>>                  stats->tx_packets = PPORT_802_3_GET(pstats,
>>>> a_frames_transmitted_ok);
>>>>                  stats->tx_bytes   = PPORT_802_3_GET(pstats,
>>>> a_octets_transmitted_ok);
>>>>          } else {
>>>> -               stats->rx_packets = sstats->rx_packets;
>>>> -               stats->rx_bytes   = sstats->rx_bytes;
>>>> -               stats->tx_packets = sstats->tx_packets;
>>>> -               stats->tx_bytes   = sstats->tx_bytes;
>>>> -               stats->tx_dropped = sstats->tx_queue_dropped;
>>>> +               stats->rx_packets = sstats->rx_packets +
>>>> dev->stats.rx_packets;
>>>> +               stats->rx_bytes   = sstats->rx_bytes +
>>>> dev->stats.rx_bytes;
>>>> +               stats->tx_packets = sstats->tx_packets +
>>>> dev->stats.tx_packets;
>>>> +               stats->tx_bytes   = sstats->tx_bytes +
>>>> dev->stats.tx_bytes;
>>>> +               stats->tx_dropped = sstats->tx_queue_dropped +
>>>> +                                   dev->stats.tx_dropped;
>>>>          }
>>>>
>>>>          stats->rx_dropped = priv->stats.qcnt.rx_out_of_buffer;
>>>> --
>>>> 1.8.3.1
>>>>
>>

  reply	other threads:[~2018-04-26 23:43 UTC|newest]

Thread overview: 6+ messages / expand[flat|nested]  mbox.gz  Atom feed  top
2018-04-26 20:37 Qing Huang
2018-04-26 21:50 ` Saeed Mahameed
2018-04-26 22:37   ` Qing Huang
2018-04-26 23:30     ` Saeed Mahameed
2018-04-26 23:43       ` Saeed Mahameed [this message]
2018-04-27  0:21         ` Qing Huang

Reply instructions:

You may reply publicly to this message via plain-text email
using any one of the following methods:

* Save the following mbox file, import it into your mail client,
  and reply-to-all from there: mbox

  Avoid top-posting and favor interleaved quoting:
  https://en.wikipedia.org/wiki/Posting_style#Interleaved_style

* Reply using the --to, --cc, and --in-reply-to
  switches of git-send-email(1):

  git send-email \
    --in-reply-to='CALzJLG_ROqv1Kby1FbGXFSgB5My4Gj=3cubw=DEts_9_3V3=3A@mail.gmail.com' \
    --to=saeedm@dev.mellanox.co.il \
    --cc=eranbe@mellanox.com \
    --cc=leon@kernel.org \
    --cc=linux-kernel@vger.kernel.org \
    --cc=linux-rdma@vger.kernel.org \
    --cc=matanb@mellanox.com \
    --cc=netdev@vger.kernel.org \
    --cc=qing.huang@oracle.com \
    --cc=saeedm@mellanox.com \
    --subject='Re: [PATCH] net/mlx5: report persistent netdev stats across ifdown/ifup commands' \
    /path/to/YOUR_REPLY

  https://kernel.org/pub/software/scm/git/docs/git-send-email.html

* If your mail client supports setting the In-Reply-To header
  via mailto: links, try the mailto: link

This is a public inbox, see mirroring instructions
for how to clone and mirror all data and code used for this inbox;
as well as URLs for NNTP newsgroup(s).