netdev - Re: [PATCH net-next 05/18] gve: Add Gvnic stats AQ command and ethtool show/set-priv-flags.

lists.openwall.net		lists / announce owl-users owl-dev john-users john-dev passwdqc-users yescrypt popa3d-users / oss-security kernel-hardening musl sabotage tlsify passwords / crypt-dev xvendor / Bugtraq Full-Disclosure linux-kernel linux-netdev linux-ext4 linux-hardening linux-cve-announce PHC
Open Source and information security mailing list archives
Hash Suite: Windows password security audit tool. GUI, reports in PDF.
[<prev] [next>] [<thread-prev] [thread-next>] [day] [month] [year] [list]
Message-ID: <CAL9ddJcDYcn+p33nKicmp7yHm6PnZ9iXnghO4AYHNmtCFCe2eQ@mail.gmail.com>
Date:   Tue, 25 Aug 2020 08:46:12 -0700
From:   David Awogbemila <awogbemila@...gle.com>
To:     Jakub Kicinski <kuba@...nel.org>
Cc:     netdev@...r.kernel.org, Kuo Zhao <kuozhao@...gle.com>,
        Yangchun Fu <yangchun@...gle.com>
Subject: Re: [PATCH net-next 05/18] gve: Add Gvnic stats AQ command and
 ethtool show/set-priv-flags.

On Tue, Aug 18, 2020 at 8:13 PM Jakub Kicinski <kuba@...nel.org> wrote:
>
> On Tue, 18 Aug 2020 12:44:04 -0700 David Awogbemila wrote:
> > From: Kuo Zhao <kuozhao@...gle.com>
> >
> > Changes:
> > - Add a new flag in service_task_flags. Check both this flag and
> > ethtool flag when handle report stats. Update the stats when user turns
> > ethtool flag on.
> >
> > - In order to expose the NIC stats to the guest even when the ethtool flag
> > is off, share the address and length of report at setup. When the
> > ethtool flag turned off, zero off the gve stats instead of detaching the
> > report. Only detach the report in free_stats_report.
> >
> > - Adds the NIC stats to ethtool stats. These stats are always
> > exposed to guest no matter the report stats flag is turned
> > on or off.
> >
> > - Update gve stats once every 20 seconds.
> >
> > - Add a field for the interval of updating stats report to the AQ
> > command. It will be exposed to USPS so that USPS can use the same
> > interval to update its stats in the report.
> >
> > Reviewed-by: Yangchun Fu <yangchun@...gle.com>
> > Signed-off-by: Kuo Zhao <kuozhao@...gle.com>
> > Signed-off-by: David Awogbemila <awogbemila@...gle.com>
>
> This patch is quite hard to parse, please work on improving its
> readability. Perhaps start by splitting changes to the stats from
> hypervisor from the stats to hypervisor.
Alright, I will split the patch as suggested.

>
> > +enum gve_stat_names {
> > +     // stats from gve
> > +     TX_WAKE_CNT                     = 1,
> > +     TX_STOP_CNT                     = 2,
> > +     TX_FRAMES_SENT                  = 3,
> > +     TX_BYTES_SENT                   = 4,
> > +     TX_LAST_COMPLETION_PROCESSED    = 5,
> > +     RX_NEXT_EXPECTED_SEQUENCE       = 6,
> > +     RX_BUFFERS_POSTED               = 7,
>
> Just out of curiosity - what's the use for the stats reported by VM to
> the hypervisor?
These stats are not used in the driver but are useful when looking at
the virtual NIC to investigate stuck queues and performance.

>
> > +     // stats from NIC
> > +     RX_QUEUE_DROP_CNT               = 65,
> > +     RX_NO_BUFFERS_POSTED            = 66,
> > +     RX_DROPS_PACKET_OVER_MRU        = 67,
> > +     RX_DROPS_INVALID_CHECKSUM       = 68,
>
> Most of these look like a perfect match for members of struct
> rtnl_link_stats64. Please use the standard stats to report the errors,
> wherever possible.
These stats are based on the NIC stats format which don't exactly
match rtnl_link_stats64.
I'll add some clarification in the description and within the comments.

>
> > +};
> > +
> >  union gve_adminq_command {
> >       struct {
> >               __be32 opcode;
>
> > +static int gve_set_priv_flags(struct net_device *netdev, u32 flags)
> > +{
> > +     struct gve_priv *priv = netdev_priv(netdev);
> > +     u64 ori_flags, new_flags;
> > +     u32 i;
> > +
> > +     ori_flags = READ_ONCE(priv->ethtool_flags);
> > +     new_flags = ori_flags;
> > +
> > +     for (i = 0; i < GVE_PRIV_FLAGS_STR_LEN; i++) {
> > +             if (flags & BIT(i))
> > +                     new_flags |= BIT(i);
> > +             else
> > +                     new_flags &= ~(BIT(i));
> > +             priv->ethtool_flags = new_flags;
> > +             /* set report-stats */
> > +             if (strcmp(gve_gstrings_priv_flags[i], "report-stats") == 0) {
> > +                     /* update the stats when user turns report-stats on */
> > +                     if (flags & BIT(i))
> > +                             gve_handle_report_stats(priv);
> > +                     /* zero off gve stats when report-stats turned off */
> > +                     if (!(flags & BIT(i)) && (ori_flags & BIT(i))) {
> > +                             int tx_stats_num = GVE_TX_STATS_REPORT_NUM *
> > +                                     priv->tx_cfg.num_queues;
> > +                             int rx_stats_num = GVE_RX_STATS_REPORT_NUM *
> > +                                     priv->rx_cfg.num_queues;
> > +                             memset(priv->stats_report->stats, 0,
> > +                                    (tx_stats_num + rx_stats_num) *
> > +                                    sizeof(struct stats));
>
> I don't quite get why you need the knob to disable some statistics.
> Please remove or explain this in the cover letter. Looks unnecessary.
We use this to give the guest the option of disabling stats reporting
through ethtool set-priv-flags. I'll update the cover letter.

>
> > +                     }
> > +             }
> > +     }
> > +
> > +     return 0;
> > +}
>
> > @@ -880,6 +953,10 @@ static void gve_handle_status(struct gve_priv *priv, u32 status)
> >               dev_info(&priv->pdev->dev, "Device requested reset.\n");
> >               gve_set_do_reset(priv);
> >       }
> > +     if (GVE_DEVICE_STATUS_REPORT_STATS_MASK & status) {
> > +             dev_info(&priv->pdev->dev, "Device report stats on.\n");
>
> How often is this printed?
Stats reporting is disabled by default. But when enabled, this would
only get printed whenever the virtual NIC detects
an issue and triggers a report-stats request.