Re: [dpdk-dev] [PATCH] net/vdev_netvsc: handle removal of associated pci device

[Thomas Monjalon <thomas@monjalon.net>]

Gaetan, Matan,
Please could you check?


25/09/2020 22:30, Long Li:
> HI Matan,
> 
> While troubleshooting a failure in DPDK on device removal when VF device briefly disappears and comes back, I notice the failsafe driver is trying repeatedly to start a sub device (after this sub device has been successfully configured, but later hot removed from the kernel). This is due to repeated alarms calling fs_dev_start(). I trace into this commit:
> 
> ae80146 net/failsafe: fix removed device handling
> 
> The implementation of fs_err() is interesting:
> 
> +fs_err(struct sub_device *sdev, int err)
> +{
> +       /* A device removal shouldn't be reported as an error. */
> +       if (sdev->remove == 1 || err == -EIO)
> +               return rte_errno = 0;
> +       return err;
> +}
> 
> If I change this function to:
> @@ -497,7 +497,7 @@ int failsafe_eth_new_event_callback(uint16_t port_id
>  fs_err(struct sub_device *sdev, int err)
>  {
>         /* A device removal shouldn't be reported as an error. */
> -       if (sdev->remove == 1 || err == -EIO)
> +       if (sdev->remove == 1 && err == -EIO)
>                 return rte_errno = 0;
>         return err;
>  }
> 
> The hung is going away. I don't know the reason why we use a || in the if(). If a call to rte_eth_dev_start() returning EIO (as the case in fs_dev_start), the best choice would be bail out and fail this sub device.
> 
> Can you please take a look?
> 
> Thanks,
> Long
> 
> ________________________________
> From: Matan Azrad <matan@nvidia.com>
> Sent: Tuesday, September 15, 2020 12:00 AM
> To: Long Li <longli@microsoft.com>; Stephen Hemminger <stephen@networkplumber.org>
> Cc: matan@mellanox.com <matan@mellanox.com>; grive@u246.net <grive@u246.net>; dev@dpdk.org <dev@dpdk.org>; Raslan Darawsheh <rasland@nvidia.com>
> Subject: RE: [dpdk-dev] [PATCH] net/vdev_netvsc: handle removal of associated pci device
> 
> Hi Li
> 
> From: Long Li <longli@microsoft.com>
> > >Subject: Re: [dpdk-dev] [PATCH] net/vdev_netvsc: handle removal of
> > >associated pci device
> > >
> > >Hi Stephen
> > >
> > >From: Stephen Hemminger:
> > >> On Sun, 6 Sep 2020 12:38:18 +0000
> > >> Matan Azrad <matan@nvidia.com> wrote:
> > >>
> > >> > Hi Stephen
> > >> >
> > >> > From: Stephen Hemminger:
> > >> > > The vdev_netvsc was not detecting when the associated PCI device
> > >> > > (SRIOV) was removed. Because of that it would keep feeding the
> > >> > > same
> > >> > > (removed) device to failsafe PMD which would then unsuccessfully
> > >> > > try and probe for it.
> > >> > >
> > >> > > Change to use a mark/sweep method to detect that PCI device was
> > >> > > removed, and also only tell failsafe about new PCI devices.
> > >> > > Vdev_netvsc does not have to keep stuffing the pipe with the same
> > >> > > already existing PCI device.
> > >> >
> > >> > As I know, the vdev_netvsc driver doesn't call to failsafe if the
> > >> > PCI device is
> > >> not detected by the readlink command(considered as removed)...
> > >> > Am I missing something?
> > >>
> > >> The original code is broken because ctx_yield is not cleared, it
> > >> keeps sending the same value.
> > >
> > >Looking on the code again, It looks like ctx->yield has no effect on
> > >the next pipe write, It is just used for log.
> > >
> > >After the PCI interface matching to the netvsc interface, the pipe
> > >write is triggered only if the readlink commands success to see the
> > >plugged-in PCI
> > >device:
> > >readlink /sys/class/net/[iface]/device/subsystem shows "pci"
> > >readlink /sys/class/net/[iface]/device shows the pci device ID.
> > >
> > >So, the assumption is when the above readlink failed on the interface
> > >the device is removed(plugged-out) and the fd write will not happen.
> > >
> > >The code will continue to retry probe again and again until success
> > >only for plugged-in pci device matched the netvsc device.
> >
> > Hi Matan,
> >
> > The original code keeps writing to pipe even it's the same PCI device.
> 
> Yes, the vdev_netvsc writes any plugged-in device to the associated netvsc device fd.
> 
> > The
> > new code writes to pipe for a new device, only once. See the following code:
> >
> > +     /* Skip if this is same device already sent to failsafe */
> > +     if (strcmp(addr, ctx->yield) == 0)
> > +             return 0;
> >
> 
> I understand you want to optimize the pipe writing to be written only after plugged-in hot event.
> 
> The current solution suffers from race: the PCI device may be plugged-out and plugged-in in short time shorter than the driver alarm delay, then the PCI device plugged-in detection will lost.
> 
> My suggestion:
> Add validation to the plugged-in device probing state and that it is owned by failsafe(using ownership API) - don't write the pipe if so.
> 
> Matan
> 
> 
> 
> > This patch also saves lots of CPU since it no longer writes to pipe all the time.
> > You are correct about the code will continue to probe on a new PCI device.
> > But someone has to do it to handle hot-add.
> >
> > Thanks,
> > Long
> >
> >
> > >
> > >> It looks like device removal and add was never tested.
> > >
> > >This is basic test we have to test plug-in plug-out and it passed every
> > >day in the last years.
> > >
> > >Maybe something new and special in your setup?
> > >
> > >> If you test removal you will see that vdev_netvsc:
> > >>  1. Sends same PCI device repeatedly to failsafe (every alarm call)
> > >>     This is harmless, but useless.
> > >>  2. When device is removed, keeps doing #1
>