Re: [dpdk-dev] [PATCH v2 2/3] net/vmxnet3: fix vmxnet3 dev_uninit() hot-unplug

[Thomas Monjalon <thomas@monjalon.net>]

31/10/2018 18:46, Luca Boccassi:
> Sorry, been otherwise busy - I can do what you and Chas have asked, but
> the problem is that v1 already did that and the VMWare maintainers
> asked to change it back. So can I assume that the v1 way is the way to
> go?

I am expecting an answer from the vmxnet3 maintainer,
or other VMware developpers.
If they don't reply, we should work without them,
and ask for a new maintainer in the community.

I assume you can now work on a v3 reproducing what was done in v1.


> On Wed, 2018-10-31 at 18:27 +0100, Thomas Monjalon wrote:
> > Any update or question for this patch?
> > If no update, it will miss 18.11.
> > 
> > 
> > 27/10/2018 17:09, Thomas Monjalon:
> > > 19/09/2018 17:47, Chas Williams:
> > > > On Wed, Sep 19, 2018 at 8:58 AM Luca Boccassi <bluca@debian.org>
> > > > wrote:
> > > > > 
> > > > > The vmxnet3 driver can't call back into dev_close(), and
> > > > > possibly
> > > > > dev_stop(), in dev_uninit().  When dev_uninit() is called,
> > > > > anything
> > > > > that those routines would want to clean up has already been
> > > > > released.
> > > > > Further, for complete cleanup, it is necessary to release any
> > > > > of the
> > > > > queue resources during dev_close().
> > > > > This allows a vmxnet3 device to be hot-unplugged without
> > > > > leaking
> > > > > queues.
> > > > > 
> > > > > Fixes: dfaff37fc46d ("vmxnet3: import new vmxnet3 poll mode
> > > > > driver implementation")
> > > > > Cc: stable@dpdk.org
> > > > > 
> > > > > Signed-off-by: Brian Russell <brussell@brocade.com>
> > > > > Signed-off-by: Luca Boccassi <bluca@debian.org>
> > > > > ---
> > > > > v2: add back extra close() call in uninit() for buggy
> > > > > applications as
> > > > >     requested by the reviewers, and add debug log noting the
> > > > > issue
> > > > > 
> > > > >  drivers/net/vmxnet3/vmxnet3_ethdev.c | 35
> > > > > +++++++++++++++++++++++-----
> > > > >  1 file changed, 29 insertions(+), 6 deletions(-)
> > > > > 
> > > > > diff --git a/drivers/net/vmxnet3/vmxnet3_ethdev.c
> > > > > b/drivers/net/vmxnet3/vmxnet3_ethdev.c
> > > > > index f1596ab19d..98e5d01890 100644
> > > > > --- a/drivers/net/vmxnet3/vmxnet3_ethdev.c
> > > > > +++ b/drivers/net/vmxnet3/vmxnet3_ethdev.c
> > > > > @@ -354,8 +354,10 @@ eth_vmxnet3_dev_uninit(struct rte_eth_dev
> > > > > *eth_dev)
> > > > >         if (rte_eal_process_type() != RTE_PROC_PRIMARY)
> > > > >                 return 0;
> > > > 
> > > > This should probably be EPERM as well.  Out of scope though.
> > > > 
> > > > > 
> > > > > -       if (hw->adapter_stopped == 0)
> > > > > +       if (hw->adapter_stopped == 0) {
> > > > > +               PMD_INIT_LOG(DEBUG, "Device has not been
> > > > > closed.");
> > > > >                 vmxnet3_dev_close(eth_dev);
> > > > 
> > > > This just seems wrong.  You have called uninit() will the driver
> > > > is
> > > > still busy.  Instead of "fixing" the state of the driver, return
> > > > EBUSY
> > > > here.
> > > 
> > > I agree.
> > > If the port is not stopped, either you stop it or you return EBUSY.
> > > 
> > > Closing the device should be done outside of this check.
> > > It is OK to close from uninit if the app did not close it.
> > > 
> > > [...]
> > > > > +static void
> > > > > +vmxnet3_free_queues(struct rte_eth_dev *dev)
> > > > > +{
> > > > > +       int i;
> > > > > +
> > > > > +       PMD_INIT_FUNC_TRACE();
> > > > > +
> > > > > +       for (i = 0; i < dev->data->nb_rx_queues; i++) {
> > > > > +               void *rxq = dev->data->rx_queues[i];
> > > > > +
> > > > > +               vmxnet3_dev_rx_queue_release(rxq);
> > > > > +       }
> > > > > +       dev->data->nb_rx_queues = 0;
> > > > > +
> > > > > +       for (i = 0; i < dev->data->nb_tx_queues; i++) {
> > > > > +               void *txq = dev->data->tx_queues[i];
> > > > > +
> > > > > +               vmxnet3_dev_tx_queue_release(txq);
> > > > > +       }
> > > > > +       dev->data->nb_tx_queues = 0;
> > > > >  }
> > > > > 
> > > > >  /*
> > > > > @@ -844,12 +869,10 @@ vmxnet3_dev_stop(struct rte_eth_dev *dev)
> > > > >  static void
> > > > >  vmxnet3_dev_close(struct rte_eth_dev *dev)
> > > > >  {
> > > > > -       struct vmxnet3_hw *hw = dev->data->dev_private;
> > > > > -
> > > > >         PMD_INIT_FUNC_TRACE();
> > > > > 
> > > > >         vmxnet3_dev_stop(dev);
> > > > > -       hw->adapter_stopped = 1;
> > > > > +       vmxnet3_free_queues(dev);
> > > > >  }
> > > 
> > > Good clean-up on dev_close.
> > > You probably want to go further and set RTE_ETH_DEV_CLOSE_REMOVE
> > > for allowing
> > > a real release of the port on close.
> > > Note: every PMDs should migrate towards this behaviour.
> > > 
> > > To make things clear (I will write a doc for -rc2):
> > > 	- "stop" should be called by the app but the PMD is allowed to
> > > force it.
> > > 	- "close" may be called by the app, and the PMD should enforce
> > > it in uninit.
> > > 		With RTE_ETH_DEV_CLOSE_REMOVE flag, it must completely
> > > release the port.
> > > 	- "remove" (implemented in PMD as uninit) is responsible of
> > > closing
> > > 		ethdev ports if not already done, and release the
> > > shared resources
> > > 		which are not specific to a port. It removes the whole
> > > EAL rte_device.
> > > 
> > > PS: for any hotplug patch or questions, feel free to Cc me.
> > 
> > 
> > 
> > 
> > 
> 
>