On Tue, May 22, 2018 at 10:19:14AM +0000, Matan Azrad wrote:
> Hi Gaetan
>
> From: Gaëtan Rivet
> > Hello Matan,
> >
> > On Mon, May 21, 2018 at 07:48:03PM +0000, Matan Azrad wrote:
> > > The fail-safe PMD registers to RMV event for each removable sub-device
> > > port in order to cleanup the sub-device resources and switch the Tx
> > > sub-device directly when it is plugged-out.
> > >
> > > During removal time, the fail-safe PMD stops and closes the sub-device
> > > but it doesn't unregister the LSC and RMV callbacks of the sub-device
> > > port.
> > >
> > > It can lead the callbacks to be called for a port which is no more
> > > associated with the fail-safe sub-device, because there is not a
> > > guarantee that a sub-device gets the same port ID for each plug-in
> > > process. This port, for example, may belong to another sub-device of a
> > > different fail-safe device.
> > >
> > > Unregister the LSC and RMV callbacks for sub-devices which are not
> > > used.
> > >
> > > Fixes: 598fb8aec6f6 ("net/failsafe: support device removal")
> > > Cc: sta...@dpdk.org
> > >
> > > Signed-off-by: Matan Azrad <ma...@mellanox.com>
> > > ---
> > > drivers/net/failsafe/failsafe_ether.c | 22 ++++++++++++++++++++++
> > > drivers/net/failsafe/failsafe_ops.c | 5 +++++
> > > drivers/net/failsafe/failsafe_private.h | 5 +++++
> > > 3 files changed, 32 insertions(+)
> > >
> > > V2:
> > > Improve the commit log and add code comments for the new sub-dev fields
> > (Ophir suggestion).
> > >
> > >
> > > diff --git a/drivers/net/failsafe/failsafe_ether.c
> > > b/drivers/net/failsafe/failsafe_ether.c
> > > index 733e95d..2bbee82 100644
> > > --- a/drivers/net/failsafe/failsafe_ether.c
> > > +++ b/drivers/net/failsafe/failsafe_ether.c
> > > @@ -260,6 +260,7 @@
> > > sdev->state = DEV_ACTIVE;
> > > /* fallthrough */
> > > case DEV_ACTIVE:
> > > + failsafe_eth_dev_unregister_callbacks(sdev);
> > > rte_eth_dev_close(PORT_ID(sdev));
> > > sdev->state = DEV_PROBED;
> > > /* fallthrough */
> > > @@ -321,6 +322,27 @@
> > > }
> > >
> > > void
> > > +failsafe_eth_dev_unregister_callbacks(struct sub_device *sdev) {
> > > + if (sdev == NULL)
> > > + return;
> > > + if (sdev->rmv_callback) {
> > > + rte_eth_dev_callback_unregister(PORT_ID(sdev),
> > > + RTE_ETH_EVENT_INTR_RMV,
> > > + failsafe_eth_rmv_event_callback,
> > > + sdev);
> > > + sdev->rmv_callback = 0;
> >
> > I agree with Ophir here, either the return value should not be ignored, and
> > rmv_callback should only be set to 0 on success, or a proper justification
> > (and
> > an accompanying comment) should be given.
> >
> > The issue I could see is that even on error, there won't be a process to
> > try again
> > unregistering the callback.
> >
> > Maybe this could be added in failsafe_dev_remove()? Something like
> >
> > FOREACH_SUBDEV(sdev, i, dev) {
> > if (sdev->rmv_callback && sdev->state <= DEV_PROBED)
> > if (rte_eth_dev_callback_unregister(...) == 0)
> > sdev->rmv_callback = 0;
> > /* same for lsc_callback */
> > }
> >
> > Does it make sense to you? Do you think this is necessary, or should we
> > ignore
> > this?
>
> The RMV\LSC event callbacks are called from the host thread and also the
> removal process is running from the host thread so I think EAGAIN is not
> expected in the removal time.
> Other error (EINVAL) may return again every attempt and probably points to
> another critical issue.
>
> Is a code comment for the above enough? Or you think we still need to check
> it?
>
>
Ok, that makes sense.
If EINVAL is possible however, I think a warning would be helpful for
the user to be aware of the issue. The callback flag would then be
meaningless anyway.
--
Gaëtan Rivet
6WIND
