[dpdk-dev] [PATCH] ethdev: fix device state on close

Shahaf Shuler shahafs at mellanox.com
Wed Aug 16 16:17:16 CEST 2017


Wednesday, August 16, 2017 3:42 PM, Gaëtan Rivet:
> On Wed, Aug 16, 2017 at 02:43:08PM +0300, Shahaf Shuler wrote:
> > Currently device state moves between ATTACHED when device was
> > successfully probed to UNUSED when device is detached or released.
> >
> > The device state following rte_eth_dev_close() operation is inconsist,
> > The device is still in ATTACHED state, however it cannot be used in
> > any way till it will be probed again.
> >
> > Fixing it by changing the state to UNUSED.
> >
> 
> You are right that simply closing the device leaves it in a unusable state.
> 
> However it seems to be by design.
> Most drivers call `rte_eth_dev_release_port` when being removed, which
> sets the state to RTE_ETH_DEV_UNUSED.
> 
> If I'm not mistaken, the API of rte_eth_dev_close is that the only available
> action should then be to detach the driver. At least PCI and vdev buses
> expects a `remove` callback from their driver, which can be called by the user
> (previously using specific API like `rte_eal_vdev_uninit` for example, now
> using `rte_eal_hotplug_remove` or `rte_eth_dev_detach` from the ether
> layer).
> 
> So, it seems that this burden lies with the driver which should call the proper
> API when removing their device.

Even though it is reasonable for driver to call the rte_eth_dev_port_release, I still think the ethdev layer should protect from such bad behavior from the application side.
It is more robust than counting on the different PMD to implement such release. 

> 
> Maybe Thomas will have a better insight about the scope of the
> `rte_eth_dev_close` function. But IMO the API is respected.
> After all, until the proper `dev_detach` function is called, the device is still
> attached, even if closed.
> 
> If you disagree, there might possibly be an argument to make about either
> adding finer-grained device states or streamlining the API. This is however a
> discussion about API design and not about its implementation anymore.

Well my first thought when I created this patch was to add fine-grained device states. However then I read the commit log which changed the device states, specifically :

" "DEV_DETACHED" is renamed "RTE_ETH_DEV_UNUSED" to better reflect that
the emptiness of a slot is not necessarily the result of detaching a
device."

Which convince me to reuse the UNUSED state to reflect that this device cannot longer be used (even though it is still attached). 

> 
> > Fixes: d52268a8b24b ("ethdev: expose device states")
> > Cc: gaetan.rivet at 6wind.com
> > Cc: stable at dpdk.org
> >
> > Signed-off-by: Shahaf Shuler <shahafs at mellanox.com>
> > ---
> >  lib/librte_ether/rte_ethdev.c | 2 ++
> >  1 file changed, 2 insertions(+)
> >
> > diff --git a/lib/librte_ether/rte_ethdev.c
> > b/lib/librte_ether/rte_ethdev.c index 0597641ee..98d9e929c 100644
> > --- a/lib/librte_ether/rte_ethdev.c
> > +++ b/lib/librte_ether/rte_ethdev.c
> > @@ -992,6 +992,8 @@ rte_eth_dev_close(uint8_t port_id)
> >  	dev->data->nb_tx_queues = 0;
> >  	rte_free(dev->data->tx_queues);
> >  	dev->data->tx_queues = NULL;
> > +
> > +	dev->state = RTE_ETH_DEV_UNUSED;
> >  }
> >
> >  int
> > --
> > 2.12.0
> >
> 
> --
> Gaëtan Rivet
> 6WIND


More information about the dev mailing list