Re: [dpdk-dev] [RFC PATCH v2 3/5] librte_ether: add API's for VF management

[Thomas Monjalon <thomas.monjalon@6wind.com>]

2016-09-23 17:02, Iremonger, Bernard:
> From: Thomas Monjalon [mailto:thomas.monjalon@6wind.com]
> > 2016-09-23 09:53, Richardson, Bruce:
> > > From: Thomas Monjalon [mailto:thomas.monjalon@6wind.com]
> > > > 2016-09-23 10:20, Bruce Richardson:
> > > > > On Thu, Sep 22, 2016 at 07:04:37PM +0200, Thomas Monjalon wrote:
> > > > > > 2016-09-15 16:46, Iremonger, Bernard:
> > > > > > > > > > Do we really need to expose VF specific functions here?
> > > > > > > > > > It can be generic(PF/VF) function indexed only through
> > > > port_id.
> > > > > > > > > > (example: as rte_eth_dev_set_vlan_anti_spoof(uint8_t
> > > > > > > > > > port_id, uint8_t on)) For instance, In Thunderx PMD, We
> > > > > > > > > > are not exposing a separate port_id for PF. We only
> > > > > > > > > > enumerate 0..N VFs as 0..N ethdev port_id
> > > > > > > > >
> > > > > > > > > Our intention with this patch is to control the VF from the PF.
> > > > > > > > >
> > > > > > > > > The following librte_ether functions already work in a
> > > > > > > > > similar
> > > > way:
> > > > > > > > >
> > > > > > > > > rte_eth_dev_set_vf_rxmode(uint8_t port_id,  uint16_t vf,
> > > > > > > > > uint16_t rx_mode, uint8_t on)
> > > > > > > > >
> > > > > > > > > rte_eth_dev_set_vf_rx(uint8_t port_id, uint16_t vf,
> > > > > > > > > uint8_t
> > > > > > > > > on)
> > > > > > > > >
> > > > > > > > > rte_eth_dev_set_vf_tx(uint8_t port_id, uint16_t vf,
> > > > > > > > > uint8_t
> > > > > > > > > on)
> > > > > > > > >
> > > > > > > > > int rte_eth_set_vf_rate_limit(uint8_t port_id, uint16_t
> > > > > > > > > vf, uint16_t tx_rate, uint64_t q_msk)
> > > > > > > >
> > > > > > > > I have a bad feeling with these functions dedicated to VF from PF.
> > > > > > > > Are we sure there is no other way?
> > > > > > > > I mean we just need to know the VF with a port ID.
> > > > > > >
> > > > > > > When the VF is used in a VM the port ID of the VF is not
> > > > > > > visible to
> > > > the PF.
> > > > > > > I don't think there is another way to do this.
> > > > > >
> > > > > > I don't understand why we could not assign a port id to the VF
> > > > > > from the host instead of having the couple PF port id / VF id.
> > > > > > Can we enumerate all the VFs associated to a PF?
> > > > > > Then can we allocate them a port id in the array rte_eth_devices?
> > > > >
> > > > > Hi Thomas,
> > > > >
> > > > > The VF is not a port visible to DPDK, though, so it shouldn't have
> > > > > a port id IMHO. DPDK can't actually do anything with it.
> > > >
> > > > You say the contrary below.
> > >
> > > Well, yes and no. The driver can manipulate things for the VF, but DPDK
> > doesn't actually have a device that corresponds to the VF. There are no PCI
> > bar mappings for it, DPDK can't do RX and TX with it etc.?
> > 
> > Very good point.
> > There are only few ethdev functions which are supported by every drivers,
> > like Rx/Tx and would not be available for VF from PF interface.
> > 
> > > > > The PCI device for the VF is likely passed through to a different
> > > > > VM and being used there. Unfortunately, the VF still needs certain
> > > > > things done for it by the PF, so if the PF is under DPDK control,
> > > > > it needs to provide the functionality to assist the VF.
> > > >
> > > > Why not have a VF_from_PF driver which does the mailbox things?
> > > > So you can manage the VF from the PF with a simple port id.
> > > > It really seems to be the cleanest design to me.
> > >
> > > While I see your point, and it could work, I just want to be sure that we are
> > ok with the results of that. Suppose we do create ethdevs for the VFs
> > controlled by the PF. Does the new VF get counted in the
> > rte_eth_dev_count() value (I assume yes)? How are apps meant to use the
> > port? Do they have to put in a special case when iterating through all the port
> > ids to check that it's not a pseudo port that can't do anything. None of the
> > standard ethdev calls from an app will work on it, you can't configure nb rx/tx
> > queues on it, you can't start or stop it, you can't do rx or tx on it, etc, etc.
> > 
> > Yes these devices would be special because their supported API would be
> > quite different. I was thinking that in the future you could add most of the
> > configuration functions through the VF mailbox.
> > But the Intel mailbox currently support only some special configurations
> > which are not supported by other devices even its own VF device (except
> > setting MAC address).
> > And when I read "set drop enable bit in the VF split rx control register", it
> > becomes clear it is really specific and has nothing to do in the generic ethdev
> > API.
> > That's why it is a NACK.
> > 
> > When we want to use these very specific features we are aware of the
> > underlying device and driver. So we can directly include a header from the
> > driver. I suggest to retrieve a handler for the device which is not a port id and
> > will allow to call ixgbe functions directly.
> > It could be achieved by adding an ethdev function like discussed here:
> > 	http://dpdk.org/ml/archives/dev/2016-September/047392.html
> > 
> 
> I have been reading the net/vhost mail thread above. The following quote is from this thread.
> 
> "It means I would be in favor of introducing API in drivers for very specific features."
> 
> At present all the PMD functions are accessed through the eth_dev_ops structure, there are no PMD API's.
> 
> Is your proposal to add API(s) to the DPDK ixgbe PMD (similar to a driver ioctl API) which can be accessed through a generic API in the ethdev?

Not exactly. I'm thinking about a PMD specific API.
The only ethdev API you need would be a function to retrieve a handler
(an opaque pointer on the device struct) from the port id.
Then you can include rte_ixgbe.h and directly call the specific ixgbe
function, passing the device handler.
How does it sound?